MBDSF-Net: An Efficient Network for In-line Inspection in Long-Distance Oil and Gas Pipelines

In long-distance oil and gas pipelines, defects are generated owing to the extension of service life and environmental impacts. In practical for in-line inspection(ILI), several challenges exist, including long pipeline mileage, large volume of inspection data, numerous abnormal signals, and low efficiency of manual defect identification. To achieve efficient processing of inspection data and automatic identification of pipeline defects, a complete pipeline defect inspection process and the Multi-Branch Depth-Sense Fusion Network (MBDSF-Net) have been proposed. For the obtained ILI data, in the stage of inspection data analysis and processing, mileage accuracy optimisation and abnormal signal elimination operations are implemented. Additionally, an efficient method for plotting inspection data is proposed. This method generates inspection images for 316.87 km within 1 h, achieving rapid conversion of inspection data into images. The detection network utilises Inception Depthwise Convolution (IDWC), which expands the receptive field while improving computational efficiency and model accuracy. Furthermore, the Multi-Scale Deep Feature Fusion Network (MSDFF-Net) is proposed to enhance the representation capabilities of features at each level. The Inception-enhanced Mixed Aggregation Network (IMANet) is developed to reinforce backbone feature extraction to achieve more discriminative representations. Experimental results demonstrate that MBDSF-Net achieves an mAP50 of 91.0% and an inference latency of 4.01 ms on the self-built PIP-813 dataset, achieving a desirable balance between accuracy and speed. The proposed inspection data processing method and MBDSF-Net improve the processing efficiency of inspection data and the accuracy of pipeline defect identification, making them highly suitable for practical engineering applications.