Abnormal Detection of Heavy Haul Train Air Whistle Based on MFCC and Lightweight Training

The air whistle is utilized for locomotive and railway communication and alarm. During the operation of heavy haul train, damage to internal mechanical components caused by body vibration can have an effect on the sound of the air whistle. Its upkeep frequently relies on manual experience, resulting in inefficient maintenance. Meanwhile, when the air whistle is stowed for repair, its high decibel level will cause hearing harm to maintenance employees. Therefore, by deconstructing the characteristic modes of sound signals, this work provides an intelligent detection algorithm for air whistle anomalies, which increases recognition accuracy and accomplishes rapid and efficient maintenance. First, normalize the on-site air whistle sound signals and clean the other on-site signals. The mel-scale frequency cepstral coefficients (MFCC) is then utilized to extract the feature coefficients of the sound signals, which is useful for further classifier recognition. Finally, for training the air whistle anomaly detection model, feed the feature coefficients into the proposed lightweight network. The experiment demonstrates that the accuracy of the four maintenance states presented by the air whistle reaches 93.81%, meeting the requirements of on-site maintenance standards and providing a realistic and feasible solution for air whistle maintenance in heavy haul train.