Research On Abnormal Vibration Early Warning And Diagnosis Method Of Strip Mill Based On Unsupervised Learning

As the main production equipment for strip and sheet,the rolling mill often experiences abnormal vibrations during the manufacturing process due to the harsh and complex production environment.These vibrations can affect the surface quality of the products,and in severe cases,damage the equipment and reduce production efficiency.Therefore,accurately monitoring and identifying the current operating state of the rolling mill in real-time,and quickly diagnosing abnormal vibrations based on their vibration characteristics can greatly reduce their negative impact.In this paper,based on unsupervised learning,three methods for abnormal vibration warning and diagnosis of rolling mills are proposed.Firstly,the basic principles and training process of unsupervised learning are elaborated,and the structures and training processes of the three unsupervised learning models used in this paper,namely autoencoder,restricted Boltzmann machine,and Transformer,are described and derived in detail.Secondly,to address the problem that it is difficult to accurately detect early-stage abnormal vibrations in rolling mills,a rolling mill chatter warning method based on sparse autoencoder and self-organizing map network is proposed.By fusing information from the original rolling mill vibration signals using a sparse autoencoder and self-organizing map network,the key information characterizing the rolling mill vibration state is obtained.Based on the fusion results,a feature indicator for rolling mill vibration state is constructed,and a reasonable threshold is set using the 3σ rule to monitor the rolling mill vibration state in real-time.Experimental results demonstrate that this method can monitor the rolling mill’s vibration state in real-time and achieve early warning of chatter vibration.Next,to address the problem of difficulty in accurately diagnosing vibration types in complex environments,a rolling mill vibration type monitoring method based on deep belief network and improved sparrow optimization algorithm is proposed.Using wavelet packet energy entropy as the signal feature,noise and other useless information are filtered out.Multiple sensor data are simultaneously input into a deep belief network for fusion diagnosis,and the improved sparrow optimization algorithm is used to optimize the parameters,further improving the diagnostic accuracy of the deep belief network.Experimental results demonstrate that this method can effectively diagnose vibration types and has higher accuracy and dynamic diagnostic capabilities compared to other methods.Finally,to address the problem that rolling mill component faults are difficult to detect,leading to abnormal vibrations,a rolling mill component fault detection method based on time-frequency masked autoencoder is proposed.Synchronous compression wavelet transform is used to transform the rolling mill bearing vibration signals into time-frequency graphs.Based on Transformer,a time-frequency masked autoencoder is constructed and trained to extract hidden fault feature information of the rolling mill bearings,and fault diagnosis is performed based on the features.Experimental results demonstrate that this method can effectively extract rolling mill bearing feature information,and has higher accuracy and robustness compared to other methods.