CN109366485B - On-site control method for on-line machine learning - Google Patents
On-site control method for on-line machine learning Download PDFInfo
- Publication number
- CN109366485B CN109366485B CN201811027921.1A CN201811027921A CN109366485B CN 109366485 B CN109366485 B CN 109366485B CN 201811027921 A CN201811027921 A CN 201811027921A CN 109366485 B CN109366485 B CN 109366485B
- Authority
- CN
- China
- Prior art keywords
- instruction
- controller
- kernel function
- data
- period
- 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.)
- Active
Links
Images
Classifications
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B25—HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
- B25J—MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
- B25J9/00—Programme-controlled manipulators
- B25J9/16—Programme controls
- B25J9/1628—Programme controls characterised by the control loop
- B25J9/163—Programme controls characterised by the control loop learning, adaptive, model based, rule based expert control
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B25—HAND TOOLS; PORTABLE POWER-DRIVEN TOOLS; MANIPULATORS
- B25J—MANIPULATORS; CHAMBERS PROVIDED WITH MANIPULATION DEVICES
- B25J9/00—Programme-controlled manipulators
- B25J9/16—Programme controls
- B25J9/1628—Programme controls characterised by the control loop
- B25J9/1653—Programme controls characterised by the control loop parameters identification, estimation, stiffness, accuracy, error analysis
Landscapes
- Engineering & Computer Science (AREA)
- Robotics (AREA)
- Mechanical Engineering (AREA)
- Feedback Control In General (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
The invention provides a field control method for online machine learning, which comprises the following steps: 1) model training: keeping a signal path as a signal receiving end-a field controller-an instruction sending end, acquiring an instruction sent by the field controller and received data, and performing multi-section fitting on the insertion model by taking the received data as an input value and the sent instruction as an output value until the received data and the sent instruction can be fitted to at least one section of kernel function, and entering the next step, wherein the linear kernel function only comprises a basic form of a linear function; 2) strategy adjustment; 3) intervention control. The invention can provide effective online learning by a multi-stage and two-stage fitting mode, is convenient for enterprises to effectively reduce the investment of manpower and material resources, shortens the research and development period, can smoothly complete the handover of the control right from the traditional field controller to the insertion controller for machine learning, and is convenient for remotely adjusting parameters by another remote communication mode.
Description
Technical Field
The invention relates to a field control method for online machine learning.
Background
At present, the application of machine learning in industrial control is gradually increased, however, the most common trouble of local enterprises at present lies in that data is seriously deficient, which leads to the difficulty in training and completing a machine learning controller. However, by adopting the method, the research and development period is too long, the manpower and material resources are extremely high, and the enterprise is not as good as directly engaging operators.
To solve the above problems, the present company has designed a site control system based on time-sharing control handover control as shown in fig. 1, which is capable of ensuring from the hardware that the control handover from the traditional site controller to the controller learned as the control core by the machine can be a gradual replacement process, however, the prior art does not provide any technical suggestion on how to accomplish the control handover.
Disclosure of Invention
In order to solve the technical problems, the invention provides the on-site control method for the on-line machine learning, and the on-site control method for the on-line machine learning can provide effective on-line learning by a multi-stage and two-stage fitting mode, is convenient for enterprises to effectively reduce the investment of manpower and material resources and shortens the research and development period.
The invention is realized by the following technical scheme.
The invention provides a field control method for online machine learning, which comprises the following steps:
1) model training: keeping a signal path as a signal receiving end-a field controller-an instruction sending end, acquiring an instruction sent by the field controller and received data, and performing multi-section fitting on the insertion model by taking the received data as an input value and the sent instruction as an output value until the received data and the sent instruction can be fitted to at least one section of kernel function, and entering the next step, wherein the linear kernel function only comprises a basic form of a linear function;
2) strategy adjustment: adding an offset value to each item in the linear kernel function, switching a signal path into a signal receiving end-inserting controller-instruction transmitting end every N times of field controller signal transceiving cycles, inserting a controller cycle every time the signal path is switched into the signal receiving end-inserting controller-instruction transmitting end, inserting the controller in each controller cycle, updating the offset value by using received data and an output instruction obtained by the last controller cycle insertion controller according to the calculation of an insertion model, entering the next step if the offset value is updated, and otherwise, repeating the step;
3) intervention control: and keeping the signal path as a signal receiving end-an insertion controller-an instruction sending end, controlling the insertion controller through an insertion model, and updating the insertion model every N signal receiving and sending periods.
The basic form of the kernel function in the insertion model at least comprises a linear function, an exponential function and a trigonometric function.
The offset values are initialized with random values when added.
In the step 1), the insertion model is subjected to multi-segment fitting in the following way:
a. taking data received in the current signal period as an input value, taking an instruction sent in the current signal period as an output value, traversing the calculated kernel function to judge whether fitting can be performed or not, if fitting can be performed, discarding the data received in the current signal period, and if not, taking the received data and the corresponding sent instruction as a data pair and putting the data pair into a database to be processed;
b. acquiring the data pair count in the database to be processed, if the data pair count is larger than M, taking out the data pair in the database to be processed, calculating the kernel function parameters as the existing data, and calculating to obtain a plurality of kernel function elementary forms;
c. and substituting the extracted data pairs into various kernel function elementary forms, judging whether any kernel function elementary form can fit more than 80% of data pairs, if so, marking the kernel function elementary form as a calculated kernel function, discarding the data pairs fitted corresponding to the kernel function, returning the rest data pairs to the database to be processed, and re-initializing the rest kernel functions.
The fitting criterion is that the difference between the kernel function calculation result and the output value in the data pair is less than 10% of any one of the kernel function calculation result and the output value in the data pair.
And N is 10-30.
And M is 10-15.
The offset value is updated in the following way:
A. calculating the instruction of the current controller period according to the data received by the current controller period;
B. calculating a difference between the instruction of the current controller cycle and the instruction of the previous controller cycle;
C. calculating an offset value reversely according to the difference value;
D. and recalculating the instruction of the current controller period according to the data received by the current controller period.
The standard of the updating completion of the deviant is that after the instruction of the current controller period is calculated according to the data received by the current controller period, the difference value obtained by calculating the instruction of the current controller period and the instruction of the previous controller period is less than 10% of any one of the instruction of the current controller period and the instruction of the previous controller period.
The invention has the beneficial effects that: through the mode of multistage, two-stage fitting, can provide effectual online study, the enterprise of being convenient for effectively reduces the input of manpower and materials, shortens research and development cycle, can smoothly accomplish the handing-over of control right from traditional site controller to the plug-in controller of machine learning, and be convenient for through other remote communication mode and remote adjustment parameter.
Drawings
Fig. 1 is a schematic connection diagram of a field control system to which the present invention is applied.
Detailed Description
The technical solution of the present invention is further described below, but the scope of the claimed invention is not limited to the described.
The invention is applied to a field control system based on time-sharing control handover control right as shown in figure 1, in particular to a field control method for on-line machine learning, which comprises the following steps:
1) model training: keeping a signal path as a signal receiving end-a field controller-an instruction sending end, acquiring an instruction sent by the field controller and received data, and performing multi-section fitting on the insertion model by taking the received data as an input value and the sent instruction as an output value until the received data and the sent instruction can be fitted to at least one section of kernel function, and entering the next step, wherein the linear kernel function only comprises a basic form of a linear function;
2) strategy adjustment: adding an offset value to each item in the linear kernel function, switching a signal path into a signal receiving end-inserting controller-instruction transmitting end every N times of field controller signal transceiving cycles, inserting a controller cycle every time the signal path is switched into the signal receiving end-inserting controller-instruction transmitting end, inserting the controller in each controller cycle, updating the offset value by using received data and an output instruction obtained by the last controller cycle insertion controller according to the calculation of an insertion model, entering the next step if the offset value is updated, and otherwise, repeating the step;
3) intervention control: and keeping the signal path as a signal receiving end-an insertion controller-an instruction sending end, controlling the insertion controller through an insertion model, and updating the insertion model every N signal receiving and sending periods.
The basic form of the kernel function in the insertion model at least comprises a linear function, an exponential function and a trigonometric function.
The offset values are initialized with random values when added.
In the step 1), the insertion model is subjected to multi-segment fitting in the following way:
a. taking data received in the current signal period as an input value, taking an instruction sent in the current signal period as an output value, traversing the calculated kernel function to judge whether fitting can be performed or not, if fitting can be performed, discarding the data received in the current signal period, and if not, taking the received data and the corresponding sent instruction as a data pair and putting the data pair into a database to be processed;
b. acquiring the data pair count in the database to be processed, if the data pair count is larger than M, taking out the data pair in the database to be processed, calculating the kernel function parameters as the existing data, and calculating to obtain a plurality of kernel function elementary forms;
c. and substituting the extracted data pairs into various kernel function elementary forms, judging whether any kernel function elementary form can fit more than 80% of data pairs, if so, marking the kernel function elementary form as a calculated kernel function, discarding the data pairs fitted corresponding to the kernel function, returning the rest data pairs to the database to be processed, and re-initializing the rest kernel functions.
The fitting criterion is that the difference between the kernel function calculation result and the output value in the data pair is less than 10% of any one of the kernel function calculation result and the output value in the data pair.
And N is 10-30.
And M is 10-15.
The offset value is updated in the following way:
A. calculating the instruction of the current controller period according to the data received by the current controller period;
B. calculating a difference between the instruction of the current controller cycle and the instruction of the previous controller cycle;
C. calculating an offset value reversely according to the difference value;
D. and recalculating the instruction of the current controller period according to the data received by the current controller period.
The standard of the updating completion of the deviant is that after the instruction of the current controller period is calculated according to the data received by the current controller period, the difference value obtained by calculating the instruction of the current controller period and the instruction of the previous controller period is less than 10% of any one of the instruction of the current controller period and the instruction of the previous controller period.
The data received by the signal receiving end is generally data of a plurality of sensors, and the received control command can be regarded as an input vector, and similarly, the command sent by the command sending end is also generally a control value of a plurality of execution mechanisms and can be regarded as an output vector, and then the mapping relation between the input vector and the output vector can be fitted by adopting a traditional machine learning mode.
In the mapping relation, the influence of the basic form of the kernel function is the largest, so that in the training and learning of the first stage, the insertion model does not intervene in control, only parameter updating is carried out according to actually generated data, the influence of the offset value on the control is relatively small, the offset value can be quickly corrected to a reasonable range under the participation of a traditional field controller, and then the optimization is completed through self-adjustment.
The invention adopts the regression method of the traditional machine learning, is convenient for remotely adjusting parameters in another remote communication mode after the modeling is finished, and is more flexible compared with the existing field controller.
Claims (8)
1. A field control method for online machine learning is characterized in that: the method comprises the following steps:
1) model training: keeping a signal path as a signal receiving end-a field controller-an instruction sending end, acquiring an instruction sent by the field controller and received data, and performing multi-section fitting on the insertion model by taking the received data as an input value and the sent instruction as an output value until the received data and the sent instruction can be fitted to at least one section of kernel function, and entering the next step, wherein the linear kernel function only comprises a basic form of a linear function;
2) strategy adjustment: adding an offset value to each item in the linear kernel function, switching a signal path into a signal receiving end-inserting controller-instruction transmitting end every N times of field controller signal transceiving cycles, inserting a controller cycle every time the signal path is switched into the signal receiving end-inserting controller-instruction transmitting end, inserting the controller in each controller cycle, updating the offset value by using received data and an output instruction obtained by the last controller cycle insertion controller according to the calculation of an insertion model, entering the next step if the offset value is updated, and otherwise, repeating the step;
3) intervention control: keeping a signal path as a signal receiving end-an insertion controller-an instruction sending end, controlling the insertion controller through an insertion model, and updating the insertion model in every N signal receiving and sending periods;
in the step 1), the insertion model is subjected to multi-segment fitting in the following way:
a. taking data received in the current signal period as an input value, taking an instruction sent in the current signal period as an output value, traversing the calculated kernel function to judge whether fitting can be performed or not, if fitting can be performed, discarding the data received in the current signal period, and if not, taking the received data and the corresponding sent instruction as a data pair and putting the data pair into a database to be processed;
b. acquiring the data pair count in the database to be processed, if the data pair count is larger than M, taking out the data pair in the database to be processed, calculating the kernel function parameters as the existing data, and calculating to obtain a plurality of kernel function elementary forms;
c. and substituting the extracted data pairs into various kernel function elementary forms, judging whether any kernel function elementary form can fit more than 80% of data pairs, if so, marking the kernel function elementary form as a calculated kernel function, discarding the data pairs fitted corresponding to the kernel function, returning the rest data pairs to the database to be processed, and re-initializing the rest kernel functions.
2. The on-site control method for on-line machine learning according to claim 1, characterized in that: the basic form of the kernel function in the insertion model at least comprises a linear function, an exponential function and a trigonometric function.
3. The on-site control method for on-line machine learning according to claim 1, characterized in that: the offset values are initialized with random values when added.
4. The on-site control method for on-line machine learning according to claim 1, characterized in that: the fitting criterion is that the difference between the kernel function calculation result and the output value in the data pair is less than 10% of any one of the kernel function calculation result and the output value in the data pair.
5. The on-site control method for on-line machine learning according to claim 1, characterized in that: and N is 10-30.
6. The on-site control method for on-line machine learning according to claim 1, characterized in that: and M is 10-15.
7. The on-site control method for on-line machine learning according to claim 1, characterized in that: the offset value is updated in the following way:
A. calculating the instruction of the current controller period according to the data received by the current controller period;
B. calculating a difference between the instruction of the current controller cycle and the instruction of the previous controller cycle;
C. calculating an offset value reversely according to the difference value;
D. and recalculating the instruction of the current controller period according to the data received by the current controller period.
8. The on-site control method for on-line machine learning according to claim 1 or 7, characterized in that: the standard of the updating completion of the deviant is that after the instruction of the current controller period is calculated according to the data received by the current controller period, the difference value obtained by calculating the instruction of the current controller period and the instruction of the previous controller period is less than 10% of any one of the instruction of the current controller period and the instruction of the previous controller period.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811027921.1A CN109366485B (en) | 2018-09-04 | 2018-09-04 | On-site control method for on-line machine learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811027921.1A CN109366485B (en) | 2018-09-04 | 2018-09-04 | On-site control method for on-line machine learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109366485A CN109366485A (en) | 2019-02-22 |
CN109366485B true CN109366485B (en) | 2021-09-10 |
Family
ID=65404606
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811027921.1A Active CN109366485B (en) | 2018-09-04 | 2018-09-04 | On-site control method for on-line machine learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109366485B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106250899A (en) * | 2016-07-29 | 2016-12-21 | 华东交通大学 | A kind of orange disease and insect pests monitoring and pre-alarming method based on distributed compression perception WSN |
CN107924384A (en) * | 2015-03-11 | 2018-04-17 | 阿雅斯迪公司 | For the system and method using study model prediction result is predicted |
CN107991877A (en) * | 2017-12-20 | 2018-05-04 | 东南大学 | A kind of Dynamic Model Identification method and system based on Recognition with Recurrent Neural Network |
CN108394769A (en) * | 2018-03-17 | 2018-08-14 | 北京化工大学 | elevator control method, server and computer readable storage medium |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7756591B2 (en) * | 2006-04-25 | 2010-07-13 | Pegasus Technologies, Inc. | System for optimizing oxygen in a boiler |
-
2018
- 2018-09-04 CN CN201811027921.1A patent/CN109366485B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107924384A (en) * | 2015-03-11 | 2018-04-17 | 阿雅斯迪公司 | For the system and method using study model prediction result is predicted |
CN106250899A (en) * | 2016-07-29 | 2016-12-21 | 华东交通大学 | A kind of orange disease and insect pests monitoring and pre-alarming method based on distributed compression perception WSN |
CN107991877A (en) * | 2017-12-20 | 2018-05-04 | 东南大学 | A kind of Dynamic Model Identification method and system based on Recognition with Recurrent Neural Network |
CN108394769A (en) * | 2018-03-17 | 2018-08-14 | 北京化工大学 | elevator control method, server and computer readable storage medium |
Non-Patent Citations (1)
Title |
---|
一种基于DCS的分时训练神经网络;陈海列等;《自动化应用》;20171025(第10期);正文第1-2节 * |
Also Published As
Publication number | Publication date |
---|---|
CN109366485A (en) | 2019-02-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108121216B (en) | Automated workshop virtual debugging method based on digital factory | |
CN107544255B (en) | State compensation model control method for batch injection molding process | |
CN107015476A (en) | A kind of position and the cooperative control method of force signal to electrohydraulic servo system | |
CN112882954B (en) | Distributed database operation and maintenance dynamic threshold alarming method and device | |
US20220309393A1 (en) | Cloud-edge collaboration forecasting system and method for aluminum oxide production indexes | |
CN111515962B (en) | Transmission error compensation control method for flexible joint with harmonic reducer | |
CN109366485B (en) | On-site control method for on-line machine learning | |
CN102073274B (en) | Expert fuzzy incremental type self-adapting parameter on-line setting and optimizing system and method thereof | |
CN110244658B (en) | Method for improving interpolation position precision of cutting bed based on improved BP neural network | |
MX2023003687A (en) | Predictive modeling and control of cell culture. | |
CN107728481A (en) | A kind of closed loop modeling method and device based on Model Predictive Control | |
CN109188904B (en) | Online training field control method based on machine learning | |
CN116977558A (en) | Three-dimensional model updating method and device, computer equipment and storage medium | |
CN111008708A (en) | Parameter adjusting method and system for quasi-proportional resonant controller | |
CN110187638A (en) | A kind of large dead time fuzzy control method | |
CN109325593B (en) | Method for on-site control time-sharing control handover control right | |
CN105955227B (en) | The controller tracking and controlling method of object-oriented on a kind of production line | |
CN114800556A (en) | Automatic processing production line based on industrial robot | |
CN109669348B (en) | Industrial robot based on iterative feedback adjustment and trajectory tracking control method thereof | |
CN110161864A (en) | A kind of large dead time Fuzzy control system | |
CN111571594A (en) | Method for improving transparency and stability of teleoperation robot | |
CN107942658B (en) | Method and system for predicting rotation period of circular knitting machine by adopting adaptive filter | |
CN110825051A (en) | Multi-model control method of uncertainty system based on gap metric | |
CN205787684U (en) | A kind of novel fuzzy controller | |
CN109001981B (en) | Reinforced learning control method for sewage treatment |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
TR01 | Transfer of patent right |
Effective date of registration: 20211130 Address after: 535300 the Guangxi Zhuang Autonomous Region Qinzhou Pubei County Town Industrial Zone Patentee after: GUANGXI KAIXING CREATIVE TECHNOLOGY CO.,LTD. Address before: 530200 No. 8, Ting Ting Road, Yongning District, Nanning, the Guangxi Zhuang Autonomous Region Patentee before: NANNING University |
|
TR01 | Transfer of patent right |