HSG-Net: A Scheduling-Aware Heterogeneous Graph Surrogate for Aviation Sensing-Data Processing Chains

Efficient end-to-end timing evaluation is essential for the design and configuration of partitioned aviation sensing-data processing chains. Task activation, partition scheduling, message dependencies, and resource sharing jointly influence flow delay, while small variations in message arrival times may introduce substantial waiting across partition execution boundaries. Conventional reference simulation can capture these interactions but becomes computationally expensive when numerous candidate configurations must be evaluated. To address this challenge, this paper proposes HSG-Net, a scheduling-aware heterogeneous graph surrogate for predicting mean end-to-end delay from system configuration information. The proposed method represents tasks, partition windows, data flows, and hardware resources as interconnected entities, integrating explicit scheduling descriptors with heterogeneous relational dependencies. Graph message passing captures scheduling and resource interactions, while an adaptive direct-flow pathway preserves flow-local timing information and combines it with graph-derived representations for multi-scale delay prediction. On 7168 held-out test flows, the three-seed HSG-Net ensemble achieves a mean absolute error of 6.372 ms relative to reference-simulator labels, with 81.18% of predictions within a 10 ms absolute-error tolerance. Architectural comparisons demonstrate improved prediction over the evaluated direct-flow-only control, supporting the complementary use of relational context and flow-local information. Furthermore, HSG-Net also supports offline candidate screening through a validation-calibrated rule that allocates reference-simulation effort. The proposed approach provides an efficient, simulation-assisted method for timing evaluation and early-stage design-space exploration of partitioned aviation sensing-data processing architectures.