Moskalenko Y. Methods of recognition by a diagnostic signal based on hybrid neural networks

Recognition by a diagnostic signal is a common problem in the implementation of automatic monitoring/contriol and diagnostics systems. Common tasks include the classification based on a characteristic signal specified and set as a time series, and semantic segmentation based on diagnostic images. With the development of convolutional neural networks, new perspectives have been opened for solving such problems quickly, in real time and with propper accuracy. However, in many cases, solving the problem of semantic segmentation requires pixel accuracy, which is not always provided by the basic methods of recognition on convolutional neural networks. Moreover, there are such cases of diagnosis that cannot be effectively solved on such networks. First of all, it happens when the values ​​of the characteristic vector are lost. Sometimes it can be noticed when anomalies occur or transients begin in the object under study. At the same time, another type of neural network — Self-organizing maps (SOM)— has the property of learning under an indefinite set of classes and forming new clusters for previously unknown classes. So, SOM provide an opportunity to solve problems of this propper type. Therefore, the study of both approaches to the research of neural network is made in order to improve the efficiency of each of them and their integration into a hybrid neural network. It was firstly proposed: - a method of increasing a receptive field of neurons of convolutional neural networks based on the aggregation of maps of features of different dimensions to increase the accuracy of classification of signals of large dimensions; - a method for determining a correspondence of the lattice neurons of the trained SOM to the input vector with lost uncertain components in order to increase the accuracy of the classification; - a model of forming averaged feature maps in convolutional neural networks based on SOM to increase the accuracy of solving classification and semantic segmentation problems. The next positions were improved: - a method of semantic segmentation on the basis of deep learning networks due to forced selection of image contours in the network decoder: FPN, PSPNet, DeepLab v3, U-Net and due to aggregation of different subsamples of the feature map to improve the process of recognition. The corresponding neural networks were implemented according to the proposed methods and their verification was performed in relation to the existing neural networks of classical architectures. The effectiveness of the proposed methods for solving diagnostic problems has been experimentally proven.