Research On Process Monitoring And Quality Diagnosis Method Of Casting And Rolling Production Via Time Series Data

Steel manufacturing is an important basic industry in China.Continuous casting and rolling are the main processes in the steel industry.The process stability plays a decisive role in the internal quality and geometric dimensions of the final product.The product quality objection caused by the fluctuation of the production process brings huge losses to the enterprise.With the increasing requirements of product quality and the vigorous development of industrial big data,improving the stability and uniformity of product quality in casting and rolling process through data-driven process monitoring and quality diagnosis has become an important research topic.In order to improve the anomaly detection and evaluation ability and accuracy of fault monitoring and diagnosis in continuous casting and rolling processes,the anomaly interval detection of single batch,state evaluation during batch transition,multi-batch process monitoring and fault batch root cause diagnosis are deeply studied based on high-resolution time series data.According to the characteristics of nonlinearity,multimodality and strong coupling of casting and rolling process,multi-scale convolution and long short-term memory network,deep clustering network with time attention,tensor locality preserving projection algorithm integrating tensor subspace and chord kernel function,causal discovery algorithm based on skipping temporal convolution and threshold multi-head self-attention are proposed and applied to quality monitoring,evaluation and diagnosis.The main research contents and results are as follows:(1)An anomaly interval detection algorithm based on multi-scale convolution long short-term memory network(MCNN-LSTM)is proposed to realize online detection of multi-scale and multi-category abnormal intervals in process variables of continuous casting process.In order to accurately detect,classify and locate abnormal intervals with large scale differences,a scale transformation layer composed of identity transformation,time-domain down sampling and low-pass filtering is designed,which integrates multi-channel one-dimensional convolution and long short-term memory network for feature extraction and classification.First,the average error of 29 benchmark data sets based on the proposed method is 2.8%,which verifies that the performance of time series classification is better than the comparison methods.Second,the proposed method is applied to the continuous casting process for anomaly interval detection.The classification accuracy is improved by 1.3%via the proposed scale transformation layer,and the detection rate of the slab with abnormal intervals during the online monitoring stage is 94.4%.The experimental and application results show that the MCNN-LSTM network can accurately detect the abnormal-intervals of different types and scales such as pulse,step,short-term and long-term,which provides data support for the stability evaluation and the comprehensive quality analysis of continuous casting process.(2)A deep clustering network with temporal attention(ADCN)is proposed to improve the ability to evaluate the multi-modal transition state when the heating system is switched in the continuous annealing process.Aiming at the multi-modal characteristics of the transition state,a clustering model is established to identify the different modal divisions,then the membership of the transition state to the mode is evaluated by the silhouette coefficient,and finally the significant transition region is given by the temporal attention coefficient.In the clustering model,the feature reconstruction based on Gated Recursive Unit and pseudo label classification tasks are proposed to enhance the ability of feature extraction,and the clustering goal of k-means algorithm is improved by spectral relaxation.The RAND coefficient of clustering experiment via benchmark data is increased by 6.1%,6.4%and 10.2%respectively.The clustering performance of ADCN is better than the traditional methods,which verifies its effectiveness and superiority.The proposed method is applied to the evaluation of the transition state during the continuous annealing process,and the validity of the evaluation results is verified by the visual analysis of the annealing temperature,which improves the evaluation ability of the transition state,and is helpful for product quality rating and judgment.(3)Tensor locally preserving projection algorithm combining tensor subspace and chord kernel function(ICK-TLPP)is proposed to effectively describe the similarity between matrix data,avoid the structure damage of matrix data caused by vectorization expansion,and improve the fault detection rate in multi-batch process monitoring.In order to accurately and efficiently measure the similarity of matrix data with nonlinear relationship,an improved chord distance of fusing subspace eigenvalues is proposed.Matrix trace operation and parallel optimization algorithm are used to improve the accuracy and calculation efficiency of process monitoring model,so as to shorten the calculation time of chord distance by nearly 20 times.The proposed method is applied to the penicillin fermentation process,and the detection rates of five types of faults are 94%,75.5%,73%,100%,and 92.5%,respectively,indicating that the process monitoring capability is superior to traditional methods such as KPLL and tensor CP decomposition.Finally,it is applied to the process monitoring of hot-rolled strip head narrowing,and the fault detection rate is 95.24%,which is 16.7%higher than that of the process monitoring method based on tensor decomposition,indicating that the ICK-TLPP method can effectively improve the monitoring capability of batch data,which can be accurately and quickly applied to industrial production processes with large-scale data.(4)A root cause diagnosis algorithm based on skipping temporal convolution and threshold multi-head self-attention(MTCMS)is proposed to improve the accuracy of root cause diagnosis under multi-scale time delay and nonlinear causality.In order to accurately and quantitatively carry out causal discovery and fault root cause diagnosis,the temporal causality is discovered through prediction tasks,and fault diagnosis is carried out considering system inherent causality.First,the network structures such as skip connection and threshold multi-head selfattention are proposed to enhance the feature extraction ability of nonlinear causal relationship with multi-scale delay time,and the average F1 score of causal discovery in the chemical simulation data set increases by 13.3%and 12%,respectively.Then,a contrastive causal graph and the root node score calculation rule are proposed for root cause diagnosis.Benchmark experiments show that the root cause diagnosis accuracy of the MTCMS algorithm is better than the traditional algorithms such as TCDF and ENCO.Finally,it is applied to the root cause diagnosis of head narrowing during hot-rolled process.Combined with the typical failure mechanism and visual analysis,the results show that the MTCMS algorithm has high accuracy of diagnosis results,which can provide support for hot-rolled product quality improvement and process optimization.