Research On Predictive Maintenance Of Key Components Of Mechanical System Based On Complex Imbalanced Data Classification Method

As the key components of the mechanical system,bearings and tools,etc.,run under high-load and high-strength conditions for a long time,and are prone to various failures or irreversible wear,which often lead to catastrophic accidents or huge economic losses.Therefore,predictive maintenance of key components(e.g.,bearings and tools)is the focus of enterprise operation and maintenance,and it is also a hot spot in academic research.At present,most of the data-driven predictive maintenance researches on key parts of mechanical systems are based balanced data.However,predictive maintenance monitoring of mechanical systems and the establishment of predictive models are typical imbalanced data classification and regression processes.Additionally,complex imbalance problems in industrial monitoring data(e.g.,noise,between-class imbalances and within-class imbalances,multi-class imbalances,and time series data imbalances,etc.),have brought huge challenges to the classification and regression modeling of key components of mechanical systems.This is because the standard classifiers are prone to failure in the face of imbalance data.To this end,based on the oversampling algorithm,this paper proposes some new algorithms to address noise problem under limited samples,between-class imbalance and within-class imbalance problem under limited samples,multi-class imbalance problem under limited samples,and time series data imbalance problem under limited samples.The main research contents of this paper are as follows:1)To address the noise problems under limited samples,an improving noise-immunity majority weighted minority oversampling technique(NI-MWMOTE)is proposed.It comprehensively considers the Euclidean distance and the density of the nearest neighbors as the probability of the "suspected" noise identified by the k-NN method as true noise for ranking;then,it adaptively selects the best noise processing strategy through iteration and misclassification errors;Finally,the AHC clustering algorithm is combined with the improved MWMOTE oversampling technology to deal with imbalanced data.To verify the effectiveness and superiority of NI-MWMOTE,it and LS-SVM(least squares support vector machine)classifier are combined to improve the condition monitoring performance of bearings and tools in scenarios with limited and noisy imbalanced data.The experimental results of public data sets and self-collected data show that the accuracy rates of NI-MWMOTE in the condition monitoring of the 2 types of key components are as high as 95.61% and 95.96%,respectively,which are 9.47% and 25.19% higher than before sampling.Obviously,The sampling effect is significantly stronger than other algorithms(e.g.,SMOTE and MWMOTE,etc.).2)To address the between-class imbalance and within-class imbalance problems under limited samples,an improving adaptive semi-supervised weighted oversampling(IA-SUWO)technique is proposed.IA-SUWO uses a semi-supervised hierarchical clustering method to segment minority instances in an iterative manner;then,it adaptively determines the sampling size of each minority sub-cluster through classification complexity and cross-validation;finally,comprehensively considers the average Euclidean dis-tance and the least squares support numerical spectrum to give weights to the minority boundary samples that are difficult to learn,and based on these weights,the k*INN(k*information nearest neighbors)-based method is used to synthesize more widely useful new minority examples.To verify the effectiveness and superiority of IA-SUWO,it and LS-SVM classifier are combined to improve the condition monitoring performance of bearings and tools in scenarios with limited and between-class imbalance and within-class imbalance data.The experimental results of public data sets and self-collected data show that the accuracy rates of IA-SUWO in the condition monitoring of the 2 types of key components are as high as 95.62% and 94.80%,respectively,which are 15.37% and 9.47%higher than before sampling.Obviously,The sampling effect is significantly stronger than other algorithms(e.g.,MWMOTE and A-SUWO,etc.).3)To address the multi-class imbalance problems under limited samples,a sample-characteristic oversampling technique(SCOTE)is proposed.SCOTE uses the "One-vs-All" strategy to convert multi-class imbalance problems into lots of binary imbalance problems.First,in each sub-problem,SCOTE uses the k-NN method to filter out the noise;then,the importance of the minority samples is ranked according to the least squares support numerical universal values;then,based on the weight sequence,a new method based on k*INN is used to synthesize a wide range of useful samples to balance the data set;finally,when all the binary sub-problems are solved,the multi-class imbalance problem is solved.To verify the effectiveness and superiority of SCOTE,it and LS-SVM classifier are combined to improve the condition monitoring performance of bearings and tools in scenarios with limited and multi-class imbalance data.The experimental results of public data sets and self-collected data show that the sampling effect of SCOTE in the condition monitoring of the 2 types of key components is significantly stronger than other sampling algorithms(e.g.,SMOTE,MWMOTE,A-SUWO,etc.)and ensemble algorithms(e.g.,SAMME,Fuzzy Imb ECOC,Im ECOC,etc.)suitable for multi-class imbalance classification monitoring.4)To address the problem of multi-class imbalance with limited image data and time series imbalance with limited data,a SCOTE oversampling technique with variable number of synthesized samples(ISCOTE)is proposed.ISCOTE introduces a scaling factor(S)to adjust the number of samples that need to be synthesized in each minority class,so that the classification and regression model can obtain the maximum amount of information without over-fitting.Finally,embedded ISCOTE into "VGG 16+Multi-class SVM(LS-SVM)" and "ICEEMDAN-Shannon energy entropy+SVR" respectively to improve the state monitoring and state prediction performance of key parts of mechanical systems in the non-eigenvector imbalance classification scenario.The experimental results of authoritative data sets and self-collected data show that the sampling(or data enhancement)effect of SCOTE in the condition monitoring and prediction of the 2 types of key components is significantly stronger than popular sampling algorithms(e.g.,SMOTE and MWMOTE,etc.).5)Integrating the above theoretical models,a set of predictive maintenance prototype system was developed.The system was applied to the predictive maintenance of the turning process of a manufacturing company in Guizhou,and realized real-time condition monitoring,remaining useful life(RUL)prediction and maintenance strategy reminders for the key part represented by tools.