Research Of SAE-MDLPP Algorithm And Its Application To Industrial Process Fault Diagnosis

As the scale and level of automation in modern industrial production processes continue to rise,the probability of production process failures increases.This makes it particularly important to design effective fault diagnosis techniques to ensure industrial production safety.The development of sensor technology and computer storage technology has enabled the recording and preservation of historical data for industrial production processes.Modern industrial production data is often high-dimensional,non-linear,and strongly coupled.Extracting key characteristic information from this process data is crucial for improving the accuracy of fault diagnosis.In response to these challenges,this paper focuses on the following main research elements.(1)The traditional Discriminant Locality Preserving Projection(DLPP)method,compared to the Locality Preserving Projection(LPP)method,uses Euclidean distances to determine whether points are near neighbors when constructing intra-class and inter-class weight matrices.Although both inter-class and intra-class discriminative information is considered,the Euclidean distance is easily affected by data distribution,high dimensionality,and noise,and does not account for differences between variables.To address this issue,this paper proposes a Discriminant Locality Preserving Projection method based on the Marxian distance(MDLPP),which selects the Marxian distance instead of the Euclidean distance in the construction of the intra-class weight matrix and inter-class weight matrix to determine whether the points are near-neighbor points,the Marxist distance can effectively avoid the effects due to the data dimension and correlation between variables,and the proposed MDLPP method can better extract the information of local features of the data.(2)Considering that industrial process data exhibits high dimensionality,nonlinearity,and strong coupling,extracting effective feature information from this process data is a key step to achieve fault diagnosis.To address this issue,this paper proposes a fault diagnosis method based on the combination of Sparse Autoencoder and improved Discriminant Locality Preserving Projection(SAE-MDLPP).This method considers both the global and local structure information of the data during the feature extraction stage and effectively ensures the results of data feature extraction.The Akaike information criterion is also used to determine the optimal dimensionality reduction order for the model.Finally,the integrated learning Ada Boost classification algorithm is employed to classify the faulty data after feature extraction,aiming to improve the fault diagnosis accuracy of the proposed method.(3)To verify the effectiveness of the proposed SAE-MDLPP fault diagnosis method,we experimentally test it using the Three-phase Flow Facility(TFF)case and the Tennessee-Eastman(TE)process case,and compare it with other common fault diagnosis methods.The experimental results demonstrate that the SAE-MDLPP-based fault diagnosis method proposed in this paper offers higher diagnostic accuracy.