A Robust Capsule Network With Adaptive Fusion of Multiorder Proximity for Intelligent Decoupling of Compound Fault

With the advancement of sensor acquisition technology and deep learning algorithms, intelligent fault diagnosis based on equipment operation data has achieved significant progress in the industrial field. However, existing deep learning methods are only aimed at recognizing a single fault, ignoring the concurrency and coupling of various types of faults in industrial scenarios. The presence of compound faults leads to an exponential increase in the number of original fault modes, posing a major challenge in fault diagnosis. To solve this issue, this paper proposes a zero-shot compound fault intelligent decoupling method based on capsule network under the framework of adaptive fusion of multi-order proximity and generalized sparse norm. Firstly, capsule network with the ability to be sensitive to spatial features is utilized to build an intelligent decoupling model. Subsequently, a dynamic routing scheme with adaptive fusion of multi-order proximity using Cauchy graph embedding is designed for learning mutual information of both local and global aspects of overlapping features of compound fault, which improves the representation learning ability of the decoupling model. Finally, the generalized sparse <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">l_p</i>/<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">l_q</i> norm is introduced to redesign the probabilistic output function for compound fault decoupling, which improves the decoupling generalization and robustness of the model to unknown compound faults under the training using only single fault samples. To verify the effectiveness of the proposed method, it was validated on a self-made airborne fuel pump (AFP) experimental platform. Extensive results show that our proposed method reaches an optimal average accuracy of 99.66% and 93.4% for decoupling compound fault under constant and varying operating conditions, respectively, without any compound fault samples involved in the model training process, and outperforms a series of existing state-of-the-art models.