Surface Defect Detection of Malleable Iron Pipe Fittings Using Convolutional Neural Network

Literature Overview

The 2024 paper by Bai Jie and Jiang Xianliang, published in the Journal of Jiangsu University, presents an improved convolutional neural network (CNN) approach for surface defect detection on malleable iron pipe fittings. Funded by the Zhejiang Provincial Basic Public Welfare Research Program (LTGN24F020003), this work addresses the practical challenge of automated visual inspection in pipe fitting manufacturing, where surface defects such as cracks, porosity, sand inclusions, and dimensional deviations must be reliably detected to ensure product quality and safety.

Core Technical Findings

The authors identified a limitation in the existing HRNetV2p algorithm: it cannot effectively balance detection accuracy across different defect scales. Large defects are detected with high accuracy, but small defects suffer from insufficient semantic information in shallow feature maps. The proposed solution introduces a channel-attention-based feature fusion module that adaptively adjusts the ratio of spatial-to-semantic information in fused features.

Performance Metric	Original HRNetV2p	Improved HRNetV2p	Improvement
Average AP50	88.7%	91.3%	+2.6%
Large defect AP50	Baseline	+2.7%	Improved
Medium defect AP50	Baseline	+2.7%	Improved
Small defect AP50	Baseline	+5.6%	Significantly improved
Dataset size (IIDD)	N/A	Custom-built	Malleable iron fittings

The improvement is most pronounced for small-scale defects, where the average detection accuracy improved by 5.6%. This is particularly significant in pipe fitting inspection, where small cracks and porosity defects are often the most critical quality concerns.

Technical Interpretation

Defect Types in Malleable Iron Pipe Fittings

Malleable iron pipe fittings (elbows, tees, couplings, caps, etc.) are subject to characteristic surface defects arising from their sand casting and heat treatment processes:

Defect Type	Cause	Typical Size	Detection Difficulty
Surface cracks	Thermal stress during cooling	0.5–5 mm length	High (thin, irregular)
Sand inclusion	Improper mold preparation	1–10 mm	Medium
Porosity	Gas entrapment during casting	0.5–3 mm	High (small, clustered)
Cold shut	Incomplete mold filling	2–15 mm	Medium
Surface roughness	Improper sand grain size	Distributed	Low
Dimensional deviation	Mold wear	Variable	Low

Feature Fusion Mechanism

The proposed CG (Channel-Guided) dense skip transmission unit and CG adaptive fusion module perform three operations:

Integration: Combining features from multiple network stages to capture both spatial detail and semantic context.
Recalibration: Using channel attention to reweight feature channels based on their importance for defect detection.
Re-integration: Re-combining the recalibrated features with adaptive weighting to produce the final fused representation.

This three-step process ensures that shallow features (which contain spatial detail but lack semantic meaning) are enriched with semantic information from deeper layers, while deep features (which contain semantic meaning but lack spatial detail) are supplemented with spatial information from shallow layers.

Engineering Practice Integration

Quality Control Workflow

The integration of CNN-based defect detection into pipe fitting manufacturing quality control follows a structured approach:

Image acquisition: High-resolution cameras (5–20 MP) with appropriate lighting (structured light or coaxial illumination) capture images of fitting surfaces.
Pre-processing: Image normalization, noise reduction, and region of interest (ROI) extraction to focus on the fitting surface.
Defect detection: The improved HRNetV2p algorithm processes the images and outputs bounding boxes with defect classification and confidence scores.
Decision logic: Defects exceeding severity thresholds trigger automatic rejection or manual review.
Data logging: All detection results are recorded for traceability and process improvement.

Comparison with Traditional NDT Methods

Method	Coverage	Speed	Cost per Part	Small Defect Sensitivity
Visual inspection (manual)	100% or sampling	Low	Low	Operator-dependent
Magnetic particle testing (MT)	Surface only	Medium	Medium	Good
Penetrant testing (PT)	Surface only	Medium	Medium	Good
Eddy current testing	Surface/near-surface	High	High	Good
CNN-based visual inspection	100%	High	Low (after setup)	Very good (91.3% AP50)

Implementation Considerations

For practical deployment in a pipe fitting manufacturing environment:

The IIDD dataset developed by the authors should be supplemented with additional defect samples from the specific production line to account for process-specific defect characteristics.
Lighting conditions must be carefully controlled and stabilized, as variations in illumination significantly affect CNN detection accuracy.
The system should be periodically validated using known defect samples to detect any degradation in detection performance.
Integration with the manufacturing execution system (MES) enables real-time quality monitoring and statistical process control (SPC).

Key Reflections

The 5.6% improvement in small defect detection accuracy is particularly meaningful for pipe fitting quality assurance. In pressure-containing applications, even small surface cracks can lead to catastrophic failure under cyclic loading. The ability to reliably detect small defects at high speed (enabling 100% inspection rather than sampling) represents a significant quality improvement. The attention mechanism approach offers a generalizable solution that could be adapted to other pipe fitting materials (cast iron, ductile iron, steel) with appropriate dataset development.

Conclusion

This research demonstrates that incorporating channel attention mechanisms into CNN-based defect detection algorithms significantly improves the detection of multi-scale surface defects on malleable iron pipe fittings. The improved HRNetV2p achieves 91.3% AP50 with the most significant gains for small defects. For manufacturing quality control, this technology enables high-speed, 100% surface inspection that surpasses the capabilities of traditional non-destructive testing methods for surface defects. The developed IIDD dataset and improved algorithm provide a solid foundation for industrial deployment, with the potential to reduce escape defects and improve product reliability in pipe fitting manufacturing.