考虑反馈控制的工业过程分层数据建模与多维故障诊断方法

数据驱动的故障诊断是维护工业过程平稳运行的重要技术。因反馈控制已是工业过程重要组成部分，数据驱动的故障诊断须能分析反馈的影响以提高精度，且能提供多维诊断信息以辅助决策。但现有诊断以分析反馈影响下的加性故障变量维度为主，故障类型以及诊断结论单一，难以满足工业实际需求。为此，拟将闭环过程分整个闭环系统、广义被控过程与核心被控对象三个层级，开展考虑反馈控制的故障诊断方法研究。针对整个闭环系统，提出复杂数据场景下融合控制约束的动态过程数据建模与故障变量诊断方法，提供变量维度的诊断；针对广义被控过程，提出基于反馈不变信息的部件特征提取方法，定位扰动、传感器、执行器及被控对象等各个故障部件；结合乘性故障模型，提出被控对象内部隐状态的去反馈方法，提供状态维度的诊断；最终形成一套考虑反馈控制的过程变量级、部件级和状态级的多维诊断框架，以提高诊断精度与丰富度，为推动数据驱动故障诊断的工业应用提供理论指导。

Data-driven fault diagnosis plays an important role in maintaining process running safety. Due to the widespread application of feedback control, data-driven fault diagnosis is required to consider the effect of feedback for improving the diagnosis precision. Moreover, available multidimensional diagnosis information is also important to make an effective decision. However, existing diagnosis approaches considering feedback are limited in analyzing the root variables for additive faults. Thus the fault type and the diagnosis result is insufficient for a real application. To solve these drawbacks, a closed-loop system will be described from three different perspectives. They are the whole closed-loop system, the generalized plant which is input with the control signals and output with the sensor measurements and the core controlled process input with the actuator signal and output with the sensor measurement. This project intends to comprehensively analyze the effect of feedback on the three-layered models. First, for the whole closed-loop system, the control algorithm is often known or easy to summarize based on the DCS so that the data model for the closed-loop system can be built supervised by the control algorithm. This will improve the variable-based contribution fault diagnosis precision since some priors can be used to reduce model uncertainty. This strategy can be applied into some complex scenarios including missing data, outliers and variable-in-errors. Secondly, the generalized plant consists of sensors, actuators, disturbances and the core process unit. A fault equipment location is necessary considering the influence of feedback. For a plant, it is still insufficient by just diagnosing the crucial variables and equipment because a plant contains a lot of inner states covering the dynamics. Thus, an effective state diagnosis model is strongly required for incisively analyzing the root causes and the effect of feedback control should also be considered when the faulty state is identified. Finally, the multidimensional diagnosis from variables, equipment and states will greatly improve the diagnosis effectiveness and precision in feedback control systems and provide a fairly complete diagnosis information for operation decisions.