Kohonen Network-Based Adaptation of Non Sequential Data for Use in Convolutional Neural Networks

Convolutional neural networks have become one of the most powerful computing tools of artificial intelligence in recent years. They are especially suitable for the analysis of images and other data that have an inherent sequence structure, such as time series data. In the case of data in the form of vectors of features, the order of which does not matter, the use of convolutional neural networks is not justified. This paper presents a new method of representing non-sequential data as images that can be analyzed by a convolutional network. The well-known Kohonen network was used for this purpose. After training on non-sequential data, each example is represented by so-called U-image that can be used as input to a convolutional layer. A hybrid approach was also presented, where the neural network uses two types of input signals, both U-image representation and the original features. The results of the proposed method on traditional machine learning databases as well as on a difficult classification problem originating from the analysis of measurement data from experiments in particle physics are presented.

The Study of Usefulness of a Set of Fractal Parameters to Build Classes of Disease Units Based on Images of Pigmented Skin Lesions

Dermatoscopic images are also increasingly used to train artificial neural networks for the future to provide fully automatic diagnostic systems capable of determining the type of pigmented skin lesion. Therefore, fractal analysis was used in this study to measure the irregularity of pigmented skin lesion surfaces. This paper presents selected results from individual stages of preliminary processing of the dermatoscopic image on pigmented skin lesion, in which fractal analysis was used and referred to the effectiveness of classification by fuzzy or statistical methods. Classification of the first unsupervised stage was performed using the method of analysis of scatter graphs and the fuzzy method using the Kohonen network. The results of the Kohonen network learning process with an input vector consisting of eight elements prove that neuronal activation requires a larger learning set with greater differentiation. For the same training conditions, the final results are at a higher level and can be classified as weaker. Statistics of factor analysis were proposed, allowing for the reduction in variables, and the directions of further studies were indicated.

Kohonen’s algorithm in problems of classification of defects in printed circuit assemblies

The article presents a new method for diagnosing the technical condition of radio-electronic components, combining the methods of thermal diagnostics with the technologies of artificial neural networks. The structure of the method is shown, and the composition of the functional blocks is determined. The implementation of the method is a symbiosis of technologies for mathematical and simulation modeling of the technical state of a radio-electronic device with its physical tests and research of characteristics. When developing the method, specialized software tools for design and circuit design were actively used, such as Altium Designer CAD, SolidWorks, NI Multisim, the FloTHERM PCB thermal analysis module, as well as the MATLAB mathematical modeling and calculation package. With the help of these tools, a number of studies were carried out, including sets of numerical values of the power of circuit elements and temperature indicators of the printing unit, both for the correct state of the device and in states with artificially introduced defects. They, in turn, became the basis of the database of electronic node failures. To implement diagnostic procedures and identify the technical condition, an artificial neural network based on selforganizing Kohonen maps was created, its structure, parameters and algorithms of functioning were determined. The diagnostic procedure is based on the analysis of information from the fault database and its comparison with experimental data obtained as a result of a physical experiment. The results of the study showed that the network automatically classifies the characteristic defects of electronic components using the algorithms embedded in it. The list of characteristic defects in the proposed diagnostic method is limited to a discrete set of the most common faults, because, as their number increases, the use of the self-organizing Kohonen network for automatic classification becomes much more complicated and ineffective in terms of performance and reliability of identification. Among the advantages of this technology, it should be noted that the Kohonen network has the ability to convert largedimensional input data into a two-dimensional array. So, the results are easy to visualize and convenient to use when generating reports and recommendations for subsequent decision-making about the possibility of using an electronic device.

Kohonen Artificial Networks for the Verification of the Diameters of Water-pipes

The design of the water distribution system is inherently linked to the execution of calculations, which aim, among other things, to determine the flow rate through individual pipes and the selection of diameters at the appropriate speed. Each step in the calculations is followed by an evaluation of the results and, if necessary, a correction of the data and further calculations. It is up to the designer to analyse the accuracy of the calculation results and is time-consuming for large systems. In this article, a diagnostic method for the results of hydraulic calculations, based on Kohonen Network, which classifies nominal diameters [DN] on the basis of data, in the form of flows, has been proposed. After calculating the new variant of the water distribution system, the individual calculation sections are assigned to the neurons of the topological map of Kohonen Network drawn up for nominal diameters. By comparing the diameter used for the calculation, with the diameter obtained on the topological map, the accuracy of the chosen diameter can be assessed. The topological map, created as a result of labelling the neurons of the output layer of the Kohonen Network, graphically shows the position of the classified diameter, relative to those diameters with similar input values. The position of a given diameter, relative to other diameters, may suggest the need to change the diameter of the pipe.

APPLICATION OF A CONVOLUTIONAL NEURAL NETWORK AND A KOHONEN NETWORK FOR ACCELERATED DETECTION AND RECOGNITION OF OBJECTS IN IMAGES

One of the most effective ways to improve the accuracy and speed of algorithms for searching and recognizing objects in images is to pre-select areas of interest in which it is likely to detect objects of interest. To determine areas of interest in a pre-processed radar or satellite image of the underlying surface, the Kohonen network was used. The found areas of interest are sent to the convolutional neural network, which provides the final detection and recognition of objects. The combination of the above methods allows to speed up the process of searching and recognizing objects in images, which is becoming more expedient due to the constantly growing volume of data for analysis. The process of preliminary processing of input data is described, the process of searching and recognizing patterns of aircraft against the underlying surface is presented, and the analysis of the results is carried out. The use of the Kohonen neural network makes it possible to reduce the amount of data analyzed by the convolutional network by 18–125 times, which accordingly accelerates the process of detection and recognition of the object of interest. The size of the parts of the input image fed into the convolutional network, into which the zones of interest are divided, is tied to the image scale and is equal to the size of the largest detectable object, which can significantly reduce the training sample. Application of the presented methods and centering of the object on training images allows accelerating the convolutional network training by more than 5 times and increasing the recognition accuracy by at least 10%, as well as minimizing the required minimum number of layers and neurons of the network by at least halving, respectively increasing its speed