Neural network analysis of vibration signals in the diagnostics of pipelines

Abstract

The article is devoted to the improvement of heat network calculation and diagnostics methods. Currently used instruments have many shortcomings for the diagnosis of pipelines, including low reliability of defect detection and subjective decision-making. The authors created an experimental stand, which allows to conduct the diagnostics of pipelines by a vibration-acoustic method. They studied steel pipes filled with water, the surface of which 50x50 mm defect and the depth of thinning of 2 mm, 3 mm, and 5 mm. Using the vibrationacoustic sensors fixed on an outer surface, the vibration signals generated by the water flow in the pipe were obtained. In order to process the large volumes of data obtained as the result of experiments, it is proposed to use artificial neural networks. Among all considered types of neural networks, the authors prefer Kohonen's networks due to the best effectiveness of a defect determination. The program for an acoustic signal processing and analyzing through a neural network was implemented in LabView 8.5 work environment. Depending on the accuracy of a problem being solved, and the details of a training sample, the program is able to produce the results of sample classification of samples for a defect-free and defective pipes of different depth of damage. The results of the classification by Kohonen's trained neural network show good abilities for the analysis of unknown samples and a high degree of their recognition reliability.

Keywords: diagnosis, corrosion, defect, pipelines, acoustic signal, neural network