Fault Detection and Isolation (FDI)

The core mechanism of FDI is the generation of residual signals. A residual is the difference between a sensor's actual measurement and the measurement predicted by a mathematical process model of the system. Under normal conditions, this residual is small and consistent with expected sensor noise. A statistically significant deviation in the residual signals the potential presence of a fault. This process requires an accurate model of the system's dynamics to generate reliable predictions for comparison.

FDI systems do not rely on simple threshold checks. They employ statistical hypothesis testing to distinguish between normal sensor noise and a genuine fault. Common methods include:

• Chi-squared tests on the squared Mahalanobis distance of the residual.
• Cumulative Sum (CUSUM) algorithms to detect small, persistent biases.
• Generalized Likelihood Ratio (GLR) tests to identify the most likely fault magnitude. This statistical framework provides a quantifiable confidence level for each fault declaration, reducing false alarms caused by transient noise.

Once a fault is detected, the system must determine which specific sensor or component is faulty. This is achieved through structural analysis of the system's equations. By designing a set of structured residuals or parity relations, each residual is made sensitive to a specific subset of faults while being decoupled from others. The unique pattern of which residuals are triggered (the fault signature) maps directly to a specific faulty sensor, enabling precise isolation. This is fundamental to maintaining system functionality by allowing the fusion algorithm to ignore or de-weight the faulty input.

Effective FDI is inherently model-based. It relies on precise mathematical representations of both the sensor model (how measurements relate to the state) and the process model (how the state evolves). These models can be:

• Linear (e.g., for a Kalman Filter-based FDI).
• Nonlinear (e.g., for an Extended or Unscented Kalman Filter-based FDI).
• Data-driven (e.g., neural networks trained to predict nominal behavior). The fidelity of these models directly determines the sensitivity and accuracy of the fault detection process.

For autonomous systems like robots or self-driving cars, FDI must operate in hard real-time. The detection and isolation logic must execute within a deterministic cycle time, often measured in milliseconds, to prevent a faulty sensor from corrupting the state estimate and causing catastrophic system failure. This demands computationally efficient algorithms and careful implementation to meet latency constraints without sacrificing detection performance.

A key challenge is designing FDI systems that are robust to model inaccuracies, environmental disturbances, and unknown inputs that are not faults. Techniques to achieve this include:

• Unknown Input Observers (UIOs) designed to decouple residuals from specific disturbances.
• Adaptive thresholds that adjust based on estimated noise levels or system operating mode.
• Robust estimation cores (like H-infinity filters) that minimize the worst-case effect of uncertainties on the residual. Without robustness, the system becomes prone to excessive false positives, undermining trust in the automation.

In self-driving cars, FDI continuously monitors the health of lidar, radar, cameras, and inertial measurement units (IMUs). A failure in a single sensor, like a camera obscured by dirt, must be isolated so the system can rely on the remaining sensors (e.g., radar) to maintain a safe occupancy grid. This prevents catastrophic failures and is a core requirement for ASIL-D (Automotive Safety Integrity Level) certification.

Modern aircraft use triple-modular redundancy with three independent sensor suites. FDI algorithms compare the outputs of these redundant channels. If one sensor drifts or fails (e.g., a pitot tube icing over), the system isolates the faulty channel and uses a voting mechanism from the remaining two to provide correct airspeed and attitude data to the fly-by-wire computers, ensuring continuous safe flight.

In chemical plants or power generation, FDI protects against costly downtime and hazardous conditions. It monitors networks of pressure, temperature, flow, and vibration sensors.

• Detection: Identifies a sensor giving a reading inconsistent with physical models (e.g., a stuck temperature gauge).
• Isolation: Pinpoints the exact sensor, allowing operators to schedule maintenance without shutting down the entire process line. This is integral to predictive maintenance strategies.

Precision robots in manufacturing use force-torque sensors, vision systems, and joint encoders. During a delicate assembly task, FDI can detect if a vision system's calibration has drifted or if a force sensor is saturated. By isolating the faulty modality, the robot can switch to a force-guided search strategy to complete the task, preventing part damage and line stoppages.

Advanced medical imaging and monitoring systems, such as MRI machines or patient vital sign monitors, use sensor fusion. FDI ensures that a malfunctioning sensor within an array does not lead to a misdiagnosis. For example, in a multi-parameter patient monitor, it can isolate a faulty pulse oximeter reading and alert staff while continuing to display valid data from ECG and blood pressure sensors.

Spacecraft rely on star trackers, sun sensors, gyroscopes, and magnetometers to determine orientation. FDI is essential due to the harsh radiation environment, which can cause transient sensor faults. The system uses model-based reasoning to detect a gyroscope experiencing drift and can reconfigure to use alternative sensors, maintaining correct attitude for communication, power, and scientific operations.

The foundational process of combining data from multiple physical sensors to produce a more accurate, complete, and reliable estimate of a system's state than is possible with any single source. FDI is a critical subsystem within a fusion architecture, responsible for identifying and mitigating sensor failures that would otherwise corrupt the fused estimate.

• Centralized vs. Decentralized: Architectures differ in where fusion occurs—centrally or across distributed nodes.
• Role of FDI: Acts as a gatekeeper, ensuring only valid, non-faulty data enters the fusion process.

The overarching goal of inferring the hidden variables (the state) of a dynamic system from a sequence of noisy observations. FDI enhances state estimation by providing diagnostic information about sensor health, allowing the estimator to down-weight or exclude data from faulty channels.

• Core Problem: Determining position, velocity, orientation, or other latent properties.
• FDI's Contribution: Improves estimation robustness by identifying which sensor streams are trustworthy, directly impacting the accuracy of the final state estimate.

A class of statistical methods designed to be insensitive to outliers and violations of modeling assumptions. While FDI explicitly identifies and isolates faulty sensors, robust estimation techniques (e.g., RANSAC, M-estimators) are often used within the fusion algorithms to tolerate a certain level of erroneous data without a formal fault declaration.

• Key Difference: Robust estimation tolerates outliers; FDI identifies and removes the source of outliers.
• Complementary Use: Systems often employ both: robust estimators handle minor noise, while FDI handles complete sensor failures.

The general probabilistic framework for recursively estimating the state of a dynamic system. Algorithms like the Kalman Filter, Particle Filter, and their variants are implementations of this framework. FDI integrates with Bayesian filtering by modifying the measurement model or innovation sequence when a fault is detected.

• Innovation Sequence: The difference between predicted and actual measurements. FDI monitors this sequence for statistical anomalies indicative of a fault.
• Model Adjustment: Upon isolation, the filter can be reconfigured to ignore measurements from the faulty sensor.

A square matrix that represents the uncertainty of the estimated state vector and the correlations between its components. In FDI, the covariance matrix is crucial for two reasons:

1. Residual Analysis: Fault detection often involves checking if the difference between predicted and actual measurements (the residual) is statistically consistent with the expected covariance.
2. Isolation Logic: The structure of the covariance helps determine which state variables are most affected by a particular sensor's failure, aiding in isolation.

It quantifies the confidence in the sensor data and the fused estimate.

A statistical distance measure between a point and a distribution, scaled by the distribution's covariance. It is a fundamental tool in FDI for fault detection gating.

• How it Works: The innovation (measurement residual) is computed. Its Mahalanobis distance from the expected distribution (typically zero-mean) is calculated. A large distance indicates the measurement is an outlier, potentially signaling a sensor fault.
• Thresholding: This distance is compared against a threshold derived from a chi-squared distribution to make a probabilistic decision about fault detection.