Dynamic Probabilistic Signed Directed Graph Model And It Based Fault Diagnosis Research In Complex Chemical Industry

Because the modern petrochemical enterprises are tending to develop large-scale equipment and automatic process, fault diagnosis is an important project in the industrial production, and numerous researches have been devoted to this problem in the past decades. The probabilistic signed directed graph (PSDG) based fault diagnosis approach is the inheritance and improvement of traditional signed directed graph (SDG) approach. The PSDG approach has the capability to express the complex causal relationship of system as well as provide the explaination of fault evolution. Meanwhile, compared with SDG, the spurious interpretations in PSDG are reduced by introducing the probabilistic parameters. However, PSDG based fault diagnosis approach still has some disadvantages, including low model accuracy, hard to reflect the time-effectiveness of process, and impossible to diagnose the faults inside the positive feedback loop. To address these issues, a dynamic PSDG (DPSDG) model is presented, and the SDG modeling methods as well as the single and multiple fault diagnosis methods are researched in accordance with the DPSDG model. The main research works are as follows:(1) Aiming at the existing data driven SDG modeling methods are unable to solve the problem of different sampling time and data asynchronization, a time-delay based SDG modeling method is proposed. In this method, the dynamic time warping (DTW) technology is firstly imployed to find the optimal misplacement correlation between two time series data and decide the delay time between them. On the basis of the delay time, the whole procedure of SDG modeling framework is illustrated. Through this SDG modeling framework, the signed directed relation of the target system is determined.(2) Aiming at the existing PSDG models are hard to reflect the features of time-effectiveness during the process fault propagation, a unified and deep fault description model named as dynamic PSDG (DPSDG) model is proposed. Based on the established SDG structure, this DPSDG model introduces the conditional probabilistic parameters to express the transmission intensity between process variables, where the conditional probability of directed edges is defined as the probability of effect node state under the double conditions of cause node state and the propagation time between two nodes. The value of conditional probability is determined by two steps:off-line training and on-line selection. Compared with the traditional PSDG model, the constructed DPSDG model realizes the flexible fault representation and powerful ability for time-effectiveness, which provides a base of following researches on the single and multiple fault diagnosis.(3) Aiming at the PSDG based fault diagnosis method is impossible to diagnose the fault inside the positive feedback loop, a DPSDG based fault diagnosis method that considers the single source fault is proposed. In this method, the probability calculation formula of serial connection structure and converging connection structure are firstly derived. Then taking the collected fault symptoms as the condition, the posterior probabilities of each searched fault is computed and arranged, so the most possible fault reason can be further distinguished.(4) Aiming at the multiple source faults indeed exist in chemical process, combine the DPSDG model, and a Bayesian theory based multiple fault diagnosis method is proposed. Because DPSDG model introduces the conditional probability of each directed edge, the multiple fault diagnosis can be realized by maximizing the posterior probability. This multiple fault diagnosis method is conducted by the criteria of interpretability and reliability. The interpretability requires the reachable relation between abnormal symptoms and fault reasons, forming the optimization constraints; the reliability requires the maximal posterior probability of fault source combinations under the condition of collected fault symptoms, forming the objective function. Finally, this multi-constraint optimization formulation is solved by the genetic algorithm to obtain the most reliable fault explanation.This research takes continuous stirred tank heater (CSTH) process and Tennesse Eastman (TE) process as application examples, and experimental results show the validity and advantages of the proposed DPSDG model as well as it based fault diagnosis method.