CN107515530A - A kind of Nonlinear Control Allocation method based on depth autoencoder network - Google Patents

A kind of Nonlinear Control Allocation method based on depth autoencoder network Download PDF

Info

Publication number
CN107515530A
CN107515530A CN201710566663.3A CN201710566663A CN107515530A CN 107515530 A CN107515530 A CN 107515530A CN 201710566663 A CN201710566663 A CN 201710566663A CN 107515530 A CN107515530 A CN 107515530A
Authority
CN
China
Prior art keywords
nonlinear
rudder
neutral net
training
method based
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.)
Pending
Application number
CN201710566663.3A
Other languages
Chinese (zh)
Inventor
黄煌
何英姿
黄盘兴
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.)
Beijing Institute of Control Engineering
Original Assignee
Beijing Institute of Control Engineering
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 Beijing Institute of Control Engineering filed Critical Beijing Institute of Control Engineering
Priority to CN201710566663.3A priority Critical patent/CN107515530A/en
Publication of CN107515530A publication Critical patent/CN107515530A/en
Pending legal-status Critical Current

Links

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
    • G05B13/0265—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion
    • G05B13/027—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion using neural networks only
    • 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

Landscapes

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

Abstract

A kind of Nonlinear Control Allocation method based on depth autoencoder network, including：(1) structure and Training Multilayer Feedforward Neural Networks, realize that rudder is biased to the fitting of the nonlinear mapping function of torque；(2) own coding neutral net is built；(3) unsupervised training is carried out to the encoder in the own coding neutral net, obtains torque to the inclined nonlinear model of rudder；(4) nonlinear model that step (3) obtains is embedded into attitude control system loop, it is expected input of the control moment as nonlinear model, the output of nonlinear model is calculated in real time, that is, the inclined desired value of rudder is obtained, so as to realize Nonlinear Control Allocation.

Description

A kind of Nonlinear Control Allocation method based on depth autoencoder network
Technical field
The present invention relates to a kind of Nonlinear Control Allocation method based on depth autoencoder network, it is possible to achieve quick high-precision Degree control distribution, is mainly used on Modern New hypersonic aircraft.
Background technology
Reached to realize that the large span whole world is quick, Modern New hypersonic aircraft shows the fusion of wing height degree Feature, while in order to obtain sufficiently large motor-driven overload, the process of reentering generally requires big angles-of-attack.In this context, The mapping that aircraft rudder is biased to torque shows nonlinearity feature, and original linearisation is assumed no longer to be applicable, corresponding line Propertyization control distribution method such as pseudoinverse technique, linear quadratic law of planning can not all handle non-linear object.
The content of the invention
The technology of the present invention solves problem：Overcome the deficiencies in the prior art, there is provided one kind is based on depth own coding net The Nonlinear Control Allocation method of network, realize the non-linear control of Modern New hypersonic aircraft with the fusion of wing height degree System distribution, has rapidity and high-precision feature, while can handle the situation of higher-dimension redundancy rudder face.
The present invention technical solution be：
A kind of Nonlinear Control Allocation method based on depth autoencoder network, step are as follows：
(1) structure and Training Multilayer Feedforward Neural Networks, realize that rudder is biased to the fitting of the nonlinear mapping function of torque；
(2) own coding neutral net is built；
(3) unsupervised training is carried out to the encoder in the own coding neutral net, obtains torque to rudder partially non-thread Property model；
(4) nonlinear model that step (3) obtains is embedded into attitude control system loop, it is expected controling power Input of the square as nonlinear model, the output of nonlinear model is calculated in real time, that is, obtain the inclined desired value of rudder, so as to real Existing Nonlinear Control Allocation.
The training data of the multilayer feedforward neural network derives from the aerodynamic data of hypersonic aircraft.
Multilayer feedforward neural network is initialized by greedy successively pre-training algorithm.
The decoder of own coding neutral net is multilayer feedforward neural network known to the weights trained in step (1), The encoder of own coding neutral net is weights depth feedforward neural network undetermined.
The encoder output layer neuronal quantity of own coding neutral net is consistent with the rudder face quantity of hypersonic aircraft.
The step (3) carries out unsupervised training to the encoder in own coding neutral net, obtains torque to rudder partially Nonlinear model, it is specially：
Evaluation function when being trained according to optimization aim setting network, carries out the training of autoencoder network, in training process The weights of decoder keep constant, and only the weights of encoder are trained, obtain torque to the inclined nonlinear model of rudder；
The evaluation function is specially：
J (θ)=L (Mdes,g(δdes))+λ||δdes||1；
Wherein, L (Mdes,g(δdes)) it is the error function that own coding neutral net is inputted between output, | | δdes||1 For all rudder deviator absolute value sums, λ is preset weights parameter.
Compared with the prior art, the invention has the advantages that：
(1) the inventive method dexterously make use of the coding of autoencoder network-decode this special construction, will control point Regard the high-level characteristic extraction procedure to it is expected torque as with problem, excavate autoencoder network with controlling the consistent of assignment problem Property, realize the Nonlinear Control Allocation problem solving independent of training sample.
(2) the inventive method utilizes encoder and decoder of the deep learning method structure with multiple hidden layers, it is ensured that The rapidity and accuracy of network training.One aspect of the present invention compensate for existing Nonlinear Control Allocation method in species Serious shortcoming, on the other hand can substitute existing Linear Control distribution method, be provided for redundancy executing agency a kind of quick, high Precision, means are easily distributed, and types of objects can be widely used in.
Brief description of the drawings
Fig. 1 is the inventive method flow chart；
Fig. 2 is autoencoder network structure chart；
Fig. 3 is gesture stability circuit system block diagram；
Fig. 4 is that depth own coding ANN Control distributes result of implementation；Wherein, Fig. 4 (a) is that own coding neutral net obtains The actual torque arrived and the comparing result for it is expected torque, Fig. 4 (b) are the convergence result of autoencoder network training error.
Embodiment
The embodiment of the present invention is further described in detail below in conjunction with the accompanying drawings.
As shown in figure 1, the present invention proposes a kind of Nonlinear Control Allocation method based on depth autoencoder network, step It is as follows：
(1) structure and Training Multilayer Feedforward Neural Networks, realize that rudder is biased to the fitting of the nonlinear mapping function of torque；Institute The training data for stating multilayer feedforward neural network derives from the aerodynamic data of hypersonic aircraft.The input of training data is height The working condition of supersonic aircraft, including attitude angle, attitude angular velocity, flying height, flying speed and angle of rudder reflection, train number According to output be the actual torque obtained of hypersonic aircraft.Training data will cover the whole of hypersonic aircraft when choosing Individual flight envelope.In the training process of multilayer feedforward neural network, using greedy successively pre-training algorithm, to improve network Convergence rate and fitting precision.
(2) own coding neutral net is built；The decoder of own coding neutral net is that the weights that train have been in step (1) The multilayer feedforward neural network known, the encoder of own coding neutral net is weights depth feedforward neural network undetermined.It is self-editing The input of the encoder of code neutral net is desired control moment, and the output of the encoder of own coding neutral net is to be solved Rudder it is inclined, therefore the output layer neuronal quantity of encoder and the rudder face quantity of hypersonic aircraft should be consistent.
(3) unsupervised training is carried out to the encoder in the own coding neutral net, obtains torque to rudder partially non-thread Property model；Specially：As shown in Fig. 2 because the input and output of own coding neutral net are of equal value, thus choose it is some can Input and output of the capable expectation torque as own coding neutral net, evaluation letter when being trained according to optimization aim setting network Count, the training of progress own coding neutral net, the weights of decoder keep constant in training process, and only the weights of encoder are entered Row training, obtains torque to the inclined nonlinear model of rudder；
The evaluation function is specially：
J (θ)=L (Mdes,g(δdes))+λ||δdes||1；
Wherein, L (Mdes,g(δdes)) it is the error function that own coding neutral net is inputted between output, | | δdes||1 For all rudder deviator absolute value sums, λ is preset weights parameter.
(4) nonlinear model that step (3) obtains is embedded into attitude control system loop, as shown in figure 3, with Input of the desired control torque as nonlinear model, the output of nonlinear model is calculated in real time, that is, obtain the rudder inclined phase Prestige value, so as to realize Nonlinear Control Allocation.
Own coding neutral net and the hard-wired feasibility of deep learning method and real-time, in machine learning field Obtained extensive checking.Therefore, the technical method that this invention is proposed, there is reliable hardware fulfillment capability.Meanwhile offline The pattern of unsupervised training, dependence of the method to training sample is greatly reduced, be the practicality of this invention, there is provided strong Ensure.
Embodiment：
Consider a kind of hypersonic aircraft containing six control rudder faces.Wherein, a pair of ailerons are substantially carried out rolling control System, a pair of V tails are used to break away and pitch control, an individual wing flap are used for pitch control, are also equipped with a flap in addition.Although Six control rudder faces have been carried out with division functionally, still, obvious coupling each other be present.
Step (1) structure multilayer feedforward neural network is first according to, symbiosis is into 90000 groups of training samples, and 20000 groups Random test sample.The hidden layer activation primitive of network chooses conventional sigmoid functions, and output layer activation primitive is chosen for linearly Function.Network is initialized using the greedy successively pre-training method in deep learning.
Secondly, using foregoing multilayer feedforward neural network as decoder, construction depth own coding neutral net.Wherein, compile The input of code device is the axle desired control torque of hypersonic aircraft three and the angle of attack, yaw angle and flight Mach number, exports and is The rudder of six rudder faces is inclined, that is, controls the solution of assignment problem.In first hidden layer of decoder, the first six neuron is decoder 1 couple 1 of input layer replicates, and rear three neurons represent the angle of attack, yaw angle and flight Mach number respectively.
Finally, in selecting step (3) evaluation function preset weights parameter lambda=0.2, own coding neutral net is instructed Practice, obtain torque to the inclined nonlinear model of rudder.
Exemplified by being followed successively by 50,30,10 with three node in hidden layer of encoder, Fig. 4 (a), 4 (b) are given random In 100 test samples chosen, the comparison diagram of actual torque and expectation torque that control distribution obtains.It can be seen by Fig. 4 (a) Go out, the actual torque shown in dotted line and expectation torque shown in solid are basically identical, and distribution error is very small.Can by Fig. 4 (b) Know, autoencoder network can Fast Convergent, illustrate that network training is effective.
The content not being described in detail in description of the invention belongs to the known technology of professional and technical personnel in the field.

Claims (7)

  1. A kind of 1. Nonlinear Control Allocation method based on depth autoencoder network, it is characterised in that step is as follows：
    (1) structure and Training Multilayer Feedforward Neural Networks, realize that rudder is biased to the fitting of the nonlinear mapping function of torque；
    (2) own coding neutral net is built；
    (3) unsupervised training is carried out to the encoder in the own coding neutral net, obtains torque to the inclined nonlinear model of rudder Type；
    (4) nonlinear model that step (3) obtains is embedded into attitude control system loop, it is expected that control moment is made For the input of nonlinear model, the output of nonlinear model is calculated in real time, that is, obtains the inclined desired value of rudder, it is non-so as to realize Linear Control is distributed.
  2. 2. a kind of Nonlinear Control Allocation method based on depth autoencoder network according to claim 1, its feature exist In：The training data of the multilayer feedforward neural network derives from the aerodynamic data of hypersonic aircraft.
  3. 3. a kind of Nonlinear Control Allocation method based on depth autoencoder network according to claim 1, its feature exist In：Multilayer feedforward neural network is initialized by greedy successively pre-training algorithm.
  4. 4. a kind of Nonlinear Control Allocation method based on depth autoencoder network according to claim 1, its feature exist In：The decoder of own coding neutral net is multilayer feedforward neural network, own coding known to the weights that train in step (1) The encoder of neutral net is weights depth feedforward neural network undetermined.
  5. 5. a kind of Nonlinear Control Allocation method based on depth autoencoder network according to claim 4, its feature exist In：The encoder output layer neuronal quantity of own coding neutral net is consistent with the rudder face quantity of hypersonic aircraft.
  6. 6. a kind of Nonlinear Control Allocation method based on depth autoencoder network according to claim 1, its feature exist In：The step (3) carries out unsupervised training to the encoder in own coding neutral net, obtains torque to rudder partially non-linear Model, it is specially：
    Evaluation function when being trained according to optimization aim setting network, the training of autoencoder network is carried out, is decoded in training process The weights of device keep constant, and only the weights of encoder are trained, obtain torque to the inclined nonlinear model of rudder；
  7. 7. a kind of Nonlinear Control Allocation method based on depth autoencoder network according to claim 6, its feature exist In：The evaluation function is specially：
    J (θ)=L (Mdes,g(δdes))+λ||δdes||1；
    Wherein, L (Mdes,g(δdes)) it is the error function that own coding neutral net is inputted between output, | | δdes||1It is all Rudder deviator absolute value sum, λ are preset weights parameter.
CN201710566663.3A 2017-07-12 2017-07-12 A kind of Nonlinear Control Allocation method based on depth autoencoder network Pending CN107515530A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710566663.3A CN107515530A (en) 2017-07-12 2017-07-12 A kind of Nonlinear Control Allocation method based on depth autoencoder network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710566663.3A CN107515530A (en) 2017-07-12 2017-07-12 A kind of Nonlinear Control Allocation method based on depth autoencoder network

Publications (1)

Publication Number Publication Date
CN107515530A true CN107515530A (en) 2017-12-26

Family

ID=60722232

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710566663.3A Pending CN107515530A (en) 2017-07-12 2017-07-12 A kind of Nonlinear Control Allocation method based on depth autoencoder network

Country Status (1)

Country Link
CN (1) CN107515530A (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109500837A (en) * 2018-12-18 2019-03-22 上海岭先机器人科技股份有限公司 A kind of joint of robot torgue measurement method based on Dual-encoder
CN109858523A (en) * 2019-01-03 2019-06-07 武汉大学 A kind of shallow sea velocity of sound profile inversion method of neural network and ray theory
CN110567558A (en) * 2019-08-28 2019-12-13 华南理工大学 Ultrasonic guided wave detection method based on deep convolution characteristics
CN111639749A (en) * 2020-05-25 2020-09-08 上海智殷自动化科技有限公司 Industrial robot friction force identification method based on deep learning
CN112180979A (en) * 2020-10-14 2021-01-05 上海航天控制技术研究所 Linear cooperative control method for speed comprehensive redundancy rudder system
CN112987695A (en) * 2021-03-12 2021-06-18 北京航天自动控制研究所 Aircraft health prediction method

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101937233B (en) * 2010-08-10 2012-05-30 南京航空航天大学 Nonlinear self-adaption control method of near-space hypersonic vehicle
CN106773691A (en) * 2016-12-19 2017-05-31 西北工业大学 Hypersonic aircraft self adaptation time-varying default capabilities control method based on LS SVM
CN106896722A (en) * 2017-03-29 2017-06-27 郑州轻工业学院 Adoption status feeds back the hypersonic vehicle composite control method with neutral net

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101937233B (en) * 2010-08-10 2012-05-30 南京航空航天大学 Nonlinear self-adaption control method of near-space hypersonic vehicle
CN106773691A (en) * 2016-12-19 2017-05-31 西北工业大学 Hypersonic aircraft self adaptation time-varying default capabilities control method based on LS SVM
CN106896722A (en) * 2017-03-29 2017-06-27 郑州轻工业学院 Adoption status feeds back the hypersonic vehicle composite control method with neutral net

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
黄煌 等: "一种基于深度神经网络的非线性控制分配方法", 《战术导弹技术》 *

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109500837A (en) * 2018-12-18 2019-03-22 上海岭先机器人科技股份有限公司 A kind of joint of robot torgue measurement method based on Dual-encoder
CN109858523A (en) * 2019-01-03 2019-06-07 武汉大学 A kind of shallow sea velocity of sound profile inversion method of neural network and ray theory
CN109858523B (en) * 2019-01-03 2022-06-14 武汉大学 Shallow sea sound velocity profile inversion method based on neural network and ray theory
CN110567558A (en) * 2019-08-28 2019-12-13 华南理工大学 Ultrasonic guided wave detection method based on deep convolution characteristics
CN110567558B (en) * 2019-08-28 2021-08-10 华南理工大学 Ultrasonic guided wave detection method based on deep convolution characteristics
CN111639749A (en) * 2020-05-25 2020-09-08 上海智殷自动化科技有限公司 Industrial robot friction force identification method based on deep learning
CN112180979A (en) * 2020-10-14 2021-01-05 上海航天控制技术研究所 Linear cooperative control method for speed comprehensive redundancy rudder system
CN112987695A (en) * 2021-03-12 2021-06-18 北京航天自动控制研究所 Aircraft health prediction method

Similar Documents

Publication Publication Date Title
CN107515530A (en) A kind of Nonlinear Control Allocation method based on depth autoencoder network
CN106997208B (en) A kind of control method towards the hypersonic aircraft under condition of uncertainty
CN109343341A (en) It is a kind of based on deeply study carrier rocket vertically recycle intelligent control method
CN108445766A (en) Model-free quadrotor drone contrail tracker and method based on RPD-SMC and RISE
CN103412488B (en) A kind of miniature self-service gyroplane high-accuracy control method based on adaptive neural network
CN106877746B (en) Method for control speed and speed control unit
CN107422741B (en) Learning-based cluster flight distributed attitude tracking control method for preserving preset performance
CN109597303A (en) A kind of composite rotor craft syntype flight control method
CN104765312B (en) Implementation method for reconfigurable aircraft control system
CN107633105B (en) Improved hybrid frog-leaping algorithm-based quad-rotor unmanned aerial vehicle parameter identification method
CN104217041A (en) Multi-constrained online Gauss pseudo spectrum type terminal guidance method
CN104880945B (en) The adaptive inverse control of rotor wing unmanned aerial vehicle based on neutral net
CN114637312B (en) Unmanned aerial vehicle energy-saving flight control method and system based on intelligent deformation decision
Dong et al. Trial input method and own-aircraft state prediction in autonomous air combat
Yu et al. Optimal design of helicopter control systems using particle swarm optimization
Khizer et al. Stable hovering flight for a small unmanned helicopter using fuzzy control
Abdulla et al. Roll control system design using auto tuning LQR technique
CN106168760A (en) Uncertain time-delayed systems determination of stability method based on convex polyhedron fault model
Omkar et al. Identification of helicopter dynamics based on flight data using nature inspired techniques
CN115407294B (en) Aircraft dynamic radar scattering cross section simulation method for double-station radar system
CN114489095B (en) Brain-like pulse neural network control method applied to variant aircraft
CN114815878B (en) Hypersonic aircraft collaborative guidance method based on real-time optimization and deep learning
CN114442673B (en) Intelligent aircraft trajectory planning method based on chaotic model and DNN
CN115796021A (en) Adaptive PID control parameter setting method for variant aircraft based on neural network
CN112698666B (en) Aircraft route optimization method based on meteorological grid

Legal Events

Date Code Title Description
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20171226

RJ01 Rejection of invention patent application after publication