WO2020244346A1 - Procédé de commande inverse directe de d-fnn et système basé sur une stratégie d'élagage. - Google Patents

Procédé de commande inverse directe de d-fnn et système basé sur une stratégie d'élagage. Download PDF

Info

Publication number
WO2020244346A1
WO2020244346A1 PCT/CN2020/087798 CN2020087798W WO2020244346A1 WO 2020244346 A1 WO2020244346 A1 WO 2020244346A1 CN 2020087798 W CN2020087798 W CN 2020087798W WO 2020244346 A1 WO2020244346 A1 WO 2020244346A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
model
network model
pruning
fnn
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2020/087798
Other languages
English (en)
Chinese (zh)
Inventor
张彩霞
王向东
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Foshan University
Original Assignee
Foshan University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Foshan University filed Critical Foshan University
Publication of WO2020244346A1 publication Critical patent/WO2020244346A1/fr
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
    • G05B13/042—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators in which a parameter or coefficient is automatically adjusted to optimise the performance

Definitions

  • the present disclosure relates to the technical fields of automatic control, artificial intelligence and neural network, and in particular to a D-FNN direct inverse control method and system based on a pruning strategy.
  • a trained neural network can be regarded as an expression of knowledge, which is different from using IF-THEN rules to express local knowledge in a fuzzy system.
  • the neural network uses its structure, more definitely, through its connection rights and local processing. Units store knowledge in a distributed or local way.
  • the feedforward calculation in the neural network plays the same role as the forward reasoning in the fuzzy system.
  • Both of these systems can perform tasks by operating the stored knowledge according to the current situation, and have obtained the desired output.
  • the core of these two systems is to respond to the new situation by giving an appropriate behavior.
  • Fuzzy systems are interpolation inferences based on logical inferences, while neural networks are algebraic calculations based on generalization capabilities.
  • Fuzzy systems usually acquire knowledge from domain experts, and this knowledge is integrated into the system with the help of fuzzy logic theory.
  • neural networks usually obtain knowledge from samples, and this knowledge is absorbed into the neural network through training.
  • the combination of fuzzy systems and neural networks forms a fuzzy neural network, which focuses on the advantages of the two methods while avoiding their respective shortcomings. Applying neural networks to fuzzy systems can solve the problem of extracting fuzzy rules in fuzzy systems, and applying fuzzy systems to neural networks can avoid arbitrary selection of initial values.
  • Neural network simulates the structure of the human brain and has large-scale parallel and distributed information processing capabilities, but it cannot process and describe fuzzy information. Fuzzy system has the reasoning process that is easy to understand, but it is difficult to realize the function of adaptive learning. If the neural network is combined with the fuzzy system, it can have the effect of learning from each other.
  • the present disclosure provides a D-FNN direct inverse control method and system technical solution based on a pruning strategy. Due to the advantages of this method, such as simultaneous adjustment of structure and parameters during learning, and fast learning speed, it can The fuzzy logic system contains low-level neural network learning and calculation functions. On the other hand, it can also provide high-level human-like thinking and reasoning fuzzy logic systems for neural networks. D-FNN (Dynamic Fuzzy Neural Network) has real-time learning and Strong control ability, simultaneous parameter estimation and structure identification.
  • a D-FNN direct inverse control method based on a pruning strategy includes the following steps:
  • Step 1 Read the input data set
  • Step 2 Construct a dynamic fuzzy neural network model
  • Step 3 Adjust the dynamic fuzzy neural network model to a hierarchical learning dynamic fuzzy neural network model
  • Step 4 Pruning and removing the fuzzy rules in the hierarchical learning dynamic fuzzy neural network model through the pruning strategy to obtain the pruning neural network model;
  • Step 5 Perform data forgetting on the input data set by pruning the neural network model to obtain the control system parameter data set;
  • Step 6 construct an inverse control model and input the control system parameter data set to the inverse control model for control.
  • step 1 the input data set is observation data acquired under natural uncontrolled conditions.
  • step 2 the method of constructing a dynamic fuzzy neural network model is:
  • x 1 ,x 2 ,L,x r are input data
  • y is output data
  • c ij is a j-th Gaussian center membership function of x i
  • ⁇ j is the Gaussian width of the j-th membership function x i
  • r is the number of input data
  • u is the number of membership functions
  • N j is the output of the jth node:
  • C i is the center of the i-th RBF unit; a i and a k are both polynomial coefficients.
  • step 3 the method of adjusting the dynamic fuzzy neural network model to the hierarchical learning dynamic fuzzy neural network model is:
  • the structure of the dynamic fuzzy neural network model is not only determined by the receiving domain but also by the system error, although the two methods can be divided into hierarchical self-organizing methods, and the output error determines whether new fuzzy rules should be added;
  • Output error is determined as described below: (X i, t i) is the i-th input data, X-i indicates the i-th input vector, T i denotes the i th desired output dynamic fuzzy neural network output y i.
  • the dynamic fuzzy neural network model adds a new fuzzy rule, and finally a hierarchical learning dynamic fuzzy neural network model is obtained.
  • the value of k e is pre-selected according to the expected accuracy of the dynamic fuzzy neural network.
  • k e is a variable:
  • the maximum error is e max
  • the accuracy of the dynamic fuzzy neural network is e min
  • the convergence constant is ⁇ (0 ⁇ 1)
  • ⁇ i is the i-th convergence constant.
  • step 4 the fuzzy rules in the hierarchical learning dynamic fuzzy neural network model are pruned and eliminated by the pruning strategy to obtain the pruned neural network model:
  • the pruning strategy uses the error reduction rate (ERR) method.
  • ERR error reduction rate
  • Different pruning strategies using ERR, SVD (singular value decomposition) or ED (eigenvalue decomposition) methods will lead to different dynamic fuzzy neural network model structures, that is, different fuzzy rules and for a specific performance, each rule The parameters are different.
  • the experiment shows that the ERR method has obvious physical meaning and stable numerical calculation, which is why this method is widely used.
  • ⁇ R (r+1)u ⁇ n and T ⁇ R n is as follows:
  • the error vector is E ⁇ R n, return amount h i, the return amount of q i. Error reduction rate
  • the threshold is set in advance as k err , the i-th fuzzy rule can be eliminated, and finally the pruned neural network model is obtained.
  • the training error without trimming is always smaller than the error after trimming, but after trimming, whether it is ERR, SVD or ED method, the system test error (judged by RMSE) is always smaller than that of the untrimmed dynamic fuzzy neural network model .
  • step 5 the method for obtaining the control system parameter data set by performing data forgetting on the input data set by pruning the neural network model is:
  • X j (x 1j ,x 2j , ⁇ ,x rj ) is the input observation data, and y i is the output data of the system, which can be written in the form of expression (16):
  • the transposition of ⁇ is ⁇ T
  • the Kalman filter algorithm is used to determine the following weights:
  • the coefficient matrix after i iterations is W i
  • the error covariance matrix of the i-th observation data is S i
  • ⁇ is a positive number
  • the i-th column of ⁇ is ⁇ i
  • u(r+1) ⁇ u( The r+1)-dimensional identity matrix is I.
  • the realization method is to add a forgetting factor ⁇ , pruning the neural network model to obtain the control system parameter data set through the regression least squares formula of the data forgetting method.
  • the regression least squares formula using data forgetting method is as follows:
  • step 6 the method of constructing the inverse control model is:
  • ⁇ p be the control system parameter data set
  • u be the change rate of the control system parameter data
  • transfer function of the relationship between them is:
  • the cyclic delay is T c
  • the initial transmission delay is T i
  • the sensitivity of the control system parameter data is expressed as k
  • the lag time constant is ⁇
  • the cyclic coefficient is ⁇ .
  • the so-called direct inverse control method is based on the reference model of the system, the fuzzy neural network learns or approximates the inverse system model of the model, and is simply connected in series as the controller and the control object, so that the synthesized system can respond in the expected response r(t ) And the output y(t) of the controlled system form a consistent mapping.
  • the role of the reference model is to generate training sample data.
  • the pruning neural network model When pruning the neural network model as the controller in the system, in order to make the output value y(t) approach the expected value r(t), the goal is to obtain the appropriate control behavior u(t).
  • This simulation experiment requires two stages: learning stage and application stage.
  • learning stage the pruning neural network model identifies the time-varying dynamic inverse model lurking in the object.
  • application stage the pruning neural network model acts as a controller to generate control behavior.
  • the NARX model can be used to directly construct the following inverse control model:
  • the present invention also provides a D-FNN direct inverse control system based on a pruning strategy.
  • the system includes a memory, a processor, and a computer program stored in the memory and running on the processor.
  • the processor executes the computer program and runs in the unit of the following system:
  • the data set acquisition unit is used to read the input data set
  • the hierarchical learning building unit is used to adjust the dynamic fuzzy neural network model to the hierarchical learning dynamic fuzzy neural network model
  • the model pruning unit is used to trim and eliminate the fuzzy rules in the hierarchical learning dynamic fuzzy neural network model through the pruning strategy to obtain the pruned neural network model;
  • the data forgetting unit is used to forget the input data set by pruning the neural network model to obtain the control system parameter data set;
  • the inverse control model control unit is used to construct an inverse control model and input the control system parameter data set to the inverse control model for control.
  • the present invention provides a D-FNN direct inverse control method and system based on a pruning strategy: the idea of hierarchical learning can ensure a more concise structure and shorter learning time, a major advantage of hierarchical learning strategies It can alleviate the shock problem in the learning process.
  • the experimental results show that the inverse controller of D-FNN (Dynamic Fuzzy Neural Network) has powerful online learning capabilities, especially when the external environment is disturbed, D-FNN has strong adaptability and robustness. The error quickly decreases to close to zero.
  • D-FNN can automatically generate or delete fuzzy rules according to the importance of the inverse control system and the complexity of the system.
  • the model does not need to be pre-set during online learning, and it can learn from the training data set sequence adaptively, thereby compensating for Linear system modeling errors deal with external disturbances.
  • dynamic fuzzy neural network will be applied in many real-time automatic control systems.
  • Figure 1 shows a flow chart of a D-FNN direct inverse control method based on pruning strategy
  • Figure 2 shows the training phase expectation in the difference between the training phase expectation and the actual injection
  • FIG. 3 shows the actual and expected changes in blood pressure
  • Figure 4 shows the comparison of expected and actual output during the training phase of the D-FNN controller
  • Figure 5 shows the root mean square error during training
  • Figure 6 shows the output error in the training phase
  • Figure 7 shows the relationship between drug injection rate and blood pressure
  • Figure 8 shows the comparison between the actual and expected MAP blood pressure changes of the D-FNN controller
  • Figure 9 shows the expected blood pressure change
  • Figure 10 shows the drug injection rate in a noisy environment
  • Figure 11 shows a D-FNN direct inverse control system diagram based on pruning strategy.
  • Fig. 1 shows a flow chart of a D-FNN direct inverse control method based on a pruning strategy according to the present disclosure. The following describes a D-FNN direct inverse control method based on a pruning strategy according to an embodiment of the present disclosure with reference to Fig. 1 Inverse control method.
  • the present disclosure proposes a D-FNN direct inverse control method based on pruning strategy, which specifically includes the following steps:
  • Step 1 Read the input data set
  • Step 2 Construct a dynamic fuzzy neural network model
  • Step 3 Adjust the dynamic fuzzy neural network model to a hierarchical learning dynamic fuzzy neural network model
  • Step 4 Pruning and removing the fuzzy rules in the hierarchical learning dynamic fuzzy neural network model through the pruning strategy to obtain the pruning neural network model;
  • Step 5 Perform data forgetting on the input data set by pruning the neural network model to obtain the control system parameter data set;
  • Step 6 construct an inverse control model and input the control system parameter data set to the inverse control model for control.
  • step 1 the input data set is observation data acquired under natural uncontrolled conditions.
  • step 2 the method of constructing a dynamic fuzzy neural network model is:
  • x 1 ,x 2 ,L,x r are input data
  • y is output data
  • c i j is the j-th Gaussian center membership function of x i
  • ⁇ j is the Gaussian width of the j-th membership function x i
  • r is the number of input data
  • u is the number of membership functions
  • N j is the output of the jth node:
  • C i is the center of the i-th RBF unit; a i and a k are both polynomial coefficients.
  • step 3 the method of adjusting the dynamic fuzzy neural network model to the hierarchical learning dynamic fuzzy neural network model is:
  • the structure of the dynamic fuzzy neural network model is not only determined by the receiving domain but also by the system error, although the two methods can be divided into hierarchical self-organizing methods, and the output error determines whether new fuzzy rules should be added;
  • Output error is determined as described below: (X i, t i) is the i-th input data, X-i indicates the i-th input vector, T i denotes the i th desired output dynamic fuzzy neural network output y i.
  • the dynamic fuzzy neural network model adds a new fuzzy rule, and finally a hierarchical learning dynamic fuzzy neural network model is obtained.
  • the value of k e is pre-selected according to the expected accuracy of the dynamic fuzzy neural network.
  • k e is a variable:
  • the maximum error is e max
  • the accuracy of the dynamic fuzzy neural network is e min
  • the convergence constant is ⁇ (0 ⁇ 1)
  • ⁇ i is the i-th convergence constant.
  • step 4 the fuzzy rules in the hierarchical learning dynamic fuzzy neural network model are pruned and eliminated by the pruning strategy to obtain the pruned neural network model:
  • the pruning strategy uses the error reduction rate (ERR) method.
  • ERR error reduction rate
  • Different pruning strategies using ERR, SVD (singular value decomposition) or ED (eigenvalue decomposition) methods will lead to different dynamic fuzzy neural network model structures, that is, different fuzzy rules and for a specific performance, each rule The parameters are different.
  • the experiment shows that the ERR method has obvious physical meaning and stable numerical calculation, which is why this method is widely used.
  • ⁇ R (r+1)u ⁇ n and T ⁇ R n is as follows:
  • the error vector is E ⁇ R n, return amount h i, the return amount of q i. Error reduction rate
  • the threshold is set in advance as k err , the i-th fuzzy rule can be eliminated, and finally the pruned neural network model is obtained.
  • the training error without trimming is always smaller than the error after trimming, but after trimming, whether it is ERR, SVD or ED method, the system test error (judged by RMSE) is always smaller than that of the untrimmed dynamic fuzzy neural network model .
  • step 5 the method for obtaining the control system parameter data set by performing data forgetting on the input data set by pruning the neural network model is:
  • X j (x 1j ,x 2j , ⁇ ,x rj ) is the input observation data, and y i is the output data of the system, which can be written in the form of expression (16):
  • the transposition of ⁇ is ⁇ T
  • the Kalman filter algorithm is used to determine the following weights:
  • the coefficient matrix after i iterations is W i
  • the error covariance matrix of the i-th observation data is S i
  • ⁇ is a positive number
  • the i-th column of ⁇ is ⁇ i
  • u(r+1) ⁇ u( The r+1)-dimensional identity matrix is I.
  • the realization method is to add a forgetting factor ⁇ , pruning the neural network model to obtain the control system parameter data set through the regression least squares formula of the data forgetting method.
  • the regression least squares formula using data forgetting method is as follows:
  • step 6 the method of constructing the inverse control model is:
  • ⁇ p be the control system parameter data set
  • u be the change rate of the control system parameter data
  • transfer function of the relationship between them is:
  • the cyclic delay is T c
  • the initial transmission delay is T i
  • the sensitivity of the control system parameter data is expressed as k
  • the lag time constant is ⁇
  • the cyclic coefficient is ⁇ .
  • the so-called direct inverse control method is based on the reference model of the system, the fuzzy neural network learns or approximates the inverse system model of the model, and is simply connected in series as the controller and the control object, so that the synthesized system can respond in the expected response r(t ) And the output y(t) of the controlled system form a consistent mapping.
  • the role of the reference model is to generate training sample data.
  • the pruning neural network model When pruning the neural network model as the controller in the system, in order to make the output value y(t) approach the expected value r(t), the goal is to obtain the appropriate control behavior u(t).
  • This simulation experiment requires two stages: learning stage and application stage.
  • learning stage the pruning neural network model identifies the time-varying dynamic inverse model lurking in the object.
  • application stage the pruning neural network model acts as a controller to generate control behavior.
  • the NARX model can be used to directly construct the following inverse control model:
  • u(t) g[y(t+1), ⁇ ,y(tn y ),u(t-1), ⁇ ,u(tn u )](25)
  • An embodiment of the present disclosure applies the method of the present disclosure to a biomedical engineering application algorithm program to simulate a direct inverse control case of a drug injection system: the simulation results show that D-FNN has strong real-time learning and control capabilities, Advantages of simultaneous parameter estimation and structure identification
  • ⁇ p is the change in blood pressure
  • SNP sodium nitroprusside
  • u is the SNP injection rate
  • the transfer function between them is: Among them, the cyclic delay is T c , the initial transmission delay is T i , the sensitivity of sodium nitroprusside is k, the lag time constant is ⁇ , and the cyclic coefficient is ⁇ .
  • ⁇ p(k) a 0 ⁇ p(k-1)+b 0 u(kd)+b 1 u(km)+n(k),
  • D-FNN can approach a dynamic time-varying system.
  • the so-called direct inverse control method is based on the reference model of the system, the fuzzy neural network learns or approximates the inverse system model of the model, and is simply connected in series as the controller and the control object, so that the synthesized system can respond in the expected response r(t ) And the output y(t) of the controlled system form a consistent mapping.
  • the principle diagram of D-FNN as a direct inverse control method is shown in Fig. 2.
  • Fig. 2 shows the expectation of the training phase in the difference between the expected training phase and the actual injection.
  • the role of the reference model is to generate training sample data.
  • D-FNN When D-FNN is used as the controller in the system, in order to make the output value y(t) approach the desired value r(t), the goal is to obtain the appropriate control behavior u(t).
  • This simulation experiment requires two stages: learning stage and application stage.
  • D-FNN identified the time-varying dynamic inverse model lurking in the object.
  • D-FNN acts as a controller to generate control behavior.
  • the inverse model is usually derived from the NARX (autoregressive nonlinear model with out-of-band input) model through simple derivation:
  • u(t) f -1 [y(t+1),y(t), ⁇ ,y(tn y ),u(t-1), ⁇ ,u(tn u )],
  • the drug injection rate u(k) is taken as:
  • f is D-FNN, which represents the inverse mapping of formula (22).
  • Figure 2 shows the training phase expectation in the difference between the training phase expectation and the actual injection (the shape in Figure 2 "—”) and actual injection rate (shape "o" in Figure 2).
  • the sampling interval of the controller is 15s.
  • the blood pressure of the hypertensive patient is set to 140mmHg.
  • the D-FNN controller is trained, it is used to control the object.
  • the patient's blood pressure requirement is reduced to 100mmHg.
  • Figure 3 shows the actual change in blood pressure.
  • the expected change the expected blood pressure change contains white noise with a variance of 1mmHg. It can be seen that the blood pressure changes steadily with the change of sodium nitroprusside injection, and there is no time delay or oscillation, which shows that D-FNN has learned the characteristics of the model very well, and the system has been well modeled. Pruning technology is very necessary for the identification of dynamic time-varying nonlinear systems. If inactive fuzzy rules are detected and eliminated during the learning process, a more compact system structure and fast learning speed can be obtained.
  • Figure 4 shows the expected (shape "+” in Figure 4) and actual (shape "o” in Figure 4) output comparison of the D-FNN controller during the training phase
  • Figure 5 shows the root mean square error during training
  • Figure 6 shows the output error in the training phase.
  • Figure 7 shows the relationship between the drug injection rate and blood pressure.
  • Figure 8 shows the comparison between the actual and expected MAP blood pressure changes of the D-FNN controller.
  • Figure 9 shows The expected blood pressure change.
  • Figure 10 shows the drug injection rate in a noisy environment, where the perturbation selects white noise with a variance of 1mmHg.
  • Table 1 is the simulation performance comparison result between the IANA method in the literature "Indirect adaptive nonlinear control of drug delivery systems" and the D-FNN. Compared with the D-FNN method, the IANC method has an ideal simulation result ( Figure 7), and there is no large delay and oscillation even in a white noise environment with a disturbance selection variance of 1mmHg. The maximum error ⁇ p max is used to evaluate the performance of the D-FNN controller with expected and actual MAP (mean arterial blood pressure) changes. The results of the comparison between the IANA method and the D-FNN method are listed in Table 1.
  • the width w of the moving window plays an important role in identifying time-varying systems.
  • the width w is smaller, some important data will be missed, and the sample data collected by the larger width w will be more and more D-FNN gradually enters saturation and loses the adjustment function, making the time-varying parameters unrecognizable.
  • D-FNN has strong online learning capabilities, especially when the external environment is disturbed, D-FNN has strong adaptability and robustness, which can quickly reduce the trajectory error to nearly zero.
  • D-FNN can automatically generate or delete fuzzy rules according to the importance of the control system and the complexity of the system.
  • the model does not need to be pre-set during online learning, and it can learn from the training data set sequence adaptively, thereby compensating for nonlinearity System modeling errors deal with external disturbances.
  • dynamic fuzzy neural network will be applied in many real-time automatic control systems, and the rapidity of D-FNN learning still has room for further improvement in the future.
  • the embodiment of the present disclosure provides a D-FNN direct inverse control system based on a pruning strategy
  • FIG. 11 is a diagram of a D-FNN direct inverse control system based on a pruning strategy of the present disclosure.
  • a D-FNN direct inverse control system based on a pruning strategy includes: a processor, a memory, and a computer program stored in the memory and running on the processor, and the processor implements the above when the computer program is executed. Steps in an embodiment of a D-FNN direct inverse control system based on a pruning strategy.
  • the system includes a memory, a processor, and a computer program that is stored in the memory and can run on the processor, and the processor executes the computer program to run in the following system units:
  • the data set acquisition unit is used to read the input data set
  • the hierarchical learning building unit is used to adjust the dynamic fuzzy neural network model to the hierarchical learning dynamic fuzzy neural network model
  • the model pruning unit is used to trim and eliminate the fuzzy rules in the hierarchical learning dynamic fuzzy neural network model through the pruning strategy to obtain the pruned neural network model;
  • the data forgetting unit is used to forget the input data set by pruning the neural network model to obtain the control system parameter data set;
  • the inverse control model control unit is used to construct an inverse control model and input the control system parameter data set to the inverse control model for control.
  • the D-FNN direct inverse control system based on the pruning strategy can run in computing devices such as desktop computers, notebooks, palmtops, and cloud servers.
  • the operable system may include, but is not limited to, a processor and a memory.
  • the example is only an example of a D-FNN direct inverse control system based on a pruning strategy, and does not constitute a limitation on a D-FNN direct inverse control system based on a pruning strategy, and may include More or fewer components than the example, or a combination of some components, or different components.
  • the D-FNN direct inverse control system based on a pruning strategy may also include input and output devices, network access devices, and buses Wait.
  • the so-called processor can be a central processing unit (Central Processing Unit, CPU), other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), ready-made Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.
  • the processor is the control center of the D-FNN direct inverse control system operating system based on the pruning strategy. This kind of interface and line connect the whole D-FNN direct inverse control system based on pruning strategy to run the various parts of the system.
  • the memory may be used to store the computer program and/or module, and the processor implements the one by running or executing the computer program and/or module stored in the memory, and calling data stored in the memory.
  • Various functions of a D-FNN direct inverse control system based on pruning strategy.
  • the memory may mainly include a storage program area and a storage data area, where the storage program area may store an operating system, an application program required by at least one function (such as a sound playback function, an image playback function, etc.); the storage data area may store Data (such as audio data, phone book, etc.) created based on the use of mobile phones.
  • the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disks, memory, plug-in hard disks, smart media cards (SMC), and secure digital (SD) cards.
  • non-volatile memory such as hard disks, memory, plug-in hard disks, smart media cards (SMC), and secure digital (SD) cards.
  • Flash Card at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage device.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Feedback Control In General (AREA)

Abstract

La présente invention concerne un procédé de commande inverse directe de D-FNN et un système basé sur une stratégie d'élagage. Au moyen d'un apprentissage hiérarchique, une structure plus concise et un temps d'apprentissage plus court peuvent être assurés. Un avantage majeur de la stratégie d'apprentissage hiérarchique est de réduire le problème d'oscillation dans le processus d'apprentissage. Lorsque des perturbations se produisent dans un environnement externe, un D-FNN présente une forte adaptabilité et une robustesse élevée, de telle sorte que l'erreur de trajectoire peut être rapidement réduite à une valeur proche de zéro. Le D-FNN peut générer ou supprimer automatiquement des règles floues selon l'importance du système de commande inverse et la complexité du système. Le modèle n'a pas besoin d'être prédéfini pendant un apprentissage en ligne, et un ordre peut être réglé selon des données d'apprentissage pour l'apprentissage de manière adaptative, ce qui compense ainsi des erreurs de modélisation de système non linéaire pour traiter des perturbations externes. Au moyen de recherches de simulation, le réseau de neurones flou dynamique est appliqué dans de nombreux systèmes de commande automatique en temps réel.
PCT/CN2020/087798 2019-06-04 2020-04-29 Procédé de commande inverse directe de d-fnn et système basé sur une stratégie d'élagage. Ceased WO2020244346A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910482044.5A CN110361968A (zh) 2019-06-04 2019-06-04 一种基于修剪策略的d-fnn直接逆控制方法及系统
CN201910482044.5 2019-06-04

Publications (1)

Publication Number Publication Date
WO2020244346A1 true WO2020244346A1 (fr) 2020-12-10

Family

ID=68215155

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/087798 Ceased WO2020244346A1 (fr) 2019-06-04 2020-04-29 Procédé de commande inverse directe de d-fnn et système basé sur une stratégie d'élagage.

Country Status (2)

Country Link
CN (1) CN110361968A (fr)
WO (1) WO2020244346A1 (fr)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115713693A (zh) * 2022-11-24 2023-02-24 常州大学 一种基于神经网络和模糊决策的果树剪枝方法及系统
CN121091829A (zh) * 2025-11-13 2025-12-09 山东大学 基于时空特征与物理信息的轮胎炼胶过程控制方法及系统

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110361968A (zh) * 2019-06-04 2019-10-22 佛山科学技术学院 一种基于修剪策略的d-fnn直接逆控制方法及系统
WO2022261946A1 (fr) * 2021-06-18 2022-12-22 舍弗勒技术股份两合公司 Procédé et dispositif de commande de démarrage de moteur pour véhicule électrique hybride

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101452258A (zh) * 2007-12-06 2009-06-10 西安电子科技大学 一种与模型无关的自适应控制器及其控制方法
US20110125685A1 (en) * 2009-11-24 2011-05-26 Rizvi Syed Z Method for identifying Hammerstein models
CN103034122A (zh) * 2012-11-28 2013-04-10 上海交通大学 基于时间序列的多模型自适应控制器及控制方法
WO2016010601A2 (fr) * 2014-04-23 2016-01-21 The Florida State University Research Foundation, Inc. Commande prédictive de modèle non linéaire adaptative au moyen d'un réseau neuronal et d'un échantillonnage d'entrée
CN110361968A (zh) * 2019-06-04 2019-10-22 佛山科学技术学院 一种基于修剪策略的d-fnn直接逆控制方法及系统

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101452258A (zh) * 2007-12-06 2009-06-10 西安电子科技大学 一种与模型无关的自适应控制器及其控制方法
US20110125685A1 (en) * 2009-11-24 2011-05-26 Rizvi Syed Z Method for identifying Hammerstein models
CN103034122A (zh) * 2012-11-28 2013-04-10 上海交通大学 基于时间序列的多模型自适应控制器及控制方法
WO2016010601A2 (fr) * 2014-04-23 2016-01-21 The Florida State University Research Foundation, Inc. Commande prédictive de modèle non linéaire adaptative au moyen d'un réseau neuronal et d'un échantillonnage d'entrée
CN110361968A (zh) * 2019-06-04 2019-10-22 佛山科学技术学院 一种基于修剪策略的d-fnn直接逆控制方法及系统

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
CAI-XIA ZHANG, LIU GUO-WEN: "Research on D-FNN Direct Inverse Control Algorithm Based on Pruning Strategy", ACTA AUTOMATICA SINICA, vol. 45, no. 8, 1 August 2019 (2019-08-01), pages 1599 - 1605, XP055764715, DOI: 10.16383/j.aas.c190038 *
MA LI, DE-FENG ZHANG, YONG XU: "A Bio-Engineering Algorithm Research Based on Dynamic F uzzy Neural Netw orks", COMPUTER ENGINEERING AND SCIENCE, GUOFANG KEJI DAXUE JISUANJI XUEYUAN, CN, vol. 32, no. 3, 1 March 2010 (2010-03-01), CN, pages 137 - 140, XP055765585, ISSN: 1007-130X *
SHIQIAN WU ET AL.: "Dynamic Fuzzy Neural Networks—A Novel Approach to Function Approximation", IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, vol. 30, no. 2,, 30 April 2000 (2000-04-30), XP011056881, ISSN: 1083-4419, DOI: 20200518175402Y *
ZHANG DEFENG, ZILONG MA: "SVD-QR D -FNN Research on D-FNN Algorithm with the Combination of Column Pivot SVD-QR Method Pruning Strategy and Parameters Adjustment", ACTA SCIENTIARUM NATURALIUM UNIVERSITATIS SUNYATSENI, vol. 52, no. 2, 1 March 2013 (2013-03-01), pages 8 - 13, XP055765597, DOI: 10.13471/j.cnki.acta.snus.2013.02.002 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115713693A (zh) * 2022-11-24 2023-02-24 常州大学 一种基于神经网络和模糊决策的果树剪枝方法及系统
CN121091829A (zh) * 2025-11-13 2025-12-09 山东大学 基于时空特征与物理信息的轮胎炼胶过程控制方法及系统

Also Published As

Publication number Publication date
CN110361968A (zh) 2019-10-22

Similar Documents

Publication Publication Date Title
WO2020244346A1 (fr) Procédé de commande inverse directe de d-fnn et système basé sur une stratégie d'élagage.
Heess et al. Actor-critic reinforcement learning with energy-based policies
Zhang et al. Dynamic extreme learning machine and its approximation capability
Nelles Axes-oblique partitioning strategies for local model networks
Chen et al. Online modeling with tunable RBF network
Sun et al. Learning to filter with predictive state inference machines
Qi et al. Stability of delayed memristive neural networks with time-varying impulses
Zanger Convergence of a least‐squares Monte Carlo algorithm for American option pricing with dependent sample data
CN118152894B (zh) 一种电力系统宽频振荡识别方法、装置、终端及介质
Silva et al. Evolving neo-fuzzy neural network with adaptive feature selection
van de Laar et al. Application of the free energy principle to estimation and control
CN108958037B (zh) 小波模糊大脑情感学习控制方法、装置、设备及存储介质
Meyer-Base et al. Local and global stability analysis of an unsupervised competitive neural network
Allouche et al. ExceedGAN: simulation above extreme thresholds using Generative Adversarial Networks
Hung Robust Kalman filter based on a fuzzy GARCH model to forecast volatility using particle swarm optimization
Zhang et al. Generalized maximum correntropy-based echo state network for robust nonlinear system identification
CN104933052B (zh) 数据真实值的估计方法及数据真实值估计装置
Sun et al. Federated continual learning based on weakly supervised diffusion models for disease diagnosis
Shapiai et al. A non-linear function approximation from small samples based on Nadaraya-Watson kernel regression
CN117493956A (zh) 跨数据库的脑电情绪识别方法、装置、设备及存储介质
CN115034371A (zh) 一种基于偏见神经元的数据去偏方法和装置
Wei Boosting wavelet neural networks using evolutionary algorithms for short-term wind speed time series forecasting
Rodrigues et al. A bounded real lemma for continuous-time linear systems with partial information on the Markovian jumping parameters
CN113920365A (zh) 多源时间序列的分类方法、装置、设备和存储介质
Ab Malek et al. Evolutionary tuning method for PID controller parameters of a cruise control system using metamodeling

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20818058

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20818058

Country of ref document: EP

Kind code of ref document: A1