Novel intelligent diagnosis method of oil and gas pipeline defects with transfer deep learning and feature fusion

Abstract

The destruction of oil and gas pipelines may result in enormous financial loss and significantly affect public safety. Hence, early defect diagnosis of oil and gas pipelines is of immense significance. In order to improve the accuracy and reliability of oil and gas pipeline defect detection and diagnosis with limitation of small and poor data sets, this paper proposes intelligent diagnosis and recognition method based on Transfer deep learning, Continuous Wavelet and Short-time Fourier Time-Frequency feature fusion, and Strengthen Convolutional Neural Network (TWSC). Oil and gas pipeline defects are converted into identifiable defect signals by the three-coil bidirectional excitation detector. Feature fusion, focusing on different feature distribution composed of Short-Term Fourier Transform (STFT) and Continuous Wavelet Transform (CWT), is introduced as the input of the TWSC model. It effectively extracts the time-frequency features of the defect signal on different feature distribution, and complements each other. A deep convolutional neural network with strengthen convolution kernel is constructed as a diagnostic model, expanding receptive field of diagnosis. For most practical engineering problems that only have a small data set with partial poor samples, transfer learning is introduced in the model to optimize the diagnostic performance. Parameter transfer from transfer network is operated to initialize the diagnostic model parameters. The limitation of insufficient training with a small data set and interference of poor samples during model initial training progress are both greatly improved. Under the conditions of oil and gas stations and laboratories, the typical oil and gas pipeline defects are collected to analyze. For simulating complex working conditions, the diagnosis performance of add-noise signal is verified. The final results show that TWSC intelligent diagnosis method proposed in this paper has a good performance on diagnosis accuracy and stability in defect diagnosis of oil and gas pipelines.