BMS software reliability prediction as a function of residual defect density using SIL dynamic metrics

Abstract

The continuously increasing complexity and computerization of automotive systems correlate directly to increased expectations from customers on the reliability of the system. Especially for battery management systems, the malfunction, defect, or loss of functionality can potentially lead to hazards that might compromise the safety of passengers and the environment. Software development being the major part of em- bedded product development, quality metrics are important in deciding the overall reliability of the system. However, unlike hardware, the software reliability is diffi- cult to estimate since reliability modeling requires a failure dataset that is collected for a substantial amount of time for accurate prediction which is not feasible during the early development phases of software. In this work, we propose the reliability prediction of the embedded software based on the residual defect density as a failure rate function. To estimate residual defect density, it is important to model the prediction of residual defects. The traditional approaches generally use static metrics such as code metrics from a single software release along with failure data to predict and forecast software defects. We propose a new approach by introducing dynamic metrics in residual defect prediction along with code and process metrics. Dynamic metrics are collected using the SIL tool which verifies the functional compliance of the software on vehicle test data. The impact of dynamic metrics on residual defect prediction and reliability prediction is evaluated by comparing the performance of regression-based machine learning algo- rithms such as Normal Regression, Decision Tree, Random Forest, Artificial Neural Network, and K-Nearest Neighbour. The faults are injected to verify the usefulness of the SIL tool in capturing the dynamic metrics and their impact on the reliability of the software. The results from the experiment are positive, indicating the improvement in the residual defects prediction with dynamic metrics. Although, some algorithms (Ran- dom Forest and Decision Tree) performed better than the others (Linear Regression and K-Nearest Neighbour). In terms of reliability estimation, the dynamics metrics are able to capture the faults introduced and slightly improved the reliability of the software compared to actual software release versions