No abstract

No abstract

Deep learning (DL) is progressively popular as a viable alternative to traditional signal processing (SP) based methods for fault diagnosis. However, the lack of explainability makes DL-based fault diagnosis methods difficult to be trusted and understood by industrial users. In addition, the extraction of weak fault features from signals with heavy noise is imperative in industrial applications. To address these limitations, inspired by the Filterbank-Feature-Decision methodology, we propose a new Signal Processing Informed Neural Network (SPINN) framework by embedding SP knowledge into the DL model. As one of the practical implementations for SPINN, a denoising fault-aware wavelet network (DFAWNet) is developed, which consists of fused wavelet convolution (FWConv), dynamic hard thresholding (DHT), index-based soft filtering (ISF), and a classifier. Taking advantage of wavelet transform, FWConv extracts multiscale features while learning wavelet scales and selecting important wavelet bases automatically; DHT dynamically eliminates noise-related components via point-wise hard thresholding; inspired by index-based filtering, ISF optimizes and selects optimal filters for diagnostic feature extraction. It's worth noting that SPINN may be readily applied to different deep learning networks by simply adding filterbank and feature modules in front. Experiments results demonstrate a significant diagnostic performance improvement over other explainable or denoising deep learning networks. The corresponding code is available at https://github.com/albertszg/DFAWnet.

Introduction: The goal of the 2021 PhysioNet/CinC challenge is diagnosing cardiac abnormalities from electrocardiograms (ECGs) and evaluating the diagnostic potential of reduced-lead ECGs. We describe the whole model created by the team "AI Healthcare" for this goal.Methods: ECGs were downsampled to 300 Hz and filtered by wavelet. Then ECGs we randomly clipped or zeropadded to 4,096 samples. To have a better representative learning ability, a modified ResNet with larger kernel sizes was used. Multi-source adversarial feature learning was used to learn domain-invariant and discriminative representations with a special gradient reversal layer (GRL). The performance with and without the domain generation methods was compared.Results: We achieved a challenge score of 0.66, 0.64, 0.65, 0.65, 0.62 on the validation data. We ranked 8th, 7th, 6th, 6th, and 12th for 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead ECGs, respectively. Testing showed that domain generation improved metric scores on the unseen domain.Conclusion: Generalized representations perform well for "unseen" data. It is a general method for other models to improve generalization performance by learning a domain-invariant feature representation.