Category: Monitoring methods in transport systems

Authors:

Bazilevsky F. Yu. , Grachev V. V. , Grishchenko A. V. , Kruchek V. A. , Schwartz M. A. , Schwartz F. M.

Annotation:
Despite the vast experience of using the neural networks for solving various machine learning problems, the numerous attempts to use them in technical diagnostics have not yet led to complete solutions so far (with rare exceptions). The reason is the specific nature of technical diagnostics that distinguishes such tasks from traditional machine learning problems. Having analyzed these specific features, the authors propose an approach to diagnosing complex technical objects that is focused on the use in built-in diagnostics systems and is based on the neural network reference diagnostic models of functionally isolated nodes and assemblies. The article describes the methodology for the synthesis of such models, their training on the data obtained by monitoring the object being tested using built-in diagnostic tools, determining the permissible response errors, and adapting to the current status of the object. The fuzzification of the diagnostic model results using the test sample proposed in the article makes it possible to standardize the approach to diagnosing complex technical objects designed for various purposes. The use of D. Trigg’s tracking control signal proposed by the authors to monitor regression residuals during the learning increases the training quality and generalization ability of models. The value of this signal determined by the model run on a test sample is an additional informative diagnostic parameter that increases the accuracy of classifying the status of the object under test. The proposed methodology applied at the complex technical object design stage allows optimizing the monitored parameters’ array and multiplying the efficiency of the diagnostic information recorded by the built-in diagnostic and monitoring tools.

Key words:
Technical diagnostics, neural network regression model, decomposition of a complex object, training sample, model response error, monitored parameters’ array, diagnostic parameters’ array, an alphabetic arrangement of object status classes, fuzzification, membership function


Download article