Data Missing Aluminum Electrolysis Multi-fault Diagnosis For Industrial Fields

Aluminum electrolysis is the most important way to produce raw aluminum in China.Its tough production environment induces variety harmful faults.To improve efficiency and safety,it's necessary to develop aluminum electrolysis fault diagnosis technology.At present,there are many problems in aluminum electrolysis fault diagnosis and prediction research,such as short prediction time,low diagnosis accuracy,single target fault and single used feature.The industrial aluminum electrolysis production data is usually of low quality,which limits the development of the diagnosis system.To solve these problems,this thesis takes real aluminum electrolysis production data as research object and carries out researches from multi feature and multi model.Firstly,the experimental data set was constructed.To solve the problems of key information missing and poor data quality,data of anode effect state,noise state and normal state was selected and normalized.Linear spline interpolation and singular value thresholding were used to complete the time series data and data in cell respectively.Then,to obtain deep information of aluminum electrolysis data,this paper uses a variety of signal decomposition and feature extraction methods to process data,obtains the deep-seated information of the signal.Based on the features,classifiers were constructed respectively.Moreover,Multi-LSTM algorithm which can use both the sequence feature of time series data and the feature of data in cell was proposed in this thesis.It applies the sequence feature to aluminum electrolysis fault diagnosis task and therefore improve the diagnosis accuracy.Finally,stacking was used to intergrade three classifiers based on sequence feature,hybrid entropy feature and energy density feature.A multi-feature intergraded aluminum electrolysis multi-fault diagnosis and prediction model based on stacking was constructed.It further improved the diagnosis accuracy and the prediction result.The experimental analysis results show that sequence feature has a good effect in aluminum electrolysis fault diagnosis and prediction task,and the performance can be further improved by integrating multiple features.Multi-LSTM classifier and the Stacking based model can achieve 86.54% and 88.48% accuracy respectively in real-time diagnosis task.In a 10 minutes advance prediction task,the accuracies of above model can achieve 74.38% and 79.57% respectively.The experimental results show that the performance of diagnosis and prediction was significantly improved.