(614e) Machine Learning-Based Identification and Accommodation of Sensor Faults in Sampled-Data Process Systems

Modern process control systems are comprised of dense sensor and actuator networks that interface through a shared communication medium to ensure process stability and performance requirements are met. Occasionally, however, faults and malfunctions can occur in the control system components, such as the inability of a control actuator to actuate the input, or the measurement sensors to transmit the correct reading. If left unaddressed, these faults can lead to degradation in the control system performance, and may even jeopardize closed-loop stability. The importance of operating processes safely and profitably, while meeting the desired performance specifications and stability requirements, has motivated significant research work on the problem of fault-tolerant control (e.g., [1]).

This problem has been studied extensively in the context of conventional feedback control systems where continuous state or output measurements are assumed to be available to the controller. Discretely sampled-data systems, on the other hand, are commonly found in industrial processes as continuous measurement of system data often proves logistically or technologically prohibitive. In sampled-data processes, mitigation of the impact of faults on the stability and performance of the closed-loop system is critically dependent on the rate at which the sampled measurements are taken. This realization has motivated a series of prior works on the detection and mitigation of control actuator faults in sampled-data processes (e.g., [2,3,4]). Compared with the available results on on the handling of control actuator faults, the problem of sensor fault-tolerant control has received less attention.

Examples of previous works on this problem include studies on passive fault-tolerant control approaches for handling sensor data losses (e.g., [5]), sensor reconfiguration-based approaches (e.g., [6]), stability-based accommodation using delayed measurements (e.g., [7]), and performance-based accommodation of systems with multi-rate sensor faults (e.g., [8]). A common theme in these studies has been the focus on mitigating the effects of sensor faults in order to maintain closed-loop stability and quality standards. The implementation of these approaches, however, requires the availability of effective schemes for the detection and identification of the sensor faults. Knowledge of the occurrence and magnitude of the faults is often required for any effective fault compensation strategy.

Motivated by these considerations, we present in this work an integrated approach for the active detection, identification and mitigation of sensor faults in a class of sampled-data process systems. The approach brings together tools from supervised machine learning (e.g., [9,10]), which are used for fault detection and identification, and model-based control techniques, which are used for fault mitigation. Initially, a model-based sampled data controller is designed, and its closed-loop stability region is explicitly characterized in terms of the fault size, the sampling rate and the controller design parameters. This characterization reveals the key parameters that can be used to actively mitigate the effects of these faults when they arise. To detect sensor fault occurrences and estimate their magnitudes, an Artificial Neural Network (ANN) is constructed and trained off-line using data collected from the control inputs under fault-free operation and the control inputs under varying magnitudes of sensor faults. Doing so, and using a sensor fault magnitude and class label mapping, enables us to generate an ANN model that is able to both classify the existence of a sensor fault as well as its magnitude. Once the training is complete and the model accuracy is verified, the ANN is used online to detect and identify sensor faults. The implementation and efficacy of the proposed strategy are demonstrated using a chemical process example.

References: