Memory-Augmented Prediction With Cross-Iteration Alignment for Anomaly Signal Detection in the Integrated Test of Aerospace Products

Abstract

Abstract. Anomaly detection in integrated test signals is crucial for ensuring the reliability and safety of aerospace products. However, this task faces significant challenges due to the scarcity of anomalous samples, the high cost of missed detection, and the subtlety of critical fault signatures. To address these issues, this article proposes a novel semi-supervised framework. First, to combat the extreme class imbalance and asymmetric misclassification costs, a cost-sensitive classification module (ASAP-C) is designed. It explicitly incorporates a misjudgment cost matrix into the loss function, significantly increasing the model's sensitivity and recall for rare anomalies. Second, to overcome the lack of point-wise labels and capture subtle anomalies, an Unsupervised Anomaly Boundary Inference (UABI) module is developed. This module leverages the strong temporal characteristics of normal signals; a Temporal Sequence Predictor with Memory Reinforcement (TSP-MR) is trained to model normal temporal dynamics, enabling the identification of subtle anomalies as significant deviations from the predicted normal pattern. The framework is trained end-to-end via a Progressive Fine-Tuning Pipeline (PFTP). The “cross-iteration alignment” is achieved through an adaptive loss modulation mechanism that progressively refocuses the model on difficult-to-classify anomalies across training iterations. Experimental results on real-world aerospace test datasets demonstrate the superiority of the proposed method, which not only achieves high accuracy in classifying abnormal signals but also excels in precisely locating subtle anomalous intervals that are often overlooked by conventional methods. This approach contributes to enhancing the efficiency and reliability of aerospace product testing.