A comparative analysis of clustering techniques for the detection of anomalies in the parameters сharacterising the functioning of an electrical railway system

Abstract

This article considers the problem of anomaly detection and the solutions to this problem by means of clustering techniques. The study was performed on railway equipment. The article describes several maintenance strategies with the predictive maintenance strategy considered to be the most promising. The main components and aims of a predictive maintenance system are discussed with regard to the railway area. The article considers in detail the implementation of the diagnostics module and formulates the anomaly detection problem based on the sample measurements of the electrical parameters of the functioning of a wired railway system. Time-series values of electrical signals are considered to be the input data for the anomaly detection problem. To preprocess the data spectral analysis techniques were used: the estimation of the power spectrum density together with a periodogram-based metric. The data was processed using the sliding window approach. The article presents the results of the comparison of support vector machines and K-means clustering when applied to the test data and evaluates the ratio of correct answers. The optimal parameters were determined.