Application Research Of Ultrasonic Recognition Method Of Weld Defects Based On High-dimensional Feature Information Fusion

With the development of production and technology,welding as an important processing method has been widely used in various fields.How to ensure the quality of welding has become an important issue to consider to ensure the safe and efficient operation of equipment.Nowadays,the ultrasonic welding method is often used to detect welds.This method is non-radiative,fast and does not damage the weldment itself.However,in traditional ultrasonic testing,manual methods are often used to determine the type of defects,which has low efficiency,poor consistency,and high requirements on the experience of the inspectors.The existing methods have reached the bottleneck for the recognition of single features,and high-dimensional features will bring to calculate the substantial increase in the amount of calculation,it is difficult to balance efficiency and accuracy.Aiming at the problem that the accuracy rate of high-dimensional sample recognition of weld defects needs to be further improved,taking three types of buried defects,such as bubble,slag nd cracks,as the main research objects,a new algorithm combining random forest algorithm and DS evidence theory(Dempster-Shafer envidence theory)is proposed.The random forest algorithm is resistant to noise and can well handle the characteristics of high-dimensional sample processing.It is combined with the decision-level information fusion capability of DS evidence theory to achieve efficient and accurate identification of weld defects.Experiments were carried out using prefabricated defect test block to verify the accuracy of the proposed algorithm.main tasks as follows:(1)Finite element simulation and experiment.Based on the basic detection principle of the ultrasonic reflection method,a finite element simulation of electrostaticsolid mechanics-pressure acoustics multiphysics coupling was carried out to study the influence of the interaction between ultrasonic and different kinds of defects on the echo signal.Experiments were performed with ultrasonic testing equipment to collect ultrasonic echo signals of three types of prefabricated weld defect patterns such as pores,cracks,and slag inclusions.(2)Feature extraction.Considering that a single type of ultrasonic echo signal characteristics cannot well characterize the properties of weld defects,three types of features are extracted in the time domain,frequency domain and time-frequency domain.Taking into account the differences in the information carried by the echo signals generated by different propagation processes,the echo signals are divided into defect echoes and bottom echoes,and feature extraction is performed separately.(3)Identification of defect types.The random forest algorithm is trained using the extracted three types of feature sample sets and tested.The results show that the random forest algorithm can better identify three kinds of defects,but there is room for further improvement in accuracy.The DS evidence theory is introduced to make decision-level information fusion of random forest recognition results,and comprehensive decision is made on the random forest recognition effect trained by three types of features to eliminate the one-sidedness brought by a single kind of feature.The experimental results show that,with the introduction of DS evidence theory,the recognition accuracy of three types of weld defects such as pores,cracks,and slag inclusions is significantly improved,and the accuracy and algorithm stability are better than those of support vector machines.