多模型传感器故障软闭环容错控制研究

Study of multi-model soft close-loop fault-tolerant control with sensor faults

针对一类非线性系统建立精确机理模型困难、且仅用单一模型进行故障检测与容错不甚可靠等问题,提出一种基于数据驱动的多模型传感器故障软闭环容错控制方法,并对非线性系统中卡死、恒增益、恒偏差等常见传感器故障进行了研究。首先采用历史数据建立了系统的RBF神经网络、最小二乘支持向量机和核部分最小二乘三种预测模型,并基于序贯概率比检验算法同时以多个模型产生的残差对传感器进行故障检测;当检测出传感器发生故障时,则用系统多个预测模型的融合值代替传感器的输出,从而以软闭环方式实现对传感器故障的容错控制。最后将所提出的方法应用于一阶水箱液位控制系统,实验结果表明多残差与序贯概率比检验算法的结合能够可靠诊断传感器故障,多预测值优化融合的软闭环可对传感器故障实现安全、高性能容错。

There were some problems in domain of fault detection of sensor and fault-tolerant control, such as difficult establishment of a precise mechanism model for a type of nonlinear system and the unreliability of the fault detection and fault tolerance with a single model. In order to solve those problems, this paper proposed a fault-tolerant control approach of multi- model soft close-loop with sensor faults based on the data-driven, and studied the stuck fault, constant gain fault and constant deviation fault in nonlinear system. Firstly, this approach established three type prediction models of RBF neural network, least square support vector machine and kernel partial least squares for the system with historical data. The residual errors generated by the multi-model could detect the sensor fault based on the SPRT（ sequential probability ratio test）algorithm. Secondly, it used the fusion values of the multi-predictive-models for the system instead of the output of the sensor when a sensor fault was detected, thus the approach of soft close-loop achieved the goal of fault-tolerant control of sensor faults for the system. And finally, the experiment applied the proposed method into the one-order tank liquid level control system. The experimental results show that sensor fault diagnosis becomes more reliable based on integration of multiple residual errors and SPRT, and the approach of soft close-loop with optimized fusion of multiple predictive values can possess high safety and performance of fault-tolerant.