Seal Ring Assembly Detection for Automotive Pipe Fittings Using Improved Faster RCNN

Literature Overview

The research by Xia Lei, Yuan Haibing, and Wu Jun, published in Computer Applications and Software (2025, Vol. 42, No. 6, pp. 93–99), presents an improved Faster RCNN-based method for detecting seal ring assembly quality in automotive pipe fittings. Conducted at Hubei University of Automotive Technology and supported by the Ministry of Education Industry-Education Collaboration Program, this study addresses a significant quality control challenge in automotive manufacturing: the efficient and accurate detection of rubber seal ring assembly status on pipe fittings. The traditional approach of manual visual inspection is described as having low efficiency and high false detection rates, motivating the development of an automated detection system.

Core Technical Findings

The improved Faster RCNN model incorporates several architectural modifications to enhance detection accuracy while reducing computational cost. The model achieves 91.45% detection accuracy on a dataset of 5500 images, which the authors state meets practical production requirements.

The detection system is designed to identify whether the rubber seal ring is correctly installed on the pipe fitting, detecting both presence/absence and potential misalignment. The improved model reduces the number of parameters and computational cost through depthwise separable convolutions and CSPNet structure, while maintaining or improving detection accuracy through the channel shuffle unit and Mish activation function.

Interpretation of Key Technical Points

Network Architecture Improvements

The replacement of VGG16 with ResNet50 as the backbone network provides a significant improvement in feature extraction capability. ResNet50's residual connections enable the training of deeper networks without degradation, allowing the model to learn more complex and discriminative features relevant to seal ring detection. The deeper architecture captures both fine-grained texture features (important for detecting seal ring material properties) and structural features (important for understanding the spatial relationship between the seal ring and the pipe fitting).

The integration of CSPNet (Cross Stage Partial Network) structure addresses the vanishing gradient problem in deep networks by splitting the input into two paths—one processed through the full network and one passed directly to the output. This design reduces computational redundancy while maintaining gradient flow, which is particularly beneficial for the real-time processing requirements of production line inspection systems.

The adoption of depthwise separable convolutions decomposes standard convolutions into depthwise and pointwise operations, reducing the number of parameters by approximately 8–10 times compared to standard convolutions. This lightweight approach is essential for deploying the detection system on edge computing devices commonly used in manufacturing environments, where computational resources are limited but real-time performance is required.

Channel Shuffle and Mish Activation

The channel shuffle unit, borrowed from ShuffleNet architecture, enables information exchange between different channel groups after depthwise separable convolutions. Without channel shuffling, each channel group operates independently, limiting the model's ability to learn cross-channel feature relationships. The channel shuffle unit restores full connectivity between channels at minimal computational cost, improving feature representation for the detection task.

The Mish activation function replaces the conventional ReLU activation. Mish is a smooth, non-monotonic activation function that provides smoother gradients during backpropagation, which can improve training stability and convergence speed. For object detection tasks where precise localization is critical, the smoother gradient landscape of Mish can lead to better convergence of bounding box regression parameters.

Detection Performance Analysis

The 91.45% accuracy on 5500 images represents a practical level of performance for industrial quality inspection. In the context of automotive pipe fitting seal ring detection, this accuracy level suggests a false detection rate of approximately 8.55%, which must be balanced against the cost of manual re-inspection and the risk of shipping defective fittings. The dataset size of 5500 images is modest by modern standards but sufficient for the specific task of seal ring detection, which involves a relatively limited variety of defect types.

Process and Standards Analysis

Automotive pipe fittings are governed by standards such as ISO 1143 (Polyamide 6 for automotive applications), SAE J1401 (Automotive fluid transfer hoses), and various OEM-specific specifications. Seal ring assembly quality is a critical safety attribute, as improper seal ring installation can lead to fluid leaks, pressure failures, or contamination in fuel, coolant, or hydraulic systems.