Abstract
Fault diagnosis is crucial for complex autonomous mobile systems, especially for modern-day autonomous driving (AD). Different actors, numerous use cases, and complex heterogeneous components motivate a fault diagnosis of the system and overall system integrity. AD systems are composed of many heterogeneous components, each with different functionality and possibly using a different algorithm. In addition, these components are subject to the vehicle's driving state and are highly dependent. This paper presents the concept of a modular fault diagnosis framework for AD systems. The concept suggests modular state monitoring and diagnosis elements, together with a state- and dependency-aware aggregation method. Our proposed classification scheme allows for the categorization of the fault diagnosis modules. The concept is implemented on AD shuttle buses and evaluated to demonstrate its capabilities.
Introduction
Modern autonomous driving (AD) systems are inherently heterogeneous, integrating traditional rule-based logic with advanced AI-driven components across diverse hardware platforms. Due to the high degree of interdependence between these modules, a single failure—such as a sensor malfunction—can trigger a "waterfall" of errors across perception, localization, and planning stacks. For safety drivers or remote operators, this often results in information overload, making it difficult to identify the root cause of a system failure. Furthermore, many diagnostic tools fail to account for the vehicle's operational context (e.g., whether the vehicle is driving or stationary), leading to unnecessary alerts.
Methodology
The proposed framework addresses these challenges through a three-pronged modular approach:
1. Classification Scheme
The authors introduce a taxonomy to categorize faults based on three dimensions: Diagnosis Location (isolated vs. contextual), Type of Information (meta-properties like frequency vs. actual data content), and Data Flow (sequential vs. parallel). This structured approach allows for more precise diagnostic module design.
2. Five-State Diagnostic Model
To move beyond binary "Pass/Fail" reporting, the framework utilizes five discrete states: OK, WARNING, ERROR, IGNORE, and UNKNOWN. The IGNORE state is a critical innovation; it is applied when a component's status is irrelevant to the current vehicle state or when its inputs are already known to be faulty.
3. Dependency-Aware Aggregation
The system utilizes a directed graph to map software dependencies. If a root component (e.g., a LiDAR sensor) fails, the framework automatically transitions dependent modules (e.g., Perception) into the IGNORE state. This prevents propagated errors from cluttering the operator's interface.
Implementation and Evaluation
The framework was implemented on autonomous shuttle buses and evaluated through real-world failure injection scenarios, including sensor outages and localization drift.
In testing, the framework successfully isolated root causes. For instance, during a sensor outage, only the Sensors group reported an ERROR, while Localization and Perception were moved to IGNORE. Additionally, the system demonstrated state-awareness: a planning timeout that would be critical while "Active" (driving) was correctly ignored when the vehicle was "Inactive" at a bus stop.
Conclusion
The modular fault diagnosis framework significantly improves system transparency by utilizing dependency-aware aggregation and context-sensitive states. By effectively filtering "noise" from propagated errors, it allows operators to focus on root causes. Future work will explore automatic dependency detection to reduce the need for manual graph configuration and refine granular emergency countermeasures.
Summary updated