PaperlayerView exact revision on arXiv

Source-linked AI summary

Few-Shot Bearing Fault Diagnosis Based on Model-Agnostic Meta-Learning

Shen Zhang, Fei Ye, Bingnan Wang, Thomas G. Habetler

arXiv:2007.12851v4cs.LGstat.ML

TL;DR

Bearing fault diagnosis often requires more samples than can safely or practically be collected for each fault type. This paper uses MAML for few-shot diagnosis and reports an advantage of up to 25% over a Siamese Network benchmark, alongside evaluation on real damage using artificial-damage data.

  • Problem

    Collecting sufficient fault samples can be unsafe, time-consuming, expensive, or impractical, limiting training data for robust bearing fault classifiers.

    Source: §I · p. 1

  • Method

    The paper proposes a MAML-based few-shot bearing fault diagnostic framework that learns to adapt to new fault scenarios from limited data.

    Source: §I · p. 2

  • Results

    The framework's advantage over the Siamese Network reaches 25% on new artificial faults, and it delivers comparable performance on real-damage adaptation while competing with few-shot methods.

    Source: §V · p. 9

  • Takeaways & Limitations

    The results offer promising prospects for identifying naturally evolved bearing failures using laboratory data collected from artificially induced faults.

    Source: §V · p. 9

    The work focuses on learning per-step rates for vanilla MAML; other optimization strategies to improve robustness remain future work.

    Source: §V · p. 9

Abstract

from arXiv · show

The rapid development of artificial intelligence and deep learning has provided many opportunities to further enhance the safety, stability, and accuracy of industrial Cyber-Physical Systems (CPS). As indispensable components to many mission-critical CPS assets and equipment, mechanical bearings need to be monitored to identify any trace of abnormal conditions. Most of the data-driven approaches applied to bearing fault diagnosis up-to-date are trained using a large amount of fault data collected a priori. In many practical applications, however, it can be unsafe and time-consuming to collect sufficient data samples for each fault category, making it challenging to train a robust classifier. In this paper, we propose a few-shot learning framework for bearing fault diagnosis based on model-agnostic meta-learning (MAML), which targets for training an effective fault classifier using limited data. In addition, it can leverage the training data and learn to identify new fault scenarios more efficiently. Case studies on the generalization to new artificial faults show that the proposed framework achieves an overall accuracy up to 25% higher than a Siamese network-based benchmark study. Finally, the robustness and the generalization capability of the proposed framework are further validated by applying it to identify real bearing damages using data from artificial damages, which compares favorably against 6 state-of-the-art few-shot learning algorithms using consistent test environments.

I. INTRODUCTION

Bearing fault classifiers often need more labeled data than real-world settings can safely or practically provide. The paper proposes MAML-based few-shot diagnosis, reporting stronger generalization to unseen artificial faults and testing transfer from artificial to real damage.

Motivation

  • Collecting enough samples for each fault condition can be unsafe, expensive, and impractical because faults may evolve over months or years and some applications cannot run in faulty states.Insufficient samples can also produce data imbalance.Source: §I · p. 1

Related work

  • Prior few-shot bearing studies include test faults already represented during training, so they do not evaluate generalization to unseen fault classes.The paper notes that the case studies in likewise use fault classes seen during training, while [23] varies defect diameters and operating conditions.Source: §I · p. 2

Approach

  • The paper uses MAML to train initial model parameters intended to perform well on a new task after one or more gradient updates using only a small amount of task data.MAML is described as learning to learn, beyond generalizing to new tasks.Source: §I · p. 2

Contributions

  • With 3 unseen classes and a small training set of 9 samples per class, the MAML-based framework achieves 90.36% average accuracy, versus 64.81% reported for the Siamese Network.Source: §I · p. 2
  • The proposed MAML framework is applied to identify real bearing damage using a model trained with artificial damage, with performance described as comparable to artificial-fault generalization and competitive against few-shot methods.The authors frame this Artificial-to-Real evaluation as a way to explore identification of naturally evolved failures using laboratory data from artificial faults.Source: §I · p. 2

II. PRINCIPLE OF MODEL-AGNOSTIC META-LEARNING

Meta-learning trains a model to adapt rapidly to new tasks using limited task-specific data. MAML does this by updating model parameters within each task and then using post-update losses to adjust the shared initialization.

A. Meta-Learning

  • Meta-learning seeks a strategy for learning quickly on new tasks by transferring knowledge across tasks and rapidly adapting to each one.For fault diagnosis, the goal is to find initial neural-network parameters that yield good performance on a new fault category after only a few gradient steps.Source: A. Meta-Learning · p. 2

B. Meta-Agnostic Meta-Learning

  • MAML is agnostic to network architecture and loss function, and optimizes parameters that adapt quickly through inner- and outer-loop gradient descent.The paper describes its use for both reinforcement-learning objectives and supervised losses such as cross-entropy and mean-squared error.Source: B. Meta-Agnostic Meta-Learning · p. 3
  • In meta-training, MAML updates parameters for each class using training data, evaluates post-update loss on validation data from that class, and aggregates the task losses to update the original parameters.The per-class updated parameter is φ_i; the classification loss is typically cross-entropy, and classes can represent bearing defects such as inner- or outer-race, ball, or cage defects.Source: B. Meta-Agnostic Meta-Learning · p. 3
  • At meta-test time, MAML adapts using a few samples from unseen classes and uses the resulting parameters to predict a test sample from the same class.The initialization is optimized for quick adaptation, rather than optimal performance only on classes directly used for training.Source: B. Meta-Agnostic Meta-Learning · p. 3

C. Meta-Agnostic Meta-Learning with Learnable Inner Loop Learning Rates

The paper addresses the fixed, shared inner-loop learning rate in vanilla MAML, which can affect generalization and convergence and can be costly to tune. It uses a variant that learns rates for each layer and update step through back-propagation.

C. Meta-Agnostic Meta-Learning with Learnable Inner Loop Learning Rates

  • Vanilla MAML uses a fixed inner-loop learning rate shared across classes and update steps, which can affect generalization and convergence and be costly to tune for a dataset.The fixed rate is assigned as α.Source: C. Meta-Agnostic Meta-Learning with… · p. 3
  • The learnable-rate variant learns different inner-loop learning rates for each network layer and update step through back-propagation.The rates form a vector whose elements can decrease during training, which may promote faster convergence and alleviate overfitting.Source: C. Meta-Agnostic Meta-Learning with… · p. 3

III. FEW-SHOT BEARING FAULT DIAGNOSTIC FRAMEWORK BASED ON MAML

The framework trains a MAML initialization across bearing-fault scenarios so it can adapt to new classes with few examples. The described use cases include both new artificial faults and real failures using data from artificial damage.

A. Proposed Few-Shot Bearing Fault Diagnostic Framework

  • Few-shot classification is formulated as an N-way K-shot task, where N counts new classes unseen during meta-training and each class has K training samples.This frames the diagnostic task around adapting to new fault classes with limited examples.Source: A. Proposed Few-Shot Bearing Fault… · p. 3
  • During meta-training, the framework optimizes a shared parameter set along aggregated gradient directions from different bearing-fault scenarios to support quick adaptation to unseen classes.The objective is quick adaptation to new classes rather than optimal performance only on the training classes.Source: A. Proposed Few-Shot Bearing Fault… · p. 3
  • The example trains on outer-race and cage defects and adapts to ball defects and lubrication failure using 5 samples in a 2-way 5-shot setting.The framework is also presented as a way to adapt to classes with limited data and potentially recognize naturally evolved defects using models trained on artificial failures.Source: A. Proposed Few-Shot Bearing Fault… · pp. 3–4

B. Objectives

The objectives are to evaluate MAML-based few-shot bearing diagnosis, including how training-set size and unseen artificial faults affect performance.

B. Objectives

  • The framework is evaluated for few-shot bearing fault diagnosis across specified performance dimensions.The passage introduces the evaluation objectives before listing them.Source: B. Objectives · p. 4
  • The study investigates training-data-size effects and prediction of previously unseen artificial bearing faults.Source: B. Objectives · p. 4

3) New Realistic Bearing Failures with Accelerated Aging:

This section describes the datasets and class selection used to test transfer from artificial bearing damage to real failures. CWRU provides the artificial-fault benchmark, while Paderborn supplies artificial and accelerated-aging real-damage classes.

3) New Realistic Bearing Failures with Accelerated Aging:

  • The CWRU study compares the MAML framework with the Siamese benchmark under a consistent test environment.The paper also identifies accelerated-aging Paderborn data as the basis for testing real failures with artificially damaged bearings.Source: 3) New Realistic Bearing Failures with… · p. 4
  • The Paderborn test rig comprises an electric motor, torque-measurement shaft, bearing module, flywheel, and load motor.Source: Figure 3 · p. 5
  • CWRU classes are based on defect location and size rather than operating speed or load, with each class containing data from three operating conditions.Segments contain 2048 × 2 points sampled at 12 kHz from fan-end and load-end accelerometers; each of 10 classes has 1,980 segments.Source: 3) New Realistic Bearing Failures with… · p. 5
  • Paderborn includes 32 bearings—6 normal, 12 artificially damaged, and 14 with accelerated-aging real damage—with inner- and outer-raceway defects represented.Rolling-element damage was not observed, and data were collected under four operating-condition combinations.Source: 3) New Realistic Bearing Failures with… · p. 5
  • The Paderborn evaluation targets real failures in categories 8–13 using healthy and artificial-fault categories 1–7.The selected Paderborn set contains 13 classes: one healthy, six manually initiated defects, and the remaining real failures from accelerated lifetime tests.Source: 3) New Realistic Bearing Failures with… · p. 5
  • Paderborn defect severity levels correspond to lengths below 2 mm, 2–4.5 mm, and 4.5–13.5 mm, respectively.The thresholds are attributed to.Source: 3) New Realistic Bearing Failures with… · p. 5

C. Model Implementations

The experiments test MAML on few-shot bearing diagnosis, including an ablation of the inner-loop learning rate and optimizer. The learnable rate with Adam performs best in the stated unseen-defect setting.

C. Model Implementations

  • The model uses four convolutional modules with 3 × 3 convolutions and 64 filters, followed by batch normalization, ReLU, and 2 × 2 max-pooling.Vibration signals are represented as 64 × 64 inputs from 4096 samples, and the final layer feeds a softmax; each N-way, K-shot gradient uses NK examples.Source: C. Model Implementations · p. 5
  • Training uses one gradient step, a meta-batch of 25 tasks, and 1,500 iterations; the fixed and learnable inner-loop rates both start at 0.01.Testing uses a meta-batch of one task for both 1-shot and 5-shot settings.Source: C. Model Implementations · p. 5

IV. EXPERIMENTAL RESULTS

  • The experiments compare MAML with the Siamese Network using a test setup intended to remain consistent with the benchmark.The reported study covers CWRU training-size and unseen-fault tests, plus Paderborn prediction of real failures from artificial-damage data.Source: §IV · pp. 5–6

A. Ablation Study

  • At least 3% mean-accuracy margin separates learnable-lr MAML with Adam at lr = 0.01 from the other tested cases in 3-way 5-shot prediction of an unseen outer-race defect.Meta-testing was repeated 10 times; TABLE III presents mean accuracy and standard deviation.Source: 1) Influence of the Inner Loop… · p. 6, Table III · p. 6

2) Influence of Training Data Size:

MAML with a learnable inner-loop rate outperforms the Siamese benchmark across the tested training-set sizes. The experiments use 9 samples per class in later tests to examine unseen fault types and retain benchmark consistency.

2) Influence of Training Data Size:

  • Training sets contain 4, 6, 9, or 12 randomly selected samples per class, and training and testing are repeated five times for each size.The benchmark includes original Siamese results for nine samples per class and results obtained using its open-source code for the other sizes.Source: 2) Influence of Training Data Size · p. 6
  • 20% to 30% improvement in average accuracy is reported for learnable-lr MAML over the Siamese Network in the 3-way few-shot setting.With 5-shot learning and 12 training samples per class, the method reaches 99.77% average accuracy.Source: 2) Influence of Training Data Size · p. 6
  • Nine samples per class are retained for later experiments to study unseen fault types and keep the setup consistent with the Siamese benchmark.The authors state that near-perfect accuracy with 12 samples could overshadow other MAML processes.Source: 2) Influence of Training Data Size · p. 6

B. Predicting New Artificially-Induced Bearing Defects

On the CWRU dataset, MAML improves few-shot classification of new artificial bearing faults over the Siamese benchmark, with stronger gains from learnable inner-loop learning rates. Results also examine unseen defect types and higher-severity faults whose categories were represented during meta-training.

B. Predicting New Artificially-Induced Bearing Defects

  • After 600 training epochs, learnable inner-update lr consistently yields higher and more stable validation accuracy than fixed lr.The authors suggest decreasing learning rates may alleviate overfitting and promote faster convergence.Source: B. Predicting New Artificially-Induced… · pp. 6–7
  • With 9 training samples per class, MAML improves average accuracy over the Siamese Network by 10%–20% with fixed lr and 20%–30% with learnable lr.The samples were randomly selected from 1,980 available per class.Source: B. Predicting New Artificially-Induced… · p. 6, Table V · p. 7
  • For three unseen artificial fault cases, MAML beats the Siamese Network by around 20% with fixed lr and 25% with learnable lr; learnable-lr 5-shot accuracy is around 90% in all three.The cases cover new inner-race and ball failures as well as the unseen outer-race example.Source: Table VI · p. 7, B. Predicting New Artificially-Induced… · p. 7
  • The unseen outer-race example uses classes 8–10, all absent from meta-training, and Table IV evaluates 3-way K-shot classification across different training-sample counts.Source: B. Predicting New Artificially-Induced… · p. 7, Table IV · p. 7
  • When fault categories are represented during meta-training, predicting three higher-severity faults improves average accuracy by around 10% over the cases with completely new fault categories.The authors attribute this to greater similarity between tasks involving the same fault type at different severities than tasks involving different fault types.Source: B. Predicting New Artificially-Induced… · p. 7

C. Predicting New Realistic Bearing Defects

Using healthy and artificially damaged bearings as source data, MAML adapts to real bearing faults in the Paderborn dataset. It achieves over 97% accuracy in 5-shot settings with three or fewer new realistic faults and performs comparably to artificial-fault adaptation and favorably against few-shot benchmarks.

C. Predicting New Realistic Bearing Defects

  • With learnable lr, MAML achieves over 97% accuracy in 5-shot tests for three or fewer new realistic bearing failures.The source data are healthy and artificially damaged bearings; real-fault classes are used for meta-testing.Source: C. Predicting New Realistic Bearing… · p. 8
  • MAML is the best-performing benchmark in the Artificial-to-Real comparison, though not by a large margin, and its standard deviation is 2–20 times lower than those of the benchmarks.MAML’s initial parameters are updated once, compared with 50 epochs for the Feature Transfer and Unfrozen Fine-Tuning Nets and 100 epochs for Relation Net.Source: Table IX · p. 9, C. Predicting New Realistic Bearing… · p. 9
  • MAML’s adaptation capability to real bearing failures delivers results similar to, if not better than, adaptation to artificially induced faults, demonstrating robust generalization across these conditions.Source: C. Predicting New Realistic Bearing… · p. 9
  • The benchmark comparison uses the same Paderborn data, preprocessing, and source and target fault categories as, with 1-, 3-, 5-, and 10-shot tests averaged over 10 trials.Source: C. Predicting New Realistic Bearing… · p. 9

V. CONCLUSION

The paper concludes that MAML can adapt from limited artificial-fault data to new artificial and real bearing fault scenarios. It reports gains over the Siamese benchmark and competitive Artificial-to-Real performance, while identifying unresolved issues in vanilla MAML.

V. CONCLUSION

  • On new artificial bearing faults, MAML’s advantage over the Siamese Network reaches 25% with learnable inner-loop learning rates and a small meta-training set.Source: §V · p. 9
  • On the Paderborn dataset, MAML performs comparably when adapting to real rather than artificial faults and competitively against several few-shot methods.The authors describe these findings as promising prospects for identifying naturally evolved failures using laboratory data from artificially induced faults.Source: §V · p. 9
  • Vanilla MAML may be sensitive to inner-loop hyperparameters, train unstably, and generalize poorly; this work focuses on learning per-step learning rates, with other optimization strategies left for future work.Source: §V · p. 9
Loading 2007.12851v4…