What Is Fault Detection?

Fault detection is the process of determining, from measured system signals, that a fault has occurred in a component or subsystem. It is the first phase of the broader fault detection and isolation (FDI) pipeline, which proceeds from detection through isolation (identifying which component is faulty) and ultimately to fault accommodation or system reconfiguration. Fault detection does not require full identification of the fault type or magnitude; it needs only to reliably distinguish normal operation from abnormal operation, typically within a bounded time window that preserves system safety margins.

The field draws on control theory, signal processing, and statistical estimation, and is closely related to fault tolerance and fault tolerant computing, where the objective is to keep systems operational despite component failures. Detection performance is measured by two competing metrics: sensitivity (the ability to detect small or incipient faults) and specificity (the ability to avoid false alarms under normal disturbances and noise).

Model-Based Detection Methods

Model-based fault detection compares actual plant outputs to predictions generated by a mathematical model of the healthy system. The difference between measured output and model prediction forms a residual signal. In fault-free operation, residuals are small and bounded by modeling error and sensor noise; when a fault occurs, residuals deviate detectably. Observer-based methods, including the Luenberger observer and Kalman filter, are widely used to generate residuals with known statistical properties. The parity space method constructs algebraic consistency checks from input-output data and the plant model. A survey of model-based and data-driven methods in process monitoring and fault diagnosis provides a structured comparison of these approaches, including their assumptions about model accuracy and their behavior under uncertainty. Model-based detection performs well when an accurate plant model is available but degrades when the model fails to capture system nonlinearities or time-varying behavior.

Data-Driven Detection Methods

Data-driven fault detection uses statistical or machine learning algorithms applied to historical and real-time process data, without requiring an explicit physics-based model. Principal component analysis (PCA) and partial least squares (PLS) are classical multivariate methods that establish a normal operating envelope in the measurement space and flag deviations from it. More recent approaches use neural networks, support vector machines, and convolutional networks trained on labeled fault and no-fault data. Recent advances in intelligent algorithms for fault detection and diagnosis surveys deep learning architectures applied to vibration, current, and thermal signals from rotating machinery, reporting detection accuracies above 98 percent for common bearing and gear faults when sufficient labeled training data are available. Data-driven methods are especially valuable for complex systems where first-principles models are unavailable or prohibitively expensive to develop, but they require representative training datasets and may generalize poorly to fault modes not represented in training.

Thresholds, Alarms, and Fault Tolerant Computing

Once residuals or anomaly scores are computed, a decision rule determines whether the current state should be classified as faulty. Simple threshold logic triggers an alarm when a residual exceeds a fixed level; adaptive thresholds adjust to changing operating conditions. In fault tolerant computing, analogous mechanisms operate at the hardware and software level: watchdog timers detect processor hangs, parity bits detect single-bit memory errors, and cyclic redundancy checks flag corrupted data in communication links. The integration of fault detection into real-time embedded systems must account for computation latency and the cost of false positives, which in safety-critical applications may trigger unnecessary shutdowns. Data-driven fault diagnosis for electric drives illustrates how detection latency and alarm rate trade-offs are managed in motor drive applications with fast dynamics.

Applications

Fault detection has applications in a wide range of engineering domains, including:

• Industrial process control, where early detection of sensor and actuator faults prevents product quality degradation
• Automotive systems, where onboard diagnostics (OBD-II) monitor engine and emissions components continuously
• Aerospace, where flight control computers monitor actuator and sensor health on every control cycle
• Power systems, where relay protection algorithms detect fault conditions within milliseconds
• Structural health monitoring, where vibration and strain sensors detect crack initiation or delamination in bridges and aircraft structures