Research On The Aided Recognition Method Of Weld Defects Based On Multi-domain Analysis Of Ultrasonic Signals And Model Self-optimization

Welding defect identification is very important for industrial activities such as special equipment pressure vessel and pressure pipeline and is an important task in industrial production.As a kind of national resources,the weld defect identification of austenitic stainless steel is an important task in welding defect identification.Therefore,it is an urgent engineering practice problem to construct weld defects of austenitic stainless steel quickly and timely.Although the current weld defect identification process has reached a certain level of mechanization,there are still the following problems: 1)Existing weld mechanization identification methods still rely on professional personnel to analyze the severity of defects and determine the nature of defects,which is highly subjective,time-consuming and laborious,and relatively high identification cost.2)In the deep learning method research for weld defect identification,there are problems such as low information richness of recognition signal feature,insufficient feature extraction of convolutional network,limited discriminance,strong dependence of artificial professional field of deep network architecture,high network redundancy,and lack of adaptive ability.In view of the above problems and difficulties,the main research carried out in this paper is as follows:(1)Multi-spatial transformation of information based on ultrasonic time domain signals.Will be collected in the ultrasonic frequency domain and time domain signal is derived to gram angular field,markov transfer field domain,such as space in the form of a multi-domain,giving more abundant signal characterization information,to identify the signal feature extraction clustering problem is transformed into different spatial domain of 2D image processing problems,more adapt to the convolutional neural network to 2D image processing.(2)Construction based on lightweight multi-scale depth separable convolution model.Multi-scale depth separable convolution was proposed.The single convolution kernel in the deep convolution was replaced with multi-scale convolution kernel of different sizes,and the convolution receptive field was enlarged.By extracting weld defect information at different scales,the richness of feature information extraction was effectively improved.(3)Self-optimization of recognition model based on sparrow search algorithm.The sparrow search algorithm was applied to the weld defect identification model based on timefrequency feature fusion and the deep multi-scale lightweight identification model based on multi-domain analysis.The self-compression of model structure and multi-dimensional hyperparameter single-objective optimization based on sparrow search algorithm were carried out.Taking recognition accuracy as fitness function,kernel function type,kernel parameter and penalty factor of support vector machine,The Mobile Net V3 network structure and the multi-dimensional hyperparameters of the LM-Mobile Net V3 network,such as multi-scale convolution kernel size,multi-scale convolution kernel ratio,learning rate,optimizer,channel expansion number of each layer of the network,were optimized adaptively.Based on the above research content,a stainless steel weld defect identification and management system was designed,which realized the functions of data processing,data query and user management,and met the needs of the practical application of weld defect identification.The research of this paper is helpful to improve the automatic level of weld defect identification,which is of great significance to reduce the economic loss caused by welding quality and reduce the occurrence of safety liability accidents.