Source-linked AI summary
A Novel GAN-based Fault Diagnosis Approach for Imbalanced Industrial Time Series
Wenqian Jiang, Cheng Cheng, Beitong Zhou, Guijun Ma, Ye Yuan
TL;DR
Industrial fault diagnosis must handle imbalanced time series in which normal samples greatly outnumber failure cases, while deep learning also faces large data and preprocessing demands. The paper combines a feature extractor with a GAN using an encoder-decoder-encoder design trained only on normal samples. Tests on rolling-bearing data from CWRU and laboratory data report excellent fault-detection performance, with abnormal samples producing larger evaluation scores.
Problem
Industrial fault diagnosis involves imbalanced time series with far fewer fault samples, while deep learning requires large labeled datasets and preprocessing of huge inputs.
Method
The approach combines a feature extractor with a GAN whose encoder-decoder-encoder generator learns normal-sample patterns and latent representations using only normal training samples.
Results
Tests on CWRU rolling-bearing data and laboratory datasets report excellent fault-detection performance, with fault samples producing larger evaluation scores than normal samples.
Takeaways & Limitations
The approach provides a fault-diagnosis architecture for imbalanced industrial time series and remains almost stable across subsample sizes while achieving higher accuracy than BiGAN.
Abstract
from arXiv · showhide
This paper proposes a novel fault diagnosis approach based on generative adversarial networks (GAN) for imbalanced industrial time series where normal samples are much larger than failure cases. We combine a well-designed feature extractor with GAN to help train the whole network. Aimed at obtaining data distribution and hidden pattern in both original distinguishing features and latent space, the encoder-decoder-encoder three-sub-network is employed in GAN, based on Deep Convolution Generative Adversarial Networks (DCGAN) but without Tanh activation layer and only trained on normal samples. In order to verify the validity and feasibility of our approach, we test it on rolling bearing data from Case Western Reserve University and further verify it on data collected from our laboratory. The results show that our proposed approach can achieve excellent performance in detecting faulty by outputting much larger evaluation scores.
I. INTRODUCTION
Industrial fault diagnosis uses monitored measurements to identify damage or impending damage, but time-series deep learning requires large labeled datasets and effective preprocessing. This paper addresses imbalanced industrial data with a GAN-based approach that combines feature extraction with training on normal samples.
- Industrial fault diagnosis analyzes sensor measurements and machine-status information to detect damage or impending damage.
- Deep learning for industrial time-series anomalies faces the need for large labeled datasets and preprocessing of huge time-series inputs.
- The proposed method combines a feature extractor with GAN to address imbalanced industrial data and preprocess large time-series inputs.
- For univariate industrial time series, the method uses a feature extractor to represent period-specific characteristics while reducing dimensionality and computing time.
- The approach is tested first on rolling-bearing data from Case Western Reserve University and then validated on laboratory datasets.
II. RELATED WORKS
Prior work includes handcrafted feature-based models and deep learning approaches for industrial time-series fault diagnosis. Recent research increasingly uses adversarial training, including GAN architectures combined with recurrent and encoder-decoder designs.
- Feature-based models: Feature-based models extract time, frequency, or time-frequency features before applying classifiers such as PCA, SVM, or random forest.
- Deep learning models: Deep learning models, including LSTM, recurrent, convolutional, and autoencoder networks, have been developed for time-series anomaly detection.
- Adversarial methods: GAN-based anomaly detection has been explored for multivariate time series by combining adversarial training with LSTM-RNNs and data augmentation.
- Adversarial methods: A prior semi-supervised GAN used encoder-decoder-encoder subnetworks and three generator losses to capture distinguishing features in input and latent spaces.
III. OUR APPROACH
The approach combines a feature extractor with an encoder-decoder-encoder GAN trained only on normal industrial time series. It detects faults through reconstruction and latent-space deviations, using adversarial and representation losses to learn normal patterns.
- Framework: The framework comprises a feature extractor, generator, and discriminator for processing univariate industrial time series.The feature extractor reduces the dimensionality of long samples before GAN processing.
- Generator: The generator uses two encoders and a decoder to learn both input-sample representations and latent-space patterns while reconstructing data.The first encoder produces a latent representation, the decoder reconstructs the input, and the second encoder represents the generated sample.
- Training and detection: Training uses only normal samples so the generator learns normal patterns; fault inputs then produce deviations from the original sample and latent representation.These deviations provide the basis for identifying faulty samples during testing.
- Losses: The generator combines fraud, apparent, and latent losses, while the discriminator uses feature matching loss for adversarial learning.Fraud loss encourages generated samples to be judged real, apparent loss measures input-to-reconstruction L1 distance, and latent loss aligns real and generated representations.
- Anomaly scoring: During testing, latent loss and apparent loss are combined to score the abnormality of each subsample.The anomaly score is based on the model’s reconstruction and latent-space discrepancies.
IV. EXPERIMENTAL SETUP
The experiments evaluate the approach on rolling-bearing datasets from CWRU and the authors’ laboratory, using normal-only training and fault-inclusive testing.
- The study evaluates feasibility first on rolling-bearing data from Case Western Reserve University and then validates the approach on laboratory-collected data.
- The CWRU dataset contains accelerometer-measured vibration signals from motor bearings with seeded faults at the inner raceway, rolling element, and outer raceway.Introduced fault diameters range from 0.007 to 0.040 inches.
- The laboratory dataset records 50 Hz Hall-sensor voltage signals under normal and three fault-location conditions using bearings with a 14-mil fault diameter.
- Normal samples are split into 80% training and 20% testing, while fault samples are included only during testing.
- Both datasets are trained for 20 epochs using PyTorch and Adam with an initial learning rate of 0.001 and momentums 0.5 and 0.999.
V. RESULTS
The proposed model distinguishes normal from faulty samples on CWRU and laboratory rolling-bearing data using evaluation scores, reconstruction comparisons, and latent representations. It also remains stable across subsample sizes and shows dataset-specific sensitivity to latent-vector size.
- CWRU results: High evaluation scores indicate abnormal samples in the CWRU binary-classification results.The scores also help classify different fault types when the weighted factor in the general loss is adjusted.
- CWRU results: Fault samples show substantially larger differences than normal samples when raw data are compared with reconstructed data and latent representations.The paper presents these comparisons as an intuitive explanation for the model’s fault-detection effectiveness.
- Laboratory results: The model achieves excellent fault-diagnosis performance on the laboratory rolling-bearing dataset.This result is reported after the initial evaluation on CWRU.
- Hyperparameter analysis: An optimal subsample length of 12000 is reported for both datasets, while a latent representation size of 64 gives the highest accuracy for the laboratory dataset.The paper associates the laboratory sampling frequency with a potential pattern interval of four minutes.
- Hyperparameter analysis: Latent-representation size does not affect final accuracy on the CWRU dataset.This dataset-specific result contrasts with the reported optimum for the laboratory data.
- Comparison with BiGAN: The proposed algorithm is nearly stable across subsample sizes and achieves higher accuracy than BiGAN.The comparison is performed using different subsample sizes.
VI. CONCLUSION
The paper presents a GAN-based architecture for imbalanced industrial time series that trains on normal samples and detects abnormalities through large deviation scores. It also identifies feature extraction for different signal modes and multivariate time series as directions for future work.
- The architecture trains only on normal samples to diagnose faults in imbalanced industrial time series.
- A dataset-specific feature extractor precedes the fault detector, while an encoder-decoder-encoder generator models normal latent and reconstructed patterns.
- Abnormal samples are detected through large deviation scores from learned normal patterns.
- Future work should improve feature extractors for different recorded signals and incorporate cross-dimensional information from multivariate time series.