Uncertainty Quantification Fault Diagnosis Method Based on Multisignal Domain Interpretable Neural Network

With the introduction of uncertainty estimation methods in deep learning, artificial intelligence (AI) has improved the reliability in various fault diagnosis and risk decision-making applications. However, the training process of these models is often treated as a black box, leading to a lack of interpretability in fault diagnosis and making the physical meaning of the models unclear. To address this issue, we propose an uncertainty quantification fault diagnosis method based on a multisignal domain interpretable neural network, termed local and global BayesCNN (LGBayesCNN). This framework uses BayesCNN as the backbone network and replaces traditional convolutional kernels with fractional Fourier transform (FrFT) and short-time Fourier transform (STFT), ensuring that the model has a clear physical meaning and enabling feature extraction and analysis from both global and local perspectives. During the training process, the network adaptively adjusts the FrFT order and the frequency of the window function in STFT, allowing it to reveal the signal features that the network focuses on in the fractional-frequency and time-frequency domain, further enhancing the interpretability of the network. Moreover, the joint cross-attention mechanism is introduced to fuse these two types of features for achieving feature complementarity and enhancement. In addition, the uncertainty-aware capability of BayesCNN enables the detection of unknown faults and the estimation and quantification of uncertainty in diagnostic results. We conducted experiments using a bearing fault diagnosis dataset from Southeast University (SEU) and Huazhong University of Science and Technology. The results show that LGBayesCNN achieves the highest fault diagnosis accuracy of 99.81% and 100%, demonstrating its effectiveness and reliability.