STFT and CNN-1D for Mechanical Anomaly Detection in Electrical Machines

Mechanical anomaly detection in electrical machines is essential for preventing costly breakdowns and improving predictive maintenance. In this work, we apply Short-Time Fourier Transform (STFT) coupled with a 1D Convolutional Neural Network (CNN-1D) to analyse the acoustic signals from machines. Unlike the traditional 2D spectrogram-based approach, we use 1D representations of the acoustic data provided by the MIMII Dataset. A CNN-1D model is trained on this data to detect mechanical anomalies in electrical machines. The results show that this approach achieves detection accuracy exceeding 97 %, providing an efficient solution for real-time fault detection. We also discuss the challenges related to false positives in detections and propose optimization strategies to improve the overall model performance.