Fault Detection and Isolation (FDI)

The foundational mechanism of FDI that computes the difference between actual sensor measurements and expected values derived from a mathematical model. A non-zero residual signals a potential fault.

• Analytical Redundancy: Uses a dynamic system model to generate expected outputs, replacing the need for physical backup sensors.
• Parity Space Methods: Transforms sensor outputs into a subspace where faults are decoupled from the system state, making anomalies immediately visible.
• Observer-Based Residuals: Employs state observers like Luenberger observers or Kalman filters to track estimation errors that diverge under fault conditions.

The decision-making layer that evaluates residual signals against statistical thresholds to declare a fault condition, balancing sensitivity against false alarms.

• Sequential Probability Ratio Test (SPRT): Accumulates evidence over time to detect subtle incipient faults while minimizing detection delay.
• CUSUM Algorithm: Detects small, persistent shifts in residual mean that indicate gradual sensor degradation rather than catastrophic failure.
• Adaptive Thresholding: Dynamically adjusts detection boundaries based on operating conditions and noise levels to maintain a constant false alarm rate.

The process of pinpointing the specific faulty component by designing residual sets where each fault produces a unique, identifiable signature pattern.

• Directional Residuals: Constrains each residual vector to respond to only a subset of possible faults, creating a decoupled fault signature matrix.
• Generalized Observer Scheme: Deploys a bank of observers, each driven by all sensor inputs except one, so a fault in the excluded sensor leaves only its corresponding observer unaffected.
• Dedicated Observer Scheme: Drives each observer with a single sensor input, so a fault in that sensor only corrupts its dedicated observer's residual.

The capacity of an FDI system to distinguish true sensor faults from model-reality mismatches caused by unmodeled dynamics, parameter drift, or external disturbances.

• Unknown Input Observers (UIO): Decouples the state estimation error from unknown disturbances, making residuals sensitive only to faults and not to environmental noise.
• H-infinity Filtering: Minimizes the worst-case gain from disturbances to estimation error, ensuring residuals remain below thresholds during normal operation despite model inaccuracies.
• Disturbance Decoupling: Algebraically removes the influence of known disturbance directions from the residual generator, preventing false trips during load changes.

The quantitative characterization of how a sensor's performance metrics drift over time, enabling predictive FDI rather than purely reactive fault detection.

• Bias Drift Estimation: Tracks the slow accumulation of systematic offset in sensor readings, distinguishing it from the zero-mean noise expected in healthy operation.
• Noise Density Monitoring: Continuously estimates the variance of the sensor's stochastic error, flagging increases that indicate impending bearing wear or electronic degradation.
• Prognostic Health Index: Fuses multiple degradation indicators into a single scalar value that predicts remaining useful life before the sensor violates its accuracy specification.

An advanced FDI strategy where a probing auxiliary signal is deliberately injected into the system to excite latent faults that are not detectable under passive observation alone.

• Optimal Input Design: Computes a minimally invasive test signal that maximizes the statistical separability of healthy and faulty hypotheses.
• Closed-Loop Probing: Injects the diagnostic signal while maintaining stability constraints, ensuring production quality is not compromised during the test.
• Multi-Model Hypothesis Testing: Evaluates the response to the injected signal against a library of fault models, selecting the most likely fault scenario based on Bayesian evidence.

In SAE Level 4 autonomous driving platforms, FDI continuously monitors the health of LiDAR, radar, and camera subsystems. When a sensor degrades due to occlusion or hardware failure, the FDI system isolates the faulty stream and reweights the fusion algorithm to rely on remaining healthy sensors.

• Detects blinding attacks and physical sensor damage in real-time
• Prevents corrupted point clouds from entering the Kalman filter state estimator
• Triggers safe fallback maneuvers when sensor redundancy is compromised

Modern turbofan engines employ FDI within their Full Authority Digital Engine Control (FADEC) units. The system cross-validates hundreds of parameters—exhaust gas temperature, rotor speeds, fuel flow—against an onboard thermodynamic model to detect incipient sensor faults before they affect thrust control.

• Isolates biased thermocouples that would cause erroneous fuel metering
• Uses analytical redundancy to replace failed sensors with model-based virtual estimates
• Prevents in-flight shutdowns by distinguishing sensor faults from genuine engine anomalies

In pressurized water reactors, FDI algorithms validate the integrity of redundant neutron flux detectors and coolant temperature sensors that feed into the reactor protection system. The FDI layer prevents spurious sensor readings from triggering unnecessary emergency shutdowns.

• Applies consistency checking across quadruple-redundant sensor channels
• Detects common-mode failures that could defeat simple voting logic
• Maintains plant availability while preserving safety instrumented system integrity

In photolithography and etching tools, FDI monitors mass flow controllers, pressure transducers, and RF power sensors that govern plasma conditions. A drifting pressure sensor can produce sub-surface damage invisible to post-process inspection.

• Uses principal component analysis to detect subtle multivariate sensor correlations
• Isolates faulty sensors before they cause yield-killing process excursions
• Reduces scrap by distinguishing sensor drift from genuine process variation

FDI systems on offshore turbines analyze vibration accelerometers, oil debris sensors, and strain gauges in the gearbox and main bearing. The challenge is distinguishing a failing sensor from an actual mechanical fault in a remote, inaccessible asset.

• Applies model-based residual generation comparing physical models to sensor outputs
• Prevents unnecessary maintenance vessel dispatches costing hundreds of thousands
• Extends asset life by ensuring only genuine mechanical faults trigger interventions

In automotive assembly lines, FDI validates the absolute encoders and torque sensors in each robot joint. A faulty encoder reporting incorrect angular position can cause collisions, tool damage, or safety zone violations.

• Cross-validates encoder readings against motor current signatures and kinematic models
• Detects intermittent faults caused by cable wear in continuous motion axes
• Enables condition-based replacement during planned downtime rather than reactive repair