PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine Learning and Deep Learning Algorithms for Bearing Fault Diagnostics -- A Comprehensive Review

Shen Zhang, Shibo Zhang, Bingnan Wang, Thomas G. Habetler

arXiv:1901.08247v3cs.LGstat.ML

TL;DR

Bearing-fault diagnosis must handle complex, noisy, and variable signals, while conventional methods depend on fault-frequency expertise and substantial operating-condition data. This survey reviews ML and DL methods, compares DL variants on the CWRU dataset, and synthesizes application guidance and future directions. DL methods are presented as capable of adaptive feature extraction without prior expertise on fault characteristic frequencies or operating conditions, although they require large datasets for training.

  • Problem

    Bearing faults are difficult to diagnose when faults are incipient, signals are noisy, operating conditions vary, and conventional approaches depend on manually specified fault-frequency knowledge and healthy commissioning data.

  • Method

    The paper systematically reviews conventional ML and DL bearing-diagnostics literature, analyzes DL feature extraction and classification, compares algorithms on CWRU, and provides application recommendations and future directions.

  • Results

    DL methods can automatically perform adaptive feature extraction without prior expertise on fault characteristic frequencies or operating conditions, while reported CWRU variants include 99.27% mean accuracy for MS-DCNN and 99.15% average classification accuracy for an ensemble deep auto-encoder.

  • Takeaways & Limitations

    DL methods are presented as promising candidates for real-time bearing-fault diagnostics, with recommendations tied to setup environment, data size, sensors, and sensor types.

Abstract

from arXiv · show

In this survey paper, we systematically summarize existing literature on bearing fault diagnostics with machine learning (ML) and data mining techniques. While conventional ML methods, including artificial neural network (ANN), principal component analysis (PCA), support vector machines (SVM), etc., have been successfully applied to the detection and categorization of bearing faults for decades, recent developments in deep learning (DL) algorithms in the last five years have sparked renewed interest in both industry and academia for intelligent machine health monitoring. In this paper, we first provide a brief review of conventional ML methods, before taking a deep dive into the state-of-the-art DL algorithms for bearing fault applications. Specifically, the superiority of DL based methods over conventional ML methods are analyzed in terms of fault feature extraction and classification performances; many new functionalities enabled by DL techniques are also summarized. In addition, to obtain a more intuitive insight, a comparative study is conducted on the classification accuracy of different algorithms utilizing the open-source Case Western Reserve University (CWRU) bearing dataset. Finally, to facilitate the transition on applying various DL algorithms to bearing fault diagnostics, detailed recommendations and suggestions are provided for specific application conditions such as the setup environment, the data size, and the number of sensors and sensor types. Future research directions to further enhance the performance of DL algorithms on health monitoring are also discussed.

I. INTRODUCTION

Bearing faults are a major source of electric-machine failures, yet practical diagnosis is difficult because signals are noisy, operating conditions vary, and faults may be complex or incipient. This review surveys conventional ML and recent DL approaches for extracting fault features and classifying bearing conditions.

  • Bearing faults account for 30% to 40% of induction-machine failures, creating maintenance costs, financial losses, and safety hazards.
  • Physical-model diagnosis relies on characteristic fault frequencies, but sliding and multiple simultaneous faults can make those frequencies inaccurate or ambiguous.
  • Incipient faults are difficult to identify when monitored signals have small signal-to-noise ratios, while multi-physics interactions and sensor placement affect vibration-analysis accuracy.
  • Motor current signature analysis avoids some vibration-sensing constraints but requires healthy-data commissioning across loads and speeds because current magnitudes vary with operating conditions.
  • The review covers conventional ML methods and more than 180 bearing-diagnosis papers, including around 80 DL studies, with emphasis on DL feature extraction, classification, and additional functionalities.

II. POPULAR BEARING FAULT DATASETS

Bearing-diagnosis datasets support algorithm development and common evaluation, but they differ in test-rig design, sensing modalities, fault realism, and operating conditions. The review introduces CWRU, Paderborn, and PRONOSTIA-related experimental resources used across the literature.

  • Published bearing datasets provide a common ground for developing, evaluating, and comparing ML and DL algorithms despite the time required for data collection.
  • The CWRU test stand combines a 2-hp induction motor, torque transducer/encoder, and dynamometer, with single-point faults introduced at multiple diameters and locations.
  • CWRU is used as a fundamental dataset for validating and comparing the performance of different ML and DL algorithms.
  • Paderborn synchronously measures motor current and vibration from 26 damaged and 6 healthy bearings, supporting multi-physics modeling and sensor-fusion studies.
  • Paderborn combines artificially damaged bearings with accelerated-life-test defects, enabling evaluation involving more realistic aging and lubrication-loss damage.

C. PRONOSTIA Dataset

PRONOSTIA is an accelerated-life dataset for bearing remaining-useful-life prediction, whereas IMS provides long-duration natural defect evolution records. Their testbeds use controlled operating conditions and multiple sensing modalities for degradation monitoring.

  • PRONOSTIA supports research on bearing remaining useful life prediction through accelerated degradation experiments under varying operating conditions.
  • PRONOSTIA characterizes operating conditions using rotating-speed and force sensors while monitoring temperature and vibration with two uni-axis accelerometers.
  • The IMS dataset records natural bearing-defect evolution during 30 days of continuous operation at 2,000 rpm before defect confirmation.
  • IMS uses two accelerometers on each bearing housing and four thermocouples on each bearing’s outer race to monitor vibration, temperature, and lubrication-related conditions.
  • IMS experiments include inner-race, rolling-element, and outer-race defects, with vibration sampled at 20 kHz in repeated tests from operation toward failure.

E. Summary

Classical machine-learning studies use domain-informed feature extraction and classifiers for bearing diagnostics, while PCA can reduce inputs and improve accuracy. The survey organizes prior methods and highlights transferability challenges.

  • Datasets: Most bearing-fault studies use the CWRU dataset because of its simplicity and popularity, while Paderborn supports current-and-vibration analysis and transfer-learning validation.The survey presents a comparative study of methods using CWRU and identifies Paderborn as a more realistic evaluation resource.
  • Classical ML workflow: Classical machine-learning workflows typically perform exploratory analysis, dimension reduction, feature extraction, and classification, requiring substantial domain expertise.PCA is described as one dimension-reduction technique used to obtain representative features before classification.
  • Limitations: Classical ML models may be difficult to transfer across contexts because feature extraction and model generalization depend on specialized knowledge.The survey notes that domain knowledge differs across applications and can hinder appropriate feature extraction and transferability.
  • Classical ML methods: The survey reviews established methods including ANN and provides a comprehensive reference list of publications applying classical ML to bearing diagnostics.ANN applications include mapping stator current and rotor speed measurements to bearing-related quantities through supervised neural-network training.
  • PCA: PCA provides lower-dimensional projections and systematic guidance for selecting representative features whose sensitivity varies across operating conditions.This supports manual feature selection for classification.
  • PCA: PCA increased bearing-fault diagnosis accuracy from 88% to 98% using identified features instead of 13 original features.The study also reported higher accuracy with fewer input features than when using all original features.

C. K-Nearest Neighbors (k-NN)

k-NN classifies bearing-fault samples through nearest-neighbor voting, while classical feature engineering uses calculated fault frequencies. These engineered features can become unreliable under realistic fault and operating conditions, motivating broader ML and DL approaches.

  • k-NN: k-NN assigns a new sample to a fault class by majority vote among its k nearest neighbors.Bearing-diagnostics studies use distance analysis on each new sample to determine its fault class.
  • Feature engineering: Classical bearing-fault diagnosis calculates characteristic frequencies from rotor speed and bearing geometry, then monitors signal amplitudes as engineered features.These features are used to train ML algorithms and identify anomalies.
  • Challenges: Sliding between rolling elements and raceways can shift calculated fault frequencies away from their real values, reducing feature informativeness.The no-sliding assumption underlying frequency calculation is seldom satisfied in practice.
  • Challenges: Multiple simultaneous bearing faults can interact, causing characteristic frequencies to add or subtract and obscure informative frequencies.This frequency interplay complicates feature-based diagnosis.
  • Challenges: External vibration can obscure useful bearing-fault features, while lubrication and roughness faults may lack characteristic cyclic frequencies.The latter faults are therefore difficult to detect using traditional characteristic-frequency methods.
  • Deep learning: Deep-learning performance depends strongly on data volume: classical ML can compete with or outperform DL on small datasets, whereas DL can outperform classical ML as data increases.Bearing-fault datasets are often expensive and time-consuming to acquire.
  • Deep learning: Deep learning learns nested representations and can remove manual feature engineering by extracting features directly from raw data.The paper contrasts this with classical ML’s reliance on sophisticated, expert-designed features.

A. Convolutional Neural Network (CNN)

CNN-based bearing diagnosis models learn hierarchical features from vibration data and have been adapted for raw 1-D signals, noise, speed variation, sensor fusion, and limited training data.

  • CNN fundamentals: CNNs hierarchically extract higher-level features from lower-level patterns through successive convolutional layers.The approach was inspired by visual-cortex processing, progressing from simple features such as edges and corners to more complex features.
  • CNN fundamentals: A typical bearing-fault CNN stacks 1-D temporal data from multiple accelerometers into a 2-D representation before convolutional feature extraction and pooling.This image-like transformation enables conventional CNN architectures to process vibration measurements.
  • Architectural variants: 99.41% average accuracy was obtained by fusing two accelerometers, compared with 98.35% using one sensor; CNNs also improved performance under ambient noise.Other variants included adaptive learning rates, dislocate layers, and denoising behavior under noisy signals.
  • Speed variation: 99.63% accuracy was achieved by LiftingNet on a four-class CWRU split, while accuracy remained 93.19% across four distinct rotor frequencies.LiftingNet targets speed variation through split, predict, and update layers; PSPP-CNN similarly reports higher accuracy at varying speeds.
  • Architectural variants: 99.19% test accuracy was achieved by AOCNN with sparse filtering on ten bearing health conditions using only 5% of samples for training.AOCNN directly processes raw 1-D vibration signals through overlapping sampling and adaptive convolution, avoiding the usual 1-D-to-2-D conversion.

B. Auto-encoders

Auto-encoders provide unsupervised feature learning for bearing diagnostics, with variants improving classification across datasets while often increasing computational cost.

  • B. Auto-encoders: Auto-encoders combine an encoder and decoder for unsupervised feature learning and layer-wise neural-network pre-training.Their one-hidden-layer training process is illustrated in Fig. 8.
  • B. Auto-encoders: 99.6% classification accuracy exceeded the 70% achieved by back-propagation neural networks using a 5-layer auto-encoder DNN.An auto-encoder-based ELM reached 99.83% average accuracy while reducing training time by around 60% to 70%.
  • B. Auto-encoders: 91.90% accuracy from DAE-CAE-LPP exceeded standard DAE, standard CAE, BPNN, and SVM baselines on a locomotive bearing dataset.The reported values were 84.60%, 85.10%, 49.70%, and 57.60%, respectively.
  • B. Auto-encoders: Auto-encoder methods can be 6 to 10 times more time-consuming than classical machine-learning methods.This cost was reported alongside the improved classification performance of a mixed DAE-CAE-LPP approach.
  • B. Auto-encoders: 99.15% average accuracy from an ensemble deep auto-encoder exceeded BPNN, SVM, and RF using manually selected 24-dimensional features.The comparison accuracies were 88.22%, 90.81%, and 92.07%, respectively, on the CWRU dataset.
  • B. Auto-encoders: 97.47% accuracy was obtained with a stacked sparse-autoencoder DNN using only 30% of the original data after 70% compression.An SAE-LCN variant reported 99.92% average accuracy for a 10-class classification problem.

C. Deep Belief Network (DBN)

Deep belief networks compose unsupervised subnetworks for layer-wise training and have been applied to bearing classification, feature fusion, and remaining-useful-life prediction.

  • C. Deep Belief Network (DBN): A DBN composes restricted Boltzmann machines or auto-encoders, with each hidden layer serving as the next subnetwork’s visible layer.This supports fast layer-by-layer unsupervised training using contrastive divergence.
  • C. Deep Belief Network (DBN): 97.82% accuracy was achieved by fusing time- and frequency-domain features from multiple 2-layer SAEs before 3-layer DBN classification.Validation used vibration data collected at different speeds.
  • C. Deep Belief Network (DBN): 94.38% average accuracy exceeded ANN, GRNN, and SVM results of 63.13%, 69.38%, and 66.88%, respectively, using the same training and test data.The result came from a 5-layer adaptive DBN after order-5 Daubechies wavelet processing.
  • C. Deep Belief Network (DBN): DBNs were evaluated beyond CWRU, including locomotive-bearing datasets and remaining-useful-life prediction.The RUL approach combined automatic feature learning with DBN and prediction with a feed-forward neural network.

D. Recurrent Neural Network (RNN)

RNNs model sequential or time-series relationships, while combined CNN–LSTM systems apply convolutional and recurrent processing to bearing-fault classification and GANs address data-generation and imbalance problems.

  • D. Recurrent Neural Network (RNN): RNNs use recurrent hidden-layer connections to capture sequential relationships in sequential or time-series data.When unrolled, the recurrent architecture can be viewed as a feed-forward network over the input sequence.
  • D. Recurrent Neural Network (RNN): An early RNN bearing-diagnostics scheme extracted and selected fault features before feeding them into the RNN for detection and classification.The reported experiment found the scheme capable of accurately detecting and classifying bearing faults.
  • D. Recurrent Neural Network (RNN): A 1-D CNN, max-pooling, LSTM, and Softmax architecture achieved 99.6% best test accuracy using raw signals without preprocessing.A DRNN reported 94.75% and 96.53% average test accuracy at 1,750 and 1,797 rpm, respectively.
  • D. Recurrent Neural Network (RNN): GAN-based data balancing with DCGAN produced better training and test accuracy than random over-sampling, random under-sampling, and SMOTE.The balanced signals were followed by time-domain and frequency-domain feature extraction and SVM classification.
  • D. Recurrent Neural Network (RNN): Adversarial adaptive CNNs were developed for changing operating conditions, with reported fault-discriminative and domain-invariant capacity.GAN variants were also used for classification and data-imbalance problems.

F. Deep Learning based Transfer Learning

Deep transfer-learning methods address bearing diagnostics when operating conditions and data distributions differ across domains by learning transferable or domain-invariant representations.

  • F. Deep Learning based Transfer Learning: Deep-learning diagnostics require massive annotated data, which is often infeasible because faulty operation is dangerous, degradation is time-consuming, and conditions vary.Different speeds and loads can create many operating conditions for training data collection.
  • F. Deep Learning based Transfer Learning: Domain adaptation transfers knowledge from a source domain to a target domain by exploring domain-invariant features.The approach is motivated by real-world cases where training and test distributions differ as operating conditions vary.
  • F. Deep Learning based Transfer Learning: 86.3% average accuracy across cross-dataset training and testing exceeded conventional CNN at 53.1% and two existing frameworks at 75.6% and 78.8%.The evaluation used CWRU, IMS, and railway locomotive bearing datasets.
  • F. Deep Learning based Transfer Learning: A two-stage deep generative domain-adaptation framework first trains a source-supervised CNN feature extractor, then trains cross-domain components and a final classifier.The later components minimize maximum mean discrepancy before cross-domain classification.

G. Other Variants

The review highlights deep-learning variants beyond standard architectures, including VAE, capsule, and siamese networks, with applications to representation learning, generalization, and limited-data diagnosis.

  • Variational Auto-Encoder: VAE uses variational inference to impose a distribution over latent variables and data, supporting unsupervised bearing-fault representation learning.A representative study applies a fully unsupervised deep VAE-based approach.
  • Variational Auto-Encoder: VAE showed greater competence than PCA for dimensionality reduction, while its generative capabilities may support semi-supervised fault diagnosis.
  • Capsule Neural Networks: Capsule networks identify feature position and orientation relationships while using relatively few parameters, promoting model generalization.
  • Capsule Neural Networks: An inception-augmented capsule network improved adaptation across workloads and added a regression branch to predict bearing-defect size.
  • Siamese Neural Networks: Siamese neural networks compare distinct inputs and rank their similarities, motivating their use for few-shot bearing-fault learning.

3) Siamese Neural Networks:

The review contrasts deep learning with engineered-feature ML and examines CWRU-based accuracy comparisons, while emphasizing that high accuracy alone does not establish general superiority.

  • Automated Feature Extraction and Selection: Deep learning learns discriminative representations directly from raw data end-to-end, unlike ML methods that manually engineer features.
  • Automated Feature Extraction and Selection: Most reviewed deep-learning methods reportedly outperform traditional ML methods, especially under external noise and changing operating conditions.
  • CWRU Performance Comparison: The review compares classification accuracy across deep-learning algorithms on the CWRU bearing dataset.
  • CWRU Performance Comparison: All evaluated deep-learning algorithms achieved test accuracy above 95% on the CWRU dataset.
  • CWRU Performance Comparison: Reported accuracy comparisons are weakened by fixed operating conditions, noise and load variation, unbalanced sampling, random splits, and accuracy saturation.

VI. SUGGESTIONS, CHALLENGES, AND FUTURE WORK DIRECTIONS

The authors recommend selecting ML or DL methods according to operating conditions, disturbances, sensor configuration, and available data characteristics.

  • Setup environment: Algorithm selection should begin by examining whether the system operates indoors or outdoors and at fixed or multiple speeds and loads.
  • Setup environment: Classical ML methods may suit simple single-point indoor setups, whereas advanced deep learning is recommended for disturbances or multiple operating points.
  • Setup environment: Noisy environments with small SNR require denoising blocks and additional hidden layers to improve neural-network noise resilience and robustness.
  • Sensors: The recommended process also examines the number and type of sensors mounted near the bearing before choosing traditional or deep-learning methods.

B. Current Challenges

The review identifies major barriers to real-world deployment, including laboratory-to-field transfer, limited labels, imbalance, and noisy industrial data, and proposes future research directions.

  • Knowledge transfer from laboratories to the real world: Most reviewed models use laboratory datasets, leaving transfer of learned structures and parameters to unseen real-world setups as an unresolved challenge.
  • Limited labels: Naturally occurring bearing faults are difficult to label precisely, particularly regarding their first trace and incipient duration.
  • Data imbalance: Healthy-condition data often dominates collections, while faulty-condition data can be costly or difficult to obtain for effective DL training.
  • Noisy data: Industrial vibration, resonance, and environmental noise challenge methods developed mainly from laboratory accelerometer data.
  • Future work directions: The authors suggest transfer learning, generative models, few-shot learning, and explainability as directions for addressing these challenges.
  • Conclusion: The review concludes that deep learning can perform adaptive feature extraction without prior expertise, although it requires large training datasets.
Loading 1901.08247v3…