Research On Classification Method Of Abnormal Sound Of Loudspeaker Based On Psychoacoustics

China has become the world’s most important production and export base of loudspeakers,the traditional loudspeaker quality testing and classification method is to diagnose by means of manual listening,due to the results of manual listening the influence of the subjective factors of the listener,it is difficult to form a stable and unified loudspeaker testing classification standards,and easy to cause hearing damage to the listener,so in the production process of loudspeakers with the signal processing method,Therefore,it is important to detect quality and classify faults in the production process of loudspeakers by means of signal processing.In order to improve the correct rate of loudspeaker classification,this paper studies the psychoacoustic characteristics of the human ear hearing system for loudspeaker abnormal sound classification proposes a fault feature extraction method based on psychoacoustics and conducts classification experiments.The main research contents of this paper are.1.First of all,analysed the mechanism of the abnormal sound generated by the loudspeaker under the three types of faults: touch ring,air leakage,and slight sound,and set up a speaker response signal acquisition platform,and then data acquisition of standard speakers and the above three types of faulty speakers.2.In order to improve the accuracy of loudspeaker anomalous sound classification and reduce the impact of problems such as high spatial dimensionality and considerable redundancy of psychoacoustic models,this paper proposes a feature extraction method that combines auditory masking with the maximum relevance minimum redundancy algorithm,drawing on the role of the masking effect of the human ear on auditory sound recognition.Firstly,the auditory mask spectrum of the speaker response signal is calculated,the singular value decomposition of the auditory mask spectrum is performed,and the obtained singular values are used as the feature set to be selected;Finally,the MRMR algorithm is used to select the optimal features from this feature set.The analysis of the t-distributed stochastic neighbor embedding algorithm shows that the method can extract speaker response signal features with better discrimination and eliminate redundant information and avoid the problem of the too high dimensionality of the classification model due to too many features in the classification process.3.Based on the standard psychoacoustic features in the subjective hearing of the human ear,the paper proposes a feature extraction method based on entropy-weighted time-varying feature loudness combined with two-dimensional principal component analysis for abnormal loudspeaker sounds.The method analyzes the loudspeaker acoustic response signal by simulating artificial listening.In order to facilitate the feature extraction of fault information,this method calculates the weighting coefficients based on the energy entropy of each frequency band of time-varying feature loudness and then enhances the features of each frequency band in time-varying feature loudness by the weighting coefficients.In order to reduce the amount of feature redundancy,this method uses two-dimensional principal component analysis to reduce the dimensionality of high-dimensional features.The analysis by the t-distributed stochastic neighbor embedding algorithm shows that the method can effectively improve the recognition rate of different faculty speakers.4.A decision fusion-based speaker detection and classification method is proposed in the paper.The method uses a voting mechanism to fuse the two feature extraction methods proposed in the paper,as well as traditional time-frequency domain features,and the paper compares weighted voting with traditional voting decision fusion methods.The experiments show that the use of the weighted voting decision fusion method can solve the problem of inadequate classification results of a single method,as well as improve the overall speaker fault recognition rate;the average recognition rate of speaker fault classification reached 98.9%.