Research On The Key Technology Of The Intelligent Monitoring System For On-line Roll Profiler Of Hot Rolling Mill

The hot rolling on-line roll profiler is generally used in the finishing mill stand of the hot rolling production line.It is a precision equipment for on-line grinding of the working roll in the rolling mill.The reliability of its operation is the basis for ensuring the continuous,stable and efficient work of the hot rolling production line.However,the on-line roll profiler is in the harsh environment of high temperature,electromagnetic interference,water vapor coverage,continuous vibration of rolling mill frame,narrow working space and so on.These factors make it difficult to observe the working state of it directly,which is very easy to cause unexpected failures,poor grinding quality and other adverse effects.If these adverse effects are not stopped in time,they will cause serious production quality problems.In order to solve the above problems and eliminate the "unstable" factors in the working process of the on-line roll profiler,this paper proposes to build its intelligent monitoring system from the three aspects of realizing the perceptible,visualized and predictable working state information of it.This paper focuses on the key technical links(working state information perception and state trend prediction and evaluation of key components)of the construction of the intelligent monitoring system to carry out relevant work research.Based on the research project of "hot rolling on-line roll profiler technology research" of the Central Research Institute(Technology Center)of Baoshan Iron and Steel Co.,Ltd.,this paper mainly does the following work:(1)Based on the analysis of the working principle and characteristics of the on-line roll profiler,this paper summarizes the current methods and methods of condition monitoring and fault prediction of mechanical equipment,and summarizes the work done by scholars at home and abroad in signal feature extraction and condition monitoring system development.This paper focuses on the classification and introduction of fault prediction methods,summarizes 29 kinds of fault prediction tools and methods,and evaluates them respectively.Combined with the development trend of fault prediction technology,the multi parameter and multi model fusion method is proposed to be applied to the residual life prediction and health evaluation of key components of on-line roll profiler.(2)This paper analyzes the causes of the failure of the on-line roll profiler from multiple perspectives,puts forward the overall framework of the on-line roll profiler intelligent monitoring system and determines its data collection scheme,and selects the hardware required by the intelligent monitoring system based on the requirements of the overall framework.(3)In order to grasp the signal features collected by the monitoring system as much as possible,extract the time-domain,frequency-domain and wavelet packet decomposition node energy features of the collected signals respectively.In view of the limitation of using single index feature to reduce the dimension,this paper puts forward the method of combining Pearson correlation coefficient and MIC information coefficient to reduce the dimension of the extracted multi features.Finally,through the data of PRONOSTIA experimental platform,it tests the validity of this method is proved.(4)The rolling bearing is the key part of the grinding head spindle of the on-line roll profiler,the safe and reliable operation of rolling bearing directly affects the working condition of the mill.In view of the limited prediction results of bearing residual life prediction,a random forest model with multiple deterioration characteristics and stronger anti noise ability is proposed to predict bearing residual life.based on the multi feature extraction and dimensionality reduction methods in Chapter 3,the bearing features are preliminarily screened,and the bearing degradation state is divided according to the RMS value of bearing vibration signal.The research object is further focused on the slowly deteriorating bearing.According to the difference of the degradation characteristic signals of the same type of bearing under the same working condition,the relative feature method is adopted to make the signal trend close,the S-G filtering method is used to smooth and remove the noise of the degraded features of the preliminary screening,and the consistency of the preliminary screening features of the same type and the same working condition bearings after the above processing is compared to filter out the deterioration features of the "tightness" is not strong.Finally,the PRONOSTIA platform rolling bearing life cycle data is used to carry out the key parameters of the random forest in a single data set,based on these parameters the performance of SVR,Lasso and CART tree regression is further compared in multiple data sets.The results show that the random forest has better accuracy in predicting the remaining life of rolling bearing.(5)The abnormal wear of the grinding head tool of the on-line roll profiler directly affects the quality of the product,and the effective wear state evaluation is an important means to avoid accidents.In view of the limitation of single model algorithm,a multi model double-layer fusion method based on stacking algorithm framework is proposed,in this model,GBDT,SVM and KNN are used as the first layer learners,and BP neural network is used as the second layer learners,finally,the final evaluation results are given by combination prediction.Based on the open data set of tool wear,the threshold filtering method and truncation method are used to clean the data acquisition signals,the method proposed in Chapter 3 is applied to feature extraction and dimensionality reduction,and the 10 times 10 folds cross validation method is used to divide the data.In view of the large number of model parameters in the multi model fusion method,the model parameters to be optimized are discretized and optimized by grid search method,finally,the fusion model in this paper is compared with 8 kinds of single model algorithms and 5 kinds of fusion model algorithms,which shows that the fusion model in this paper has higher accuracy and stronger stability.