TAPA: A Transformer-Driven Framework for Predictive Anomaly Detection via Parallel Learning and Pattern-Guided Analysis

Anomaly detection in sequential data has become increasingly critical across various domains, with increasing demands for predictive detection capabilities. While recent advances in deep learning have led to various sophisticated methods, the complex nature of real-world anomalies requires more effective integration of different learning paradigms. This paper introduces a novel Transformer-driven framework that combines parallel learning streams with pattern-guided analysis for predictive anomaly detection. The framework achieves efficient predictive detection through two parallel processes: sequence pattern learning via an improved Transformer architecture and direction-referenced scoring probability distribution learning of sequence behaviors, complemented by pattern-guided analysis with optimized tree structure for pattern recognition. Through extensive experiments on comprehensive sequential datasets, our framework demonstrates superior performance with average precision of 83.13%, recall of 86.24%, and F1-score of 84.54%, consistently outperforming existing methods across different datasets. The results show significant improvements over traditional methods, particularly in capturing complex temporal dependencies and sequential anomaly patterns.