A Novel Anomaly Detection Framework for DevOps Pipelines Using Semantic Code Analysis and Build Log Correlation

Modern Continuous Integration and Continuous Deployment (CI/CD) pipelines serve as the backbone of cloud-native software engineering, yet diagnosing intermittent build failures and silent runtime regressions remains a labor-intensive challenge. Traditional pipeline monitoring approaches typically rely on isolated rule-based heuristic log parsers or separate static code analyzers, neglecting the critical contextual interplay between source code semantics and runtime build events. In this paper, we propose SemLog-AD, a novel multimodal anomaly detection framework that correlates abstract syntax tree (AST)-derived semantic code representations with semi-structured build log sequences. SemLog-AD integrates a CodeBERT-based semantic differential encoder with an attention-augmented Bidirectional Long Short-Term Memory (Bi-LSTM) network that processes streaming log events parsed via an adaptive Drain template engine. By projecting code delta embeddings and sequential log representations into a unified metric space, the framework captures subtle functional deviations and environmental anomalies during automated build execution. We evaluate SemLog-AD across a benchmark dataset of 14 large-scale cloud-native enterprise repositories comprising over 45,000 build executions. Experimental results demonstrate that SemLog-AD achieves an F1-score of 94.6% in identifying anomalous pipeline executions, outperforming state-of-the-art baselines by 8.4% while reducing developer diagnostic lead time by 41.3%.