Research On Application Of The Classification Algorithm Of Vehicle Abnormal Noises Based On Mechine Learning

When abnormal noise appears on the vehicle,it is often a harbinger of component failure,and if not located and eliminated in a timely manner,it is likely to cause functional failure of vehicle components or even a series of safety problems.At present,technical service personnel generally use manual listening to troubleshoot abnormal noise problems,inevitably low efficiency and poor precision.furthermore,deficient dataset restricts the development of intelligence in this field.In this paper,we address the above phenomenon and complete the intelligent classification of abnormal noise based on artificial intelligence methods.To address the problem of the insufficient dataset of abnormal noises,seven types of abnormal noises were collected under various working conditions,and the labeling of the collected abnormal noises was completed by subjective and objective evaluation,and two data expansion techniques,audio cropping and data enhancement,were proposed to increase the size of the dataset.For the feature extraction of vehicle noise,the time domain analysis is used to understand the time domain direction of the abnormal noise;the time-frequency domain analysis is used to obtain the comprehensive performance of the noise features in the time and frequency domain,and the logarithmic energy,Mel cepstrum coefficients and wavelet packet Mel cepstrum coefficients are extracted by this method,and the feature fusion of logarithmic energy,Mel cepstrum coefficients and wavelet packet Mel cepstrum coefficients is completed in a tandem manner.Based on three classical machine learning models: k-nearest neighbor,support vector machine and multi-layer perceptron,their applications in vehicle abnormal noise classification are studied.Using sklearn machine learning library,the classification performance and optimization mode of the three types of models are explored based on the fusion features;The effects of data expansion techniques and feature types on classification performance are analyzed.To address the problem of low categorization accuracy,based on convolutional neural network and transformer encoder,a deep learning classification model with parallel mechanism is proposed to extract the spatial and time sequence information of abnormal noise features at the same time.The results of performance experiments show that when Mel cepstrum coefficient feature is used as input,the classification accuracy of this model can reach 93.80% in the test set.