Multimode Industrial Process Modeling And Monitoring Based On Statistical Machine Learning

With the change of the external environment or internal properties,many industrial processes are characterized by multiple running conditions,i.e.,multimode process.Statistical machine learning methods have yielded abundant research results in general process monitoring,however,there remains the challenge in multimode process modeling and monitoring due to multimode processes accompanied by nonlinear,non-Gaussian and time-varying dynamic characteristics.At present,although many research results on steady modes have been achieved,there are still problems of insufficient feature extraction,low state identification and fault detection accuracy;the research of transition mode is relatively scarce,and the dynamic time-varying characteristics make it difficult in mode identification,modeling and monitoring;the theoretical research of the full-flow multimode（containing multiple steady and transition modes）is in the preliminary stage of development;in addition,operational faults are an equally important source of faults in the actual process compared to running faults,but they have not been studied sufficiently.To solve the above problems,this dissertation makes in-depth research on multimode stage division,feature mining,model construction,monitoring strategy optimization,operational fault analysis,monitoring accuracy and algorithm robustness improvement,the details are as follows:For nonlinear multiple steady modes,a modeling and monitoring strategy based on statistics pattern analysis and improved weighted support vector description is proposed,to solve the problem of low fault detection accuracy due to insufficient feature extraction and poor standardization strategy in the single-model based methods.The strategy first maps the original data into multiple statistics patterns space to fully exploit the potential features of different modes.Then,a new weight factor（i.e.standardization strategy）is constructed for the traditional weighted support vector description algorithm by considering the local spatial and temporal information,which not only normalizes the multimode data（i.e.removes the multimodality of process data）,but also reduces the influence of outliers on the control limits.The proposed method improves the fault detection accuracy,while effectively enhances the robustness of the model.Industrial multimode datasets are usually high in dimensionality and there is a mixture of normal/faulty data in different modes.To address the resulting problem of not being able to automatically and accurately identify mode/health states,a state identification framework based on sparse representation and modified density peak clustering is proposed.First,a sparse representation-based key variable selection strategy with_1,2norm is designed to eliminate redundant variables with high correlation before clustering.Then,a new distance measurement method with a time factor is proposed for density peak clustering to improve the separability of similar states,so that clustering centers with similar states can all stand out.Further,the sum of squared error-based method is developed to determine the optimal number of clusters and clustering centers,which truly enables automatic state identification.Finally,considering the problem that the mode attributes may cover the fault attributes,a two-step strategy of“Coarse division-Fine division”is designed to further improve the accuracy of state identification.The time-varying characteristics of transition modes will lead to difficulties in modeling and monitoring,and operational faults are prone to occur in this process,but the relevant research is insufficient.To solve these problems,a trajectory-based transition mode identification and operation anomaly monitoring strategy is proposed.Firstly,the slowest slow factor reflecting the inherent evolution law of the whole process is constructed,which can accurately judge the starting point and ending point of the transition mode,and eventually realize mode identification.Then,the“position”and“velocity”test statistics representing the operating trajectory are designed to constitute the process monitoring method.The method can not only determine whether a fault occurs,but also infer that the fault trajectory is more aggressive than normal or atrophy.At the same time,8 kinds of operational faults in multimode process are summarized and categorized according to the theory of hazard and operability analysis and the actual process characteristics,and the operational dataset of transition mode is generated in the case studies and used as the monitored object to verify the effectiveness of the proposed method.In addition,the nature of the catastrophic fault is deeply analyzed,and new indicators are exploited to judge the practical application value of the proposed method,which provides a reference for the subsequent related research works.In the full-flow multimode process,the stationary and non-stationary characteristics will appear alternately under multiple operations.To address the resulting problem of modeling and monitoring,a generalized monitoring strategy is proposed.The strategy first constructs two indicators based on variable correlations and spatial distance,that not only classify stages but also identify repetitive stages.Then,in each identified stage,all non-stationary variables with different integrated orders are handled by the proposed cointegration analysis and detrended fluctuation regression method,which ensures that process features are sufficiently extracted.Furthermore,a comprehensive similarity index is designed for online stage identification and a local learning strategy is used for real-time monitoring to obtain a compact control range.The proposed method uses a unified way to model and monitor each stage of the full-flow multimode process,and ensures the accuracy of fault detection.The above-proposed methods are applied to numerical cases,classical Tennessee-Eastman process simulation,extended Tennessee-Eastman process simulation,wastewater treatment process,penicillin fermentation process and actual semiconductor etching process.Compared with the existing methods,the monitoring results demonstrate the effectiveness and superiority of methods in the dissertation.