A review on deep learning based condition monitoring and fault diagnosis of rotating machinery

Electric motors

Motors utilize more than half of all electrical power produced worldwide, accounting for nearly 60% of all industrialized electricity. Various types of AC and DC motors have been utilised in industry. Motors have been evolved as modern world’s backbone, as they are essential to the manufacturing, transportation and other industries. Conveyors, compressors, blowers, generators, fans, machine tools, cranes, and electric vehicles are all examples of common motor applications. ²⁰

Gear and gear box

Gear is a major component in automatic transmission mechanisms in wind turbines, aircraft, automobiles, and a variety of other mechanical machines. The failure of the gear results in unnecessary downtime, production losses, and human casualties. Consequently, diagnosing gear defects early is crucial. ²¹

Rotor-bearing system

A rotor bearing system might fail either due to rotor or bearing defects. Bearing damage is the most predominant type of damage in a rotor-bearing system or rotating machinery. Depending on the type and size of the machine, the percentage of bearing defects in the total number of machine faults ranges from 40 to 90%. Consequently, diagnostics of the bearing is crucial in rotating machines. ²² In addition to bearing damage, rotor fault is another major fault in any rotating machines.

Wind turbine

Large-scale power systems such as wind turbines are subject to changeable weather and hostile conditions. As a result, wind turbines are prone to frequent breakdowns and malfunctions, resulting in large expenses. The development of condition monitoring systems (CMS), which could monitor wind turbine parameters in real time and detect faults in the near future, has aroused the interest of the wind energy industry. ²⁵

Pump

Pumps are used to move fluids and are controlled by an electric motor. They are used to pump water from wells, in filter aquariums and ponds, and provide ventilation. They are also utilised in the power industry to pump oil and typical gases, as well as in cooling towers and other industries for water cooling and fuel injection. Pumps come in a variety of sizes, ranging from microscopic pumps (used in medical applications) to massive industrial pumps. Their failure results in the shutdown of manufacturing processes and lines, thus monitoring and diagnosing them is critical. ²⁸

Air compressor

Compressors are similar to pumps in that they increase the force applied on a fluid while also allowing it to be transported through a pipeline. Compressors are vital machinery in a variety of industry applicaton, therefore diagnosing and managing them is essential. ³¹

DL based motor fault diagnosis

Deep learning is emerging in the field of motor condition monitoring and fault diagnosis. Different types of DL, such as DNN, CNN, RNN, encoder, and others, have been developed and successfully used to detect various electrical and mechanical faults in motors.

In a work, Sun et al. 2016 ⁴³ investigated a sparse auto-encoder with deep neural network (SAE-DNN) system for categorizing induction motor faults. This method can be trained features directly from unprocessed datasets that distinguish various induction motor operational conditions. The results revealed that learning and a thorough investigation validate the SAE-DNN’s efficiency in analyzing induction motor faults. In other study, Xiao et al. 2016 ⁴⁴ presented adaptive function mining with stacked denoising auto-encoders (SDAE) designed for asynchronous motor fault diagnosis. The SDAE approach improved the asynchronous motor fault diagnosis evaluation with stacked auto-encoder (SAE), which solved the problem with the conventional MCSA technique.

Shao et al. 2017 ⁴⁵ presented a DL approach based on DBN for fault analysis of induction motors, where frequency domain data of the measured dataset is used as input data. In this paper, a DBN-based approach is used to model and analyse multiple layers of representation on a high-dimensional vibration dataset. In other study by Sun et al. 2017 ⁴⁶ investigated convolutional discriminative feature learning (CDFL) for induction motor fault diagnosis. The rapid convolutional pooling structural architecture was used in this study to extract discriminative and invariant properties from an unprocessed vibration dataset. The weights of the small filters were learned using a back-propagation neural network (BPNN). Then support vector machine (SVM) is used to effectively categorise the learned features.

Xia et al. 2017 ⁴⁷ conducted an intellectual fault study of motor bearing using unsupervised characteristic learning with stacked denoising autoencoders (SDA). The proposed approach, which is derived from SDA and DNN, can learn representative characteristics from an enormous amount of unlabeled dataset without the need for human interference. Wang et al. 2017 ⁴⁸ presented the motor fault diagnosis based on short-time fourier transform (STFT) and CNN. Proposed deep learning methodology effectively categorizes health states with increased diagnostic accuracy and can extract prominent fault characteristics in a consistent manner. Ghosh et al. 2017 ⁴⁹ used DNN to detect a fault caused by a short circuit in the stator winding or an unstable inverter deliver by predicting magnetic flux. The DNN prognostic harmonics compensated control drive is first simulated and used for dataset analysis and finite element investigation.

Zhang et al. 2017 ⁵⁰ used a 1-D stacked dilated convolutional neural network (1D-SDCNN) to create an intelligent fault study for rolling bearings of induction motor under changeable operational condition. Stacked dilated convolutional allows the device to contain large accessible areas with a limited number of coats, allowing for improved overall features. The results showed that the proposed approach is more practical and powerful for analyzing induction motor faults even when different load conditions are used for training and testing. Lee et al. 2017 ⁵¹ presented the DNN and GAN to diagnose induction motor faults in industrial maintenance. The effect of oversampling on the fault diagnosis of a motor was demonstrated in this study. By resolving the complexity of dataset instability with GAN and conducting intellectual fault diagnosis and analysis with DNN, the results demonstrated the likelihood of data-driven maintenance solution. However, other problem of data imbalance like under-sampling and ensemble learning need to be consider for the study.

Yao et al. 2018 ⁵² presented a fault diagnosis for a self-sensing motor drive structure using CNN. The motor drive structure was used as a multivariable sensor at the drive-side of a mechanical device in this paper, which measures torque, speed, location, and power in addition to current. Therefore there is no need of extra sensors and data acquisition, in the proposed work, the raw data of healthy and faulty conditions from self-sensing motor drive are first converted into color and polar images for different operating conditions. The result demonstrated that CNN can be used for accurate diagnosis. Cipollini et al. 2019 ⁵³ worked on unsupervised deep learning for induction motor bearing’s fault diagnosis. Results showed that the proposed technique has the potential to derive a dense and important representation of the bearing state from the stator current signal. Shao et al. 2019 ⁵⁴ studied the deep convolutional neural network (DCNN)-with multi-signal (vibration and current) for induction motor fault diagnosis. deep convolutional neural network is being developed as a building block for multi-signal representation, demonstrating the ability to study discriminative characteristics from time-frequency distribution (TFD) images without human interference.

Afrasiabi and Afrasiabi. 2019 ⁵⁵ introduced an accelerated deep learning approach to analyse real-time bearing defects in induction motors. In this work, they developed an accelerated CNN (ACNN) by avoiding less significant connections and sharing the weights in order to compress and speed up the conventional CNN. Results showed that the developed ACNN perform better and faster than the conventional CNN, SVM and ANN. Zou et al. 2020 ⁵⁶ used deep learning for bearing fault diagnosis of the traction motor in high speed trains. The discrete wavelet transforms (DWT) and enhanced DBN were used for the diagnosis. The DWT was first used to extract 2D time–frequency map from the healthy and faulty dataset. After that, these maps were used to train the enhanced DBN for correlating the fault features and fault types. In comparison to other conventional approaches like BPNN and SVM, the proposed approach achieved the better accuracy.

Kumar and Hati. 2020 ⁵⁷ introduced an adaptive gradient optimizer with deep convolutional neural network (ADG-dCNN) for successful bearing and rotor defects detection in squirrel cage induction motors. In this work, adaptive gradient optimizer was successfully used for optimizing the parameter for efficient training and deep CNN was successfully used for efficient feature learning. Rodriguez et al., 2020, ⁵⁸ performed a diagnostic for damaged BRB of various severity levels in induction motors based on a CNN and MCSA. The STFT was initially used to observe left sideband frequency components associated with BRB. Then CNN was used to achieve automatic image categorization of different BRB fault conditions.

Skowron et al. 2020 ⁵⁹ investigated the CNN for stator winding faults of an inverter-fed induction motor using raw transient stator current signals. In this work, they used only 2000 samples of transient stator current for the diagnosis thus avoiding other signal processing techniques like FFT, WPT and HHT. The proposed technique had a high degree of precision in determining the location of the breakdown and the severity levels, as well as the ability to discern even individual shorted turns. Grezmak et al. 2020 ⁶⁰ introduced a multi-stream convolutional neural network (MS-CNN) for induction motor in variable frequency drive. They showed that the proposed technique could learn patterns in the inputs from FFTs of combined vibration and current signals obtained at different line frquency. Results showed that the MS-CNN identifies the harmonics of the motor’s line and rotational frequencies for differnet fault conditions.

Kumar et al. 2020 ⁶¹ introduced a novel transfer learning (TL) and a DCNN to detect bearing fault and BRB faults (individually and jointly) in squirrel cage induction motors. TL extracts critical information from the already trained model to improve the learning of a new model therefore it is useful for less labelled data cases. Guo et al. 2020, ⁶² presented the use of DNN to predict the winding temperature of a permanent magnet synchronous motor (PMSM). In this work, temperature (ambient, coolant surface and tooth temperature), voltage, current, speed, and torque were considered for the input and winding temperature was considered for the output. The method has the potential to play a significant role in PMSM high temperature identification to avoid any upcoming faults, as well as provide industrial support for high temperature caution and PMSM protection.

Liang et al. 2020 ⁶³ introduced a novel gate-structure dilated convolutional capsule network (GDCCN) with a deep learning methodology for motor fault diagnosis. It is combined by the long short-term memory (LSTM) system’s enter gateway arrangement, extended convolution, and capsule neural network. Result showed that the GDCCN provides better performance than conventional DL methods even in noisy enviroments and different workloads. Espitia et al. 2020 ⁶⁴ & Espitia and Soto. 2020 ⁶⁵ used SAE based deep learning methodology for effective fault diagnosis in electromechanical systems. They considered bearing, demagnetized, and eccentricity-related induction motor faults. GA was used to optimize the hyperparameter of a SAE and LDA was used to maximize distance between different data sets.

Singh et al. 2021 ⁶⁶ performed the fault diagnosis of electrical and mechanical faults in induction motor based on DNN using vibration and current time domain signals. In this paper, critical features such as standard deviation, skewness and kurtosis were extracted to characterize motor faults. They performed the optimization of batch size and number of epochs by trial-and-error method and suggested that increasing both hyper parameters is not a better proposal. Zou et al. 2021 ⁶⁷ used deep learning techniques for fault diagnosis of bearing of traction motor in high-speed trains. In this work, they obtained 2 –D time-frequency map from 1-D vibration signals based on DWT. Then they used improved DBN for performing the bearing fault diagnosis and showed that it is better than SVN and BPNN. Kumar and Hati 2021 ⁶⁸ performed the multiple fault diagnosis of induction motor using transfer learning-based CNN (TL-CNN). In this work, they converted current signals acquired from hall sensors into images which are further used as input to TL-CNN. They showed that the transfer learning reduces the computation time required for building effective CNN model with high depth.

A summary of the research papers for motor fault diagnosis based on DL techniques is added in the Table 1.

OPEN IN VIEWER

Table 1.

Summary of Deep Learning (DL) based motor fault diagnosis.

Ref. No.	Author, year	Machine faults	Signals	Technique	Remarks
44	Sun et al. 2016	Rotor faults, bearing faults and stator winding faults	Vibration and current signals	SAE-DNN	SAE-DNN has high efficiency for induction motor fault diagnosis
45	Xiao et al. 2016	Stator-winding faults, rotor-broken faults and bearing faults	Current signals	SDAE-DNN	The feature extracting method with SDAE is efficient in asynchronous motor fault diagnosis
46	Shao et al. 2017	Induction motor faults	Vibration signals	DBN	DBN method with frequency domain signal reduces training error and improves classification accurateness
47	Sun et al. 2017	Rotor faults, bearing faults and stator winding faults	Vibration and current signals	CDFL	CDFL method shows ability of mining the discriminative, robust and more efficient features from the fault signals for induction motor fault diagnosis
48	Xia et al. 2017	Rotating machinery faults	Vibration and current signals	SDA-DNN	The method demonstrates the efficiency in RUL prediction and achieves superior prediction accurateness than ANN
49	Wang et al. 2017	Rotor faults, bearing faults and stator winding faults	Vibration and current signals	STFT-CNN	STFT-CNN is efficiently categorize the health states with elevated diagnosis accurateness and capable to compatibly extract prominent fault characteristics
50	Ghosh et al. 2017	Stator winding faults	Current signals	DNN	DNN technique established like a feed forward slab for predicting flux with superior accurateness in the fault diagnose
51	Zhang et al. 2017	Bearing faults	Vibration signals	1D-SDCNN	1D-SDCNN is extra realistic and strong in favor of motor fault diagnosis along with it had high-quality presentation in the case of insufficient dataset
52	Lee et al. 2017	Broken rotor bar fault and bearing fault	Current signals	DNN with GAN	The DNN perform well by resolving the dataset unstable difficulty with GAN
53	Yao et al. 2018	Permanent magnet synchronous motors (PMSMs) faults	Vibration and current signals	CNN	Multisensory dataset fusion technique with CNN perform wellare validated in imitation and experimentation
54	Cipollini et al. 2019	Bearing faults	Current signals	DNN	The present method is promising to extract features from the stator current signal a dense and significant representation of the bearings state
55	Shao et al. 2019	Induction motor faults	Vibration and current signals	DCNN	Present appraoch attained the finest execution
56	Afrasiabi and Afrasiabi, 2019	Bearing faults of induction motor	Vibration signals	ACNN	ACNN technique is a quick and robust than CNN, SVM and ANN
57	Zou et al. 2020	Bearing faults of traction motor	Vibration signals	DWT-DBN	DWT-DBN is a superior than SVM and BPNN
58	Kumar and Hati, 2020	Bearing faults and rotor faults of induction motor	Vibration signals	ADG-dCNN	ADG-dCNN perform efficient training and feature learning
59	Rodriguez et al. 2020	Broken rotor bar faults	Current signals	CNN	CNN with STFT successfully performs BRB fault diagnosis
60	Skowron et al. 2020	Stator winding faults	Current signals	CNN	The CNN has high precision in identify the localization of the breakdown and the fault stages
61	Grezmak et al. 2020	Rotor faults and bearing faults	Vibration and current signals	MS-CNN	The MS-CNN can learn patterns in the inputs coupled with the harmonics of the motor’s line and rotational frequencies
62	Kumar et al. 2020	Bearing faults and BRB faults	Current signals	DCNN	DCNN with TL method shows better performance
63	Guo et al. 2020	Stator winding faults	Current signals	DNN	The DNN system can detect high temperature
64	Liang et al. 2020	Short circuit faults, air-gap eccentricity faults, BRB, bearing cage broken fault and bearing abrasion faults	Vibration signals	DGCCN	DGCNN perform better than conventional DL method even in noisy environment
65, 66	Espitia et al. 2020 & Espitia and Soto, 2020	Bearing, demagnetized, and eccentricity fault induction motor	Vibration signals	DNN based on SAE	The DNN-SAE perform well with small no. of samples
67	Singh et al. 2021	Mechanical and electrical faults of induction motor	Vibration and current signals	DNN	DNN with critical features like standard deviation, skesness and kurtosis of time domain vibration and current signals perform well
68	Zoa et al., 2021	Bearing fault of traction motor	Vibration	DBN	DBN with DWT features of vibration perform well and better than SVM and BPNN
69	Kumar and hati 2021	Broken rotor and bearing fault of induction motor	Current	CNN	CNN with 2D image of current signals perform well

DL based rotor-bearing system fault diagnosis

Deep learning have been also gaining interest in the diagnosis of rotor-bearing faults. Various researchers have looked into various types of DL to detect rotor and bearing faults. The work on DL fault diagnosis of rotor-bearing models reported in the last 5 years is discussed here.

Lu et al. (2015) ⁶⁹ investigated a novel feature extraction approach for rolling bearing fault diagnosis using DNN in a paper. They demonstrated that DNN is currently the most powerful machine learning tool, and that its benefit is suitable for resolving high nonlinearity issues in bearing signal features. The proposed method achieves features that accurately and clearly depict bearing signal details, which are very useful for successful fault diagnosis. Ahmed et al. (2016) ⁷⁰ employed DNN for bearing fault identification based on sparse autoencoder. This allows to extract fault features in unsupervised manner. They showed that the system structural design and regularization parameters, as well as DNN parameters, have a major impact on bearing fault identification performance. Results showed that identification performance is higher in the re-training phase as compared to the pre-training phase of DNN.

In a work, Xia et al. 2017 ⁷¹ employed DNN for bearing fault classification based on stacked denoising autoencoder. Critical fault features were extracted from unlabeled data using denosing autoencoder, and then DNN was trained with few labelled data. Results showed that the present approach can extract critical features based on large unlabeled data and correctly classify bearing fault based on few labeled data of a new condition. Xia et al. 2017 ⁴⁷ used a hierarchical DNN to calculate the remaining useful life (RUL) of bearings in rotary machines based on vibration and temperature data. The RUL of bearing was evaluated using a hierarchical DNN with RUL predictor (DNNRULP) technique. The DNNRULP technique consists of three main components: a DNN with Health Stage Classifier (DNNHSC), an ANN with RUL predictor (ANNRULP), and a DNN with an even operative. Result showed that the proposed approach demonstrates efficiency in RUL prediction and achieves higher prediction accuracy than a single ANN device.

Zhang et al. 2017 ⁷² presented a novel method called deep CNN with wide First-layer kernels (WDCNN) learning method for bearing fault diagnosis. The WDCNN was divided into two parts: a large first-layer convolutional kernel and a deep network formed by small convolutional coats. In addition to the wide kernels in the primary convolutional coat was used for extracting characteristic and restraining elevated frequency tone, the proposed method used unprocessed vibration signals as input. Chen et al. 2017 ⁷³ presented a deep sparse auto-encoder method with noise (DSAE-N) for rolling bearing fault severity detection. The DSAE-N was built by loading many SAEs into a classifier coat in order to achieve automated attribute extraction and fault diagnosis. The DSAE-N was trained with noise samples to avoid the overfitting issue due to limited training data. Result showed that the DSAE-N performs well in comparison to manual feature extraction technique, SAE and DSAE without noise samples.

Sohaib et al. 2017 ⁷⁴ employed hybrid feature pool in deep-learning-based bearing fault diagnosis. In this study, they combined features from different domains i.e. time, frequency and time-frequency domain. Finally they combined the sparse stacked auto encoder (SSAE) with DNNs and showed that the present method perform very well to analyze a variety of bearing faults with different severity levels. Li et al. 2017 ⁷⁵ investigated a combined deep CNN architecture and enhanced Dempster–Shafer theory (IDSCNN) for bearing fault diagnosis. Result showed that the present method successfully perform the diagnosis by effective fusing signals from multiple vibration sensors.

Chen et al. 2017 ³⁷ worked on DNN based rolling bearing fault diagnosis. In this work, performance of DBN, DBM, and stacked SAE were evaluated. Result showed that all three methods perform very well with the combination of time-frequency and time-frequency domain features in comparison with raw data. In addition, SAE perform slightly better than DBN and DBM, and DBN need more training time. Chen and Li (2017) ⁷⁶ investigated the multi-sensor characteristic information for bearing fault diagnosis based on SAE and a DBN. Time and frequency characteristics obtained from various accelerometers signals were fused in a single stream with SAE. Then, DBN was trained using these combined characteristics and found effective identification of bearing faults. In other work, Chen and Li, 2017 ⁷⁷ investigated DNN with a denoising auto-encoder (DAE) in order to perform fault diagnosis on rotary machines. In this work, they considered critical fault of rotor-bearing system like unbalance, misalginment, rubbing and pedestal loosness faults. This paper used a DNN with a denoising auto-encoder (DAE). After obtaining a DNN via unsupervised DAE training, the dropout approach is used to change the network parameters, reducing the issue of over-fitting. The results showed that the DNN can efficiently learn the frequency domain features in unsupervised mode and adaptively extract the characteristics of different faults in the rotor-bearing system.

Ren et al. 2017 ⁷⁸ used a deep learning method to estimate the RUL of several bearings. From the vibration signals of a rolling bearing, the technique used elevated value of degression samples. Three time domain features and one proposed six-dimensional frequency domain feature (Frequency Spectrum Partition Summation (FSPS)) among the features of the collected bearing vibration signal used here. Results showed the promising performance of deep learning approach for RUL prediction of multi-bearing. Xin et al. 2018 ⁷⁹ considered a combination of time-frequency features obtained from STFT and deep CNN for rotary machines fault diagnosis. Here, sparse auto encode, convoluting, pooling, and softmax classifier were used to build a new DNN called the DCNN scheme. The results showed that the present methodology is capable of recognizing various faults in gear and bearings from time-frequency images efficiently.

Eren et al.2018 ⁸⁰ used a dense adaptive 1D CNN classifier to investigate bearing faults. The result showed that 1D CNNs can extract extremely discriminative characteristics from the unprocessed input sensor dataset and an effective fault diagnostics can be achieved by using a simple-dense CNN arrangement. Yang et al. 2018 ⁸¹ worked on deep learning based fault diagnosis of bearing. In this work, they divided the whole spectrum into several sections and used for the input of DNN. In addition, they assigned labels in data sets randomly to avoid pre-affixing a specific class label to a sample. Xia et al. 2018 ⁸² presented a two-stage outlook for calculating bearings’ RUL based on deep learning. The degradation method was divided into various physical conditions in the first step. Then a stacked denoising auto encoder was used to initialize a DNN classifier. In the second stage, a shallow neural network system was built and learned in each physical condition of fault in order to perform the middle RUL evaluation. The experiments showed that the industrial technique could achieve adequate RUL calculation with a limited dataset.

Wang et al. 2018 ⁸³ and Wang et al. 2018, ⁸⁴ employed Gaussian radial basis kernel function based autoencoder (KAEs) to enhance the feature extraction technique especially for non-linear component in bearing fault diagnosis of aircraft engine. A DNN was built with one kernel based autoencoder and multiple KAEs and result showed that the present methodology has higher accuracy because of its better feature clustering effect in feature extractions. Nguyen et al. 2018 ⁸⁵ used DNN to identify different bearing fault severities. In this work, they extracted characteristic frequency from envelop power spectrum of acoustic emission signals. Then, Adam optimization-based backpropagation algorithm with Xavier initialization method was used in DNN and results showed that better performance as compared to state of the art approaches. Chunfeng et al. (2018) ⁸⁶ introduced heterogeneous transfer learning (HTL) based on stack sparse auto-encoder (SSAE). In order to solve the problem of small target domain data, they introduced a concept of distance to the center of the source and target domain to find the similarity of distribution using heterogeneous characteristics representation in HTL. Result showed that the present methodology has better performance than the conventional machine learning approaches especially for limited labeled data.

Saufi et al. 2018 ⁸⁷ performed the bearing fault diagnosis based on wavelet transform based images and flexible SSAE. A differential evolution and a resilient back propagation technique were used to optimize SSAE hyper constraints to prevent overfitting problem. Result showed that, even with limited datasets, the present SSAE based method is capable of accurately diagnosing bearing fault conditions and has better performance than DNN, CNN, SVM and KNN. Appana et al. (2018) ⁸⁸ used the envelope spectrum (ES) of acoustic signals and CNN to perform a robust fault diagnosis of bearings under real life operating condition of changeable rotating speed. They performed the diagnosis by training CNN with signal obtained at one speed and testing on the reaming signal obtained at other speeds.

Sadoughi and Hu (2018) ⁸⁹ introduced a physics based CNN (PCNN) for rolling bearing fault diagnosis. The traditional CNN was customized to include useful details from physical experience about bearings and their fault characteristics. Novel physics bases kernels were created based on the bearing fault frequencies and shaft speeds and used in convolutional layers of the PCNN. Espitia et al. 2019 ⁹⁰ used SAE based feature reduction analysis to track the condition of electromechanical systems and compared the performance with other technique like PCA and LDA. In this work, first, statistical time-domain features analysis using the vibration signal is demonstrated. Then SAE, PCA and LDA methods employed for feature reduction. Finally, they showed that the DNN based approach using SAE represents better discriminative capabilities than PCA and LDA in bearing fault diagnosis.

Gao et al. 2019 ⁹¹ combined 1D convolutional neural system and generative adversarial networks (ASM1D-GAN) to develop an intelligent bearing fault diagnosis technique. In this work, in order to overcome the problem of small data of natural occurring fault, an ASM1D-GAN approach was developed to generate effective novel fault features which are consistent with natural fault features obtained from the raw data. Wang et al. 2019 ⁹² presented bearing fault diagnosis using a multiscale learning neural network based on CNN. Two channel inputs were considered in this study i.e. one-dimensional CNN and a two-dimensional CNN to extract a variety of vibration features. 1D and 2D CNN learn informations from adjacent and nonadjacent intervals, respectively. Result showed that diagnosis with 1D CNN perform better than with 2D CNN. The performance of the diagnosis was significantly increase when both 1D and 2D CNN were used simultaneously.

Zhou et al. 2019 ⁹³ used CNN to study real-time fault diagnosis in rolling bearings based on 2D image input. DNN with stacked AutoEncoder do not efficiently get the spatial neighborhood features (SNFs) from 1D vibration signal because this 1D vibration signals show the amplitude only. Therefore, in this study, CNN was used with 2-D image of the vibration waveform as input. Results showed that present approach fully utilized the SNFs of the vibration signals in achieving the accurate bearing diagnosis. Enshaei and Naderkhani, 2019 ⁹⁴ presented role of deep learning in bearings of induction machine. This study focused on a deep bi-directional long short-term memory (BiD-LSTM) method fed by unprocessed signals in order to overcome the shortcomings of traditional machine learning (ML) techniques. The proposed method’s results validated the system’s efficiency with high accuracy and low latency.

Yan et al. 2019 ⁹⁵ used a multi-domain indicator and an optimised stacked denoising auto encoder (MIOSDAE) to recognize the current state of rolling bearings. The grasshopper optimization algorithm (GOA) was used to optimise the significant parameters of the stacked denoising auto encoder (SDAE) model, which could improve error categorization precision. Result showed that the present technique is valuable for identifying bearing current state. Chen et al. 2019 ⁹⁶ presented a new deep learning methodology called RNN scheme with encoder-decoder arrangement for estimating the remaining useful life (RUL). Here, 5 band-pass energy magnitude of frequency spectrum was used as input in RNN. Finally health indicator (HI) is obtained from this encoder-decoder arrangement using long historic data to extract critical degradation facts to calculate the RUL. Ma et al. 2019 ⁹⁷ investigated a lightweight deep CNN for bearing fault diagnosis of rotary machine. In this paper, the WPT was used to extract excellent information from the time-frequency domain, while the secondary layer constructs a relatively lightweight CNN. In terms of noise suppression and advanced training potential, the experiment shows that the present algorithm is superior to other algorithms. Zhou et al. 2019 ⁹⁸ proposed a deep learning fault diagnosis technique for unstable datasets of bearing using a global optimization GAN (GOGAN). It is noted that GOGAN can generate additional qualifying dataset required for the ultimate purpose of fault diagnosis.

Zhao et al. 2020 ⁹⁹ used FFT to perform fault diagnosis of a motor-rotor system using multi-manifold deep extreme learning machine (MDELM). In this work, they performed multichannel vibration data fusion, and random and pure unsupervised feature mapping for identifying the faults by MDELM. Results showed that the present methodology has superior performance in detecting critical faults like unbalance, misalginment, loosness and rubbing in the motor-rotor system. Chen et al. 2020 ¹⁰⁰ used a deep DRN to recognise the severity as well as the location of the bearing faults at the same time. They used frequency and energy features of vibration obtained in rasonance zone. To visualize inside of the DNN, they devloped a mathematical model by maximizing activation value. The result showed that the devloped methodology has high potential to detect severity as well as location of the bearing fault even under noisy environment.

Zhang et al. 2020 ¹⁰¹ summarised application of DL in bearing fault diagnosis. They discussed the advantage of DL over ML such as the capability of automatic feature extraction, unsupervised learning, to handle imbalance data and noisy data, and classifier’s performance. To handle small data or imbalanced data problem in fault diagnosis, techniques like data random sampling, GAN and etc. have been successfully used in order to create additional faulty data in the training. Tang et al. 2020 ¹⁰² developed a deep CNN approach based on knowledge fusion for analysing bearing defects in a variety of operational scenarios. Via the use of multi-sensors and narrow-band decomposition methods, an information fusion technique was implemented in this learning to increase the characteristic demonstration capability with the transferability of analysis systems. Han et al. 2020 ¹⁰³ used CNN to identify bearing faults based on different input modes like time, STFT, grayscale, CWT and time domain color feature (TDCF) diagram. In this work, they added red colour to TDCF in one input mode of CNN and found enhanced fault feature characteristic of the vibration signals. Results showed that the CNN with TDCF input mode maintained high performance as compared with time, STFT, gray scale and wavelet diagram even under the high noise conditions.

Xin et al. 2020 ¹⁰⁴ developed an automated fault diagnosis for bearings and gears using vibration signal analysis and multi-object deep convolutional neural networks (MO-DCNN). The time series, frequency dataset, and time–frequency dataset was structured like the multi-object to mine characteristics. The present method could boost overall feature characteristics and diagnostics performance while avoiding the drawbacks of using a single time or frequency field dataset. Zeng et al. 2020 ¹⁰⁵ investigated a bearing fault study based on denoising autoencoders for small number of labelled trial cases. In this work, they combined a CNN with an auto encoder (AE) and skip connect and finally trained the model with pre-training and fine tuning. Results showed that the present method reduces training time, improves the capability of the traditional fully connected auto encoder to extract features and diagnosis performance.

Xiong et al. 2020 ¹⁰⁶ combined WPT and CNN and developed a new continuous fault diagnosis approach for rolling bearings. Time-frequency analysis such as WPT was employed to find the richer information from the 2D coefficient matirix obtained from 1D raw signals. The CNN with WPT coefficient showed better and robust performance as compared to CNN with 1D raw signals. Li et al. 2020 ¹⁰⁷ developed a rolling bearing fault diagnosis with WPT and CNN. In this work, WPT was used in the first step to extract 1-D time-frequency coefficients from vibration signals, which were then converted into 2-D grey images using a careful data-to-image conversion technique. In the next step, a CNN system with three convolutional coats was used to find characteristics features from grey pictures and the present method yielded good results.

Zhao and Jia, 2020 ³⁶ presented a novel unsupervised deep learning netwrok (UDLN) method for rotary machine fault diagnosis. In this work, they used vibration spectrum as input to the UDLN where first layer is constructed with sparse filtering for feature extractor and second layer with weighted euclidean affinity propagation for the clusturing extractor. Zhang et al. 2020 ¹⁰⁸ showed a new unsupervised field adaptation with DNN for bearing and gear fault diagnosis. The key contributions were such as implementing maximum mean discrepancy (MMD) in DNN to reduce the variation of the source and target data distribution, and applying manifold regularizations, while simultaneously improving the delegate knowledge of the unprocessed dataset.

Zhao et al. 2020 ¹⁰⁹ used a semi-supervised deep sparse auto-encoder (SSDSAE) for rotary machine fault diagnosis using few labeled data. In this work, they used vibration spectrum signals to extract local structural informations (by minimizing interclass compactness) and non-local structural informations (by maximizing interclass separability). In addition, weighted cross entropy methods was adapted in this work for improving generalization accuracy of the SSDSAE model. Li et al. 2020 ¹¹⁰ studied the fault diagnosis of bearing fault of rotary machine using deep learning and dataset augmentation. In this work, five dataset augmentation methods were employed such as extra Gaussian noise, masking noise, signal translation, amplitude shifting, and time stretching. In addition, they showed they signal translation and time stretching methods are promising as compared to other data augmentation methods.

Yan et al. (2021) ¹¹¹ performed fault diagnosis of rotor-bearing system based on a new method called deep regularized variational autoencoder (DRVAE). In this work, the hyper parameter DRVAE was optimally selected based on bird swarm algorithm (BSA). They used frquency domain vibration signals acquired from multiple sensors as input to DRVAE. Che et al. 2021 ¹¹² performed the fault diagnosis of rolling element bearing based on hybrid multimodel fusion of DL models like CNN and DBN. In this work, vibration signals are converted into grayscale images and time series data for processing into CNN and DBN, respectively. Then three CNN and three DBN model were combined to obtain multi model fusion. Result showed that the prsent methdology based on multimodel fusion perform better than individual DL models. Xin et al. 2021 ¹¹³ developed Gaussian convolution DBN (GCDBN) for fault diagnosis of rotor-bearing system based on infrared thermal analysis. As vibration signals sometime diffcult to acquire and affetced by time varying speeds, infrared thermal images are utilized to discretize the rotor bearing faults to avoid the affect of varying speed. Results showed that GCDBN perform very well for fault diagnosis with thermal images.

A summary of the research papers for fault diagnosis rotor-bearing system based on DL techniques is added in the Table 2.

OPEN IN VIEWER

Table 2.

Summary of DL based rotor-bearing system fault diagnosis.

Ref. No.	Author, year	Machine faults	Signals	Technique	Remarks
70	Lu et al. 2015	Rolling bearing faults	Vibration signals	DNN	DNN attains features that can illustrate the bearing signal data accurately and clearly
71	Ahmed et al. 2016	Bearing faults	Vibration signals	SAE-DNN	Present method attains high levels of accurateness for diverse ranges of constraint values in the re-training stage of the DNN compared to the classification accuracies obtained in the pre-training stage
74	Zhang et al. 2017	Rolling element bearing faults	Vibration signals	WDCNN	WDCNN method showed superior classification accurateness and it is also extremely robust against load changes and noise
72	Xia et al. 2017	Bearing faults	Vibration signals	SDA-DNN	SDA-DNN is useful for few labelled data sets
75	Chen et al. 2017	Rolling bearing faults	Vibration signals	DSAE-N	DSAE-N is better than manual feature extraction technique, SAE and DSAE without noise samples
76	Sohaib et al. 2017	Bearing faults	Vibration signals	SSAE-DNN	The SSAE-based DNN successfully diagnose the bearing faults and severity of it
77	Li et al. 2017	Bearing faults	Vibration signals	IDSCNN	IDSCNN performs better even in changeable load situation through multi-sensors signals
78	Chen et al. 2017	Rolling bearing faults	Vibration signals	DBM, DBN and SAE	SAE is better than DBM and DBN. DBN need more training time
79	Chen and Li, 2017	Bearing faults	Vibration signals	SAE-DBN	SAE-DBN can efficiently learn from time and frequency domain signals
81	Ren et al. 2017	Bearing faults	Vibration signals	DNN	DNN can estimate the RUL of muli-bearing
82	Xin et al. 2018	Bearing faults and gear faults	Vibration signals	DCNN	DCNN efficiently identifies bearing fault with time-frequency images
83	Eren et al. 2018	Bearing faults	Vibration signals	CNN	1D CNNs is able to study extremely discriminative characteristics straight from the unprocessed input sensor dataset
84	Yang et al. 2018	Bearing faults	Vibration signals	DNN	DNN with FFT of sub signals perform well
85	Xia et al. 2018	Bearing faults	Vibration and current signals	DNN	DNN successfully estimates RUL of bearings for the diverse operational situation
86, 87	Wang et al. 2018 & Wang et al. 2018	Bearing faults	Vibration signals	KAE	The proposed method has a superior accurateness, better feature extraction capability with improved clustering results
88	Nguyen et al. 2018	Bearing faults	Vibration signals	DNN	The proposed DNN method yielded superior standard accurateness and enhanced performance compared to other methods
89	Chunfeng et al. 2018	Bearing faults	Current signals	SSAE	SSAE with HTL shows good performance even with small labelled data
90	Saufi et al. 2018	Bearing faults	Vibration signals	SSAE	The present technique is capable to precisely diagnosing bearing fault condition as well as provided a featureless methodology
91	Appana et al. 2018	Bearing faults	Acoustic signals	ES and CNN	ES of acoustic signals with CNN improve the diagnosis accuracy
92	Sadoughi and Hu 2018	Bearing faults	Vibration signals	PCNN	PCNN is capable to identify and restrict the faults with high accuracy
93	Espitia et al. 2019	Bearing fault, demagnetized faults and eccentricity faults	Vibration signals	DNN-SAE	The proposed method shows enhanced discriminative abilities than other linear techniques like PCA AND LDA
94	Gao et al. 2019	Bearing faults	Vibration signals	ASM1D-GAN	The proposed approach demonstrates efficient performance by producing significant novel fault features
95	Wang et al. 2019	Bearing faults	Vibration signals	1D-2D CNN	CNN performance improves with combined 1D and 2D signals
96	Zhou et al. 2019	Bearing faults	Vibration signals	CNN	CNN with 2D image of vibration performs well
97	Enshaei and Naderkhani 2019	Bearing faults	Vibration signals	BiD-LSTM	The proposed method outcomes authenticate the efficiency of the system with high accurateness and low latency
98	Yan et al. 2019	Bearing faults	Vibration signals	MIOSDAE	The technique is precious in favor of physical condition identification on steady pace
99	Chen et al. 2019	Bearing faults	Vibration signals	RNN	RNN is useful for estimating correct RUL
100	Ma et al. 2019	Bearing faults of rotating machinery	Vibration signals	WPT-CNN	The proposed method can efficiently enhance the recognition accurateness and it is comparatively lightweight
101	Zhou et al. 2019	Rotating machinery faults	Current signals	DNN based on GAN	GAN can produce extra eligible dataset intended for the effective DNN fault diagnosis with unbalanced data
102	Zhou et al. 2020	Unabalnce, misalginment, loosness and rubbing in the motor-rotor system	Multichannel vibration data	MDELM	MDELM with FFT effeicently diagnose a no. of faults in motor-rotor system
103	Chen et al. 2020	Rotating machinery bearing faults	Vibration signals	DRN	The proposed DRN method achieves good results in detecting fault severity and location
104	Zhang et al. 2020	Bearing faults	Vibration signals, current, acoustic	Deep learning	DL perform better than conventional ML
105	Tang et al. 2020	Induction bearing faults	Vibration signals	CNN	Multi sensor data fusion with CNN can boost the diagnosis performance
106	Han et al. 2020	Bearing faults	Vibration signals	CNN	CNN with TDCF technique has higher performance and acceptance to noise environment
107	Xin et al. 2020	Gear faults and bearing faults	Vibration signals	MO-DCNN	MO-DCNN with combined time, frequency and time-frequency data achieves good fault identification
108	Zeng et al. 2020	Bearing faults	Vibration signals	CNN-AE	CNN-AE speed up the training, increases the capability of the traditional fully-connected auto encoder to extract features
109	Xiong et al. 2020	Bearing faults	Vibration and current signals	WPT and CNN	CNN with WPT attains superior accurateness than with 1D signal
110	Li et al. 2020	Bearing faults	Vibration signals	WPT and CNN	CNN with WPT images yielded good results
111	Zhao and Jia 2020	Rotating machinery faults	Vibration signals	UDLN	UDLN is able to completely exploit the benefits of unsupervised learning along with enhance the accurateness of fault diagnosis with the ability of automatic analysis
112	Zhang et al. 2020	Gear faults and bearing faults	Vibration signals	DNN	DNN achieves good cross-domain fault recognition
113	Zhao et al. 2020	Bearing faults and rotor faults	Vibration signals	SSDSAE	SSDSAE offers the improvement of semi-supervised fault diagnosis knowledge of rotary machines
114	Li et al. 2018	Bearin fault of rotating machinery	Vibration signals	Deep learning using data augmentation	Signal translation and time stretching methods are promising as compared to other data augmentation methods
115	Yan et al. (2021)	Bearing faults of a rotor-bearing system	Vibration signals from multiple sesnors	Deep regularized variational autoencoder (DRVAE)	DRVAE perform better than other deep learning methods like DBN, CNN and SAE
116	Che et al. 2021	Bearing fault diagnosis	Vibration signals	Multimodel fusion of CNN and DBN	Present multimodel fusion perform better than individual DL models
117	Xin et al. 2021	Rotor-bearing system fault	Infrared thermal images	Gaussian convolution DBN	GCDBN outperform other methods and infrared thermal images gives a alternative to vibration signals in rotor-bearing fault diagnosis

DL based gear and gearbox fault diagnosis

Various researchers have explored deep learning techniques in Gear fault diagnosis over the last 5 to 6 years. This subsection provides an overview of various DL techniques that have been developed and used to detect various Gear faults.

Lu et al. 2016 ¹¹⁴ introduced a deep model based domain adaptation (DAFD) technique for gear well as bearing fault diagnosis. DAFD was used with MMD and weight regularization to learn portable features of vibration spectrum that link cross-domain discrepancies between source and target domain while supporting the minimal identifiable information in novel datasets. Heydarzadeh et al. 2016 ¹¹⁵ presented application of DWT and DNN to develop a robust gear fault diagnosis based on vibration, acoustic and torque signals. Results showed that the DWT-DNN approach is not only data-driven and needs little prior knowledge for feature extraction, but it is also non-sensitive to various load conditions and does not require precise assumptions for signal measurements. Jia et al. 2016 ¹¹⁶ demonstrated the capability of DNN for fault characteristic extraction and intelligent diagnosis of planetary gearbox as well as rolling element bearing faults with large amounts of data. Finally, they showed that DNN can extract the discriminative features from the vibration spectra with fine tuning and suggested to use hyperbolic tangent function as active function for higher accuracy.

Qi et al. 2017 ¹¹⁷ created a deep network based on stacked sparse auto encoders (SAE) for fault diagnosis of rotating machineries like gearbox and bearing. The time field novel signals were preprocessed using ensemble empirical mode decomposition (EEMD) and autoregressive (AR) systems to obtain AR constraint like inputs to the analysis method. These novel signals' low-level characteristics are used in the stacked SAE to obtain extra conceptual and sparse elevated high level characteristics feature. Xia et al. 2017 ¹¹⁸ introduced the fault diagnosis of rotating machinery (gear and bearing) based on different sensor signals and CNN. Sensor synthesis was achieved in the dataset stage and then both temporal and spatial information was obtained in this study to improve diagnosis accuracy and quality by integrating the unprocessed signals as the CNN system’s input. Jing et al. 2017 ¹¹⁹ developed a CNN-based function learning and fault detection technique for gearbox fault diagnosis. In this work, they used frequency domain vibration signals of planetary gearbox and compared the performance of methods with time domain and time-frequency domain vibration signals. The result showed that the CNN method performed better with frequency domain vibration signals in comparison with time and time-frequency techniques. Also CNN showed better performance than fully connected neural network (FNN), SVM and random forest (RF).

Rao and Zuo, 2018 ¹²⁰ developed a novel fault diagnosis techniques using order tracking DNN (OT-DNN) for gear and bearing fault diagnosis especially for varying speed condition. In this work, first the OT was used to resample the vibration signals to avoid frequency smearing then the spectra obtained from resample signals was used as input to DNN. In addition, the OT-DNN method showed much higher accuracy for fault diagnosis than Fourier transform-DNN (FT-DNN) especially under varying speed conditions. Han et al. 2019 ¹²¹ used characteristic selection to investigate the intellectual fault diagnosis of rotary machines using deep learning. In this work, they used frequency domain features of original vibration signals, and time and frequency domain features of intrinsic mode function (IMF) component obtained from original signals. It has the advantage of requiring a few labelled samples to be entered in the fine-tuning process to change the DNN’s constraints for effective diagnosis. Yu and Zhou 2019 ¹²² introduced gear fault diagnosis using one-dimension residual convolutional auto-encoder (1-DRCAE) based feature learning. In this work, 1-DRCAE developed a deep 1-D convolutional auto-encoder to suppress noise and extract fault characteristics from noisy vibration signals in order to achieve unsupervised learning. Sun et al. 2019 ¹²³ used adaptive separation and deep learning to diagnose compound fault of a planetary gearbox. An adaptive fault separation improved particle swarm optimization variation mode decomposition (IPVMD) method was used to decompose the vibration spectrum signals of combined fault to many single fault signals.

Fu and Wang 2020 ¹²⁴ developed a new deep learning approach for gear and bearing fault diagnosis using vibration signals based on dataset augmentation. This work introduced a hybrid fault diagnosis technique with a GAN and a SDAE to improve the issue of minimal fault data in real engineering scenarios. The GAN technique was used to supplement the limited actual measured dataset, which was mostly in defective states. The SDAE fault diagnosis framework then modified the created dataset. Ye et al. 2020 ¹²⁵ introduced a deep learning model for diagnosis of gear faults based on characteristic synthesis of multi-channel sensory signals. Initially, a deep formation multiple DNN (MDNN) comprised of various auto-encoders was constructed to adaptively extract important characteristics from sensory signals as well as to extract compound connections between indication and error samples. Second, MDNNs were used to combine delegate deep characteristics learned from multi-channel sensory datasets using locality preserving projection (LPP). Finally, the combined deep characteristics in softmax were used to build an intellectual investigation method.

Ha and Youn, 2021 ¹²⁶ performed fault diagnosis of planetary gearbox with the use of CNN and maximum classifier discrepency (MCD). In this study, vibration signals were first converted into image map using time syncronous averaging (TSA) which visualizes the fault toothwise. Then MCD a typical domain adaption method was employed to solve the domain shift issue in the data map. They considerd discrepency scale factor and ensemble appraoch to improve the performance and robustness of the CNN. Mallikarjuna et al. 2021 ¹²⁷ worked on fault diagnosis of aircraft gearbox using a DL technique. In this work, they used time and frequency domain vibration signals in long short term memory (LSTM) and Bi-directional LSTM (BLSTM) models to classify the gearbox condition. BLSTM model is more effective than LSTM with both time and frequency domain signals. In other work, Shi et al. 2022 ¹²⁸ applied Bi-directional convolution LSTM (BiConvLSTM) for fault diagnosis of planetary gearbox using vibration and rotational speed signals. Result showed that the BiConvLSTM model can not only detect the fault types but also the location of the fault even under, and it has high accuracy than other methods like LSTM, ConvLSTM, BiLSTM, CNN and CNN-BiLSTM.

A summary of the research papers for gear fault diagnosis based on DL techniques is added in the Table 3.

OPEN IN VIEWER

Table 3.

Summary of DL based gear fault diagnosis.

S. No.	Author, year	Machine faults	Signals	Technique	Remarks
118	Lu et al. 2016	Gearbox faults and bearing faults	Vibration signals	DAFD	The experimental outcomes on genuine world datasets prove the advantage of proposed model over LR, KNN, SVM and others
119	Heydarzadeh et al. 2016	Gearbox faults	Vibration, acoustic and torque signals	DWT-DNN	The proposed method is not only essentially data driven and requires negligible previous information for feature extraction, but also sensitive to different conditions and does not require detailed assumptions for signal measurements
121	Qi et al. 2017	Gearbox faults and bearing faults	Vibration and current signals	SAE	The proposed method shows enhanced diagnosis performance, which demonstrated the capacity of DL in without human intervention mining delegate characteristics
122	Xia et al. 2017	Roller bearing faults and gearbox faults	Vibration signals	CNN	CNN attains superior performance along with extra consistent analysis execution for the diverse kinds of machines as well as diverse faults
123	Jing et al. 2017	Gearbox faults	Vibration signals	CNN	The proposed method is capable to attain realistic detection outcomes for gearboxes with different faults and learn features efficiently
124	Rao and Zuo, 2018	Bearing faults and gearbox faults	Vibration signals	OT-DNN	OT-DNN performs superior rotating machinery fault diagnosis for the equally varying and constant speed conditions
125	Han et al. 2019	Bearing faults and gearbox faults	Vibration signals	DNN	DNN with time and frequency domain features of intrinsic mode function (IMF) reduces network size and performs efficiently
126	Yu and Zhou. 2019	Gearbox faults	Vibration signals	1-DRCAE	The proposed method has good fault denoising capability and enhanced feature extraction performance than other typical DNN methods
127	Sun et al. 2019	Planetary gearbox faults	Vibration signals	IPVMD and I-CNN	The proposed method is superior and effective than traditional methods like CNN.
128	Fu and Wang. 2020	Gear faults and bearing faults	Vibration signals	GAN-SDAE	The GAN with SDAE improves the diagnosis accuracy
129	Ye et al. 2020	Gear faults	Vibration signals	M-DNN	Multi-channel sensory signals overcomes the problem with single channel sensor data
131	Mallikarjuna et al. 2021	Aircraft gearbox	Vibration	CNN and MCD	Combination of CNN and MCD is robust and effective under stationary and non-stationary speed conditions
130	Ha and Youn, 2021	Planetary gear box	Vibration	LSTM and BiLSTM	BiLSTM is better than LSTM, CNN, SVM, RF, DT and other methods
132	Shi et al. 2021	Planetary gearbox	Vibration and rotating speed signals	BiConvLSTM	BiConvLSTM is better than LSTM, ConvLSTM, BiLSTM, CNN and CNN-BiLSTM

DL based turbine fault diagnosis

A study of DL-based techniques developed for wind turbines in the last 5 years is addressed in this sub-section.

Wang et al. 2016 ¹²⁹ explored the deep auto encoders to detect wind turbine blade breakage. The supervisory control and data acquisition (SCADA) dataset for each wind turbine condition was calculated, and a 1-step contrastive divergence (CD) algorithm was used to train the DA model. The results showed that the proposed method could be used to identify blade breakages in the near future using SCADA data. Wang et al. 2016 ¹³⁰ used DNN to investigate the failure of a wind turbine (WT) gearbox. The proposed method was built in two phases in this study: modelling and tracking the lubricant strain. The first phase focused on the development of a precise lubricant pressure prediction model of WT with healthy gearboxes’ using SCADA data. The prediction model was expanded using a DNN algorithm, which was compared to other models. An exponentially weighted moving average (EWMA) manage graph was used in the next step to obtain guidelines for monitoring lubricant strength. The present DNN method produced more accurate prediction outcomes and a more effective model for monitoring lubricant pressure.

In a study, Cheng et al. 2017 ¹³¹ used frequency analysis and a DL classifier to create a rotor-current-based fault diagnosis for drive train gearboxes of doubly-fed induction generator (DFIG) wind turbines. In this work, a classifier was projected in favour of gearbox defect categorization by extracting fault characteristics using a deep structural architecture that included a stacked auto encoder (SAE) and a SVM. The tests were carried out on a DFIG wind turbine drive train test rig in the suit, with four different gear defect states. Teng et al. 2018 ¹³² presented a novel approach to fault detection in a direct drive wind turbine (DDWT) based on a DNN model. For a defective wind turbine, the multi-layer learning characteristic of DNN for the input features exhibits brilliant fitting ability between input variables and output variables, allows for the detection of the degree of deviation of the online data from the normal state. The present DNN model effectively detected the fall off of the permanent magnets in the DDWT generator.

Han et al. 2018 ¹³³ used a new novel deep adversarial convolutional neural network (DACNN) for diagnosis of mechanical defects in wind turbines. In this study, they considered bearing faults, misalignment, and variance within the airfoil of blades, as well as Yaw faults. The present technique was found to be efficient and advantageous in fault diagnosis, especially when it came to identifying undetected states through system learning. Qian et al. 2019 ¹³⁴ used a long short-term memory (LSTM) neural system to conduct a fresh state analysis of wind turbines. Supervisory control and data acquisition (SCADA) data from a commercial wind farm was collected over a 12-month span to verify the anticipated system’s execution. Gearbox oil and bearing temperature were used as inputs to the LSTM in this case. The experiment shows that LSTM outperforms traditional back propagation neural network algorithms in terms of prediction accuracy.

Afrasiabi et al. 2019 ¹³⁵ used GAN and temporal convolutional neural networks (TCNN) to investigate wind turbine (WT) fault diagnosis. The proposed approach is split into two parts. GAN was used to mine and pick useful characteristics from signals, as well as to inscribe machine fault distinguisher learning using a limited number of samples. Tan et al. 2019 ¹³⁶ provided an analysis of smart condition monitoring with dropouts and cross-validation using deep neural networks (DNN). In this work, the present approach used real datasets of wind turbines to determine the best number of epochs and nodes for the DNN. They showed that the number of nodes in the predicted technique is the determining factor in the neural system’s outcome.

Han et al. 2020 ¹³⁷ combined deep transfer network (DTN) with joint distribution adaptation (JDA) technique to develop a new intelligent fault diagnosis technique for industry. They analysed experimental data from bearings, gearboxes, and wind turbines in this study. Inspired from tranfer learning, JDA was used to analysis disrimination features associated with source domain labelled data and to adjust the conditional distribution of unlabelled data in target domain in order to get effective distribution matching. Prosvirin et al. 2020 ¹³⁸ used an auto encoder with nonlinear function approximation and a DNN to identify blade rub-impact defect in turbine. The deep undercomplete denoising auto encoder (DUDAE) based on the resample vibration signals was used to learn the nonlinear function approximation of the system condition for normal working conditions. Then the remaining signals from unknown fault states were used as inputs to the deep neural system, which was used to create the residual signals for producing discriminative features based on auto encoder. The results showed that the amplitude of residual signals with DNN model is very effective in identifying complex fault like blade rub-impact. In a study, Nie et al. 2020 ¹³⁹ introduced a new autoencoder with an active characteristic enhanced aspect for efficient wind turbine fault diagnosis. A model which incorporates a noise addition approach and denoising stacked feature enhanced autoencoders based on dynamic feature improved aspect (DSFEAE-DF) was proposed to extract extra discriminative features from raw signals. The influence of neurons was also visualized to demonstrate the perspectives of function enhancement. Results showed that the present approach worked very well for turbine fault diagnosis even under noisy environments compared to SAE, k-sparse SAE and others.

Zhang et al. 2021 ¹⁴⁰ developed a hybrid attention improved residual network (HA-ResNet) based fault diagnosis of a turbines of a wind farm. In this work, they combined the WPT obtained from the raw vibraiton singals and the ResNet an improved CNN model for further improving the diagnosis perforamnce by giving attention to key frequency band of wavelet coefficients. Wang et al. 2021 ¹⁴¹ introduced a novel collaborative deep learning framework for fault diagnosis of a wind turbine. In this work, they cosidered a number of fault condtions like misalignment, loosninng, Outer-ring beairng fault, mass unbalance of wind wheel, variation in airfoil of blades, yaw fault and aero-asymmetry of wind wheel. Result showed that the present method of fault diagnosis is better than local learning scheme and also avoids many problems of covnetional learning like big data, unbalanced data, big computational burden and overfitting.

Yu et al. 2021 ¹⁴² introduced a fast deep graph convolution network (FDGCN) for fault diagnosis of wind turbine fault diagnosis. In this work, they obtained time-frequency graph of vibration signals based on WPT. Finally they employed particular pooling operation for optimizng the training time and graph convolutional kernels for extracting features from the nodes and edges of the graph. Yang et al. 2021 ¹⁴³ performed DL based blade damage detection of wind turbine using image recognition. In this work, they used transfer learning to improve features extraction and ensemble learning using random forest to improve the CNN perforamance. Result showed that the present appraoch has potenial to detect the blade damage based on unmanned aerial vehicle (UAV) images. In other work, Li et al. 2021 ¹⁴⁴ performed successful fault diagnosis of wind turbine based on transfer learning model and convolution autoencoder (CAE) even with small samples. Guo et al. 2021 ¹⁴⁵ developed a effective fault diagnosis appraoch in which combined informations from WPT of vibraton signals, domain knowledge and operating conditions of turbine gear box were used as rienforced input to CNN.

A summary of the research papers for turbine fault diagnosis based on DL techniques is added in the Table 4.

OPEN IN VIEWER

Table 4.

Summary of DL based turbine fault diagnosis.

Ref. no.	Author, year	Machine faults	Signals	Technique	Remarks
133	Wang et al. 2016	Wind turbine blade breakages	Current signals	Deep auto encoders	Results established that the proposed method is appropriate to classify in the near future blade breakages based on SCADA data
134	Wang et al. 2016	Gearbox faults	Vibration signals	DNN	DNN method presents more precise prediction outcomes and more appropriate model in monitoring the lubricant pressure
135	Cheng et al. 2017	Wind turbine drivetrain gearbox faults	Current signals	Deep learning (SAE-based multiclass SVM deep Classifier)	Present method is efficient to identify as well as enhanced than the conventional multiclass SVM technique to recognize gear faults under different operating conditions
136	Teng et al. 2018	Direct drive wind turbine (DDWT) faults	Vibration signals	DNN	The proposed DNN-based approach is valuable for the maintenance program and cost management in a wind farm
137	Han et al. 2018	Mechanical defects in wind turbine	Vibration signals	DACNN	The proposed method shows the efficiency and advantage in fault diagnosis, mainly when diagnosing the unseen conditions through model training
138	Qian et al. 2019	Wind turbine faults	Current signals	LSTM	The proposed method is appropriate to be applied in the condition monitoring system, which can decrease the cost of condition monitoring and enhance monitoring accurateness
139	Afrasiabi et al. 2019	Wind turbine faults	Vibration signals	GAN-TCNN	The proposed method gives the highly accurate data in the noisy conditions and very efficient performance in fault condition identification compared with other methods
140	Tan et al. 2019	Wind turbine faults	Vibration and current signals	DNN	In the proposed method number of nodes plays the deciding aspect of the neural network for the accurateness
142	Prosvirin et al. 2020	Blade rub-impact faults and rotor faults	Vibration signals	DUDAE-DNN	The projected fault diagnosis system established steady meeting performance beneath diverse training–testing datasets variations as well as outperformed further techniques used for the evaluation in conditions of the micro-averaged presentation metrics
143	Nie et al. 2020	Wind turbines faults	Vibration signals	DSFEAE-DF	The proposed method achieves better performances in noise surroundings faults diagnosis and the reactions of neurons are visualized to illustrate the insights of feature improvement
144	Zhang et al. 2021	Wind turbine gearbox and bearing	Vibration	HA-ResNet	HA-ResNet is much better than CNN-raw, WDCNN-raw, CNN-wave, WDCNN-wave, and ResNet model
145	Wang et al. 2021	Wind turbine faults	Vibration	Collaborative deep learning framework	Distributuion based present method is better than conventional raw data based local learning scheme
146	Yu et al. 2021	Wind turbine bearing and gearbox faults	Vibration	FDGCN model	FDGCN model is better to overcome noise than WDCNet, LeNet, ResNet, and GCNet models
147	Yang et al. 2021	Blade damage of turbine	UAV images of blade	CNN	CNN with UAV images perform better blade diagnosis
148	Li et al. 2021	Wind turbine faults	Vibration	CAE	CAE with transfer learning perform better even with small no. of samples
149	Guo et al. 2021	Gearbox of wind turbine	Vibration and operating condition signals	CNN	Reinforced input in CNN gives better performance

DL based pump and compressor fault diagnosis

This section reviews the literature published in the last 5 years on DL-based pump and compressor fault diagnosis.

Thirukovalluru et al. 2016 ¹⁴⁶ proposed using a denoising stacked auto-encoder to generate feature datasets for air compressors, drill bit monitoring and steel plate monitoring. In this paper DNN was used to create useful feature from handcrafted features extracted from conventional techniques such as FFT and WPT. Yan et al. 2016 ¹⁴⁷ considered CNN for diagnosis hydraulic pump faults such as cylinder wear, valve wear, roller wear, loose slipper, sliding boot wear and spring wear under stable and variable pump speeds. They showed that the present CNN based method is simple and effective even for varying speed conditions. In addition, the performance of CNN was found higher than DBN, SVM and BPNN. Wen et al. 2017 ³⁸ employed a new LeNet-5 based CNN method for fault diagnosis of motor bearing, centrifugal and axial piston hydraulic pump. They converted 1D vibration signals in to 2-D images and then CNN was used to extract features and perform the diagnosis. Result showed that the present method has higher potential than other methods like deep CNN, DBN, sparse filter and SVM.

Wang et al. 2018 ¹⁴⁸ used DBN to perform multiple fault diagnosis in axial piston pump. In this work, they used time, frequency and time-frequency domain data for characterizing pump faults and then DBN was used to automatic learn the features using RBM. Results showed that the performance of DBN based methodology is higher than the SVM and ANN for multiple fault diagnosis of pump. Yang et al. 2019 ¹⁴⁹ performed centrifugal pump fault diagnosis for impeller and bearing faults based on hierarchical symbolic analysis (HSA) and CNN. In this work, HSA was employed to decompose the raw time domain signals into a no. of low and high frequency components. Result showed that the combination of HSA-CNN performed better than SVM, RF, STFT-CNN and CWT-CNN.

Tang et al. 2019 ¹⁵⁰ presented a review of DL techniques in condition monitoring rotary machines like pumps, bearing and gears. They showed the various advantages of DL such as automatic feature extraction, effective dimensional reduction, imbalanced data solution, handling varying fault information and etc. in condition monitoring over other AI techniques like, SVM, ANN and others. Maurya et al. 2020 ¹⁵¹ used DNN to conduct condition-based monitoring of rotating machines under variable system running conditions. They used vibration data to track bearing faults and acoustic data to monitor faults in air compressors using DNN. They showed that the DNN’s computational complexity was reduced through a related enhancement in execution by using low-level features or derived features from time, frequency, and time-frequency domain. Finally result showed that the present low-level feature based DNN technique improves diagnosis accuracy with little computational time.

Hasan et al. 2021 ¹⁵² successfully used grayscale images from 2D time-frequency graph called scalogram obtained using CWT of vibration signals as input to adaptive deep CNN (ADCNN) for centrifugal pump fault diagnosis. Jiang et al. 2021 ¹⁵³ combined empirical wavelet transform (EWT) and 1D-CNN for fault diagnosis of axial pump using vibration and pressure signals. They fused time and frequency domain features obtained from denoised signals by EWT and further used as input to CNN for effective fault diagnosis. In other work, Bie et al. 2021 ¹⁵⁴ used singular spectral entropy of intrinsic modal function (IMF components of vibration signals as input to LSTM DNN for performing successful diagnosis of reciprocating pump.

A summary of the research papers for fault diagnosis of pump and compressor based on DL techniques is added in the Table 5.

OPEN IN VIEWER

Table 5.

Summary of DL based pump and compressor fault diagnosis.

S. No.	Author, year	Machine faults	Signals	Technique	Remarks
150	Thirukovalluru et al. 2016	Air Compressor faults, drill bit faults and steel plate faults and bearing faults	Vibration signals	DNN	DNN based features is capable to provide enhanced categorization execution along with FFT characteristics and perform better than SVM and RF
151	Yan et al. 2016	Axial Pump faults (cylinder wear, valve wear, roller wear, loose slipper, sliding boot wear and spring wear)	Vibration	CNN	CNN perform well evenin varying operating speeds and better than DBN, SVM and BPNN
152	Wen et al. 2017	Motor bearing, centrifugal and axial piston hydraulic pump	Vibration	CNN	CNN has better perforamnce than deep-CNN, DBN, sparse filter and SVM
153	Wang et al. 2018	Multiple fault in axial pump	Vibration	DBN	DBN performs better than SVM and ANN for multiple fault diagnosis
154	Yang et al. 2019	Impeller and bearing faults in centrifugal pump	Vibration	CNN	CNN used with HSA perform better than STFT-CNN, CWT-CNN and others
155	Tang et al. 2019	Rotating machinery faults (pump, bearing and gear faults)	Vibration signals	DL	DL has advantages such as automatic feature extraction, effective dimensional reduction, and imbalanced data solution
156	Maurya et al. 2020	Bearing faults, air compressor faults	Vibration and acoustic	DNN	Low level features from time, frequency, time-frequency domain improve the DNN performance
157	Hasan et al. 2021	Impeller crack, mechanical seal scratch and hole in centrifugal pump	Vibration signals	ADCNN	ADCNN with scalogram images perform well
158	Jiang et al. 2021	Axial pump faults (single shoe loosening, slipping and wear, swash plate wear, and center spring failure	Vibration and pressure signals	1D-CNN	EWT presents better signal denoising effect. EWT-CNN perform well
159	Bie et al. 2021	Reciprocating pump faults (piston wear, bearing wear, valve disc wear)	Vibration signals	LSTM DNN	LSTM DNN perform better than RBF, PNN and BP