Research On Recognition Method Of Abnormal Noise Of Automobile Interior Parts

The abnormal noise problem in the interior of the car is an important factor that affects the NVH performance of the car.The level of abnormal noise control reflects the comprehensive capabilities of the OEM in vehicle design,processing and assembly,and gradually becomes a factor that consumers consider when buying a new car.With the continuous advancement of intelligent technology,engineers have put forward higher requirements for the diagnosis method of abnormal sound problems.In recent years,intelligent vehicle fault recognition methods have developed rapidly,but the application of voice recognition technology is less,and with the development of deep neural network models,most of their characteristic parameters still follow the parameters in the shallow model,while in the car The diagnosis direction of the abnormal sound source of decorative parts lacks a systematic sound recognition solution.In view of the above problems,this paper proposes a method for identifying abnormal sounds of automotive interior parts based on FBank maps and convolutional neural networks,and builds a recognition process that includes signal noise reduction,preprocessing,and endpoint detection algorithms.The main contents are as follows:(1)Select the abnormal noise of window resonance,seat rail collision,glove box buckle and seat friction as research objects,analyze its main mechanism.Use the abnormal noise semi-anechoic laboratory and the four-column laboratory to carry out abnormal noise signal acquisition experiments,combined with the UrbanSound8 K data set to establish the sample set used in this paper.Aiming at the problem of large background noise interference in the real vehicle environment,a wavelet threshold noise reduction algorithm based on Bayesian theory optimization is studied.The laboratory signal is superimposed with road noise and the noise reduction experiment is performed while using the SNR and Corr values to evaluate the noise reduction effect.The experiment shows that the method can effectively suppress the noise,and the effect is significant under the condition of low signal to noise ratio.(2)Signal preprocessing is an indispensable process in the voice recognition process.This paper normalizes,pre-emphasizes,and frames the windowing preprocessing steps of the signal,and proposes to use a dual-based energy based on short-term energy and short-term zero-crossing rate.The algorithm performs endpoint detection on abnormal sound signals,verifies the detection effect under different signal-to-noise ratios,and finds that the endpoint position of abnormal sound signalscan be accurately detected when used in conjunction with noise reduction algorithms.Extract the signal's time-frequency spectrum and Mel frequency cepstrum coefficients.For the characteristics of low correlation between the Mel frequency cepstral coefficients dimensions,it is proposed to remove the discrete cosine transform in the calculation process to obtain a new feature parameter: FBank map for subsequent comparison.Verify the effectiveness in abnormal sound recognition.(3)Set up the network structure for three different characteristic parameters and perform model training based on Tenosrflow/Keras,observe the test accuracy and training time of different models,explore the optimal convolution kernel number setting in the model.Use data augmentation and add a dropout layer to optimize the overfitting phenomenon that occurs in spectrogram model.The results show that the FBank map model achieves the best results in the test accuracy and training duration of the laboratory signal test set,and the values are 89.5% and 24 s per round.The pre-trained model is called to identify the real vehicle signal test set before and after noise reduction.The results show that the FBank map model can achieve a recognition rate of 83.3% in the test set after noise reduction.The combination of the proposed wavelet Bayesian threshold noise reduction algorithm and the FBank maps CNN model can achieve the best recognition effect under high and low signal-to-noise ratio.(4)Integrate various algorithms in this paper and develop recognition software platform based on Matlab App Designer.Analyze the functional requirements of the software and devide it into four sub-modules.Each module is programmed and interface designed.Finally,the software function test is carried out to verify the effectiveness of the system.