Analytic Redundancy for Automatic Control Systems: Recursive Neural Network Based Virtual Sensors

Abstract

Listen

Translate article

Advanced Automatic Control and Fault Detection systems are being developed for Navy Submarines and Surface Ships. Within this context a critical concern is the vehicle automatic control system response to the environment, damage that may significantly degrade the vehicle performance, as well as unexpected sensor, actuator, or control surface failures. Such an occurrence may well lead to an automatic control system response that is either inadequate or inappropriate given the current state of the vehicle. In order to maintain mission effectiveness, changes in the vehicle dynamics, as well as component failures must be rapidly detected and recovery actions must be promptly initiated within the automatic control loop. The overall system consists of combining three state-of-the-art techniques—Robust/Reconfigurable Control, Recursive Neural Networks (RNNs), and Fault Detection and Isolation (FDI) algorithms—into a real-time system that provides vehicle monitoring, fault detection, and automatic control of the vehicle from within the executive control loop. The combination of these technologies provides an innovative approach for the identification of both discrete and continuous failures, for differentiating between component failure and environmental influences, and for incorporating model-based fault protection into the autonomous control loop. The current paper describes the use of the RNNs as Virtual Sensors (VS) to provide real-time analytic redundancy of the sensor readings provided to the automatic control system. Since sensor failures will feed directly into the automatic control system, it is critical to check and verify the sensor readings prior to utilization in the automatic control loop since all subsequent commands are predicated on the assumption that the sensed information is correct. As such, a virtual sensor system has been developed which provides complete analytic redundancy of the sensor measurements utilized by the automatic control system. Each sensor reading is checked against real-time simulation predictions of the “true” sensor values and a decision is made as to the validity of the measurement. The sensed values are then either passed on as correct or flagged as being in error and a Virtual Sensor estimate of the “true” sensor reading values are provided to the automatic control system. The results show that the typical sensor failure modes, sensor drift, sensor lock-up, sensor drop-out, sensor data spikes, and sensor noise can be detected and corrected analytically using this approach with zero false positives. A key requirement in developing this system was the capability to avoid reporting false positives, sensor problems, when none actually existed. These real-time systems for Advanced Control and Monitoring can mean the difference between safe, continued operation and potentially catastrophic failures.