Leveraging Machine Learning for CRC Diagnostic Coverage Estimation

This paper presents a novel machine learning-based approach to improve the accuracy of estimating the diagnostic coverage of Cyclic Redundancy Check (CRC) in automotive systems, crucial for adhering to functional safety standards like ISO 26262. By focusing on the prediction of Hamming weights - indicative of CRC’s error detection capabilities - for large code and message lengths, our methodology addresses the limitations of traditional brute-force and approximate estimation methods. Utilizing supervised learning with known Hamming weights as training data, we evaluate the effectiveness of seven machine learning models, with LightGBM demonstrating superior performance after hyperparameter tuning and training data filtering. Our experiments show that this approach can estimate normalized Hamming weights with a relative error margin of approximately $\pm 15 \%$, significantly enhancing the precision of CRC’s diagnostic coverage estimation compared to conventional techniques.