Эволюционный алгоритм для автоматической генерации нейросетевых систем подавления шума

Полный текст

In the modern world, there are many sources and receivers of information signals, such as wired and wireless Internet, different access points, a huge range of radio waves, mobile sources, etc. All these sources are sensitive to various kinds of noises and disturbances that are associated with the signals mutual influence and the external factors influence that contributes to the mismatch of the transmission line, resonance phenomena, etc. [1]. The noise filters theoretical foundation is a spectral analysis, algorithmic basis is the fast Fourier transformation. Application of spectral analysis and classical filters requires careful adjustment of the parameters set, which makes it very difficult to implement the design automation of noise reduction systems. It results into necessity of finding new approaches. One of such approaches might be the use of intelligent information technologies intensively developing the last twenty years [2]. To date, there are several methods for noise reduction: shielding, grounding, signal filtration, noise reduction with adaptive filters, the signal wavelet analysis method, etc. All of them have limitations and some disadvantages (the requirement of a priori information about the signal or noise, the complex technology structure, expensive equipment, complex mathematical tools, the developers’ high qualification, etc.). Noise reduction system should be created to relax these restrictions and eliminate disadvantages, so that they will not require changes in the environment, a priori information about the signal or noise, expensive equipment, highly qualified developers and could be designed in an automated mode. Evolutionary methods (EM) are able to search in a complex space where the solution is a hierarchical structure or a combinatorial circuit. It does not use a priori information about the optimized function that significantly expands the application field of such methods. Neural network models are another famous method of data mining. Neural networks (ANN) are able to process large data sets, are resistant to noise, adapting to the changing problem conditions. In this paper, we propose the use an evolutionary algorithm for automatically generating the neural network based noise suppression system. The evolutionary algorithm automatically generating neural network based noise suppression system. The process of ANN models implementing and preparing for work consists of two main steps: neural networks structure selection (including activation functions threshold values adjustment) and tuning the neurons connections weight values. Moreover, neural network model can be adapted by adjusting the weights when new data would be received or problem conditions would be changed. Researchers seek to implement minimal neural networks architecture. In this case, the network generalizing properties are higher, the result obtained is more predictable and less time is required for signal processing. We propose to use evolutionary algorithm to generate such neural networks, in which the structure of the neural network is configured by a genetic programming algorithm (GP), and the weights and activation functions thresholds are adjusted by genetic algorithm (GA) followed by a hill climbing method [3]. Neural network structure configuring. The terminal and functional sets should be defined to generate the neural network structure using genetic programming algorithm (GP). Neurons or neuron blocks interconnected in a certain way can be chosen as the terminal set in the problem of generating the neural network structure. Then the operators that combine these neurons and their blocks in the network will be included in a functional set. The chosen encoding method must satisfy two conditions: insularity and sufficiency. The insularity condition requires that the admissible solutions would be obtained for any combination of functional and terminal elements. Two operators can be included in a functional set to satisfy these requirements, operators such as the installation of the terminal elements in a single layer and link layers. Sufficiency condition requires that the terminal and functional elements are sufficient for the task. A number of different activation functions and their combinations are included in the terminal set to satisfy this condition. Weights and activation function threshold values optimization. In this paper, we use a genetic algorithm followed by local search method for weights optimization. Studies show that the GA focuses individuals in attraction areas of local extremum points on the first iterations. Local search is easier to produce with help of the conjugate gradient algorithm. This algorithm is comparable to the effectiveness of second-order methods while using the first order derivative. The derivative numerical calculation extends this coefficients optimization method application on the neural networks with arbitrary structure. Results of the study. A complete investigation of all types of noise filtering problems formulations is not possible in one paper, therefore studies were carried out under following restrictions: - the test signals are periodic harmonic signals; - the test noise is a constant broadband noise (white noise); - the signal spectral analysis method is taken as a basis for the implemented model. After preliminary examination of existing intellectual information technologies, it has been established that the most appropriate technology for the initial research are artificial neural networks because of their ability to be automatically trained for solving the problem and to adapt to the external influences changes. The software environment MATLAB ® Neural Network Toolbox™ was used for pre-adaptation neural network technology to the noise suppression problem. Input data for training the neural network is a noisy harmonic spectrum of a periodic test signal. Test signal is a sine wave with 100 Hz frequency and amplitude of 1. Noise is broadband constant noise with average power equal to 4. Four neural network structures available in the Matlab were chosen for the comparative analysis based on the specific solved problem. The following structure neural networks have been the chosen: a cascading direct propagation network, an Elman network, a feedforward network with error back propagation; an autoregressive dynamically trained neural network. The hidden layers number, number of neurons in the hidden layer, training function and the neurons activation function are varied for each structure. The comparison results are presented in Table 1. The most effective structures were selected on the basis of the analysis. They are shown in bold in Table 1. Elman network and the feedforward network with back error propagation can be regarded as the best performance options. At the same time the latter network has only half the training time and a simpler structure that is significantly in terms of practical implementation. Therefore, we can assume that in our case direct distribution network with back propagation error must be considered as the best neural network that solves the noise suppression problem. Best Network has the following characteristics: 1 hidden layer, 5 neurons in the layer, bipolar sigmoid as activation function, the average training time is equal to one second, training error is equal to 0.01, the average signal/noise ratio is equal to 9,2 dB before training and 16,3 dB after training, the average variance of the processed signal is equal to 0.0179, the number of false positives alarms is equal to 55. Neural networks performance comparison Network Hidden layers number Neurons number Average training time, second Average variance Average signal/ noise ratio, dB False positives alarms number Cascading direct distribution network 3 10 1 0,0226 15,4 107 2 10 1 0,0231 15,4 109 1 10 1 0,0235 15,2 116 3 15 1 0,0214 15,4 89 Elman network 3 10 13 0,0263 16 112 2 10 4 0,0223 14,8 92 1 10 2 0,0191 15,6 51 1 10 3 0,0179 16,3 53 Feedforward network with back propagation error 3 10 1 0,5007 -0.02 1000 2 10 1 0,0221 15,4 91 1 10 1 0,0226 15,4 94 1 5 1 0,0176 16,3 55 Autoregressive dynamically trained neural network 3 5 0,1821 9,2 451 2 5 1 0,2285 9,3 449 1 5 1 0,0555 9,9 446 1 8 1 0,0193 15,9 62 [Image] The typical network structure The following settings were chosen for the program system [4] generating neural networks with arbitrary structure using the genetic programming algorithm: 1. The running time was equal to 4 generations with 20 individuals. 2. Tournament selection with three individuals was selected. 3. The initial depth of the tree is equal three. The trees are growing with full growth method. The experiments established that an efficient neural network successfully solving the noise suppression problem can be obtained after every run. The typical obtained structure is shown in Figure 1. The neural network has the following characteristics: average signal/noise ratio is 9,2 dB (before processing), 19,8 dB (after processing), the processed signal average variance with respect to the standard is equal to 0.0163, the false positives alarms number (signal/noise ratio is less than 10dB) reached 20. It can be concluded that the automatically generated network is the most effective because it has a very simple structure and the best results of signal processing. 3 Conclusions In this paper we proposed an approach to solving the noise suppression problem based on the ANN, presented the neural network structure automatically generated with the help of the genetic programming algorithm, conducted the statistical analysis of the results and substantiated the practical application possibility of the noise suppression neural network based method in digital communication systems.

Об авторах

Е. А. Попов

Сибирский государственный аэрокосмический университет имени академика М. Ф. Решетнева

Email: epopov@bmail.ru
доктор физикоматематических наук, профессор кафедры системного анализа и исследования операций Сибирского государственного аэрокосмического университета имени академика М. Ф. Решетнева. Окончил Красноярский государственный университет в 1973 г. Область научных интересов - физика твердого тела, технологии обработки экспериментальных данных.

М. Е. Семенкина

Сибирский государственный аэрокосмический университет имени академика М. Ф. Решетнева

Email: semenkina88@mail.ru
магистр техники и технологий, аспирантка кафедры системного анализа и исследования операций Сибирского государственного аэрокосмического университета имени академика М. Ф. Решетнева. Окончила Сибирский государственный аэрокосмический университет имени академика М. Ф. Решетнева в 2012 г. Область научных интересов - интеллектуальные информационные технологии, эволюционные вычисления

Л. В. Липинский

Сибирский государственный аэрокосмический университет имени академика М. Ф. Решетнева

Email: lipinskiyl@mail.ru
кандидат технических наук, доцент кафедры системного анализа и исследования операций Сибирского государственного аэрокосмического университета имени академика М. Ф. Решетнева. Окончил Сибирский государственный аэрокосмический университет имени академика М. Ф. Решетнева в 2006 г. Область научных интересов - интеллектуальные информационные технологии, эволюционные вычисления, поддержка принятия решений.

Список литературы

Дополнительные файлы