CN108873695A - The full Connection Neural Network control system of multilayer - Google Patents
The full Connection Neural Network control system of multilayer Download PDFInfo
- Publication number
- CN108873695A CN108873695A CN201810554501.2A CN201810554501A CN108873695A CN 108873695 A CN108873695 A CN 108873695A CN 201810554501 A CN201810554501 A CN 201810554501A CN 108873695 A CN108873695 A CN 108873695A
- Authority
- CN
- China
- Prior art keywords
- neural network
- multilayer
- control system
- full connection
- output
- 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
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
- G05B13/04—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)
- 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
The invention discloses a kind of using the full Connection Neural Network of multilayer as the control system of primary control, which has good adaptability to multiple-input and multiple-output control.System realization has the following steps：(1) ideal input-output curve is determined.(2) sampled point is taken on ideal input-output curve.(3) multilayer fully-connected network is established, and trains neural network with sampled point.(4) trained network is added to controller.(5) input-output curve of observation-well network, if want to meet with ideal input-output curve.Through the above way, establish the full Connection Neural Network control system of multilayer, by the way that neural network is combined with control system, the property of arbitrary curve can be fitted using multilayer neural network, the Complex Response curve in the scene of multiple-input and multiple-output is simulated, has good adaptability to the complication system of multiple-input and multiple-output.
Description
Technical field
The invention belongs to control engineering and computer field more particularly to artificial intelligence neural networks control system flow chart
With implementation method and device.
Background technique
Neural network is widely used according to its powerful study and adaptability in the current situation, especially more
Layer neural network is quickly grown at artificial intelligence, image recognition, signal processing aspect, in recent years classical control theory and modern control
It makes and is slightly having deficiency in face of increasingly complicated control occasion, therefore gone to realize that control is to be worth research with neural network.
The property the most outstanding of one of neural network can be achieved on the function of any function, even for only one
This conclusion of the neural network of hidden layer is still set up, and is not that can accurately calculate the value of original function completely, but lead to
Cross increase hidden layer and hidden layer neuron, can increasingly approximating function, i.e., the function for needing to realize for one,
It is required that realizing that precision is, that is, the output for needing enough hidden layer neurons to make neural networkMeet, for allAll meet；And traditional control system is substantially the physics realization of transmission function, institute
To realize that control system is feasible with the full Connection Neural Network of multilayer.
Summary of the invention
The object of the present invention is to provide a kind of using neural network as the Control System Design mode of primary control and control
Device solves the problems, such as that multi-input multi-output control system design and implementation is difficult.
Above-mentioned target is realized by the following technical solution：
（1）Obtain the ideal control curve
（2）Controller network training
（3）Control flow design, control system building
（4）Whether observation control response is consistent or meets required precision with ideal response.
Detailed description of the invention
Fig. 1 System control structures figure
The basic block diagram of Fig. 2 nerve network controller
Fig. 3 network controller trains flow chart
Beneficial effect
The present invention, using the full Connection Neural Network of multilayer as sole controller, realizes more using neural network as primary control
The design for inputting multi output complication system is simplified, and designer does not need the relationship between attention location system internal signal, only needs
To be reached according to the ideal input/output signal Training Control neural network of system by adjusting the weight and deviation of neural network
The purpose of control system is designed, system not only can satisfy control and require, and can also be the extension and upgrading of subsequent control system
It is convenient to provide, it is only necessary to which input layer and output layer at control network increase neuron, and obtain ideal response data from new instruction
Practice control neural network, the raising for controlling precision also only needs to increase the number of plies of hidden layer or increases hidden layer
Neuron number can be realized.
Specific embodiment
Steps are as follows for the full Connection Neural Network Control System Design of multilayer：
(1) it requires to draw according to the control of controlled device and determines ideal input and output response signal, can be continuous multidimensional and ring
The multidimensional response vector for answering space either discrete, input signal is multi-C vector, output signal is more
Dimensional vector, ideal response vector corresponding to above-mentioned input and output obtain as far as possible multiple groups as training sample
This, sample, which trains more greatly the nerve network controller output come, under normal circumstances can more accurately track input signal；
(2) determine control system process, according to some signals amplifications of the addition Bu Tong appropriate of controlled object and input signal or
The signal energy of person's signal driving device, such as output is smaller, and controlled device needs more energy to drive, then needs to increase signal
Amplifying device；
(3) it is required according to control, builds nerve network controller, including input layer and output layer neuron number, hide the number of plies
With hidden layer neuron number and activation primitive, wherein input layer and output layer neuron number are by input signal and output
Signal number determines, hides the number of plies and hidden layer neuron number and is determined by control precision, and the selection of activation primitive need according to
The signal type needed according to controlled device determines generally just like minor function：
(1)
(2)
(3)
If output signal is that analog quantity generally uses the activation primitive in formula (2), if output signal only has positive value and in section,
Then with the activation primitive in formula (1), the activation primitive in formula 3 is otherwise used.Activation primitive is not unique, but has to continuously, no
Then the retrospectively calculate in network training cannot achieve；It is certain anti-interference in order to there is nerve network controller to external disturbance
Ability, neural network input layer should increase one or more disturbing signal, range of the training sample value in disturbance in input layer
It is interior to take at random；
(4) training method of neural metwork training, controller neural metwork training and general networking is essentially identical, is divided into forward direction
Calculating and retrospectively calculate；First to sampled value data prediction according to the following steps：One, eliminate mean value；Two, decorrelation；Three, association
Variance is balanced；Secondly netinit to the weight of network and is biased to initialize, when network weight be endowed one it is larger or compared with
When small initial value, the neuron of network is likely to tend to be saturated, and causes retrospectively calculate local gradient that a very little is presented
Value, keeps network learning procedure slack-off, therefore weight initialization will make its mean value be 0, the linking number of variance and neuron at
Inverse ratio；Followed by learning rate be arranged, general learning rate should not be arranged it is excessive or too small, if network may be made not restrain greatly very much,
Training can not be normally carried out；If the too small slack-off training of convergence rate needs the long period, therefore learning rate：FromDrop to
Linear Annealing；On-line study is finally used, training, reason is that on-line study easily performs, and saves resource and is not easy to fall into part
Extreme point, repetitive exercise, stopping criterion are when the Euclidean Norm of gradient vector reaches a sufficiently small Grads threshold, it is believed that
Retrospectively calculate has restrained；The setting of threshold value is related with the control precision of controlled device, and precision is higher, and threshold value setting can be appropriate
Reduce, vice versa；
(5) network will be trained to be added in control system, and will test whether that meeting control will require, if being unsatisfactory for continuing to train；
Stop if meeting.
Claims (8)
1. the full Connection Neural Network control system of multilayer, it is characterized in that:
（1）Master controller is only made of the full Connection Neural Network of multilayer
（2）Feedback channel can not be nerve network controller
（3）The acquisition modes of ideal input-output curve
（4）The acquisition modes of sampled point
（5）Determine that neural network inputs neuron number according to the number of input signal
（6）Neural network output neuron number is determined according to the number of output signal
（7）The number of plies of neural network is determined according to the requirement of control precision
（8）The training full Connection Neural Network controller of multilayer
（9）Multilayer fully-connected network control system is established, controlled device is controlled.
2. the full Connection Neural Network control system of multilayer according to claim 1, feature（1）：With other control system
It compares, the not traditional PID controller of this system, the primary control of entire control system only has a multilayer neural network,
Functional relation between outputting and inputting all is expressed by multilayer neural network.
3. the full Connection Neural Network control system of multilayer according to claim 1, feature（2）：The feedback of this system is logical
Road includes neural network, but is not limited only to neural network.
4. the full Connection Neural Network control system of multilayer according to claim 1, feature（3）：The ideal of this system is defeated
Entering curve of output is by as under type obtains, if being easy to decouple between input and output signal, ideal input and output are bent
Line is after input signal decouples, and control variable obtains；If input signal is not easy to decouple, need to construct multidimensional signal sky
Between, output signal is decoupled to obtain.
5. the full Connection Neural Network control system of multilayer according to claim 1, feature（4）：Sampled point acquisition be
On ideal input-output space, simple random sampling, but sample wants sufficiently large.
6. the full Connection Neural Network control system of multilayer according to claim 1, feature（5）,（6）：This system master control
The multilayer fully-connected network of device processed output and input layer neuron number only it is related with system input and output signal, with other nothings
It closes.
7. the full Connection Neural Network control system of multilayer according to claim 1, feature（7）：Due to neural network pair
It is a kind of approximate fits that ideal, which outputs and inputs curve, always there are some errors, but in the total different controlled device of engineering
To error requirements difference, for the convenience of network training, under the premise of meeting required precision, the network number of plies is reduced as far as possible；
4 layers of neural network can fit all functions of modern control theory description under normal circumstances, therefore will be according to control
Precision determines the network number of plies.
8. the full Connection Neural Network control system of multilayer according to claim 1, feature（8）：The training of neural network
It is divided into forward calculation and retrospectively calculate two parts.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810554501.2A CN108873695A (en) | 2018-06-01 | 2018-06-01 | The full Connection Neural Network control system of multilayer |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810554501.2A CN108873695A (en) | 2018-06-01 | 2018-06-01 | The full Connection Neural Network control system of multilayer |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108873695A true CN108873695A (en) | 2018-11-23 |
Family
ID=64335898
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810554501.2A Pending CN108873695A (en) | 2018-06-01 | 2018-06-01 | The full Connection Neural Network control system of multilayer |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108873695A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109902741A (en) * | 2019-02-28 | 2019-06-18 | 上海理工大学 | A kind of breakdown of refrigeration system diagnostic method |
| CN109948778A (en) * | 2019-02-28 | 2019-06-28 | 上海理工大学 | A kind of refrigeration equipment trouble-shooter and system |
-
2018
- 2018-06-01 CN CN201810554501.2A patent/CN108873695A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109902741A (en) * | 2019-02-28 | 2019-06-18 | 上海理工大学 | A kind of breakdown of refrigeration system diagnostic method |
| CN109948778A (en) * | 2019-02-28 | 2019-06-28 | 上海理工大学 | A kind of refrigeration equipment trouble-shooter and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106774379B (en) | Intelligent supercoiled strong robust attitude control method | |
| Han et al. | Adaptive neural control for a class of switched nonlinear systems | |
| CN110989639B (en) | Underwater vehicle formation control method based on stress matrix | |
| CN107272403A (en) | A kind of PID controller parameter setting algorithm based on improvement particle cluster algorithm | |
| Gibson et al. | Improved transient response in adaptive control using projection algorithms and closed loop reference models | |
| WO2023019601A1 (en) | Signal modulation recognition method for complex-valued neural network based on structure optimization algorithm | |
| CN107766936A (en) | Artificial neural networks, artificial neuron and the control method of artificial neuron | |
| CN108873695A (en) | The full Connection Neural Network control system of multilayer | |
| US11494641B2 (en) | Reinforcement learning for human robot interaction | |
| Wei et al. | Adaptive neural dynamic surface control of MIMO uncertain nonlinear systems with time-varying full state constraints and disturbances | |
| CN110426951A (en) | A kind of robust distribution average tracking control method applied to swarm intelligence system | |
| Naoual et al. | Fuzzy sliding mode control for the two-link robot | |
| CN116560475A (en) | Server fan control method and computer equipment | |
| Zhang et al. | ANFIS-based course-keeping control for ships using nonlinear feedback technique | |
| Zhu et al. | Adaptive fixed-time tracking control for a class of nonlinear pure-feedback systems: A relative threshold event-triggered strategy | |
| Guo et al. | Active adaptive continuous nonsingular terminal sliding mode controller for hypersonic vehicle | |
| CN104991446A (en) | Unmanned plane thrust direction-changing intelligent control method based on brain emotion learning | |
| Hwang et al. | Internet-based smart-space navigation of a car-like wheeled robot using fuzzy-neural adaptive control | |
| JPH09237260A (en) | Device for learning multi-layer neural network | |
| Yu et al. | Bionic swarm control based on second-order communication topology | |
| CN113919217B (en) | Adaptive parameter setting method and device for active disturbance rejection controller | |
| CN114666729A (en) | Underwater acoustic communication error compensation method for cooperative navigation and positioning of underwater cluster UUV | |
| CN110928185A (en) | Quantitative control method of multi-agent system | |
| CN110568756A (en) | Adaboost prediction robust control-based control method for continuous rotary motor electro-hydraulic servo system | |
| CN109799765A (en) | Steering engine electrohydraulic servo system intelligent control method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181123 |