TL;DR
This paper introduces the Domain-Adaptive Prototype-Recalibrated Network (DAPRN), a novel approach for intelligent fault diagnosis in industrial machinery when data is scarce. DAPRN leverages transductive learning and a prototype recalibration strategy to effectively adapt knowledge from source domains to target domains, achieving superior performance compared to existing few-shot learning methods.
What Problem Does This Paper Address?
In modern industrial settings, the accurate and timely diagnosis of equipment faults is crucial for safety and economic reasons. However, collecting sufficient labeled data for training robust fault diagnosis models is often challenging due to the difficulty in simulating or capturing rare fault events, leading to data scarcity. Traditional deep learning and transfer learning methods struggle in these limited data scenarios. Few-shot learning (FSL) offers a promising solution by enabling models to learn from very few examples. Yet, existing FSL methods often assume that source and target domains share similar distributions, which is not always true in real-world industrial applications where operating conditions can vary significantly. Furthermore, metric-based FSL methods can suffer from biased classifiers due to limited target data. This paper addresses the critical need for intelligent fault diagnosis methods that can generalize effectively to unseen fault conditions with minimal labeled data, even when there are significant domain shifts between training and testing data.
Methods
The proposed Domain-Adaptive Prototype-Recalibrated Network (DAPRN) is an end-to-end model designed for few-shot fault diagnosis. It integrates three key components:
- Feature Extractor: A 1D Convolutional Neural Network (CNN) is used to extract robust features from sequential vibration signals. This extractor is jointly optimized to minimize classification loss and maximize domain discriminative loss.
- Domain Discriminator: This component works in conjunction with the feature extractor to learn a domain-invariant representation space. By maximizing the domain-discriminative loss, it encourages the feature extractor to produce representations that are indistinguishable between the source and target domains, thus achieving domain adaptation. A Gradient Reversal Layer (GRL) is employed to facilitate this adversarial training.
- Prototype Recalibration Strategy (PRS): During the meta-testing phase, DAPRN uses a novel PRS to refine the "naïve" prototypes (average representations of support samples). This strategy incorporates pseudo-labeled query samples based on prediction confidence and uses a weighted approach to reduce prototype bias, leading to more accurate classification of unseen target samples.
The training process employs a transductive learning paradigm, meaning both labeled source data and unlabeled target data are used during training. This contrasts with inductive learning, where models are trained solely on labeled source data and then applied to unlabeled target data.
Datasets
The proposed DAPRN was evaluated on two benchmark datasets:
- CWRU Bearing Dataset: This dataset contains vibration signals from bearings under normal and nine different fault conditions (outer race, inner race, and ball faults) with varying damage severities. Experiments were conducted under four different working loads (0, 1, 2, and 3 hp).
- Gearbox Dataset: This dataset comprises vibration signals from a two-stage planetary gearbox with three types of faults (sun gear, planetary gear, and ring gear faults) under three different working loads (50, 150, and 250 Nm).
These datasets were used to simulate various few-shot learning scenarios, including transfer learning (source and target domains share label spaces but differ in operating conditions) and cross-domain few-shot learning (source and target domains have disjoint label spaces).
Key Insights and Ideas
- Domain Adaptation is Crucial for Real-World Scenarios: The paper strongly emphasizes that assuming domain similarity between source and target data is unrealistic in industrial settings. DAPRN's domain adaptation component is shown to be vital for achieving high performance when operating conditions change.
- Prototype Recalibration Mitigates Bias: Naïve prototypes in metric-based FSL can be biased by limited target data. The PRS effectively addresses this by incorporating more information from the query set, leading to more robust classification.
- Transductive Learning Enhances Generalization: By allowing the model to see unlabeled target data during training, DAPRN can learn more effective representations that are better suited for the target domain, even with limited labels.
- Feature Extractor Depth Matters: The experiments on feature extractor depth reveal that while deeper networks generally improve performance, the effectiveness is also tied to the specific task and the ability of the network to capture relevant features.
Key Results
- Superior Performance: DAPRN consistently outperformed seven popular few-shot learning methods across various transfer learning and cross-domain fault diagnosis tasks.
- High Accuracy on Bearing Tasks: DAPRN achieved up to 89.94% accuracy on bearing tasks (e.g., task AB2) and 99.98% on gearbox tasks (e.g., task AG1) in one-shot transfer learning scenarios.
- Excellent Cross-Domain Results: For cross-domain few-shot tasks, DAPRN achieved 100.00% accuracy on bearing task BB3 and 98.65% on gearbox task BG1, significantly outperforming other methods.
- Robustness Across Shots: DAPRN showed minimal performance degradation across different "shot" settings (1, 3, 5, 10 shots), indicating its robustness in few-shot scenarios.
- Ablation Studies Confirm Module Effectiveness: Ablation experiments showed that both the domain discriminator (NoDA) and the prototype recalibration strategy (NoPR) contribute significantly to DAPRN's performance, with the domain discriminator playing a particularly vital role.
Limitations and Future Work
- Offline Predictions: The current DAPRN model is designed for offline predictions, meaning it cannot perform real-time fault diagnosis.
- Requirement for Annotated Data: The method still requires a high-quality annotated related dataset (base set) for effective training.
- Future Work: The authors suggest extending the method to handle more challenging scenarios, such as meta-knowledge generalization across different machines and few-shot fault diagnosis with even more insufficient base datasets. They also aim to enable online predictions.
Summary updated