Sensor faults can cause serious problems in control systems, such as performance degradation, safety hazards, or false alarms. Therefore, it is important to detect and diagnose them in real-time, before they affect the system output or damage the equipment. In this article, you will learn some basic concepts and methods for sensor fault detection and diagnosis in control engineering.
A sensor fault is a malfunction or deviation of a sensor from its normal operation, resulting in inaccurate or unreliable measurements of the physical variables of a system. Sensor faults can be classified into different types, such as bias, drift, noise, outage, or saturation. Sensor faults can be caused by various factors, such as aging, wear, environmental conditions, interference, or tampering.
Real-time detection is imperative for many reasons, such as maintaining the quality and efficiency of the system operation and avoiding waste or errors, preventing or mitigating the consequences of sensor faults on the system safety and stability, as well as providing timely and accurate information for fault diagnosis and recovery. This can help to avoid accidents or failures, and facilitate the maintenance and repair of the sensors.
One of the biggest causes of inefficiency is 'Latency'. Latency here means delay between the trigger and response. There are four links in a chain - event detection, data collation, depiction or visualisation, and decision making. The very first step - event detection, is most crucial, it sets into motion other steps. Sensors are placed to detect and report any events in the process. Sensor malfunction is the most serious kind of event and this should be detected with close-to-zero latency.
If a LiDAR sensor goes awry you could end up killing someone. In a practical example, in wafer processing, if a wafer is not detected properly, one could lose $100,000 worth of sensors, devices or read heads in data storage. Can you imagine if a fighter aircraft fails to detect the signature of an enemy aircraft or incoming missile? Sometimes redundancy has to be built in, depending on the cost/benefit analysis. Can you really imagine losing a $150 million F-35 because a home sensor malfunctioned?
To detect sensor faults, you need to have a model of the system and the sensors, and compare the actual measurements with the expected ones. A model is a mathematical representation of the system dynamics and the sensor characteristics, based on physical laws, empirical data, or statistical methods. A model can be linear or nonlinear, deterministic or stochastic, continuous or discrete, depending on the system complexity and the sensor behavior.
A fault detector is a device or algorithm that compares the actual measurements with the expected ones, and generates a signal or an indicator when a sensor fault is detected. It can be based on various principles, such as residual generation (calculating the difference between the actual and expected measurements and comparing it with a threshold or a confidence interval), parity relations (using redundant or independent measurements to check the consistency or compatibility of the sensor outputs), parameter estimation (estimating the parameters of the sensor model and detecting changes or anomalies in their values), and state estimation (estimating the state variables of the system and detecting discrepancies or errors in the sensor measurements).
Fault diagnosis is the process of identifying the type and location of a sensor fault, based on the fault detection signal or indicator. To do so, various methods can be employed, such as fault isolation by using logic rules or decision trees, fault identification with regression or classification techniques, and fault estimation with optimization or filtering techniques. All of these methods can help to accurately identify a faulty sensor, identify the fault type from a set of predefined categories, and estimate the fault magnitude or duration from the measurements.
Evaluating a fault detection and diagnosis system requires considering a variety of criteria, such as sensitivity for detecting small or subtle faults, robustness for dealing with uncertainties and disturbances, accuracy for diagnosing the fault type and location, speed for detecting and diagnosing a fault, and complexity in terms of computational or hardware resources.