Dynamic Process Monitoring Method With Uneven Time Delays Based On Improved LSTM

Process monitoring is a key means to ensure safe,efficient and green manufacturing.With the rapid development of big data technology,data-driven process monitoring method has become the hot spot and frontier of process monitoring technology research.Due to the wide application of closed-loop controller,dynamic characteristic has become the most essential characteristic in industrial process.Hence,it is of great significance to conduct data modeling and monitoring for dynamic processes.Due to the uneven time delays and other complex disadvantageous factors in the actual dynamic processes,the effectiveness of model feature extraction is affected,which degrades the model performance.Therefore,this paper proposes two kinds of improved long short-term memory(LSTM)neural network models under the framework of deep learning to solve the problems existing in actual dynamic industrial processes.Aiming at the problem that complex dynamic features are difficult to extract adaptively due to the uneven time delays characteristic causing by the different autocorrelation of variables in dynamic process data,a multivariable-oriented method for dynamic process monitoring with uneven time delays is proposed in this paper.The size of moving windows for dynamic process time series can be quantitatively determined by introducing the autocorrelation coefficient function(ACF)and threshold judgment.Then the LSTM neural network structure and Autoencoder(AE)neural network structure are deeply integrated to improve the ability of adaptive feature extraction for complex dynamic features and other hidden features of the model by using the idea of data reconstruction.Aiming at the problem that it is difficult to extract quality-related features in dynamic process monitoring methods,which means that faults can only be found but cannot be judged whether faults are quality-related,this paper proposes a quality-related fault detection method for dynamic process with uneven time delays based on Quality-Driven LSTM and AE(QLSTM-AE)model.The dynamic feature is extracted adaptively by reconstructing training data.At the same time,the quality variables are embedded in parallel to extract quality-related features.After that,the quality-related fault detection is realized by the strategy of quality monitoring as the main and process reconstruction error monitoring as the auxiliary.Finally,the proposed method is applied to industrial process datasets.Compared with the traditional methods,LSTM-AE method can adaptively extract dynamic features with uneven time delays and improve the accuracy of monitoring.QLSTM-AE method can comprehensively extract quality-related features on the basis of LSTM-AE method and effectively realize quality-related fault detection.Figures(31),Tables(9),References(94).