An enhanced sparse autoencoder for machinery interpretable fault diagnosis

Abstract The interpretability of individual components within existing autoencoders remains insufficiently explored. This paper aims to address this gap by delving into the interpretability of the encoding and decoding structures and their correlation with the physical significance of vibrational signals. To achieve this, the Sparse Coding with Multi-layer Decoders (SC-MD) model is proposed, which facilitates fault diagnosis from two perspectives: the working principles of the model itself and the evolving trends of fault features. Specifically, a sparse coding protocol to prevent L1-norm collapse is proposed in the encoding process, regularizing the encoding to ensure that each latent code component possesses variance greater than a fixed threshold on a set of sparse representations given the input data. Subsequently, a multi-layer decoder structure is designed to capture the intricate mapping relationship between features and fault patterns. Finally, the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) is employed as the solver for the SC-MD model, enabling end-to-end updates of all parameters by unfolding FISTA. The coherent theoretical framework ensures the interpretability of SC-MD. Utilizing aeroengine bearing data, we demonstrate the exceptional performance of our proposed approach under both normal conditions and intense noise, as compared to state-of-the-art deep learning methods.