|
Original research
THE DETERMINATION OF DEFECTS IN STEEL PIPES USING COMPUTER VISION METHODS BASED ON THERMOGRAPHIC IMAGESPages 425-436
Abstract:
Ensuring the reliability and safety of pipeline transportation systems in industry is a critical area of technical service. This is especially important when steam, hot water, oil, or other hazardous liquids are transported through pipes, as leakage can lead to serious emergencies. The development of new non-destructive testing methods and computer systems for analyzing data and detecting defects has significantly improved the accuracy and speed of pipeline inspections. For this research, a dataset of 210 thermographic images was collected from five types of cooling system pipes in a six-cylinder four-stroke engine. Existing methods for processing thermographic data were reviewed and neural network models were selected for training. The data was preprocessed and labeled. During dataset processing, computer vision methods based on convolutional neural networks were applied. The YOLOv8 neural network architecture was used. As a result, a neural network was developed that allows for the identification of volumetric and extended defects in thermographic images of pipes. In the determination of volumetric defects, an accuracy of 92.2% was achieved with a response time of 870 milliseconds. For extended defects, the mask accuracy was 58.6% and the response time was 800 milliseconds. The identification of point defects was hindered by interference in the thermograms, but a mask accuracy of 51.9% was still achieved with a response of 286 milliseconds.
Keywords:
Thermodiagnostics, steel pipes, pipe defects, defect segmentation, neural network modeling, online augmentation.
|
Shcherban Pavel Sergeyevich, 
PDF 