A structural-functional model of constructing the receiving system IR-UWB signals in the range of very high frequencies of the elements of intelligent control, which is based on a separate plane of physical infrastructure and plane management that allows you to automate efficiently manage the process of joint use of resources physical infrastructure and methods of artificial intelligence. In contrast to existing models receiving of IR-UWB signals terahertz (THz) range, it is able to provide protocol and infrastructural collect the necessary data for intelligent algorithms. The proposed physical infrastructure has a module training and optimization, which involves the use of existing simulation model radio THz range from 110 to 170 GHz for testing of intelligent algorithms control power capacity radioliniy IRUWB signals in the range of very high frequencies. The developed algorithm collecting data involves tracking the status of blocks receiving complex for efficient collection of data actually to use changing values as metrics Euclidean distances well and metrics functional technical parameters in relation to the number of clusters.
The development of 5G networks and the further increase in the density of small cells and access points leads to the need to pay significant attention to the radio channel between the access point and the Internet. In many cases, the use of a wired connection to organize such a communication channel is impractical for economic reasons. In these circumstances, a wireless connection that works in the terahertz (THz) range and frequency are an attractive resource for building a high-speed wireless network connection.
At the same time, national regulators of many countries have been identified terahertz range of frequencies and set limitations on the spectral power radiation ultrawideband signals (NSHS) for their unlicensed first use. In virtue of limitations on the spectral density range radiation NSHS device connection is 10-50 meters, so the main area of application NSHS connection - this wireless network communications and wireless sensor networks of small radius of action (local and personal networks). For that to expand the scope of application of technology NSHS terahertz communication should increase the range of action of receiving and transmitting devices in the mode of "point to point".
The steady increase in the variety and volume of information flows in a telecommunication network, the urge to solve scientific and practical task of developing infrastructure gear -pryymalnyh of a range of very high frequencies to ensure effective management radioliniy IR-UWB signals THz range on the basis of algorithms, machine learning and neural networks with regard setting energy efficiency.
For you to solve this problem is of interest complex (combined) application: - methods of receiving a multibeam signal; - principles of construction and algorithms of signal formation and processing in multiposition communication systems; - methods of artificial intelligence and machine learning.
Development of innovative methods to increase the efficiency of managing the use of energy
potential of radiohertz range lines to increase noise immunity and range of low-orbit communication
system is a new scientific direction [
The aim of this work is to develop methods increase the efficiency of managing the use of the energy
potential radioliniy on the basis of spatially distributed devices [
There is a method of receiving a multi-beam signal [
The disadvantage of this method is the fact that the choice of threshold h, which is used to identify clusters of rays based on the initial evaluation of pulse characteristics of the channel. However, the disadvantage of this approach is the fact that the estimates obtained from the analysis of the received test signal and is virtually in all the ways there is a serious engineering problem selection threshold adoption decision is either not indicated or she is given enough attention.
In [
In addition, the implementation of this method does not take into account the fact of distortion of
the time form of the pulses of the receiving system IR-UWB signals THz bands [
The expansion of the pulses leads to a decrease in their amplitude, which reduces the efficiency of pulse selection against the background of noise and interference. This is due to the fact that in a nontime of the pulse duration in the signal, will be equal to: periodic sequence of pulses, the average duty cycle ̅, as the ratio of the signal duration to the total where V is the transmission rate, n is the number of pulses in the signal, is the pulse duration.
And accordingly, the average duty cycle of the signal should be as high as possible 1. Otherwise, each of the signals will be a dense flow of pulses, which will prevent code division of signals in a network where many terminals operate simultaneously. The average duty cycle of the signals, in essence, determines the possibility of ensuring their orthogonality, ie the possibility of signal separation.
One should also note that the value of the average duty cycle ̅ determines the average power of the transmitted signal ̅, if you know the peak power of the pulses i by the following formula: ̅ =
=
, ̅ = ⁄ ̅ (1) (2)
Therefore, the only way to increase the signal power with restrictions on the duration and number of pulses is to increase the amplitude (peak power) of the transmitted pulses.
The authors in [
The authors propose the use of intelligent algorithms based on trained models of artificial intelligence, which will be used at the physical level of the proposed architecture of the receiving system IR-UWB signals THz bands using algorithms based on artificial intelligence, as well as at the control plane level (Fig. 1). algorithms based on artificial intelligence
The work of these algorithms is focused on managing the use of energy of radio lines with IR-UWB signals and as a result of optimizing the use of resources of their energy potential. For example, the use of neural network-based algorithms to implement wavelet transform avoids a large number of calculations and greatly speeds up the search for wavelet decomposition coefficients to make more efficient use of the spectral resources of the communication channel.
Intelligent control algorithms on the SDN controller can perform optimization at the level of the whole complex. However, for such algorithms that operate at the physical level and the management level must have the appropriate infrastructure for data collection, training, testing and updating of relevant trained models.
Figure 2 shows the neural network infrastructure of the developed innovative solution. Its components are the following modules: operation of ML algorithms on SDN controller; neural network training; training and optimization of the neural network; operation of algorithms using a neural network on blocks of an innovative solution.
Maintaining this infrastructure should cost less than the benefits it should provide. Many works
present the use of single algorithms based on neural networks that optimize a particular process or part
of the device [
The existing methods [
The search unit (see Fig.1) has a device that determines the value of the decision function for a given discrete time delays of the multipath, compares the value of the generated decision function with a given threshold h and generates estimates of search time delays of the beam component over the threshold h.
As is well-known, when receiving broadband multi-beam signals, a search procedure is also performed, which, as a rule, is a scan of the uncertainty area with the detection of the signal at each of its points.
The disadvantage of the known treatment approaches multipath signal along through the use of search procedure is that during this search procedure signal radiation is not considered the impact of multipath signal components to each other. As a result, the probability of erroneous detection of beam signals increases. In addition, there is no optimization of the number of beam signals used to obtain soft decisions about information symbols, which leads to inflated requirements for hardware implementation without increasing the quality of the selected information.
From this point of view, it is advisable to introduce a tracking procedure and identify changes in the properties of non-stationary processes in the signal search algorithm, as well as adaptation to changes in the noise level.
The block diagram of the signal pre-processing system using wavelet processing to obtain the input image of the neural network of the proposed approach is presented in Fig. 3.
transformation
The peculiarity of the proposed approach is that in order to adapt digital signal processing to a timevarying noise, in solving the problem of recognizing the presence or absence of a signal at a given interval, a method based on wavelet transform and neural network is proposed.
Wavelet transform allows you to more accurately localize the frequency properties of the signal over time and does not increase the amount of data in the transition from the time representation of the signal to its representation in the wavelet region. The best frequency separation is provided by the filter with the highest frequency response. Such a trait is characteristic of Dobeshi's wavelet.
As the order of the Dobeshi filter increases, its frequency response tends to be ideal. But it is
necessary to consider that application of the filter with the long impulse characteristic leads to very
appreciable distortion. Therefore the most appropriate use of the filter Daubechies 4 kofleta 2 and
simmleta. Biortohonalni wavelets Daubechies allow to reduce the amount of computation by the
decomposition by using short filters [
The algorithm for dividing the intervals of signal presence or absence is constructed taking into account the peculiarities of the multipath signal propagation, which are described perceptually by the model. The model divides the spectrum of a multibeam signal into frequency bands, into so-called critical intervals.
The developed algorithm uses a wavelet - the transformation of a multibeam signal on the biotogonal basis of Dobesha, and to decide on the types of interval (segment) of the signal—a neural network on a multilayer perceptron.
Compared with traditional spectral methods, wavelet transform provides a more accurate localization of the signal in time and frequency (in the decomposition subbands) and has a fast implementation algorithm. The biotogonal basis preserves the phase ratios of the frequency components of the signal after its recovery by inverse wavelet transform.
Further improvement of the proposed method was achieved by reducing the dimension of the input vector of the perceptron, which allowed to reduce the number of training samples and speed up the learning process. Reducing the dimensionality of input vectors in neural network learning, based on the principal components algorithm.
The results of the study of the dependence of the probability of recognition error on the dimensionality of the feature vector after the transformation by the method of principal components is shown in Fig. 4. The presented results show that the dependences for the total probability of recognition error (curve above) and root mean square error (RMS) transformation are almost monotonically decreasing, with the selection of principal components and reducing on this basis the dimensional space of features gives a more significant effect with increasing correlation signs.
The Levenberg-Marquardt Algorithm (LMA) algorithm was used to train the multilayer perceptron, which is the most common algorithm for minimizing quadratic deviations. Its advantages, in comparison with the method of gradient descent, are the high speed of calculations.
In the developed architecture it is proposed to use the module of training and testing which allows to adjust intelligent control algorithms, namely processes of preparation, testing and an estimation before their deployment in the innovative decision.
Simulated data or data from a real receiving system of IR-UWB signals of THz bands can be used for training or testing of such algorithms. This approach allows you to better prepare for the launch of such algorithms on real networks and reduce the associated risks (Fig. 1).
To evaluate such algorithms, it is necessary to introduce certain functional metrics of the developed system, on the basis of which these algorithms will be evaluated. The training and testing module includes a simulation model of a wireless telecommunications system in the terahertz range based on the use of IR-UWB multibeam signals of picosecond duration with appropriate components for maximum proximity to the operation of the actual developed receiving device.
As a quality environment simulation modeling of selected software complex NI Multisim 13.0 [
Functional metrics of the developed system of evaluation of overtrained neural network should be the usual functional parameters of this solution: signal-to- noise ratio, pulse duration of UWB signal, efficiency of use of spectral resources of communication channels and others. If the overtrained model for a certain algorithm has led to deterioration of functional parameters then this model should be returned or should be removed from the storage module.
For that to intelligent algorithms for control given the correct result should collect sufficient set of data, by which is meant the optimum amount of data with which training models considered complete and is not observed so -called process overfitynhu.
An SDN controller is usedt collect functional data, which directly carries out the process of collecting the relevant data from the blocks of the receiving system IR-UWB signals of the THz band.
Each unit receiving system is a source of information for ML algorithms, as can be and the purpose of their use. The main data processing before training neural networks is carried out on the SDN controller. All the collected functional parameters that are necessary for the training of neural networks are stored in the FE cloud. Only intelligent algorithms on the SDN controller have access to these parameters. If there is a change in the state of the receiving system IR-UWB signals in the range of extremely high frequencies, which requires retraining of the respective models, the corresponding algorithms carry out the procedure of retraining them on the basis of new FE parameters. After that, the corresponding trained models in the cloud are replaced. When a new version of the model for the corresponding block appears, the block downloads the updated version of the model.
In work used cluster approach to determine states receiving system IR-UWB signals in band very
high frequency of training without the supervision of the use of ML algorithms k- means (k - average)
and c- means (c - average) for developed algorithm collecting data [
This approach allows you to determine on the basis of a certain period of time the necessary states of the proposed device or the emergence of new states. Besides that, this approach allows to take into account the greater number of functional technical parameters with minimal change in software security.
An important part of intelligent algorithms management is the collection of direct data for the study. One of the features were using these algorithms in telecommunication radio networks THz range is the variability of conditions in the sector, as well as appearances and new and disappearance of current conditions that require additional collection of data and overtraining neural networks.
In the study [
The developed algorithm of data collection provides tracking of a condition of blocks of a receiving complex for rational data collection actually using change of values, both metrics of Euclidean distances, and metrics of functional technical parameters, in relation to number of clusters.
The novelty of this approach is in contrast to the classical implementation that the introduced metrics signal-to-noise ratio and pulse duration UWB signal, instead of Euclidean distance metrics, which allows to take into account the spatial characteristics of signal propagation in the process of selfoptimization of the developed receiving structure. 4. Conclusion
1. An algorithm for tracking a multibeam signal of a system for receiving signals from spatially spaced low-power transmitters has been developed, the feature of which is the specification in the process of tracking not only the time positions of components, but also their number. A distinctive feature of the developed algorithm is that it is built using wavelet - processing to obtain the input image of the neural network.
2. The structural-functional model of construction of the receiving system of IR-UWB signals of THz bands with elements of intelligent control of the receiving complex at the physical level and control of the complex is offered. Unlike the existing infrastructures of the receiving system IR-UWB signals THz bands, these radio networks are not able to provide protocol and infrastructure collection of the necessary data for intelligent algorithms. The proposed infrastructure has a training and optimization module, which provides for the use of the existing simulation model of the THz radio line in the range from 110 to 170 GHz to test intelligent algorithms for controlling the energy potential of IR-UWB radio lines of the THz band. 3. Further research is aimed at modeling and studying the effectiveness of the proposed solutions based on the developed simulation model of energy potential control of IR-UWB radio lines THz signals, as well as the development of a web application that will reflect the real state of the developed receiver. functional parameters to obtain the desired results.