Abstract:

In order to solve the problem of false alarms and missed alarms in pipeline gas leakage detection using a single sensor, timely warning and feedback of leakage status, a multi-source data fusion pipeline leakage detection method based on cross-attention was proposed. Firstly, the pre-trained ShuffleNetV2 model was used to extract spatial features from thermal imaging data. Then, a 1DCNN BiGRU model was constructed by combining a one-dimensional CNN (1DCNN) and a bidirectional gated recurrent unit (BiGRU) to extract temporal features from gas sensors. Finally, cross-attention was used to capture the spatiotemporal correlation of the data and obtain the feature representations of the two data sources. The residual method was used to connect the features and input them into the classification layer to obtain the recognition results. The results show that the constructed SCGA model has a gas recognition accuracy of 99.22%, and the loss value fluctuates between 0-0.04. Compared with support vector machines (SVM), 1DCNN, and BiGRU models that only use gas sensor data, the accuracy is improved by at least 4.12%. Compared with MobileNetV3, ShuffleNetV2, and ResNet18 models that only use thermal image sensor data, the accuracy is improved by at least 1.14%. Compared with the multi-source data fusion model SCG, which simply connects temporal and spatial features, the accuracy is improved by 1%. It was verified that the SCGA model has high accuracy.

Key words: cross-attention, multi source data fusion, gas leak detection, convolutional neural network (CNN), bidirectional gated recurrent unit (BiGRU)