Diagnostics of Rotating Machinery through Vibration Monitoring: Signal Processing and Pattern Analysis

1. Introduction

Rotating machinery, ubiquitous in industrial applications, is vital for numerous processes, so ensuring their reliability, efficiency and safety is of primary importance. Diagnosing faults in rotating machinery has traditionally been a challenge due to the complexity of the systems and the multitude of potential failure modes. Vibration monitoring, coupled with advanced signal processing and pattern analysis techniques, has emerged as a powerful diagnostic tool, enabling early detection of faults and reducing unplanned downtimes. The present Special Issue titled “Diagnostics of Rotating Machinery through Vibration Monitoring: Signal Processing and Pattern Analysis” [1] was meant to explore the methodologies and benefits of using vibration monitoring for the diagnostics of rotating machinery, emphasizing the roles of signal processing and pattern analysis.

In particular, vibration monitoring (VM) involves the measurement of mechanical oscillations in machinery components. These oscillations, or vibrations, are often indicative of the machine’s “health” state. Deviations from normal vibration patterns can signal the onset of various faults, such as imbalance, misalignment, bearing and gear damage, fault or wear. Sensors, typically accelerometers, are mounted on the machinery to capture vibration data, which are then analyzed to assess the machine’s health.

The raw vibration data collected from machinery are often complex and require sophisticated signal processing techniques to extract meaningful information. The primary goal of signal processing is to transform the raw data into a form that highlights the characteristics of potential faults. This data mining procedure falls within big data management, and data compression turns out to be an important secondary effect of such a feature extraction [2].

Analysis and feature extraction can be tackled in different domains. Examining the vibration signal in the time domain and extracting features such as peak amplitude, root mean square (RMS) value, and crest factor is often the first step, which can sometimes provide an insufficient result due to its inability to isolate specific frequency-related characteristics of the signal. A common second step involves the exploitation of the Fourier Transform (FT) which reveals the signal’s frequency components. Faults in rotating machinery often manifest as distinct frequency peaks, corresponding to specific fault frequencies when the rotational speed can be considered constant, or can be order-tracked [3,4]. For instance, bearing defects typically generate characteristic frequencies that can be identified in the frequency spectrum after further detail analysis, such as envelope demodulation [5,6,7,8,9]. Sometimes, a more comprehensive analysis of transient or non-stationary events can be obtained by examining the signal in both the time and frequency domains at the same time, exploiting time–frequency techniques (e.g., Short-Time Fourier Transform (STFT), Wavelet Transform (WT) etc.) [10,11].

Once the relevant features are extracted through signal processing, pattern analysis techniques are employed to diagnose the specific type of fault. In particular, pattern recognition techniques such as Principal Component Analysis (PCA) and cluster analysis can be used to classify the operational state of the machinery comparing the extracted features against known patterns of faults. This is commonly performed either by statistical methods involving the analysis of the statistical properties of the vibration signal features. Control charts and statistical hypothesis testing are commonly applied in this regard to detect deviations from normal behavior [12]. The alternative involves advanced machine learning algorithms, such as support vector machines, neural networks, and decision trees, which can be trained on historical vibration data to recognize patterns associated with specific faults, enabling automated and accurate fault identification [13,14].

The integration of vibration monitoring with signal processing and pattern analysis offers several benefits:

• Early Fault Detection: Allows for the identification of faults at an early stage, preventing catastrophic failures and reducing maintenance costs.
• Improved Reliability: Enhances the reliability and efficiency of machinery by enabling condition-based maintenance rather than time-based schedules.
• Data-Driven Decisions: Provides actionable insights based on real-time data, facilitating informed maintenance decisions.

However, challenges remain in the implementation of these techniques. The complexity of rotating machinery and the variability of operating conditions can complicate the analysis. Furthermore, the initial setup of monitoring systems and the development of accurate diagnostic models require significant expertise and resources.

Also, a gap exists between research advancements and industrial adoption in diagnostic models. Researchers often focus on cutting-edge technologies that are in early development stages, while industries prefer mature, proven solutions. High implementation costs, complexity, and specialized knowledge requirements can deter companies from adopting new research-based methodologies. Additionally, the extensive data requirements for advanced diagnostics pose a challenge for industries with limited data collection capabilities. While academia often pursues novel discoveries, industrial interests are driven by practical benefits like cost reduction and efficiency improvement. Customization and specificity of research solutions may not directly translate to the diverse environments of industrial applications, necessitating adaptable technologies.

It is then the aim of this Special Issue to provide some novel tools to tackle the diagnostic of different rotating machinery from several different areas, driven by industrial needs. In particular, it was decided to prefer works on available benchmark datasets coming from both research laboratories and real industrial applications. For example, the NASA Prognostics Data Repository [15] collects a wide variety of datasets that are essential for the development and validation of prognostic algorithms and methodologies. These datasets include real-world and simulated data on system performance, health, and failures, and are widely used for research in predictive maintenance and fault detection. Gearboxes, bearings electric motors, pumps and structural components data are available. Additionally, universities such as the Case Western Reserve University [16] and the Politecnico di Torino [17] shared rolling elements bearings datasets from laboratory test rigs. Also, partial datasets from International Conferences such as Surveillance 8 and CMMNO 2014 are freely available online, featuring the Instantaneous Angular Speed (IAS) estimation of a SAFRAN aircraft engine and a wind turbine gearbox [18,19,20], while large repositories called Data Lakes are aggregated from the U.S. Department of Energy’s Programs, Offices, and National Laboratories under the name of Open Energy Data Initiative (OEDI) and store information from a variety of sources: private industry, laboratories, analytic tools, use cases, research reports, and more [21].

Research on open datasets presents numerous opportunities for cost-effective, transparent, and collaborative research across various fields, so it was an objective of the present Special Issue to push for the exploitation of such available information. In the next section, a brief overview of the works presented in the Special Issue, some of them based on the proposed benchmark datasets, is provided hereinafter.

2. An Overview of Published Articles

Based on the presented contributions, and according to the specified requirements, two main areas of interest may be identified: signal processing techniques for feature extraction, and machine learning-driven diagnostics.

Within the first group, the proposed methods include texture features from vibration signals (Paper 1), Teager–variational mode decomposition for incipient fault detection (Paper 2), wavelet representation and denoising (Papers 4 and 10), envelope demodulation (Paper 8) and detrended fluctuation analysis (Paper 9). At the same time, machine learning is prominently featured across multiple papers for fault classification and feature extraction. Particular applications involve support vector machine (SVM) and k-nearest neighbors (k-NN) classifiers (Paper 1), K-means clustering, decision tree, random forest, SVM, and AdaBoost algorithm (Paper 4), and neural networks (NN). In particular, generative adversarial networks for augmenting the available dataset and allowing for the training of further diagnostics NNs are proposed in Paper 5, while a depthwise separable convolution-based network such as YAMNet, commonly used for everyday life sound classification, is applied to machine diagnostics through transfer learning in Paper 6.

Each paper addresses specific applications within industrial contexts, such as induction motors (Paper 1), large rotating metallurgic machinery (Paper 2), turbochargers (Paper 3), bearings (Papers 4, 5, 6, 8, 9), wind turbines’ gearboxes (Paper 7), and pumping stations (Paper 10).

The papers emphasize the practical implications of their findings, such as improved fault detection accuracy and reliability, enhanced maintenance practices and reduced downtime, optimization of operational efficiency in various industrial settings, potential for noise reduction and compliance with environmental regulations and application of advanced data analysis techniques to real-world challenges.

These common themes and main arguments highlight the diversity of approaches and applications in fault detection and machinery health monitoring, showcasing advancements in both methodology and application across different industrial sectors. Also, it appears clear that the most recent artificial intelligence (AI) and deep learning papers are not just focused on classification effectiveness but also tries to solve the well-known issues related to the need of large training datasets with labeled data, which are often scarce in practical industrial applications and may lead to overfitting and generalization issues. In this regard, transfer learning (Paper 6) efficiently utilizes pre-trained models to reduce data requirements and training time, while generative networks (Paper 5) can augment and balance datasets, enhancing model performance and generalization.

3. Conclusions

Vibration monitoring, augmented by advanced signal processing and pattern analysis, is a powerful approach for diagnosing faults in rotating machinery. By transforming raw vibration data into actionable insights, these techniques enable early fault detection, improve machinery reliability, and optimize maintenance practices in the industrial field [22]. As technology advances, further integration of machine learning and real-time data analytics will continue to enhance the capabilities and effectiveness of vibration-based diagnostics and prognostics.