On-line Monitoring Of Welding Quality Of Stainless Steel GTAW In Spent Fuel Pool Based On Deep Learning

As a high quality and stable welding method,GTAW is widely used in aerospace,nuclear power construction and other industries.As the place of fuel storage and transportation in spent fuel pool,the quality reliability and stability are particularly important.Therefore,the quality inspection of weld is particularly important to ensure the stable load of nuclear fuel pool.However,most of the current welding detection methods are post welding detection,which has some defects such as poor real-time performance and limited detection effect.To solve this problem,this paper proposes a GTAW welding quality monitoring system based on weld pool vision sensor to study how to realize the real-time monitoring of welding deviation and penetration state.The system can ensure the welding quality and reliability of spent fuel pool steel cladding effectively and promote the development of welding process sensing technology.First of all,the online monitoring system of molten pool is constructed,which is the passive vision sensing system.The calibration method of internal and external parameters of the camera is studied,and the optical center coordinates and pose information of the camera are obtained.Under different welding parameters,8492 images of weld pool and front weld are collected.Then the collected molten pool images are processed and classified,and the database is established according to the requirements.Then an image processing algorithm suitable for weld pool image in welding process is proposed.The image processing methods such as filtering,Hough line detection,gamma transform and edge detection in Open CV are studied to improve the feature recognition effect and determine the better image processing methods.The information of the center position of the weld pool and the center of the weld seam are calculated to recognize the welding deviation.Aiming at the complexity and instability of welding process,the image processing algorithm is improved to improve the adaptability and stability of the algorithm.The average error of of the final algorithm for recognition of welding deviation is 0.21 mm,which meets the actual production demand enough.Finally,the prediction model of penetration state based on CNN is established.The acquired two-dimensional image information is used as input and four kinds of penetration states are output.In the process of training,automatic function optimization and manual adjustment are used to optimize the super parameters and model structure parameters.The final convolutional neural network structure consists of 14 convolution layers,7 pooling layers and 3 fully connected layers.The learning rate of model is 0.0009 and the last convolution layer contains 128 convolution kernels.After 200 iterations,the accuracy of penetration prediction in training set reaches 100% and the accuracy of penetration prediction in test set reaches 99.49%,which achieving the expected goal.Then,the characteristics of the molten pool are visualized and the working mechanism of the prediction models under several penetration states is described.The TIG welding pool monitoring software is developed and the integration mode of each module is studied.The overall program diagram and the construction scheme of the software is output.