Fault Detection And Diagnosis Of Industrial Processes Based On Extreme Learning Machine

Fault diagnosis technology as a significant means to ensure the safety and reliability of industrial process,has received the extensive attention of scholars and professionals.Extreme learning machine(ELM),which is a supervised training method for the single-hidden layer feedforward neural network based on balanced data,has been applied in fault diagnosis filed because of its fast learning speed and good generalization performance.However,the existing ELM methods for fault diagnosis are all implemented by fault classification to determine the type of fault,lacking an effective fault detection method,so this paper proposes a fault detection method based on unsupervised kernels ELM.Further,considering the effect of data characteristic,including the lack of labeled data and the imbalance of the data,on building the fault diagnosis models,fault diagnosis methods based on semi-supervised ELM and selective ensemble ELM are studied in this paper.The simulation results on TE process verify the effectiveness of the proposed methods.The specific work is as follows:To deal with the poor performance in fault detection caused by the unsupervised ELM(UELM),which only concentrates on the local structure information of the observation data and the inherent randomness,a fault detection method based on unsupervised kernels ELM(UKELM)is proposed through the analysis and researches on data structure and kernel technology.By integrating the global structure analysis of data into the target function of UELM,and using the kernel mapping instead of the random feature mapping in hidden layer,UKELM can extract feature information more effectively and solve the instability problem of the output.Based on the data feature obtained by UKELM,the monitoring statistics can be established with the k-neighbor detection(FD-kNN)method,and then fault detection can be performed.The simulation experiments on TE process show that this method can not only detect the fault effectively but also improve the performance of the fault detection.To copy with the fault diagnosis problem in the industrial processes when the labeled data is insufficient because labeling data is difficult,a fault diagnosis method based on semi-supervised ELM is proposed by studying the ELM network structure and semisupervised learning theory.Firstly,the reconstruction based semi-supervised ELM(RSELM)uses the ELM auto-encoder(ELM-AE)to obtain the input weight of hidden layer,and then based on the principle of minimizing the reconstruction error,a reconstruction graph which can determine the number of nearest-neighbor adaptively is built in the feature space,and at the same time,the connection weight is optimized by using the category information of labeled data;Finally,a new objective function with local keep function is constructed in the output space to obtain the output weight,completing the training of fault classifier.The simulation results on standard data sets and TE process show that this method can achieve higher classification accuracy than traditional semi-supervised ELM methods,improving the precision of fault diagnosis.To deal with the fault diagnosis problem of industrial process when data is imbalanced caused by the inconsistent occurring frequency of fault,a fault diagnosis method based on imbalanced ELM is proposed by analyzing and studying the characteristics and the solutions of imbalanced data.The hybrid sampling based selective ensembel ELM(HSE-ELM)adopts the hybrid sampling strategy,which combines the random undersampling and SMOTE method,to obtain the balanced fault data subset used for training ELM base classifiers;Meanwhile considering the different classification performance of classifiers in different data fields,selective ensemble method is employed to select the proper classifiers for each sample under test to ensemble learning.The simulation experiments on TE process verify the effectiveness of the proposed method.