RU193607U1 - A device for controlling an electric drive of an object under the influence of external disturbances based on neural networks - Google Patents

Abstract

A device for controlling an electric drive of an object under the influence of external disturbances, based on neural networks, relates to the field of automatic control systems with feedback of electric drives of objects characterized by changes in parameters, non-linearity of characteristics, as well as exposure to random undefined external disturbances presented, for example, in the form of white or color noise. The utility model is aimed at increasing the stability and quality of operation of the electric drive control system of an object under the influence of external disturbances. The technical result, which consists in increasing the accuracy and dynamics of the control system, is achieved due to the fact that the control device containing the serially connected master device, the first summing unit, the optimizer unit, the drive converter unit, the output of which is designed to connect to the input of the control object located under the influence of external perturbations, as well as a block of the quality functional, the output of which is connected to the second input of the optimizer block, the block of predictive models, the output to of which is connected to the second input of the first summing block, the second summing block, the output of which is connected to the input of the weight coefficient block, and the state neuroidentifier block, the second input of which is designed to connect to the output of the drive converter block, and the third input is connected to the output of the weight coefficient block, the output the state neuroidentifier unit is connected to the second input of the second summing unit, an external disturbance neuroidentifier unit is additionally introduced, the first input of which is connected To the output of the drive converter unit, the second input is designed to connect to the output of the control object and is connected to the first input of the second summing unit, the output of the external disturbance neuroidentifier unit is connected to the first input of the state neuroidentifier unit, and a linearization block is introduced, the output of which is connected to the input of the unit predictive models, the first input of the linearization block is connected to the output of the state neuroidentifier block, and the second input is connected to the output of the external neuroidentifier block disturbances.

Description

The utility model relates to the field of automatic control systems with feedback by electric drive of objects characterized by changes in parameters, non-linearity of characteristics, as well as exposure to random undefined external disturbances presented, for example, in the form of white or color noise.

Currently, a known system for automatic control of objects with a self-tuning PID controller (patent No. 2419122, IPC: G05B 13/02 (2006.01), published on 05/20/2011), containing sequentially connected PID controller connected to the input of the control object and to the first input of the identifier, the output of the control object is connected to the second input of the identifier, the output of which is connected to the input of the synthesizer, as well as an I-regulator, a control unit and a self-tuning unit for the amplitudes of the test signal, the test signal generator is configured so that General periods of test frequencies were multiples of the period of discreteness.

The disadvantage of this control system is the limited range of stability of the system with large changes in external disturbances associated with the lack of optimal control.

A control device with adjustable parameters based on neural networks is also known (patent No. 111914, IPC: G05B 13/00, publ. 12/27/2011). This device contains an optimal control unit, to the input of which there is a reference signal and the output signal of a neuroclassifier, an intelligent controller whose output is connected to the input of the control object, the input of the intelligent controller is connected to the output of the optimal control unit, to the output of the control object and to the output of the time delay unit, the input of the time delay block is connected to the output of the control object, the input of the neuroclassifier is connected to the output of the control object and to the output of the time delay block , introduced an intelligent regulator based on a three-layer perceptron and a neuroclassifier based on the Kohonen neural network.

A disadvantage of the known device is the slow optimization rate and the tendency to premature convergence due to falling into the region of a local minimum during training of Kohonen neural networks with an error back propagation algorithm.

The closest in technical essence to the claimed utility model is a device for controlling an electric drive of nonlinear objects based on neural networks (patent No. 186950, IPC: G05D 3/12, G05B 13/00, G06N 3/02 publ. 02/11/2019). This device contains a weighting unit, as well as a serially connected master, a first summing unit, an optimizer unit and a drive converter unit, the output of which is connected to the control object, and a quality functional unit, the output of which is connected to the second input of the optimizer unit, a block of predictive models the output of which is connected to the second input of the first summing block, the second summing block, the output of which is connected to the input of the weighting block, and the block is neuroen a state identifier, the first input of which is intended to be connected to the output of the control object and to the first input of the second summing unit, the second input is connected to the output of the drive converter unit, and the third input is connected to the output of the weighting unit, the output of the neuroidentifier unit is connected to the second input of the second summing unit and to the input of the block of predictive models.

A disadvantage of the known device is the decrease in the accuracy of the dynamic characteristics of the control system in the presence of external disturbances associated with the lack of their assessment.

The task to which the claimed utility model is directed is to increase the stability and quality of the electric drive control system of an object under the influence of external disturbances.

The technical result consists in increasing the accuracy and dynamics of the control system by restoring the estimate of the values of external disturbances acting on the control object and introducing a feedback signal for evaluating the current external disturbances into the automatic control law to eliminate their negative impact on the quality of the control system.

The specified technical result is achieved due to the fact that the control device containing a serially connected master device, a first summing unit, an optimizer unit, a drive converter unit, the output of which is designed to connect to the input of a control object that is under the influence of external disturbances, as well as a functional unit quality, the output of which is connected to the second input of the optimizer block, the block of predictive models, the output of which is connected to the second input of the first block a second summing unit, the output of which is connected to the input of the weight coefficient block, and a state neuroidentifier block, the second input of which is designed to connect to the output of the drive converter block, and the third input is connected to the output of the weight coefficient block, the output of the state neuro identifier block is connected to the second input of the second summing unit, an additional block of the neural identifier of external disturbances is introduced, the first input of which is connected to the output of the drive converter unit, the second in The od is designed to connect to the output of the control object and is connected to the first input of the second summing unit, the output of the external disturbance neuroidentifier unit is connected to the first input of the state neuroidentifier unit, and a linearization unit is introduced, the output of which is connected to the input of the predictive model unit, the first input of the linearization unit is connected with the output of the state neuroidentifier block, and the second input is connected to the output of the neuroidentifier block of external disturbances.

The essence of the technical solution is illustrated by the drawing (Fig.), Which shows the structural diagram of the device for controlling the electric drive of the object. The drawing shows:

1 - master device (memory);

2 - the first block summation (PBS);

3 - optimizer block (BO);

4 - block functional quality (BFC);

5 - block predictive models (BPM);

6 - block drive converter (BPP);

7 - control object (OS);

8 - the second block summation (VBS);

9 - state neuroidentifier unit (BNIS);

10 - block weighting factors (BVK);

11 - block neuroidentifier external disturbances (BNIVV);

12 - block linearization (BL).

The device for controlling the electric drive of the object, (see Fig.), Contains a serially connected master device 1 (memory), the first summing unit 2 (PBS), the optimizer block 3 (BO), the drive converter unit 6 (BPP), the output of which is intended for connection to the control object 7 (OS), the block of functional quality 4 (BFC), the output of which is connected to the second input of the optimizer block 3 (BO), the block of predictive models 5 (BPM), the output of which is connected to the second input of the first summation block 2 (PBS) , linearization block 12 (BL), the output of which is connected nen with the input of the block of predictive models 5 (BPM), the block of the neuroidentifier of external disturbances 11 (BNIVV), the first input of which is connected to the output of the block of the drive converter 6 (BPP), the second input is designed to connect to the output of the control object 7 (OS) and to the first the input of the second summing unit 8 (VBS), the output of the external disturbance neuro identifier block 11 (BNIVV) is designed to connect to the first input of the state neuroidentifier block 9 (BNIS) and to the second input of the linearization block 12 (BL), the second summing block 8 (VBS), you One of which is connected to the input of the block of weight coefficients 10 (BVK) and the block of the state neuroidentifier 9 (BNIS), the second input of which is connected to the output of the block of the drive converter 6 (BPP), the third input is connected to the output of the block of weight coefficients 10 (BVK), the block state neuroidentifier 9 (BNIS) is connected to the second input of the second summing unit 8 (VBS) and to the first input of the linearization block 12 (BL).

In this case, the optimizer block 3 (BO), the quality functional block 4 (BFK), the linearization block 12 (BL) and the predictive model block 5 (BPM) perform the function of a digital adaptive predictive controller. The control object 7 (OS), is under the influence of external disturbances presented, for example, in the form of white or color noise. To restore the estimate of the values of external disturbances, a block of the neuroidentifier of external disturbances AND (BNIVV) is used. The neural network observer of the state of the control object consists of a block of state neuroidentifier 9 (BNIS) and a block of weighting factors 10 (BVK).

Elements of the device can be implemented in software and hardware based on an industrial base. As a master device 1 (memory), a digital generator or industrial computer can be used, the control object 7 (memory) can be technological equipment, the operation of which is described in time by a system of differential and algebraic equations.

As a block of drive converter 6 (BPP), transistor inverters based on IGBT transistors (frequency converter) from OMRON (Japan) can be used.

To implement the summation blocks 2 (PBS), 8 (VBS), the optimizer block 3 (BO), the quality functional block 4 (BFC), the predictive model block 5 (BPM), the state neuroidentifier block 9 (BNIS), and the weight coefficient block 10 ( BVK), external disturbance neuro-identifier block 11 (BNIVV) and linearization block 12 (BL), computational and program modules of the FASTWEL microcontroller (Russia) can be used.

The device controls the electric drive of the object as follows. The control signal is generated in the master 1 (memory), which is connected to the first summing unit 2 (PBS) to subtract the predictive negative feedback signal. The predictive signal is generated by the block of predictive models 5 (BPM), in the block of which predictive models are synthesized by the output signals of the control object for a certain number of steps forward during the prediction horizon. These predictive models are constructed as expressions of future values of control actions and multiply connected linearized models with parameters depending on the state of the control object. To calculate the linearized model coefficients in the linearization block 12 (BL), an external disturbance estimation signal from the output of the external disturbance neuroidentifier block 11 (BNIVV) and a control object state variable estimation signal from the output of the state neuroidentifier block 9 (BNIS) are used. The synthesis method for linearized models with state-dependent parameters is determined based on the theory of extended linearization of nonlinear objects [Afanasyev V.N. Control of nonlinear indefinite dynamic objects // 1st ed. M .: Lenard, 2015. - 224 p.].

The predictive mismatch signal at the output of the first summing block 2 (PBS) is sent to the optimizer 3 (BO) block, in the block of which the optimization of program control is performed according to the quasi-Newton criterion expressed in the block of quality functional 4 (BFC). The optimization criterion in the block of quality functional 4 (BFK) is synthesized in such a way as to minimize the standard error of the regulation between the driving control signal and the output signal of the block of predictive models 5 (BPM) with a given accuracy. The result is the optimal control signal found, connected through the drive converter unit 6 (BPP) to the control object 7 (OS).

External disturbances acting on the control object 7 (OS) are determined by the nature of the uncertain non-linearities. The mathematical model of external perturbations is successfully described by the theory of random processes with a zero mean, which can be determined by the Gaussian probability distribution law. The signal from the output of the control object 7 (OS) is fed to the first input of the second summing unit 8 (VBS) and the second input of the external disturbance neuroidentifier unit 11 (BNIVV) to estimate the values of external disturbances.

An estimate of the values of current external disturbances is generated in the block of the neuroidentifier of external disturbances 11 (BNIVV) using well-known identification algorithms using neural networks. The construction of a neuroidentifier for assessing current external disturbances is associated with a number of the following tasks: choosing the structure of a neural network, choosing the number of layers, choosing the number of neurons in layers, choosing an activation function, choosing a learning algorithm, and choosing a training set. The structure of the neural network in the block of the external disturbance neuroidentifier 11 (BNIVV) is a partial recurrent neural network, which is called the Elman network [M. Burakov Neural networks and neurocontrollers: textbook. manual // SPb .: GUAP, 2013. - 284 p.]. The Elman neural network has demonstrated that it is optimally suited for identifying random indefinite non-linearities. The Elman neural network belongs to the class of a multilayer perceptron with local delay feedbacks closed from hidden neurons to context neurons. In this case, hidden neurons record their previous states, which allows you to save the output information of hidden neurons for one or more clock cycles, and then transfers them to the input layer. Consider the Elman neural network, consisting of m inputs, n neurons of the hidden layer, covered by feedbacks through delay elements and p neurons of the output layer. The vector of the input layer of the Elman network x (k) at iteration k is determined by the expression

x (k) = [u ₀ (k), u ₁ (k), ..., u _m (k), ν ₁ (k-1), ..., ν _m (k-1)],

where u ₀ (k), u ₁ (k), ..., u _m (k) is the vector of the input network signal; ν ₁ (k-1), ..., ν _m (k-1) is the state of the neurons of the hidden layer at the previous iteration k-1.

The state of the neurons of the hidden layer is described as follows

- синаптические весовые коэффициенты нейронов скрытого слоя;

- функции активации нейронов скрытого слоя.Where

- synaptic weights of neurons of the hidden layer;

- activation functions of neurons of the hidden layer.

The state of the neurons of the output layer is determined similarly.

- синаптические весовые коэффициенты нейронов выходного слоя;

- функции активации нейронов выходного слоя.where y ₀ (k), y ₁ (k), ..., y _p (k) is the network output signal vector;

- synaptic weights of the neurons of the output layer;

- activation functions of neurons of the output layer.

The sigmoidal activation functions in the form of the hyperbolic tangent tansig are used as activation functions of the two-layer Elman network in the hidden layer, and the purelin linear functions are used in the output layer. The Elman recurrent neural network learning method is the Levenberg-Marquardt backpropagation algorithm. Note that during the training of the Elman network, the Levenberg-Marquardt optimization method has a high ability and quickly converge to the optimal error value. As a result of the training process, we obtain a signal for evaluating the current external disturbances at the output of the external disturbance neuro identifier block 11 (BNIVV), which is connected to the first input of the linearization block 12 (BL) for synthesizing the control law and to the first input of the state 9 neuroidentifier (BNIS) to restore the state vector management object.

The full state vector of the control object for the synthesis of linearized models is restored by a state observer using well-known identification algorithms using neural networks. The structure of state 9 neuroidentifier unit (BNIS) is a recurrent neural network based on the “Nonlinear Auto Regressive with Exogenous inputs” model, which is called the NARX recurrent neural network or NARX model [Khaikin S. Neural networks: full course // Per. from English M .: William, 2006. S. 919-989]. The architectural structure of the NARX model can take many different forms, but, the simplest form, using an architectural neural network with feedback. This model has a single input, which is applied to memory on delay lines consisting of d _u elements. It has a single output, closed to an input through memory on delay lines, which also contain d _y elements. The dynamics of the NARX model is described by the following non-linear input-output equation:

y (k) = σ {u (k), u (k-1), ..., u (kd _u ), y (k-1), y (k-1), ..., y (kd _y) },

where u (k) is the input of the NARX model; y (k) is the output of the NARX model; σ (⋅) activation functions. Moreover, the structure of a recurrent neural network NARX contains two layers. The first layer is a hidden layer with sigmoidal activation functions in the form of a tansig hyperbolic tangent, the second layer is an output layer with purelin linear activation functions. The training method of the recurrent neural network NARX is an algorithm for the back propagation of errors in time. This algorithm is an extension of the standard gradient descent backpropagation algorithm and allows you to conduct the training process much faster and get good convergence and a better generalization. The mismatch signal of the state vector of the control object at the output of the second summing unit 8 (VBS) is connected to the input of the block of weighting factors 10 (BVK). The block of weighting coefficients 10 (BVK) contains a matrix of weighting coefficients of the activation functions of neural networks NARX for each class from the set of changes in the state parameters of the control object 7 (OS). This matrix of weighting coefficients is assigned to each class in a one-to-one correspondence. When the structure of the control object 7 (OS) is transferred from one class to another, the matrix of weight coefficients is replaced by the matrix of weight coefficients corresponding to the new class. As a result, the recoverable signal of the complete state vector of the control object is determined at the output of the state neuroidentifier unit 9 (BNIS) connected to the input of the linearization block (BL) and to the second input of the second summing unit 8 (VBS). The process of operation of the control device continues until the master control action is completely worked out.

As a result, we obtain improved stability and ensuring the quality of operation of the electric drive control system of an object under the influence of external disturbances.

Claims (1)

A device for controlling an electric drive of an object under the influence of external disturbances, based on neural networks, containing a serially connected master device, a first summing unit, an optimizer block, a drive converter unit, the output of which is intended to be connected to the input of a control object under the influence of external disturbances, and also a quality functional block, the output of which is connected to the second input of the optimizer block, a block of predictive models, the output of which is connected about the second input of the first summing block, the second summing block, the output of which is connected to the input of the weight coefficient block, and the state neuroidentifier block, the second input of which is connected to the output of the drive converter block, and the third input is connected to the output of the weight coefficient block, the output of the state neuro identifier block is connected to the second input of the second summing unit, characterized in that the device additionally includes a neural identifier block of external disturbances, the first input of which is connected the output of the drive converter unit, the second input is used to connect to the output of the control object and is connected to the first input of the second summing unit, the output of the external disturbance neuroidentifier unit is connected to the first input of the state neuroidentifier unit, and a linearization block is introduced, the output of which is connected to the input of the block of predictive models , the first input of the linearization block is connected to the output of the state neuroidentifier block, and the second input is connected to the output of the external neuroidentifier block scheny.

Images