MTGFormer: A novel multi-task gated transformer with multi-head selective fusion attention for aeroengine gas-path parameter deviations parallel prediction

The variations in gas-path parameter deviations (GPDs) provide a comprehensive reflection of an aeroengine’s health state; therefore, GPDs are considered key indicators by airlines for monitoring the aeroengine’s health. However, airlines face challenges in autonomously obtaining them due to commercial restrictions. Although many data-driven methods have been developed to predict the GPDs from routinely collected multidimensional time-series (MTS) monitoring data, they often model GPDs separately, neglecting their internal correlations, which limits prediction accuracy and efficiency. To address this issue, a novel transformer’s variant, the so-called MTGFormer, is designed to capture both distinct and inner-relationships of different GPDs for accurate parallel prediction. Differently from existing transformer-based architectures that treat the temporal features extracted by all self-attention heads equally, a modified attention mechanism is designed, enabling MTGFormer to capture long-term temporal dependencies in MTS data while selectively focusing on the temporal features extracted by important self-attention heads based on downstream tasks, ultimately further enhancing the feature learning capability and interpretability of Transformer. Furthermore, a novel multiple-task learning module namely GMMoE is designed to handle the parallel prediction of multiple GPDs, which effectively addresses the problem that existing transformer-based architectures fail to capture both distinct and inner-relationships of different tasks in multiple-task scenarios. GMMoE treats the extracted temporal features as underlying shared representations, and employs multiple lightweight expert networks to independently extract aeroengine’s performance status information focusing on different aspects. Dedicated gate branches are then configured for each task to adaptively weight these state features, optimizing allocation and leveraging relationships between different GPDs, thereby significantly improving the prediction accuracy of each specific task. The effectiveness and superior performance of MTGFormer are validated by extensive comparative experiments on two real-world aeroengine maintenance datasets. Practically, MTGFormer empowers airlines with a reliable tool for autonomous health monitoring, reducing dependency on original equipment manufacturers. However, the model’s efficacy relies on the quality of routinely collected MTS monitoring data, implying that data completeness and consistency remain essential prerequisites.