Research On Intelligent Diagnosis Methods Of Rolling Mill Health Conditions Based On Non-Completed Datasets

The rolling and production of high-strength,high-performance,high-precision steel strip mainly depends on the cooperation and smooth operation of all types of rolling mills in the production line.However,with the continuous increase in rolling speed and rolling intensity,the rolling mill equipment is prone to frequent health state failure failures during the production rolling process.At the same time,it is particularly important to solve the problem of non-complete data sets because the mill equipment is prone to non-complete data sets under actual working conditions,which can easily cause the performance of traditional health state diagnosis models to collapse.In this paper,three new methods are proposed for monitoring and diagnosing the health status of mill units under three typical cases of non-complete data sets,and then three typical cases of non-complete data sets are designed on the mill experimental platform according to the actual conditions of the mill equipment,and finally the superiority of the proposed methods in solving these three cases is demonstrated by experimental data.First,the basic concepts of deep learning,migration learning and non-complete datasets are introduced,followed by three models of the underlying health monitoring and diagnosis models used in this paper,namely,convolutional neural networks,deep confidence networks and graph neural networks.Convolutional neural networks use convolution and pooling operations to extract deep features from the raw data,and then realize the health monitoring and diagnosis of mechanical equipment.Deep confidence network is a mathematical model that combines supervised and unsupervised learning to extract features from the raw data layer by layer to output the health status of the target object.Graph neural networks,on the other hand,improve their monitoring and diagnostic performance by learning the graph topology on non-Euclidean spatial data and its node features to capture the interdependencies between different health states of mechanical equipment.Second,when the traditional deep learning diagnosis model of rolling mill equipment faces the unbalanced data set under non-complete data set,the inconsistency of its health state ratio easily leads to the reduction of the generalization ability of the model and thus causes the breakdown of the diagnosis performance.A method of rolling mill equipment health state diagnosis under unbalanced data set based on improved convolutional neural network is proposed.Firstly,the method applies group normalization to both 1D and 2D convolutional neural networks to improve the robustness of the model.Immediately afterwards,both 1D and 2D convolutional neural networks use global average pooling instead of the traditional fully connected layer to improve their spatial feature extraction and generalization ability.Immediately after that,the acquired vibration signals are converted into two-dimensional spectral cliff images by the fast-spectral cliff method and input into the improved two-dimensional convolutional neural network,while the acquired sound signals are input into the improved one-dimensional convolutional neural network.After that,the extracted feature vectors are spliced and fused to achieve the monitoring and diagnosis of mill equipment health status.Finally,experiments are designed and tested on the mill experimental platform with different proportions of unbalanced datasets for the proposed method to evaluate its performance in tolerating unbalanced datasets.Then,for the problems such as overfitting easily due to the inadequate training of traditional deep learning diagnostic models for rolling mill equipment when finite datasets under non-complete datasets,an improved single-sensor deep confidence network is proposed,which combines previously hidden features and visible features as the next stage input,and thus repeats the execution to fully exploit potentially valuable information in finite datasets and alleviate the model’s The problem of information loss during layer-by-layer feature extraction is alleviated.After that,a multi-sensor improved deep confidence network is proposed on top of the improved single-sensor deep confidence network to obtain complementary and rich health status features from different sensors to effectively handle the health status monitoring and diagnosis tasks with limited sample data sets.Finally,due to the problem of cross-domain dataset under non-complete dataset due to the variable working conditions of mill equipment in real working environment,a multi-source adversarial graphical convolutional network based on local maximum mean difference index is proposed to achieve cross-domain health condition monitoring and diagnosis of mill equipment through domain adaptation of mill equipment with multiple source domains.Firstly,a shared feature extraction module is used to perform feature extraction for the multi-source domain and the target domain.After that,the feature data in Euclidean space are converted into non-Euclidean space graphs and the graphs are modeled using graph convolutional networks to achieve health state monitoring and diagnosis of rolling mill equipment under cross-domain data sets.Finally,experiments are designed and the feasibility of the proposed model is verified in the mill experimental platform.