Processing heterogeneous diagnostic information on the basis of a hybrid neural model of Dempster-Shafer

Abstract

In the this work it is emphasized that fusion of the diverse data obtained from sources of primary information (sensors, the measuring equipment, systems, subsystems) for adoption of diagnostic decisions at a research of faults of devices, is one of the main problems in information processing. A generalized scheme of fusion of diverse data reflecting features of this process is considered. Classification of levels, modern methods of fusion of diverse data in the conditions of incomplete, indistinct basic data is also considered. The article develops a new hybrid approach to the diagnosis of technical objects based on multisensory data in terms of heterogeneity of the original information. We consider a new class of adaptive network models focused on the implementation of the procedures of logical-probabilistic inference using the Dempster-Shafer methodology and fuzzy logic. The adaptive Dempster-Shafer model (DS model) is a multilayered network of neurons mapped to the elements of the hypothesis space together with the current values of their base probabilities, on the basis of which the confidence probabilities of hypotheses are calculated. The original training algorithm for the neural network model with the attraction of experimental data is based on the principle of the distribution of the total error in the neural network in proportion to their confidence probabilities. The network model of Dempster-Shafer functions is trained together with the neural network model, which simulates the process of forming empirical estimates of hypotheses on the basis of subjective preferences of experts for the influence of various factors on diagnostic solutions. The principal advantage of the hybrid system is the ability to jointly adapt the parameters of both models in the learning process, which increases the reliability of the results of calculations due to the diversity of the used expert statistical information. The adaptability of the hybrid system also makes it possible to implement a new approach to the calculation of probability estimates of hypotheses based on a combination of several evidence by training a hybrid system based on data from several sources.