Imbalanced Flight Test Sensor Temporal Data Anomaly Detection

The functioning of flight test sensors is crucial for aviation safety, but previous methods often overlooked the impact of data imbalance on model performance while exploring anomalies in aviation data. Imbalances in aviation time-series data can lead to the following issues: 1) Varying sizes of feature spaces lead to the majority class overshadowing the minority class; 2) small interclass differences in the model result in a bias toward predicting the majority class; and 3) imbalanced algorithms causes the model to lack learning of general features. To alleviate these problems, this article proposes the imbalanced flight test sensor temporal data anomaly detection (IFAD) model. First, a dual-branch network is designed where one branch is trained on original data distribution for general feature learning, and the other on resampled data to enhance learning of anomalous aviation data features. Second, a normalization module is introduced to standardize feature and weight vectors for a quantifiable decision boundary, along with an angle constraint for greater class differentiation. Finally, an adaptive reweighting algorithm (cosine-variance-aware loss reweighting algorithm) is implemented to balance the loss between different categories using cosine sample variance. Extensive ablation studies demonstrate that each proposed module enhances model performance. The IFAD model, when benchmarked against the baseline network, notably excels in the flights dataset, achieving improvements in accuracy, recall, and F1 score by 18.93%, 27.16%, and 12.43%, respectively, affirming its efficacy and superiority in detecting anomalies in flight test data.