RU2573766C1 - Artificial neural network - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: device comprises a layer of input nodes and at least two neuron layers connected in series in a forward direction, which form a forward channel for processing control signals, a neural subnetwork for processing adaptation signals, consisting of a layer of input nodes and at least one neuron layer, the output signals of which set weight coefficients of links of at least one neuron layer of the forward channel for processing control signals.

EFFECT: providing efficient adaptation of an artificial neural network to external environment parameters and to control object state parameters.

4 cl, 7 dwg

Description

The invention relates to intelligent control systems and can be used as the core of digital neurocontrollers, neuroprocessors and neuroregulators.

It is known that the core of neuroregulators for complex dynamic objects (DOs) are usually artificial neural networks (ANNs) of direct distribution [1]. Possessing the ability to generalize, the use of ANN allows effective control of BS in a wide range of changes in phase variables (state variables) of the latter. Known ANNs containing a set of artificial neurons and devices for transmitting information between them [2] (Application for invention No. 2006116775 RF, IPC G06N 3/04. Hardware-implemented artificial neural network. / G.A. Sorokin, A.E. Rogov. - Stated May 16, 2006; Published on November 27, 2007).

The closest in technical essence to the claimed device is a multilayer direct distribution ANN, consisting of a layer of input nodes, hidden layers of artificial neurons and an output layer of neurons, connected by connections called synaptic, sequentially in the forward direction and not containing connections between elements inside the layer and feedbacks between layers [3] (http://www.osp.ru/os/1997/04/179189/ - Introduction to artificial neural networks. / Zhianchang Mao, Enil J. // ″ Open Systems ″, No. 04, 1997. Fig. 4). The known device, having the ability to generalize, does not have adaptability to the parameters of the external environment of DO and the parameters of its state. Adaptation can be implemented by restructuring the weighting coefficients of synaptic connections according to the library of typical tuning algorithms, which requires additional equipment (components).

The problem to which the claimed invention is directed is to create an effective adaptation of the ANN to the parameters of the external environment in which the control object is located, and to its state parameters (values of phase variables of DO) without increasing the equipment (number of components) of the neurocontroller, neuroprocessor or neuroregulator.

The expected technical result is achieved through the implementation of a new direct distribution ANN architecture, containing a layer of input nodes and at least two layers of neurons connected in series in the forward direction, forming a direct channel for processing control signals, by introducing adaptation signal processing neural subnet into its structure, which forms a channel of parameters consisting of a layer of input nodes for adaptation signals and at least one layer of neurons whose output signals establish weight coupling coefficients of at least one neuron layer of the direct channel for processing control signals, i.e. the weighting coefficients of the synaptic connections of at least one neuron layer of the direct control signal processing channel are equal to the monotonic functions of the signals of the output neuron layer of the parameter channel.

This ensures the effective adaptability of the ANN with an unchanged basic set of weight coefficients obtained during the training of the ANN, and therefore without increasing the number of components (equipment), since there is no library (database) of typical algorithms for setting weight coefficients, as is realized in the known adaptive neurocontrollers [4].

An additional technical result is the elimination of the static error of DO control due to the execution of artificial neurons of the last hidden layer of the direct channel linear, and the introduction of trainee proportionally integrating links into them.

The implementation of the activation functions of all artificial neurons, except for the linear neurons of the last hidden layer of the direct channel, identical and passing through the zero point, provides unification of the ANN elements, which allows to simplify the design of the device. At the same time, the source of the trained bias is introduced in the input layer of neurons of the parameter channel, which allows shifting the reference point of activation functions, providing an acceleration of the ANN learning process.

Another additional technical result is the acceleration of the learning process of an adaptive ANN due to the clustering of adaptation signals by physical nature, for example, the separation of ANS adaptation signals to the parameters of the external environment in which the control object is located and to its state parameters (values of phase variables of DO), and introduction for each cluster of the corresponding subnet, which forms its own channel of parameters, each of which consists of a layer of input nodes and at least one layer of neurons, the output whose signals establish the weighting coefficients of the connections of different layers of neurons of the direct channel for processing control signals.

In FIG. 1 shows the architecture of the proposed adaptive ANN and the notation is introduced: 1 - layer of input nodes of the direct channel; 2 - layers of neurons of the direct channel for processing control signals; 3 - layer of input nodes of the parameter channel for adaptation signals; 4 - layer of neurons of the parameter channel for processing adaptation signals; 5 is a vector of weight coefficients of the synaptic connections of the neurons of the direct channel established by 4 adaptation signals processed in the neuron layer.

; 12 - датчики параметров внешней обстановки ДО 6, формирующие вектор сигналов $$\overline{EC}$$

; 13 - блок мультиплексирования (обозначение как принято в пакете Simulink MATLAB и используется на фиг. 2-6).In FIG. 2 shows the structure of the adaptive control system for complex DO, in which the neuroregulator is included in the feedback circuit; designations introduced: 6 - DO; 7 - neuroregulator; 8 - ANN; 9 is a block for calculating the derivatives of signals with respect to time; 10 is a memory unit that stores a set of weighting factors (MVK) for ANN 8; 11 - vector of adaptation signals $$\overline{SV}$$

; 12 - sensors of environmental parameters up to 6, forming a vector of signals $$\overline{EC}$$

; 13 - multiplexing unit (designation as adopted in the Simulink MATLAB package and used in Figs. 2-6).

; 12 - датчиков параметров внешней обстановки ДО 6, формирующий вектор сигналов $$\overline{EC}$$

; 13 - блок мультиплексирования; 14 - база данных (БД), содержащая библиотеку типовых алгоритмов настройки весовых коэффициентов ИНС 8.In FIG. 3 shows the structure of the adaptive control system for complex DO, in which the neuroregulator is included in the feedback circuit; designations introduced: 6 - DO; 7 - neuroregulator; 8 - ANN; 9 is a block for calculating the derivatives of signals with respect to time; 11 - vector of adaptation signals $$\overline{SV}$$

; 12 - sensors of environmental parameters DO 6, forming a vector of signals $$\overline{EC}$$

; 13 - block multiplexing; 14 is a database (DB) containing a library of typical algorithms for setting the weight coefficients of ANN 8.

In FIG. 4 presents a fragment of the implemented structure of the neuroregulator with the main fragment of the ANN structure for automatic control of the aircraft in the Simulink MATLAB package; designations are introduced: 1 - layer of input nodes of the direct channel; 2 - layers of neurons of the direct channel for processing control signals; 3 - layer of input nodes of the parameter channel for adaptation signals; 4 - layer of neurons of the parameter channel for processing adaptation signals; 10 - a memory unit that stores MVK; 15 - source of the trained bias (source with a single signal) for the input layer of neurons 4 channel parameters.

In FIG. 5 shows a fragment of the structure of the neuroregulator implemented in the Simulink MATLAB package, and the main fragment of the ANN structure; designations introduced: 2 - a hidden layer of neurons of the direct channel for processing control signals; 16 - linear neurons of the last hidden layer of the direct channel; 17 - blocks for calculating the signal integral over time; 18 - input source of constant bias; 19 - output layer of neurons; 20 - blocks that implement the activation functions of the output layer of neurons 19.

; 12 - датчики параметров внешней обстановки ДО 6, формирующие вектор сигналов $$\overline{EC}$$

.In FIG. 6 shows the structure of the adaptive control system for complex DO, in which a neuroregulator containing an embodiment of the inventive device is included in the feedback circuit; designations are used: 6 - DO; 7 - neuroregulator; 8 - ANN; 9 is a block for calculating the derivatives of signals with respect to time; 10 is a memory unit that stores a set of weighting factors (MVK) for ANN 8; 11 - vector of adaptation signals $$\overline{SV}$$

; 12 - sensors of environmental parameters up to 6, forming a vector of signals $$\overline{EC}$$

.

In FIG. 7 shows a variant of the architecture of the proposed adaptive ANN and the notation is introduced: 1 - layer of input nodes of the direct channel; 2 - layers of neurons of the direct channel for processing control signals; 3 - layers of input nodes of parameter channels for adaptation signals; 4 - layers of parameter channel neurons for processing adaptation signals; 5 - vectors of weight coefficients of synaptic connections of neurons of the direct channel established by 4 adaptation signals processed in layers of neurons.

. Функции от выходных сигналов слоя нейронов 4, значения которых присваивают вектору весовых коэффициентов, зависят от конкретной физической реализации ИНС. Так, в цифровых обучаемых нейроконтроллерах это просто линейная функция с единичным коэффициентом, а, например, в аналоговых нейропроцессорах (необучаемых) синаптические связи могут быть реализованы полевыми транзисторами, включенными как управляемые резисторы, на затворы которых поступают сигналы установки весовых коэффициентов соответствующих синаптических связей ИНС, при этом функции нелинейные, но монотонные и знать их математическое представление не требуется.The inventive device - ANN direct distribution - contains (Fig. 1) a layer of input nodes 1 and at least two layers of neurons 2 connected in series in the forward direction and not containing connections between elements inside the layer and feedbacks between layers. These layers of neurons 2 form a direct channel for processing control signals. The introduced adaptation signal processing neural subnet forms a parameter channel consisting of a layer of input nodes 3 for adaptation signals and a layer of neurons 4, the output signals of which establish link weights 5 of the neurons of the direct control signal processing channel, i.e. assign values of arbitrary monotonic functions from the output signals of the neuron layer to 4 vector weights $${\overline{W}}_P$$

. The functions of the output signals of the neuron layer 4, the values of which are assigned to the vector of weights, depend on the specific physical implementation of the ANN. So, in digital trained neurocontrollers, it’s just a linear function with a unit coefficient, and, for example, in analog neuroprocessors (untrained), synaptic connections can be implemented by field-effect transistors included as controlled resistors, to the gates of which the signals for setting the weight coefficients of the corresponding synaptic connections of the ANN are received, in this case, the functions are nonlinear, but monotonic, and it is not required to know their mathematical representation.

. Этот вектор составлен как из самих сигналов ошибки управления, так и из их временных производных, реализуемых блоком 9. Множество весовых коэффициентов (МВК) $$\overline{W}$$

, значения которых предварительно получены в процессе обучения ИНС 8, сохраняются в блоке памяти 10. Совокупность сигналов адаптации составлена из вектора сигналов $$\overline{SV}$$

11, получаемых из датчиков параметров состояния ДО 6, и из вектора сигналов $$\overline{EC}$$

, получаемых из датчиков 12 параметров внешней обстановки, в которой находится объект управления. Эти векторные сигналы $$\overline{SV}$$

и $$\overline{EC}$$

объединяют в один общий вектор сигналов адаптации (параметров) $$\overline{P}$$

с помощью блока мультиплексирования 13 (обозначение как принято в пакете Simulink MATLAB). Вектор сигналов адаптации $$\overline{P}$$

поступает на слой входных узлов канала параметров ИНС 8. Таким образом, осуществляется эффективное адаптивное управление сложным ДО 6.The operation of the inventive device is shown in FIG. 2 in the adaptive control system of complex DO 6, in which the neuroregulator 7 is included in the feedback circuit. Input ANN 8, the proposed architecture and which is the core of the neuroregulator 7, receives a vector of control error signals $$\overline{x}$$

. This vector is composed both of the control error signals themselves and of their temporary derivatives, implemented by block 9. The set of weight coefficients (MVK) $$\overline{W}$$

, the values of which were previously obtained during the training of ANN 8, are stored in the memory unit 10. The set of adaptation signals is composed of a vector of signals $$\overline{SV}$$

11 obtained from sensors of state parameters DO 6, and from the signal vector $$\overline{EC}$$

obtained from the sensors 12 of the parameters of the external environment in which the control object is located. These vector signals $$\overline{SV}$$

and $$\overline{EC}$$

combine into one common vector of adaptation signals (parameters) $$\overline{P}$$

using multiplexing unit 13 (designation as adopted in the Simulink MATLAB package). Adaptation Signals Vector $$\overline{P}$$

enters the layer of input nodes of the channel of the ANN 8 parameters channel. Thus, effective adaptive control of complex DO 6 is carried out.

, поступающего на БД 14. Заявляемое устройство использует единожды полученное МВК $$\overline{W}$$

, хранящееся в блоке памяти 10 (фиг. 2), что требует меньшего оборудования (количества компонентов) нейроконтроллера 7, чем это необходимо при реализации БД 14 в нейроконтроллере 7 (фиг. 3), содержащей целую библиотеку типовых алгоритмов настройки весовых коэффициентов ИНС 8.The expected technical result follows from a comparison of the claimed invention with a well-known similar adaptive control system for complex DO 6, in which the neuroregulator 7 is included in the feedback circuit shown in FIG. 3. The essence of the known adaptive control method [4] is to create a library of typical algorithms for setting the weight coefficients of ANN 8, stored in a database (DB) 14 created during the training of ANN 8. The selection of specific algorithms for setting weight coefficients is carried out according to the value of the signal vector adaptations $$\overline{P}$$

arriving at the database 14. The inventive device uses once received MVK $$\overline{W}$$

stored in the memory block 10 (Fig. 2), which requires less equipment (number of components) of the neurocontroller 7 than is necessary when implementing the database 14 in the neurocontroller 7 (Fig. 3), containing a whole library of typical algorithms for setting the weight coefficients of ANN 8.

A fragment of the implemented structure of the neuroregulator with the main fragment of the ANN structure, the proposed architecture, for automatic control of the aircraft in the Simulink MATLAB package is shown in FIG. 4. The layers of neurons 2 of the direct channel and the layer of neurons of 4 channels of parameters are implemented by ″ dotprod ″ blocks, which scalarly multiply the vector of parameters by the vector of weight coefficients stored in the memory block 10. The activation functions of all neurons are implemented by ″ tansig ″ blocks that perform hyperbolic tangent transformation:

where F is the value of the activation function (output signal of the neuron);

x is the result of adding the input signals of the neuron multiplied by the corresponding weights.

The activation functions of all neurons are monotonically increasing and passing through the zero point, i.e. if the input signal is zero, the output signal is zero. The necessary shift of the activation functions of the input layer of neurons 4 channel parameters relative to the zero point is provided by the introduction of the source of the trained bias 15.

. Источники постоянного смещения 18 могут быть либо обучаемыми, либо с жестко выставляемыми значениям их весов.In FIG. 5 shows the output fragment of the implemented structure of the ANN of the proposed architecture. In order to eliminate the static control error DO, the neurons of the last hidden layer 16 of the direct channel are linear and include trained proportionally integrating links based on blocks 17 for calculating the signal integrals over time. Sources of constant bias 18, introduced in parallel with proportional-integrating links, the output layer of neurons 19 of the direct channel and blocks 20 that implement their activation functions, provide the formation of the necessary region of variation of the values of the vector of output control signals $$\overline{y}$$

. Sources of constant bias 18 can be either trained, or with hard-set values of their weights.

11, получаемых из датчиков параметров состояния ДО 6, и вектора сигналов $$\overline{EC}$$

, получаемых из датчиков 12 параметров внешней обстановки. Архитектура варианта предложенной адаптивной ИНС 8 (фиг. 6) с двумя входами векторных сигналов адаптации $${\overline{P}}_1$$

и $${\overline{P}}_2$$

и с двумя каналами параметров представлена на фиг. 7. При таком варианте исполнения сигналы выходных слоев нейронов 4 разных каналов параметров устанавливают весовые коэффициенты разных слоев связей 5 прямого канала обработки сигналов управления. Предложенное разделение сигналов адаптации (кластеризация) обуславливает ускорения процесса обучения адаптивной ИНС.Figure 6 shows the structure of the adaptive control system for complex TO 6 when using several groups of adaptation signals, in particular a signal vector $$\overline{SV}$$

11 obtained from sensors of state parameters DO 6, and the signal vector $$\overline{EC}$$

obtained from sensors 12 environmental parameters. The architecture of the variant of the proposed adaptive ANN 8 (Fig. 6) with two inputs of vector adaptation signals $${\overline{P}}_{\mathrm{one}}$$

and $${\overline{P}}_2$$

and with two parameter channels is shown in FIG. 7. With this embodiment, the signals of the output layers of neurons of 4 different parameter channels set the weights of different layers of links 5 of the direct channel for processing control signals. The proposed separation of adaptation signals (clustering) determines the acceleration of the adaptive ANN learning process.

Information sources

1. Burakov M.V. Multi-agent neural regulators. // Neuroinformatics - 2004. Part 2. Pages 130-138.

2. Application for invention No. 2006116775 of the Russian Federation, IPC G06N 3/04. Hardware-implemented artificial neural network / G.A. Sorokin, A.E. Horns. - Declared May 16, 2006; Publ. 11/27/2007.

3. http://www.osp.ru/os/1997/04/179189/ - Introduction to artificial neural networks. / Zhianchang Mao, Enil J. // ″ Open Systems ″, No. 04, 1997. Fig. 4 - (prototype).

4. Tyukin I.Yu., Terekhov V.A. Neural network solution to the problem of adaptive control of a class of nonlinear dynamic objects with non-convex parameterization // Neurocomputers: development and application. 2004. No. 7-8. S.4-16.

Claims (4)

1. An artificial neural network of direct propagation, containing a layer of input nodes and at least two layers of neurons connected in series in the forward direction, forming a direct channel for processing control signals, characterized in that a neural subnet for processing adaptation signals is introduced, which forms a channel of parameters consisting of layer of input nodes for adaptation signals and from at least one neuron layer, the output signals of which establish link weights of at least one neuron layer ryamogo control channel signal processing. 2. The artificial neural network according to claim 1, characterized in that the neurons of the last hidden layer of the direct channel are linear and trained proportionally integrating links are introduced into them. 3. The artificial neural network according to claim 1, characterized in that the activation functions of all neurons, except the linear neurons of the last hidden layer of the direct channel, are implemented identically and passing through the zero point, and a source of trained bias is introduced in the input layer of neurons of the parameter channel. 4. The artificial neural network according to claim 1, characterized in that at least a second neural subnet for processing adaptation signals of the second group is additionally introduced, which forms a second channel of parameters, also consisting of a layer of input nodes and at least one layer of neurons, the weekend whose signals establish the weighting coefficients of the connections of another layer of neurons of the direct channel for processing control signals.

Images