PaperlayerView exact revision on arXiv

Source-linked AI summary

Applications of Unsupervised Deep Transfer Learning to Intelligent Fault Diagnosis: A Survey and Comparative Study

Zhibin Zhao, Qiyang Zhang, Xiaolei Yu, Chuang Sun, Shibin Wang, Ruqiang Yan, Xuefeng Chen

arXiv:1912.12528v2eess.SPcs.LG

TL;DR

UDTL-based IFD addresses diagnosis when target data are unlabeled and distributions differ from training data, while comparable code and baselines remain limited. The paper surveys and categorizes the field, evaluates representative methods and datasets in a unified framework, and releases code. The study reports dataset- and method-dependent performance patterns and identifies feature transferability and backbone influence as important open issues.

  • Problem

    UDTL-based IFD lacks established open-source code, baseline accuracy, and fair comparative settings despite domain shifts and unlabeled target data.

  • Method

    The paper constructs a task-based taxonomy and review, gathers five public datasets, and evaluates representative methods in a unified test framework with released source code.

  • Results

    CWRU and JNU achieve accuracy above 95%, while joint-distribution assumptions, adversarial training, and dataset-dependent input types show useful performance differences.

  • Takeaways & Limitations

    The framework and comparative results provide basic reusable resources for more comparable and reproducible UDTL-based IFD research.

  • Takeaways & Limitations

    The study identifies feature transferability and backbone influence as insufficiently studied, limiting direct interpretation of transfer performance.

Abstract

from arXiv · show

Recent progress on intelligent fault diagnosis (IFD) has greatly depended on deep representation learning and plenty of labeled data. However, machines often operate with various working conditions or the target task has different distributions with the collected data used for training (the domain shift problem). Besides, the newly collected test data in the target domain are usually unlabeled, leading to unsupervised deep transfer learning based (UDTL-based) IFD problem. Although it has achieved huge development, a standard and open source code framework as well as a comparative study for UDTL-based IFD are not yet established. In this paper, we construct a new taxonomy and perform a comprehensive review of UDTL-based IFD according to different tasks. Comparative analysis of some typical methods and datasets reveals some open and essential issues in UDTL-based IFD which are rarely studied, including transferability of features, influence of backbones, negative transfer, physical priors, etc. To emphasize the importance and reproducibility of UDTL-based IFD, the whole test framework will be released to the research community to facilitate future research. In summary, the released framework and comparative study can serve as an extended interface and basic results to carry out new studies on UDTL-based IFD. The code framework is available at \url{https://github.com/ZhaoZhibin/UDTL}.

I. INTRODUCTION

UDTL-based IFD addresses domain shifts and unlabeled target data that limit conventional deep-learning diagnosis. This paper surveys and categorizes the field, establishes comparative resources, and formalizes the transfer setting.

  • Motivation: Conventional deep-learning IFD assumes matching source and target distributions and plentiful labeled training data, assumptions often violated in real deployments.Simulation or fault-seeding data may not strictly match real test data, weakening generalization.
  • Motivation: UDTL-based IFD transfers knowledge from labeled source data to classify unlabeled target-domain samples under domain shift.The paper treats UDTL and unsupervised domain adaptation as equivalent terms for its purposes.
  • Comparative framework: The framework and source code support reproducible comparisons by providing common settings, datasets, parameter choices, and baseline results for UDTL-based IFD.The study addresses the field’s shortage of open-source code and comparable baseline accuracy.
  • Scope and taxonomy: The paper establishes a task-based taxonomy and comprehensive review of UDTL-based IFD, organizing methods by domains, target-data usage, labels, label-set relations, and methodological level.The taxonomy distinguishes single-domain and multi-domain UDTL from a macro perspective.
  • Comparative framework: It collects publicly available datasets, discusses their adaptability and data splitting, and integrates typical methods into a unified framework tested on five datasets.The authors argue that splitting data into training and test sets is appropriate regardless of source or target domain.
  • Formal setting: UDTL is defined as minimizing target risk for unlabeled target samples using supervision from labeled source data.The source and target domains follow distributions P and Q, respectively, and the model predicts target labels.

1) Single-domain UDTL:

Single-domain UDTL addresses domain shifts across working conditions, fault types, locations, or machines by transferring models through shared features. Its taxonomy distinguishes label relationships, target-data usage, and methodological strategies, while fair comparisons require a common backbone.

  • 1) Single-domain UDTL:: Label-consistent UDTL uses matching source and target label spaces, whereas label-inconsistent transfer includes partial, open set, and universal tasks.Partial transfer uses a target label subset; open set transfer includes unknown target labels; universal transfer combines both conditions.
  • 1) Single-domain UDTL:: UDTL distinguishes multi-domain adaptation, where unlabeled target samples join training, from domain generalization, where they do not.
  • 1) Single-domain UDTL:: Single-domain UDTL targets distribution differences caused by working conditions, fault sizes, fault types, locations, and machines.
  • 1) Single-domain UDTL:: The backbone performs feature extraction and can substantially affect test accuracy, making comparisons across studies with different backbones unfair.The framework therefore evaluates different UDTL methods using the same CNN backbone for fair comparison.
  • 1) Single-domain UDTL:: Methodologically, label-consistent UDTL comprises network-based, instance-based, mapping-based, and adversarial-based approaches.

2) Applications to IFD:

Mapping-based UDTL methods transfer fault-diagnosis features by measuring or aligning source–target distributions. The study represents this family with MK-MMD, JMMD, and CORAL implementations.

  • Instance-based UDTL: Fine-tuning strategies transfer pretrained networks using limited target data, while AdaBN adapts batch-normalization statistics without target labels.AdaBN replaces source-domain batch statistics with target-domain statistics during testing and can update them batchwise or over all target data.
  • Mapping-based UDTL: MMD measures squared distance between source and target marginal-distribution embeddings in a reproducing kernel Hilbert space.MK-MMD combines multiple kernels to improve two-sample testing and mapping estimation.
  • Mapping-based UDTL: The framework implements MK-MMD by adding a multi-kernel discrepancy loss to the objective for source–target feature alignment.It uses five Gaussian kernels, with bandwidths set by median pairwise distances on training data.
  • Mapping-based UDTL: JMMD addresses joint-distribution shifts by aligning source and target feature-label distributions through a tensor-product Hilbert-space mapping.The implemented model adds JMMD to the loss, with λJMMD controlling its contribution.
  • Mapping-based UDTL: CORAL aligns second-order source and target statistics by matching their covariance matrices through an added CORAL loss.The implementation uses the Frobenius norm, with λCORAL as a trade-off parameter.

2) Applications to IFD:

Adversarial UDTL reduces source–target feature-distribution discrepancies with a domain discriminator. DANN performs unconditional alignment, whereas CDAN conditions adversarial alignment to address more complex joint distributions.

  • Adversarial-based UDTL: Adversarial UDTL uses a domain discriminator and feature extractor to reduce discrepancies between source and target feature distributions.The study represents this family with DANN and CDAN.
  • DANN: DANN jointly trains a feature extractor, domain discriminator, and class predictor so extracted features hinder domain discrimination.Its objective minimizes prediction loss while maximizing adversarial loss, implemented with a gradient reversal layer.
  • DANN: DANN is limited because it cannot capture complex multi-modal structures and makes safe conditioning of the domain discriminator difficult.These limitations motivate conditional adversarial adaptation.
  • CDAN: CDAN uses a multilinear map and cross-covariance representation to model joint distributions and address source–target joint-distribution differences.The framework embeds conditional adversarial training into its loss, with entropy-aware weighting that lowers weights for hard-to-transfer samples.

2) Applications to IFD:

Label-inconsistent UDTL handles cases where source and target fault-label sets differ. The study examines partial transfer, open-set transfer, and universal transfer, including a partial-adaptation model that down-weights source-private classes.

  • Label-inconsistent UDTL: Real applications motivate label-inconsistent UDTL because source and target label sets may not be consistent.The study examines partial, open-set, and universal UDTL settings.
  • Universal UDTL: Universal transfer combines partial and open-set conditions, requiring adaptation when shared, source-private, and target-private classes coexist.The supplied discussion identifies universal transfer as the simultaneous occurrence of the first two cases.
  • Partial UDTL: Partial UDTL assumes the target label set is a subset of the source label set, Ct ⊂ Cs.PADA estimates source-class contributions from average target predictions and reduces source-private-class influence.
  • Partial UDTL: PADA applies normalized class-level weights to classifier and discriminator losses, reducing contributions from source samples belonging to source-private classes.The resulting model adds the weighted partial-adaptation loss to the objective.

3) Applications to IFD:

Open-set UDTL addresses target domains containing classes absent from the source domain. OSBP adds an unknown class and adversarially constructs a boundary separating known from unknown target samples.

  • Open-set UDTL: Open-set domain adaptation assumes the source label set is a subset of the target label set, Cs ⊂ Ct.Its goal is to classify known target classes and recognize source-absent target classes as unknown.
  • OSBP: OSBP uses a feature extractor and a C + 1 classifier whose extra output dimension represents the unknown class.Softmax converts classifier outputs into known- and unknown-class probabilities.
  • OSBP: OSBP trains the classifier to assign target samples an unknown-class probability τ while training the feature extractor adversarially around that threshold.A binary cross-entropy loss supports construction of the known–unknown decision boundary.
  • OSBP: The framework implements OSBP by introducing the C + 1 classifier and adversarial loss to construct a pseudo decision boundary for unknown classes.The resulting saddle-point problem is solved through min-max optimization.

3) Applications to IFD:

Universal domain adaptation addresses target domains whose labels may include both shared and private classes. UAN uses instance-level transferability to align shared classes while identifying unknown target samples.

  • Universal domain adaptation allows source and target label sets to contain common and private classes.
  • UAN combines domain similarity and prediction uncertainty to define instance-level transferability.
  • UAN adds a non-adversarial domain discriminator to its DANN-like model.
  • Training UAN maximally aligns source and target distributions in the shared label set and reduces the category gap.
  • During testing, target samples exceeding threshold ω0 are assigned to the unknown class; otherwise, label prediction is used.

3) Applications to IFD:

Multi-domain UDTL uses multiple source domains to learn domain-invariant features through adaptation or generalization. The reviewed methods primarily reduce domain gaps with adversarial training or select similar source domains.

  • Multi-domain UDTL: Multi-domain UDTL includes multi-domain adaptation using target data and domain generalization without target data during training.
  • Multi-domain adaptation: Traditional single-source UDTL may fail to exploit relationships and domain-invariant features available across multiple source domains.
  • Multi-domain adaptation: Multi-domain adaptation seeks domain-invariant features or identifies the source domain most similar to the target domain.
  • Multi-domain adaptation: Adversarial training confuses the domain discriminator, reducing feature gaps through a minimax objective.
  • Related methods: Reviewed multi-domain methods include adversarial, feature-and-task alignment, classifier-alignment, and joint-matching approaches.

1) Basic concepts:

Domain generalization learns from multiple source domains and evaluates on an unseen target domain without target samples during training. The comparative study spans several open-source bearing and gearbox datasets with varied operating conditions and transfer settings.

  • Domain generalization: Domain generalization learns shared knowledge from multiple source domains and excludes unlabeled target samples from training.
  • Domain generalization: DG performance depends on features that are both discriminative and domain-invariant.
  • Domain generalization: IAN applies DANN-based adversarial training to help the feature extractor learn domain-invariant representations.
  • Datasets: The study evaluates UDTL methods on five datasets, including CWRU, PU, JNU, PHM2009, and SEU.
  • Datasets: CWRU defines twelve transfer tasks across four motor-load operating conditions.
  • Datasets: PU contains artificially induced and real bearing damages across four operating conditions, with twelve transfer settings.
  • Datasets: PHM2009 experiments use four shaft speeds under high load, producing twelve transfer-learning settings.

5) Southeast University (SEU) dataset:

The SEU evaluation uses two gearbox sub-datasets and two speed-load conditions, while the framework controls preprocessing, splitting, training, and accuracy evaluation. Across datasets, performance varies substantially and adversarial methods generally outperform simpler alignment baselines.

  • SEU dataset: SEU contains bearing and gear sub-datasets with five conditions each, using eight channels collected from a drivetrain simulator.
  • SEU dataset: SEU defines two transfer tasks from 20 Hz–0 V and 30 Hz–2 V speed-load configurations.
  • Preprocessing and splitting: The framework separates 80% training samples from 20% test samples to prevent overlap between target alignment data and test data.
  • Preprocessing and splitting: Models use either direct time-domain signals or frequency-domain inputs, with Z-score normalization applied to inputs.
  • Training and evaluation: Methods are implemented in PyTorch, trained for 300 epochs with Adam, batch size 64, and progressive trade-off weighting.
  • Training and evaluation: Overall accuracy is averaged and maximized across five runs to evaluate model performance.
  • Results: CWRU and JNU exceed 95% accuracy, whereas the other datasets reach only around 60%; these values are lower bounds.
  • Results: JMMD generally outperforms MK-MMD, while DANN and CDAN generally outperform MK-MMD, supporting joint-distribution and adversarial alignment.

4) Results of input types:

Across datasets, input-domain preferences vary, while backbone choice affects the observed accuracy gap. Evaluation fluctuations and test leakage make Last-based accuracy more suitable for judging generalization when instability is large.

  • Time-domain input performs better on CWRU, JNU, and SEU, whereas frequency-domain input performs better on PU and PHM2009.
  • Accuracy gaps between input types can be large, so neither domain can be declared universally superior independently of the backbone.
  • The study therefore recommends testing both input types and prioritizing stability, robustness, and Last-based accuracy when Best–Last fluctuation is large.
  • Different experiments sometimes show large fluctuations, especially on datasets with lower overall accuracy, indicating limited algorithmic stability and robustness.
  • Best-based evaluation uses the test set to select the best model, creating test leakage; Last may better represent generalization accuracy.

2) Dataset settings:

The experiments evaluate partial, open set, and universal transfer settings on CWRU using multiple accuracy summaries and TOPSIS. Results are generally favorable on many tasks, but later-epoch degradation, negative transfer, and unknown-class errors remain important limitations.

  • Dataset settings: CWRU tasks are constructed with randomly selected classes to evaluate partial, open set, and universal transfer learning across different label sets.
  • Results of partial UDTL: PADA performs well on most tasks, but Last-Mean is lower than Best-Mean for tasks 3-1, 2-3, and 3-2.
  • Results of open set UDTL: OSBP performs relatively well on most tasks, but later-stage performance degrades, especially for unknown-class accuracy.
  • Results of open set UDTL: For OSBP, the lowest shared-class accuracy is about 50%, showing that unknown-class detection does not ensure accurate shared-class classification.
  • Results of universal UDTL: UAN achieves excellent performance on CWRU, while later-stage degradation, wrong feature alignment, and insufficient shared-class classification remain concerns.
  • Evaluation: TOPSIS is used to combine five evaluation metrics into a final score for open set and universal transfer learning.

2) Results of multi-domain adaptation:

Multi-domain UDTL improves CWRU accuracy over direct transfer, while backbone choice substantially changes results across datasets and input types. Bottleneck removal has little consistent effect, but freezing backbone parameters sharply reduces target-domain accuracy.

  • Multi-domain adaptation: MS-UADA consistently improves CWRU accuracy over Basis for direct transfer from multiple source domains.Time-domain input is slightly better than frequency-domain input, but the overall difference is very small.
  • Multi-domain adaptation: IAN performs similarly to Basis on most CWRU tasks but decreases greatly on task 012-3 F.The paper attributes this result to IAN’s focus on domain-invariant features without target-domain information.
  • Influence of backbones and bottleneck: Different backbones make published UDTL-based IFD results difficult to compare directly and can strongly influence accuracy.The paper recommends selecting a strong, suitable backbone and keeping it fixed across comparisons.
  • Influence of backbones and bottleneck: ResNet18 significantly improves every PHM2009 algorithm’s accuracy but reduces SEU accuracy with time-domain input.With frequency-domain input on SEU, ResNet18 significantly improves accuracy, showing dataset- and input-dependent backbone behavior.
  • Influence of backbones and bottleneck: Removing the bottleneck layer produces no consistent advantage for either time- or frequency-domain CWRU inputs.The comparison considers original models, models without the bottleneck, and models with frozen pretrained backbones.
  • Influence of backbones and bottleneck: Freezing backbone parameters during transfer learning dramatically decreases target-domain accuracy.Only bottleneck-layer parameters are updated in this setting, indicating that source-trained backbones cannot be transferred directly to the target domain.

C. Negative transfer

The study examines whether transfer tasks produce negative transfer and identifies broader design issues affecting UDTL-based IFD. Results emphasize that transfer suitability, shared features, label consistency, and physical priors require careful attention.

  • Negative transfer: Transferring across damage-generation methods on the PU dataset produced negative transfer for every method with both time- and frequency-domain inputs.The authors therefore question whether this constructed task is suitable for transfer learning.
  • Negative transfer: Transfer tasks should first be checked for suitability and for shared features between the source and target domains.The paper links unsuitable transfer settings to unreliable transfer results.
  • Physical priors: Published UDTL-based IFD studies often use methods from other fields while paying less attention to physical phenomena and rules in the monitored systems.The authors call for incorporating physical prior knowledge into more targeted methods.
  • Label-inconsistent transfer: Label-inconsistent transfer methods may overfit: acceptable best average accuracy can coexist with a large drop in last average accuracy.The suggested explanation is that models fail to focus effectively on shared classes, producing poor domain alignment.
  • Multiple source domains: Most studies use a single source domain, but multiple real-world source domains may have different distributions and require balancing their contributions.The paper identifies multi-domain adaptation and domain generalization as common alignment strategies, while noting that source weighting remains unresolved.
  • Data limitations: Class imbalance remains a natural concern because most machines operate normally, leaving fault data scarce in some conditions.The paper highlights imbalanced learning or few-shot learning combined with transfer learning as possible directions.

Appendix A. Evaluation Results

The appendix reports evaluation results for UDTL methods across PU, JNU, PHM2009, CWRU, and PU-Types settings. The tables expose substantial variation across datasets, tasks, methods, input domains, and evaluation statistics.

  • CWRU PADA: For CWRU PADA, time- and frequency-domain inputs both achieve several 100.00% best accuracies, but some last mean accuracies are substantially lower.For task 0-3, the time-domain input falls to 80.12% last mean accuracy and the frequency-domain input to 79.40%.
  • PU-Types: PU-Types frequency-domain results are generally much lower than the strongest PU and JNU results, with best mean accuracies reaching only 68.80% in the shown tasks.This setting is associated with negative transfer in the paper's discussion.
Loading 1912.12528v2…