NEURAL NETWORK SUPPORT FOR INTROSCOPY OF INTERNAL STRUCTURE AND PROPERTIES OF THE BUILDING CONSTRUCTIONS

Abstract

Introscopy is a process of contactless, non-destructive analysis of the internal structure of an object or processes in it using X-ray radiation, optical, acoustic, ultrasonic, seismic, electromagnetic waves of various ranges, modulation and coding principles. Its implementation involves methods for obtaining shadow, tomographic, radar and other images of the object of study. The resultingimage contains information about the object. Image analysis and decision- making about the object structure or its condition is carried out by an expert (operator). Obviously, the decision is made subjectively;its effectiveness depends on the qualification of the expert and can be significantly reduced because of the increasing number of errors and analysis time. In real conditions, the classification of the state of the object of study with a significant number of signs, with their unstable or uninformative degree of knowledge extraction, seems to be not a trivial task. To date, however, image recognition technologies based on artificial intelligence technologies have been developed and implemented that make it possible to synthesize neural network classifiers in vision systems that are invariant to the physical features of feature spaces of the studied object images. For introscopy technology, it has been prepared a neural network information-analytical, software and instrumental basis for solving the task of automating of the process of image visualization and its identification in the paradigm of designing and recognizing images in the space of shadow, tomographic, multi-view signs using statistical decision rules.The developed technology is represented in the form of an ensemble of neural network classifier models that implemented as independent software applications in the main code of an existing technical analysis package, for example, the neuro-emulator of StatSoft environment. The synthesis of classifier models according to the input data of images based on shadow and tomographic raster sweepsin a standard package of neuroemulators allows us to solve the problem with minimal cost and required quality indicators. Studies of the characteristic spaces of the introscopy process, the possibilities for the correct application of statistical decision rules, algorithms for the compulsory training of synthesized neural network models in the basis of existing technical data packages can improve the productivity of introscopy equipment by automating the analysis process, reducing the impact of subjective decisions, and reducing reaction times