Source-linked AI summary
Pre-Processing-Free Gear Fault Diagnosis Using Small Datasets with Deep Convolutional Neural Network-Based Transfer Learning
Pei Cao, Shengli Zhang, Jiong Tang
TL;DR
Gearbox fault diagnosis needs adaptive methods that work with limited training data, because conventional feature engineering is application-specific while deep networks typically require substantial datasets. The paper develops transfer learning with a deep CNN to classify raw accelerometer data without preprocessing. The approach outperformed locally trained CNN and AFS-SVM methods, with reported accuracy gains of up to 50%.
Problem
Gearbox diagnosis requires feature extraction and classification, but conventional methods are application-specific and deep neural networks generally require substantial training data.
Method
The paper transfers pretrained deep CNN stages for adaptive feature extraction and trains a task-specific classification stage using raw accelerometer data without preprocessing.
Results
Accuracy outperformed locally trained CNN and angle-frequency analysis-based SVM by as much as 50%.
Takeaways & Limitations
The outcome indicates potential applicability of the proposed approach to other fault diagnosis practices.
Abstract
from arXiv · showhide
Early fault diagnosis in complex mechanical systems such as gearbox has always been a great challenge, even with the recent development in deep neural networks. The performance of a classic fault diagnosis system predominantly depends on the features extracted and the classifier subsequently applied. Although a large number of attempts have been made regarding feature extraction techniques, the methods require great human involvements are heavily depend on domain expertise and may thus be non-representative and biased from application to application. On the other hand, while the deep neural networks based approaches feature adaptive feature extractions and inherent classifications, they usually require a substantial set of training data and thus hinder their usage for engineering applications with limited training data such as gearbox fault diagnosis. This paper develops a deep convolutional neural network-based transfer learning approach that not only entertains pre-processing free adaptive feature extractions, but also requires only a small set of training data. The proposed approach performs gear fault diagnosis using pre-processing free raw accelerometer data and experiments with various sizes of training data were conducted. The superiority of the proposed approach is revealed by comparing the performance with other methods such as locally trained convolution neural network and angle-frequency analysis based support vector machine. The achieved accuracy indicates that the proposed approach is not only viable and robust, but also has the potential to be readily applicable to other fault diagnosis practices.
1. Introduction
Gearbox fault diagnosis commonly uses measured vibration signals, extracted fault-related features, and a classifier, but conventional methods are application- and condition-specific. Deep neural networks offer adaptive extraction and classification, yet their data demands remain problematic for mechanical systems with limited samples.
- Signal-based fault diagnosis measures vibration signals, extracts fault-related features, and classifies fault occurrence by type and severity.
- Conventional diagnosis methods are considerably specific to applications and working conditions.
- Deep neural networks are investigated for adaptive feature extraction and fault classification with minimal tuning.
- Deep neural-network approaches require substantial training data, limiting their use in mechanical-system applications where data are insufficient.
2 Approach Formulations
The approach formulates CNN feature extraction and transfer learning for small-data fault diagnosis, using transferred convolutional stages and task-specific classification layers. Its architecture combines a pretrained deep CNN with newly trained layers, optimized using regularized cross-entropy and stochastic gradient descent.
- 2.1 Convolutional Neural Network (CNN): CNNs combine convolutional, ReLU, pooling, fully connected, and softmax layers to extract features and classify inputs.The example architecture accepts image input, transforms features nonlinearly, reduces data size, and produces predefined labels.
- 2.1 Convolutional Neural Network (CNN): Convolution filters slide across inputs to form feature maps, with multiple filters learning different feature perspectives from training data.The formulation assumes stride 1 and zero padding for the illustrated two-dimensional operation.
- 2.1 Convolutional Neural Network (CNN): ReLU models nonlinear relationships after convolution, while max pooling selects regional maxima to reduce parameter size and sensitivity to feature location.The pooling operation downsamples an activation-map sub-region to its maximum value.
- 2.2 Transfer Learning: Deeper CNNs can improve performance but require more trainable parameters and substantial valid training samples, which mechanical applications often lack.
- 2.2 Transfer Learning: Transfer learning copies the first n layers from a pretrained network and trains the remaining m-n layers with data from the novel task.The previous-task and novel-task datasets are represented using inputs X and output labels L.
- 2.2 Transfer Learning: Transferred convolutional, ReLU, and pooling stages provide general input features, whereas fully connected and classification stages are more task-specific.
- 2.3 Proposed Architecture: The proposed architecture uses a pretrained deep CNN base with five convolution stages and three fully connected stages, transferring parameters for its first 21 of 24 layers.
- 2.3 Proposed Architecture: Training minimizes cross-entropy with an l2 normalization term and updates parameters through stochastic gradient descent.The normalization term is described as preventing over-fitting; the update uses a learning rate and prior gradient contribution.
3. Experimental Studies
Experiments on a replaceable-gear two-stage gearbox compare transfer learning with local CNN and AFS-SVM across training-set sizes and sampling rates. Transfer learning consistently performs best, including with raw signals, limited data, and down sampling.
- Experimental setup: The study uses a two-stage gearbox with replaceable gears and records vibration signals from multiple gear health conditions.Each condition provides 104 signals, with 3,600 angle-even samples over four gear revolutions and a 900-point down-sampled case.
- Compared methods: The proposed transfer learning approach is compared with a locally trained CNN and an angle-frequency analysis-based support vector machine.Training data are randomly selected across scenarios ranging from 80% to 2% of the available signals.
- Case 1: 3,600 sampling points: 94.90% accuracy is achieved with only five training signals per condition, increasing to 99%-100% when 10% or more training data are used.The transfer learning approach is the best performer for all evaluated training-data sizes in the 3,600-point case.
- Case 1: 3,600 sampling points: 27.99% to 97.57% is the local CNN accuracy range as training data increase, while AFS-SVM reaches a plateau after 20% training data.The results indicate that training-set size strongly affects local CNN performance, whereas transfer learning maintains high accuracy with less data.
- Feature extraction: The transferred convolutional layers produce increasingly abstract feature maps across five convolution layers when processing gearbox vibration signals.The case study supports transferability of the base architecture to raw time-domain signal inputs.
- Case 2: 900 sampling points: 60.11% transfer-learning accuracy at the smallest reported training size reaches 95.88% with 20% training data after down sampling.At that 20% setting, local CNN and AFS-SVM achieve 43.56% and 70.07%, respectively; lowering the sampling rate deteriorates overall performance.
4. Concluding Remarks
The paper develops a deep CNN transfer-learning approach for gear fault diagnosis using raw accelerometer data and small training sets. Its reported accuracies outperform local CNN and AFS-SVM by as much as 50%, supporting potential applicability to other fault-diagnosis practices.
- Approach: The approach combines a pre-trained neural-network portion for automatic feature extraction with a fully connected classification stage trained on experiment data.This architecture provides adaptive feature extraction without preprocessing while limiting task-specific training to the classification stage.
- Conclusion: The method performs gear fault diagnosis directly from raw accelerometer data and requires only a small training set compared with locally trained CNNs.The conclusion frames preprocessing-free operation and reduced data requirements as central properties of the proposed approach.
- Conclusion: Up to 50% higher classification accuracy is reported than for locally trained CNN and angle-frequency analysis-based SVM methods.The reported outcome is based on experimental gear fault-diagnosis studies.
- Conclusion: The achieved outcome indicates potential applicability beyond gear diagnosis to other fault-diagnosis practices.The paper characterizes the approach as viable and robust within the reported study scope.