Continual learning fault diagnosis method based on grid-based artificial immune system

Abstract

In recent years, continual learning for class increments has attracted a great deal of attention. The ontinual-learning classification method (CLCM based on an artificial immune system (AIS) can identify unknown faults during testing. However, the CLCM still has the problem of excessive runtime consumption. Therefore, it is crucial to improve the efficiency of the immune algorithm and take advantage of its continual learning mechanism in the field of fault diagnosis. In this paper, a continual learning fault diagnosis method based on sparse grid and the AIS, which called sparse grid classification method (SGCM), is proposed, which is inspired by grid-based techniques and the CLCM based on an AIS. Firstly, a new cell generation strategy is proposed to reduce the time complexity and improve the diagnosis efficiency; therefore, the problem of dimension explosion is avoided. In addition, the memory cell coding capabilities of the SGCM increases the utilization rate of cells so as to simplify the calculation of affinity. At the same time, the conceived cell backtracking strategy enhances the continual learning ability of the algorithm so that new fault types can be quickly identified through the existing learning results. Ultimately, the model adaptive adjustment method inspired by a single-layer feed-forward neural network improves the generalization power and the accuracy of classification. We conduct experiments on well-known datasets from the UCI repository to assess the performance of the SGCM. To evaluate the fault diagnosis performance of the SGCM, experiments on a reciprocating compressor experimental dataset and the XJTU-SY rolling element bearing dataset were performed. The results show that theSGCM is a fast fault diagnosis method with low time complexity and continual learning ability.