Neural network modeling and control method of a double flexible nozzle with electric servo mechanism

Abstract

The present invention discloses a neural network modeling and control method for a dual-flexible nozzle of an electric servo mechanism, relating to the field of simulation technology. The method comprises: a neural network identifier employing a BP neural network algorithm identifies model parameters of a multi-loop system based on the swing angle to obtain a dual-flexible nozzle neural network model; a Gray Wolf optimized PID controller uses the deviation between the multi-loop system's command swing angle and the dual-flexible nozzle neural network model as an error signal based on parameter changes of the dual-flexible nozzle neural network model and the disturbance experienced by the electric servo mechanism, thereby determining the neural network system learning algorithm. The output of the Gray Wolf optimized PID controller is the control voltage of the electric servo mechanism system. The present invention can establish an accurate load model for the dual-flexible nozzle under different conditions and disturbances, addressing the problems of time-varying parameters, severe interference from nonlinear factors, and structural defects of conventional models.

Description

Technical Field

[0001] The present invention relates to the technical field of ground simulation of thrust vector control systems, and more particularly to a neural network modeling and control method for a double-flexible nozzle of an electric servo mechanism. Background Art

[0002] The dual-flexible nozzle thrust vectoring control system primarily consists of a dual-flexible joint, a diffuser, a mounting flange, and a matching servo control system. The dual-flexible joint is constructed from a laminated polymer rubber material and a metal support. The servo system drives the movable body, achieving system posture changes through the dual-flexible joint. The dual-flexible joint's material and structural composition make the dual-flexible nozzle a typical nonlinear hysteresis system.

[0003] Typical physical models for describing hysteresis phenomena include the constant stiffness and constant damping model, the linear viscous damping model, and the mixed damping model. Among them: the constant stiffness and constant damping model ignores the influence of conditions such as swing amplitude, frequency, and working pressure on the torque, and cannot meet the needs of accurate modeling of the dual-flexible nozzle thrust vector control system; the linear viscous damping model adapts to the swing torque of the flexible nozzle, but its generalization ability under different working conditions is still insufficient; the mixed damping model has good generalization ability, but there is a problem of divergence of the damping component coefficient under high frequency conditions.

[0004] Due to the viscoelastic properties of the rubber components in the dual-flexible joint, hysteresis occurs between the swing angle and the restoring torque during swing. Furthermore, the hysteresis curve is affected by factors such as swing amplitude, swing frequency, operating pressure, and environmental factors, making its dynamic characteristics very complex. Therefore, establishing an accurate load model for the dual-flexible nozzle under different conditions and disturbances, and employing a control method suitable for this system, is key to improving the response accuracy and control quality of the dual-flexible nozzle and is a technical problem that those skilled in the art urgently need to solve. Summary of the Invention

[0005] In view of this, the present invention provides a neural network modeling and control method for a dual-flexible nozzle of an electric servo mechanism, which can establish an accurate load model of the dual-flexible nozzle under different conditions and different disturbance conditions, and comprehensively improve the response accuracy and control quality of the dual-flexible nozzle.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A neural network modeling and control method for a double-flexible nozzle of an electric servo mechanism includes the following steps:

[0008] The neural network identifier uses the BP neural network algorithm to identify the model parameters of the multi-loop system according to the swing angle and obtains the neural network model of the dual flexible nozzle;

[0009] The dual-flexible nozzle neural network model is used as the controlled object of the multi-loop system. The Grey Wolf Optimized PID controller (GWO-PID) uses the deviation between the multi-loop system command swing angle and the dual-flexible nozzle neural network model as the error signal according to the parameter changes of the dual-flexible nozzle neural network model and the disturbance of the electric servo mechanism to determine the learning algorithm of the neural network system. The output of the Grey Wolf Optimized PID controller is the control voltage of the electric servo mechanism system.

[0010] The technical effects achieved by the above technical scheme are: the dual flexible nozzle can be identified and a dual flexible nozzle neural network model can be obtained, which effectively solves the problems of time-varying parameters, serious interference from nonlinear factors and structural defects of conventional models; combining the functions of the neural network identifier and the Gray Wolf optimized PID controller, it is applied to the control of the dual flexible nozzle system, comprehensively improving the control effect and control quality of the system, and is suitable for ground simulation experiments of dual flexible nozzles under experimental conditions.

[0011] Optionally, the multi-loop system includes an angular position controller, a speed controller, a current controller, a motor, a speed reduction mechanism, a dual flexible nozzle, a tension and pressure sensor, a current sensor, a speed sensor, a posture sensor, and a data acquisition card; wherein:

[0012] The angular position controller, the speed controller, the current controller, the motor, the speed reduction mechanism, and the double flexible nozzles are connected in sequence;

[0013] The deceleration mechanism, the tension and pressure sensor, and the data acquisition card are connected in sequence;

[0014] The motor, current sensor, data acquisition card and current controller are connected in sequence;

[0015] The motor, speed sensor, data acquisition card and speed controller are connected in sequence;

[0016] The double flexible nozzles, attitude sensor, data acquisition card and angular position controller are connected in sequence.

[0017] The technical effects achieved by the above technical solution are: the control structure of the multi-loop system is disclosed, the controller has the functions of current control, speed control, and posture control, the acquisition range of each sensor meets the requirements, and the returned data are all current or voltage signals, which can be cascaded with the controller and driver to form a three-closed-loop system.

[0018] Optionally, when performing model parameter identification of the multi-loop system, based on the multi-loop system test excitation signal input, the nozzle swing amplitude, swing frequency, swing speed, restoring torque and system working pressure output by the sensor are used as input variables, and the swing angle is used as the output variable;

[0019] Methods for model parameter identification include online identification and offline identification. In a further solution, an offline identification method based on experimental data is adopted.

[0020] Optionally, obtaining a dual flexible nozzle neural network model specifically includes the following steps:

[0021] The forward dynamics model of the nonlinear system is approximated by a neural network, and the neural network model of the dual flexible nozzle is discretized to obtain a discretized model; the discretized model is:

[0022] y N (k+1)=f[y(k),...,y(k-n+1),u(k),u(k-1),...,u(k-m+1)];

[0023] Where: f is an unknown nonlinear function that describes the system characteristics; y(k) is the system output; u(k) is the system input, y N (k+1) is the output of the neural network, that is, the output swing angle after N iterations; n and m represent different sampling points; k is the kth sampling point;

[0024] Set the sampling time and collect the discretized experimental data of the multi-loop system; determine the number of neurons in the input layer, hidden layer, and output layer of the BP neural network to be 4-10-1; select the nozzle swing amplitude, swing frequency, working pressure, and restoring torque as input neurons, and select the swing angle as output neuron;

[0025] Assume that the weights and thresholds of the BP neural network are random numbers in [-1, 1]. During the training of the mth group of learning samples, calculate the hidden layer input function net of the BP neural network. jm , and then use the S function as the activation function, and the output function of the hidden layer neuron is:

[0026]

[0027] Where, ω ij is the connection weight between the input layer and the hidden layer; x im is the input signal of BP neural network; θ j is the threshold of hidden layer neurons;

[0028] Similarly, the neuron function of the BP neural network output layer is:

[0029]

[0030] Where: net m is the input function of the output layer neurons of the BP neural network, ω j is the connection weight between the hidden layer and the output layer, and θ is the neuron threshold of the output layer;

[0031] Define the expected output function and calculate the global error;

[0032] Determine the weight correction criteria of BP neural network, adopt the steepest descent method, introduce the learning step length to reversely calculate the weight of BP neural network;

[0033] The total number of BP neural network learning samples, the number of training times for each group of samples, the momentum factor α, the learning step size, and the allowable error are set, and the BP neural network is trained to obtain a dual flexible nozzle neural network model.

[0034] Optionally, the BP neural network is a feedforward network, and the identification structure of the dual-flexible nozzle BP neural network includes two modules: a measured system and an identification model; the measured system is the dual-flexible nozzle to be identified, and the identification model is a dual-flexible nozzle neural network model;

[0035] The BP neural network learns the input and output of the system under test to minimize the required error function and filter out the relationship between the input and output data of the system under test; when the error is less than the preset threshold, the identification is considered successful.

[0036] Optionally, the identification structure of the dual flexible nozzle BP neural network is a "series-parallel" type;

[0037] The input of the neural network with a "series-parallel" structure is composed of the input u(k) and output z(k) of the system under test and the time-delayed signals of the two. The "series-parallel" structure feeds back the output of the system under test to the identification model, which is conducive to ensuring the stability and convergence of the identification model and has a simple structure.

[0038] Optionally, the dual flexible nozzle neural network model is obtained by using a load perturbation modeling method to fuse the pseudo-random sequence into Gaussian noise, specifically:

[0039] The Gaussian signal is used as the main interference signal, the pseudo-random signal is used as the auxiliary inspiration signal, and the sequence generation time is used as the superposition basis to superimpose the pseudo-random signal and the Gaussian signal;

[0040] in, is the characteristic polynomial generated by the pseudo-random sequence; x i To specify the coefficient a i The value of x itself has no practical meaning;

[0041] The mathematical formula of Gaussian signal is G(x,y)=f(x,y)+n(x,y), and the power spectrum obeys the normal distribution. Where μ is the mean and σ is the variance;

[0042] According to the time series, at each pseudo-random signal true value sampling point, the pseudo-random signal result is multiplied by the Gaussian signal, and the formula is expressed as

[0043] The technical effect achieved by the above technical solution is: Gaussian noise is integrated into the pseudo-random sequence signal. Since the pseudo-random sequence has good randomness and a correlation function close to white noise, and has predetermined determinism and repeatability, these characteristics enable the pseudo-random sequence to be used to simulate environmental noise and interference, and the intensity can be controlled at will. Therefore, by integrating the pseudo-random sequence into the Gaussian noise, the environmental noise and disturbance can be changed by changing the pseudo-random sequence intensity, so as to detect the control quality of the system under different disturbances.

[0044] Optionally, the Gray Wolf optimized PID controller algorithm includes the following steps:

[0045] Initialize the gray wolf population, where each gray wolf position represents a set of PID parameters;

[0046] Calculate the fitness value of each gray wolf;

[0047] Determine the optimal solution, optimal solution and suboptimal solution of the current PID according to the fitness value, and calculate the distance control parameters;

[0048] When the maximum number of iterations is reached, the optimal parameter system response is output.

[0049] The technical effect achieved by the above technical solution is: combining the Gray Wolf optimization algorithm with PID control, constructing the Gray Wolf optimized PID control algorithm, applying it to the control of the electric servo mechanism-dual flexible nozzle system, and adaptively adjusting the PID parameters to ensure that the system works in a good state and achieve satisfactory control effects.

[0050] It can be seen from the above technical solution that compared with the prior art, the present invention discloses a neural network modeling and control method for a dual-flexible nozzle of an electric servo mechanism, identifies the dual-flexible nozzle and obtains a neural network model of the dual-flexible nozzle, effectively solving the problems of time-varying parameters, serious interference from nonlinear factors and structural defects of conventional models; secondly, the present invention proposes a Gray Wolf optimized PID controller, which combines the functions of a neural network identifier and a Gray Wolf optimized PID controller, and is applied to the control of the dual-flexible nozzle system, comprehensively improving the control effect and control quality of the system, thereby effectively solving the working performance of the electric servo mechanism-dual-flexible nozzle under actual working conditions. This method is suitable for ground simulation experiments of dual-flexible nozzles under experimental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0052] Figure 1 It is the control structure diagram of the multi-loop system;

[0053] Figure 2 It is the structural diagram of BP neural network model;

[0054] Figure 3 This is a schematic diagram of the BP neural network identification of dual flexible nozzles;

[0055] Figure 4 This is a schematic diagram of the neural network model of the electric servo mechanism-dual flexible nozzle;

[0056] Figure 5 This is the flow chart of the GWO-PID algorithm;

[0057] Figure 6 This is the structure diagram of the GWO-PID algorithm. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0059] The embodiment of the present invention discloses a neural network modeling and control method for a double flexible nozzle of an electric servo mechanism, comprising the following steps:

[0060] The neural network identifier uses the BP neural network algorithm to identify the model parameters of the multi-loop system according to the swing angle and obtains the neural network model of the dual flexible nozzle;

[0061] Taking the dual-flexible nozzle neural network model as the controlled object of the multi-loop system, the Grey Wolf Optimized PID controller (GWO-PID) uses the deviation between the multi-loop system command swing angle and the dual-flexible nozzle neural network model as the error signal according to the parameter changes of the dual-flexible nozzle neural network model and the disturbance of the electric servo mechanism (referred to as the actuator) to determine the neural network system learning algorithm. The output of the Grey Wolf Optimized PID controller is the control voltage of the electric servo mechanism system.

[0062] Further, refer to Figure 1 The multi-loop system control structure diagram shown in the figure includes an angular position controller, a speed controller, a current controller, a motor, a reduction mechanism, a dual flexible nozzle, a tension and pressure sensor, a current sensor, a speed sensor, a posture sensor, and a data acquisition card; wherein:

[0063] The angular position controller, the speed controller, the current controller, the motor, the speed reduction mechanism, and the double flexible nozzles are connected in sequence;

[0064] The speed reduction mechanism, the tension and pressure sensor, and the data acquisition card are connected in sequence;

[0065] The motor, current sensor, data acquisition card and current controller are connected in sequence;

[0066] The motor, speed sensor, data acquisition card and speed controller are connected in sequence;

[0067] The double flexible nozzles, attitude sensor, data acquisition card and angular position controller are connected in sequence.

[0068] According to the above analysis, Figure 1 The controller has the functions of current control, speed control and posture control. The acquisition range of each sensor meets the requirements, and the returned data are all current or voltage signals. It can be cascaded with the controller and driver to form a three-closed-loop system.

[0069] Furthermore, when the model parameter identification of the multi-loop system is performed, the nozzle swing amplitude, swing frequency, swing speed, restoring torque and system working pressure output by the sensor are used as input variables according to the multi-loop system test excitation signal input, and the swing angle is used as the output variable;

[0070] Methods for model parameter identification include online identification and offline identification. In this embodiment, an offline identification method based on experimental data is adopted.

[0071] Furthermore, a dual flexible nozzle neural network model is obtained, which specifically includes the following steps:

[0072] The forward dynamics model of the nonlinear system is approximated by a neural network. The neural network model is parallel to the actual system in structure. The neural network model of the dual flexible nozzle is discretized to obtain a discretized model. The discretized model is:

[0073] y N (k+1)=f[y(k),...,y(k-n+1),u(k),u(k-1),...,u(k-m+1)];

[0074] Where: f is an unknown nonlinear function that describes the system characteristics; y(k) is the system output; u(k) is the system input, y N(k+1) is the output of the neural network, that is, the output swing angle after N iterations; n and m represent different sampling points; k is the kth sampling point;

[0075] Set the sampling time and collect the discretized experimental data of the multi-ring system; refer to Figure 2 The structural diagram of the BP neural network model is shown in Figure 1. The number of neurons in the input layer, hidden layer, and output layer of the BP neural network is determined to be 4-10-1. The nozzle swing amplitude, swing frequency, working pressure, and restoring torque are selected as input neurons, and the swing angle is selected as the output neuron.

[0076] Assume that the weights and thresholds of the BP neural network are random numbers in [-1, 1]. During the training of the mth group of learning samples, calculate the hidden layer input function net of the BP neural network. jm , and then use the S function as the activation function, and the output function of the hidden layer neuron is:

[0077]

[0078] Where, ω ij is the connection weight between the input layer and the hidden layer; x im is the input signal of BP neural network; θ j is the threshold of hidden layer neurons;

[0079] Similarly, the neuron function of the BP neural network output layer is:

[0080]

[0081] Where: net m is the input function of the output layer neurons of the BP neural network, ω j is the connection weight between the hidden layer and the output layer, and θ is the neuron threshold of the output layer;

[0082] Define the expected output function and calculate the global error;

[0083] Determine the weight correction criteria of BP neural network, adopt the steepest descent method, introduce the learning step length to reversely calculate the weight of BP neural network;

[0084] The total number of BP neural network learning samples, the number of training times for each group of samples, the momentum factor α, the learning step size, and the allowable error are set, and the BP neural network is trained to obtain a dual flexible nozzle neural network model.

[0085] Furthermore, the BP neural network is a feedforward network, and its structure is as follows Figure 3 As shown, u(k) is the input, z(k) is the output, v(k) is the noise input, e(k) is the error between the identification model output and the true output, and P is the dual flexible nozzle to be identified.

[0086] Specifically, the identification structure of the dual-flexible nozzle BP neural network includes two modules: the measured system and the identification model. The measured system is the dual-flexible nozzle to be identified, and the identification model is the dual-flexible nozzle neural network model.

[0087] The BP neural network learns the input and output of the system under test to minimize the required error function and filter out the relationship between the input and output data of the system under test. Identification does not focus on how the neural network approaches the system under test, but only cares about the difference between the network output and the output of the system under test. When the error is less than the preset threshold, the identification is considered successful.

[0088] From the perspective of the input-output relationship between the identification model and the system under test, the identification structure can be divided into two types: parallel type and "series-parallel" type, such as Figure 3 As shown, the input of the neural network of the "series-parallel" structure is composed of the input u(k) and output z(k) of the system under test and the time-delayed signals of the two. Unlike the parallel mode, the "series-parallel" structure feeds back the output of the system under test to the identification model, which is conducive to ensuring the stability and convergence of the identification model, and has a simple structure. Therefore, this embodiment adopts the "series-parallel" identification model.

[0089] In addition, if Figure 3 As shown in the figure, both the input, output and feedback links have a time lag. In this model, the error caused by the time lag can be ignored during offline identification. The neural network corrects the neural network by the error between the actual system output and the neural network output. Combined with the established dual flexible nozzle neural network identification model, a system simulation experiment is designed, as shown in the figure. Figure 4 As shown, the four groups of data 1, 2, 3, and 4 represent the input parameters such as frequency, pressure, torque, and amplitude respectively.

[0090] Further, if Figure 4 As shown, the dual flexible nozzle neural network model is obtained by adopting the load disturbance modeling method, and the Gaussian noise is integrated into the pseudo-random sequence signal. Since the pseudo-random sequence has good randomness and a correlation function close to white noise, and has predetermined determinism and repeatability, these characteristics enable the pseudo-random sequence to be used to simulate environmental noise and interference, and its intensity can be controlled at will. Therefore, in this embodiment, the pseudo-random sequence is integrated into the Gaussian noise, and the environmental noise and disturbance can be changed by changing the pseudo-random sequence intensity, so as to detect the control quality of the system under different disturbances.

[0091] Specifically, a pseudo-random sequence is a deterministic sequence with certain random properties. It is generated by a shift register, yet possesses certain random properties. During system testing and simulation experiments, random interference signals are often introduced to verify the anti-interference capabilities of the system control algorithm. However, to ensure the randomness and non-repetitive nature of the interference, the pseudo-random sequence introduced in this embodiment is fused with a Gaussian noise signal in the following manner:

[0092] The Gaussian signal is used as the main interference signal, the pseudo-random signal is used as the auxiliary inspiration signal, and the sequence generation time is used as the superposition basis to superimpose the pseudo-random signal and the Gaussian signal;

[0093] in, is the characteristic polynomial generated by the pseudo-random sequence; x i To specify the coefficient a i The value of x itself has no practical meaning, and there is no need to calculate the value of x;

[0094] The mathematical formula of Gaussian signal is G(x,y)=f(x,y)+n(x,y), and the power spectrum obeys the normal distribution. Where μ is the mean and σ is the variance. Gaussian noise is a type of noise whose probability density function (power spectrum) follows a Gaussian distribution (i.e., normal distribution) and whose amplitude follows a Rayleigh distribution.

[0095] According to the time series, at each sampling point of the pseudo-random signal true value (defined as 0 for false and 1 for true), the pseudo-random signal result is multiplied by the Gaussian signal, and the formula is expressed as

[0096] like Figure 4 The figure shows an enhanced random search (ARS) algorithm used in this embodiment. ARS is a simplified learning algorithm and a linear strategy for solving continuous control problems. It seeks the maximum system output angle and provides feedback to the control end to compensate and correct control parameters. This algorithm improves system stability and control accuracy. Compared with many difficult-to-reproduce optimization algorithms, this algorithm has a simple structure, high efficiency, and can be applied in real-world environments.

[0097] also, Figure 4 The actuator is composed of a motor and a reduction mechanism mathematical model, and Figure 1 different, Figure 1 The reduction mechanism is a gear reduction ball screw mechanism; Figure 4 The control drive will Figure 1 The control algorithm of the three closed-loop controllers shown is integrated into one. Figure 1 The speed controller and current controller both adopt PI control strategy, while the actuator stroke control and angular position controller apply the gray wolf optimized PID control strategy of this embodiment.

[0098] The PID control algorithm is based on the deviation between input and output to form the control deviation, and the output control quantity acts on the actuator through proportional integral differential operation. Due to its advantages such as simple principle and easy implementation, it has been widely used in the field of flexible nozzle vector control. However, the traditional PID control algorithm parameters (K p , K i , K d ) is complex to tune and has poor adaptability to nonlinear systems, which in turn affects the stability and control accuracy of the electric servo mechanism-dual flexible nozzle system. Based on the above problems, the Grey Wolf optimization algorithm is combined with PID control to construct the Grey Wolf optimization PID control algorithm (GWO-PID), which is applied to the control of the electric servo mechanism-dual flexible nozzle system. The PID parameters (K p , K i , K d ) to make adaptive adjustments to ensure that the system works in a good state and achieves satisfactory control effects. The process of the GWO-PID algorithm is as follows: Figure 5 As shown, the specific steps include:

[0099] Initialize the gray wolf population, where each gray wolf position represents a set of PID parameters;

[0100] Calculate the fitness value of each gray wolf;

[0101] Determine the optimal solution, optimal solution and suboptimal solution of the current PID according to the fitness value, and calculate the distance control parameters;

[0102] When the maximum number of iterations is reached, the optimal parameter system response is output.

[0103] Furthermore, GWO is a metaheuristic algorithm proposed to simulate the hierarchy and hunting behavior of gray wolf packs. A gray wolf pack is usually composed of four levels of gray wolves: α, β, δ, and ω. Among them, α wolf is the gray wolf with the highest level in the pack, and is mainly responsible for major decisions such as leadership, hunting, and food distribution; β wolf is the gray wolf subordinate to α wolf, and is mainly responsible for assisting α wolf in making decisions; δ wolf is the gray wolf subordinate to α and β wolf, and is mainly responsible for sentinel and reconnaissance; ω wolf is the gray wolf with the lowest level, and is mainly responsible for maintaining the internal balance of the gray wolf pack. The hunting behavior of gray wolf packs can be divided into three stages: (1) tracking and approaching prey; (2) surrounding and forcing prey to stop moving; (3) attacking and capturing prey. Assume that the number of gray wolf population is N, and the position of the i-th gray wolf is X i , the location of wolf α is the optimal solution for the group, the location of wolf β is the second optimal solution, and the location of wolf δ is the third optimal solution. The mathematical model of gray wolf predation and hunting behavior can be described as follows.

[0104] (1) The mathematical model of tracking and encirclement behavior is:

[0105] D=|CX p (t)-X(t)|

[0106] X(t+1)=X p (t)-A·D

[0107] Among them, t is the current iteration number, D represents the distance between the wolf and the prey, P represents the prey, X p (t) is the location of the prey at the tth iteration, X p (t) is the location of the gray wolf individual, A and C are parameters, which can be expressed as:

[0108] A=2ar1-a

[0109] C=2r2

[0110] Among them, r1 and r2 are random numbers between [0,1], and a is the algorithm convergence factor, which can be defined as:

[0111]

[0112] Among them, t max is the maximum number of iterations.

[0113] (2) The mathematical model for forcing the prey to stop moving and attacking to capture the prey is:

[0114] D α =|C1X α -X|

[0115] D β =|C2X β -X|

[0116] D δ =|C3X δ -X|

[0117] Among them, the above formulas represent the distances between α wolf, β wolf, δ wolf and ω wolf respectively;

[0118] X1=X α -A1·(D α )

[0119] X2=X β -A2·(D β )

[0120] X3=X δ -A3·(D δ )

[0121] X α 、X β and X δdenote the position vectors of α wolf, β wolf and δ wolf respectively;

[0122]

[0123] The above formula is the position update method of ω wolf.

[0124] The electric servo mechanism dual flexible nozzle neural network model system is selected. The (stroke) displacement of the electric servo mechanism and the nozzle swing angle are respectively taken as the difference with the target value as the control quantity. The control effect is achieved by superposition of the stroke PID controller and the swing angle PID controller, and the optimal PID control parameters are screened by using GWO, such as Figure 6 shown.

[0125] When establishing a system simulation model, this embodiment proposes a modeling method that simulates aerodynamic noise into a Gaussian noise signal and integrates a pseudo-random sequence disturbance signal to simulate the interference situation of the dual-flexible nozzle under ground experimental conditions, thereby reflecting the true simulation capability of the entire system (electric servo mechanism-dual-flexible nozzle model) in the simulation, and is used to verify the control algorithm's requirements for technical indicators such as the response speed, control accuracy, and tracking capability of the electric servo mechanism-dual-flexible nozzle system.

[0126] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. The above description of the disclosed embodiments enables professionals and technicians in this field to implement or use the present invention. Various modifications to these embodiments will be apparent to professionals and technicians in this field, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A neural network modeling and control method for a double flexible nozzle of an electric servo mechanism, characterized in that: The following steps are involved: The neural network identifier uses the BP neural network algorithm to identify the model parameters of the multi-loop system according to the swing angle and obtains the neural network model of the dual flexible nozzle; The dual-flexible nozzle neural network model is used as the controlled object of the multi-loop system. The Gray Wolf optimized PID controller uses the deviation between the multi-loop system command swing angle and the dual-flexible nozzle neural network model as the error signal according to the parameter changes of the dual-flexible nozzle neural network model and the disturbance of the electric servo mechanism to determine the neural network system learning algorithm. The output of the Gray Wolf optimized PID controller is the control voltage of the electric servo mechanism system. The neural network model of the dual flexible nozzle is obtained by using the load perturbation modeling method, which fuses the pseudo-random sequence into the Gaussian noise. Specifically: The Gaussian signal is used as the main interference signal, the pseudo-random signal is used as the auxiliary inspiration signal, and the sequence generation time is used as the superposition basis to superimpose the pseudo-random signal and the Gaussian signal; in, is the characteristic polynomial generated by the pseudo-random sequence; x i To specify the coefficient a i The value of x itself has no practical meaning; The mathematical formula of Gaussian signal is G(x,y)=f(x,y)+n(x,y), and the power spectrum obeys the normal distribution. Where μ is the mean and σ is the variance; According to the time series, at each pseudo-random signal true value sampling point, the pseudo-random signal result is multiplied by the Gaussian signal, and the formula is expressed as 2. The neural network modeling and control method of a double flexible nozzle of an electric servo mechanism according to claim 1 is characterized in that: The multi-loop system includes an angular position controller, a speed controller, a current controller, a motor, a reduction mechanism, a dual flexible nozzle, a tension and pressure sensor, a current sensor, a speed sensor, a posture sensor, and a data acquisition card; among which: The angular position controller, the speed controller, the current controller, the motor, the speed reduction mechanism, and the double flexible nozzles are connected in sequence; The speed reduction mechanism, the tension and pressure sensor, and the data acquisition card are connected in sequence; The motor, current sensor, data acquisition card and current controller are connected in sequence; The motor, speed sensor, data acquisition card and speed controller are connected in sequence; The double flexible nozzles, attitude sensor, data acquisition card and angular position controller are connected in sequence.

3. The neural network modeling and control method of a double flexible nozzle of an electric servo mechanism according to claim 1 is characterized in that: When performing model parameter identification of a multi-loop system, the experimental excitation signal of the multi-loop system is input, the nozzle swing amplitude, swing frequency, swing speed, restoring torque and system working pressure output by the sensor are used as input variables, and the swing angle is used as the output variable; The methods of model parameter identification include online identification and offline identification.

4. The neural network modeling and control method of a double flexible nozzle of an electric servo mechanism according to claim 1 is characterized in that: The dual flexible nozzle neural network model is obtained, which specifically includes the following steps: The forward dynamics model of the nonlinear system is approximated by a neural network, and the neural network model of the dual flexible nozzle is discretized to obtain a discretized model; the discretized model is: y N (k+1)=f[y(k),...,y(k-n+1),u(k),u(k-1),...,u(k-m+1)]; Where: f is an unknown nonlinear function that describes the system characteristics; u(k) is the system input; y(k) is the system output; y N (k+1) is the output of the neural network, that is, the output swing angle after N iterations; n and m represent different sampling points; k is the kth sampling point; Set the sampling time and collect the discretized experimental data of the multi-loop system; determine the number of neurons in the input layer, hidden layer, and output layer of the BP neural network to be 4-10-1; select the nozzle swing amplitude, swing frequency, working pressure, and restoring torque as input neurons, and select the swing angle as output neuron; Assume that the weights and thresholds of the BP neural network are random numbers in [-1, 1]. During the training of the mth group of learning samples, calculate the hidden layer input function net of the BP neural network. jm , and then use the S function as the activation function, and the output function of the hidden layer neuron is: Where, ω ij is the connection weight between the input layer and the hidden layer; x im is the input signal of BP neural network; θ j is the threshold of hidden layer neurons; Similarly, the neuron function of the BP neural network output layer is: Where: net m is the input function of the output layer neurons of the BP neural network, ω j is the connection weight between the hidden layer and the output layer, and θ is the neuron threshold of the output layer; Define the expected output function and calculate the global error; Determine the weight correction criteria of BP neural network, adopt the steepest descent method, introduce the learning step length to reversely calculate the weight of BP neural network; The total number of BP neural network learning samples, the number of training times for each group of samples, the momentum factor α, the learning step size, and the allowable error are set, and the BP neural network is trained to obtain a dual flexible nozzle neural network model.

5. The neural network modeling and control method of a double flexible nozzle of an electric servo mechanism according to claim 1 is characterized in that: The BP neural network is a feedforward network. The identification structure of the dual-flexible nozzle BP neural network includes two modules: the measured system and the identification model. The measured system is the dual-flexible nozzle to be identified, and the identification model is the dual-flexible nozzle neural network model. The BP neural network learns the input and output of the system under test to minimize the required error function and filter out the relationship between the input and output data of the system under test; when the error is less than the preset threshold, the identification is considered successful.

6. The neural network modeling and control method of a double flexible nozzle of an electric servo mechanism according to claim 5 is characterized in that: The identification structure of the dual flexible nozzle BP neural network is a "series-parallel" type; The input of the neural network of the "series-parallel" structure consists of the input u(k), output z(k) of the system under test and the time-delayed signals of the two. The output of the system under test is fed back to the identification model of the "series-parallel" structure.

7. The neural network modeling and control method of a double flexible nozzle of an electric servo mechanism according to claim 1 is characterized in that: The Gray Wolf algorithm for optimizing the PID controller consists of the following steps: Initialize the gray wolf population, where each gray wolf position represents a set of PID parameters; Calculate the fitness value of each gray wolf; Determine the optimal solution, optimal solution and suboptimal solution of the current PID according to the fitness value, and calculate the distance control parameters; When the maximum number of iterations is reached, the optimal parameter system response is output.

Images