A Study on Vehicle Abnormal Noise Event Detection Based on Convolutional Neural Network

To address the challenges of low real-time performance and high false alarm rates in vehicle abnormal sound detection within industrial scenarios, this study proposes an end-to-end detection framework integrating acoustic feature optimization and lightweight deep learning. By enhancing the Mel-frequency cepstral coefficient (MFCC)-based feature extraction process through a Dynamic Differential Cepstral Coefficient Enhancement module, we effectively improve the time-frequency representation capability for transient abnormal sound events. Lightweight Depth Separable Convolutional Network (LDSCNN) is designed to achieve adaptive feature learning under the constraint of merely 1.2 M parameters. Experiments conducted on a collected vehicle abnormal sound dataset employ a 300 ms segmentation strategy to balance detection real-time performance and event coverage. Results demonstrate that the proposed model achieves 90% accuracy in complex noise environments, 93.75% F1-score for abnormal sound detection, and 35 ms single-sample inference time, significantly outperforming conventional methods. This research breaks through the collaborative optimization bottleneck between feature extraction and model architecture in industrial scenarios, providing an intelligent solution with high interpretability and low deployment costs for vehicle noise, vibration, and harshness (NVH) performance evaluation.