Methodology of neural modelling in fault detection with the use of chaos engineering

The paper deals with the problem of robust fault detection using recurrent neural networks and chaos engineering. The main part of the proposed approach is a locally recurrent neural network that is composed of complex dynamic neural units for which chaotic behaviour can be obtained. Selected global and local optimization methods are connected to have diphase strategies for training this kind of neural networks. And beyond this, chaos engineering is incorporated into both the evolutionary and simulated annealing algorithm in order to improve the efficiency of the tuning procedure. The problem of selecting the most relevant input variables is solved by means of extended Hellwigs method of integral capacity of information. Criteria isolines and a sensitive-based method are used to identify the suitable architecture of a neural network. Moreover, the issue of stability analysis of neural models is also considered in this paper. Recurrence quantification analysis is proposed for residual evaluation in order to have the comprehensive methodology of neural model-based fault detection. The preliminary verification of the elaborated methodology in modelling tasks was carried out for both simulation and industrial data. The fundamental verification was conducted for the data made available within DAMADICS benchmark problem. The achieved results confirm the effectiveness of the proposed approach.