Research On The Separation And Detection Of Gearbox Sound Source Based On Multimodal Fusion

As a mechanical component for power transmission,torque transfer,speed and direction change,gearbox is widely used in rotating machinery equipment such as lathes and milling machines.Due to its high load and high speed working characteristics,it is prone to failures.Therefore,carrying out gearbox fault detection is of great significance to ensure the normal operation of rotating machinery equipment such as lathes and milling machines.Vibration signals and audio signals can characterize the operating status of mechanical equipment and are widely used in fault detection.However,the installation of contact vibration sensors is difficult in some scenarios,and vibration signals are difficult to collect.Video surveillance,as a multi-modal data acquisition equipment for audio and video,has the advantages of simple installation and non-contact data acquisition,and has been widely used in security,industrial production and other fields.Therefore,this thesis studies gearbox fault detection based on audio and video data.The main research contents can be summarized as follows:(1)A self-supervised learning-based gear sound source separation algorithm is proposed to address the problem of non-uniform results in single-modal blind source separation of mixed audio.The visual features of mechanical equipment are combined with the mixed audio features,and the corresponding sound source separation is achieved based on visual modality features.To improve the sound source separation effect,an audio feature extraction network CA-UNet with embedded coordinate attention mechanism and a visual feature extraction network Res2Net18 with multi-scale feature extraction capability are constructed.Experimental results demonstrate that the proposed gear sound source separation algorithm can effectively separate the gear audio in mixed audio signals,laying the foundation for gear fault detection.(2)A two-stage audio-visual sound source localization algorithm based on comparative learning is proposed to address the problem of traditional sound source localization methods that cannot visualize the localization results.In the training stage,by constructing positive and negative sample pairs for comparative learning,in order to improve the sound source localization effect,supervisory loss is introduced into the comparative learning loss to obtain the initial localization result;In the inference stage,the pre trained Res Ne St-50 is used as the target guidance module to extract visual features,which are integrated into the initial positioning results to improve positioning accuracy.The experimental results show that the proposed sound source localization algorithm can effectively locate the sound source of the gearbox.(3)A gearbox fault detection algorithm based on adversarial autoencoders is proposed to address the problem of limited fault detection accuracy in classical autoencoder models.By adding the identification network in the self encoder model,the encoder can map the input sample data to the standard normal distribution while minimizing the reconstruction error,so as to improve the expression ability of the model and improve the accuracy of gearbox fault detection.The experimental results show that the proposed adversarial autoencoder has an AUC improvement of 12.6% compared to the classical autoencoder,demonstrating the superiority of the adversarial autoencoder model,which can be used for gearbox fault detection.