An intelligent approach for vehicle-whistle recognition based on variational mode decomposition and ReliefF fusion features

This paper proposes a vehicle-whistle recognition method based on variational mode decomposition (VMD) and the ReliefF algorithm. A vehicle-whistle dataset is first constructed, and a VMD-based signal reconstruction approach is developed for transient whistle enhancement. A double-threshold endpoint detection method based on short-time energy and zero-crossing rate is employed to segment whistle signals. The optimal number of VMD decomposition modes is determined using a kurtosis-based criterion, and the results indicate that transient whistle signals can be effectively represented using the first four intrinsic mode functions (IMFs). To improve recognition performance, a feature fusion strategy based on the ReliefF algorithm is proposed by integrating Mel-frequency cepstral coefficients (MFCC) and linear prediction cepstral coefficients (LPCC). Two classification models, namely a back-propagation neural network (BP-NN) and a support vector machine (SVM), are established for recognition. Compared with conventional MFCC and LPCC features, the proposed ReliefF-based fusion features improve both recognition accuracy and computational efficiency. In addition, comparative experiments among BP-NN, SVM, and CNN demonstrate that the BP-NN model achieves superior classification performance and more stable results. Finally, the proposed recognition framework is experimentally validated. The results demonstrate that the integration of VMD, ReliefF-based feature fusion, and BP-NN achieves high recognition accuracy, a low false alarm rate and efficient vehicle-whistle detection. This study proposes a practical framework for transient sound source recognition, which may be applied to traffic monitoring and illegal vehicle-whistle detection.