GLD-Net: a lightweight detection method for ultrasonic signals of gas leakage with different apertures

Abstract In the industrial production process, safety accidents caused by the leakage of hazardous gases from pressure vessels are common. The assessment of leakage orifice detection is crucial. Traditional detection methods have low sensitivity. This study proposes a network model called gas leakage detection net (GLD-Net) for gas leakage detection, which utilizes an array of acoustic sensors for data collection of gas leakage signals. The network is based on the Yolov5n model as the baseline, and the C3 module of the backbone network structure incorporates the swift parameter-free attention block (SPAB) structure from the fast parameter self-attention network mechanism. The C3-faster-SPAB module is designed to achieve lightweight feature extraction. In terms of the neck network, GLD draws inspiration from the efficient reparameterized generalized-feature pyramid network (FPN) concept and designs the vision transformer with BI-level routing attention block (Biformer-Block), which is suitable for capturing narrow-band features of gas leaks. It integrates the lightweight biformer cross stage partial stage (Bi-CSPStage)module to enhance the expressive power of the FPN. The network replaces traditional convolution operations with depthwise separable convolution and reduces the scale of detection in the neck network. Experimental results show that after adjusting the depth of the network, the parameter size and computational complexity of GLD-Net are reduced by 96.7% and 96.1%, respectively. It exhibits higher average accuracy than other lightweight model algorithms on datasets with different noise levels, and achieves a frame rate of up to 500 FPS. It possesses fast and accurate detection capability, providing a reference basis for real-time leakage detection in industrial production and manufacturing.