Research On Quality Inspection Method Of Anchor Chain Flash Welding Based On Deep Learning

The anchor chain is made up of many chain links.It is an important device used to buffer the external forces on ships and offshore engineering platforms.The quality of the anchor chain often affects the safety of life and property of the relevant personnel.In the production process of the anchor chain,flash welding is the most important link,which determines the mechanical properties of the anchor chain.Even if only one chain link is broken due to quality problems,the entire anchor chain will fail,and the loss will be huge.At present,the quality of the anchor chain flash welding is mainly detected by the tensile test and the break test before leaving the factory,but this method makes the cost of replacing the unqualified chain links extremely high.Therefore,how to timely and effectively detect the quality of the anchor chain flash welding is very important.Aiming at the imbalance of anchor chain samples,this paper proposes the nearest neighbor piecewise interpolation sampling,and based on deep learning,studies the problems such as how to obtain the labels of samples and how to learn new samples.The main research contents are as follows:First of all,because there are far more qualified samples than unqualified samples in the anchor chain flash welding,the problem of data imbalance between different samples is caused.Aiming at this problem,combining with the characteristics of the electrode-position curve and the current curve during the anchor chain flash welding,a new sampling method is proposed: the nearest neighbor piecewise interpolation sampling.New samples are synthesized through a series of processes such as normalization,dynamic time warping,piecewise linear interpolation,and random extraction,which effectively increase the number of unqualified samples.Secondly,a convolutional neural network is built to realize the real-time detection of the quality of the anchor chain flash welding.The nearest neighbor piecewise interpolation sampling proposed in this paper is compared with random under-sampling and random oversampling,and analyzed the impact of sample imbalance on model classification performance.To further improve the recognition ability of the model,the model is optimized from the aspect of the dropout rate.Thirdly,a large number of the anchor chain flash welding samples without labels are easy to obtain,but currently,they are mainly marked manually,which is inefficient and subjective.To this end,the model of detecting whether the human heart rhythm is abnormal is fine-tuned by the method of transfer learning,and the second transfer is realized on the four types of anchor chain with fewer samples.Transfer learning not only saves the time of model training but also makes up for the shortcomings of the number of labeled the anchor chain flash welding samples.Finally,in response to the problem that new samples are generated and the data is becoming more and more in the anchor chain flash welding,this paper uses incremental learning method to train the model,which only needs to learn the knowledge of new samples each time,and compares with the traditional batch learning.A 10-fold cross-validation method is used to train the model of the electrode-position signal,current signal,and combined signal,and the accuracy of incremental learning on old samples is also detected.