A control law automatic design method based on long short-term memory network

Abstract

The application discloses a control law automatic design method based on a long short-term memory network. The method draws on the experience of network architecture search method in deep learning, utilizes the advantage of a recurrent neural network in time sequence correlation exploration, converts a control law structure design problem into a directed acyclic graph topological relation automatic search problem, realizes automatic generation of a flight control law control structure, and realizes automatic setting of parameters under a given control law structure based on a genetic algorithm. The method aims at overcoming the limitation that current control law automatic optimization can only utilize a heuristic algorithm to automatically set controller parameters for a known controller structure, reducing the artificial design workload, and improving the control effect under complex design input conditions.

Description

Technical Field

[0001] This invention relates to an automatic control law design method based on long short-term memory networks, belonging to the field of aircraft control system design. Background Technology

[0002] Currently, control law design is mainly done manually. The control law structure is first selected according to different functional requirements, and then the control law parameters are manually adjusted by the designers according to the characteristics of the controlled object to adapt to the specific control task. This requires a high level of design experience from the designers, and the parameter adjustment process mainly relies on trial and error, which is inefficient, time-consuming and labor-intensive.

[0003] Artificial intelligence technology is developing rapidly, with deep learning-based target recognition algorithms being widely used. However, similar to the design process of control laws, the design of the structure and weights of various target recognition neural networks often requires manual trial and error, consuming a lot of manpower and time. To address these pain points, Neural Architecture Search (NAS) technology, which combines exploration-trial-error heuristic search, has emerged, enabling automatic optimization of neural network architecture and weights, thus reducing the difficulty of manual design.

[0004] However, the above methods are only suitable for the automatic design of neural network architecture and weights, and cannot solve the problem of automatic design of controller structure and parameters. Some existing automatic controller parameter tuning methods, such as the Grey Wolf algorithm and the Cuckoo algorithm, can only adjust parameters under the premise of a given controller structure. The overall design efficiency needs to be improved. It is necessary to draw on the ideas of automatic design of neural network architecture and weights to give an integrated automatic controller design method. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an automatic control law design method based on long short-term memory networks, which realizes the integrated automatic design of controller structure and parameters, reduces the design difficulty, and improves the control effect under complex design input conditions.

[0006] The technical solution of this invention is as follows: an automatic control law design method based on long short-term memory networks, comprising:

[0007] The amplitude-frequency characteristics of the controlled object are analyzed using Bode plots. Based on rigid body and elastic stability conditions, the basic controller form and the number of filters are obtained, and the controller structure is transformed into N... l The typical form of cascading layer units;

[0008] Based on the basic controller and filter, the available structural modules in each cascaded layer are separated to form the feasible search space of the cascaded controller.

[0009] Based on the obtained feasible search space of the cascaded controller, a structure generator based on the Long Short-Term Memory (LSTM) network is established to complete the cyclic generation of each layer of the cascaded controller structure; at the same time, the number of hidden neural networks in the LSTM network is set, and the number of output classifications of the LSTM network structure generator is determined according to the number of optional structures in each layer of the cascaded controller.

[0010] For any controller structure composed of feasible search space combinations, a unified and feasible parameter tuning method is designed. The parameter tuning algorithms for the basic controller and filter parts are given respectively. The number of parameters to be tuned and the possible value range of the parameters are determined. A fitness function that can evaluate the quality of controller parameters under arbitrary structure is constructed.

[0011] By combining the structure generator and parameter tuning method based on the Long Short-Term Memory Network (LSTM) obtained in the previous step, an integrated training and learning architecture with step-by-step optimization of the inner and outer loops is established. The parameters of the LSTM structure generator of the outer loop and the control law parameters of the inner loop are trained respectively, and the integrated automatic generation of the cascaded controller structure and parameters is completed.

[0012] Extract the cascaded controller with the highest fitness score generated automatically, analyze the amplitude-frequency characteristics after adding the controller using methods such as Bode plots, complete time-domain simulation and data analysis, and analyze whether it meets the design requirements. If it does not meet the requirements, adjust the hyperparameter settings and redesign until the design requirements are met.

[0013] The method involves analyzing the amplitude-frequency characteristics of the controlled object using Bode plots and other techniques. Based on rigid body and elastic stability conditions, the basic controller form and the number of filters are obtained, transforming the controller structure into N... l Typical cascading forms of layer units include:

[0014] All control elements are normalized to the forward path and written as several typical modules cascaded in series. For any closed-loop feedback system, assuming C(s) is the controller transfer function, and considering that any controller can be written as N... l The transfer function is multiplied together, i.e., the controller is set to N. l In the case of serially cascaded layer controllers, the controller transfer function C(s) is:

[0015]

[0016] in, Ncl represents a transfer function selected by the l-th layer control unit from the typical modules, and Ncl represents the betweenness number of the transfer functions that can be selected for each layer.

[0017] The system consists of two parts: a basic controller and filters. The cascaded controller is composed of a basic controller and multiple filters cascaded together. For actual aircraft, in addition to rigid body characteristics, there are also multiple high-frequency elastic resonant modes. First, a basic controller is determined to complete the stable control of the rigid mode of the aircraft. Then, multiple filters are designed for the elastic resonant mode. Finally, the controller parameters are repeatedly adjusted to achieve the optimal overall performance.

[0018] Depending on the specific characteristics of the controlled object and the different control performance indicators, basic controllers include PID controllers, robust controllers, LQR controllers, and sliding mode controllers.

[0019] The process involves establishing a structure generator based on a Long Short-Term Memory (LSTM) network to cyclically generate the structure of each layer of the cascade controller. The number of hidden neural networks in the LSTM network is set, and the number of output classifications of the LSTM network structure generator is determined based on the number of selectable structures for each layer of the cascade controller. This includes: considering the correlation between the selection of each layer's structure, a classifier is constructed using the LSTM network to classify the structure type of each layer in an autoregressive manner to complete the layer-by-layer generation of the cascade controller structure. The decision from the previous step is used as input to generate the output of the next layer, while 0 is used as input in the first step. The generator cyclically generates the filter structure selection for each layer in a time sequence. The number of LSTM network iterations equals the number of cascade controller layers to be generated. The number of hidden neural networks in the LSTM network is also set according to the complexity of the problem.

[0020] The proposed design provides a unified and feasible parameter tuning method, outlining parameter tuning algorithms for both the basic controller and the filter. It determines the number of parameters to be tuned, the possible value ranges of these parameters, and constructs a fitness function to evaluate the quality of controller parameters under arbitrary structures. This includes automatically calculating the controller parameter η for each possible structure. C The design of the controller parameter η; C It is divided into two parts, including the basic controller parameter η. base And other parameters η oth The parameters need to maximize the control bandwidth while meeting the specified gain and phase margins. After determining the types of controller parameters, the possible range of values ​​for the parameters to be tuned is given based on experience, and a fitness function suitable for evaluating the quality of controller parameters under any structure is constructed. The fitness function is specifically set according to the characteristics and index requirements of the controlled object.

[0021] For η base and η oth The two sets of parameters are tuned using different optimization methods; for the basic controller parameter η... baseThe controller parameters are tuned using the corresponding self-tuning theory, employing the PID parameter automatic tuning method. For other parameters η... oth Parameter tuning is performed using heuristic methods, such as those based on genetic algorithms.

[0022] The step establishes an integrated training and learning architecture with step-by-step optimization of the inner and outer loops. The parameters of the LSTM (Long Short-Term Memory) network structure generator in the outer loop and the control law parameters in the inner loop are trained separately to complete the integrated automatic generation of the cascaded controller structure and parameters. This includes: the entire training process includes two sets of learnable parameters, namely the LSTM parameters θ and the controller parameters ω; the training process adopts a step-by-step integrated training method, including two interleaved stages. The first stage trains the controller parameters ω of the filter, optimizing the parameters as a whole through automatic parameter tuning; the second stage trains the LSTM parameters θ. The two stages alternate, and the total number of training rounds is set according to the complexity of the problem and the convergence status.

[0023] The process involves extracting the cascaded controller with the highest automatically generated fitness score, analyzing its amplitude-frequency characteristics after its addition using methods such as Bode plots, completing time-domain simulation and data analysis, and analyzing whether it meets the design requirements. If it does not, the hyperparameter settings are adjusted and the design is redesigned. This includes: after the integrated training and learning is completed, extracting the group of controllers with the highest reward score, analyzing the changes in amplitude-frequency characteristics after the controllers are added using Bode plots, completing time-domain simulation and data analysis, and analyzing whether the design index requirements are met using traditional analysis methods. If they are not met, the hyperparameter settings are adjusted and training and learning iterations are restarted until the requirements are met, thus completing the automatic design of the control law.

[0024] The hyperparameters include parameters in the Long Short-Term Memory (LSTM) network structure generator and various parameters during the training process.

[0025] The advantages of this invention compared to the prior art are:

[0026] (1) Based on the characteristics of the controlled object of the aircraft and combined with prior experience, the control law design problem is transformed into a directed acyclic graph generation and optimization problem, realizing the integrated optimization of control law structure and parameters, reducing the dependence on the experience of designers, overcoming the problem of numerous links in the traditional design process, and improving efficiency.

[0027] (2) The controller structure was generated cyclically by a structure generator based on Long Short-Term Memory Network (LSTM), the control law structure was decomposed into a cascaded controller form, and the controller structure design was transformed into a cyclic layer-by-layer classification and selection problem, which improved the intelligence level of the controller design process.

[0028] (3) By decomposing the inner and outer loops, the controller structure and controller parameters are simultaneously tuned. The inner loop completes the parameter tuning under the specified architecture through heuristic search such as genetic algorithm, and the outer loop completes the parameter learning of the structure generator through reinforcement learning mechanism. Through step-by-step iteration, global integrated optimization is achieved. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the automatic control law design method of the present invention;

[0030] Figure 2 Automatically generate schematic diagrams for cascade controller structures;

[0031] Figure 3 This is a schematic diagram of the LSTM-based structure generation process.

[0032] Figure 4 Here is a diagram of the LSTM-based structure generator.

[0033] Figure 5 To automatically design Bode comparison charts. Detailed Implementation

[0034] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Figure 1 As shown, the present invention proposes an automatic control law design method based on long short-term memory networks, the specific steps of which are as follows:

[0035] (1) Analysis of the characteristics of the controlled object

[0036] Because the control theories used by the controllers differ, their structures will vary significantly. Furthermore, in engineering, the controller structure is often very complex due to the need to address the rigidity and elasticity constraints of the controlled object. However, as shown in the attached... Figure 1 As shown, all control links can be normalized to the forward channel, and the classic multi-loop control law can be transformed into the form of several typical modules cascaded in series. Then, the automatic design problem of control law for complex controlled objects can be simplified by using a cascade controller.

[0037] like Figure 1 As shown, for any closed-loop feedback system, we can assume that C(s) is the controller transfer function, considering that any controller can be written as N l The transfer function is multiplied together, which allows the controller C(s) to be set to N. l The layer controllers are cascaded in series.

[0038] That is, the transfer function C(s) can be written as

[0039]

[0040] here This represents a transfer function selected by the l-th layer control unit from the typical optional modules, i.e., the attached... Figure 1 The transfer function of each connected part in the layer; N cl This represents the optional betweenness of the transfer function for each layer.

[0041] (2) Determining the search space

[0042] The optional modules in step (1) can be set from order 0 to N. cl The general form of the transfer function of order N can also be flexibly set according to the specific problem, with a minimum number of optional modules. Therefore, the controller C(s) can be decomposed into N... l The layer transfer function is in serial cascade form.

[0043]

[0044] However, for actual aircraft, in addition to possessing rigid body characteristics, they also exhibit various high-frequency elastic resonant modes. Therefore, as shown in the attached... Figure 2 As shown, a basic controller can be first determined to achieve stable control of the aircraft's rigid modes. Then, multiple filters (correction networks) are designed for the elastic resonant modes. Finally, the controller parameters are repeatedly adjusted to achieve optimal overall performance. Depending on the control performance indicators, the basic controller can be a PID controller, robust controller, LQR controller, sliding mode controller, etc.

[0045] Taking launch vehicle control as an example, the basic controller often uses a PID controller. However, due to the existence of elastic modes requiring filtering, a cascaded PID controller for rigid stability and multiple filters (correction networks) for elastic resonant mode stability are needed. Prior knowledge indicates that the filters (correction networks) required for launch vehicle attitude control generally use high-order low-pass filters. A common form is:

[0046]

[0047] Depending on the specific characteristics of the controlled object, the mathematical expression of C(s) in the continuous time domain can be further simplified as shown in the following equation.

[0048]

[0049] The basic controller uses a PID controller, and its transfer function is as follows:

[0050]

[0051] Therefore, a structure of two filter (correction network) modules cascaded is adopted. Each filter (correction network) has 3 possible forms, and the search space has a total of 9 possible structures. The transfer function forms of the 3 filters are shown in the table below.

[0052] Table 1. Transfer function forms of filters

[0053]

[0054] (3) Structure Generator Design

[0055] Considering the correlation between the selection of each layer of the structure, in order to complete the layer-by-layer generation of the cascade controller structure, as shown in the attached figure... Figure 3 As shown, a classifier is constructed using a Long Short-Term Memory (LSTM) network to classify the structure type of each layer in an autoregressive manner. That is, the decision of the previous step is used as input to generate the output of the next layer, while 0 is used as input in the first step. The generator generates the filter structure selection of each layer in a time-series cyclical manner.

[0056] Without loss of generality, assuming each filter layer can use one of the three filter modules listed in Table 1, and the number of cascaded layers is N, then the input size of the LSTM is 1, and the number of LSTM loops is N, meaning that N filters need to be generated iteratively. (Appendix) Figure 3 The bottom layer represents the output of the LSTM loop cell during a certain generation, while the middle layer represents the filter structure corresponding to the output of the LSTM loop cell.

[0057] In the specific example of launch vehicle control shown in Table 1, the number of output categories of LSTM is 3, and the number of filter layers to be generated is 2. Therefore, the input of the initial recurrent unit is 0, and the output is the probability of 3 categories. According to the problem size, the number of hidden neurons in the recurrent network LSTM can be set to 10.

[0058] (4) Parameter tuning method design

[0059] For steps (3) where any structure can be formed, the controller parameter η needs to be automatically completed. C As can be seen from the design, the controller parameter η C It is divided into two parts: basic controller parameters η base And other parameters η oth That is, η C ={η base ,η oth These parameters need to meet the specified gain margin gm (dB) and phase margin. At the same time, maximize the control bandwidth.

[0060] As shown in Table 1 of step (2), the basic controller parameter η in this example is... baseFor η PID including K P K I and K D Other parameters η oth This includes ω1, ω2, ω3, ω4, ω5, ω6, ω7, ω8, ζ1, ζ2, ζ3, and ζ4. Based on experience, the possible ranges of values ​​for the parameters to be tuned are given, and a fitness function F is constructed that can expand the control bandwidth while satisfying specific stability margins, and is suitable for evaluating the quality of controller parameters under arbitrary structures. The fitness function needs to be specifically set according to the characteristics and performance requirements of the controlled object.

[0061] Different optimization methods are used to obtain the parameters of the cascaded controller for the two sets of parameters mentioned above. For the parameter η of the basic controller... base The controller parameters are tuned using the corresponding self-tuning theory. In this example, η PID The parameters are tuned using the built-in MATLAB function pidtune; however, for other parameters η... oth The parameters were tuned using a metaheuristic method based on a genetic algorithm (GA). The detailed parameter tuning process is as follows.

[0062] Step 1): GA randomly generates the initial population η oth (A population has N) ind Individuals) as the first generation (i gen =1).

[0063] Step 2): Obtain the appropriate value for η using the pidtune function. oth N ind Candidate PID parameters η PID .

[0064] Step 3): Based on the defined fitness function (which reflects control bandwidth and stability margin), evaluate the fitness score f of all individuals in GA to obtain the elite parameter η of elite individuals. PID and η oth .

[0065] Step 4): If i is generated gen Less than the specified number N gen , then i gen =i gen +1 and go to step 5); otherwise, go to step 6).

[0066] Step 5): Perform genetic operations, such as selection, crossover, and mutation, and generate a new population η for the next generation. oth . when i gen ≤N gen Repeat steps 2) to 5).

[0067] Step 6): Use Elite η base and η oth Obtain the required controller C(s).

[0068] (5) Integrated training and learning

[0069] As attached Figure 4 As shown, in this example, the structure generator is an LSTM with 10 hidden units, and the decision is sampled in an autoregressive manner using a Softmax classifier. During the entire training process, there are two sets of learnable parameters: the LSTM parameters are denoted by θ, and the shared parameters of the filters are denoted by ω. The training process includes two interleaved stages. The first stage trains the shared parameters ω of the filters, optimizing the parameters as a whole using the parameter tuning method in step (4). The second stage trains the LSTM parameters θ. These two stages alternate during training, with the total number of rounds typically set between 100 and 200. The specific process is as follows.

[0070] Step 1): Train the shared parameter ω of the filter.

[0071] In this step, the policy π(m; θ) of the structure generator is fixed, and ω is iteratively tuned to maximize the fitness function F. The filter structure m samples from π(m; θ). We can update ω using any single model m sampled from π(m; θ). We train ω throughout the process of finding better fitness function values.

[0072] Step 2): Train the structure generator parameters θ.

[0073] In this step, ω is fixed and the policy parameter θ is updated, with the goal of maximizing the expected fitness. The gradients are computed using the classic REINFORCE algorithm in reinforcement learning, employing either the SGDM or Adam optimizer, and a moving average baseline is used to reduce variance.

[0074] The fitness F(m,ω) is calculated by applying the controller to the controlled object to encourage the selection of a well-performing controller architecture. The formula for the moving average baseline is:

[0075] baseline=decay_bl×baseline+(1-decay_bl)×coeff×F(m,ω)

[0076] Where, decay_bl is the sliding ratio, typically 0.9, coeff is the coefficient of the fitness function, typically 0.001; and the initial value of baseline is 0.

[0077] Step 3): Determine the optimal architecture and parameters.

[0078] After performing overall training for controller structure search and parameter optimization, we first select the controller structure with the highest probability from the trained policy π(m,θ), and then we select the model parameter ω with the highest fitness from the shared parameters.

[0079] (6) Analysis and confirmation of design results

[0080] After the integrated training is completed, the controller group with the highest reward score is extracted, as shown in the attached figure. Figure 5 As shown, the amplitude-frequency response changes after adding the controller are analyzed using Bode plots, and time-domain simulation and data analysis are performed to determine if the design requirements are met. If not, the hyperparameter settings in the above steps are adjusted, and training is restarted; if the requirements are met, the automatic design of the control law is complete.

[0081] Example:

[0082] In this project, η PID The parameters are tuned using MATLAB's built-in function `pidtune`, and η... oth The parameters were tuned using a genetic algorithm (GA). Based on the object's Nyquist curve and repeated experiments, the final fitness function was determined as follows:

[0083]

[0084]

[0085]

[0086] Considering both the convergence and computational time complexity of the genetic algorithm, the population size was set to 50, the elite number to 7, the crossover number to 40, the mutation number to 3, and the number of iterations to 3000. Other settings remained at their default values. The SGDM optimizer was used for parameter learning of the structure generator. Given that the goal in launch vehicle control missions is to achieve the best control performance and avoid overfitting, the moving average baseline was changed to use the current best result as the baseline.

[0087] Considering that the reward function may increase sharply during the search process, in order to prevent the structure search from getting trapped in local optima, the following function is used to limit the generator's reward, which can limit the reward to between 0 and 1.

[0088]

[0089] Ultimately, the average iteration time for each structure was 63.0 seconds. The output of the LSTM structure generator is shown in the table below. It can be seen that the first filter converges to a doublet, and the second filter converges to a doublet.

[0090] Table 2 LSTM Structure Generator Output

[0091] First filter The second filter (with a dual input) 0-fold filter 0.1086 0.0378 1st filter 0.2117 0.0603 2-stage filter 0.6797 0.9019

[0092] After automatic search and parameter optimization, the transfer function of the first filter can be obtained as follows:

[0093]

[0094] The transfer function of the second filter is:

[0095]

[0096] A comparison of Bode plots before and after adding a filter / correction network to the controlled object is shown in the attached figure. Figure 5 As shown, the gain margin and phase margin are improved, the elastic mode is suppressed, and the goal of automatic design of the cascade controller control law is achieved.

[0097] The above description is only the best specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the protection scope of the present invention.

[0098] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A control law automatic design method based on a long short-term memory network, characterized in that, include: The amplitude-frequency characteristics of the controlled object are analyzed by means of Bode diagram, and the basic controller form and the number of filters are obtained according to the rigid body and elastic stability conditions, so as to convert the controller structure into a form of cascade of typical units of layers; Based on the basic controller and filter, the available structural modules in each cascaded layer are separated to form the feasible search space of the cascaded controller. Based on the obtained feasible search space of the cascaded controller, a long short-term memory network is established. LSTM A structure generator is used to cyclically generate each layer of the cascade controller structure; simultaneously, a long short-term memory network is configured. LSTM The number of hidden layers in the neural network is determined, and the long short-term memory network is defined based on the number of optional structures in each layer of the cascaded controller. LSTM The structure generator outputs the number of categories; For any controller structure composed of feasible search space combinations, a unified and feasible parameter tuning method is designed. The parameter tuning algorithms for the basic controller and filter parts are given respectively. The number of parameters to be tuned and the possible value range of the parameters are determined. A fitness function that can evaluate the quality of controller parameters under arbitrary structure is constructed. Combined with the obtained long short-term memory-based network LSTM The structure generator and parameter tuning method are used to establish an integrated training and learning architecture with step-by-step optimization of the inner and outer loops, respectively, for the long short-term memory network of the outer loop. LSTM The structure generator parameters and the inner loop control law parameters are trained to complete the integrated automatic generation of the cascade controller structure and parameters; Extract the highest fitness score of the automatically generated cascade controller, through bode The amplitude-frequency characteristic after adding the controller is analyzed in a graph mode, time domain simulation and data analysis are completed, whether the design requirements are met is analyzed, if the design requirements cannot be met, the hyperparameter setting is adjusted for redesign until the design requirements are met; The integrated training learning architecture of the inner and outer ring step-by-step optimization respectively trains the long short-term memory network of the outer ring LSTM The structure generator parameters and the control law parameters of the inner ring are trained to complete the integrated automatic generation of the cascade controller structure and parameters, including: during the whole training process, two groups of learnable parameters, i.e. LSTM The parameters of the long short-term memory network And the controller parameters ​ The training process employs a step-by-step integrated training method, comprising two interleaved stages. The first stage trains the filter's controller parameters. The parameters are optimized as a whole through automatic parameter tuning methods, and the long short-term memory network is trained in the second stage. LSTM parameter The two phases are carried out alternately, and the total number of training rounds is set according to the complexity of the problem and the convergence.

2. The automatic control law design method based on long short-term memory networks according to claim 1, characterized in that: The method of analyzing the amplitude-frequency characteristics of the controlled object using Bode plots, and obtaining the basic controller form and number of filters based on rigid body and elastic stability conditions, transforms the controller structure into... Typical cascading forms of layer units include: Normalize all control elements to the forward path and write them as several typical modules cascaded in series. For any closed-loop feedback system, assume... The pass function for the controller, considering that any controller can be written as The transfer function is multiplied together, that is, the controller is set to... In the case of serially cascaded layer controllers, the controller transfer function... : Indicates the first The layer control unit selects a transfer function from a typical module. .

3. The automatic control law design method based on long short-term memory networks according to claim 1, characterized in that, The system consists of two parts: a basic controller and filters. The cascaded controller is composed of a basic controller and multiple filters cascaded together. For actual aircraft, in addition to rigid body characteristics, there are also multiple high-frequency elastic resonant modes. First, a basic controller is determined to complete the stable control of the rigid mode of the aircraft. Then, multiple filters are designed for the elastic resonant mode. Finally, the controller parameters are repeatedly adjusted to achieve the optimal overall performance.

4. The automatic control law design method based on long short-term memory networks according to claim 3, characterized in that, Depending on the specific characteristics of the controlled object and the control performance indicators, the basic controller includes PID Controller, Robust Controller LQR Controller, sliding mode controller.

5. The automatic control law design method based on long short-term memory networks according to claim 1, characterized in that, The establishment of a long short-term memory network LSTM A structure generator is used to cyclically generate each layer of the cascade controller structure; simultaneously, a long short-term memory network is configured. LSTM The number of hidden layers in the neural network is determined, and the long short-term memory network is defined based on the number of optional structures in each layer of the cascaded controller. LSTM The structure generator outputs the number of categories, including: considering that the selection of each layer of structure is related, a long short-term memory network is used to complete the layer-by-layer generation of the cascade controller structure. LSTM A classifier is constructed to classify the structure type of each layer in an autoregressive manner. This means the decision from the previous step is used as input to generate the output of the next layer, while 0 is used as input in the first step. The generator iteratively generates the filter structure selection for each layer in a time-series manner; Long Short-Term Memory (LSTM) network. LSTM The number of iterations equals the number of cascade controller layers to be generated; simultaneously, the long short-term memory network is configured according to the complexity of the problem. LSTM Number of hidden layers in a neural network.

6. The automatic control law design method based on long short-term memory networks according to claim 1, characterized in that, The proposed design provides a unified and feasible parameter tuning method, outlining parameter tuning algorithms for both the basic controller and the filter. It determines the number of parameters to be tuned, the possible value ranges of these parameters, and constructs a fitness function to evaluate the quality of controller parameters under any given architecture. This includes automatically completing the controller parameter tuning for each possible architecture. The design of the controller parameters; It is divided into two parts, including basic controller parameters. And other parameters The parameters need to maximize the control bandwidth while meeting the specified gain and phase margins. After determining the types of controller parameters, the possible range of values ​​for the parameters to be tuned is given based on experience, and a fitness function suitable for evaluating the quality of controller parameters under any structure is constructed. The fitness function is specifically set according to the characteristics and index requirements of the controlled object.

7. The automatic control law design method based on long short-term memory networks according to claim 6, characterized in that, against and The two sets of parameters are tuned using different optimization methods; for the parameters of the basic controller... The controller parameters are tuned using the corresponding self-tuning theory, employing an automatic PID parameter tuning method. For other parameters... Parameter tuning is performed using heuristic methods, such as those based on genetic algorithms.

8. The automatic control law design method based on long short-term memory networks according to claim 1, characterized in that, The cascaded controller with the highest automatically generated fitness score is extracted, through... bode The amplitude-frequency characteristics after adding the controller are analyzed using a graphical method. Time-domain simulation and data analysis are completed to determine if the design requirements are met. If not, the hyperparameter settings are adjusted and the design is redesigned. This includes: extracting the controller with the highest reward score after integrated training and learning; analyzing the changes in amplitude-frequency characteristics after adding the controller using a Bode plot; completing time-domain simulation and data analysis; and analyzing whether the design specifications are met using traditional analysis methods. If not, the hyperparameter settings are adjusted and training and learning are restarted iteratively until the requirements are met, thus completing the automatic design of the control law.

9. The automatic control law design method based on long short-term memory networks according to claim 8, characterized in that, The hyperparameters include the Long Short-Term Memory network. LSTM The parameters in the structure generator and various parameters during the training process.

Images