Pipeline leak detection based on multiscale convolution neural network and improved symmetric dot pattern optimized by grasshopper optimization algorithm

Abstract

In the engineering of pipeline condition identification, the complexity of pipeline signal components often results in insufficient feature extraction with traditional feature extraction-machine learning methods, thereby affecting the recognition performance. In order to effectively address the aforementioned issues, based on deep learning, we propose a multiscale convolution neural network (MCNN) to effectively identify pipeline conditions by classifying improved symmetry dot pattern (ISDP) images of one-dimensional negative pressure wave signals of pipelines. First, we propose the ISDP transformation method, considering that negative pressure wave signals of pipes with different leakage degrees have different amplitude changes. The ISDP transformation method transforms the negative pressure wave signal of the pipeline from one dimension to two dimensions. Then the grasshopper optimization algorithm (GOA) was employed to optimize the parameters of the ISDP algorithm. Second, we build the MCNN depth network to train and classify the ISDP image. The MCNN can simultaneously learn both the global and local features of an image. The corresponding evaluation indicators show that the proposed method of working condition recognition using MCNN to classify and recognize the ISDP image of pipeline signal has higher accuracy and robustness than traditional machine learning methods and common deep learning methods. The evaluation results prove that the proposed algorithm is effective in pipeline signal classification.