End-to-end evaluation and error propagation in the two-stage cascaded system

Experimentally verify the error propagation characteristics between the Stage 1 LiteInception+Transformer fault detection module and the Stage 2 LiteInception fault identification module by conducting a unified end-to-end evaluation of the two-stage cascaded diagnostic system, including characterizing the behavioral patterns of Stage 1 false positive samples upon entering Stage 2.

Background

The paper proposes a two-stage cascaded diagnostic architecture for general aviation fault diagnosis: Stage 1 performs high-recall fault detection using a LiteInception+Transformer hybrid, and Stage 2 performs fine-grained fault identification using LiteInception. While both stages are evaluated individually, the overall system behavior when stages are composed has not been assessed.

The authors emphasize that the first stage acts as a safety bottleneck and that understanding how errors propagate from Stage 1 to Stage 2 is essential for deployment. They explicitly note that an end-to-end evaluation and experimental verification of error propagation characteristics remain outstanding.

References

Nevertheless, this work currently reports only the independent performance of the two stages separately, and has not yet conducted an end-to-end evaluation of the two stages in series on a unified test set. The error propagation characteristics between the two stages - for example, the behavioral patterns of Stage 1 false positive samples upon entering Stage 2 - remain to be experimentally verified, representing an important direction for future work.

— LiteInception: A Lightweight and Interpretable Deep Learning Framework for General Aviation Fault Diagnosis  (2604.01725 - Wei et al., 2 Apr 2026) in Section 5.1 (Discussion)

Because the chaining condition used ground-truth task categories, the reported gains are an upper bound under correct classification; end-to-end performance would additionally depend on Stage~1 accuracy, which we have not measured on this benchmark.

— Design and Evaluation of LLM Chaining-Based Task Planning for General Purpose Service Robots  (2609.29043 - Bruno et al., 24 Sep 2026) in Section 5, Limitations

An additional stress test injects 10\% false-positive anomaly boxes during Stage-II training, targeting otherwise valid hands and objects. The available assessment of this experiment is qualitative; no separate aggregate scores or uncertainty estimates were reported. It therefore does not establish a quantified tolerance to a 10\% diagnostic error rate. The Confidence-Decay ablation supplies the quantitative evidence for the safeguard, while systematic robustness to different error rates remains unmeasured.

— Diagnose-to-Generate: Eliminating Structural Distortions in Flow Models via Patch-Level Grounded Alignment  (2610.17882 - Li et al., 13 Oct 2026) in Appendix, Section “Additional Results and Analysis,” subsection “Grounded Supervision and Diagnostic Robustness”