Research On Intelligent Fault Diagnosis Method Of Rolling Mill Based On Convolutional Neural Network Framework Under Complex Working Conditions

The status of iron and steel industry in social construction and economic development continues to rise,and rolling mill equipment gradually tends to be larger,more sophisticated and automated,so the safety of the rolling mill system has put forward higher demands,and it becomes more and more indispensable to carry out fault diagnosis for the rolling mill,but the rolling mill production environment is complex,in which the long time stable rolling production activities lead to serious class imbalance phenomenon in the collected data,and different rolling speeds required by different production operations lead to deviation of data distribution.These problems bring challenges to traditional fault diagnosis methods,so it is urgent to put forward practical and intelligent fault diagnosis methods to solve the problem of rolling mill fault diagnosis under complex working conditions.In this paper,the convolutional neural network is used as the main framework,the rolling mill equipment condition monitoring method under complex working conditions is taken as the main line,and the research is gradually carried out from the aspects of feature extraction,fault diagnosis model establishment and performance evaluation.The main research contents are as follows:Firstly,a dual attention-guided feature enhancement network is proposed to solve the problem of different data amounts in different fault categories of rolling mill.To obtain the time-frequency images,wavelet transform is used to extract the time-frequency feature information of the data sample.Then,the separable convolution layer and pooling layer are stacked to extract the fault feature adaptively from the time-frequency image.Then,a feature enhancement module is designed to integrate and enhance the current feature representation of the convolution layer by using the attention mechanism.It improves the ability of network model structure to judge and control the trend of fault information,enhances the relevance of information sharing,and adopts the strategy of global average pooling and multiple learning rate variation,which not only reduces the calculation amount in the process of network operation,but also prevents and controls the risk of overfitting in advance.Different datasets are designed and collected from rolling mill experimental system,the experimental results show that this method can achieve efficient diagnostic performance under both data balance and data imbalance conditions.Secondly,a multi-scale feature adaptive alignment network is proposed to solve the problem of cross-domain fault diagnosis of rolling mill under variable speed conditions and the difficulty of acquiring label information in actual industry.Firstly,One-dimensional convolutional neural network is directly used to extract the fault features in the source and target domains,which reduces the work of data preprocessing.Then,the multi-level learning structure is designed to automatically learn the high-level expressions at different scale levels,and then LMMD is used as the subdomain distribution loss term of the source domain feature distribution and the target domain feature distribution at each scale respectively,while a square class confusion loss function is proposed to reduce the classification divergence in the classification layer.Experiments on laboratory rolling mill datasets and public datasets prove that the proposed method can achieve excellent fault diagnosis performance under cross-domain unlabeled data.Finally,aiming at the coupling problem of data imbalance and variable speed conditions,a rolling mill fault diagnosis method based on multi-representation feature enhancement alignment network is proposed.Firstly,the multi-scale fault feature information is extracted from the input signal.Then uses the feature enhancement module to increase the reference information of fault detection.Then,the LMMD is used to align the distribution on the corresponding scale of the enhanced features,and self-constraint is introduced to reduce the bias caused by pseudo-labeling.Finally,the classification loss,distribution loss and class confusion loss are combined to adjust the model parameters.The experimental results show that the method can achieve high fault recognition rates under the condition of variable working conditions and data imbalance.In summary,different fault diagnosis models are designed in this paper to solve the problem of rolling mill fault diagnosis under different complex working conditions,and the experiment proves that the diagnosis performance is better than other methods,and has a certain practical engineering application value.