This paper presents the results of a study aimed at developing a method for semantic segmentation of thermal images using a modified neural network algorithm that differs from the original neural network algorithm by a higher speed or processing graphic information. As part of the study, a modification of the DeepLabv3+ semantic segmentation neural network algorithm was carried out by reducing the number of parameters of the neural network model, which made it possible to increase the speed of processing graphic information by 48% – from 27 to 40 frames per second. A training method is also presented that allows to increase the accuracy of the modified neural network algorithm; the accuracy value obtained was 5% lower than the accuracy of the original neural network algorithm.

Keywords: neural network algorithms, semantic segmentation, machine learning, data augmentation

The article presents the results of an on-site test spatial metal farm of an aircraft hangar with a span of 72 m. A comparative analysis of the data obtained by the simulation of the farm from the predicted loads (taking into account the actual geometry) and the on-site tests is performed. During the assessment of the technical condition and testing, significant errors were identified in the design and manufacture of gate frames structures,which could lead to loss of bearing capacity or unsuitability for normal operation. Based on the test results, measures were developed and implemented to strengthen the elements of the farm, repeated full-scale tests were conducted. Based on the foregoing, the span structure is recognized as suitable for safe operation.

Keywords: test, metal truss, bolted connections, compliance, bearing capacity, deformability, gate frame, reliability, static calculation