Nomaly Detection And Sequence Status Analysis In Wleding Process Based On Deep Learning

A large seam welder is a key piece of equipment for the continuous operation of a steel mill production line,and its reliable and smooth operation is an important factor in ensuring efficient production in a steel mill.Due to the complexity of the welding machine system,the monitoring of the welding process of the welding machine in actual production still needs to be manned around the clock to keep an eye on the welding process of the welding machine,which is time-consuming and labor-intensive.Therefore,in order to save labor and improve the efficiency of the welding machine welding process state abnormal detection,it is necessary to implement automatic monitoring of the welding process signal.Traditional detection means using modeling and machine learning methods,these methods require mathematical modeling of the working mechanism of the welding machine system,simulation and manual extraction of features,more complex,while the current rise of deep learning can be directly from the data,independent mining data internal information,without excessive human intervention.For this reason,this paper uses deep learning methods to carry out research related to the automatic detection of abnormal state of the welding machine signal.Due to the limited a priori knowledge of abnormalities in factory production and the scarcity of abnormal data,it is difficult to research the identification and diagnosis of abnormal types of welding process signals,so this paper focuses on the detection of abnormal state of the welding process signal,that is,to determine whether the signal is abnormal.The main contents of this paper and related work are as follows.(1)To achieve automatic detection of welding machine welding data anomalies,three deep learning-based anomaly detection methods that perform better in the field of anomaly detection-the EGBAD model,the DCAE model and the DSVDD model-are applied to the welding machine welding process signal anomaly detection problem studied in this paper.First determine the detection object is the temperature and pressure signals generated by the welding process,and then build three network models,using Bayesian optimization method to optimize the hyperparameters of each of the three models to achieve the best detection performance of their respective models,in order to prevent the number of samples is too small to divide the validation set and the training set have too much deviation,which leads to the model to anomaly detection accuracy evaluation bias,using a five-fold cross The average AUC value is used as the evaluation index to test the performance of the models.The results show that the DCAE model is simple to build,fast training and high detection accuracy,so the DCAE model was selected as the basic method for signal anomaly detection in the welding process of the welder.In order to further determine the location of the defect of the abnormal signal,a method based on the Otsu method of the DCAE model reconstruction error matrix for the division of abnormal regions is proposed,and its division results can assist maintenance personnel to repair the equipment.(2)In order to solve the problem of the detection accuracy of the trained model to be further improved due to the small sample size of the welding process signal data,this paper proposes two methods of model migration:a model migration method based on data generation and a model migration method based on the domain variability metric.The data generation-based model migration method uses a confrontation generation method to make the welding process data in the target domain closer to the welding data distribution in the source domain,improving the similarity between the target domain and the normal data in the source domain,so that the detection model with stronger generalization capability trained by the source domain data samples can be effectively applied to the target domain welding data anomaly detection task,avoiding the target domain samples due to The model overfitting problem caused by fewer samples in the target domain is avoided.The model migration method based on the domain variability metric expands the difference between abnormal and normal data by introducing a domain variability metric loss in the objective function of the original DCAE model,and expands the difference in data distribution between the normal data of each welding specification in the training set,thus further improving the generalization ability of the abnormality detection model,so that under the condition that no data of the specification to be detected are involved in the training The two methods were compared with the original DC-DC method.Finally,the two methods were compared with the original DCAE model for testing,proving the effectiveness of the proposed two methods.(3)In order to identify the trend of progressive deterioration of the welding process over time as well as the dynamic adjustment of the anomaly detection threshold,the state of the time series consisting of the abnormal fraction of the welding process signal data is analyzed,using Bayesian variable point detection method for the time series consisting of the abnormal fraction of the welding process data is divided into states,and proposed a method for identifying the deterioration trend based on MK detection and the method based on Shapiro-Wilk smoothness detection,the feasibility and effectiveness of the method was verified through experiments.(4)Designed and developed a welding machine welding process signal monitoring system,first of all,the system’s function and structure of modular design,including data acquisition,data analysis and data management of the three major functional modules,and then the paper proposed based on DCAE anomaly detection method and timing analysis method for encapsulation,the development of the system’s function of each module.