Abnormal Pattern Recognition Of Binary Autocorrelation Processes In Imbalanced Data

In the quality control of complex manufacturing processes,the simultaneous monitoring of multiple quality characteristic variables that exist with both inter-correlations and autocorrelations is defined as a multivariate autocorrelated process.In recent years,the anomaly detection of multivariate autocorrelated processes introduced by machine learning techniques has shown more advantages than traditional statistical methods.Due to the high cost of obtaining abnormal pattern data in actual production,this field usually conducts theoretical research based on synthetic simulation datasets,often ignoring the class imbalance characteristics that can usually be obtained in actual production data.Therefore,this thesis proposes research on the anomaly pattern recognition of binary autocorrelated processes under data imbalance conditions,introducing cost-sensitive deep learning algorithms and multi-layer cost-sensitive classification structures to construct more effective anomaly detection models for binary autocorrelated processes.The main research work of this thesis is as follows:(1)A cost-sensitive binary autocorrelated process pattern recognition model is proposed under data imbalance conditions.First,two methods,direct and decomposition,are studied to construct cost-sensitive multi-classification models.Then,the Particle Swarm Optimization(PSO)algorithm is introduced to search for the optimal cost parameters,and PSO optimization algorithms for both direct and decomposition methods are designed.The impact of the two typical multi-classification performance indicators,G-mean and F1-score,as PSO optimization fitness values on model performance is studied,and 1D-CNN,Long Short-Term Memory(LSTM),Random Forest(RF),and Support Vector Machine(SVM)are compared as base classifiers.Experimental results show that the model with cost-sensitive LSTM as the base classifier and G-mean as the fitness value driven by PSO optimization of cost parameters has advantages over other models.(2)A model for synchronous recognition of binary autocorrelated process anomaly patterns and anomaly amplitudes under conditions of imbalanced data for both pattern type data and anomaly amplitude data is proposed.First,the modeling method and performance evaluation indicators under the hierarchical classification concept are explained.On this basis,a cost-sensitive hierarchical multi-classification model,MMCS-LSTMPSO,is proposed.Then,comparative experiments are conducted on a Monte Carlo simulation dataset to validate the performance of the model,including the performance of anomaly pattern recognition and anomaly amplitude recognition under offline operation,as well as the comprehensive performance verification as a hierarchical classification model.In addition,two online detection performance indicators,average running chain length and dynamic recognition accuracy,are determined based on simulated data streams.Experimental results show that the proposed MMCS-LSTMPSO model has advantages over comparison models.(3)Based on the theoretical research results of this thesis,an application prototype system is analyzed,designed,and developed according to software engineering processes to explore the application approach of the proposed model.The system supports functions such as data collection and management,model cost parameter optimization,model training,and anomaly recognition through model loading,centered around the requirements of anomaly pattern and anomaly amplitude recognition in binary autocorrelated processes.The system is tested and evaluated.Theoretical research results show that the proposed model can effectively solve the problem of binary autocorrelated process anomaly recognition under data imbalance conditions,and the developed application prototype system demonstrates its feasibility for application.