Real-Time Identification Of Natural Gas Pipeline Leak Apertures Based On 1DCNN

As the world enters the double carbon era,the demand for clean energy such as natural gas is increasing,and the risk of tiny pipeline leaks is getting bigger and bigger.If the size of pipeline leakage aperture can be detected quickly and accurately in real time,it is beneficial for the relevant departments to take corresponding measures quickly according to the severity of the leakage,so the research on real-time identification of natural gas pipeline leakage aperture is of great significance.To address the problems of large number of parameters,large memory consumption,poor noise immunity,and unsuitable for real-time identification,a method based on One-Dimensional Convolutional Neural Networks(1DCNN)is proposed for real-time identification of pipeline leak apertures.The main research contents are as follows:1.This paper builds a real-time detection platform composed of signal acquisition module,piezoelectric sensor and jetson nano,introduces the structure and hardware composition of the system in detail,analyzes and explains the pipeline leakage experimental device,and compares the real-time detection platform with the proposed The combination of algorithms verifies the real-time and effectiveness of the real-time detection platform in pipeline leakage aperture identification,especially under the condition of limited hardware resources,and the identification accuracy and execution time of the system are studied.2.Deep learning has been widely used in the identification of pipeline leakage apertures.However,most of them are designed and implemented for offline data.There are problems such as complex model structure and poor anti-noise performance.In order to solve this For the problem,a lightweight residual convolutional neural network,L-Resnet,is proposed.Based on the deep separable technology,two new separable residual modules are constructed to realize the feature extraction of pipeline leakage signals;a more efficient activation function is applied to high-dimensional space to enhance the nonlinear ability of the model,and a lightweight attention module is used The extracted feature weights are distinguished,the weights of important features are obtained,and finally the size of the leakage aperture is obtained through a classifier.Experimental results show that this method can identify the leakage aperture of pipelines in real time and accurately.Compared with other models,it has higher computing efficiency and good robustness.3.Aiming at the problems of L-Resnet model that the model is not lightweight enough,real-time performance and anti-noise performance can still be improved,a pipeline based on dynamic depth separable convolutional neural networks(DDS-CNN)is proposed A method for real-time identification of leak apertures.Construct a new dynamic depth separable convolution module,through the organic combination of dynamic convolution and separable convolution,to realize the feature extraction of leakage signals;obtain better feature receptive fields by alternating large and small convolution kernels;Based on the neural architecture search(NAS)method,the number of channels in the network layer is optimized,and three efficient network architectures are constructed;finally,the classification results are obtained through the classifier.Experimental results show that the method is verified in three different positions and obtains good results.While maintaining a high recognition accuracy,the method in this paper has a low number of parameters and flops,shortens the training time,improves the recognition speed,and is suitable for leak monitoring of industrial pipelines.