To address the limitations of traditional anomaly detection mechanisms in financial transaction systems, particularly in handling dynamic loads and complex failure patterns, this study proposes an optimization method for an intelligent anomaly detection model designed for financial infrastructure. The objective is to improve early warning capability and fault localization efficiency to ensure system stability. Based on the data characteristics and operational scenarios of financial transaction systems, the proposed method adjusts the fusion weights between system metrics and logs, prioritizing high-frequency and continuous system metrics for constructing an early warning baseline. A multi-label anomaly classification mechanism is also introduced to enable accurate identification and labeling of composite anomalies by leveraging a set of predefined typical failure scenarios. Experimental results show that the optimized model maintains stable performance in simulated financial transaction scenarios, with an accuracy rate exceeding 95%, demonstrating high applicability.