1. Certification Challenges in the V&V of Bayesian Networks for Safety Critical Fault Diagnosis in Aerospace Systems Speaker: Mark Douthwaite Supervisor: Tim Kelly

2. Overview Introduction to Bayesian Networks – Capabilities, Use Cases and Benefits Assurance According to DO-178B/C – How Assurance is Approached in DO-178 Challenges in Bayesian Network Assurance – Overview of Identified Challenges Bayesian Networks and DO-178B/C – How the Assurance Challenges map to DO-178

3. Introduction to Bayesian Networks Figure 1: Visual representation of medical Bayesian Network used to aid diagnoses of patients in an ICU.

4. Introduction to Bayesian Networks Use Cases: Diagnostic/Prognostics (Medicine, Aerospace, Environment) Navigation/Time Series Analysis (Aerospace, Robotics) Text Processing/Filtering (Email, Consumer Products) Benefits: Robust to error & missing data Intelligible to human experts Designed to cope with uncertainty But – no formal safety critical guidance.

5. Assurance According to DO-178B/C System Requirements High-Level Requirements Low-Level RequirementsSoftware Architecture Source Code Executable Object Code Figure 2: Visual interpretation of implicit safety lifecycle in DO-178

6. Assurance According to DO-178B/C Non-prescriptive: Safety lifecycle is implicit Designed to be flexible Testing & Traceability: Strong focus on testing and V&V activities Stresses importance of traceability to requirements

7. Challenges in BN Assurance High level failure modes – Variations of Type I and Type II errors. Model Parameterisation – Defining the probability distribution of network. Model Structure – Defining the independence relations within the network. Model Evaluation – Safety-focused evaluation of model performance. Implementation – Implementing a software-based network system. Assurance of abstract Bayesian Network model is key.

8. Bayesian Networks and DO-178B/C Model Structure Data Artefacts System Requirements High-Level Requirements Low-Level RequirementsSoftware Architecture Source Code Executable Object Code Figure 3: Visual interpretation of safety lifecycle in DO-178 with envisioned mapping of BN-specific consideration.

9. Bayesian Networks and DO-178B/C Existing Provisions Data parameterisation items guidance in DO-178B/C (2.5.1) 2.5.1: ‘A data set that influences the behaviour of the software without modifying the Executable Object Code… Examples include configuration tables and databases.’ Data Assurance Need to provide high levels of confidence in underlying data artefacts Model Assurance Need to provide high levels of confidence in abstract model

10. Thanks for listening