RU2775510C1 - Method for operation of an adaptive system for controlling the process of guidance of complex objects and apparatus for implementation thereof - Google Patents

Abstract

FIELD: computing technology.

SUBSTANCE: technical result is achieved by means of an apparatus for implementing a method for operation of an adaptive system for controlling the process of guidance of complex objects, containing series-connected blocks of the reference model and a third controller, wherein the first input of the block of the reference model and the third input of the third controller are connected to the effect from the signal of the first controller, and the first output of the third controller is connected to the control object, the output whereof is connected to the second input of the block of the first stage of current identification and bandpass filters, to the third input of the block of the second stage of current identification, and to the second input of the third controller, wherein the composition of the reference model includes an artificial neural network (ANN), the input whereof has an attribute of a particular goal, and the output of the ANN has a class whereto the target is attributed, wherein the output of the block of the reference model is linked with the output of the control object and is connected to the first input of the third controller via an adder, and the second output of the third controller is connected to the third input of the second controller.

EFFECT: reduction in the volume of control information and increase in the stability of the control system when recognising images of ground-based targets.

1 cl, 21 dwg

Description

The invention relates to the field of technical means for controlling the process of targeting complex objects, such as guided aircraft weapons and unmanned aerial vehicles in autonomous flight mode (hereinafter referred to as the control object (OC)).

The most popular branch of control theory is the theory of adaptive control. Achievements in the application of vision systems, namely artificial neural networks (ANNs), make it possible to synthesize an adaptive control system for complex objects.

An adaptive control system with a two-stage identifier and an implicit reference model is known, designed to generate such a control law (CL) so that the CO behaves like a reference model (EM) [see, patent RU 2258951, C1, IPC G05B 13/04, publ. 20.08.2005].

The disadvantages of this adaptive system are that there are no:

- EM block;

- block of the third regulator, in which the adaptation error is minimized;

- INS for pattern recognition of ground targets.

The technical result of the invention is to reduce the amount of control information and increase the stability of the control system when recognizing images of ground targets.

The specified technical result is achieved by the fact that in the known adaptive missile control system with a two-stage identifier and an implicit reference model shown in Fig. 1, consisting of an adder 1, the second 2 and the first 3 regulators, a low-pass filter 4, a regulator tuning unit 10; the block of the first stage of the current identification and band-pass filters 8, the block of the second stage of the current identification 9, as well as the block that contains the adder 1, the first and second inputs of which are connected to the setting action, and the output to the first input of the first controller 3, the output of which is connected to the input of the low-pass filter 4, the output of which is connected to the input of the block of the reference model 5, the control object 7 and the third regulator 6, to the first input of the block 8 of the first stage of the current identification and band-pass filters and to the first input of the block 9 of the second stage of the current identification. The fourth output of the block of the control object 7 is connected to the second input of the block of the first stage of the current identification and band-pass filters 8 and to the third input of the block of the second stage of the current identification 9, the second output of which is connected to the second input of the regulator setting block 10, the first output of which is connected to the fifth input of the first controller 3, and the second output to the fourth input of the second controller 2, the output of which is connected to the second input of the adder 1, according to the invention, serially connected blocks of the reference model 5 and the third controller 6 are additionally introduced. The first input of the reference model 5 and the third input of the third controller 6 are connected to effect from the signal of the first regulator 3, and the first output of the third regulator 6 is connected to the control object 7, the output of which is connected to the second input of the block of the first stage of the current identification and bandpass filters 8, to the third input of the block of the second stage of the current identification 9 and to the second input of the third regulator 6. Etal composition On-line model 5 includes an ANN, the input of which has a sign of a particular goal, and the output of the ANN is the class to which the goal belongs. The output of the reference model is connected to the output of the control object 7 and is connected through the adder to the first input of the third controller 6. The second output of the third controller 6 is connected to the third input of the second controller 2.

The essence of the invention lies in the fact that the blocks of the reference model 5, the third controller 6 and the artificial neural network included in the reference model 5 are introduced. The first input of the reference model 5 and the third input of the third controller the output of the third regulator 6 is connected to the control object 7, the output of which is connected to the second input of the block of the first stage of the current identification and band-pass filters 8, to the third input of the block of the second stage of the current identification 9 and to the second input of the third regulator 6. The reference model 5 includes ANN , the input of which has a sign of a particular goal, and the output of the ANN is the class to which the goal belongs. The output of the reference model 5 is connected to the output of the control object 7 and is connected through the adder to the first input of the third controller 6. The second output of the third controller 6 is connected to the third input of the second controller 2.

Consider the construction of such a control system. Let the CO on the current time interval t∈[t ₀ , ∞] be described by the following matrix differential equation [1]:

- непосредственно измеряется и эффективности динамики или аналитически вычисляется по x(t).where x∈R ⁿ is the directly measured CO state vector; u∈R ⁿ - control vector (control law); ƒ - vector of uncontrolled external perturbations, limited in norm; A, B - matrices of unknown parameters of the CO (matrices of the object's own control) with the corresponding dimensions, in the general case, variables in time and state;

- directly measured and the efficiency of the dynamics or analytically calculated from x(t).

An adaptive system must form such a control law; so that the DT behaves like a reference model (EM), which is given in the form of the following differential equation:

where x _m is the state vector of the model; u _m - norm-limited input action of the model (dimensions correspond to equation (1)); A _m , B _m - matrix parameters of the model in the general case, variable in time, and the operator A _m - asymptotically stable.

The quality of the adaptive control system will be judged by the vector value e=x-x _m , which we will call the adaptation error. Accurate tracking of the op amp for the EM can be ensured only when the condition of full compliance of the models [2] is met:

rankB=rank(B,А _m -А)=rank(В, В _m )=rank (B, ƒ), or, which is identical,

where B is a pseudo-inverse matrix to B [1, 3]. When condition (3) is satisfied, the control, which we call exact:

will provide asymptotic properties of the adaptation error, which follows from the joint consideration of equalities (1) and (2) [1]. According to the condition of the matrix A, B and external disturbances ƒ are not measured, therefore, instead of (4), the real control law will be in the form:

- оценки матриц, доставляемые блоком текущей идентификации.where

- matrix estimates delivered by the current identification block.

As a current identification algorithm, a recurrent algorithm of the stochastic approximation type is used, which is described as follows [4]:

- оценка матрицы С, С=[А, В] - блочная матрица неизвестных параметров; i - номер дискретного момента времени с шагом Δt;

- невязка, называемая в дальнейшем ошибкой идентификации;

- расширенный вектор состояния ОУ (совокупность в общем случае переменных, участвующих в идентификации); Г_i - переменная положительно определенная квадратная матрица размерностью (n+m) или положительный скаляр.where

- evaluation of the matrix C, C=[A, B] - block matrix of unknown parameters; i is the number of a discrete moment of time with a step Δt;

- discrepancy, hereinafter referred to as the identification error;

- extended state vector of the CO (generally a set of variables involved in the identification); Г _i is a variable positive-definite square matrix of dimension (n+m) or a positive scalar.

- это справедливо для подавляющего большинства практических задач. Тогда можно показать, что при достаточно большой норме матрицы Г с течением времени ε→0, причем без каких-либо дополнительных условий. Напротив, оценки параметров для сходимости к истинным значениям требует соблюдения целого ряда условий, в том числе отсутствия неконтролируемых внешних возмущений [4].Assume that the following norms of vectors and matrices are bounded:

- this is true for the vast majority of practical problems. Then it can be shown that for a sufficiently large norm of the matrix Г over time ε→0, and without any additional conditions. On the contrary, estimation of parameters for convergence to true values requires compliance with a number of conditions, including the absence of uncontrolled external perturbations [4].

In works [1, 5] it is shown that under the following conditions:

or

the equation for the adaptation error of a closed-loop dynamic system using memory (5) has the form:

where K is the correlation coefficient.

The left side of the differential equation (9) is stable, therefore, when ε>0, the goal of adaptation is achieved: |ε|→0.

To meet the conditions (7) or (8), it is proposed to introduce a block diagram of the control system with a block of a priori information about the matrix of the management efficiency of the object, which gives information about the matrix B ₀ to the block of current identification, with a dimension of n × m .. This matrix takes into account a priori information about management efficiency matrix in the form of equality [1]:

чтобы скорректированная оценка

удовлетворяла условиям (7), (8).In the current identification block, matrix B ₀ is used to correct the matrix

so that the adjusted estimate

satisfies conditions (7), (8).

However, it is assumed that there is a priori information about the management performance matrix, which is not always the case. For example, at the stage of missile control on a highly maneuverable target at high angles of attack, when the control efficiency changes significantly, moreover, according to a complex ambiguous dependence. The objective of this invention is to eliminate this disadvantage. To solve it, it is proposed to use signal filtering, and the identification process is divided into two stages:

Для формирования этого этапа целесообразно в структурной схеме системы выделить отдельным блоком процедуру оценивания матрицы эффективности управления. В состав системы также включить блок полосовых частотных фильтров.1. Determining the estimates of the management performance matrix

To form this stage, it is advisable to single out the procedure for evaluating the management efficiency matrix in the structural diagram of the system as a separate block. The system also includes a block of band-pass filters.

Этот этап целесообразно организовать в блоке текущей идентификации прототипа, который в дальнейшем будет называться блоком второго этапа текущей идентификации.2. Obtaining an estimate of the matrix of own dynamics of the CO

It is advisable to organize this stage in the block of the current identification of the prototype, which will be referred to as the block of the second stage of the current identification.

We justify the need for the proposed changes. Obviously, a block of a priori information about the matrix of the management efficiency of the object is not needed if the specified matrix is determined with sufficient accuracy:

и условия (7), (8) выполняются автоматически. При этом уравнение динамики ошибки адаптации будет иметь еще более простой, чем выражение (9), вид [5]:In the limit

and conditions (7), (8) are satisfied automatically. In this case, the equation for the dynamics of the adaptation error will have an even simpler form than expression (9) [5]:

способствует следующее:Improved Accuracy

contributes to the following:

1. Unlike changes in the state vector components, information about which often contains significant dynamic, fluctuation and constant errors, the measurement of the control vector components is provided with high accuracy. For example, the measurement of the angular position of the control surface of a rocket using a position sensor with a discrete output is carried out with an error of no more than a minute of arc [6].

2. For a real technical control system, the frequency range of control signals, as a rule, is significantly wider than the frequency range of the OS dynamics. The latter is, most often, a low-frequency link.

3. The accuracy of the estimated parameters depends on their number. The more parameters the mathematical model of the CO (control object) contains, the more difficult it is to achieve high accuracy of their estimation and vice versa [5]. If the OU (1) evaluates only the elements of the matrix B, then the number of estimates is significantly reduced and thereby facilitates the conditions for their precise determination.

To clarify what has been said, consider one of the lines of system (1):

(термин регрессионного анализа) является реакцией на все сигналы в правой части уравнения (10). Для сокращения числа оцениваемых параметров необходимо выделить из сигнала отклика составляющую, являющуюся реакцией только на сигналы управления.where index k indicates the line number of equation (1). Response

(regression analysis term) is the response to all signals on the right side of equation (10). To reduce the number of estimated parameters, it is necessary to isolate from the response signal a component that is a response only to control signals.

To do this, we use the fact of differences in the operating frequency spectra of the OS (control object) and control signals. It is proposed to pass all the variables of equation (10) through a band-pass filter, with the frequency response (AFC) shown in Fig. 2, where are indicated: Α ₀ - frequency response of the control object; And _y is the spectrum of control signals; A _pf is the frequency response of the bandpass filter. When ensuring the narrowness of the frequency response of filter 1, the latter passes signals practically only with a frequency ω _pf , which we will call the filter selection frequency.

- передаточные функции колебательных звеньев с разными относительными коэффициентами затухания:The simplest implementation of such a bandpass filter can be provided according to the scheme shown in figure 3a, where

- transfer functions of oscillatory links with different relative attenuation coefficients:

where p is the differentiation operator; ω _pf - natural frequency of the band-pass filter; ξ ₁ , ξ ₂ - relative attenuation coefficients, such that ξ ₁ →0, ξ ₂ →0, ξ ₁ <ξ ₂ . The index "pf" denotes the corresponding output signals of the bandpass filter.

изображены на фиг.3б.Frequency response of oscillatory links (Α _f1 , Α _f2 ) and the resulting bandpass filter (A _pf ) relative to u and

shown in Fig.3b.

Equation (10), written through the output signals of the bandpass filter, will have the form:

where g _k - interference, due to incomplete suppression of the signals of the object's own dynamics on the bandpass filter and external disturbances.

As you can see, the number of estimated parameters in equation (12) is much less than in model (10).

The custom model for the current identification will accordingly look like [5]:

The algorithm of the first stage of identification in accordance with expressions (6), (12) and (13) has the form:

- положительно определенная матрица.where

is a positive definite matrix.

A natural practical requirement for the possibility of using the considered adaptive control approach is the requirement for the smallness of the noise norm g at the selection frequency. For an op amp with scalar control, the identification algorithm is greatly simplified, since in this case, instead of (12), it will be:

требуется, чтобы при u^пф≠0 выполнялось g_k→0, то есть, чтобы отношение |u_пф/g_k| достаточно большим.Hence it follows that in order to improve the accuracy

it is required that for u ^pf ≠0 g _k →0 be satisfied, that is, that the ratio |u _pf /g _k | large enough.

The latter can be used in the system in the following way: it is not always calculated according to the proposed approach, but only in cases where the norm of the vector and exceeds the experimentally specified threshold value.

доставленной первым этапом, и предназначен для определения оценки

Она может определяться в соответствии с алгоритмом (6) по зависимости:The second stage of identification is based on the assessment

delivered by the first stage, and is intended to determine the assessment

It can be determined in accordance with the algorithm (6) according to the dependence:

- положительно определенная матрица.where

is a positive definite matrix.

Since x is not measured by the condition, we propose to build the second stage of identification according to the indicated dependencies, but on filtered signals:

- выходной векторный фильтра реального - сигнал дифференцирующего звена с передаточной функцией (р/(τ_ρ+1)), когда на его вход подается х_i;

- выходные векторные сигналы фильтра- апериодического звена с передаточной функцией (1/(τ_р+1)), когда на его вход подаются u_i·и x_i; τ - постоянная времени фильтров, выбирается достаточно малой, чтобы выполнялись соотношения:where

- the output vector filter of the real - the signal of the differentiating link with the transfer function (р/(τ _ρ +1)), when x _i is applied to its input;

- output vector signals of the filter - aperiodic link with the transfer function (1/(τ _p +1)), when u _i ·and x _i are fed to its input; τ is the time constant of the filters, is chosen small enough to satisfy the following relations:

where a and b are arbitrary elements of the matrices A, B.

When the last relations are satisfied, it is easy to show that equation (1) corresponds to the following notation:

indicating the possibility of building identification on the specified filtered signals.

Since the identification process takes place in two stages, it is necessary to divide the current identification block into two blocks: the block of the first stage, where the evaluation of the management efficiency matrix of the object will be carried out; block of the second stage for evaluation of the matrix of own dynamics of the OS.

The block of the first stage of identification includes a logical device that starts the identification algorithm of this stage only when the norm of the vector u ^pf exceeds a predetermined threshold value. Also, a block of bandpass filters based on links (11) is introduced into the system, designed for high-frequency bandpass selection of control signals and the derivative of the OS state vector.

Since with such an implementation of the identification process, the conditions for estimating the control efficiency matrix with high accuracy are met, the block of a priori information about the control efficiency matrix of the object becomes unnecessary and is excluded from the adaptive system. No correction procedure is required either.

An adaptive control system was built using Simulink (Matlab) with a reference model for a continuous plant, the operation algorithm of which is described by the expression (control law):

The block diagram of the adaptive system is shown in Fig. 4. Structural diagram of the adaptive missile guidance system contains the adder 1, the second 2 and the first 3 regulators (see Fig. 5); low-pass filter 4, its parameters (see Fig. 6); control object 5, third controller 6, reference model 7 (see Fig. 10); the first stage of the current identification and bandpass filters 8 (see Fig. 9); the second stage of the current identification 9 (see Fig. 8); controller tuning unit 10 (see Fig. 7).

The reference model block contains an artificial neural network for pattern recognition of ground targets (jeep, tank). The output of the ANN is connected to the input of the reference model (see Fig. 11). The INS consists of two layers (see Fig. 12), where the first and second layers are shown in Figs. 13 and 14.

The first layer consists in connecting the input and hidden layers (see Fig. 15), and the second layer - hidden and output layers (see Fig. 16).

со второго выхода блока 2 на четвертый вход блока 1 второго регулятора. С первого выхода блока 2 поступает псевдообратная матрица (В⁺) на пятый вход блока 1 первого регулятора, в результате этого блок 1 формирует закон управления (17).The adaptive system works as follows. The master action in the form of a signal (B _M u _M ) _i is applied to the first and second inputs of the adder of block 1. The third input of block 1 of the second regulator receives signals x _i from the third output of block 5 and the signal

from the second output of block 2 to the fourth input of block 1 of the second regulator. From the first output of block 2, a pseudo-inverse matrix (B ⁺ ) arrives at the fifth input of block 1 of the first controller, as a result of which block 1 forms the control law (17).

связан со входом блока 5, а также с первым входом блока 3 второго этапа текущей идентификации и с первым входом блока 4 первого текущей идентификации и полосовых фильтров.The output of block 1 is connected to the input of a low-pass filter configured to cut high frequencies. Filter output

connected with the input of block 5, as well as with the first input of block 3 of the second stage of the current identification and with the first input of block 4 of the first current identification and bandpass filters.

поступает в блок 5 эталонной модели и объекта управления. С четвертого выхода блока 5 снимается информация об измеренном значении вектора состояния

поступающая на третий вход блока 3 второго этапа текущей идентификации и на второй вход блока 4 первого этапа текущей идентификации и полосовых фильтров.Control signal

enters block 5 of the reference model and the control object. Information about the measured value of the state vector is taken from the fourth output of block 5

coming to the third input of block 3 of the second stage of the current identification and to the second input of block 4 of the first stage of the current identification and bandpass filters.

и производной

составляющие в узком диапазоне частот около частоты выделения каждого фильтра (ω_пф), формируя сигналы

Частота ω_пф выбирается заранее по исходной априорной информации для удовлетворения соотношений АЧХ собственной динамики объекта и спектра управляющих сигналов (см. фиг. 2). В этом блоке вычисляется оценка матрицы эффективности управления В.Bandpass filters have a structure (see Fig. 8) and are based on links with transfer functions (11). They extract from control signals

and derivative

components in a narrow frequency range around the selection frequency of each filter (ω _pf ), forming signals

The frequency ω _pf is selected in advance according to the initial a priori information to satisfy the ratios of the frequency response of the object's own dynamics and the spectrum of control signals (see Fig. 2). In this block, the estimate of the control efficiency matrix B is calculated.

From the output of block 4 of the first stage of the current identification and bandpass filters, the signal B is fed to the second input of block 3 of the second stage of the current identification. In this block, the estimate of matrix A is calculated. The first and second output of block 3 is connected to the first and second output of block 2 for adjusting the regulators, in which the pseudo-inverse matrix B ⁺ is calculated and

на четвертый вход блока 1 второго регулятора.From the first output of block 2, the pseudoinverse matrix B ⁺ enters the fifth input of block 1 of the first controller and

to the fourth input of block 1 of the second regulator.

извлекает с помощью алгоритма SURF для пока неизвестного объекта и поступает на вход искусственной нейронной сети, до момента пока объект не будет распознан [7]. Используя выходной вектор

объект можно классифицировать по классу, к которому он принадлежит (jeep или tank) (см. фиг. 17) и этот сигнал подается на вход эталонной модели. С выхода эталонной модели сигнал (x_м) отправляется на сумматор, вычитающий из этого сигнала - сигнал (x_i) вектора состояния объекта управления, формируя сигнал адаптивной ошибки (ε) (см. фиг. 9), который подается на первый вход третьего регулятора (см. фиг. 18). В этом блоке формируется сигнал u_ОУ объекта управления [8]:Block of artificial neural network (ANN). Feature vector

extracts using the SURF algorithm for a yet unknown object and enters the input of an artificial neural network until the object is recognized [7]. Using the output vector

the object can be classified according to the class to which it belongs (jeep or tank) (see Fig. 17) and this signal is applied to the input of the reference model. From the output of the reference model, the signal (x _m ) is sent to the adder, which subtracts from this signal the signal (x _i ) of the state vector of the control object, forming an adaptive error signal (ε) (see Fig. 9), which is fed to the first input of the third controller (see Fig. 18). In this block, the signal u _OC of the control object is formed [8]:

where λ is a positive constant that determines the desired dynamic properties of the closed system; ξ - adjustable controller parameter generated by the adaptation algorithm from the condition of minimizing the tracking error ε; γ is the coefficient that determines the speed of setting the parameter ξ and is called the adaptation coefficient.

Based on the analysis (see Fig. 19, 20), we conclude that with the value of the adaptation coefficient (γ=25), the trajectory of the control object in the minimum short time (t=5 s.) coincides with the line of the reference model. This indicates the positive dynamics of the OS in the conditions of a complex background-target situation in the area of the observed ground target.

Exploring the static principle, which consists in the rejection of a priori information about the dynamics of change in the CO during the guidance process, leads, for example, to the fact that existing algorithms contain a single hypothesis about the nature of the target’s movement, which combines all possible target maneuvers. Although such a model of maneuver is, in the general case, of a statistical nature, it does not take into account the main tactical principle of countermeasures - the suddenness and speed of a change in the situation.

The implementation of this principle in target recognition algorithms leads to the fact that the Neyman-Pearson criterion turns out to be the rule for choosing a solution. According to this criterion, a priori information about the probabilities of hypotheses is not used at all. This significantly reduces the possibility of recognition in the face of opposition.

In an adaptive control system, a priori information about the dynamic properties of the filtered process serves to smooth measurements distorted by noise. If the useful signal is suppressed by interference, then this a priori information is the basis for obtaining predictable estimates. If the information interruption time is within acceptable limits, then the adaptive system remains operational.

Recognition is in a sense analogous to evaluation. In the event of a sharp sudden change in guidance conditions due to enemy opposition, a priori information about dynamic characteristics can help increase the recognition probability.

Using the a priori predictable matrix A = A _M -1, we eliminate the disadvantage - a low level of adaptation of guidance algorithms (adaptation error is minimized) under conditions of a priori uncertainty to the characteristics of the dynamics of the behavior of the OS and to slowly changing parameters of enemy counteraction (see Fig. 21) .

The proposed method and device for its implementation is an effective tool for managing the guidance process of a complex system.

The proposed technical solution is industrially applicable, since elements used in the field of optoelectronic systems and electrical engineering can be used for its implementation.

Sources of information:

1. Methods of classical and modern theory of automatic control: Textbook in 5 tg.; 2nd ed., revised. and additional T .5: Methods of modern theory of automatic control / Ed. K.A. Pupkova, N.D. Yegupova. M.: Publishing house of MSTU im. N.E. Bauman, 2004. - 784 p.

2. Utkin V.N. Sliding modes in control optimization problems. M.: Nauka, 1981. - 364 p.

3. Gantmakher F.R. Matrix theory. M.: Nauka, 1988. - 576 p.

4. Tsypkin Ya.Z. Fundamentals of Information Theory of Identification. M.: Nauka, 1984. - 336 p.

5. Ljung L. Identification of systems. Theory for the user: Translation from English. / Ed. Ya.Z. Tsypkin. M: Nauka, 1991. - 432 p.

6. Bukov V.N. Flight and navigation systems. Part 2. Pilot systems. M.: VVIA, 1986. - 383 p.

7. F.Kh. Al Saftli, S.T. Balanyan. Investigation of an algorithm for detecting singular points in an image and its use in missile guidance systems. Aerospace forces, theory and practice. ISSN 2500-4352. Peer-reviewed electronic periodical #15, September 2020, pp. 193-208.

8. A.Yu. Oshchepkov. Automatic control systems: theory, application, modeling in MATLAB. Tutorial. St. Petersburg - Moscow-Krasnodar. 2018. - 208 p.

9. RF patent for the invention No. 2258951, C1, IPC G05B 13/04, published on August 20, 2005.

Claims (1)

A device for implementing a method for functioning of an adaptive control system for the process of pointing complex objects, containing blocks of a reference model (5) and a third controller (6) connected in series, where the first input of the reference model block (5) and the third input of the third controller (6) are connected to the impact from signal of the first controller (3), and the first output of the third controller (6) is connected to the control object (7), the output of which is connected to the second input of the block of the first stage of the current identification and bandpass filters (8), to the third input of the block of the second stage of the current identification ( 9) and to the second input of the third controller (6), where the reference model (5) includes an artificial neural network (ANN), the input of which has a sign of a particular target, and the output of the ANN has the class to which the target belongs, while the output of the reference model block (5) is connected to the output of the control object (7) and is connected through the adder to the first input of the third regulator (6), and the second output of the third regulator regulator (6) is connected to the third input of the second regulator (2).

Images