Research On Online Detection Method For Abnormalities During Continuous Casting Based On Machine Learning

During the development of high-speed continuous casting technology,the high steel flow caused a significant increase in mold thermal load,and various defects and abnormalities such as bulging,longitudinal crack,and breakout caused by high-load casting appeared endlessly,which has become a bottleneck factor that affects casting smooth and slab quality.The development of high-efficiency continuous casting technology and the demand for high-quality slab put forward higher requirements for slab quality and process control,there is an urgent need to develop appropriate detection technology and integrated equipment.With the rapid increase in the system structure,instrumentation and data volume of continuous casting,it has become more and more difficult to monitor the production process with traditional modelling methods.Therefore,it is necessary to introduce a data value mining method based on machine learning to find suitable coping strategies and means for controlling complex continuous casting process.This work mainly focuses on the detection and prediction approaches of slab bulging,longitudinal crack and sticking breakout.The methods of feature selection,dimension reduction,classification and clustering are comprehensively used to explore the algorithm combination for specific abnormalities so as to provide support for the development of intelligent monitoring technology based on machine learning.The main research contents include:(1)Research on prediction method for slab bulge’s position and deformation.HilbertHuang Transform is applied to investigate the periodic fluctuation rules and characteristics of mold level.The difference between the decomposition/transformation results and Hilbert spectrum characteristics of mold level under different casting conditions is analyzed.When slab bulging occurs,the amplitude of the fluctuation component of mold level is 5-6 times than that of normal casting condition,and the values of the Hilbert amplitude spectrum,energy spectrum and marginal spectrum are 3-10 than times that of normal casting condition.By extracting and grasping the correspondence between the frequency of Hilbert marginal spectrum,casting speed and roll pitch,the bulging frequency was determined.The bulging component is separated from the mold level and its fluctuation amplitude is obtained.According to the equivalent relationship between the volume change of mold level fluctuation and that of liquid core cavity caused by the bulging,a prediction method for bulge’s position and deformation based on Hilbert-Huang Transformation is proposed.The prediction results of bulge’s positioning and deformation are verified.Using the proposed method to predict bulging instances in production,the frequency of bulging is between 0.043 Hz and 0.056 Hz,the amplitude of bulged-component is between 3.21 mm and 5.85 mm.The bulging deformation is between 0.123 mm and 0.201 mm,which shows an increasing trend with the increase of slab width and casting speed.(2)Research on prediction method for longitudinal cracks.Aiming at the one-dimensional propagation characteristics of longitudinal facial cracks in the process of formation and propagation,this paper extracts 17 relevant features that can characterize the‘falling-rising’change trend of longitudinal crack temperature.Principal component analysis is used to reduce the features from 17 dimensions to 6 dimensions to remove redundant information and compress the input parameters of detection model.Support vector machine algorithm is used to train the principal components of temperature features after dimension reduction.Then,the separation hyper-plane and classification decision function are obtained,which can distinguish and identify the temperature patterns of different casting conditions.Based on the above work,a longitudinal facial crack detection model based on principal component analysis and support vector machine is developed.The sample set composed of 50 normal casting conditions and 50 longitudinal crack temperature is tested,the average accuracy of the model is 92%～96%,and all the longitudinal crack samples can be detected when the accuracy reaches 96%.(3)Research on prediction method of sticking breakout based on density clustering.Processing the sticking breakout temperature by the method of temperature rate,difference,and standardization,the ’rising-falling’ trend,’time lag’ and ’inversion’ characteristics of the temperature are extracted to obtain the timing temperature pre-processing sample.After that,dynamic time warping is used to measure distances of the timing temperature samples,and density clustering algorithm is used to cluster the timing temperature sample sets.The temperature modes under different casting conditions are distinguished and identified according to the clustering results,and a sticking breakout prediction model based on dynamic time warping and density clustering is developed.The test results show that all the breakout samples are gathered into one cluster,and the model can correctly distinguish the samples of breakout and normal conditions,which preliminarily confirms the feasibility of applying the clustering method to the prediction of sticking breakout.(4)Research on the prediction method of sticking breakout based on K-Means clustering.Aiming at temperature linkage changes of thermocouples in the same column during longitudinal propagation and adjacent columns during transverse propagation and the ’V’shaped propagation characteristic,9 main features including temperature,temperature rate and time lag are extracted.Using feature selection and correlation analysis,the typical feature vector(TFV)that can characterize the temperature change trend in the same and adjacent columns is filtrated and constructed.After that,K-Means clustering algorithm is used to cluster the TFVs of sticking breakout and normal casting conditions to obtain the region where breakouts’ TFVs are of concentrated distribution.On this basis,a sticking breakout prediction method based on feature selection and K-Means clustering algorithm is proposed.Using the proposed method to test and analyze the historical data of a steel plant in the past three years,the model’s report rate of sticking breakout is 100%,and the number of false alarms is greatly reduced compared with foreign developed system,showing good application potential.