GenAuto HARA
General Autonomy Inc.
Abstract
Ensuring vehicular safety and reliability in autonomous vehicles requires comprehensive hazard analysis and risk assessment (HARA). Traditional HARA processes are often labor-intensive and prone to human error, necessitating a more advanced solution. This white paper introduces our cutting-edge software, GenAuto HARA, which leverages generative artificial intelligence (AI) to revolutionize HARA for autonomous driving safety analysis. By harnessing generative AI’s predictive modeling and data generation capabilities, our software automates the most demanding aspects of HARA, significantly speeding up the process and enhancing the precision of safety evaluations. This results in remarkable efficiency gains, shorter assessment times, and a broader analytical scope, making a compelling case for integrating generative AI at the early stages of safety system development.
1. Introduction
Autonomous driving technologies present complex risks and uncertainties, requiring detailed safety analysis. HARA plays a crucial role in identifying potential hazards, evaluating associated risks, and implementing effective safety measures [1, 2]. However, conventional HARA processes are laborintensive and susceptible to human error. This paper introduces a generative AI framework, supervised by a safety expert, to automate HARA, thereby enhancing operational efficiency while maximizing precision in coverage.
2. HARA Process
HARA is a systematic process for identifying potential hazards, assessing risks, and implementing mitigation strategies. Fig.1 shows the overall framework of the HARA process [3]. The HARA steps accompanied by a brief description of each, are outlined as follows:
Figure 1. HARA steps to obtain safety goals.
2.1 Item Definition
Defining the item under analysis, specifying functions and operating environments based on product requirements and the Operational Design Domain (ODD).
2.2 Function Extraction
Identifying primary behaviors or actions of system components, establishing relationships between functions, and ensuring accurate representation of system behavior.
2.3 Malfunction Derivation
Identifying potential faults using guide words such as “no,” “unintended,” “early,” “late,” “more,” “less,” “inverted,” and “intermittent,” assessing their impact on the system.
2.4 Hazards Identification
Examining combinations of faults and functions to identify potential hazards, understanding vehicle-level behavior, and detailing hazards and their impacts.
2.5 Risk Assessment
Evaluating risks based on controllability (C), severity (S), and exposure (E), calculating the Automotive Safety Integrity Level (ASIL), and categorizing risks to define safety requirements.
2.6 Safety Goals
Establishing high-level requirements to mitigate hazardous scenarios, guiding the development of technical safety requirements and design measures.
3. Leveraging Generative AI in HARA
Generative AI automates repetitive and data-intensive tasks in HARA, thereby reducing human error and enhancing traceability. GenAuto proposed the following framework in Fig. 2 to assist the safety engineers in repetitive and data-intensive tasks using the capabilities of Generative AI. The key advantages of this framework include, but are not limited to:
Figure 2. The overall framework of GenAuto HARA.
3.1 GUI
A user-friendly Graphical User Interface (GUI) facilitates interaction between AI and safety engineers, supporting prompts for each HARA step. Fig. 3 shows the GUI of the GenAuto software for the generative AI-driven HARA.
Figure 3. GenAuto GUI for AI-driven HARA.
3.2 Supervision
A human-in-the-loop approach allows safety engineers to revise and edit AI-generated content, ensuring expert judgment and domain knowledge application.
3.3 Traceability
Maintaining comprehensive records of interactions, modifications, and decisions, ensuring compliance with safety standards like ISO 26262.
3.4 Compatibility
Handling diverse data sources and formats, ensuring compatibility with existing tools and datasets used in safety engineering.
4. Case Study: Autonomous Emergency Braking (AEB)
HARA is a systematic process for identifying potential hazards, assessing risks, and implementing mitigation strategies. Fig.1 shows the overall framework of the HARA process [3]. The HARA steps accompanied by a brief description of each, are outlined as follows:
4.1 Objective
To evaluate the efficacy of our proposed AI-driven HARA tool, we conducted a benchmark study on an Advanced Driver Assistance System (ADAS) feature: Autonomous Emergency Braking (AEB). The primary objective was to compare the results of our AI-driven HARA tool with those derived from a traditional manual HARA process.
4.2 Methodology
The study involved two phases:
1. Conducting a traditional manual HARA on the AEB system by experienced safety engineers.
2. Utilizing our AI-driven HARA tool to perform the same analysis by an experienced safety engineer.
4.3 Results
The results show significant improvements in both time efficiency and analysis coverage with our AI-driven approach.
4.3.1 Time Efficiency
The AI-driven HARA reduced the required time for performing the analysis by up to 80%, compared to the traditional manual process. This is illustrated in Table 1 and Figure 4.
4.3.2 Coverage Improvement
Our AI-driven platform improved the hazard analysis coverage by 20% compared to the average performance of safety engineers. This is detailed in Table 1 and Figure 4.
Table 1. Coverage and Time Improvement Comparison
Method | Coverage (%) | Time Required (hours) |
Traditional Manual HARA | 80 | 100 |
AI-driven HARA | 100 | 20 |
Figure 4. Comparison of Time Required and Coverage: Traditional vs AI-driven HARA
4.3.3 Traceability and Auditing
Additionally, the platform provides comprehensive traceability and maintains change logs throughout the analysis, facilitating easier auditing and review processes.
5. Conclusion
Generative AI transforms Hazard Analysis and Risk Assessment (HARA) by streamlining workflows, mitigating oversight, and broadening hazard perception. Integrating generative AI in the early stages of safety system development ensures proactive identification and mitigation of latent system vulnerabilities. This fosters a robust safety culture aligned with the complexities of autonomous driving technologies. The benchmark study on the Autonomous Emergency Braking (AEB) system demonstrates that our AI-driven HARA tool significantly enhances the efficiency and effectiveness of hazard analysis. Compared to the traditional manual process, the proposed AI-driven approach reduced the required time for performing HARA by up to 80% and improved the coverage by 20%. Additionally, our platform maintains comprehensive traceability and change logs, facilitating easier auditing and review processes.
These improvements highlight the potential of generative AI in advancing safety analysis processes. By leveraging generative AI, safety engineers can achieve more thorough and timely hazard assessments, ultimately contributing to the development of safer and more reliable autonomous driving systems.
References
[1] “ISO 26262-1:2018, road vehicles functional safety,” https://www.iso.org/standard/68383.html, accessed: December 2023.
[2] P. Koopman, The UL 4600 Guidebook: What to Include in an Autonomous Vehicle Safety Case.
Carnegie Mellon University, 2022.
[3] K. Beckers, M. Heisel, T. Frese, and D. Hatebur, “A structured and model-based hazard analysis
and risk assessment method for automotive systems,” in 2013 IEEE 24th International Symposium
on Software Reliability Engineering (ISSRE). IEEE, 2013, pp. 238–247