Audio-Driven Predictive Maintenance for Electric Vehicles: A Convolutional Neural Network Approach

The traditional fault detection methods in a vehicle are either based on the experience of the mechanics or the service center diagnostic tools. They are expensive, timeconsuming, and also provide dubious accuracy most of the time. To avoid the above drawbacks, this work suggests a new sound-based early fault detection system for electric vehicle motors using Convolutional Neural Networks. The motor's acoustic signals recorded for different operating conditions will be utilized in the system so that fault patterns would be identified and possible failures would be predicted. To this end, the major input features for the CNN model which would enhance the accuracy and reliability of the diagnosis are MFCCs and spectral features derived from the sound records. Such a system will be able to differentiate normal operation, heavy loaded state, and motor fault with high accuracy through providing adequate training on the CNN by larger, diverse sets of motor sounds both for healthy and faulty motors. Results we got like Accuracy 93%, Precision 85%, Recall 90%, F1-Score 0.87, AUC 0.92 Processing Time 200 ms per audio clip and Cross-Validation Accuracy 92%-94%. It concluded from the result that our system is good in terms of F1-score, precision, recall, and accuracy with an excellent overall classification accuracy of more than 90%. Another contribution of this research work is thorough frequency analysis, which lead to identification of prominent sound features indicative of some faults, thus proving helpful in the development of more sophisticated diagnostic tools. Finally, this research finds a low-cost, non-intrusive, and potentially reliable solution for early fault detection in electric vehicle motors-which would be used as the basis for real-time condition monitoring and for predictive maintenance-offering a means to further refining ideas on automobile diagnostics.