Artificial Intelligence Methodologies in Smart Grid-Integrated Doubly Fed Induction Generator Design Optimization and Reliability Assessment: A Review

Abstract

The reliability assessment of smart grid-integrated distributed power-generating coordination is an operational measure to ensure appropriate system operational set-ups in the appearance of numerous issues, such as equipment catastrophes and variations of generation capacity and the connected load. The incorporation of seasonable time-varying renewable energy sources such as doubly fed generator-based wind turbines into the existing power grid system makes the reliability assessment procedure challenging to a significant extent. Due to the enormous number of associated states involved in a power-generating system, it is unusual to compute all possible failure conditions to determine the system’s reliability indicators. Therefore, nearly all of the artificial intelligence methodology-based search algorithms, along with their intrinsic conjunction mechanisms, encourage establishing the most significant states of the system within a reasonable time frame. This review’s finding indicates that machine learning and deep learning-based predictive analysis fields have achieved fame because of their low budget, simple setup, shorter problem-solving time, and high level of precision. The systems analyzed in this review paper can be applied and extended to the incorporated power grid framework for improving functional and accurate analytical tools to enrich the power system’s reliability and accuracy, overcome software constraints, and improve implementation strategies. An adapted IEEE Reliability Test System (IEEE-RTS) will be applied to authenticate the relevance and rationality of the proposed approach.

1. Introduction

Recently, countries worldwide have been battling the growing energy demand with a wide-ranging misuse of fossil fuels. The world energy outlook dataset from 2021 [1] indicates a steady increase in electricity demand. The statistics show a self-effacing exile from persistent coal-based energy and predict the growth of renewable energies (REs) from below 30% of electricity production in 2020 to more than 45% in 2030. As claimed by International Renewable Energy Agency (IRENA) statistics, in 2021, the generated wind power was 732 GW, 26% of the electricity generated worldwide, and exhibits continuous growth [2]. With the increase in the wind power ingress onto the grid and the extent of wind-generated power increases, further attention needs to be paid to the stimulus of grid-integrated wind generation to ensure electrical power system stability [3]. However, with the stable growth of wind power infiltration, the applicable grid codes are frequently updated and have become complex [4]. Concurrently, wind turbine generators are moving from the ground- to offshore-based locations to moderate the ecological effect and improve ambient circumstances. Due to the high operational and maintenance costs of offshore wind turbine generators, the lifecycle of the DFIG-based wind turbine generation structure is approximately 20–25 years. This is considerably lengthier than the customized engineering norm for the power electronic converters used for DFIG control applications [4].

Centralized electricity generation facilities (multi-megawatt large stations) using a primary energy source are typically situated far away from end consumers and dispense the electricity through transmission lines to a substation; from there, it is distributed to meet end users’ growing energy demands [1]. The phasing out of greenhouse gas emissions by primary energy sources is a vital part of the energy conversion required to eliminate environmental damage [2]. The centralized electricity generation model contradicts the distributed generation model [1]. The distributed electricity generation model involves smaller-scale electricity generation to fewer clusters of end users. Recently, the distributed generation system using REs—either wind or solar power—has experienced a growth in acceptance, including self-sufficiently produced electricity, due to its inclination to use RE generation methodologies [3]. The usage of REs has significantly reduced the overall transmission costs and distribution losses due to the locally installed base being closer to the end users, at a lower voltage or, less often, at a medium voltage level. Presently, the technical standards regarding integrating distributed wind power generation systems into the existing grid are minimal [4]. Hence, there is a need to refine the power system’s technical arrangements using human intelligence, such as simulation processes by computer systems, to incorporate distributed power generation systems without any negative impacts.

Artificial intelligence (AI) appeared as a research subset of computing science, happening to grow in the direction of software engineering in the 1970s due to its reduced diagnostics time and consistent results. AI systems generally work by consuming significant quantities of the categorized training dataset, evaluating the dataset for associations and arrays, and using these arrangements to make forecasts about future circumstances. AI subsidiary machine learning (ML) is now becoming an essential and precise dynamic research arena that acknowledges complex principles and theoretical problems [5]. Electric power systems (EPSs) are a universal approach and have become progressively computerized since the 1970s. In recent times, EPSs endured a revolution aiming to acknowledge social and ecological challenges. The deviations distinguishing such a process assert the standing logical methodologies for electrical power system reliability assessment to their restrictions [5]. Industrial productions have been forward-thinking in terms of advancing more reliable techniques and processes for consistent equipment performance in an approach called “design for reliability” (DFR) [6]. DFR is a procedure involving the design of an element or structure that allows it to achieve the prerequisite stage of reliability. It is intended to be used to understand and resolve any availability complications ahead of time in the designing phase [7].

The researchers in [8,9] described that the deterioration of capacitors causes around 25–30% of power electronic converter failures, and 10–15% of the power electronic converter failures are due to IGBTs. The capacitors and IGBTs are the feeblest links in power electronic converters [9]. The bipolar IGBTs and electrolyte capacitors are significant links in the mainstream electronic power converters. They are subtle to the electrical load, switching pattern, and heat strains and have the core shortcoming of a restricted lifecycle and a higher deterioration catastrophe frequency [9,10,11]. Because of the explicit features and challenges of power converter systems, e.g., great susceptibility in reliability assessment for decaying revelation, etc., there is a persistent requirement for machine and deep learning application-based diagnostic tools in power electronic converters to accelerate collaborative research in multidisciplinary applications [8]. The functionality of the machine and deep learning methods can support the building of appropriate, accurate, and accessible analytic tools using three inclusive stages: the comprising element stage, the power converter stage, and the process stage. Dynamic stress–strength analysis (SSA) is based on the failure of component dependability forecasting [8,9] to screen and detect abnormalities and handle complicated optimization difficulties [10,11] on grid-integrated wind turbine-connected DFIG systems, such as varying wind speeds [12], non-linear load variants, energy dispatch, and load management at the connection point [13]. Here, the base connected load and the wind are correlated random variables. Hence, the cluster procedures that designate a certain wind power for each load value that happens within a specific time measure the time association between the wind-generated power and the connected load [14]. AI methodology uses a database to produce intelligent machine-handling systems that can carry out tasks that usually entail human intellect [15,16]. There are numerous ways to attain AI systems, including machine/deep learning, fuzzy logic (FL), neural networks, expert systems (ES), natural language processing, and robotics. Essentially, AI techniques facilitate decision-making with speed and accuracy. In grid-integrated renewable energy reliability assessment, AI can be defined as mirroring operators’ rational functionalities using workstations to achieve self-healing abilities [16].

In nature, the wind speed is inconsistent and unstable. The fluctuating nature of RE sources is highly dependable on climate conditions [17] and is one of the important features deemed as a downside. To precisely estimate wind power generation, artificial intelligence computational procedures such as ANN can abstract random wind speed dataset characteristics [18] to create desirable forecasts to manage the usage of ANN in wind power system reliability. A comprehensive overview is provided in [19,20,21,22,23,24,25] on AI performances for power converter systems, their distinctive life-processing phases, design, optimization, and preservation, and a discussion about their reliability. The electrical system’s dependability is a portion of its capability to manage customized loads. Theoretically, competency guidelines, such as expected energy not supplied (EENS) and loss of load expectation (LOLE), quantify the electrical power system’s dependability [21,23]. Here, the LOLE is stated as the number of hours or days where the electrical power system cannot provide the required electrical load during the given period. An EENS has reduced energy due to the power generation shortfall [8]. Moreover, in the reliability analysis of conventional power systems, the consistency of power converter-based wind energy systems has been widely studied in [8,9,20,21,22,23].

The review described above provides an assessment and study of the conditioning monitoring techniques of power electronics components. However, it does not debate the new state-of-the-art approach, and there is a lack of a summary of monitoring principles, the IEC 61400-26-1/2/3 timeline, and production-centered accessibility for wind-driven turbine standards configuration to evaluate the reliability of the power system [26,27]. The conception of wind-driven turbine accessibility as “not available” has not yet been well defined by the trade. In 2007, IEC TC-88 and IEA task 33 were arranged to commence work concerning a standard classification of wind turbine accessibility. IEA task 33 was designed to deliver an open policy on the standardization of database pooling for wind-driven turbine reliability and operational set-up and maintenance assessments [28]. The wind turbine accreditation institution provides a general system review using the electrical simulation methods of IEC 61400-27 to ensure that the results work in conjunction with the planned safety and steadiness [29].

The researchers in [16] found that the applications of black-box-type AI practices to load forecasting, grid stability appraisal, and safety barriers can enhance the accessibility and strength of smart grid procedures. An added comprehensive overview was suggested in [30], centering on the accumulative difficulties and reservations in power systems and arguing for a possible optimization solution using reinforcement learning algorithms and their applications. In [17,31] the diverse methodologies and their uses of the machine and deep learning methodologies in fault-finding and analysis in grid-incorporated energy systems were investigated and argued.

As an outcome, this paper aims to provide an assessment and association of the machine and deep learning methodologies to enable systems with intellect. Therefore, it is essential to provide precisely and rapidly functioning AI analytical tools that can be cost-effective and continuous, readily available and accessible, with a rapid diagnostics processing time and increased throughput, thereby increasing the wind-driven turbine generation availability, which includes EPSs such as DFIG reliability systems, uses, procedures, critical performance measures, and techniques. Though other research articles have concisely studied these methodologies, comprehensive assessment and investigation still need to be made available. Section 2 produces a summary and narrative of the IEC 61400-26 standard. Section 3 of this article evaluates the overall composition of the AI approaches, focusing on machine and deep learning methodologies, and describing the frameworks analyzed. The AI functions and techniques applied in DFIG design optimization applications, the merger speed, training period, modeling period, and sensitivity are important elements in studying AI approaches. Henceforth, Section 4 intends to discuss these elements and their importance in the DFIG power electronic converter’s design optimization and reliability. The evaluation of the wind energy conversion system of several AI modeling practices is limited in Section 5. Lastly, Section 6 provides a conclusive proposal for extending the functions of AI approaches and uses in DFIG reliability systems, followed by the future scope of work in Section 7.

2. Smart Grid-Integrated Wind Energy Reliability Challenges Confined to IEC 61400-25-26/1/2/3 Standards

The smart grid-integrated wind power system affects power stability, a conventional variable clarifying the power system elements such as current, frequency, active and reactive power, and voltage provided to the end users in typical operational conditions [32,33,34].

To confirm consistency and adhere to the accreditation of wind turbines’ power supply stability characteristics, the International Electro-Technical Commission (IEC) acted to facilitate power supply stability in 1996 [35]. Consequently, IEC 61400-26 interest was established in the accessibility of wind-driven turbines to build a scientific explanation of availability for wind turbines. The typical operational circumstances of WTGS, though, allow for internal and peripheral conditions and procedures for reporting performing gauges founded on a portion of the timeframe and production [36,37]. IEC 61400-25 and 26 were introduced as an advancing system pulled from IEC 61850 to correspond with wind turbines [38]. IEC 61400-26 obliges applicable configurations and citations. Hence, no deviations from the stand-up IEC standards are allowable. The IEC 61400-26 specification further divides the object into three sub-divisions: IEC 61400-26-1 for time-associated WTGS accessibility, IEC 61400-26-2 for production line-associated WTGS accessibility, and IEC-61400-26-3, for production- and time-associated WTGS availability [36]. Analyzing and evaluating the power supply features of grid-incorporated wind energy converters was established by IEC technical board 88. These days, most wind turbine capitalists supply power feature data accordingly [39]. These certifications are a powerful base for power system services to evaluate the grid-incorporated power quality and grid code concerns of wind turbines [40,41].

Table 1. Reliability standards of the DFIG-connected wind-driven turbine systems.

Type	Description	Reference
IEC-61400	WTG design requirements	[42]
IEC-61400-12	Power curves
IEC-61400-25	Communication protocol	[43]
IEC-61400-26	Accessibility of wind-driven turbines and wind turbine plants	[38]
IEC-61400-26-1	Time-based WTG accessibility
IEC-61400-26-2	Production line-based WTG accessibility
IEC-61400-26-3	Wind power plant availability (time and production)
IEC-61131-3	WTG Automation language	[38]
IEC-61400-22	WTG Project certification	[44]

demonstrates the reliability standards of the DFIG-connected wind-driven turbine systems. The obligations of wind–incorporated power system stability systems and related matters are emphasized in Section 2.1, Section 2.2, Section 2.3 and Section 2.4.

2.1. Fatigue Life Estimation

The inconsistency of fatigue load conditions due to probabilistic airstreams and the isotropic rigidness and robustness characteristics of compound ingredients utilized to attain a firmer, sturdier, and more resilient wind turbine blade design assess the fatigue lifecycle as a thought-provoking mission. More specific to the application of this paper, the authors reviewed the fatigue design prerequisite of the IEC 61400-1 norm and the Germanischer Lloyd (GL) guidelines [44], which stipulate that the wind turbine rotor blades must be produced to endure their designed lifecycle of 20 years [27,28,45]. The wind-driven turbine designing applications in [45] explain how the compound material fatigue categorization conceded out applying the SN graphs and the constant life diagram (CLD) [45]. The SN graphs demonstrate the fatigue life measured in failure phases, N_f, as the functionality of the fatigue load at a specific ratio of stress R. The stress ratio R is described by Equation (1) [45]:

$$R=\frac{{\sigma }_{\min }}{{\sigma }_{\max }}\frac{{\sigma }_m-{\sigma }_a}{{\sigma }_m+{\sigma }_a}$$

(1)

where σ_min, σ_max, σ_a, and σ_m signify minimum stress, maximum stress, mean stress, and stress magnitude, respectively. The distinctive values of R are 0.1, demonstrating an untainted tension (T-L) load, a value of 1, demonstrating tension-compression (T-C) for fully compressed loading, and a value of 10 for clean compression (C-C) loading [45].

Researchers in [46] displayed a straightforward CLD, labeling the association between the magnitude and mean elements of fatigue stress. The number of fatigue phases N_f to catastrophe for an amalgamation of magnitude and mean aspects of the fatigue stress (σ_1,m, σ_1,a) can be calculated by Equation (2) as [46]:

$$N_f={\left[\frac{X+X^{\prime }-\left|2{\gamma }_{Ma}{\sigma }_{1,m}-X+X^{\prime }\right|}{2\left(\raisebox{1ex}{${\gamma }_{Mb}$}\!\left/ \!\raisebox{-1ex}{$C_{1b}$}\right.\right){\sigma }_{1,a}}\right]}^m$$

(2)

where X and X′ are tensile and compressed shield forces f, and γM_a and γM_b are the fractional safety factors in the forte and fatigue-based investigation. The C_1b limits (i.e., C_1b = N₁/m) describe the fatigue graph for the number of sequences N and grade constraint m.

2.2. Availability

Numerous standardized descriptions are supplied by IEC 61400-26-1 on wind turbine availability [47]. The researchers in [48] established that the timeline-basis availability (A_t) of a wind-driven turbine power generation delivers knowledge on the timeshare where a wind turbine (WT) is functioning or capable of performing in association with the actual timeline. The wind turbine’s startup and switching operation can originate from the power supply system’s timeline basis and technical availability. The researchers studied the “System Operational Availability” (SOA) of the IEC-Standard report in this review paper. The entire WT downtime, excluding lower wind speed, is measured as unavailable. The timeline-basis availability is described by Equation (3) [48]:

$$A_W=\frac{{\overline{W}}_{actual}}{{\overline{W}}_{potential}}$$

(3)

The technical availability (A_tech) of a wind-driven turbine is a deviation of the timeline-basis availability. It delivers facts on the time-sharing where a WT is accessible from a technological perspective. Here, the non-available time caused by external factors, such as power grid failures or lightning caused by thunderstorms and heavy rains, is measured as available. In addition, some causes are omitted from the calculation, such as planned preventive maintenance or inevitable accident. The research scholars defined technical availability in Equation (4), which supports the IEC 61400-26-1 definition [48]:

$$A_{tech}=\frac{t_{available}}{t_{available}+t_{unavailable}}$$

(4)

Researchers in [49] reviewed wind-driven turbine reliability based on database resources and challenges concerning using the database. The research scholars discussed the standard definitions of IEC 61400-26-2 in [50]. They established that the SOA is used as a description for production-based availability, as described in Equation (5) [48], identical to the timeline base availability.

$$A_W=\frac{{\overline{W}}_{actual}}{{\overline{W}}_{potential}}$$

(5)

The research scholars also explored production-based availability in [48], which leads to the wind-driven turbine energy produced being associated with the prospectively designed output, and thus focuses on the longest unavailable time throughout the higher wind speed stages and derated operational practices. Therefore, all variances concerning probabilistic and real productivity are presumed to be losses. In computing the production-based accessibility, the purpose of the prospective power is a distinct contest where reasonable airstream computations and power plots are mandatory to achieve rational outcomes.

2.3. Failure Rate

According to the ISO 14224 standard and IEC 60050-191 standard, the research scholars in [37] found that the rate of failure frequency (λ) is the likelihood of an unrepairable element failing within a definite time. If an article is the same as a novel after restoration, it can also be measured using failure frequency design computing. The diverse conclusive rates (year, day, hour) could furnish the failure frequency rate in the event of a continuous failure frequency rate. Equation (6) defines the correlation between the mean time to failure (MTTF) and the rate of catastrophe frequency [48]:

$$MTTF=\frac{1}{\lambda }$$

(6)

The research scholars in [48] established that it is difficult to reinstate the redeemable articles to a condition close to new. Hence, the MTTF is only appropriate for the initial failure of an item. The mean time between failures (MTBF) measures the successive defects. According to ISO/TR 124899:2013 [37], MTBF comprises mean uptime (MUT) as well as mean downtime (MDT), whereas the mean operating time between failures (MOTBF) is a standby for MUT. Both the MUT and MOTBF descriptions are repeatedly interconnected in the study. The difference between MTTF and MTBF involves a maintenance database, particularly concerning whether a measurement was a restoration or replacement. Within this study, such particulars are unavailable, which is why all failures in the system are measured as a general loss.

2.4. Wind Farm Data Automation Architecture

Research scholars have comprehensively reviewed the IEC 61131-3 standard in [38]. They proposed a customized software module that permits dispensing of the database collected in all three categories from the wind-driven turbines, climate masts, and intelligent electrical devices (IEDs) used on substation and production parameters and transmitted periodically to the SCADA system for forecasting and impact analysis purposes. The researchers in [49] reviewed the wind-driven turbine reliability based on database resources and challenges concerning using the database, how the accessibility of the database affected the qualitative analysis of the studies, and how informal they were. The researchers compared the benefits and the limitations of the maintenance activities log, SCADA alarm and fault log system, SCADA’s 10-min operating and accessible database, maintenance team, and spare element cost. The authors established that maintenance activity logs have precise information about failure, equipment unavailability, and repairs expense accessible as hard copies, which are challenging to read. The alarm system dataset comprises failures and periods. However, alarm coding can be difficult to decode and frequently contains numerous alarms or stoppages associated with a similar loss.

Researchers in [38] explained how the integrated modules form the wind farm automation architecture. Here, the rule base module handles the calculation of the counter module, which uses the turbine statuses (IEC 61400-25) and the electrical equipment statuses (using various communication protocols such as DNP3, IEC61850, and Modbus). These intended counters are transmitted in real-time to utility SCADA and historian servers for display in the local subscribed HMI. The researchers established the accessibility of wind farm data, allowing the advancement of automatic reports to examine the influence of specific weather conditions on the farms’ generation. Such a database is needed to lower the cost of integrating wind turbines since it delivers improved tools for data assessment and energy production forecasting.

Researchers suggested some fundamental operational systems based on the principles of database analysis in [50]. Database analysis is applied to resolve the challenges associated with the Internet of Things (IoT) and to target opportunities for further improvement. The IoT has not been deemed a revolutionary trend in the industrial domain, as SCADA systems have allowed connectivity from remote means. Still, the wind industry has seen the enhanced quality and reliability of networking structures, making them available for advanced data management at wind sites. Researchers have focused on precise lost time reporting that can back up the handbook database. However, the log sheets do not continuously cover 100% of events, as unplanned outages can be lost and provide a means for root cause investigation while evaluating abnormalities in turbine performance. The automated lost time case reporting system helps to observe and capture the following factors: the total interval of lost time incidents, the total number of lost time incidents, fault reasoning computer code, personnel comments, and topographical information.

3. Methodology and Review Structure

The reliability, fault diagnosis, and design optimization of DFIG-based wind turbines have experienced remarkable dynamism over the past few years. The current innovation in big data analytical prospects [50], IoT, frame computing, and other factors [51,52] provides extensive, diverse, and valuable information for power converter reliability-centered maintenance systems through numerous stages of their lifecycle. The growth in database size provides a solid basis for using AI in power converters. It is quite evident from Figure 1 that the implementation of AI methodologies such as machine and deep learning has drastically increased in wind-driven turbine-based DFIG reliability-centered maintenance. The number of publications regarding design, fault, and reliability-based maintenance has been growing continuously since 2005. This study included all journal articles to achieve universal quality research expertise and to accomplish more prominence and discoverability to achieve advanced research assessment.

Research scholars discussed the reliability of power electronic systems in [53], including condition checking, reliability, and the forecasting of remaining useful life. Numerous papers that have been assessed over the recent past can be seen in [5,51,52,53]. Research scholars in [51] presented a novel investigation of condition checking and failure recognition in power converters. This review included merely the AI-based parameter identification methodologies.

The research scholars in [54,55] summarized the methodologies used in Prognostics and Health Management (PHM) from databases and electronically abundant arrangements. These studies only discussed the grouping of AI processes in the PHM region. However, the researchers did not address the algorithm details or conduct any comparative analysis. The summary used in the reliability management of power systems in [5] only emphasizes the ML methodology and the maintenance activities, and the desired facts of the AI procedures and their appraisals are unavailable.

Hence, from a lifecycle perspective, this paper intents to seal the gap between an inclusive assessment of the machine and deep learning procedures and applications for the reliability of power converters, and a comprehensive assessment of the available research on power converter reliability utilizing AI methods, which requires a methodical alliance. The scholars studied machine and deep learning procedures in power converter reliability from a lifecycle perspective, and discovered relationships among the appropriate AI algorithms, their basic functionality, and suitable applications.

A timeline mapping the qualitative information of the methodology utilization percentage and the application trending demonstrates the critical indicators of the machine and deep learning algorithms and the associated power electronic applications. The contribution systematically investigates the benefits and restrictions of the machine and deep learning algorithms. The studies provide the standard applications for AI in each life phase and discuss the challenges and possible future investigations.

3.1. Functionalities and Methodologies of AI for DFIG and Power Converter Availability

This section elaborates on artificial intelligence’s functions and methodologies in power converter-based DFIG reliability-centered maintenance. As an efficient, practical-level linking of AI and the purpose of power electronic converters, the scholars classified the vital purposes of AI as:

• Optimization.
• Regression.
• Classification.
• Data structure exploration.

These are beneficial means for advancing aspects of database assessment, such as database analysis, logic, and planning prospects, as shown in Figure 2.

• Optimization: Research scholars have established that the optimization function provides an understanding of the most suitable solution by exploiting or diminishing the unbiased functions, including a selected set of available alternatives that are given limits, independence, or disproportions to satisfy the outcomes [56].
• Classification: Research scholars have established that abnormality finding and fault diagnostics in the maintenance of a specific arrangement task can select fault tags with condition monitoring data. However, the authentic database might consist of noise labeled in [57], comprising non-logical errors [58]. Researchers have also widely studied machine and deep learning methodologies and found that they specifically deal with conveying input data facts with a tag representing the k discrete modules [19].
• Regression: Historic databases transmit much information to consumers. Moreover, the precision of the predicted approach is crucial [59]. Subsequently, an interchange between accuracy and effectiveness is needed for the dataset-activated energy forecasting methodologies [60,61].

The researchers in [62] discovered that the more extended and shorter-term memory (LSTM) and convolutional neural network (CNN)-integrated method breaks the discrete technique in reducing the error while trying to integrate. This method acknowledges the correlation between erratic input and output parameters to estimate the implication of additional active parameters on the expected input parameters. The scholars in [63] found that amongst the polynomial, random forest (RF) procedures, and SVM, the RF procedure predicted the most minor daily error at 0.58%. However, the RF algorithm’s methods are inappropriate for long-term forecasting. For example, in [64] the scholars identified that the regression model could facilitate an intellectual regulator among the varying electrical input- and output-regulated parameters.

4.

Data structure exploration: The research scholars suggested that data structure comprises database gathering. It determines clusters of similar data within a dataset, density approximation that governs the supply of the dataset within the input planetary, and database density that plans a higher-dimension dataset down to a lower-dimension dataset for the futuristic decline.

The applications of machine/deep learning, fuzzy logic, and an expert system are studied and debated more carefully to identify the collective understanding, practical suggestions, and added research scenarios relating to AI means for grid-integrated DFIG-based wind turbine reliability-centered maintenance. In essence, the analysis intends to recognize the suitable uses, set of rules, and AI techniques in the narrative. An appropriate scoping can develop efficient AI tools to effectively detect abnormalities in power systems to increase the reliability of grid-integrated DFIG power electronic converters. These tools will result in complete support and usage in the RE system.

The Pareto assessment presented in Figure 3 reveals several characteristics of AI’s influence on power converter-based grid-incorporated DFIG wind-driven turbine reliability-centered maintenance, the most effective design, and fault finding in the downhill order of 60%, 23%, and 17%, correspondingly. This assessment reveals that most AI function assignments apply to optimization and regression, with a slight emphasis on classification. The scholars assert that the AI approaches are broadly sensitive to the power converter model utilized for the grid-incorporated DFIG-based wind-driven turbine reliability engineering.

3.1.1. Expert Systems

The expert system, recognized as a classical computation program, uses AI technologies such as machine learning and deep learning to implement modern industrial reliability-centered maintenance applications [65]. Elements of the expert system include database checks, a conclusive device, a graphical operator interface, and a rationalization subsystem for deriving a solution and for user conclusion [62]. The attributes of expert systems are extreme implementation, stability, receptiveness, and logic. The research scholars established that the expert system’s methodological features are essentially a dataset that integrates professional knowledge using a matrix of IF-THEN rules in human intelligence analysis, simulated using the Boolean variables given in [17,65,66,67,68,69]. Several exemplary applications can be found in [16,17,65,66,67,68,69,70,71,72].

Figure 4 shows that the usage of expert-based systems in the application of DFIG reliability-centered maintenance is as low as 4%. In principle, the expert system has some constraints in replacing human judgment-making, generating precise output, processing intelligence, and human abilities and acceptance. Furthermore, due to the rapid expansion of the set of rule stages, fuzzy logic, machine learning, and deep learning capabilities can supplement the functionality of expert systems in implications and estimates.

3.1.2. Fuzzy Logic

This segment represents the preferred articles’ elements, involving fuzzy logic assessment associated with the grid-incorporated DFIG reliability-centered maintenance. Fuzzy logic, unlike Boolean logic, is a guideline-based methodology. Although fuzzy logic is a perfect means to challenge system reservations and noisy measurements, in most artificial intelligence-based DFIG reliability uses, a fuzzy logic methodology mainly comprises four elements [73]: fuzzification, rule-based assumption, data-based assumption, and defuzzification. Primarily, fuzzification is accomplished by inputting language-based parameters with membership functions involving trapezoidal, bell-shaped, Gaussian, singleton, triangular, and other custom-made profiles. Furthermore, the reasoning module combines the signal elements corresponding to IF-THEN fuzzy rules in the intelligence base obtained from expert knowledge. Then, defuzzification is achieved on the signal for output. One example of the fuzzy-based practice is the forerunner: if X is medium and Y is zero, consequential: Z is positive. For both the originator and consequential, the research scholars have established the degree of accomplishment by the membership functions. The researchers found that more fuzzy sets were required for the same task for the Mamdani-type system [73] associated with the TSK-type system [74,75] For the TSK-type fuzzy interface scheme, the membership function of the forerunner part is similar to the Mamdani-type system. At the same time, that of the consequential is singleton at several constant values. The membership function in the TSK-type system can be active as direct or steady, which is further robust and precise in non-linear estimation. Research scholars discuss more theoretical particulars of fuzzy logic in [76].

3.1.3. Metaheuristic Methods

Research scholars have established that the optimum solution for a specific application can be obtained by either quadratic or linear programming or a metaheuristic methodology. The acceptable programming methodologies must compute the Hessian and gradient matrices [40], which is thought-provoking for the best optimization tasks in power converters due to their complication. The metaheuristic methodologies [77,78] are established with simulations from a genetic algorithm [79] through the regular selection of an end-to-end tool and an ant colony optimization algorithm (ACO) [80] by aping ants in finding a practical pathway for food. The trial-and-error procedure inspires the investigation of the optimum resolution.

The metaheuristic methodologies are considered trajectory-based methods (tabu search method) [81], population-based methods (genetic algorithm) [79,82,83,84,85,86,87,88] particle swarm optimization [89,90,91,92] ant colony optimization, and differential evolution [93]. The trajectory-based method performance depends on the feature and effectiveness of the definite rule. As a result, the merging speed is slow, and the result is susceptible to local instead of comprehensive solutions for open-ended optimization tasks. Conversely, multiple candidate solutions are generated randomly for the population-based methods. These candidature solutions are merged and substituted with novel candidature resolutions for each iterative assessment. As a result, the population-based methodologies are better in terms of global searching conjunction speed. They are more beneficial for large-scale optimization tasks associated with trajectory-based methodologies. Due to these enormous advantages, population-based techniques are more suitable for optimization tasks in power converters.

Table 2. Applications of metaheuristic methods in power converters design and reliability.

Type	Conventional Algorithms	Applications		Variant	Advantage	Challenges
		Design	Reliability
Population-based methods	Particle swarm optimization (PSO)	[89,90,92]	[91,94]	Velocity clamping	Velocity clamping lessens the step velocity dimension to regulate the element’s development	If all velocity turns into equivalent, the particle will remain to perform searches in a hypercube and possibly stay in the targets. It will still not meet in the regional area
				Inertia weight	A considerable inertia weight at the end of the hunt will raise the merging capability	Attain optimality union intensely inclined by the inertia weight
				Limitation coefficient	Similar to inertia weight	When the process joins, the stable values of the constraints might lead to the needless variation of elements
				Synchronous and asynchronous information	Better merging rate	Advanced output: more refined finite section designs with advanced accuracy (web concentrations)
	Genetic algorithm (GA)	[79,82,84,85,87,95]	[88]	Parallel abilities	It has better parallel abilities	Not effective in resolving modest problems
				Merging	It can enhance numerous glitches such as distinct functions, multi-objective glitches, and constant functions	A lack of suitable application may make the algorithm obtain a result that is not best
					It offers answers that advance over the period	The feature of the final result is not certain
				Implementation ease	GA does not need derivative information	Computational challenges due to repetitive calculation of fitness values
	Ant colony optimization (ACO)	[96]		Capability	Capable of clustering and constructing routes	It is laborious to lay pheromone on trails used by ants as a communication standard
				Implementation ease	Easy to apply	Glitches in parameter selection due to its poor assessment
				Parallel abilities	Inherent parallelism	Experimental rather than theoretical
				Applications	It can be used in dynamic applications	Probability distribution changes by duplications
Solution with AI implementation: 1. Achieves pre-training with an attractive smaller learning rate to achieve rapid integration
Trajectory-based methods	Tabu search method (TSM)	[81]		Capability	Can outflow local targets by selecting non-improving results	The number of iterations can be very high BEST
				Ease of implementation	The tabu list can be utilized to evade cycles and return to old results	A lot of tunable parameters
				Applications	It can be helpful for both distinct and constant solutions	The actual application depends on the difficulty level
Solution using AI implementation: 1. Works on indefinite junction location. 2. Less prone to a spontaneous mergers. 3. Minimum potential to become stuck in the local area.

presents the advantages and challenges of the metaheuristic method.

3.1.4. Neural Network Learning

This section explains that machine learning (ML) and deep learning (DL) neural network methodologies are the most synchronized methodologies among all the artificial intelligence techniques. The researchers in reference [6] assumed that the fault evaluation process on the DFIG modeling is authenticated by combining the state–space vector (SVM) model that can effectively administer the fault event. The researchers in [10] used a pure ANN modeling method. Here, an ANN methodology projected improved data of wind gust spans at mid-energy points and during cut-off incidents at the peak wind gust associated with a hybrid computing fluid dynamics method. The model also suggested the demand for health monitoring for a rotor side converter and fault detection utilizing neuro-fuzzy to ensure that the operation of the DFIG is entirely viable.

ML robotically determines codes and symmetries with know-how from composed datasets or interfaces by analysis and are categorized as:

1.

Supervised learning.

2.

Unsupervised learning.

3.

Reinforcement learning.

1.

Supervised learning (SL)—This is explained by using tagged data (input and output pair) to train a set of rules to accurately set up the plotting and relevant connections between the inputs and outputs. This standpoint is exclusively correct for conditions in power electronic converters in which the formulation of structural shapes is encouraged. SL consists of classification and regression types of assignments. For example, the research scholars in [97] established a typical classification task of fault diagnostics for a multilevel inverter in which a different fault is discovered, given the input fault report. Another illustration of regression is the outstanding functional lifetime estimation of IGBTs [98], where the throughput, i.e., the good due time, is a permanent variable. The accomplished model is equipped to assess new database elements that vary from the training dataset. The model’s ability to trade with more unique database points, i.e., the ones in the analysis dataset, is called generalization.

The enhancements mentioned above are from two aspects of applications in electronic converters configured with different features for users. Usually, the supervised learning methodology is considered method-oriented (i.e., the ANN technique), a probabilistic graphic procedure (Bayesian networks), and a thought-oriented methodology (i.e., conventional and sparse kernel methodology). The primary component of the SL methodology connects with allowing the uncertainty competence in managing the weak signal of the NN to progress the method’s strength [99]. Such competence quality is enabled by absorbing the fuzzy control system logic towards NN labeled as the fuzzy neural network (FNN) and wavelet packet and formants neural networks (WPFNN) or its divergences, e.g., the adaptive neuro-fuzzy inference system (ANFIS) [99]. The resulting element is advantageous to successfully achieving the development of the NN to contest time-series data set circumstances, e.g., illustrative functional lifetime forecast.

Associated with the conventional neural network (CNN), where the network weights are liberated, the transitory performance is enabled by distributing weights between various layers and network cells. The weight distribution is accomplished on a narrow scale with a convolutional structure time-delayed neural network (TDNN) [98]. Generally, the modeling competence of the recurrent unit execution is better than the one with a convolutional system.

The probabilistic graphic methodologies acquire information from the database by visually interpreting input and output sets. The visual interpretation suggests the local dependency association between the variable judgment quantities. The fundamental association in the model is articulated in the Bayesian framework [100], which can be assumed to be a distinctive probabilistic methodology [101]. Therefore, the model’s interpretability is associated with neural network methodologies in a better way. Further, probabilistic visual modeling is better in trade with ambiguity and partial knowledge. When the training of the dataset in the machine learning methodology is accomplished, the training dataset is discarded for the NN and the graphic methodologies. Though the training dataset in kernel methodology is reserved and utilized in the testing phase, the studied knowledge is accelerated by classifying significant database elements in the support vector machine (SVM) [102] or a subset in the training dataset. One distinctive kernel method is the Gaussian approach, which is suitable for the leftover practical life estimation of IGBTs [103]. Here, the conventional kernel methodology (e.g., Gaussian processes) is numerically thorough because the entire training dataset is employed in the assessment phase. Sparse solutions such as SVM and relevance vector machine (RVM) are suggested to prevent the extreme computational burden. The parameter approximation is enhanced and established on Bayesian methodologies. With the sparse resolution, only a subsection of the training dataset is used in the testing phase, and therefore it is more effective than the conventional kernel methodologies. Generally, the requirement of the training dataset for the kernel methodologies is inferior to the neural network methodologies. Therefore, the kernel methodologies are more appropriate for cases with a smaller dataset.

Table 3. Application of supervised learning methods in power converter design and reliability.

Methods	Type	Conventional Algorithms	Applications	Advantages	Drawbacks
Machine/deep learning neural network (NN) methods	Clustering	Feed-forward neural network (FFNN) Radial basis function network (RBFN)	Reliability [104,105,106,107,108,109,110,111,112] Design [113,114]	Compared to feed-forward NN
				1. Simple network shape 2. Greater training speed	1. Longer training time 2. Less sensitive 3. Low ability to manage ambiguity
	Association	Adaptive Fuzzy neural interface system (ANFIS)	Reliability [115,116]	Compared to conventional FNN
				1. Automatic fuzzy logic rule production
	NN with recurrent unit	Recurrent neural network (RNN) or Elman NN (ENN)	Reliability [117,118]	Compared to conventional NN
				1. Improved transient ability 2. Enhanced responsiveness	1. Lower long-term dependence dealing 2. Slower training speediness
		Echo state network (ESN)	Reliability [54,119]	Compared to RNN or ENN
				1. Only hidden output weights need to be established	1. Less issue of gradient explodes
		Non-linear autoregressive network with exogenous inputs (NARX)	Reliability [120,121]	Compared to RNN or ENN
				1. Improved training speed 2. Improved simplification 3. Greater long-term dependence on trading	1. Less sensitive
	NN with convolutional structure	Time-delayed neural network (TDNN) or1D convolutional NN (CNN), FSMPC	Reliability [98,122]	Compared to conventional RNN
					1. Softer time series model
				1. Better liability	1. Composite calculation 2. Theoretical output with arbitrary factors
Kernel-based approach	Sparse kernel method	Support vector machine (SVM)	Reliability [102,123,124,125,126]	Compared to the conventional kernel method
				1. Improved estimation 2. Improved computing effectiveness	1. Probabilistic output 2. Longer training period
		Relevance vector machine (RVM)	Reliability [127,128,129]	Compared to the conventional kernel method
				1. Better sparser than SVM	1. Longer training period
	Conventional kernel method	Gaussian processes	Reliability [103,130,131]	Compared to NN methods
					1. Probabilistic output 2. Uncertain quantification
Way out with AI implementation: 1. Probabilistic findings are overcome with certainty 2. Using a database to estimate the frequency of past achievements and unsuccessful activities 3. Resolve training ranges locally
Probabilistic graphical methods		Bayesian networks	Reliability-centered maintenance [55,61,101,125,126,132]	Compared to NN methods
				1. Better interpretability 2. Intensive computational	1. Probabilistic output 2. Uncertain quantification

summarizes the advantages and drawbacks of supervised learning methodologies in power converter-based RCM and design optimization applications.

2.

Unsupervised learning (USL)—The unsupervised learning tasks are classified as database gathering and data compression. Unlike supervised learning, where the dataset is input and output sets, unsupervised learning does not have an output database for the learning objective during the learning procedure. For the dataset grouping, investigates the symmetries from the spread dataset and categorizes the dataset into many distinct clusters corresponding to their relationships. This way, the database features in one cluster differ from those in another group.

Research scholars in [49,61,132,133,134,135] identified a typical database-grouping application that identifies the distinct health status from the constant degrading database in the condition checking of power electronics. The database density removes the extreme data in the dataset to decrease the number of database structures. Research scholars in [83,127,136] used the principal component analysis (PCA) methodology. Here, the researchers obtained a compact demonstration of the dataset with fewer structures and managed the reliability of the dataset. Typically, these unsupervised learning techniques provide the database processing before the successive evaluation. Even though this stage is not obligatory, decreasing the computing burden and improvising the analysis precision is advantageous.

3.

Reinforcement learning (RL)—In contrast to supervised and unsupervised learning, reinforcement learning does not need a training dataset. As an alternative, it finds an appropriate action that maximizes the incentive for a specific task, effectively an optimization task. Research scholars in [30,137] discovered that a trial run and error procedure prepares this objective-oriented approach from interfaces with simulation models. This way, it accrues a progressive experience and finds a particular system that extends the predefined objective. Ideally, RL is a Markov decision process [137]. The training of RL involves working out a Q-table. The Q-table is an instructive matrix that documents the optimum action for a given specific condition variable, which can increase the total estimated incentives over time.

3.1.5. AI Events Timeline

Figure 5 analyses the statistics of the appropriate AI approaches and their applications in electronic converter reliability systems applicable to REs [138], involving the period when the computing practice is initially suggested. Only the procedures that have the potential to be integrated into power converter reliability are listed; it is noted that:

• Currently, expert systems and fuzzy logic have become more acceptable, particularly expert systems [112]. Until 2005, their hands-on applications were established in the partial presentation of computing calculations and hardware [139], which have been meaningfully developed to date [140]. This fast improvement facilitates the acceleration of the application of several AI procedures for substituting the expert system and fuzzy logic [141].
• Metaheuristic methodologies are constantly developing and are practiced with power converters [77] and utilized for complex tasks with added ML methodologies.
• Bayesian frameworks, including kernel methods and probabilistic models, retain improved simplification and interpretability [142]. Furthermore, their algorithm liability is undertaken with available advanced platforms [143].
• Neural network approaches are the most excellent hands-on sector for AI realization in power converter systems for the following reasons:

• The substantial growth of computational hardware advances the capabilities of neural network methodologies.
• The authors in [144,145] proposed that the NN structure is fairly compliant with merging different applicable AI procedures for system accomplishment.

Figure 5. Review of AI publication timeline.

Figure 5. Review of AI publication timeline.

4. Design and Optimization System

Fundamentally, electromechanical machinery and reliability assessment system research and development are facing the problem that the field of applications is so widespread that only local optima can be found, and new research and development efforts must be continuously undertaken. The research scholars in [146,147,148,149] acknowledge that designing power converters incorporating network topology collection, module sizing, circuit combination, reliability concerns, etc., is fundamentally an optimization assignment [150].

The optimization method includes several steps: describe a system statistically, classify its variables and the terms they must comply with, specify the properties of the system, and then strive for the status of the plan (that is, the values of the variables) that produces the most-needed properties, either maximum or minimum [77]. The standard practice for the design of power systems consists of four stages [23]:

• An empirical interpretation: Empirical tasks are appropriate for maximizing or minimizing objectives. Mostly, the design objectives in power electronic converters consist of element limitation [79,151] weight [152,153] volume [154,155,156] cost [152], heatsink pattern [95,157] area [158,159,160] and power loss [86]. The research scholars in [150] established that it is critical to articulate the anticipated design constraints to several obvious statistical expressions as a particular purpose, as specified in Equation (7), or numerous objectives, as presented in Equation (8) [77,150]:

  $$\underset{x}{\max }f\left(x\right)$$

  (7)

  $$\underset{x}{\max }{w}^{T}f\left(x\right),\underset{x}{\max }f\left(x\right)$$

  (8)

  $$g\left(x\right)\le 0,h\left(x\right)=0,x\in \left[x_l,x_u\right]$$

  (9)
• where g(x) and h(x) are inequities and similarities, X_l and X_u are the minor and the higher limits for ruling variable quantity x, respectively. Here, the maximization can be applied as the objective to the minimization case. Note that multiple goals in Equation (8) can be explained by either maximizing a scalar functionality W^T f(x) by weighting numerous objectives altogether or by directly enhancing the objective vector function f(x). The researchers illustrated a common 2D Pareto optimization problem in [150] in a graphical manner about capitalizing efficiency while curtailing cost. Here, the researchers highlighted an enhanced point as a Pareto front of the optimum solution. The study explains the signs that give the Pareto-optimal facts, repeatedly termed as non-dominated and dominated points. It is up to the designer to first-rate one of these proposals by trading cost vs. efficiency. Space constraint: The research scholars in [152,153] outlined that the space constraint in the designing aspect is subject to the reasonable spacing, limits, association, and conditions that the design function meets the linear or nonlinear similarities and variations. The functional model constraints, shape, capacity, life expectancy attributes, costing, etc., are derived from the design space constraints.
• Discovering a solution: A well-defined optimization problem is to extend (or reduce) objective functionalities by altering the outcome variable quantities in the constraint spaces. AI methodologies, particularly metaheuristic techniques, can be employed here.
• Performance evaluation: The research scholars in [148,149,150] established that using simulation hardwired loop testing and a practical prototype, the applied solution can be assessed contrary to the predefined objectives. They referenced the results of previous steps for further performance improvement and optimization.

The designing and optimization task is an iterative aspect trial run and inaccuracy procedure used as an alternative to a progressive approach. The task is reformulated based on the assessment outcome at every phase, e.g., altering the objectives, adapting the constraint space, and reconfiguring the programming methodologies. Standard design in power electronic converters is time-consuming and requires numerous iterative steps. For instance, the element alignment and the model selection depend on professional knowledge and perception without abundant measurable references. This way, the design performance will come together gradually to meet the mandatory standards. AI methodologies can alleviate this drawback by applying an empirical interpretation for the design time reduction and solution assessment for the model and optimization.

4.1. Design Timeline Saving

The research scholars in [114] recognized the prerequisites for the design objectives. These requirements are to be enhanced if its assessment is computationally thorough. One of the usages of AI methodologies is substitute modeling as the actual formulation to reduce computational efforts. The AI-based substitute model functions as a substitute in the iterative aspect of the designing procedure that substantially eases the computational endeavor.

The research scholars reviewed the application of the design for reliability (DfR) in [113,114,120]. The authors in [120] reviewed the superior tackling capability of time series data on thermal modeling. They proposed the nonlinear autoregressive network (NARX) for power converters’ thermal modeling, allowing for the thermal cross-coupling impacts. The proposed NARX-based modeling outcome has a faster time response on task completion of 109 s than the conventional modeling time of 1005 s. In addition, the proposed NARX model resulted in a minor error value of less than 1 degree centigrade between the estimated and actual temperatures. The authors in [113] applied a feed-forward neural network (FFNN) to conceptualize the element behavior modeling of MOSFETs without an in-depth understanding of the design configuration. Under steady-state conditions, the authors established the drain-to-source voltage, output current, junction temperature, and gate-to-source voltage variables using the neural network. The applied model significantly enhanced the design replication procedure with equivalent precision.

The DfR application in [114] used two feed-forward neural networks (FFNNs) to design the power electronic system reliability. The first FFNN served as a substitute model following the thermal attributes of power electronic converters to plot the design constraints to the correspondence data variations of the semi-conductor junction temperature. Consequently, the second FFNN was applied to map the annually objective outlines to the corresponding lifecycle consumption. In this way, the authors quantitatively described the nonlinear association between the proposed boundaries and the annual lifecycle utilization, which can accelerate the iterative design process.

4.2. Design Modeling and Optimization

The research scholars in [77,87,95], established that the design modeling and optimization of power converter systems is around stipulating the network topology, element model, and model parameters. These parameters are system weight, dimensions, and operating and design frequency, resulting in ideal attributes given the model’s design limitations.

The authors in [95] discussed the genetic algorithm (GA) in a finite element analysis for the automatic heatsink model design to create a dense cell format of the heatsink. The authors could not use more cell designs here due to the substantially expanded solution space, the potentially high-pressure drop, and concerns about the lengthier calculation time. The simulation and experimental results found that the planned methodology accomplishes a heatsink volume reduction of 27% and a 6% lower device junction temperature.

The research scholars in [86] formulated a multi-objective optimization task by planning a 500 KW solar power-based microgrid to amplify the standard power distribution, decrease the system mass, and provide operation throughout the crucial failure modes. The system explored the parameters’ optimum value for the multi-objective optimization tasks, including battery terminal voltage, maximum power, MPPT voltage, and panels per string.

The researchers in [90] applied particle swarm optimization (PSO) to the power electronic circuit (PEC) amalgamation. Here, the optimal elements valuation is utilized for its easy application, faster merging, and higher global hunt ability, which can substantially enhance the effectiveness of PEC optimization and provide highly optimized values of circuit elements. In this case, the simulation results show that the PSO produces an exceptional solution with less computing effort than GA. The PSO algorithm is proficient in resolving comprehensive optimization difficulties with constant variables. Its usage in the operational reliability arena presents not only the benefit of its application ability but also the likelihood of achieving the design fact and the catastrophe possibility with decent accuracy. The low computational time of the zero-order algorithm greatly benefits the resolution of optimization challenges. The reliability evaluation for probabilistic restrictions frequently includes an iterative process; therefore, binary-loop procedures are used in a probabilistic optimization. A new methodology, sequential optimization and reliability assessment (SORA), are also used to enhance the effectiveness of a probabilistic design. The SORA methodology engages a single-loop approach where sequential optimization cycles and reliability evaluation are employed to modify the limits of disrupted deterministic limitations (with low reliability) to the possible way established on the reliability information acquired in the preceding cycle. Therefore, the strategy is rapidly enhanced from process to cycle, and the computing effectiveness is enhanced considerably.

Local search algorithms are among the standard methodologies for resolving severe problems from the numerous capacities of artificial intelligence [161]. For the local search stochastic for the satisfiability problem (SAT), the most effective and influential algorithms are founded on stochastic local search families such as GSAT and WalkSAT architectures. GSAT marks the alteration that reduces the number of displeased sections in the new task. In contrast, WalkSAT chooses a displeased section using the present study and then flip-flops a parameter within that section. These algorithms perform using a standard set that comprises occurrences from randomized allocations as well as SAT-encoded difficulties from several domains. Stochastic local search (SLS) methodologies can also solve countless real-world difficulties that frequently include large-scale occurrences at rational computation outlays while providing good-quality solutions [162]. The authors also studied a novel empirical SLS procedure termed an “adaptive variable wisdom” SLS for PMSAT problem solving, established on an active confined search framework such as parameter tuning and variable depth neighborhood search (VDS). This empirical evaluation proves that VDS is a more efficient WRT single neighborhood search. Another novel method using local search for justification along with original finishing time heuristics is suggested by researchers in [163]. A local search stochastic for satisfiability checking meaningfully improves the development performance in terms of the make-span and schedule length ratio (SLR). A new task arrangement heuristic, the local search on critical path (LSCP) model, is assembled on heterogeneous earliest finish time (HEFT) heuristics and, through simulation results, has shown an improved performance associated with an iterative greedy algorithm for local search.

5. Reliability-Centered Maintenance

Reliability-centered maintenance (RCM) is an ongoing systematic procedure that confirms that maintenance tasks are achieved in a well-organized, reliable, economical, and safe way to maximize overall reliability. Maintenance tasks may be predictive maintenance (PdM), preventive maintenance (PM), run-to-failure (RTF), real-time monitoring (RTM), proactive maintenance techniques, and non-destructive inspections to identify or monitor flaws in the system. In this framework, the maintenance activity provides a framework capable of increasing the equipment lifespan, or the mean time between failures (MTBF), along with a potentially expensive restoration, the plant’s overall performance, product quality, safety, or ecological reliability [164]. The RCM process must meet the following technical criteria: (i) IEC 60300-3-11 [165], (ii) SAE-JA1012 [166], and (iii) SAE-JA1011 [167]. The failure of the equipment is a fundamental feature of overall equipment efficiency (OEE). OEE deals with the efficiency of different assets in the installation base [168].

$$OEE=availability rate\times performance rate\times quality rate$$

(10)

The consistency and protection of power electronic modules, converters, and arrangements are significant for ground uses in DGIF reliability systems. The research scholars measured the reliability features in [169] during the initial project stages and control stages. Power electronic converter systems assume numerous endangered and uniform calamitous disasters because of the compound and acute ambient situations [169,170]. The authors identified that preventive actions, condition monitoring, abnormality diagnosis, failure analysis, RUL forecast, etc., are active methodologies to confirm that preventive planned tasks are appropriately implemented. The research scholars used prognostic metrics in [169,170] to quantify the usefulness of the selected PHM scheme. They used metrics to assess prognostics, a thru function of the application arena. The metrics are centered on algorithm presentation (accuracy, sensitivity, robustness, etc.), computational density, and cost-effective metrics. The researchers systematically explained the preventive maintenance system repair actions in [171,172] in power electronic converters that comprise the below sections:

• Offline training and knowledge sharing: It participates in numerous features of information comprising a historic observing dataset, model dataset, enhanced mature test experimentation, and fault mode and effect analysis (FMEA). Moreover, the authors in [172] used collaborative methodologies in the section on execution enhancement.
• Condition checking and healthiness valuation: This section connects the conditional checks of the component to the online and offline condition checking in-ground applications. The functions of parameter documentation do not involve introducing test instruments into the test body, dataset scrubbing, characteristic removal, abnormality observation, failure analysis, and RUL forecast. The authors customized the disconnected pattern with the provisional element between the model variable fine-tuning levels by regulating the ground-functioning capacity to ensure the removal of sensitive information on decision-making from the constant condition-check statistics.
• Administration and decision making: In this section, the authors deal with the compassionate understanding of a health review on ideal decision-making. This health review feedback can improve the governing strategies to improve system usage given the real-time asset health condition. Subsequently, the authors broadly debated the practical AI applications in maintenance.

5.1. Condition and Health Monitoring

Condition and health monitoring (C&HM) is an efficient way of improving accessibility and controlling the lifetime value of power electronic components, converter components, and the overall structural system of the asset. Actual health analysis delivers benefits such as enhanced safety, better-quality reliability, and the economical operation and maintenance of compound engineered systems [173]. The research scholars in [52,174,175] identified that the research dynamics in the condition monitoring information of power electronics, including system variable evaluation, dataset processing, and characteristic extraction, act as a base for the related PHM uses. A recent ECPE industry-wide study [176] indicates that the active components and capacitors are the most common in power electronic converter failures [177].

5.2. System Parameter Identification

Identifying reliability system parameters deal with the maintenance modeling and the failure understanding tool. It delivers a guideline to combine the essential competence and conditional exchanges to enhance the system design. The authors observed the structure’s design tools in [77,178] and utilized them to improve how the system parameter recognition deals with information obtained for crucial elements. These tools have evolved specific hardware requirements for parameter recognition to enhance the quantity and position of the instruments in energy conduction. The authors allowed actual-time “what-if” analysis to regulate the trade-off influence in [151,175], given the challenges of a very intimate space for elements in a power segment, faster IGBT switching frequency, comparatively irrelevant variable deviations relevant to the capacitor and IGBT aging, etc. The authors considered the system parameter identification a pattern-based and pattern-free methodology, seeing whether the system-changing aspects and patterns are essential. The obvious way out is a non-invasive methodology, where the data of choice ultimately settled on the accessible trending signaling parameters. Consequently, sensorless condition monitoring is a cost-effective and favorable solution for industrial specialists.

The researchers in [104] identified that a pattern-free system dynamics method does not need the previous data information. Fundamentally, it connects with the regression competence of AI procedures to perceive associations among the inputs and targets. Training an FFNN example on a bridge rectifier is explained in [105,106]. The authors illustrated the potential of an FFNN assembled by the frequency domain facts of the DC-link voltage ripple [100]. This way, the association between the drive current and DC-link ripple voltage as input signal and the DC-link capacitance is recognized. Hence, the capacitance is ultimately concluded.

The researchers in [117] suggested that an impedance recognition methodology should be established on an RNN to allow the steadiness assessment to form a pattern to make a similar target to the A-phase current i_a as the physical system specified similar inputs, including three-phase voltages v_a, v_b, and v_c. Consequently, the RNN-founded design holds similar frequency features achieved for impedance recognition, as the physical one prevents the process from being disturbed. The authors practiced an enhanced control logic in [179] to overcome the challenges of conservative boost converter current source emulating the disruption experienced by load deviations and uncertainties in practical applications. A four-layer FNN comprising the input, association, instruction, and output levels is accepted to carry through the AFNNC arrangement. The input level conveys the input language parameters x_i (i = 1, …, n) to the following group, in which x_i represents the input components of the AFNNC control arrangement. The association level signifies the input standards with the following Gaussian membership functions:

$${\mu }_i^j\left(x_i\right)=\exp \left[\frac{-{\left(x_i-m_i^j\right)}^2}{{\left(c_i^j\right)}^2}\right]$$

(11)

The ruling layer performs the fuzzy logic interpretation, where every node in the level grows the input signals and targets the product outcome. Moreover, the FNN target is rephrased as:

$$y_0=u_{FNN}\left(x_i,w,m,c\right)=wl$$

(12)

The projected AFNNC arrangement contains the primary FNN control and related system variable fine-tuning procedures, which result in the intellect of the Lyapunov stability theorem [180], and the forecast procedure [181] to confirm stable control performance as well as network convergence. However, the dataset gathering is laborious and expensive.

The researchers established an improved pattern for the healthiness checking of MPF capacitors and supercapacitors (SCs) on the NFN model [116]. They suggested evaluating the supercapacitor’s capacitance and equivalent series resistance (ESR). During the observing period t, the ANFIS inputs comprise the source voltage v_t, the supercapacitor temperature Ɵ_t, and a time series ESR_{t−400:100:t} containing earlier ESR dataset facts. The ANFIS output is the ESR approximations in the forthcoming p stages [116]. The practical study specifies that the ESR of the supercapacitor can be precisely assessed, and the standardized root mean square fault of the ESR estimate is as little as 0.025 at a condition check time of 2600 h. An additional class of system variable credentials is the pattern-oriented method [77,178]. The authors acknowledged the system physics and patterns in the early stages and expressed the credential pattern with unknown variables. Here, the metaheuristic method is utilized as an improviser to identify the best outcome. Several techniques, such as PSO [91], the crow search algorithm [182], and GA [88], can be used. The authors identified that a review of the precision and strength in the compound ambient is needed for system variable credential methodology in power electronics. They recognized that the number of mandatory datasets for the valuation was compacted for the pattern-oriented methods due to the union of system dynamics and patterns. In [134], the power MOSFET element is measured as unsuccessful if there is an upsurge of 0.08 for the degrading indication of drain-to-source on-state resistance RDS (on). Consequently, additional research is essential to discover the sensibility of the AI-based system variable credential methodologies in computing accountability for field requests.

5.3. Anomaly Fault Discovery and Failure Analysis

Automated fault discovery (AFD) systems can help alleviate many unexpected machine faults by monitoring sensor data and flagging anomalies before the problems become more serious [183]. The anomaly detection builds a binary conclusion and identifies any uncharacteristic behavior. It indicates when the evaluated system features or insignificant parameters surpass the rated specifications [184]. Once an abnormality occurs, the fault analysis [51] recognizes and discovers the detailed failure modes later. The authors categorized the anomalies as point anomalies and contextual anomalies. While the point anomalies have limited violations of standard operating parameters, contextual abnormalities are only detectable concerning the surrounding values or readings from other sensors [183]. Abnormality discovery and failure diagnosis consist of time series forecasting, classifications, regression, or clustering jobs. Since it is established on the knowledgeable association using a training phase, it regulates the failure tag on the occurrence of a new fault signature. The authors considered the abnormality discovery and failure diagnosis methods as subsections of supervised learning and unsupervised methodology [74,185].

5.3.1. Supervised Learning Methods

The research scholars in [107] explain how an integrated grid system’s variable loading situations impact the fault diagnosis algorithm. Hence, a signal processing method is proposed that reduces the essential particulars for the failure illustration and suppresses the influence of the loading alteration. The computational load of the planned methodology is decreased to 10% of that of the current procedures to use FFNN as a diagnostic category device. Failure discovery notions comprise ANFIS to decide the degree of seriousness of a filtering capacitor in the DC link [115].

In [186], scholars used a kernel function in SVM to plot the novel model to the highest dimension featuring the capacity to attain the optimum lined category plane. The radial basis function (RBF) consists of fewer parameters with improved outcomes than different kernel functions. Consequently, the study used the RBF kernel function. The equation is as below:

$$k\left(x,x_i\right)=\exp \left(-\frac{{\Vert x-x_i\Vert }^2}{{\sigma }^2}\right)$$

(13)

The parameter selection penalty factor C and the σ in RBF-SVM training define the classifier presentation. Hence, it is essential to select the best C and σ. Here, σ indicates the breadth size of a kernel function. The authors in [119] suggest a multiple-switch failure analysis procedure for an inverter voltage source. An echo state network (ESN) is employed to classify the low-frequency fault dataset. The switching-state-featured datasets are used for ESN input and system training. It is noticed that ESN is an enhanced version of RNN, whose hidden layer is changed with storage to evade gradient discharge. The diagnostic show of ESN is associated with FFNN, in conjunction with RBFN, and a wavelet activation function. It specifies the ESN’s greatness in sensibility, training speed, and designing procedures. The discrete ESN model is defined as [184]:

$$x\left(n+1\right)=f\left(W^{in}u\left(n+1\right)+Wx\left(n\right)+W^{fb}y\left(n\right)\right)$$

(14)

$$y\left(n\right)=g\left(W^{out}\left[x\left(n\right);u\left(n\right)\right]\right)$$

(15)

In [187], the research scholars used a 1-D convolutional neural network (CNN) for the failure analysis of segmental multistage converters. The feature mining and analytical category of the 1-D CNN model are combined to allow the failure symptomatic in the new dataset. The empirical outcomes show that the planned methodology is exceptionally consistent in delivering a finding precision of 98.9% and a failure detection accuracy of close to 99.7% within the 100 ms timescale. In [111], the research scholars suggested an FFNN to form the intermittent association of the inputs and targets of a full-bridge diode rectifier. Study [111] indicates that the FFNN training used as the regression means is completed at the normal operation mode of the rectifier. Consequently, the ideologies and plotting association between the input voltage v_i(t), input current i_i(t), target current i₀(t), and the target signal of the target voltage v₀(t) is categorized, considering the alphanumeric simulator, indicating the usual operational way of the bridge rectifier. Once the observed rectifier output voltage meaningfully departs from the FFNN target, it runs the rectifier irregularly to ease the abnormality identification.

The authors used a relevance vector machine (RVM) in [127] to diagnose a cascaded H-bridge multilayer inverter failure and used principal component analysis (PCA) to abstract the failure indicator characteristic. The practical study specifies that the RVM has a lengthier training time than the SVM and overtakes the SVM and FFNN with 100% analytical precision.

5.3.2. Unsupervised Learning Methods

The need for improved accuracy and flexibility in the condition monitoring approaches of electrical machines has prompted the transition from model-based to data-driven approaches. A k-means modeling approach was used to develop diagnostic models for the machine under different incipient fault conditions. The clustering results indicate that stator current harmonics provide the best results among the single-signature harmonics feature sets [186].

The study in [167,188] proposed an unsupervised trajectory anomaly density detection approach based on the clustering algorithm DBSCAN30 to cluster and the cosine similarity to measure the distance. DBSCAN can find clusters of any shape and irregular groups without knowing the number of sets in advance. This model approach is conducive to reducing the algorithm’s complexity and has practical significance. The cosine similarity of the mixed feature sequence is:

$$F{M}_i \mathrm{and} F{M}_j, 1\le i,j\le n.$$

(16)

$$S_{ij}=\frac{F{M}_i\cdot F{M}_j}{F{M}_{i}F{M}_j}$$

(17)

In [135], PCA has suggested the abnormality detection of SiC MOSFETs. This detecting apparatus is comparable with [93]. Numerous arithmetical structures, comprising kurtosis, skewness, etc., are chosen as PCA procedure inputs, and the target has a few compressed characteristics and an alteration matrix. The recently accessible dataset is utilized in the alteration matrix for the abnormality table calculation of ground uses. Exceptional performance is advised as soon as the abnormality table surpasses a defined edge. The methodology is confirmed by a central processing unit familiar with the experimentation. Similar learning methodologies in abnormality recognition and failure analysis, including SOMs and k-means, are suggested in [126].

5.3.3. Discussion

Table 4. Summary of the distinct AI procedures on abnormality detection and failure analysis.

Methods	References	Fault Diagnosis	Accuracy	Applications	Benefits	Drawbacks
Support vector machines (SVM)	[123]	Yes	98.1%,	Reliability-centered maintenance (RCM) and health management	1. Simple implementation 2. Better interpretability 3. Improved diagnostic accuracy 4. Discovers more fault patterns 5. Better optimization	1. Sensitive to outliers 2. Enhanced training required
Fuzzy logic (FL)	[74,190]	Yes	100%			1. High cost
Eco-state Network (ESN)	[119]	Yes				2. Enhanced training required
Solution with AI implementation: 1. Replace outliers with a suitable value using the quantile method
Relevance vector machine (RVM)	[83,127,191]	Yes	97.30%	RCM and health management	1. Can only be cast off with clustering and for smart moves	1. Enhanced training required
Solution with AI implementation: 1. Substitute database minimization methods with filtering and normalization attempts

summarizes the structures of distinct AI procedures and their variations in abnormality detection and failure analysis. Each AI algorithm has benefits and limitations. To obtain the benefits of each process, it is necessary to associate various algorithms with a governance blend to advance the analytical precision and sturdiness. The tabled summary indicates that the SVM methodologies are more beneficial in terms of simple implementation, better interpretability, improved diagnostic accuracy, ability to discover more fault patterns, and better optimization for the reliability evaluation of the wind turbine power converters. Considering the dynamic capability of RNN, the stability analysis of the power converter system is achievable over a wide frequency range. The empirical analysis of the ANFIS and FFNN model indicates that the smallest equivalent series resistance of the supercapacitor can be accurately estimated over a wide frequency domain of the monitoring time. Governance blending on IGBT failure analysis and more collective methods associated with multiple procedures are available in Chapter 14 [1], [125].

Numerous AI approaches and variations are effectively used for abnormality identification and failure analysis. Relevant concerns regarding the type of available datasets for various uses have been recorded as a significant feature of the hands-on services of AI. The “Prognostic Maintenance Toolkit” is accessible in MATLAB [189], containing numerous abnormality identification and symptomatic procedures. The AI perspective indicates an insignificant change between power electronic equipment and additional manufacturing capacities (e.g., electromechanical uses) on abnormality identification and failure analysis assignments [184].

The use of AI is advantageous for methodical progress and formal investigation. The vast majority of the methodology is related to the equivalent presentation, which concerns evaluation precision. Additional research needs to be dedicated to the space between theoretic procedures and hands-on applications:

• In addition to singular element failure, the fault mode of several elements breaking down concurrently also has to be measured. The interdependency and linking effect amongst the failed elements should be integrated into the analytical procedures.
• In particular, the difficulties in the dataset acquirement of power electronic converter systems and the drilling database for practical use are usually restricted. This situation is much worse for a database when there are unstable failure labels, i.e., the sample dataset of the usual process and the insufficiency of the dataset with failure tags because of disastrous catastrophes. Therefore, the algorithm applies to a given partial dimension of the database, and an inferior database should be examined.
• The practicability, computational load, adjustable ability, sturdiness, the designing and correcting of the algorithm, and application budget must be extensively measured [119].

5.4. Remaining Useful Life Prediction

The typical network of a wind energy system (WES) founded on a doubly fed induction generator (DFIG) consists of the turbine, the geared transmission, the DFIG, the power transformer, and the power electronic converters [61]. Lifespan estimation during the designing stage supports reliability, which states the character of a population of elements. The RUL does not forecast the lifespan of a collection of parts; rather, it forecasts the remaining lifespan of a specific element in facility-oriented condition checking information [192]. There are similar reservations for lifespan forecast, pattern standardization faults, industrial acceptance, variants of operative ambient, and amount of effort. These reservations result in imprecise reliability evaluations for a specific element in field applications [192]. Typically, the assignment shape of a unit is categorized as longer-term, medium-term, and shorter-term cycles. The research scholars in [61] explained the systematic processes for the RUL forecast for an electrical catastrophe that seems roundabout in milliseconds and is categorized as a shorter-term catastrophe. Hence, safeguarding the WES from electrical disasters by exploring a predictive methodology is essential. RUL forecast is used as an added tool to ease the reservations for consistency, guard, or accessibility uses based on physics-based, dataset-based, and hybrid-based methods.

The research scholars recognized the probability density function (PDF) regression pattern based on a historic database [193]. The PDF of the degradation level derives from the PDF of RUL. Given that the system is adequately operative at condition checking time t, its RUL 1, a random variable, is well defined as the remaining lifespan during the degradation process D(t) exceeds the failure threshold w.

$$l=\inf \left\{l:D\left(t+l\right)\ge w|D\left(t\right)<w,D_{1:j}\right\}$$

(18)

Here, D_1:j indicate the collective condition monitoring details relevant to the time t. Besides its predictable significance, the unpredictability indicators with the top and bottom self-reliance intervals (l_lo, l_up) are also of outstanding significance. The AI methodologies in the RUL forecast are usually traded with an illogical regression within the deterioration information [54]. Hence, an equivalent RUL is established on the training database in [193], and the degradation arrangements can be categorized and learned. They can be openly anticipated as specified in the regression pattern to enable the upcoming deterioration stage forecast, resulting in an assessed RUL. In [131], an experiential deterioration pattern is proposed with the deterioration method ∆R_DS(ON) noted during the five training apparatuses.

$$\Delta R_{DS\left(ON\right)}=\alpha \left(e^{\beta t}-1\right)$$

(19)

where t is denoted as time and α, β are model limitations that could be fixed or assessed online during the Bayesian tracking frame. An experiential deterioration pattern is used to pattern the deterioration procedure when a physics-centered deterioration pattern is unavailable. The pattern variables are dissimilar for various devices. Consequently, the defined state variables α and β need to be assessed online to confirm that the accuracy of the experiential degradation pattern is modeled as a dynamic system.

$$\mathrm{Let} R=R_{DS\left(ON\right)}, \mathrm{then} \frac{dR}{dt}=R\beta +\alpha \beta .$$

(20)

In [54], an echo state network is used for the RUL forecast of power MOSFETs. The input to the echo state network is the deterioration pointer drain to basis on-state resistance ∆R_DS,(ON) at times k, and k − 1 the target at a time k + 1 is the ∆R_DS,(ON). An element filter is used to frequently refresh the target weights when an established element’s novel condition-checking dataset is accessible to accommodate the echo state network. Alternative neural network methods linking a time-delayed neural network on the RUL forecast of IGBT devices are available in [98].

In [103], Gaussian procedure regression is used for the RUL forecast of IGBT devices. For the deterioration pattern, the nonlinear association between the decrease of on-condition collector-emitter voltage ∆V_ce,on and the condition-checking timeline is fixed by the Gaussian procedure regression. Since the Gaussian process is expressed using the Bayesian context, essentially, it can forecast the ambiguity of deviation ∆V_ce,on. The error bar of the development ∆V_ce,on is established to utilize the calculation of the assurance recess of RUL. The authors in [194] have established that a maintenance plan is well defined as a developed maintenance methodology to accomplish maintenance objectives (DIN EN 13306:2015-09, 2.4). Such maintenance strategies determine the needs to be assumed, the element on which it is to be carried out, the occurrence at which it must be completed, and on what schedule it needs to be done. Overlooking maintenance could lead to an extreme number of expensive catastrophes and deprived system performance, resulting in compromised reliability. However, carrying out maintenance too frequently would improve the system’s reliability but at a more significant cost of care since the residual useful life (RUL) of the element would not be thoroughly utilized.

5.4.1. Ambiguity Presentation and Clarification

The initial pace of RUL consists of the understanding and presentation of uncertainties, directed by the option of patterning and simulation structures. Some shared theories comprise probabilistic theory, classic set theory, fuzzy logic theory, fuzzy measures (plausibility and belief), and rough set theory (top and bottom estimates). Among these, the probabilistic approach is extensively applied in the domain of PHM.

5.4.2. Ambiguity Quantification

Compared to further regression-associated assignments, the competence of uncertainties assessment is vital for RUL forecasting. Research scholars in [19] indicated the RUL as an arbitrary variable; therefore, a measurement of the assurance interval is essential for governance. These reservations originate from data size variations, quantification noise level, variable operative settings, etc., broadly measured for a hands-on outcome. AI methodologies are relatively attractive for these uncertainties, with the measurement of forecast outcomes seeing the black-box characteristics. Numerous possible methods are discussed, including the Monte Carlo techniques in [98], Bayesian-centered AI approaches (e.g., Gaussian procedure, RVM), and the stochastic dataset-based methodology [136,170,195] that can essentially deliver the PDF of the RUL for computing the assurance interval.

5.4.3. Uncertainty Propagation

The first step involves spreading the numerous causes for uncertainties throughout the forecast pattern. The next step is calculated by assessing uncertainties within the upcoming conditions alongside a Boolean threshold function to specify an element’s fatigue status. The last step is uncertainties spread, which is primarily applicable to predictions accounting for all of the formerly measured uncertainties and practices. These statistics forecast (1) futuristic conditions and the relevant uncertainties; and (2) outstanding valuable lifetime and the relevant uncertainties.

5.4.4. Adaptive Capability

The uncertainties pattern and the relevant parameter-tuning level interconnect the online and offline models, which is a crucial stage for hands-on requests. Suppose AI methodology lacks an adjustable ability. In that case, its use is restricted. One precondition is that the training and test dataset is produced within comparable circumstances and contributes to the highest range of comparison [196]. For power electronic components, the operational settings of the testing dataset are very unlike the training data, usually achieved along with enhanced test experimentations. The mainstream approach to the study [103,124] involves ensuring that the operational limits of the existing system are equal to the training data, which could differ in field usage. Therefore, the AI-centered RUL forecast methodology’s adjustable ability is vital to linking theoretical study and built-up engineering requests. Similar encouraging ways of pattern limit fine-tuning comprise the clear plotting association derivation [197] and transfer knowledge [198,199] of deterioration features within numerous operational limits (voltage, humidity, temperature, etc.).

Table 5. Different kinds of abnormalities in the power grid incorporated the DFIG reliability system.

Abnormalities in the Smart Grid-Integrated Power Converter-Based DFIG	Configuration	AI-Based Solutions	Sensitivity	Speed
Uncertainties prediction	Patterning and simulating structures	Probability theory, classic set theory, fuzzy logic, rough set theory	Moderately sensitive	Sensible speed
Uncertainties quantification	Regression associated assignments	NN, Monte Carlo methods, Bayesian approaches, RVM, stochastic-based methods	Less sensitive	High speed
Uncertainties propagation	Forecast	Boolean threshold	Moderately sensitive	Sensible speed

summarizes the different abnormalities in the smart grid-integrated DFIG reliability system.

6. Overview of AI for Power Electronic Reliability-Centered Maintenance Systems

This paper aims to evaluate the wind energies transformation system, underlining its electrical reliability aspects, including records on reliability-centered maintenance. The study explores the newest technologies relevant to wind energy systems and upcoming study routes. Based on the algorithm’s outlook, it is essential to examine the characteristics of the machine and deep learning methodology relating to the various life-span stages of electronic power converters. Employing an electronics power converter system in a power grid-integrated DFIG-based wind-energy turbine, detailed illustrations are used to demonstrate the necessities of AI methodology for the various life-span stages of the RCM aspects.

The metaheuristic methodologies are functional to the optimization that includes an iterative trial-and-error process. For the heatsink designing of a power converter, an enormous number of conclusive variables, e.g., design, capacity, and burden, need to be determined, which is fundamentally an optimization assignment. Though the computing work is rigorous, the designing task is usually achieved in offline mode. There is a lesser prerequisite on the algorithm speed in this case. Though the metaheuristic methodology-based optimization does not confirm a universal resolution, the suboptimal heatsink design is still more acceptable in most circumstances.

Considering the reliability-centered maintenance of the DFIG system, the assignment profile of an element is categorized as a long, medium, or short-term phase. Thus, the algorithm accuracy is also not acute. The training dataset and interpretability of the optimization procedure are not essential.

For the RUL forecast of power converter switching strategies, the prerequisite of the algorithm speed is modest since the device degradation is gentle and the more extended period of judgment making is usual. The degradation modeling for the RUL forecast can be arranged in offline mode and professionally tuned into an online approach, and the computational work in this application is reasonable. Since the model accurateness is highly reliant on the dataset, the dataset prerequisite, e.g., dataset superiority, dataset sizing, and label stability, is critical. Likewise, the interpretability of the RUL forecast outcomes with ambiguity is also acute.

Table 6. Evaluation of AI algorithms in each phase of the lifecycle of power electronic systems.

Conditions of AI in Power Electronic Converter in Model Applications	Dataset Requirement	Accuracy	Sensitivity	Speed	Interpretation	Computing Effect
RCM and health management	High requirement	High accuracy	Moderately sensitive	Sensible speed	High	Moderate effect

evaluates the AI algorithms in each phase of the lifecycle of power electronic systems.

This review has recognized that AI possesses immense possibilities in power converter reliability assessment systems. Abundant opportunities and disputes are still to be discovered, as discussed in the following section.

6.1. Rationalizations of AI Applied to Power Converter Systems

Though there have been several revisions on AI for power electronic converter systems in the studied narrative since the 1990s, the hands-on presentations in industries are still partial, which is a high-pitched divergence related to the requested AI capabilities. It is necessary for profound studies to be conducted into tasks where AI can fundamentally outclass conservative methodologies. The rationalizations of AI-based explanations should be openly recognized by those associated with conventional methodologies from engineering standpoints.

6.2. AI Performances through Lifecycle Stages

The application of AI at every lifecycle stage of designing and maintenance will enable flexible, practical interfaces. This quality is favorable to complete performance optimization and process overview. Consequently, additional consideration should be made concerning the linked collaborations driven by AI systems. It allows the system to handle data flow between electrical and mechanical aspects. For example, maturity data acquired by the AI-based system’s variable identification can be adapted to the AI-based control for reliability enhancement.

6.3. Informative Association

Safety, reliability, and toughness are the foundations of self-sufficient artificial intelligence know-how. Building a safe and robust advanced AI-based power converter reliability system is so complex that no one producer can single-handedly acquire all of the essential technology. As a substitute, there is a broad ecosystem of partners, each working on a different facet of the required expertise. If these statistics bases and designs are concurrently misused, likely partialities can be eased to improve the system’s sturdiness. Multilayer information combinations can be achieved at the dataset level [195,200], governance-level, characteristic-level [125], and their groupings to accomplish the visions of every detail base. Likewise, the firmly entrenched differential calculations of the power electronic converter system are combined with AI as an amalgamated result for condition checking. Consequently, the advances in the model-impelled edge and the dataset-impelled edge are increased for improved precision and sturdiness.

6.4. Database Confidentiality

Based on the existing progress of AI knowledge, it is predicted that AI methodologies used in grid-incorporated REs will overcome the problem of deficient dataset model accumulation. The heap dataset base assessment in the incorporated power grid is still in the early phase, and the dataset base growth in different application circumstances is not the same. Database set models that can overcome constraints of various AI methodologies applications are not sufficient, so understanding AI applications driven by a minor database set samples is an approach that requires continuous planning.

The privacy of AI-based datasets involves an alliance between the General Data Protection Regulation (GDPR) [201] and artificial intelligence (AI). The GDPR measures the innovation in AI methodology and directs the application of AI in terms of personal data use. It partners with individuals and society in cases of dispute, estimates how risks can be mitigated, and empowers opportunities through technical knowledge and governing regulations. It analyzes how AI is controlled in the GDPR and observes the extent to which AI follows the GDPR’s theoretical structure, particularly in terms of purpose restriction and dataset minimization. It evaluates dataset subjects’ privileges, such as rights involving access, removal, transferability, and objectives. It sets competently trending dataset confidentiality guidelines on the application of AI-based outcomes. Along with the disapproving standards, the training of standard AI methods is encouraging since compacted dataset collection may not be possible right away. Thus, for power converter applications, it is prudent to develop and expand a collective learning system for AI techniques deprived of mutually combined datasets from various localities, e.g., merged knowledge [196]. The study undertakes this by recognizing that AI can be arranged in a way that is following the GDPR, but also in such a way that the GDPR does not provide adequate guidelines for power converter controllers and that its formulations need to be prolonged and emblematized.

7. Conclusions

Centered on the research and analysis of AI procedures in powered converter reliability systems for power grid-incorporated DFIG-based wind-driven turbines and renewable energies, it is established that the current machine learning computational processes such as fuzzy control logic (FNN and ANFIS), metaheuristic-type methods (PSO and GA), and machine/deep learning (RNN, SVM, CNN, and AE) play a significant role in helping researchers to evaluate research leanings and the hotspots.

This review determines that machine learning computation progressions could aid in exploring the existing work by evaluating a sizeable number of papers, interpreting trends, presentation tendency, the practice fraction, conditions, and structures of AI in the life-span stage of the power converter applicable in the smart grid-integrated DFIG reliability system. Based on the functionality viewpoint, AI-related applications are fundamentally handled with improvisation, classification, regression, and dataset arrangement investigation. By learning the above discoveries, unconventional control approaches will be advanced to improve wind turbines’ functional performance and power system integrity and reliability, considering an elementary reliability strategy for standard and uninterrupted operation for smart grid abnormalities and grid code achievement.