Rupich S. Multi-class recognition of objects technical condition by a classifier based on Neural Network

The thesis describes that complex spatial objects are usually operated in difficult accessible places in zones with increased external influences and dynamic loads. It is described that multi-site damage can be aroused under such conditions of operation due to imperfection of the elements of the construction that are used under the action of complex loading. It is shown that in order to ensure the safe operation of complex spatial objects that are characterized by multi-class of possible technical states, it is necessary to carry out continuous monitoring of structural integrity and provide multi-class diagnostics. The thesis describes that the most rigid conditions for the preservation of integrity put forward to the welded tanks with ecologically dangerous substances. It is presented the general structure of the functional diagnostics system for monitoring the technical condition of the tank. In the dissertation the subsystem of decision-making is substantiated and developed to improve the functional diagnostics system of the technical condition of the welded tank. It is substantiated that using Probabilistic Neural Network for the development of the classifier. The general structure of the neural network classifier is developed. Models of processes for forming sets of input vectors of diagnostic features for such diagnostic tasks as localization of single damage, localization of multi-site damage, monitoring of damage development and monitoring of structural degradation are developed. A generalized information model of the system of multi-class recognition, which combines the diagnostic tasks, is developed. As a result, the probability of recognition from the network influence parameter, which shows the effectiveness of the neural network classifier for diagnostic tasks, was established. It has been established that a error-free multi-class recognition of the object’s status is achieved. The influence of the parameters of a neural network and characteristics of diagnostic vectors on the probability of multi-class recognition of the object’s state for monitoring of structural degradation is investigated. It is shown that the classifier provides an error-free recognition depending on established parameters of the neural network. It is identified that using a classifier based on Probabilistic Neural Network is effective for the recognition of cracks in welds of tanks. It is shown possibility of using the designed neural network classifier for recognizing real objects has been confirmed. Keywords: complex dimensional object, tank with welded joints, multi-focal damage, Structural Health Monitoring, multi-class recognition, neural network classifier, Probabilistic Neural Network, vector of diagnostic signs, classification efficiency.