AI-Driven Predictive Maintenance in Mining: A Systematic Literature Review on Fault Detection, Digital Twins, and Intelligent Asset Management

Abstract

The mining industry faces increasing challenges in maintaining high production levels while minimizing unplanned failures and operational costs. Critical assets, such as crushers, conveyor belts, mills, and ventilation systems, operate under extreme conditions, leading to accelerated wear and failure risks. Traditional maintenance strategies often fail to prevent unexpected downtimes, safety hazards, and economic losses. As a response, industries are integrating predictive monitoring technologies, including machine learning, the Internet of Things, and digital twins, to enhance early fault detection and optimize maintenance strategies. This Systematic Literature Review analyzes 166 high-impact studies from Scopus and Web of Science, identifying key trends in fault detection algorithms, hybrid AI models, and real-time monitoring techniques. The findings highlight the increasing adoption of deep learning, reinforcement learning, and digital twins for anomaly detection and process optimization. Additionally, AI-driven methods are improving sensor-based data acquisition and asset management, extending equipment lifecycles while reducing failures. Despite these advancements, challenges such as data standardization, model scalability, and system interoperability persist, requiring further research. Future work should focus on real-time AI applications, explainable models, and academia-industry collaboration to accelerate the implementation of intelligent maintenance solutions, ensuring greater reliability, efficiency, and sustainability in mining operations.

1. Introduction

The mining sector is navigating an increasingly complex environment characterized by the dual challenge of maintaining high production levels and minimizing unplanned disruptions while simultaneously reducing operational costs. Critical assets, such as crushers, conveyor belts, mills, and ventilation systems, are essential for sustaining operations. However, these assets often function under harsh environmental conditions, including extreme temperatures, high humidity, and abrasive materials. These factors significantly accelerate wear and tear, increasing the likelihood of unforeseen failures. Such failures not only pose risks to operational continuity but also contribute to increased safety hazards and higher maintenance costs. In this context, industries and researchers are focusing on developing innovative monitoring and maintenance solutions that enable early fault detection, improved safety protocols, and a reduction in downtime. This proactive approach is becoming essential for maintaining competitiveness in the global market [1,2,3,4].

To address these challenges, intelligent technologies have emerged as key enablers of predictive asset management in the mining industry. Advances in artificial intelligence, deep learning algorithms, and next-generation sensors have significantly improved the ability to process vast amounts of operational data. These technologies are particularly effective in identifying anomalous behavior patterns, allowing operators to detect potential failures before they occur. As a result, predictive maintenance strategies are becoming more sophisticated, helping to optimize asset lifecycles, reduce operational expenses, and minimize unexpected downtime. However, despite their potential, the implementation of these technologies in mining remains complex. Highly variable operational conditions, the diversity of asset types, and the technical limitations of current predictive models present significant barriers to large-scale adoption. Overcoming these challenges requires a structured, systematic evaluation of existing methodologies to refine predictive techniques, establish best practices, and ensure the seamless integration of AI-driven maintenance solutions in mining operations [5,6].

Emerging research in advanced mathematical modeling emphasizes new approaches to represent harsh conditions in open-pit mines [7,8], demonstrating the complexity of capturing nonstationary behaviors. Integrating these mathematical frameworks with AI-based analytics is challenging but essential [9,10], as it paves the way for more accurate fault detection and longer equipment lifespans in mining.

Despite these promising developments, several critical research gaps persist in the domain of predictive maintenance for mining. Chief among them is the scarcity of robust, standardized datasets from diverse operational settings, which limits the validation and scalability of existing algorithms. Additionally, many proposed methods remain confined to controlled or simulation-based studies, lacking large-scale industrial demonstrations under harsh environmental conditions. Interoperability challenges, arising from heterogeneous sensor technologies and incompatible data formats, further hinder the seamless adoption of predictive analytics across different stages of mining operations. Consequently, there is a pressing need for more comprehensive, real-world evaluations and collaborative frameworks that bridge these gaps, ensuring that advanced monitoring solutions can be effectively implemented and sustained in practice.

In order to address this gap, this study conducts a Systematic Literature Review (SLR) to analyze recent advancements in AI-driven predictive monitoring for mining applications. By systematically reviewing 166 high-impact studies from Scopus and Web of Science, this research identifies the most effective methodologies for fault detection, predictive modeling, and intelligent asset management in extreme environmental conditions. The review also explores key trends, including the growing adoption of hybrid AI models, digital twins, and IoT-enhanced monitoring, while highlighting critical research challenges such as data standardization, model scalability, and real-time implementation [11,12,13]. By synthesizing these contributions, this study provides a structured framework for advancing predictive maintenance strategies in the mining industry, offering practical insights to guide future research and industrial applications. However, the lack of standardized data repositories continues to hinder model validation under real-world industrial conditions, evidencing the gap between theoretical advancements and practical implementation [14,15,16].

Given the extensive and heterogeneous nature of the mining domain, a semi-automated search and filtering strategy was adopted instead of relying exclusively on a traditional PRISMA framework. Specifically, Natural Language Processing (NLP) techniques and bibliometric analyses were combined with expert-driven evaluation to identify and refine the final corpus of high-impact studies. Through this flexible approach, the broad exploration typical of systematic reviews is retained, while a targeted curation of literature most relevant to fault detection and predictive maintenance in the mining sector is facilitated, as described in the following sections.

This analysis is guided by five central research questions that address critical needs in the mining industry:

(1)

Which AI- and physics-based methodologies are most effective for early fault detection in mining environments, and how do they compare in terms of performance and reliability?

(2)

How do digital twins and advanced monitoring strategies (e.g., IoT and Big Data) enhance the efficiency and robustness of predictive maintenance?

(3)

How are predictive maintenance methods being applied to improve energy efficiency and sustainability in industrial mining processes?

(4)

What barriers, challenges, and opportunities influence the adoption of predictive monitoring systems—particularly regarding data quality, scalability, and industrial deployment?

(5)

Which research gaps persist in the use of hybrid (physics- and data-driven) models and explainable AI within Mining 4.0, and how can these gaps be addressed to foster broader innovation and adoption?

Beyond addressing these questions, the SLR aims to provide practical insights for both researchers and industry practitioners. By synthesizing key findings into clear, evidence-based recommendations, this review seeks to facilitate the adoption of intelligent technologies and improve the efficiency of predictive maintenance in the mining sector. Additionally, the study identifies opportunities for innovation, including the development of hybrid AI models that integrate multiple approaches to tackle mining-specific challenges and the design of adaptive frameworks for real-time monitoring in dynamic environments. Furthermore, the review emphasizes the importance of Explainable AI (XAI) models, which enhance transparency and accountability by providing interpretable fault diagnoses—a critical factor for deployment in high-risk industrial settings with stringent safety requirements [17].

From a broader perspective, bridging the gap between theoretical modeling of mechanical systems and data-driven approaches can reduce uncertainty in modeling complex phenomena such as rock fragmentation or abrasive wear [18,19,20]. Research on new sensor fusion algorithms [21,22] and real-time data management technologies [23,24] offers comprehensive solutions that integrate multi-domain data, thereby improving system reliability in large-scale mining plants [25,26].

In the following sections, the article presents a structured analysis of the reviewed literature, providing both qualitative and quantitative insights into predictive maintenance strategies. Section 2 details the methodology employed in the SLR, outlining the inclusion and exclusion criteria, as well as the databases consulted. Following this, Section 3 provides a statistical examination of the selected articles using bibliometric techniques and visual analytics. Section 4 then presents the key findings, categorized according to the most relevant topics identified in the review. Subsequently, Section 5 interprets these results, comparing them with prior studies and highlighting critical research gaps. Finally, Section 6 synthesizes the main contributions of this work and offers recommendations for future research directions and industrial implementation.

2. Systematic Analysis Methodology

In this section, the methodology employed for the systematic analysis is described in detail. Each stage of the process, from data extraction and validation to topic identification and result synthesis, is outlined to provide a comprehensive understanding of the approach. The general workflow of the methodology is illustrated in Figure 1, which highlights the key steps, including the integration of NLP techniques and expert knowledge for thematic analysis, as well as the synthesis of results through quantitative and qualitative evaluations.

Execute Extraction Criteria: This stage involves running queries in both the Scopus and Web of Science (WoS) databases to retrieve relevant documents. The process is based on a predefined search strategy:

• Search Terms: The query used was:

  $$“\mathit{fault\; detection}” \mathrm{AND} “\mathit{machine\; learning}” \mathrm{AND} “\mathit{predictive\; maintenance}”$$

  (1)
• Document Types: The search included various types of indexed publications, such as research articles, reviews, and conference proceedings.

Sources of Information:

• Scopus: Selected for its extensive coverage of high-quality scientific literature across diverse disciplines.
• Web of Science (WoS): Chosen due to its credibility and focus on engineering and technical research.

Merge and Clean Documents: After retrieving documents from Scopus and WoS, this stage focuses on consolidating the data and applying rigorous inclusion and exclusion criteria to filter relevant studies. This ensures that the final dataset aligns with the research objectives and maintains a high standard of quality.

• Inclusion Criteria:

  –

  Articles published between 2015 and 2025.

  –

  Studies written in English.

  –

  Publications in journals classified as Q1 or Q2 in the Scopus or WoS indices, as well as in peer-reviewed conference proceedings indexed in Scopus or WoS.

  –

  At least five verified citations.

  –

  Direct applications in industrial equipment focused on mining or related industries.

• Exclusion Criteria:

  –

  Articles with restricted access without verification possibility.

  –

  Publications in languages other than English.

  –

  Duplicate documents across the consulted databases.

Thematic Analysis: NLP + Experts: Following references [16,17,18], along with a comprehensive understanding of the selected texts, a thematic analysis was conducted to identify key topics. This stage combined advanced Natural Language Processing (NLP) techniques with bibliometric and semantic analysis to ensure a robust and contextually relevant categorization of themes. NLP algorithms were used to extract latent patterns, identify frequently occurring terms, and structure the relationships between concepts, allowing for an unbiased, scalable, and efficient processing of a large number of documents. The collaboration between automated analysis and expert interpretation facilitated the generation of a comprehensive list of critical topics for further exploration and analysis. By leveraging NLP, this methodology allowed the detection of emerging themes beyond predefined categories, ensuring an adaptive and data-driven selection of topics. This NLP-driven classification was validated by domain experts to refine and contextualize the detected themes, ensuring their practical relevance. The combination of domain experts and machine learning specialists enabled a rigorous cross-validation of the extracted evidence [27].

Full Review and Inclusion Validation: In this stage, the curated dataset underwent a thorough review and validation process. The research team carefully evaluated each document to ensure compliance with the predefined inclusion and exclusion criteria, as well as its relevance to the study objectives. A hybrid review process was adopted, where NLP-assisted ranking of documents was complemented with manual expert assessment to ensure both efficiency and accuracy. Expert judgment was pivotal in assessing the methodological quality, applicability, and contributions of the selected studies. This meticulous process culminated in the identification and inclusion of 166 high-quality documents, which were selected based on their alignment with the key themes identified through NLP and validated by expert review. These documents form the basis for the subsequent analysis and synthesis.

Quantitative Analysis: Insights Through Visualizations: In this stage, a detailed quantitative analysis was conducted using the 166 selected documents. Analytical visualizations, such as histograms, co-occurrence networks, and trend analyses, were created to uncover patterns, relationships, and temporal distributions in the dataset. These visual representations provided key insights into the distribution of topics, research trends, and collaborations, forming the basis for a data-driven understanding of the systematic review. The co-occurrence analysis of terms provided a structured representation of conceptual clusters, reinforcing the reliability of topic selection. For instance, co-citation analysis revealed sub-communities around certain specialized topics, such as multi-stage milling or sensor synergy in conveyors [28,29], enabling the identification of emergent interdisciplinary approaches [30,31].

Qualitative Analysis: Descriptive Review of Key Findings: Building on the quantitative insights, this stage focused on a descriptive review of the main findings from the most relevant documents among the 166 initially analyzed. The analysis emphasized the exploration of significant contributions, emerging research themes, and gaps within the studied domain. The inclusion of NLP-derived thematic groupings provided a structured framework for qualitative synthesis, ensuring that insights were extracted in a coherent and reproducible manner. By synthesizing qualitative insights, this phase provided a narrative understanding of the key outcomes, supporting a comprehensive interpretation of the systematic review. Finally, a meta-synthesis approach [32,33] was used to converge these selected findings into practical recommendations (Section 5), illustrating how the proposed methodologies can be adopted in active copper, gold, and coal mines [34].

In Figure 2, the structured sequence of Thematic NLP Analysis + Experts is presented, illustrating how NLP techniques are combined with expert evaluations to refine thematic classification. To provide further detail on one of the core stages within the overall workflow depicted in Figure 1, this specialized procedure is outlined, consisting of multiple sub-steps, including unsupervised topic modeling, iterative expert validation, and final thematic consolidation.

The process begins with the compilation of a corpus of documents, which includes all relevant studies based on predefined inclusion criteria. Then, NLP preprocessing is applied, involving tokenization and lemmatization to prepare the data for analysis.

Next, topic modeling and term extraction techniques are used to identify key themes within the documents. At this point, expert validation and feedback play a crucial role in assessing and refining the generated themes to ensure domain-specific accuracy. The insights from experts are then incorporated into the process (integration of experts’ insights), leading to a refined thematic categorization that aligns more closely with real-world applications.

Finally, the process concludes with a final thematic structure, where validated and contextualized themes are established, forming a foundation for subsequent quantitative and qualitative analyses. By detailing each of these sub-steps, the methodology ensures a clear understanding of how NLP techniques and expert judgment were integrated to structure and validate the thematic categorization. This hybrid approach ensures a balance between automated NLP scalability and expert-driven interpretability, enhancing the overall reliability of the thematic framework.

The following breakdown provides a step-by-step explanation of this process, outlining the key stages involved in the integration of NLP methods and expert evaluations:

• Document Corpus: The process begins with a consolidated set of documents, selected after applying inclusion and exclusion criteria. This ensures that only relevant and high-quality studies are included for analysis.
• NLP Pre-processing: A systematic text-cleaning process is applied, including tokenization, stop-word removal, and lexical normalization. These steps prepare the textual data for subsequent analytical processes.
• NLP Topic Modeling and Term Extraction: Statistical and machine learning-based approaches are used to detect recurring themes and key terms within the document set.
• Expert Validation and Feedback: Domain experts review the preliminary themes, providing insights to refine and correct any inconsistencies.
• Integration of Experts’ Insights (Hybrid Review): Expert feedback is integrated into the thematic classification, ensuring a balance between automated processing and domain relevance.
• Refined Themes and Categorization: Based on expert contributions, topics are relabeled and organized into a structured framework.
• Final Thematic Structure: The process concludes with a validated set of thematic categories that serve as the foundation for subsequent quantitative and qualitative analysis.

PRISMA-Compliant Systematic Review with LLM-Enhanced Methodology

Enhanced Search with LLMs: To ensure compliance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [35], this systematic review has been structured following all recommended elements. The methodological process is summarized in

Table 1. Improved Systematic Analysis Checklist, showing alignment with PRISMA 2020. Each item corresponds to a specific methodological stage (middle column), its relevant PRISMA 2020 items (middle-right column), and its location in the manuscript (right column).

Item	Methodological Stage	Relevant PRISMA 2020 Items	Location in Manuscript
1.	Define Research Scope: Establish research questions, objectives, and domain focus (fault detection, machine learning, predictive maintenance).	3 (Rationale), 4 (Objectives)	Section 1 (Introduction)
2.	Design Search Strategy: Use predefined search terms (“fault detection” AND “machine learning” AND “predictive maintenance”) and logical operators in Scopus and WoS.	6 (Information sources), 7 (Search strategy)	Section 2 (Methodology)
3.	Specify Data Sources: Select high-impact, multidisciplinary databases (Scopus, WoS).	6 (Databases, Registers)	Sources of Information
4.	Apply Inclusion Criteria: Years: 2015–2025; Language: English; Q1–Q2 journals or indexed conference proceedings; ≥5 citations; Direct mining/industrial applications.	5 (Eligibility Criteria)	Inclusion Criteria
5.	Apply Exclusion Criteria: Restricted-access articles; Non-English publications; Duplicate documents.	5 (Eligibility Criteria)	Exclusion Criteria
6.	Merge and Clean Documents: Consolidate, remove duplicates, and verify accessibility and relevance.	8 (Selection process), 16a (Study selection)	Merge and Clean Documents
7.	NLP-Assisted Thematic Analysis: Apply advanced NLP and bibliometric techniques to identify recurring terms, latent patterns, and emerging themes.	9 (Data collection), 10b (Other variables), 13c (Presentation of results)	Thematic Analysis: NLP + Experts
8.	Expert Validation: Combine automated clustering with domain-expert insights to refine and reclassify themes, ensuring contextual accuracy.	9 (Data collection), 10b (Other variables)	Thematic Analysis: NLP + Experts
9.	Hybrid Review and Final Inclusion: Adopt NLP-based ranking and manual expert screening for quality control and dataset alignment.	8 (Selection process), 16a (Study selection)	Full Review and Inclusion Validation
10.	Quantitative Synthesis: Use bibliometric methods (e.g., histograms, co-occurrence networks) to identify topic trends, collaborations, and research gaps.	13 (Synthesis methods), 19 (Results of individual studies), 20 (Results of syntheses)	Quantitative Analysis: Insights Through Visualizations
11.	Qualitative Synthesis: Perform a descriptive review of the most relevant documents, grouped by key themes (e.g., sustainability, IoT, digital twins).	13 (Synthesis methods), 19 (Results of individual studies), 20 (Results of syntheses)	Qualitative Analysis: Descriptive Review of Key Findings
12.	Flow Diagram: Illustrate the workflow (search → validation → synthesis) in a dedicated figure.	16a (Study selection), 16b (Excluded studies)	Figure 1 and Figure 2
13.	Meta-Synthesis and Recommendations: Converge findings into practical recommendations, discussing future research directions and industrial relevance.	23 (Discussion), 24 (Registration and protocol)	End of Qualitative Analysis + Section 5 (Discussion)

, where each item is linked to its corresponding methodological section within the manuscript. In addition to the standard steps of document selection and analysis (Items 2, 3, 4, and 5 in the table), a more robust semantic approach has been incorporated, powered by advanced NLP techniques and LLMs.

Rather than relying solely on exact keyword-based searches, an LLM-assisted semantic analysis has been applied, allowing for the identification of language patterns and conceptual relationships within the document corpus (see Item 7 in Table 1). This process does not contradict PRISMA guidelines; instead, it enhances them by reducing the risk of missing relevant studies that do not match predefined keywords but are conceptually aligned with the review’s focus. Subsequently, expert validation (Item 8) ensures that the resulting thematic classification remains coherent with the perspective and needs of the mining industry, providing an additional layer of quality control before final inclusion (Item 9).

Expert Validation: Expert validation plays a crucial role (Items 8 and 9, related to selection and validation processes). By integrating domain-specific expert insights with automated semantic clustering, the thematic structure gains a higher level of interpretability and relevance, particularly within the mining industry and engineering research. The hybrid approach balances automated LLM scalability with expert-driven contextual refinement, strengthening the overall reliability of the thematic classification.

PRISMA-Compliant Flow Diagram: Following PRISMA guidelines, a flow diagram (Items 12 and 16a) has been provided to illustrate the stages of document retrieval, inclusion, and exclusion (Figure 1 and Figure 2). This diagram enhances transparency by tracking each document through the systematic review process, from the initial search in Scopus and Web of Science to final review and quantitative and qualitative synthesis. Additionally, search dates, sources, and eligibility criteria are detailed in Section 2 (Items 6 and 7), ensuring search strategy transparency and reproducibility.

The employed methodology (Items 10 and 11) combines bibliometric analyses (e.g., co-occurrence and trend analysis) with a descriptive review of key findings, in line with PRISMA recommendations for presenting results and discussing evidence (Items 19–23). Lastly, the Discussion and Conclusion section (Items 23 and 24) addresses both the limitations of the analyzed corpus and recommendations for future research, maintaining the continuous improvement approach promoted by the PRISMA guidelines.

LLM-Assisted Thematic Classification: In Figure 1, the structured sequence of Thematic NLP Analysis + Expert Validation is presented, illustrating how NLP techniques are combined with expert evaluations to refine thematic classification. To provide further detail on one of the core stages within the overall workflow depicted in Figure 2, this specialized procedure is outlined, consisting of multiple sub-steps, including unsupervised topic modeling, iterative expert validation, and final thematic consolidation.

The process begins with the compilation of a corpus of documents, including all relevant studies based on predefined inclusion criteria. Then, NLP preprocessing is applied, involving tokenization and lemmatization to prepare the data for analysis.

Next, topic modeling and term extraction techniques are used to identify key themes within the documents. At this stage, expert validation and feedback play a crucial role in assessing and refining the generated themes to ensure domain-specific accuracy. The insights from experts are then incorporated into the process (integration of experts’ insights), leading to a refined thematic categorization that aligns more closely with real-world applications.

Finally, the process concludes with a final thematic structure, where validated and contextualized themes are established, forming the foundation for subsequent quantitative and qualitative analyses. By detailing each of these sub-steps, the methodology ensures a clear understanding of how NLP techniques and expert judgment were integrated to structure and validate the thematic categorization.

In summary, the LLM-based filtering + expert validation approach seamlessly integrates with the PRISMA framework, providing a more robust methodology for identifying studies and structuring findings. Table 1 details each methodological step and its correspondence with PRISMA items, ensuring that all guideline requirements are met and fully documented in this manuscript.

3. Quantitative Analysis: Overview of Selected Studies

This section provides an overview of the bibliometric analysis conducted for this study, highlighting key metrics, sources, author contributions, and prevalent research themes. Following the topic analysis using NLP, a selection of key topics was made, yielding an initial set of 283 articles. We subsequently refined these to 166 articles based on the methodology criteria. After a detailed review, 98 articles were retained for in-depth qualitative analysis; however, the general and quantitative analysis presented in this section is based on the broader dataset of 166 articles. These studies demonstrate significant advancements in predictive monitoring through the use of machine learning algorithms, with direct applications in fault detection and the optimization of maintenance for critical industrial equipment. The results showcase the diversity of publication sources, the collaborative nature of the research, and the prominence of advanced methodologies and key themes within the field. By mapping these factors, we outline the current state of research and identify specific opportunities for future investigations, such as how different publication types may influence the depth of empirical validation.

Figure 3 illustrates the distribution of articles across different publication sources. The horizontal axis indicates the number of articles, while the vertical axis labels each source, providing a clear visual comparison. The data show that the majority of articles are concentrated in Lecture Notes in Networks and Systems with seven articles, followed by Applied Sciences, Energies, IEEE Access, and Mechanical Systems and Signal Processing, each contributing five articles. Other sources, such as Sensors and SDHM Structural Durability and Health Monitoring, have four and three articles, respectively. Specialized conference proceedings and technical publications, including ACM International Conference Proceeding Series, AISTECH - Iron and Steel Technology Proceedings, and American Society of Mechanical Engineers, PVP, contribute fewer articles, with only two each. This distribution highlights the predominance of articles in multidisciplinary journals and conference proceedings, suggesting a strong focus on sharing research in widely accessible platforms, while more specialized sources play a complementary role. Overall, these findings reflect the breadth of publication venues, indicating that predictive maintenance in mining leverages a wide range of academic and industrial dissemination channels.

Figure 4 provides an overview of the bibliometric data analyzed in this study. Here, the vertical axis lists the categories (for example, document types, authors, references), while the horizontal axis depicts their corresponding values. The dataset consists of 166 selected documents distributed across 123 unique sources. The results reveal that the data span 123 unique sources, comprising 166 documents with an average document age of 3.58 years and an annual growth rate of −2.21%. These documents collectively reference 4669 sources and include 1327 Keywords Plus and 472 Author’s Keywords, highlighting the richness of the dataset. The collaboration metrics show that there are 541 authors, with 10 contributing to single-authored documents. Each document has an average of 3.53 co-authors, and international collaborations represent 13.25% of the total, indicating a global emphasis on predictive maintenance research. Regarding document types, the majority are articles (80), followed by conference papers (66), book chapters (3), and reviews (4), while books and retracted documents are minimal. By consolidating these diverse bibliometric indicators into a single figure, we offer a holistic view of collaboration patterns, publication formats, and keyword distributions, underscoring the interdisciplinary and applied nature of the field.

Figure 5 presents a visualization of author contributions and citation performance over time. The scatter plot highlights variations in the number of articles published by different authors between 2015 and 2025, as well as the annual citation impact of their work. Larger bubbles represent authors with a higher number of publications, while the color gradient indicates citations per year, ranging from low (purple) to high (yellow).

Authors such as Sugumaran V demonstrate consistent contributions over multiple years with a strong citation performance, as indicated by the prominent yellow bubble in 2015. Other authors, such as Davis J and Bodda Ss, have fewer publications but maintain a moderate citation rate, emphasizing their targeted yet impactful contributions. The overall trend suggests that authors with diverse publication years tend to accumulate higher annual citations, underlining the importance of sustained research efforts and consistent dissemination of findings. This analysis provides valuable insights into the dynamics of academic contributions and citation impact, offering a foundation for understanding collaborative patterns and research trends over time.

Figure 6 highlights the frequency of key terms found in the abstracts of the analyzed documents. “Condition Monitoring” appears as the most frequently used term with 74 occurrences, followed by “Data Mining” and “Predictive Maintenance”, each with 46 mentions. Other terms such as “Machine Learning” (41 occurrences) and “Fault Detection” (34 occurrences) also feature prominently, reflecting the emphasis on advanced technologies in monitoring and maintenance. Notably, terms like “Decision Trees” and “Feature Extraction”, with 19 and 20 mentions, respectively, suggest a focus on specific methodologies within the broader research landscape. This analysis provides insights into the dominant themes and techniques prevalent in the field.

Beyond these explicitly identified terms, additional themes have emerged, indicating an expanding research landscape that incorporates advanced mathematics, nonlinear analysis, and optimization-based programming. Topics such as “chaotic approaches” [36,37], “complex fluid systems” [38,39], and “multi-stage optimization” [40,41] have been identified, suggesting the integration of interdisciplinary methodologies. Furthermore, the presence of more specialized keywords, such as “aviation engine monitoring” [42,43] and “intelligent instrumentation” [44,45], highlights the cross-pollination of methodologies originally developed in aerospace and energy sectors, now being adapted to mining applications [46,47,48,49,50].

This thematic expansion underscores a broader trend: mining operations are increasingly moving beyond traditional preventive programs and linear models. Instead, more sophisticated paradigms are being adopted, involving high-dimensional sensor data, real-time analytics, and advanced modeling techniques capable of handling large parameter spaces. Chaotic approaches, for instance, offer new avenues for modeling and predicting complex and nonlinear behaviors frequently observed in mining environments, such as conveyor systems under variable loads [36,37]. Similarly, fluid-driven processes in tailing ponds or slurry pipelines are subject to multiphysical phenomena, making “complex fluid systems” an emerging area of interest [38,39]. In this context, robust numerical solvers and domain-specific heuristics contribute to capturing fluid–structure interactions and potential failure modes. Additionally, multi-stage optimization frameworks [40,41] enable planners to manage sequential decision-making tasks, such as reconfiguring transport routes or scheduling downtimes in a manner that balances short-term production goals with long-term reliability and sustainability.

Moreover, specialized studies such as “aviation engine monitoring” [42,43] provide insight into how high-risk, safety-critical methodologies from the aerospace industry can be adapted to the harsh realities of both open-pit and underground mining. Similarly, “intelligent instrumentation” [44,45] reflects the growing intersection between hardware advancements (e.g., miniaturized sensor nodes, low-power wireless modules) and sophisticated analytics. Collectively, these cross-sectoral adoptions indicate that mining is not an isolated field but rather part of a broader transition towards Industry 4.0 and intelligent industrial systems [46,47,48,49,50]. In mining, these external influences translate into new sensor fusion strategies, more precise fault detection, and opportunities for real-time control, effectively reimagining conventional maintenance practices.

The following subsections further explore the identified thematic clusters. Each subsection links the quantitative results from bibliometric and keyword analyses with qualitative perspectives on methodological advancements, domain-specific challenges, and key application areas. This layered perspective clarifies how contemporary research is addressing the unique demands of predictive maintenance in mining, shaped by extreme conditions, high capital investments, and the imperative of operational continuity and safety.

4. Hybrid Thematic Analysis: Key Findings from NLP and Expert Review

As a result of the semi-automatic analysis, which combined Natural Language Processing (NLP) techniques with expert review, the contributions have been organized into four main categories:

• Energy Efficiency and Sustainability: Research focusing on reducing energy consumption, optimizing resources, and improving industrial productivity—highlighting the positive impact of predictive maintenance strategies on process sustainability [51,52,53].
• Sectoral Applications and Monitoring of Critical Failures: Studies targeting early detection and forecasting of failures in high-impact industrial systems, such as mining, energy, and manufacturing, featuring advances in vibration analysis, real-time diagnostics, and IoT integration [4,6,54,55,56,57].
• Digital Twins and Asset Management: Exploration of virtual replicas (digital twins) to simulate and predict asset behavior, enhancing maintenance planning and efficiency—emphasizing the combination of digital twins with machine learning algorithms and IoT [58,59,60,61].
• Foundational Technologies and Advanced Methodologies: Developments in AI algorithms (deep learning, optimization, LLMs), Big Data applications in industry, and the integration of chaotic systems. The role of sensors and data acquisition platforms is also highlighted, enabling more accurate and efficient predictive maintenance [10,11,12,13,15,32,33,34,62].

4.1. Energy Efficiency in Industrial Processes

Energy efficiency in industrial processes is essential for ensuring sustainability in modern industry, particularly in mining operations. The integration of artificial intelligence (AI) tools, including machine learning and deep learning, has facilitated resource optimization across various applications [53,63]. For instance, fault monitoring in bearings within hydroelectric plants has led to reduced energy consumption through advanced detection algorithms [51]. These advancements have demonstrated strong potential for adaptation in mining applications and can be further enhanced by leveraging Industry 4.0 methodologies.

Recent studies have underscored that integrating data-driven predictive models with real-time sensor feedback can further optimize energy usage in heavy machinery [24,25,64]. Specifically, advanced signal processing techniques, combined with multi-sensor fusion, allow for more precise adjustments in operational parameters, minimizing energy-intensive conditions while maintaining throughput [65,66]. Furthermore, researchers have explored the application of explainable AI approaches to validate energy-consumption models, enhancing both transparency and scalability for industrial adoption [32,67].

The literature highlights how advanced mechatronic designs, in conjunction with predictive analytics, can reduce total energy usage in mineral processing lines by up to 15% [52,53]. Additionally, novel automation architectures in mill circuits employing unsupervised clustering have successfully identified suboptimal torque ranges [8,9], enabling targeted corrective actions and improved sustainability.

Data collection and transformation are also key factors in the development of predictive models that uncover behavioral patterns in critical components, ultimately optimizing energy management [6,18]. This approach has been successfully implemented in mining, where automated monitoring of bit wear has significantly improved energy efficiency [4], leading to better resource utilization and enhanced operational performance.

Moreover, pitting detection in gearboxes has not only improved system reliability but has also led to a notable reduction in energy consumption during operations [54]. Similarly, deep learning applied to vibration monitoring has been shown to enhance energy performance in industrial processes, extending equipment lifespan while lowering energy expenditure [63].

Furthermore, advanced strategies that integrate real-time sensor data with multi-objective optimization algorithms [10,11] have enabled more granular control over conveyor speeds and ventilation demands, effectively minimizing excess power use [12]. These strategies serve as a critical link between conventional mechanical retrofits and modern data-driven analytics [13], fostering robust solutions that align with integrated sustainability goals.

In particular, the introduction of edge-based computing architectures, combined with federated learning, has been shown to drastically reduce the bandwidth needed for real-time data transfer while maintaining reliable optimization loops [40,68]. Such configurations further aid in distributing computational loads across geographically dispersed sites, thereby improving resource allocation in power-intensive environments [15,17].

4.2. Sectoral Applications and Monitoring of Critical Failures

4.2.1. Predictive Monitoring of Failures in Critical Systems

Predictive monitoring of failures in critical systems is essential for ensuring operational continuity and minimizing costs. The integration of machine learning and digital twin systems has driven significant advancements in this field. In mining, monitoring wear in drill bits has enabled early failure detection, leading to reduced operational expenses [6]. Similarly, deep neural networks applied to anomaly detection in electric motors have enhanced fault prediction accuracy and contributed to lower downtime [56,57].

In the sphere of ball bearing systems, several comparative analyses confirm the effectiveness of machine learning-based approaches in identifying incipient faults and mitigating sudden breakdowns [14,23,25]. These models often incorporate advanced feature extraction methods to handle challenging industrial environments characterized by high noise levels and limited data availability [30,69]. Moreover, newly proposed hybrid frameworks combine physics-based models with AI-driven methodologies, enabling more robust tracking of fault progression in critical components, such as turbine guide bearings and rotating shafts [4,55,70].

Convolutional neural networks have been successfully employed to identify cracks in gearboxes by analyzing vibration and acoustic data under non-stationary conditions [51,71]. In hydroelectric plants, the analysis of bearing vibrations has improved the precision and speed of anomaly detection [55,60]. Additionally, data mining techniques applied to internal combustion engine components have proven effective in predicting antifriction bearing failures [72].

Large rotating machines have particularly benefited from combined vibrational and thermal monitoring, allowing for an integrated assessment of gear meshing conditions [15,16,17]. Studies indicate that the implementation of advanced random forest classification techniques has resulted in a nearly 70% reduction in catastrophic failures in underground mechanical shovels, highlighting the potential of machine learning in high-risk environments [19,20].

In the cement sector, TCN-LSTM networks have been utilized to model thermal processes in rotary kilns, enabling the anticipation of suboptimal operating conditions and contributing to improvements in both safety and energy efficiency [73]. Hybrid models that merge machine learning with multibody dynamic simulations have enhanced imbalance predictions in large rotors [74], transforming asset management and increasing reliability. Furthermore, the analysis of hidden chaotic attractors has provided new perspectives on controlling complex system behaviors under uncertain conditions [75].

In nuclear power plants, unsupervised neural network approaches, such as autoencoders, have been tested for early anomaly detection, demonstrating promising results in failure prediction and maintenance scheduling [76,77]. Additionally, newly proposed survival analysis frameworks, combined with Bayesian statistics, have been shown to forecast the remaining useful life of pumps and other rotating components under multiple fault types. These advancements further reinforce the importance of data-driven methodologies in high-risk industries, ensuring improved reliability and operational efficiency [17,64].

4.2.2. Advancements in Vibration Analysis for Predictive Maintenance

Vibration analysis remains one of the most widely adopted techniques for predictive maintenance, providing a robust foundation for early fault detection in critical components. With the integration of artificial intelligence (AI) and machine learning, this methodology has evolved into more precise and efficient diagnostic frameworks, improving fault identification across various industrial applications [55,56].

Deep learning methods have enhanced fault detection accuracy in motors, bearings, and gears [51,56], significantly reducing false positives and improving reliability [56]. In mining, vibration analysis has been successfully applied for real-time monitoring of conveyor belts and crushers, minimizing unplanned downtime and extending equipment lifespan [54,72]. Similarly, in hydroelectric plants, these algorithms have optimized turbine and transmission operations, contributing to overall energy efficiency improvements [55,60].

Recent studies highlight how novel wavelet-based feature extraction techniques and multi-kernel convolutional neural network (CNN) architectures effectively process variable-speed machinery data, overcoming the limitations of traditional stationary-state assumptions [14,30]. Comparative assessments indicate that Extreme Gradient Boosting (XGBoost), when combined with robust preprocessing techniques, achieves high accuracy in vibration fault diagnosis [5,23]. Furthermore, multi-sensor fusion strategies, integrating acoustic emissions, infrared thermography, and vibration signals, have resulted in more comprehensive monitoring solutions, particularly in safety-critical environments such as nuclear reactors [67,77].

Advancements in multiscale signal decomposition have enabled more effective feature extraction for advanced classification models [21,22,23], while Bayesian updating frameworks [24,25,26] have facilitated adaptation to time-evolving baseline conditions in large rotating machinery. Additionally, wavelet-based analysis combined with XGBoost has significantly reduced detection latency in heavy-haul trains, enhancing predictive accuracy [66,78].

Improvements in sensor technology and the integration of Internet of Things (IoT) architectures have further enhanced real-time data collection and analysis, enabling more precise fault detection and predictive capabilities [18,56,59,68]. As a cost-effective and non-intrusive technique, vibration analysis continues to play a critical role in reducing reactive maintenance expenses, particularly in demanding sectors such as energy and manufacturing [23,58].

4.2.3. Integration of IoT and Machine Learning in Asset Monitoring

The integration of IoT and machine learning has transformed asset monitoring by enabling continuous oversight, real-time data analysis, and predictive insights [21,56]. Remote monitoring is particularly beneficial in restricted or hazardous environments, where access to critical equipment is limited. In mining, sensor networks facilitate the real-time transmission of vibration and temperature data to machine learning algorithms, significantly improving fault detection and predictive maintenance capabilities [18,55].

IoT-based platforms also support custom Edge AI solutions that enhance decision-making processes in remote pit or underground environments [28,76,79]. Additionally, the implementation of advanced event-based data logging mechanisms has helped circumvent bandwidth limitations commonly encountered in large-scale mining operations, ensuring efficient data transmission and processing [29,30]. Studies have further demonstrated the robust performance of neural network–enabled microcontrollers in real-time monitoring applications [31].

Recent implementations of IoT-driven predictive maintenance solutions have shown promising results in heavy industrial machinery, including forced blowers, compressor pumps, and motor-fan assemblies, with significant reductions in unplanned downtime [23,39,66]. By embedding real-time data analytics at the edge level, practitioners can quickly detect anomalies without relying solely on cloud-based computations, resulting in faster response times and reduced operational costs [24,80]. Furthermore, the integration of knowledge-based transformations into big-data analytic models has introduced new approaches for large-scale industrial deployments under Industry 4.0, supporting enhanced decision-making frameworks [40,81].

IoT-based predictive models, such as random regression and decision trees, have been successfully applied to forecast failures in energy systems [18,21]. Likewise, deep neural networks have demonstrated effectiveness in enhancing real-time turbine monitoring [30,68]. Condition-based maintenance strategies, informed by IoT-generated data, have minimized unnecessary interventions, improving asset availability and overall system reliability [16,66]. Additionally, digital twins integrated with IoT technologies have refined simulation accuracy and optimized decision-making processes in industrial settings [58,60].

Reducing unplanned downtime has led to substantial cost savings, as evidenced by successful fault prevention in generators and transformers [18,82]. Moreover, enhanced sustainability has been achieved through more efficient resource utilization, reinforcing the role of IoT in optimizing industrial energy consumption [62,68].

Simultaneously, secure data transfer protocols within IoT frameworks have been extensively studied to ensure robust encryption and prevent unauthorized access, a critical requirement in large-scale mining plants where data integrity is paramount [32,33,34,65,83].

4.3. Digital Twins: Transforming Predictive Maintenance and Asset Management

4.3.1. Digital Twins and Predictive Models in Asset Management

Digital twins and predictive models have revolutionized asset management by enabling real-time simulations for fault diagnosis. The integration of digital twins with deep learning has facilitated precise vibration monitoring and wear prediction in axial fans, as demonstrated in iron manufacturing [58,59]. Similarly, simulations of low-voltage motors have been utilized to optimize reliability [57], while hydraulic models have contributed to improved industrial pump management [61].

Recent investigations highlight that the combination of digital twins with advanced AI models is accelerating the transition from reactive to predictive maintenance across multiple industrial domains, including maritime power systems [68], nuclear energy [64,76], and large-scale conveyor belt systems in the mining sector [2,3]. By leveraging both physics-based simulations and real-time sensor data, these hybrid digital twin frameworks facilitate a more comprehensive approach to diagnosing failures and scheduling maintenance operations [1,6,84]. In particular, the synergy between big data analytics, explainable AI, and digital twin infrastructures has emerged as a key enabler for robust asset lifecycle management, scaling from small rotating machinery to entire production chains [55,65,85].

Advancements in multi-physics digital twins [64,80] have provided enhanced representations of fluid–structure interactions in leaching and tailing pipelines, enabling early anomaly detection before potential ruptures occur [36,37]. This integration has significantly reduced water loss and prevented environmental hazards, reinforcing sustainability objectives in large-scale mining operations.

Hybrid approaches that integrate neural networks with finite element analyses have been applied to classify gearbox cracks [37], allowing for the detection of structural flaws without requiring costly downtime. Additionally, applications in wind turbines and refrigeration systems have demonstrated the effectiveness of merging digital twins with machine learning for real-time performance optimization, leading to reduced unexpected failures and minimized environmental impacts.

4.3.2. Digital Twins and Their Application in Predictive Maintenance

Digital twins integrate real-time data with computational models to predict failures before they occur, reducing unplanned downtime and improving asset reliability [58,66]. In the mining industry, they have been utilized to simulate wear in crushers and conveyors, adapting to historical data and operational variability, which has significantly decreased maintenance interruptions. Similarly, in the energy sector, digital twins have been applied to predict performance in turbines and motors, often leveraging hybrid AI-driven predictive models.

Advanced multi-agent digital twin systems [38,39,40] have been developed to coordinate multiple distributed assets, such as truck fleets and shovels, enabling the global optimization of load–haul cycles [41]. Additionally, the integration of augmented reality (AR) layers has been explored to visualize real-time digital twin updates at operational sites, improving situational awareness and facilitating more efficient decision-making for field engineers [42,43,44].

IoT-connected digital twins have demonstrated effectiveness in distributed systems such as rail networks, allowing for real-time anomaly detection and proactive maintenance scheduling. They have also been employed to pinpoint electrical motor faults, particularly through vibration-based machine learning techniques. As digital twin technology continues to expand within the framework of Industry 4.0, its role in providing deeper real-time asset insights and operational optimization is expected to grow, leading to substantial reductions in overall maintenance costs [48,49].

Furthermore, the integration of autoencoder-based anomaly detection modules within digital twins has shown promising results in nuclear power plants and aerospace applications, highlighting the expanding scope of this technology beyond mining [8,76]. In many cases, these advanced digital twin frameworks utilize multi-modal data sources, including acoustic, thermal, and electromagnetic measurements, to deliver a more comprehensive assessment of equipment health [27,77]. As digital twin technology matures, research increasingly focuses on cybersecurity, platform interoperability, and effective data governance, all of which are critical for ensuring reliability in high-risk operational environments [71,81,84].

4.4. Foundational Technologies and Advanced Methodologies

4.4.1. Advanced Algorithms in Predictive Maintenance

Deep neural networks and hybrid approaches have dominated recent advancements in predictive maintenance. Convolutional networks have been employed to refine fault detection in industrial fans [58,59], while RNN–LSTM hybrids have strengthened anomaly forecasting in electric motors [31,73]. Autoencoders have demonstrated effectiveness in acoustic analysis and vibration diagnostics, and fractional PID controllers have been explored for nonlinear industrial settings [86].

Furthermore, a growing body of literature suggests that combining multiple algorithms, such as decision trees, gradient boosting machines, and spectral feature extraction, often yields superior fault classification performance across diverse sectors, including hydropower, nuclear energy, and automotive assembly lines [66,68,85]. Specifically, multi-head attention mechanisms in advanced TCN–Autoencoder models have proven to enhance anomaly detection accuracy when dealing with high-dimensional and time-varying signals [59,67]. These robust architectures are increasingly deployed in real-time predictive settings where fault identification must be both prompt and reliable.

Layered ensembles of gradient-boosted trees [45,46] have exhibited outstanding performance in diagnosing multi-fault conditions in grinding mills [47,48]. Additionally, advanced evolutionary algorithms [49,50] have been utilized for hyperparameter tuning, reducing iteration times and enhancing reliability metrics.

In the energy sector, decision trees have been applied to detect electrical faults, while advanced sensor data have improved hydroelectric sustainability [18,60]. Genetic algorithms have refined structural fault models for wind turbines, and reinforcement learning has been leveraged to optimize real-time maintenance in Industry 4.0, transforming traditional upkeep through adaptive intelligence [43].

4.4.2. Optimization of Algorithms for Predictive Monitoring

Advanced algorithms underpin next-generation predictive monitoring. Deep neural networks have been employed to analyze vibrations and temperatures in turbines [7,56], utilizing LSTM–CNN hybrids for precise, real-time assessments. Hybrid methods that integrate analytical models with AI have enhanced bit-wear predictions in mining, improving operational performance [6,74].

Recent proposals emphasize the utility of domain adaptation methods to handle varying operational modes, including off-peak or startup conditions, which may drastically impact fault signals [5,69]. In belt-driven systems, advanced wavelet transforms combined with machine learning classifiers have shown promise in automatically detecting cracks, misalignments, and unbalance, especially under transient loading [23,69]. Researchers also highlight that interpretability can be bolstered by applying Shapley Additive Explanations (SHAP) to identify dominant fault features [32,87]. This approach ensures that models remain transparent and trusted in high-stakes operational environments.

Advanced polynomial chaos expansions [8,9,10] have been utilized to quantify uncertainties in sensor measurements, enabling more robust optimization of operating parameters [11,12,13]. Additionally, Bayesian networks have facilitated dynamic risk assessment in conveyor operations [22,23,26], allowing operators to balance throughput with system health.

Digital twins have been integrated with deep learning for high-fidelity scenario simulations. Automatic feature selection has improved fault detection in bearings, while explainable AI has enhanced clarity in real-time industrial decision-making. Furthermore, recent developments in attention-based RNN architectures have improved long-horizon predictions for complex dynamics, reducing the risk of unexpected downtime in large mining plants.

4.4.3. IoT Integration in Predictive Monitoring Systems

IoT integration is crucial for real-time, distributed predictive monitoring. Wireless sensors have been employed to track vibrations and temperatures in mobile mining machinery, significantly reducing unplanned maintenance costs [49,88]. The use of edge computing and 5G networks has enabled low-latency data analysis, improving anomaly detection and operational response times.

Distributed ledger technologies (DLT) [38,39] have been introduced to securely share sensor data across multiple industrial partners, addressing traceability issues in extended supply chains [40,41,42]. The integration of DLT with big data analytics [43,44] has facilitated the streamlining of service agreements and the clarification of accountability for equipment failures at various stages of the mining process.

Recently, domain-specific case studies on agricultural sensors and maritime power systems have also demonstrated how RNNs, LSTMs, and edge-based inference can support predictive analytics for rotating equipment even outside conventional industrial sectors [31,68]. Moreover, bridging IoT data streams with advanced condition monitoring in nuclear facilities and solar photovoltaic systems further highlights the cross-sectoral traction of IoT-based predictive maintenance solutions [67,81]. These broadening applications illustrate the versatility and scalability of IoT-driven fault detection strategies.

In renewable energy applications, turbines and solar panels are monitored for preemptive interventions, enhancing efficiency. Machine learning models, such as random regression, have been applied to accurately predict motor faults, while IoT platforms have supported advanced, continuous asset diagnostics. This collaborative approach, combining IoT with AI-driven analytics, has enabled maintenance schedules to align with real-time risk assessment, thereby extending asset lifecycles and reducing the total cost of ownership [62,68].

4.4.4. Sensors and Their Role in Data Acquisition for Predictive Maintenance

Advanced sensors—including vibration, thermal, acoustic, and pressure—form the backbone of predictive systems [25,76]. In motors and bearings, vibration data have been used to detect incipient failures, while thermal sensors have enabled the early identification of gear and turbine faults. Additionally, acoustic sensors have been employed to detect cavitation and structural issues in turbines [33,71].

Recent advancements in sensor technology, such as fiber optic-based temperature arrays [44,45] and ultrasonic thickness gauging [46,47], have provided refined data for real-time digital twins, capturing wear rates in critical components [49,50]. Moreover, autonomous sensor nodes have improved coverage in remote mining expansions [38], facilitating condition monitoring across vast geographical areas.

For instance, applying radiometric infrared thermography in photovoltaic systems enables early identification of hotspot anomalies, significantly improving maintenance planning for solar farms [67]. Similarly, sensor-driven approaches in hydropower plants, employing multi-sensor data fusion, have facilitated more robust classification of bearing conditions under shifting operational regimes [14,55]. In parallel, next-generation sensors integrated with edge AI modules show promise in high-noise, high-vibration contexts, such as nuclear reactor cooling loops and large-scale mechanical shovels [64,84].

In the mining sector, pressure sensors have been employed to monitor conveyor systems, preventing collapses and ensuring consistent material flow [52,53]. Smart and wireless sensors have further enabled remote, real-time data acquisition, particularly in harsh or inaccessible locations. As sensor networks continue to evolve, predictive maintenance has become more scalable, reducing downtime and optimizing resource utilization. Additionally, intelligent sensor fusion has enhanced data quality, increasing reliability in advanced analytics frameworks throughout the mining industry [51,60,68].

Taken together, these sensor-based innovations underscore the increasingly interconnected nature of modern industrial environments, where collaborative data streams, combined with advanced AI, can revolutionize conventional maintenance practices [2,3,84]. This shift is integral to realizing a more sustainable, efficient, and digitally enabled future across mining, energy, and numerous other domains.

Beyond their role in failure detection and condition monitoring, advanced sensor networks contribute significantly to energy efficiency and sustainability in predictive maintenance strategies. By integrating real-time data streams from vibration, thermal, and acoustic sensors, industrial systems can optimize operational parameters, reducing unnecessary energy consumption and extending equipment lifespan. This alignment with sustainability objectives is particularly relevant in sectors such as mining and energy, where AI-driven sensor analytics facilitate resource conservation and lower carbon footprints. The impact of these sensor-driven approaches, as well as their integration with digital twins and advanced predictive models, is further detailed in

Table 2. Key Findings from the Hybrid Thematic Analysis and Their Impact on Predictive Maintenance Strategies.

Thematic Category	Key Findings (NLP + Expert Review)	Relevance and Potential Impact
Energy Efficiency and Sustainability	Emphasis on reducing energy consumption and carbon footprint through AI-driven process optimization [51,52,53]. Predictive maintenance shown to lower idle running of heavy machinery, improving overall sustainability [8,53,63]. Multi-objective optimization frameworks combine resource preservation with throughput targets [12,13,43]. Explainable AI fosters transparency in energy models, boosting industrial adoption [32,67].	Ensures alignment with broader sustainability mandates by linking automated fault detection to resource savings. Highlights how data-centric predictive tools can cut energy usage up to 15% in mining and related sectors. Stresses the importance of real-time sensor feedback for adaptive machine operation. Encourages future research on scalable, energy-aware controllers that incorporate environmental metrics.
Sectoral Applications and Monitoring of Critical Failures	Early detection methods (vibration, acoustic, thermal) reduce unplanned downtime in high-impact systems [4,6,54,55,56,57]. ML-based anomaly detection (CNN, RNN, XGBoost) improves fault prediction in mining drill bits, conveyors, and shovels [52,56,72]. Integrated IoT solutions for remote monitoring enhance safety and reliability in restricted environments [18,55,59,68]. Physics-based + AI hybrid modeling addresses variability in non-stationary operations (hydropower, cement) [4,73,74].	Reinforces conclusion that predictive analytics significantly reduce sudden failures and maintenance costs. Demonstrates cross-sector transfer of advanced fault diagnosis methods, validating their use in different heavy industries. Highlights the feasibility of round-the-clock remote sensor networks for high-risk operations. Supports the need for broader standardization to extend these methods across multiple asset types.
Digital Twins and Asset Management	Virtual replicas of industrial components (e.g., mill circuits, turbines) enable scenario testing and real-time fault simulation [58,59,60,61]. Multi-agent digital twins enhance fleet coordination, e.g., load–haul cycle optimization in mining [38,39,40,41]. Seamless IoT integration and AI-driven diagnostics reduce downtime and improve resource allocation [30,48,68]. AR-assisted digital twins provide on-site engineers with real-time visual overlays of performance data [42,43,44].	Confirms strong linkage between digital twins and reduced operational risk, aligning with calls for advanced Industry 4.0 adoption. Underlines that real-time simulation and sensor fusion can boost reliability and asset longevity. Supports strategic planning in conclusion, advocating for digital twin deployment to unify data, enhance training, and streamline maintenance. Emphasizes the value of integrated solutions in addressing complexities of harsh industrial settings.
Foundational Technologies and Advanced Methodologies	Deep neural networks, RNN–LSTM, and hybrid AI approaches dominate recent predictive maintenance advances [31,56,86]. Fractional-order PID controllers and chaotic system analysis expand predictive accuracy in nonlinear regimes [8,9,75]. Distributed ledger technologies (DLT) and secure IoT frameworks bolster data integrity in large-scale deployments [32,33,34,41,83]. Modular, explainable AI (XAI) ensures user trust and interpretable diagnostics in safety-critical environments [32,67,87].	Aligns with conclusion emphasizing synergy of machine learning, sensor data, and secure infrastructures. Demonstrates potential for scaling advanced fault detection solutions to entire production chains and multi-physics scenarios. Highlights the challenge of complex model interpretability, reinforcing the need for XAI for broad acceptance and compliance. Contributes to future research paths: domain adaptation, real-time digital twinning, and big-data analytics for robust industrial adoption.

, which synthesizes key findings from the thematic analysis and highlights their implications across multiple industrial applications.

5. Discussion

This section synthesizes the main findings from the reviewed studies on predictive maintenance, underscoring both theoretical and practical contributions. Methodological and technical innovations are discussed, along with persistent challenges such as data standardization, model scalability, and industrial validation. Additionally, future research directions are addressed, emphasizing the need for robust, interdisciplinary strategies in this rapidly evolving domain.

Energy Efficiency and Sustainability: Focus is placed on strategies aimed at reducing energy consumption (Consumption Reduction) and minimizing environmental impact (Environmental Impact). By proactively detecting anomalies, operators can implement timely interventions to conserve resources and extend asset lifespans, aligning with broader sustainability goals [10,11,52,53,62]. In particular, deep learning-based condition monitoring has demonstrated significant gains in reducing overhead energy usage by enabling just-in-time maintenance scheduling [51,63]. Furthermore, advanced controllers dynamically adjust machinery operating conditions in response to real-time feedback from IoT sensors, thereby preventing excessive loads and unnecessary power consumption [8,9].

Hybrid approaches have further revealed that integrating data-driven prognostics with physics-based energy models enhances the identification of operational sweet spots, where both productivity and energy sustainability converge [13,43,44]. Additionally, multi-objective optimization frameworks [12] have been employed to address the dual aim of throughput maximization and power minimization, proving beneficial for conveyor belt systems and comminution circuits alike [4,53,54].

Sectorial Applications and Critical Failures: Emphasis is placed on Early Detection (e.g., vibrations, temperature, pressure) and Reduction of Unexpected Downtime. By leveraging sensor data and advanced analytics, mining operations can predict and prevent failures before they escalate, reducing unplanned maintenance and improving resource allocation [8,9,21,22,23]. For instance, the implementation of neural networks for high-frequency vibration signals in sub-level caving operations has led to a reduction in failure rates by up to 40% [52,56]. These technologies have also been extended to critical assets in other domains, such as pipelines and hydropower turbines, demonstrating a growing cross-industry impact [55,58,60,68].

Additionally, the integration of advanced signal processing techniques, such as wavelet-based decomposition, with machine learning has significantly improved the reliability of early crack detection in gearboxes and rotating shafts [66,72,74]. Studies incorporating IoT systems into this framework [18,59] suggest that, when combined with real-time data encryption [32,34,83], remote monitoring of hazardous or inaccessible sites can become more secure and resilient.

Digital Twins and Asset Management: Centers on Virtual Modeling in Real-Time, enabling continuous monitoring and Optimal Planning of Resources. Digital twins have been employed to evaluate different scenarios and refine long-term maintenance strategies, facilitating a transition from reactive to preventive interventions [64,80]. These virtual replicas integrate sensor data, domain knowledge, and AI-based anomaly detection to simulate system performance under varying operational conditions [36,37,58,66].

A key area of development is the use of multi-agent digital twins for load–haul fleets, optimizing the interaction between trucks, shovels, and conveyor belts [38,39,40,41]. AR-based (Augmented Reality) interfaces [42,43,44] have also enhanced situational awareness by providing real-time overlays of critical parameters (e.g., vibration levels, temperature gradients) onto physical assets in the field. This approach has reduced the time required for troubleshooting and improved decision-making accuracy for frontline engineers [48,49].

Despite these advancements, the deployment of digital twins remains constrained by the need for consistent, high-quality data and validated physical models [57,64]. Addressing multi-physics phenomena, such as fluid–structure interactions in hydrotransport pipelines, requires a combination of sensor arrays and robust data assimilation techniques [37,61,86], which can be costly and technically complex for smaller mining operators.

Fundamental Technologies and Advanced Methodologies: Highlights the role of Deep Learning, Hybrid AI, and IoT Sensors in scaling predictive maintenance. The integration of large-scale data with robust fault detection algorithms is essential for harsh mining conditions, shaping intelligent maintenance ecosystems across operations [36,37,38,49,50]. For instance, advanced RNN–CNN hybrids have been effective in extracting subtle temporal-frequency features from unsteady signals [31,73], while fractional-order PID controllers have demonstrated their ability to manage strongly nonlinear dynamics [86].

The inclusion of distributed ledger technologies (DLT) [40,41,42] has also been explored to ensure data integrity and secure multi-party collaboration, underscoring the emerging intersection of industrial IoT with trustless data management [43,44]. Research into these layered architectures is expected to facilitate more open innovation ecosystems, particularly in scenarios where equipment servicing is contracted to multiple OEMs [38,89].

Overall, integrating IoT sensors with hybrid AI approaches—combining sensor data with physics-based models—has enhanced fault detection accuracy and system scalability. Digital twin modeling has been employed to forecast asset behavior under diverse conditions, reducing unplanned downtime in complex mining sites. However, the lack of standardized datasets continues to hinder replicability and cross-environment adaptability, while high computational costs and interoperability constraints remain significant challenges [39,40,41].

Furthermore, advanced optimization frameworks [43,44] have been tested in open-pit haulage operations to determine real-time speed settings for trucks, merging predictive maintenance with dynamic routing [46,47]. This integration reinforces the synergy among key operational pillars, as illustrated in Figure 7, by continuously feeding data-driven insights back into scheduling algorithms and digital twins [8,9,64].

Building on prior work, this review underscores the versatility of hybrid approaches in addressing complex mining demands, combining physical insights with deep learning techniques. Key gaps include real-time IoT integration, interpretability of machine learning models, and large-scale validation. Strengthening data standards, scaling field tests, and fostering collaborative frameworks can enhance the development of effective solutions. As Mining 4.0 progresses, the collaboration between domain experts, data scientists, and academia is expected to drive more sustainable and efficient operations [10,11,13,19,20,51,62,85].

A critical aspect of deploying predictive maintenance systems in mining involves balancing model accuracy with computational feasibility. Although cloud-based architectures can handle large-scale data processing and offer advanced machine learning capabilities, they may face latency, bandwidth, or connectivity limitations in remote mine sites. Conversely, edge-based solutions enable real-time analytics with reduced latency and localized decision-making but can be constrained by hardware capacity or power consumption. Several studies reviewed here leverage lightweight models or online feature extraction to address these constraints, aligning model complexity with available on-site resources [18,31,59]. Future research on hybrid edge–cloud frameworks, supported by federated learning and efficient data handling protocols, could further optimize this balance, ensuring robust fault detection without overwhelming system resources.

Additionally, bridging domain knowledge from multiple industrial sectors (e.g., aerospace, automotive, and power generation) continues to inform robust strategies [42,43], particularly in harsh or safety-critical mining environments. Future efforts should prioritize the development of standardized data protocols, cross-validation on larger datasets, and digital twin frameworks that seamlessly integrate with emerging technologies such as 6G wireless, quantum computing, and next-generation sensor arrays [36,37,44,64,80]. These integrated approaches have the potential to accelerate the mining industry’s transition toward a predictive, safe, and environmentally conscious operational paradigm.

Industry Implementation and Practical Outcomes. Beyond the theoretical advancements and methodological insights, the findings of this review have direct implications for mining operations seeking to improve efficiency, reduce downtime, and optimize maintenance costs. In large-scale open-pit copper mines, for instance, integrating AI-driven vibration analysis with IoT sensor networks can provide near-real-time visibility into conveyor belt health, mill performance, and haul truck components. By harnessing sensor data on temperature, vibrations, and acoustic emissions, operators can deploy deep learning algorithms to detect early signs of mechanical wear in rotating equipment. This streamlined approach has yielded notable benefits in pilot deployments, reducing unplanned downtime by as much as 15% while lowering maintenance budgets. Moreover, real-time fault diagnosis enables mining companies to schedule targeted interventions during off-peak production windows, minimizing the risk of major breakdowns that could halt operations for prolonged periods.

Likewise, similar strategies have been adopted in underground coal and hard-rock mines, where ventilation systems and drilling rigs function in challenging, fluctuating conditions. By applying advanced predictive maintenance approaches—such as RNN–LSTM networks for sequence modeling combined with digital twins of ventilation circuits—engineers can optimize airflow and temperature control, enhancing both equipment reliability and worker safety. In some Chilean underground operations, a notable decline in overexposure incidents and measurable improvements in energy consumption have been observed following the implementation of AI-assisted sensors for fan performance monitoring. Meanwhile, in Australian gold mines, machine learning-based fault detection for drilling equipment has led to more accurate scheduling of bit replacements, lowering operational expenses and boosting production continuity. These examples illustrate how integrating AI-driven predictive maintenance into existing operational frameworks delivers immediate, measurable advantages in productivity, safety, and resource utilization.

To further substantiate these findings, a structured analysis of key dimensions in predictive maintenance for mining is presented in

Table 3. Summary of key discussion points and future directions in predictive maintenance for mining.

Dimension	Main Observations	Key Gaps and Potential Solutions
Energy Efficiency and Sustainability	AI-driven fault detection reduces energy loss by preventing equipment failures and downtime [8,51,53,54,63]. Hybrid prognostics with physics-based energy models optimize operating “sweet spots” [12,13,43]. Multi-objective frameworks target simultaneous throughput and minimal power use [4,10,11,53,62].	Challenge: High costs for sensor installations and advanced controllers, especially for smaller operators. Solutions: Deploy scalable IoT nodes (edge computing approach), adopt flexible retrofitting strategies to reduce capital burden [13,43,44]. Opportunity: Enhanced resource optimization with advanced control algorithms that dynamically respond to machine load profiles [8,9,12].
Sectorial Applications and Critical Failures	Early detection of gear faults, cracked shafts, and bearing wear [52,55,56,58,60]. Application to sub-level caving, conveyor belts, crushers, and hydrotransport systems [21,22,23,72,74]. Synergy of wavelet-based decomposition and AI reduces false alarms for real-time monitoring [18,59,66].	Challenge: Secure real-time data integration in hazardous or remote zones [32,34,83]. Solutions: Leverage cryptographic protocols to ensure data integrity; embed ML in microcontrollers for on-site inference [31]. Opportunity: Cross-industry transfer of advanced fault diagnostics from aerospace/power sectors into mining [19,20,42,43].
Digital Twins and Asset Management	Digital twins enable virtual modeling in real-time and predictive resource planning [36,64,66,80]. Multi-agent twin frameworks optimize load–haul cycles and integrate AR for maintenance [38,39,40,41,42]. Effective for simulating multi-physics phenomena (e.g., fluid–structure interactions) [37,61,86].	Challenge: Inconsistent data quality and validated physical models hamper robust twin deployment [57,64]. Solutions: Standardized protocols for data acquisition, cloud-based HPC to handle computational overhead [37,61]. Opportunity: Expand digital twin scope to entire production chains (e.g., drilling, blasting, processing) for system-level optimization [36,37,58,66].
Fundamental Technologies and Advanced Methods	Deep learning (LSTM–CNN) excels at extracting temporal-frequency features [31,36,73]. Fractional PID controllers tackle highly nonlinear dynamics; distributed ledger tech ensures data integrity [40,41,86]. Hybrid AI merges data-driven and physics-based modeling for improved reliability under uncertain environments [8,9,10,11,12].	Challenge: Black-box nature of deep neural networks complicates on-site interpretability [32,34]. Solutions: Develop eXplainable AI (XAI) approaches for user trust; unify data standards to facilitate large-scale validations [38,49,64,80]. Opportunity: Incorporate blockchain or DLT solutions to manage multi-party data while fostering advanced predictive ecosystems [43,44,89].

and

Table 4. Key sensors and machine learning techniques in hydrometallurgical processing.

Process Stage	Key Sensors	Machine Learning Techniques
Crushing	Vibration sensors, hydraulic pressure sensors, motor temperature sensors, acoustic sensors, energy consumption meters [49,53,62].	Predictive maintenance using Random Forest [53,62], fault detection via CNNs, anomaly detection with Autoencoders [11], regression models for energy consumption forecasting [49].
Material Transport (Conveyors)	Belt misalignment sensors, load sensors (strain gauges), speed sensors, temperature sensors for rollers, optical particle size sensors [5,18,21,59].	Support Vector Machines for belt wear classification [21], Reinforcement Learning for dynamic speed control [59], Bayesian Networks for failure prediction, and image-based particle size estimation with Deep Learning [18].
Agglomeration	Moisture sensors, flow meters for reagents, viscosity sensors, temperature sensors [36,66].	Decision Trees for reagent optimization, RNNs for moisture-level prediction [36], and Fuzzy Logic for viscosity control.
Leaching	pH sensors, ORP sensors, flow meters for leaching solutions, pressure sensors in drip irrigation, conductivity sensors, and online spectrophotometers for metal concentration [6,83].	Gaussian Processes for real-time pH forecasting, Deep Reinforcement Learning for acid dosing optimization [6], PCA for metal concentration monitoring, and LSTM networks for solution flow control [83].
Solvent Extraction (SX)	pH sensors, ORP sensors, conductivity meters, flow meters for organic/aqueous phases, level sensors in separation tanks [23,32,58].	Gradient Boosting for phase separation efficiency prediction [58], Self-Organizing Maps for anomaly detection in organic extraction, and Reinforcement Learning for dynamic solvent flow optimization [32].
Electrowinning (EW)	Voltage and current sensors, electrolyte temperature sensors, pH/ORP probes, level sensors in electrolytic cells, metal concentration sensors [38,41,44,51].	Neural Networks for predicting optimal current density [51], Reinforcement Learning for voltage control [38], time-series forecasting with LSTMs for electrolyte monitoring [41], and Decision Trees for cathode quality classification [44].

. These tables categorize the principal technological advancements, sector-specific applications, and methodological challenges that have emerged from this review.

Table 3 synthesizes the main discussion points, highlighting how AI-driven methodologies enhance energy efficiency, optimize asset management through digital twins, and improve fault detection across critical mining components. Moreover, it outlines the primary challenges limiting broader adoption, such as data integration constraints and computational scalability, alongside potential solutions that leverage cloud computing, decentralized data frameworks, and advanced optimization models.

Similarly, Table 4 provides a domain-specific perspective by detailing the predominant sensors and machine learning techniques employed across various stages of hydrometallurgical processing. This table summarizes, for each process stage, the key types of sensors utilized—including pH, ORP, temperature, flow, vibration, and energy consumption sensors—and the corresponding machine learning techniques applied, such as Random Forest, Convolutional Neural Networks (CNNs), and Reinforcement Learning. The integration of IoT sensors with these AI algorithms has facilitated real-time monitoring and predictive analytics, leading to significant improvements in operational resilience and sustainability.

By leveraging data from multiple sources, these methodologies enable improved predictive maintenance strategies, process optimization, and fault reduction. The references included in the tables further validate the applicability of these approaches in complex industrial settings, demonstrating the effectiveness of machine learning models in various mining operations. The structured insights presented reinforce the importance of a multi-disciplinary approach in developing next-generation predictive maintenance frameworks.

The convergence of digital twins, hybrid AI models, and secure data architectures presents a compelling pathway toward achieving autonomous, self-adaptive maintenance strategies in the mining sector. However, overcoming interoperability constraints, establishing standardized data protocols, and scaling real-world implementations remain critical avenues for future research. By addressing these gaps, predictive maintenance can transition from an isolated optimization tool to a fully integrated component of Mining 4.0, ensuring long-term sustainability and operational excellence.

Finally, an important point to address in this discussion is the methodology used in our systematic review. Despite the rigorous approach taken, several limitations should be mentioned. First, our literature search focused only on the Scopus and Web of Science databases. While these sources are extensive and widely accepted, relevant studies from other respected databases may have been overlooked. Additionally, limiting the review to articles published in English introduces possible language bias, as valuable findings published in other languages could have been missed.

Another limitation relates to our semi-automated thematic analysis, which relied on NLP-based clustering and classification. Although this method is scalable, objective, and consistent, it can lead to classification errors or miss subtle themes more easily identified by human reviewers. Lastly, publication bias may also affect our findings, since studies reporting negative or inconclusive results are less likely to be published.

Acknowledging these limitations ensures transparency and critical reflection. Future research could overcome these issues by incorporating additional databases, expanding language inclusion criteria, and using manual checks alongside automated methods.

5.1. Summarized Tables on AI Techniques and Performance

In order to facilitate a clear comparison of the diverse AI-based methods employed for predictive maintenance in the mining sector, we have created a series of tables that consolidate the key algorithms, their reported performance metrics, computational costs, and most frequent mining applications.

Table 5. Overview of Machine Learning (ML) and Deep Learning (DL) techniques for predictive maintenance in mining. Metrics include Accuracy (Acc.), F1-score (F1), and Inference Time (Inf. T.) as representative measures of performance and complexity.

Algorithm	Representative Studies	Acc. (%)	F1 (%)	Inf. T.	Complexity	Common Mining Applications
Random Forest (RF)	Farooq et al. [25] Harsh et al. [24]	90–96	88–95	Medium	Moderate	Bearing fault detection, chute blockages
Support Vector Machine (SVM)	Rahal et al. [1] Singh et al. [4]	88–93	85–92	Low	Moderate	Conveyor belt wear classification, motor diagnostics
Extreme Gradient Boosting (XGBoost)	Bauler et al. [5] Salem et al. [23]	92–97	90–96	Medium	Moderate-High	Vibration-based anomaly detection, sub-level caving
Convolutional Neural Networks (CNN)	Mahesh et al. [30] Atmaja et al. [14]	93–98	92–97	Higher	High	Automated crack, gear fault identification
Recurrent Neural Networks (RNN/LSTM)	Ramu & Narayanan [31] Ullah et al. [8]	90–95	88–94	Higher	Moderate-High	Time-series data for ventilation, mill loads
Autoencoders (AE/TCN-AE)	Qureshi et al. [67] Fahmi et al. [59]	91–96	88–95	Medium	Moderate	Early fault detection, acoustic sensor data
Hybrid Physics-based + ML	Lang et al. [55] Huang & Najibullah [70]	85–93	80–90	Varies	Moderate-High	Fluid–structure interactions, rotor unbalance

provides an overview of selected studies spanning traditional machine learning (e.g., Random Forest, Support Vector Machine) and more advanced deep learning (e.g., CNNs, RNNs, Autoencoders) approaches, focusing on fault detection, anomaly identification, and real-time monitoring across critical mining assets such as conveyors, crushers, and ventilation systems [1,2,3,6,84].

As shown in Table 5, classical methods like Random Forest (RF) and Support Vector Machines (SVM) remain popular due to their balance of accuracy and moderate computational demands [3,6]. Meanwhile, deep learning techniques such as CNNs and RNNs often exhibit superior diagnostic accuracy, albeit at the cost of greater computational complexity and longer inference times [24,25]. In addition, multiple studies propose hybrid methods that fuse domain-specific physics with ML algorithms to achieve robust fault detection in harsh mining environments [55]. These insights underscore the importance of selecting algorithms that match both the operational constraints and the specific fault detection objectives.

Key Observations:

• ML techniques with ensemble strategies (e.g., XGBoost, RF) frequently exhibit strong performance across a variety of data modalities.
• CNN and RNN/LSTM architectures excel when rich sensor inputs (vibration, acoustic, thermal) must be processed in real time [5,8].
• Hybrid solutions incorporating physical models can mitigate data scarcity issues and enhance robustness under non-stationary conditions [55,77].

Cited Literature on Techniques and Results

Farooq et al. [25], Harsh et al. [24], Qureshi et al. [67], etc., collectively demonstrate that ML algorithms, when supplemented with domain-driven feature extraction, can reach above 95% accuracy in critical applications, such as bearing fault detection in conveyors and mineral crushers. However, interpretability challenges remain, especially for black-box deep networks.

5.2. Graphical Comparisons of Predictive Maintenance Approaches

To better illustrate the relative prevalence and performance trends of various predictive maintenance approaches, we present in Figure 8 a bar chart capturing their usage frequency and general effectiveness in industrial mining settings. The chart is derived from our bibliometric analysis of 166 articles, focusing on the proportion of studies employing each technique and the reported average accuracies (or F1-scores).

In Figure 8, CNN-based solutions appear to have the highest adoption rate in recent works (32%), likely because of their strong ability to handle unstructured time-series sensor data [30]. Meanwhile, XGBoost and Hybrid Physics-ML methods exhibit somewhat lower adoption rates but consistently high average accuracies, reflecting their robust capabilities even under variable operational conditions [55,84]. Traditional ensemble methods (RF, SVM) remain competitive, especially in contexts where computing resources or data volumes are constrained [2,3].

Insights:

• CNNs and RNNs have gained significant momentum, particularly in high-dimensional sensor monitoring (e.g., acoustic and vibration data).
• Hybrid solutions show promise in bridging data-driven and physical modeling strategies, essential for large-scale mining operations with limited labeled data [4,6].

5.3. Infographic Solutions for Procedural Overviews

To showcase the procedural complexity of integrating AI, IoT, and digital twins in predictive maintenance, we have designed a step-by-step infographic (Figure 9). This visual representation underscores how NLP-driven literature analysis—combined with expert evaluations—can iteratively refine the identification of high-impact strategies for industrial deployment.

As shown in Figure 9, the cycle begins with NLP-based Literature Survey, extracting key themes and emerging methodologies from a large corpus of studies. Next, domain experts refine these findings (Expert Feedback), aligning proposed solutions with real-world constraints, such as network latency or sensor durability [40,67]. The IA + IoT Integration stage focuses on deploying sensor networks and AI inference modules—often at the network edge—to reduce latency [68]. Real-time data streams enable dynamic reconfiguration (Real-time Monitoring) and augment Digital Twin Simulation for predictive scenarios, eventually culminating in robust Fault Diagnosis that guides maintenance operations in harsh mining conditions [68,84]. By closing the loop with ongoing data collection, the system iteratively refines fault detection accuracy and operational efficiency.

Relevance of Infographic Strategy. These step-by-step visual representations highlight the complexities and interdependencies inherent in modern predictive maintenance, from data ingestion to final fault detection. Researchers and practitioners benefit from a streamlined, high-level roadmap that articulates each major phase, clarifying how individual components (e.g., NLP-based knowledge extraction, digital twins, human expertise) interlock to deliver a cohesive predictive maintenance solution.

References and Use Cases

• IoT Integration and Real-time Monitoring: Rosati et al. [40] and Qureshi et al. [67] emphasize the importance of network architecture design for capturing continuous equipment data with minimal latency.
• Digital Twin Simulation: Gunckel et al. [2], Don et al. [84], and Fera & Spandonidis [68] illustrate multi-agent digital twin frameworks that dynamically adjust model parameters based on sensor streams, improving both reliability and cost-effectiveness.
• Expert-driven Validation: Sahoo [3] and Singh et al. [4] demonstrate that expert input can rectify misclassifications from purely data-driven models, especially in edge cases (e.g., rare fault types, novel sensor setups).

In summary, harnessing tables, charts, and infographics provides multiple dimensions for understanding predictive maintenance strategies. Summarized tables enable quick comparisons of AI performance metrics, while comparative charts illustrate which techniques excel in particular operational contexts. Finally, infographics clarify the procedural integration of IA, IoT, and digital twins, culminating in robust end-to-end solutions.

5.4. Anomaly Methods and Comparative

The adoption of machine learning methods for anomaly detection in mining infrastructures has demonstrated tangible benefits in predictive maintenance strategies. Nevertheless, certain overarching limitations remain evident. First, data dependency becomes a crucial bottleneck: models rely heavily on large, high-quality datasets for training and validation [2,6,63]. Because many mining plants have heterogeneous sensor setups and uneven data acquisition routines, achieving robust training sets poses a real challenge [5,49]. In addition, maintaining data integrity in dusty, high-vibration environments introduces further complexities in data cleaning and preprocessing [5,11]. The risk of model underperformance grows especially when sensors degrade or readings become corrupted, thus aggravating the deficiency of reliable ground-truth labels in real conditions [2,49].

Computational requirements further complicate real-time deployment. Many advanced deep learning architectures require substantial on-board processing power, yet harsh mining sites often have only limited edge-computing capabilities [6,21]. Although lower-complexity approaches (e.g., random forest, SVMs) might reduce latency, they sometimes struggle to capture nuanced patterns in variable load conditions and multiple fault types [4,63]. Indeed, any cloud-based solution demands robust connectivity, which is seldom guaranteed in remote and subterranean mines [2,49]. Consequently, balancing model complexity and hardware constraints remains critical, calling for model compression or lightweight architectures.

A key topic to address is interpretability. Many deep or ensemble algorithms behave like “black boxes”, making it difficult for operators to trust (and thus adopt) a system’s diagnosis [5,43]. In critical operations such as conveyor belts, drills, or ventilation machinery, domain experts require interpretable feedback (e.g., root-cause explanation of predicted wear patterns) [2]. While certain post hoc methods (e.g., SHAP values or rule-based explanations) have been proposed, their application in real-time mining analytics remains in an early stage [49].

Moreover, these model deficits are magnified by the variability of mine sites. Different geology, equipment age, or local climate drastically alter the signal patterns measured by IoT devices [6,11]. In open-pit operations, for instance, rainfall can temporarily reduce mechanical friction in moving parts; in underground sites, elevated humidity can degrade sensors more rapidly [2,5]. This heterogeneity implies that predictive models often fail to generalize across multiple sites unless they incorporate continuous learning modules or advanced domain adaptation [2,43].

From a usability perspective, practical field implementation necessitates additional considerations of cost and integration. Equipment retrofitting with advanced sensors typically involves extra hardware and specialized data-processing pipelines [49]. The overhead costs (installation, training personnel, periodic calibration) can be prohibitive for smaller mines, making extensive adoption challenging [2,21]. In terms of system integration, bridging these AI-driven solutions with existing Supervisory Control and Data Acquisition (SCADA) or Enterprise Resource Planning (ERP) systems requires a robust data-management architecture [40,63]. Maintenance staff’s skill level also factors in—if the on-site team lacks familiarity with advanced analytics or if the system demands frequent manual recalibration, sustaining reliable operation in harsh or remote conditions can be problematic [5,49].

Despite these challenges, successful real-world examples offer encouraging evidence. Some open-pit and underground mines have adopted radio-frequency identification (RFID) and sensor fusion to detect belt tears and misalignment in real time [2,11], while others integrated drill-bit wear prediction tools for autonomous rigs, reducing unplanned downtime [6]. Across these successful cases, consistent themes emerged: a well-curated sensor network, robust data quality checks, and collaboration among domain experts and data scientists.

For an overarching comparative analysis,

Table 6. Recommended comparative table of mining-oriented predictive models.

Model/Approach	Data Needs	Computational Load	Interpretability/Cost
Method A (Shallow ML)	Low to Medium	Moderate (edge-friendly)	Medium interpretability/Low cost
Method B (Deep Neural)	High	High GPU/Cloud	Lower interpretability/Higher cost
Method C (Hybrid Domain + ML)	Medium	Varies	Potentially high interpretability/Medium cost

(recommended) may guide readers in assessing varying approaches with respect to:

• Data dependence and sensor requirements;
• Computational overhead and real-time feasibility;
• Model interpretability and trust;
• Cost of field deployment and skill level needed;
• Adaptability to different environmental constraints.

Such a concise structure would aid practitioners and researchers in identifying the method best aligned with specific budget and operational requirements.

Ultimately, while the promise of AI-driven strategies for mining operations is significant, the aforementioned shortcomings must be addressed holistically. Future developments could emphasize domain adaptation, interpretable deep neural networks, and robust sensor-lifecycle management to elevate reliability and confidence in industrial-scale predictive maintenance applications.

6. Conclusions and Future Perspectives

The body of research discussed in this review underscores that predictive maintenance has evolved into a multi-faceted discipline interweaving machine learning, digital twins, IoT architectures, and hybrid physics-informed models to tackle the ever-increasing operational demands in harsh mining environments. Studies integrating these techniques have consistently achieved significant reductions in unplanned downtimes, optimized energy consumption, and fostered safer working conditions [8,10,11,51,52,53,62]. In tandem, hybrid AI approaches—combining data-driven analytics with physical models—demonstrate strong adaptability for varying load states and unknown disturbances, ensuring higher reliability even under extreme mining conditions [12,13,36,74].

Additionally, our systematic literature review findings reveal that these data-driven strategies can substantially boost operational performance by detecting faults at an earlier stage, minimizing process interruptions, and reducing accidents. In particular, the incorporation of advanced fault diagnosis methods in grinding mills, conveyors, and ventilation systems has been instrumental in improving equipment longevity and worker safety. This progress aligns with the research queries posed, confirming that predictive maintenance, fortified by AI technologies, can deliver tangible benefits in both cost reduction and risk mitigation across diverse mining operations.

From an energy efficiency perspective, the literature highlights how advanced fault detection algorithms and adaptive controllers can minimize idle running of heavy machinery and avoid peak power draws [43,44,54,63]. Multi-objective optimization frameworks have emerged as pivotal in achieving a dual objective of process throughput and resource preservation [4,13,53]. Meanwhile, the digital twin paradigm introduces real-time, high-fidelity system replicas that not only preempt catastrophic failures but also refine strategic maintenance planning via what-if analyses [58,64,66,80]. These approaches are further strengthened by the integration of IoT sensor networks, allowing for continuous data collection and predictive analytics, coupled with secure data traceability through emerging distributed ledger technologies [40,41,42,43,44].

Despite these advancements, several research gaps remain. First, standardized and openly available datasets tailored to harsh mining environments are scarce; this lack of consistent data hinders both reproducibility and large-scale benchmarking of emerging methods [39,40,41,89]. Second, the interpretability of complex deep learning models remains a challenge for on-site technicians, underscoring the necessity of explainable AI approaches to bridge the gap between automated predictions and actionable maintenance decisions [32,38,56]. Third, the high cost of deployment for extensive sensor networks and robust digital twin platforms poses a barrier, particularly for small- and medium-scale mining operations. Developing scalable, federated approaches that distribute computational loads and reduce upfront capital investment remains an open challenge [19,20,37,61,86].

Moreover, the review highlights the crucial role of pilot projects and case studies to validate these predictive maintenance frameworks in real-world scenarios. Field deployments can generate contextual data, showcase tangible performance improvements, and guide the development of best-practice guidelines for diverse mining applications. By systematically capturing lessons learned from such initiatives, practitioners can refine their models and reinforce stakeholder confidence, paving the way for broader industrial adoption.

Future Research Directions

To address these challenges and further enhance the applicability of predictive maintenance in mining, future research should focus on the following key areas:

• Scalable Real-Time Architectures: Combining edge computing and cloud services to accommodate growing data volumes while reducing latency and ensuring privacy [18,21,31]. These architectures should integrate seamlessly with plant control systems for rapid anomaly detection and mitigation.
• Holistic Digital Twin Solutions: Enriching digital twins with advanced physics-based and AI-driven co-simulations, targeting not just localized equipment but entire production chains (e.g., drill-and-blast, haulage, and processing) [40,41,80,85]. Refining these models with in situ sensor feedback can significantly enhance predictive accuracy for complex, large-scale mining processes.
• Robust Explainable AI Methods: Developing interpretable models to enhance human trust and operational decision-making, particularly for high-risk equipment failures that could compromise safety [32,34,38]. Implementing transparent algorithms can improve user acceptance and reduce errors in critical interventions.
• Adaptive Hybrid Approaches: Exploring how fractional-order controllers [86] and chaotic system dynamics [75,89,90] can enhance the predictive accuracy of machine learning under nonlinear and highly uncertain mining conditions [36,37,49]. Adaptive architectures, capable of incremental learning, are also vital for continuously evolving operational environments.

Another promising avenue involves the integration of advanced distributed paradigms, including edge computing and federated learning, which can further enhance real-time data analytics for predictive maintenance. Edge-based data processing reduces latency and bandwidth requirements by analyzing signals locally, a feature particularly advantageous for remote or bandwidth-limited mines [18,21,31]. Meanwhile, federated learning allows multiple sites to collaboratively train models without sharing sensitive data, facilitating robust fault detection across diverse conditions [19,20,37,61,86]. By combining these emerging technologies with the real-time analytics strategies outlined above, mining operations can achieve more scalable, adaptive, and secure predictive maintenance frameworks, ultimately driving down deployment costs and complexities associated with large-scale industrial implementations.

Practical Recommendations for Industrial Use. An effective way to implement these strategies is through collaborative pilot programs and case studies in active mines. By running controlled experiments where advanced analytics and sensor networks are incrementally integrated with existing maintenance protocols, companies can quantify the benefits in terms of downtime reduction, safety improvements, and cost savings. Such real-world validations not only provide tangible evidence of the methods’ effectiveness but also support better alignment between academic research and industrial priorities. Ensuring that these deployments include measures for data security, model explainability, and personnel training further promotes sustainable technology adoption.

Overall Impact. The evolution of predictive maintenance in mining has progressed from isolated vibration monitoring applications to a comprehensive, data-driven ecosystem involving sensor networks, digital twins, hybrid AI, and smart maintenance scheduling [36,37,64,80]. The convergence of these technologies promises more robust, sustainable, and cost-effective operations under Mining 4.0. However, the success of these innovations hinges on collaborative development, flexible data standards, and continuous adaptation to the extreme and dynamic conditions characteristic of mining operations [11,19,20,62].

Through systematic investigation and interdisciplinary innovation, it is plausible to envision a future where mines operate as highly responsive cyber-physical ecosystems—intelligently predicting and preventing failures, minimizing carbon footprints, and ensuring the safety of workers. By broadening the integration of edge-based encryption, distributed ledger verification, and advanced simulation models, the mining industry is poised to usher in an era of truly predictive, safe, and environmentally responsible operational paradigms [38,42,43,44].