Adaptive flow control method

Abstract

The invention provides a self-adaptive flow control method. The self-adaptive flow control method comprises the following steps: step 1, establishing an RBF neural network model suitable for flow regulation and control; 2, calculating the output of the PID controller at the t moment, and transmitting a control signal output by the PID controller to a motor; 3, the motor changes the rotating speed by receiving the control signal, then the output flow of the medicament metering pump is changed, and the actual output flow reaches the input target flow; step 4, Jacob ian identification information of the RBF neural network is obtained through calculation; 5, the adjustment amount of PID parameters is obtained based on a gradient descent method; step 6, carrying out iterative optimization calculation on the output weight vector W, the network base width vector B and the node center vector; and 7, returning to the step 1 to carry out flow control at the next moment. According to the invention, on-line filling and mixing of different agents in injection-accompanying and augmented injection construction are realized, the method adapts to great changes of parameters such as pressure, and on-line self-adaptive regulation and control with a wide operation range, strong robustness and precision are realized.

Description

Technical Field

[0001] The invention relates to the technical field of oil and natural gas exploitation, and in particular to an adaptive flow control method. Background Art

[0002] During the injection and additional injection construction, it is necessary to measure and control the flow rate of the chemical agent medium. At present, the automatic injection of agents adopts a fixed frequency or flow rate, and the variable flow rate mainly relies on manual control, which has problems such as slow feedback response and large fluctuations. The problem of online injection of agents with different flow rates can be solved by implementing adaptive control, which can effectively cope with large changes in parameters such as pressure. Through the adaptive algorithm, only a small part of the parameters need to be adjusted to achieve the flow control of online injection of mixed agents, which can greatly reduce costs, improve efficiency, and has strong engineering application value.

[0003] In the Chinese patent application with application number: CN202010418794.9, a method for intelligent control of cooling water flow of an internal combustion engine based on an RBF neural network is involved. The method designs an RBF neural network model suitable for predicting the water temperature of an internal combustion engine, and uses the temperature data set of the internal combustion engine obtained by measurement to train and test the designed RBF neural network model; the predicted water temperature of the internal combustion engine is calculated by the RBF neural network model, and then the temperature difference between the predicted temperature and the target water temperature is calculated, and the cooling water flow is regulated according to the temperature difference, and the operating temperature of the internal combustion engine is temperature compensated, so that the operating water temperature of the internal combustion engine is in the target water temperature range in real time. This invention applies modern electronic information technology to the control of the cooling circulation system of the internal combustion engine, so as to ensure that the internal combustion engine always maintains or works close to the optimal temperature range under any climatic environment, thereby truly achieving the purpose of reducing the aggravation of parts wear, reducing exhaust emissions, improving the working mode of the internal combustion engine, and thus saving energy and reducing emissions.

[0004] In the Chinese patent application with application number: CN202111427658.7, a control method for quickly adjusting the valve of a flow controller based on pressure fluctuation is involved, which includes the following steps: S1: Start the flow controller, collect upstream and downstream pressures, valve control voltage and flow setting value; S2, establish a model with upstream and downstream pressure difference, flow setting value and valve control voltage as parameters; S3: Use the CMAC neural network model to obtain the predicted upstream and downstream pressure difference; S4: When the upstream pressure fluctuation exceeds the threshold, combine the upstream and downstream pressure difference predicted by S3 to generate a new upstream and downstream pressure difference, substitute the model of S2 to predict the valve control voltage after the upstream pressure fluctuation and give it to the valve opening control module, and correct the CMAC neural network model. The upstream pressure is detected by the pressure detection module in the flow controller, and the flow control is performed according to the pressure fluctuation to compensate for the influence of pressure fluctuation on the flow, effectively preventing the flow from being out of control.

[0005] In the Chinese patent application with application number: CN202111553051.3, a throttling back pressure multi-stage control method based on multivariable intelligent decoupling is involved, including selecting a single / multi-spool throttling valve mode of a throttling valve; calculating the pressure loss ΔPiL from the wellhead to the front end of the throttling valve, the pressure loss ΔPLo from the rear end of the throttling valve to the outlet, the required throttling pressure loss ΔPLs, the required controlled throttling back pressure PLs, and the required real-time throttling valve flow area SLc; if it is a single-spool throttling valve mode, the fuzzy control area and the fuzzy position in the fuzzy control area are determined according to the required real-time throttling valve flow area SLc; the valve core S is controlled to the fuzzy position; the neural network adaptive approximation method is used to regulate the valve core S from the fuzzy position to the precise position A1 point; if it is a multi-spool throttling valve mode, the fuzzy control area is determined according to the required real-time throttling valve flow area SLc; the fuzzy valve core DA is controlled to the fuzzy position of the fuzzy control area; the neural network adaptive approximation method is used to regulate the precise valve core DB from the fuzzy position to the precise position A1 point.

[0006] Due to the large differences in medium viscosity, density and other parameters of different injections during mixing, changes in these parameters of heterogeneous fluids can easily cause large changes in plunger force. Manual control is prone to problems such as slow feedback response and large fluctuations, and there are also safety hazards. At the same time, operating parameters are changeable, and existing flow control methods cannot adapt to large changes in parameters such as pressure. The parameter search of the flow control system converges slowly and is prone to local optimality and other problems. Therefore, it is necessary to design an adaptive auxiliary agent flow control method to achieve real-time tracking and precise control of changes in agent flow. To this end, we have invented a new adaptive flow control method. Summary of the invention

[0007] The purpose of the present invention is to provide a self-adaptive flow control method for online injection and mixing of different chemical agents in accompanying injection and augmentation construction.

[0008] The object of the present invention can be achieved by the following technical measures: an adaptive flow control method, the adaptive flow control method comprising:

[0009] Step 1: Establish an RBF neural network model suitable for traffic control;

[0010] Step 2: Use an incremental PID controller to input the target drug flow rate Q at time t in (t), the output of the PID controller at time t can be calculated, and the control signal of the output of the PID controller is transmitted to the motor;

[0011] Step 3: The motor can change the speed by receiving the control signal, thereby changing the output flow of the pharmaceutical metering pump so that the actual output flow can reach the input target flow;

[0012] Step 4: Calculate the RBF neural network Jacobian identification information;

[0013] Step 5: The adjustment amount of PID parameters can be obtained based on the gradient descent method;

[0014] Step 6: Based on the gradient descent method, iteratively optimize the output weight vector W, network base width vector B, and node center vector;

[0015] Step 7: Return to step 1 to perform flow control at the next moment.

[0016] The purpose of the present invention can also be achieved by the following technical measures:

[0017] The adaptive flow control method further includes, before step 1, modeling the motor-medicine metering pump system.

[0018] When modeling the motor-pharmaceutical metering pump system, the closed-loop system state space equation is:

[0019]

[0020]

[0021] Q out (t) = [1 0] q

[0022] Where: t is a certain moment of system operation, q is the system state value; Q out is the actual output flow of the pump, u is the input voltage of the motor; T L is the electromechanical time constant of the motor, T S is the electrical time constant of the motor; k u is the voltage transfer coefficient, K A is the metering pump gain coefficient.

[0023] In step 1, an incremental PID controller is used to input the target drug flow rate Q at time t. in (t), determine the data sampling time ts, and let the control error at time k be:

[0024] e(t)=Q in (t)-Q out (t)

[0025] Record the target flow rate Q at time t, time t-1, and time t-2 in (t) and output flow Q out The controller input vector is Xc, and the components of Xc are linear combinations of the deviations e(t), e(t-1), and e(t-2):

[0026]

[0027] In step 1, the controller output at time k is calculated based on the input signal:

[0028] u(t)=u(t-1)+Δu(t)

[0029] Δu(t)=k p x1+k i x2+k d x3

[0030] In the formula, u(t) is the control quantity at the current time t; u(t-1) is the control quantity at the previous time, that is, time t-1; Δu(t) is the adjustment amount of the control quantity; kp is the proportional coefficient, ki is the integral coefficient, and kd is the differential coefficient.

[0031] In step 2, the designed RBF neural network model is a three-layer feedforward neural network with only one hidden layer, including an input layer, a hidden layer and an output layer. The center and local receptive field of each neuron in the hidden layer determine the position and width of the RBF. The number of nodes in the input layer is designed to be three layers, and the input vector is taken as the controller input X=[ΔuQ out (t) Q out (t-1)] T , where Q out (t) is the output flow at time t, Q out (t-1) is the output flow at time t-1.

[0032] In step 2, the weight from the input layer to the hidden layer is set to 1, the number of hidden layer nodes is determined to be m, and the radial basis vector of the RBF neural network is set to H = [h1,h2,…,h j ,…,h m ] T , where h j is the Gaussian basis function:

[0033]

[0034] Determine the center vector C of the jth node in the network j =[c j1 ,c j2 ,c j3 ] T , let the network base width vector B = [b1, b2, ..., b m ] T , b j is the base width parameter of node j, b j >0; network output weight vector W = [w1,w2,…,w j ,…,w m] T .

[0035] In step 3, the RBF neural network Jacobian identification information is calculated as:

[0036] Q m =w1h1+w2h2+…+w m h m

[0037]

[0038] Information Q is identified by the k-time network Jacobian m (t) and the actual output flow rate Q of the pharmaceutical metering pump out (t) The performance index function of the identifier is calculated as follows:

[0039]

[0040] In step 4, the adjustment amount of PID parameters can be obtained based on the gradient descent method:

[0041]

[0042]

[0043]

[0044] The adjusted PID parameter values ​​are:

[0045] k p ′=k p +Δk p

[0046] k i ′=k i +Δk i

[0047] k d ′=k d +Δk d

[0048] The output of the PID controller after adjusting the parameters is the change in motor speed, which is Δn:

[0049] Δn=k p ′x1+k i ′x2+k d ′x3.

[0050] Where: Δkp, Δki, Δkd are the adjustment values ​​of the proportional coefficient kp, the integral coefficient ki, and the differential coefficient kd respectively; η p , η i , ηd are the learning rates of the proportional coefficient kp, integral coefficient ki, and differential coefficient kd respectively; k p ′, k i ′, k d ′ is the value of the adjusted proportional coefficient kp, integral coefficient ki, and differential coefficient kd.

[0051] In step 5, given the learning rate η and momentum factor α, the output weight vector W, network base width vector B, and node center vector are iteratively optimized based on the gradient descent method:

[0052] w j (t) = w j (t-1)+η(y(t)-y m (t))h j +α(w j (t-1)-w j (t-2))

[0053]

[0054] b j (t) = b j (t-1)+ηΔb j +α(b j (t-1)-b j (t-2))

[0055]

[0056] c ji (t) = c ji (t-1)+Δc ji +α[c ji (t-1)-c ji (t-2)]

[0057] Where: w j (t), b j (t), c ji (t) are the weight, base width and center vector of node j at time t, w j (t-1), b j (t-1), c ji (t-1) are the weight, base width and center vector of node j at time t-1, w j (t-2), b j (t-2), c ji (t-2) are the weight, base width and center vector of node j at time t-2, Δb j , Δc ji are the changes of the base width and center vector of node j respectively.

[0058] In step 3, the output flow rate of the medicine metering pump can be changed by changing the motor speed through the above process, so that the actual output flow rate can reach the input target flow rate. In steps 4 to 6, the control parameters are iteratively updated and calculated to achieve adaptive flow rate regulation.

[0059] The purpose of the present invention can also be achieved through the following technical measures: an adaptive flow control system, which adopts an adaptive flow control method to perform online injection and mixing of different reagents during the accompanying injection and augmentation construction.

[0060] The adaptive flow control method in the present invention includes a parameter search method for an adaptive flow control system of a mixing device and an online adaptive precision control algorithm for variable flow of a mixing device. The parameter search method enables the adaptive flow control algorithm to approach an accurate state model in a timely manner, thereby reducing the impact of complex changes in operating parameters on the flow control of the reagents. The control method adopts an adaptive algorithm, and the mixing device only needs to adjust a small part of the parameters to achieve flow control for online injection of different reagents, thereby improving efficiency and reducing costs. The control algorithm adopts a radial basis function (RBF) neural network adaptive algorithm combined with a PID control algorithm to improve the stability and accuracy of the flow control system. The adaptive flow control method of the present invention realizes online injection and mixing of different reagents during injection and augmentation construction, adapts to large changes in parameters such as pressure, and realizes online adaptive precision control with a wide operating range, strong robustness, and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A flowchart of an adaptive flow control method in a specific embodiment of the present invention;

[0062] Figure 2 It is a topological structure diagram of RBF neural network and PID control in a specific embodiment of the present invention;

[0063] Figure 3 This is a target flow rate variation diagram in a specific embodiment 2 of the present invention;

[0064] Figure 4 A control error diagram in a specific embodiment 2 of the present invention;

[0065] Figure 5 This is a control effect diagram in a specific embodiment 2 of the present invention;

[0066] Figure 6 This is a target flow rate variation diagram in a specific embodiment 3 of the present invention;

[0067] Figure 7 A control error diagram in a specific embodiment 3 of the present invention;

[0068] Figure 8 This is a control effect diagram in a specific embodiment 3 of the present invention. DETAILED DESCRIPTION

[0069] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0070] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations and / or combinations thereof.

[0071] Neural networks have been widely used in many fields due to their powerful data processing capabilities. They can not only solve nonlinear problems well, but also have the advantages of self-adaptation, self-learning and fault tolerance. They can effectively find data features from a large amount of data and become a powerful tool for finding data patterns. RBF neural network, or radial basis function neural network, is a novel feedforward neural network and a commonly used three-layer feedforward network. It can be used for both function approximation and pattern classification. Compared with other types of artificial neural networks, RBF network has a physiological basis, simple structure, fast learning speed, excellent approximation performance and generalization ability. It has the best approximation and global optimal performance. At the same time, the training method is fast and easy, and there is no local optimal problem. These advantages make RBF neural network have unique advantages in adaptive flow control.

[0072] An adaptive flow control method, such as Figure 1 As shown, the specific steps include:

[0073] Step 1: Model the motor-pharmaceutical metering pump system. The state space equation of the closed-loop system is:

[0074]

[0075]

[0076] Q out (t) = [1 0] q

[0077] Where: t is a certain moment of system operation, q is the system state value; Q out is the actual output flow of the pump, u is the input voltage of the motor; T Lis the electromechanical time constant of the motor, T S is the electrical time constant of the motor; k u is the voltage transfer coefficient, K A is the metering pump gain coefficient.

[0078] Step 2: Use an incremental PID controller to input the target drug flow rate Q at time t in (t), determine the data sampling time ts, and let the control error at time k be:

[0079] e(t)=Q in (t)-Q out (t)

[0080] Record the target flow rate Q at time t, time t-1, and time t-2 in (t) and output flow Q out The controller input vector is Xc, and the components of Xc are linear combinations of the deviations e(t), e(t-1), and e(t-2):

[0081]

[0082] Step 3: According to the input signal, the controller output at time k can be calculated as:

[0083] u(t)=u(t-1)+Δu(t)

[0084] Δu(t)=k p x1+k i x2+k d x3

[0085] In the formula, u(t) is the control quantity at the current time t; u(t-1) is the control quantity at the previous time, that is, time t-1; Δu(t) is the adjustment amount of the control quantity; kp is the proportional coefficient, ki is the integral coefficient, and kd is the differential coefficient.

[0086] Step 4: Design an RBF neural network model suitable for traffic control. Its topology is as follows: Figure 2 As shown in the figure, the RBF neural network model is a three-layer feedforward neural network with only one hidden layer, including an input layer, a hidden layer, and an output layer. The center and local receptive field of each neuron in the hidden layer determine the position and width of the RBF. The number of nodes in the input layer is designed to be three layers, and the input vector is taken as the controller input X = [Δu Q out (t) Q out (t-1)] T , where Q out (t) is the output flow at time t, Q out(t-1) is the output flow at time t-1.

[0087] Step 5: Take the weight from the input layer to the hidden layer as 1, determine the number of hidden layer nodes to be m, and set the radial basis vector of the RBF neural network H = [h1,h2,…,h j ,…,h m ] T , where h j is the Gaussian basis function:

[0088]

[0089] Determine the center vector C of the jth node in the network j =[c j1 ,c j2 ,c j3 ] T , let the network base width vector B = [b1, b2, ..., b m ] T , b j is the base width parameter of node j, b j > 0. Network output weight vector W = [w1,w2,…,w j ,…,w m ] T .

[0090] Step 6: The RBF neural network Jacobian identification information is calculated as:

[0091] Q m =w1h1+w2h2+…+w m h m

[0092]

[0093] Information Q is identified by the k-time network Jacobian m (t) and the actual output flow rate Q of the pharmaceutical metering pump out (t) The performance index function of the identifier is calculated as follows:

[0094]

[0095] Step 7: Based on the gradient descent method, the adjustment amount of PID parameters can be obtained:

[0096]

[0097]

[0098]

[0099] The adjusted PID parameter values ​​are:

[0100] k p ′=k p +Δk p

[0101] k i ′=k i +Δk i

[0102] k d ′=k d +Δk d

[0103] The output of the PID controller after adjusting the parameters is the change in motor speed Δn:

[0104] Δn=k p ′x1+k i ′x2+k d 'x3

[0105] Where: Δkp, Δki, Δkd are the adjustment values ​​of the proportional coefficient kp, the integral coefficient ki, and the differential coefficient kd respectively; η p , η i , η d are the learning rates of the proportional coefficient kp, integral coefficient ki, and differential coefficient kd respectively; k p ′, k i ′, k d ′ is the value of the adjusted proportional coefficient kp, integral coefficient ki, and differential coefficient kd.

[0106] Step 8: Given a learning rate η and momentum factor α, perform iterative optimization calculations on the output weight vector W, network base width vector B, and node center vector based on the gradient descent method:

[0107] w j (t) = w j (t-1)+η(y(t)-y m (t))h j +α(w j (t-1)-w j (t-2))

[0108]

[0109] b j (t) = b j (t-1)+ηΔb j +α(b j (t-1)-b j (t-2))

[0110]

[0111] c ji (t) = c ji (t-1)+Δc ji +α[c ji (t-1)-c ji (t-2)]

[0112] Where: w j (t), b j (t), c ji (t) are the weight, base width and center vector of node j at time t, w j (t-1), b j (t-1), c ji (t-1) are the weight, base width and center vector of node j at time t-1, w j (t-2), b j (t-2), c ji (t-2) are the weight, base width and center vector of node j at time t-2, Δb j , Δc ji are the changes of the base width and center vector of node j respectively.

[0113] Through the above process, the output flow rate of the medicine metering pump can be changed by changing the motor speed, so that the actual output flow rate can reach the input target flow rate.

[0114] The following are several specific embodiments of the present invention.

[0115] Example 1

[0116] In the specific embodiment 1 of the present invention, the adaptive flow control method includes the following steps:

[0117] Step 1: Determine the input layer, hidden layer and output layer structure of the RBF neural network, including:

[0118] (1) The first layer is the input layer, and the number of input layer node layers is 3;

[0119] (2) The second layer is the hidden layer. The number of hidden layer nodes is m, and the transformation function of the neuron is the Gaussian basis function: Among them, C j is the center vector of the jth hidden node, which has the same dimension as the input vector X, and b j >0 is the basis width parameter of the hidden layer node j; the node vector of the hidden layer is: H = [h1,h2,…,h j ,…,h m ] T ;

[0120] (3) The third layer is the output layer. The number of nodes in the output layer is 1, and the weight vector of the output layer is: W = [w1, w2, ..., w j ,…,w m ] T .

[0121] Step 2: The neural network performs a learning process, which includes:

[0122] (1) The input layer receives the input signal X = [Δu Q(t) Q(t-1)] T And passed to the hidden layer, the hidden layer realizes the nonlinear mapping from X to H;

[0123] (2) Hidden layer output vector H = [h1,h2,…,h j ,…,h m ] T To the output layer, the Jacobian identification information of the neural network is obtained through weighted operation: Q m =w1h1+w2h2+…+w m h m .

[0124] Step 3: Initialize the controller parameters, give the learning rate η, momentum factor α, sampling time t s and the initial value k of the PID controller p0 , k i0 , k d0 ;

[0125] Step 4: Expected input flow Q of the device at time t in (t) and the actual flow sensor output flow Q out (t) is sampled and the control error at that moment is calculated as e(t) = Q in (t)-Q out (t), and then obtain the input vector Xc=[x1,x2,x3] of the PID controller T ,in:

[0126]

[0127] Step 5: Calculate the control quantity at time t according to the incremental PID control algorithm:

[0128] u(t)=u(t-1)+Δu(t)

[0129] Δu(t)=k p x1+k i x2+k d x3

[0130] Step 6: Calculate the neural network tuning index

[0131] Step 7: Adjust the three parameters of the PID controller according to the gradient descent method:

[0132]

[0133]

[0134]

[0135] Step 8: Use the gradient descent method to derive the output weight vector W and node center vector C j And the node base width vector B is iteratively updated:

[0136] w j (t) = w j (t-1)+η(y(t)-y m (t))h j +α(w j (t-1)-w j (t-2))

[0137]

[0138] b j (t) = b j (t-1)+ηΔb j +α(b j (t-1)-b j (t-2))

[0139]

[0140] c ji (t) = c ji (t-1)+Δc ji +α[c ji (t-1)-c ji (t-2)]

[0141] Step 9: Perform control sampling at the next moment and return to step 1 to continue.

[0142] Example 2

[0143] In the specific embodiment 2 of the present invention, the control starts when the flow rate is 1500L / min, and the target flow rate is reduced to 1000L / min after 1s. The control algorithm is simulated and analyzed by using a step signal, such as Figure 3 shown.

[0144] Given the initial PID control parameter k p =0.22, ki =0.13, k d =0, learning rate η = 0.01, momentum factor α = 0.05.

[0145] The neural network input layer nodes are set to 3 layers, the hidden layer nodes are set to 6 layers, and the output layer nodes are set to 1 layer.

[0146] Execute the above algorithm process and compare the simulation results of RBF-PID control algorithm with those of traditional PID control. The simulation results are as follows: Figure 4 As shown in Figure 2. The maximum overshoot of RBF-PID control is significantly lower than that of traditional PID control, and the time to reach steady state is also shorter. Figure 5 The error analysis diagram of the two control effects is shown in Figure 2. Obviously, the error of the RBF-PID control method is much smaller than that of the PID control method.

[0147] Example 3

[0148] In the specific embodiment 2 of the present invention, the control starts when the flow rate is 200 L / min, and the target flow rate is adjusted to 900 L / min after 1 second. The control algorithm is simulated and analyzed by using a step signal, such as Figure 6 shown.

[0149] Given the initial PID control parameter k p =0.22, k i =0.13, k d =0, learning rate η = 0.01, momentum factor α = 0.05.

[0150] The neural network input layer nodes are set to 3 layers, the hidden layer nodes are set to 6 layers, and the output layer nodes are set to 1 layer.

[0151] Execute the above algorithm process and compare the simulation results of RBF-PID control algorithm with those of traditional PID control. The simulation results are as follows: Figure 7 As shown in Figure 2. The maximum overshoot of RBF-PID control is significantly lower than that of traditional PID control, and the time to reach steady state is also shorter. Figure 8 The error analysis diagram of the two control effects is shown in Figure 2. Obviously, the error of the RBF-PID control method is much smaller than that of the PID control method.

[0152] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0153] Except for the technical features described in the specification, all other technical features are known technologies to those skilled in the art.

Claims

1. An adaptive flow control method, characterized in that: The adaptive flow control method includes: Step 1: Establish an RBF neural network model suitable for traffic control; Step 2: Use an incremental PID controller to input the target drug flow rate Q at time t in (t), the output of the PID controller at time t can be calculated, and the control signal output by the PID controller is transmitted to the motor; Step 3: The motor changes the speed by receiving the control signal, thereby changing the output flow rate of the pharmaceutical metering pump so that the actual output flow rate can reach the input target flow rate; Step 4: Calculate the RBF neural network Jacobian identification information; Step 5: The adjustment amount of PID parameters can be obtained based on the gradient descent method; Step 6: Based on the gradient descent method, iteratively optimize the output weight vector W, network base width vector B, and node center vector; Step 7: Return to step 1 to perform flow control at the next moment.

2. The adaptive flow control method according to claim 1, characterized in that: The adaptive flow control method further includes, before step 1, modeling the motor-medicine metering pump system.

3. The adaptive flow control method according to claim 2, characterized in that: Before step 1, when modeling the motor-pharmaceutical metering pump system, the closed-loop system state space equation is: Q out (t)=[1 0]q Where: t is a certain moment of system operation, q is the system state value; Q out is the actual output flow of the pump, u is the input voltage of the motor; T L is the electromechanical time constant of the motor, T S is the electrical time constant of the motor; k u is the voltage transfer coefficient, K A is the metering pump gain coefficient.

4. The adaptive flow control method according to claim 1, characterized in that: In step 1, an incremental PID controller is used to input the target drug flow rate Q at time t. in (t), determine the data sampling time ts, and let the control error at time k be: e(t)=Q in (t)-Q out (t) Record the target flow rate Q at time t, time t-1, and time t-2 in (t) and output flow Q out The controller input vector is Xc, and the components of Xc are linear combinations of the deviations e(t), e(t-1), and e(t-2):

5. The adaptive flow control method according to claim 4, characterized in that: In step 1, the controller output at time k is calculated based on the input signal: u(t)=u(t-1)+Δu(t) Δu(t)=k p x1+k i x2+k d x3 In the formula, u(t) is the control quantity at the current time t; u(t-1) is the control quantity at the previous time, that is, time t-1; Δu(t) is the adjustment amount of the control quantity; kp is the proportional coefficient, ki is the integral coefficient, and kd is the differential coefficient.

6. The adaptive flow control method according to claim 1, characterized in that: In step 2, the designed RBF neural network model is a three-layer feedforward neural network with only one hidden layer, including an input layer, a hidden layer and an output layer. The center and local receptive field of each neuron in the hidden layer determine the position and width of the RBF. The number of nodes in the input layer is designed to be three layers, and the input vector is taken as the controller input X = [Δu(t) Q out (t) Q out (t-1)] T , where Q out (t) is the output flow at time t, Q out (t-1) is the output flow at time t-1.

7. The adaptive flow control method according to claim 6, characterized in that: In step 2, the weight from the input layer to the hidden layer is set to 1, the number of hidden layer nodes is determined to be m, and the radial basis vector of the RBF neural network is set to H = [h1,h2,…,h j ,…,h m ] T , where h j is the Gaussian basis function: Determine the center vector C of the jth node in the network j =[c j1 ,c j2 ,c j3 ] T , let the network base width vector B = [b1, b2, ..., b m ] T , b j is the base width parameter of node j, b j >0; network output weight vector W = [w1,w2,…,w j ,…,w m ] T .

8. The adaptive flow control method according to claim 7, characterized in that: In step 3, the RBF neural network Jacobian identification information is calculated as: Q m =w1h1+w2h2+…+w m h m Information Q is identified by the k-time network Jacobian m (t) and the actual output flow rate Q of the pharmaceutical metering pump out (t) The performance index function of the identifier is calculated as follows:

9. The adaptive flow control method according to claim 1, characterized in that: In step 4, the adjustment amount of PID parameters can be obtained based on the gradient descent method: The adjusted PID parameter values ​​are: k p ′=k p +Δk p k i ′=k i +Δk i k d ′=k d +Δk d The output of the PID controller after adjusting the parameters is the change in motor speed, which is Δn: Δn=k p ′x1+k i ′x2+k d ′x3 Where: Δkp, Δki, Δkd are the adjustment values ​​of the proportional coefficient kp, the integral coefficient ki, and the differential coefficient kd respectively; η p , η i , η d are the learning rates of the proportional coefficient kp, integral coefficient ki, and differential coefficient kd respectively; k p ′, k i ′, k d ′ is the value of the adjusted proportional coefficient kp, integral coefficient ki, and differential coefficient kd.

10. The adaptive flow control method according to claim 1, characterized in that: In step 5, given the learning rate η and momentum factor α, the output weight vector W, network base width vector B, and node center vector are iteratively optimized based on the gradient descent method: w j (t)=w j (t-1)+η(y(t)-y m (t))h j +α(w j (t-1)-w j (t-2)) b j (t)=b j (t-1)+ηΔb j +α(b j (t-1)-b j (t-2)) c ji (t)=c ji (t-1)+Δc ji +α[c ji (t-1)-c ji (t-2)] Where: w j (t), b j (t), c ji (t) are the weight, base width and center vector of node j at time t, w j (t-1), b j (t-1), c ji (t-1) are the weight, base width and center vector of node j at time t-1, w j (t-2), b j (t-2), c ji (t-2) are the weight, base width and center vector of node j at time t-2, Δb j , Δc ji are the changes of the base width and center vector of node j respectively.

11. The adaptive flow control method according to claim 1, characterized in that: In step 3, the output flow rate of the medicine metering pump can be changed by changing the motor speed through the above process, so that the actual output flow rate can reach the input target flow rate; in steps 4 to 6, the control parameters are iteratively updated and calculated to achieve adaptive adjustment of the flow rate.

12. An adaptive flow control system, characterized in that: The adaptive flow control system adopts the adaptive flow control method described in any one of claims 1 to 11 to perform online injection and mixing of different reagents during the accompanying injection and augmentation construction.

Images