Abstract: In order to detect valve leakage in gas pipelines, an improved AlexNet network architecture is studied, an ultrasonic signal recognition method for valve leakage based on an improved convolutional neural network (CNN) is proposed. Due to short-term and narrow-band line spectrum features of the leakage signals, the "square" convolution kernel, commonly used in image recognition, is changed to "flat" based on the perspective of image neighborhood information density. At the same time, the AlexNet layers are optimized, the number of convolution kernel and neurons in the fully connected layers are re-determined, and the small-scale convolution kernel is selected to increase the network capacity and model complexity while reducing the number of parameters to prevent model overfitting. The two-class and multi-class models with different leakages are established respectively, and the data set is collected through experiments to generate corresponding time-frequency diagram samples as well, including leakage signals at different valve openings and pipeline pressures and background acoustic signals. It is shown that the improved CNN classifier achieves higher recognition performance on the test set than the traditional CNN classifier.

Key words: convolution kernel, short-time Fourier transform, convolutional neural network, valve, leakage detection