Research On Abnormal Detection Method Of Anchor Chain Welding Process Based On Convolution Neural Network

As the key equipment of marine transportation,ship anchor chain is responsible for marine exploration,marine economic development and personnel safety.Due to the structural characteristics of the anchor chain,the quality of the anchor chain depends on the worst single link in the whole chain.If any link has quality problems,it will directly lead to the unqualified quality inspection of the whole anchor chain,and the work of replacing the defective anchor chain and reorganizing the new anchor chain in the later stage is undoubtedly expensive and inefficient.Therefore,how to detect the defects in the flash welding stage of anchor chain is very important.In order to detect the flash welding quality of anchor chain efficiently and accurately,this thesis decomposes the problem from two aspects of data preprocessing and deep learning.In the direction of data preprocessing,from the perspective of data enhancement,this thesis proposes a data enhancement algorithm in time domain for a small number of label data,and uses the random window slicing algorithm to improve the efficiency;In the direction of deep learning,this thesis proposes a joint network architecture,and carries out experimental research on the setting of network super parameters.According to the characteristics of flash welding data,the input method of this kind of data is studied experimentally.The main research contents are as follows:Firstly,in order to solve the imbalance problem of time sequence data label in flash welding of anchor chain,a DDA data enhancement algorithm is proposed;In order to further improve the efficiency of data enhancement algorithm,a window slicing algorithm for flash welding data is proposed.In order to verify the algorithm proposed in this thesis,two different types of convolutional neural networks are built as the experimental network model,and flash welding data and public data set are used as the experimental data.The experimental results show that the convergence speed of the test set is significantly accelerated after data enhancement,and the detection accuracy of anchor chain anomaly is also significantly improved.Then,by introducing confusion matrix as the evaluation index of unbalanced data,the flash welding data are discussed experimentally.For the data of flash welding of anchor chain,the qualified sample data can be used as auxiliary data for data balance processing.The experimental results show that the accuracy and recall of deep learning network are improved.According to the characteristics of flash welding data with multiple samples and the same label,the data input method is tested.The experimental results show that the multi-channel input can significantly improve the accuracy.Finally,in order to solve the problem of anomaly classification in the welding process of anchor chain,an improved joint network framework of perception net is proposed,and a weight output method based on joint network is proposed,which improves the recognition accuracy of the network.In order to further improve the network performance,the direction of super parameter optimization is studied and analyzed.According to the optimization results of super parameter experiment,the recall rate of abnormal detection in flash welding process of marine anchor chain can reach 82.07%,and the recognition accuracy rate can reach 83.73%.