Distilled Edge Intelligence for Bolt Defect Detection in Communication Towers

Technical Summary: Distilled Edge Intelligence for Bolt Defect Detection in Communication Towers

Problem Statement
The paper addresses the challenge of deploying high-accuracy deep learning models for bolt defect detection (specifically loosening and corrosion) on resource-constrained edge devices at communication tower sites. A fundamental trade-off exists between accuracy and efficiency: cloud-side models offer reliable detection but incur prohibitive transmission latency and bandwidth costs, while lightweight edge models reduce resource consumption but often suffer from poor recall on minority classes (loose and corroded bolts), which are safety-critical. Furthermore, field-collected datasets exhibit extreme class imbalance (dominated by normal samples), and defects present visually heterogeneous patterns (geometric deformation for loosening vs. texture changes for corrosion) that are difficult to capture with a single imaging modality.

Methodology
The authors propose a Distilled Edge Intelligence Framework that integrates cloud-side high-accuracy perception with edge-side lightweight inference through three core components:

1. Heterogeneous Dual-Stream Teacher Model:

   • A cloud-based teacher model ($M_t$) utilizes a dual-stream architecture fusing RGB texture and monocular depth geometry. The depth maps are generated from single RGB images using a pre-trained foundation model (Depth Anything), avoiding the need for dedicated depth sensors.
   • To handle extreme class imbalance, the teacher is trained with a defect-aware focal loss that amplifies the weight of the "Loose" class and employs class-specific decision thresholds.

2. Two-Phase Training Strategy for the Student Model:

   • Phase 0 (Architecture Selection): A lightweight student architecture ($M_s$) is selected from a candidate family (e.g., MobileNetV2 variants) based on strict resource budgets (parameters, FLOPs, and storage) before training begins. In the experiments, MobileNetV2-0.5× was selected to fit a CPU-only edge node.
   • Phase 1 (Knowledge Distillation): The student is trained using logit-based knowledge distillation (response-level) rather than feature-level distillation. The authors empirically found that feature-level distillation degraded minority-class recall because the teacher's intermediate features were dominated by the majority "Normal" class, reinforcing bias. Instead, the student learns from the teacher's softened logit predictions ($T^2$ scaled) combined with a focal loss, effectively recovering the teacher's defect recognition capability at a fraction of the computational cost.

3. Confidence-Aware Edge-Cloud Collaborative Inference:

   • The system employs an asymmetric offloading policy. The edge model acts as a "conservative normal-sample filter."
   • Local Processing: If the edge model predicts "Normal" with high confidence (above a threshold $\tau$), the sample is resolved locally.
   • Cloud Offloading: If the prediction is "Loose" or "Corroded," or if the confidence for "Normal" is low, the sample is offloaded to the cloud for verification by the teacher model.
   • This mechanism is governed by a risk-constrained threshold selection process that optimizes for latency while enforcing a hard constraint on minimum defect recall ($R_{min}$).

Key Contributions

1. Framework Integration: A unified framework combining cloud-side high-accuracy perception with edge-side lightweight inference specifically for communication tower bolt inspection.
2. Two-Phase Training: A strategy that decouples structural resource constraint satisfaction (architecture selection) from performance optimization (knowledge distillation), utilizing logit-based distillation to improve minority-class recall without feature-level alignment.
3. Risk-Sensitive Collaboration: A confidence-aware inference mechanism that treats the edge model as a filter for normal samples, offloading only uncertain or potentially defective samples to the cloud to balance bandwidth usage with safety-critical detection reliability.

Experimental Results
Experiments were conducted on a field-collected dataset of 3,886 images (75% Normal, 9.5% Loose, 13.5% Corroded).

• Teacher Performance: The dual-stream teacher achieved a balanced accuracy of 0.9308 and a corrosion recall of 0.988, effectively handling class imbalance.
• Distillation Efficacy: Logit-based distillation significantly improved the student model. The "Loose" recall increased from 0.365 (undistilled student) to 0.692 (+32.7 percentage points), and balanced accuracy rose from 0.618 to 0.735.
• Edge Efficiency: The distilled student model (MobileNetV2-0.5×) achieved 56.8 FPS on a CPU-only edge node (including depth estimation), satisfying real-time requirements.
• Collaborative Performance:
  • At a balanced operating point ($\tau=0.75$), the system achieved a balanced accuracy of 0.923 (only 0.8 pp below cloud-only) while reducing the cloud upload ratio to 80.8% and latency to 48.48 ms.
  • At a safety-prioritized point ($\tau=0.90$), the system achieved zero local defect leakage (no false negatives resolved at the edge) with an upload ratio of 98.7%, nearly matching cloud-only performance.

Significance and Claims
The paper claims that the proposed framework effectively navigates the accuracy–latency–bandwidth trade-off inherent in edge-based tower inspection. By using a lightweight distilled model as a conservative filter, the system achieves near-cloud detection accuracy while significantly reducing unnecessary data transmission. The authors emphasize that this approach is particularly suitable for safety-critical tasks where false negatives are unacceptable, demonstrating that combining knowledge distillation with confidence-aware edge-cloud collaboration is a viable solution for deploying high-precision visual inspection on resource-constrained edge devices. The study does not claim to solve all edge computing challenges but specifically targets the accuracy-efficiency gap in structural health monitoring under extreme class imbalance.