Process industrial components intelligent fault diagnosis based on multi-sensor information interaction and multi-level feature adaptive fusion framework

Abstract With the increasing complexity and digitization of industrial processes, the safety and reliability of process industrial components (PICs) are facing significant challenges. Due to the limitations of single-state signals in comprehensively reflecting the operating conditions of PICs, this paper considers the correlation and complementarity among multiple sensor signals, proposing a multi-sensor information interaction and multi-level feature adaptive fusion method (M2F2N) for PICs fault diagnosis. Firstly, considering the characteristics of faults at different stages, a shallow-middle-deep multi-level feature extraction network is designed to parallelly extract independent features of different sensors at multiple stages. Then, an information interaction network is developed to extract complementarity features between heterogeneous sensors, enhancing the fault representation capability of the network through mutual attention among signals. Lastly, the adaptive fusion network adaptively weights the features at all stages layer-by-layer to generate the final recognizable fusion features. The experimental results show that the average accuracy of the M2F2N method reaches 96.20% on the plunger pump dataset, and the convergence speed is 75% faster than that of the traditional method. On the control valve dataset, its robustness is reflected in that the error rate is the lowest compared with other methods under varying noise interference levels, which proves the superiority and robustness of the proposed M2F2N method.