⚡ electrical engineering

Distributed Observer-based Fault Detection over Intelligent Networked Multi-Vehicle Systems

This paper proposes a distributed, observer-based fault detection and isolation strategy for connected autonomous vehicles in mixed traffic that utilizes probabilistic thresholds on residuals to identify sensor anomalies without requiring central processing or assumptions of bounded noise.

Original authors: Mohammadreza Doostmohammadian, Hamid R. Rabiee

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammadreza Doostmohammadian, Hamid R. Rabiee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a busy highway where two types of drivers are sharing the road: Human-Driven Vehicles (HDVs) and Autonomous Vehicles (CAVs). The human drivers are unpredictable; they might speed up, slow down, or change lanes without warning. The autonomous cars, however, are smart and connected. They talk to each other to stay safe and keep traffic flowing smoothly.

The problem is: What if the sensors on the smart cars get "confused" or "hacked"?

If a smart car's sensor lies about where a human car is, that smart car might make a dangerous decision. If it tells its neighbors, "Hey, that human car is right here!" when they are actually far away, the whole group could crash. Usually, to fix this, all the cars would have to send their data to a giant central computer (like a traffic control tower) to check for errors. But that's slow, and if the tower breaks, everyone is in trouble.

This paper proposes a smarter, decentralized way to handle this. Here is how it works, broken down into simple concepts:

1. The "Group Chat" Strategy (Distributed Observation)

Instead of waiting for a central boss to tell them what's happening, the smart cars form a group chat.

The Setup: Each smart car can only "see" one human car directly through its own sensors. It's like being in a dark room where you can only see the person standing right next to you.
The Trick: Even though Car A can't see Human Car #4, it can talk to Car B, who can see Human Car #4. By sharing what they see and what they think is happening, all the smart cars can build a complete picture of the entire highway.
The Innovation: The authors created a new, faster way for these cars to share information. Previous methods were like a slow relay race where cars had to pass a message back and forth many times before agreeing on a fact. This new method is like a quick group huddle where they agree on the facts in a single step. It saves time and battery power.

2. The "Lie Detector" (Fault Detection)

Now, imagine one of the smart cars has a broken sensor (or is being tricked by a hacker). It starts sending false data to the group chat. How do the others know it's lying without asking a central boss?

The paper introduces a "Residual"—think of this as a discrepancy score.

Every car predicts where the human cars should be based on the group chat.
Then, it looks at what its own sensor actually sees.
If the prediction and the sensor reading match, the discrepancy score is low (like a quiet whisper).
If the sensor is lying, the discrepancy score spikes (like a loud shout).

3. Two Ways to Catch the Liar

The paper suggests two different "lie detection" methods for the cars to use:

Method A: The "Snap Judgment" (Stateless Detection)
This is like looking at a photo and immediately saying, "That looks fake!" It checks the current error right now. If the error is too big, it sounds an alarm.
- Pros: Super fast, doesn't need to remember the past.
- Cons: It might get fooled by a sudden, temporary glitch (like a bird flying in front of a camera).
Method B: The "Pattern Watcher" (Stateful Detection)
This is like watching a video over the last few seconds. It looks at the history of errors. If the error is just a one-time spike, it ignores it. But if the error stays high consistently, it sounds the alarm.
- Pros: Much harder to fool; it ignores temporary noise.
- Cons: It takes a little longer to decide, and the car needs to remember more data (like having a bigger memory).

4. The "Probabilistic" Safety Net

The authors didn't just set a rigid rule like "If the error is greater than 5, it's a lie." That's too simple because real-world sensors are messy (wind, rain, static).

Instead, they used probability, like a weather forecast. They calculated: "Given the usual noise of the sensors, there is a 95% chance that an error this big is a real lie, not just random noise."

This allows the system to be flexible. It can say, "I'm 99% sure this is a fault," or "I'm only 68% sure, so I'll wait and see."
This is more realistic than older methods that assumed noise was perfectly predictable.

The Bottom Line

The paper proves that a group of smart cars can:

Track all the human cars on the road by sharing information, even if no single car can see everyone.
Detect if one of their own sensors is broken or hacked, without needing a central computer to tell them.
Isolate the bad sensor so the rest of the group can ignore it and keep driving safely.

They tested this in simulations with both small groups and large networks (up to 50 cars), showing that their method is faster, more efficient, and just as accurate as older, more complex methods. It's a step toward making self-driving traffic safer and more independent.

Technical Summary: Distributed Observer-based Fault Detection over Intelligent Networked Multi-Vehicle Systems

Problem Formulation
This paper addresses the challenge of state estimation and fault detection in mixed-traffic Intelligent Transportation Systems (ITS) comprising Connected Autonomous Vehicles (CAVs) and Human-Driven Vehicles (HDVs). The core problem involves a group of CAVs attempting to track the global state (positions and velocities) of a group of HDVs using only partial, local sensor measurements. A critical constraint is that no single CAV possesses a local measurement set sufficient to render the entire HDV system observable (i.e., no local $(A, C_i)$ -observability). Furthermore, the system must operate without a centralized processing unit, relying instead on distributed communication. A significant additional challenge is the potential for sensor faults, biases, or cyber-attacks (e.g., false data injection) on the CAVs, which could corrupt the distributed estimation process and lead to erroneous tracking across the network.

Methodology
The authors propose a two-stage distributed framework: a single time-scale distributed observer for state estimation and localized Fault Detection and Isolation (FDI) mechanisms based on residual analysis.

Distributed Observer Design:
The paper introduces a distributed observer where each CAV $i$ estimates the global state $x_k$ of all HDVs. The observer operates in two steps:
- Prediction: CAVs perform a consensus step on a-priori estimates using a row-stochastic consensus matrix $W$ derived from the communication network topology $\mathcal{G}$ .
- Innovation Update: CAVs update their estimates using local measurements and information shared from neighbors.
  Crucially, this observer is single time-scale, meaning it performs only one consensus step between consecutive system dynamics time steps ( $k$ and $k+1$ ). This contrasts with existing multi time-scale approaches that require multiple inner consensus loops ( $L$ steps) between system updates, thereby reducing communication and computational overhead. The observer stability relies on distributed observability, defined by the observability of the pair $(W \otimes A, D_C)$ , where $D_C$ is a shared observation matrix. This condition requires the CAV network to be strongly connected but does not require local observability at any individual node.
Fault Detection and Isolation (FDI):
The paper proposes localized FDI strategies that detect anomalies in sensor measurements without central coordination. Two logic types are developed:
- Stateless Detection: Based on instantaneous residuals ( $r_k = |y_k - \hat{y}_k|$ ). Decisions are made immediately based on the current residual value.
- Stateful Detection: Based on the history of residuals over a sliding time window $T$ . This utilizes a distance measure (sum of squared residuals) following a Chi-square distribution to integrate temporal information, reducing sensitivity to transient noise spikes.
A key methodological advancement is the use of probabilistic threshold design rather than deterministic bounds. The thresholds are derived from the statistical properties of the estimation error covariance and noise statistics (assuming Gaussian noise), allowing for the calculation of False Alarm Rates (FAR). The gain matrix $K$ is designed via Linear Matrix Inequalities (LMIs) to ensure that the residual at a faulty sensor is significantly larger than at neighboring sensors, enabling isolation.

Key Contributions

Distributed Observability without Local Assumptions: The work demonstrates that CAVs can track HDV states globally even when no single CAV can observe the system locally, provided the network is strongly connected and distributed observability holds.
Single Time-Scale Efficiency: The proposed observer eliminates the need for inner consensus loops, offering a more computationally and communicationally efficient solution compared to multi time-scale methods found in prior literature.
Probabilistic FDI: The paper introduces probabilistic thresholding for fault detection, which is more realistic for ITS applications than deterministic thresholding, as it does not assume noise has bounded support. It provides explicit formulas for FAR based on detection probabilities.
Localized Isolation: The strategy enables every CAV to independently detect and isolate faulty sensors in its neighborhood, preventing fault propagation across the network.

Results
The paper validates the proposed methods through simulations involving mixed traffic networks:

Observer Performance: In a 4-CAV/4-HDV scenario, the distributed observer successfully tracked HDV positions and velocities, reaching consensus despite partial measurements.
Scalability: In a large-scale simulation (25 CAVs, 25 HDVs), the proposed single time-scale observer demonstrated lower Mean Square Error (MSE) and significantly reduced computational/communication overhead compared to a multi time-scale observer with inner consensus loops (where $L=7, 10, 15$ ).
Fault Detection: Simulations with injected sensor faults (biases) showed that both stateless and stateful FDI strategies could detect and isolate faults.
- Stateful detection with a sliding window ( $T=15$ ) achieved a False Alarm Rate (FAR) of less than 0.3%, demonstrating high reliability against false alarms compared to the stateless approach.
- The methods remained effective even under large noise and uncertainty conditions, though detection thresholds increased, leading to higher FARs in extreme noise scenarios.

Significance and Claims
The authors claim that this work fills a gap in the literature regarding distributed observer-based FDI for mixed-traffic tracking. The significance lies in enabling decentralized decision-making for safety and security in heterogeneous traffic systems. By removing the reliance on a central processor and local observability assumptions, the approach enhances scalability and resilience. The probabilistic nature of the FDI is highlighted as a more realistic approach for real-world transportation systems where noise is not strictly bounded. The paper concludes that these decentralized strategies improve the reliability and safety of ITS by allowing vehicles to collaboratively estimate states and locally identify compromised sensors, preventing cascading failures. Future research directions identified include delay-tolerant observer design, resilient network topology design (survivable networks), and integrating these FDI methods with cooperative control strategies.