Sarcouncil Journal of Engineering and Computer Sciences

Abstract: Bayesian networks provide powerful causal reasoning capabilities, yet face two significant barriers in industrial settings: computational scalability with high-frequency sensor data and integration of domain expertise. The Knowledge-Infused GPU Bayesian Network (KI-GPU-BN) framework addresses these challenges through a dual approach - extracting causal relationships from organizational documents via natural language processing and accelerating structure learning algorithms on graphics processing units. Domain knowledge serves as both soft priors and hard constraints to guide graph discovery toward physically plausible edges, while CUDA-optimized conditional independence testing delivers substantial speed improvements over traditional implementations. Evaluations across industrial workloads demonstrate that knowledge integration reduces false-positive edges with minimal runtime overhead. Deployments in refinery operations and ESG reporting environments showcase operational cost reductions through improved energy efficiency and streamlined compliance processes. The architecture represents a practical advancement toward enterprise-scale causal modeling that balances computational efficiency with domain expertise integration.