LiteInception: A Lightweight and Interpretable Deep Learning Framework for General Aviation Fault Diagnosis

Abstract: General aviation fault diagnosis and efficient maintenance are critical to flight safety; however, deploying deep learning models on resource-constrained edge devices poses dual challenges in computational capacity and interpretability. This paper proposes LiteInception--a lightweight interpretable fault diagnosis framework designed for edge deployment. The framework adopts a two-stage cascaded architecture aligned with standard maintenance workflows: Stage 1 performs high-recall fault detection, and Stage 2 conducts fine-grained fault classification on anomalous samples, thereby decoupling optimization objectives and enabling on-demand allocation of computational resources. For model compression, a multi-method fusion strategy based on mutual information, gradient analysis, and SE attention weights is proposed to reduce the input sensor channels from 23 to 15, and a 1+1 branch LiteInception architecture is introduced that compresses InceptionTime parameters by 70%, accelerates CPU inference by over 8x, with less than 3% F1 loss. Furthermore, knowledge distillation is introduced as a precision-recall regulation mechanism, enabling the same lightweight model to adapt to different scenarios--such as safety-critical and auxiliary diagnosis--by switching training strategies. Finally, a dual-layer interpretability framework integrating four attribution methods is constructed, providing traceable evidence chains of "which sensor x which time period." Experiments on the NGAFID dataset demonstrate a fault detection accuracy of 81.92% with 83.24% recall, and a fault identification accuracy of 77.00%, validating the framework's favorable balance among efficiency, accuracy, and interpretability.

• The paper introduces a two-stage cascaded diagnostic architecture that decouples high-recall fault detection from fine-grained classification to optimize precision and recall.
• It employs a multi-method sensor channel selection and a pruned 1+1 branch time series design, reducing sensor dimensions and achieving 70% model compression with 8x CPU speedup.
• Using knowledge distillation, the framework tunes the precision-recall trade-off while providing dual-level interpretability essential for regulatory compliance in aviation.

LiteInception: Technical Advancements in Lightweight and Interpretable Time Series Fault Diagnosis for General Aviation

Problem Context and Motivation

General aviation (GA) maintenance presents a unique challenge for machine learning approaches due to edge computational constraints, a high prevalence of weak and noisy fault signatures, a stringent need for interpretability to satisfy airworthiness, and the necessity for adaptive, safety-critical trade-offs between false positives and false negatives. The "LiteInception" framework directly targets these requirements by proposing a compressed, cascaded deep learning system with explicit interpretability and scenario-adaptive training, implemented and validated on real-world multivariate sensor data from the NGAFID Cessna 172 dataset (2604.01725).

Two-Stage Cascaded Diagnostic Architecture

Decoupling Detection and Classification for Alignment with Maintenance Practice

LiteInception introduces a two-stage, cascaded architecture that mirrors GA maintenance workflows: Stage 1 is high-recall binary fault detection (normal vs. anomalous), and Stage 2 performs fine-grained, multi-class fault classification but only on anomalous cases. This decoupling enables stage-wise, asymmetric optimization—prioritizing false negative minimization in alerting, and false positive reduction in diagnosis—matching real cost structures in aviation safety scenarios.

This cascaded design also provides computational advantages, especially when the proportion of normal flights is high, as inference resources can be allocated on demand. Objective separation further lowers the engineering burden for system updates, as only Stage 2 requires retraining for new fault types.

Multi-Method Sensor Channel Selection

Redundancy/Noise Reduction Through Statistical and Model-Based Fusion

To address the high input dimensionality typical of aircraft flight data, a three-pronged channel selection mechanism is developed, leveraging: (1) mutual information to measure statistical dependence, (2) gradient-based saliency to capture model reliance, and (3) Squeeze-and-Excitation (SE) attention weights for optimization-driven feature usage. Selection is finalized using cross-method consensus with expert physical causality constraints, eliminating channels with spurious correlations, causality errors, or redundancy.

This pipeline reduces 23 sensor channels to 15 diagnostically informative dimensions, yielding superior fault identification macro-F1 (from 69.6% to 78.2%) and efficiency versus standard feature sets.

LiteInception: Lightweight 1+1 Branch Time Series Architecture

Systematic Topological Pruning of InceptionTime

Classic InceptionTime features multi-branch modules for multi-scale temporal feature extraction, at the cost of substantial parameter and compute overhead. LiteInception proposes a theoretically motivated pruning to a 1 (kernel size=3 convolution) + 1 (MaxPool) branch design, supported by receptive field analysis: deep stacking of small-kernel convolutions achieves sufficient context coverage, while the retention of MaxPool is essential for extreme event detection (e.g., sensor spikes), which convolution cannot substitute.

Compared with the strong (tuned) InceptionTime baseline, LiteInception reaches 70% model compression (639K params vs. 2.15M), an 8x CPU inference speedup, and over 30% less memory, with only a 3.8% F1 performance dip—maintaining state-of-the-art diagnostics in 19-way classification and outperforming all transformer-based or hybrid baselines, which are shown to underperform in this time series setting due to their unsuitable inductive bias.

Knowledge Distillation for Precision-Recall Modulation

Exploiting Soft-Label Smoothing for Safety-Critical Adaptivity

A key technical finding is the use of knowledge distillation not merely for compression, but as a lever to modulate the precision-recall frontier. By varying distillation temperature and mixing coefficients, the student LiteInception can be tuned for high recall (safety-first, with more false alarms) or high precision (assisted diagnosis, workload reduction), controlled purely from training strategy, without architecture alterations. Distillation is shown to increase recall by ~2.4pp and decrease precision by ~1.5pp, with overall macro-F1 stable or even slightly improved relative to direct supervised training.

This property is especially relevant for aviation, where the cost of missed detections far exceeds that of false positives. LiteInception's training regime thus provides an essential operational tuning knob for deployed safety systems.

Dual-Level Interpretability with Multi-Method Attribution

Traceable Evidence Chain for Compliance

LiteInception incorporates dual-level explainability across sensors and temporal windows using a four-method ensemble (Input Gradient, Occlusion Sensitivity, Grad-CAM, Integrated Gradients). Attribution cross-validation and noise-perturbation analysis ensure that highlighted features correspond to physically plausible, model-relevant evidence, not dataset bias.

This framework enables maintenance engineers to localize faults to both "which sensor" and "which flight segment," supporting traceability essential for regulatory compliance. Attribution findings are consistent with domain knowledge—primary battery current and cylinder head temperatures universally carry high diagnostic weight, whereas fault type-specific responses are captured for exhaust gas and oil pressure features.

Experimental Evaluation

Validation on Real-World NGAFID Multivariate Time Series

• Fault detection (Stage 1, binary): LiteInception+Transformer achieves 81.9% accuracy and 83.2% recall, with a balanced P/R profile and substantial speedup and memory reductions.
• Fault identification (Stage 2, 19-class): LiteInception attains 77.0% accuracy/75.4% macro-F1 post-compression, with the Inception family architectures outperforming all attention-based baselines.
• Ablation analyses confirm that multi-branch pruning is justified, MaxPool is essential, and time-domain data augmentation (not amplitude-domain) is required for class balance.
• Sensor selection: Pruning to 15 channels improves F1 and reduces noise.
• Interpretability: Attribution analysis is physically consistent, robust to noise, and identifies structured discriminative patterns correlated with real engine behaviors.

Implications and Future Directions

Practical impact includes immediate applicability for edge-deployed, compliance-ready GA maintenance systems. Theoretical implications extend to the design of lightweight time series architectures, cautioning against direct adaptation of 2D image-domain strategies (e.g., MobileNet) and attention mechanisms for structured sensor data.

Limitations are non-trivial: experimental validation is constrained to Cessna 172 data, generalization to other aircraft types and real-time in-flight monitoring is not yet addressed, and true open-set/novelty detection is not covered. Further, end-to-end cascaded performance under realistic workflow (full error propagation) remains to be quantified.

Promising research trajectories include transfer learning for cross-aircraft adaptation, open-set diagnostic pipelines, online streaming architectures, and full-system evaluation with operationally relevant safety metrics.

Conclusion

The LiteInception framework presents a technical synthesis of lightweight deep time series modeling, scenario-adaptive distillation, physically motivated feature selection, and compliant interpretability, validated on large-scale GA flight data. The explicit exploitation of knowledge distillation for P-R tuning, together with strong empirical results, positions LiteInception as a reproducible benchmark for practical, deployable, and interpretable aviation fault diagnosis (2604.01725).