Abstract
<jats:p>Analysis of recent research and publications. Modern information-measuring systems for monitoring the technical condition of transport objects require effective methods for proc-essing large volumes of measurement data. In railway engineering, particular attention is paid to the timely detection of defects in wheelsets during train operation, since such defects directly affect traffic safety and the reliability of rolling stock. Various methods of vibration, acoustic and acoustic-emission diagnostics are widely used in modern studies of railway ve-hicle condition monitoring. Researchers have demonstrated the effectiveness of spectral anal-ysis, wavelet transforms and statistical signal processing for identifying diagnostic features of mechanical systems. Recently, machine learning and deep neural networks have been increas-ingly applied to automatic classification of signals and fault detection in technical systems. However, the use of spectrograms of acoustic signals as input data for neural networks in the tasks of monitoring wheel–rail interaction remains insufficiently investigated. Purpose of the research. The purpose of this study is to develop an approach for proc-essing acoustic signals of wheel–rail interaction using neural networks for automated identi-fication of characteristic operating modes of railway rolling stock. The approach is based on the use of spectrograms of acoustic signals as informative features for training a machine learning model capable of classifying different types of acoustic events.Presentation of the main research material. Experimental acoustic measurements were performed during the passage of railway rolling stock under real operating conditions. The study included recordings from passenger wagons, freight wagons and electric train cars at different speed ranges. The measurements were conducted using a digital recording device with a frequency range of 30 Hz–20 kHz and a sampling frequency of 96 kHz. To ensure reli-able results, the measurement conditions were controlled to minimize external acoustic inter-ference.Based on the obtained recordings, spectrograms of acoustic signals were constructed and analyzed. Expert analysis and manual labeling of spectrograms allowed identification of characteristic patterns corresponding to different noise sources such as rolling noise, aerodynamic noise, braking friction, wheel–rail impacts and squeal in curved track sections. These patterns were used to form a labeled dataset for training a convolutional neural network.For automatic classification of acoustic events, a convolutional neural network archi-tecture based on a pre-trained MobileNet feature extractor was used. Spectrograms were di-vided into equal square fragments, which allowed the continuous acoustic signal to be trans-formed into a standardized set of images containing local time-frequency features. The trained model was able to classify four main types of acoustic signals: wheel–rail impact, rolling noise, aerodynamic noise and curve squeal. Experimental results demonstrated that the neural network successfully identified characteristic patterns of acoustic signals with high classification probability.Conclusions. The conducted study confirms the effectiveness of applying neural net-works for the analysis of spectrograms of acoustic signals generated during wheel–rail inter-action. The proposed approach allows automatic identification of characteristic acoustic events associated with different operating conditions of railway rolling stock. The obtained results demonstrate the potential of integrating machine learning methods into intelligent in-formation-measuring systems for railway diagnostics and monitoring. The proposed method can contribute to improving the efficiency of acoustic monitoring systems and to the develop-ment of predictive maintenance technologies for railway vehicles.</jats:p>