The Transfer Entropy-based Causality Analysis Of Operating Failures In The Industrial Processes

Operating failures refer to the typical abnormal operating state of complex industrial processes during the normal manipulations,which are mainly caused by significant process disturbances and transferred between equipment units through the channel of material and energy related process parameters.Generally,four types of faults are included in the operating failures,including the oscillation,short disturbance,incipient fault and multiple operation modes.Due to the propagation characteristics of operating failures,the safety,smothness and optimization of industrial processes could be affected significantly.In this context,the timely and efficient identification,detection and root tracing of operating failures are of great significance and importance.Transfer entropy(TE)is an effective causality analysis method regarding the propagation characteristics of industrial operating failures.However,considering the complex dynamic timing features of production processes and the diversity of operating fault types,conventional TE algorithms are rarely encouraged for applications in practical industrial processes.In addition,multiple parameters are involved in the TE estimations,which causes the poor performance of real-time fault detections and low accuracy of causal relationship extractions.Considering the complicated characteristics of complex industrial processes,this thesis conducts an in-depth research on the TE-based causal feature analysis along with extraction methods for operating failures.The causal inference topology is constructed accordingly for timely and accurate root positioning and propagation path identification of operating failures.The TEbased causality analysis methods for oscillations,short disturbances,incipient faults and multiple operation mode conditions are proposed,respectively.The relevant parameters could be adjusted adaptively according to the unique timing features and process dynamics of different faults.This contribution effectively improves the accuracy,reliability and practicability of the TE-based causal analysis method for operating failure monitoring.The innovative achievements are presented as follows.1.In response to time delays of oscillation propagations,a time domain adaptive transfer entropy(TDATE)is proposed.Based on the periodic characteristics of oscillations,the oscillation detection and period extraction methods are given.By analyzing the relationship between the dynamic characteristic parameters of the process and the time-domain parameters of the transfer entropy,the parameter tuning ranges are suggested.The significance test method of the causal features is clarified.The accuracy of the transfer entropy algorithm for causal feature extraction of oscillating faults is improved.2.To deal with the problem that the short duration of important data districts are generally neglected in the short disturbance fault diagnosis,an attention transfer entropy(ATE)algorithm is proposed.The industrial process expert knowledge is illustrated by an evaluation function for parameter optimization.The fixed sliding time window in the transfer entropy calculation is upgraded to an adaptive window with adjustable sliding step size,which effectively improves the feature extraction and causality analysis ability of the conventional TE regarding to the operating failures.3.Regarding the problem that the transition time of incipient fault is long and the fluctuation of incipient fault is insignificant,the data stream clustering based fault detection method is proposed.On the basis,a dynamic data stream transfer entropy(DDSTE)algorithm is proposed and the DDSTE-based causality analysis method is presented.The suitable length of data selection window is matched with the evolution feature of the incipient fault and the time scale is applied to highlight the sequence change characteristics,reduce the influence of noise signals,and improve the online calculation speed of transfer entropy.4.In response to the problem that poor performances of the real-time root tracing of multiple operating condition faults,a novel convolution neural network for production indices prediction is proposed.On this basis,a regional adaptive transfer entropy(RATE)algorithm that can distinguish the data distribution characteristics of different operating conditions is proposed.Through adaptive segmentation and symbolic processing,the dynamic threshold is matched for causal feature analyses,which reduces the complexity of the algorithm and the loss of key information in the causality analysis.