CN113188715A - Multi-dimensional force sensor static calibration data processing method based on machine learning - Google Patents
Multi-dimensional force sensor static calibration data processing method based on machine learning Download PDFInfo
- Publication number
- CN113188715A CN113188715A CN202110287377.XA CN202110287377A CN113188715A CN 113188715 A CN113188715 A CN 113188715A CN 202110287377 A CN202110287377 A CN 202110287377A CN 113188715 A CN113188715 A CN 113188715A
- Authority
- CN
- China
- Prior art keywords
- output
- model
- calibration
- learning model
- learning
- 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
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01L—MEASURING FORCE, STRESS, TORQUE, WORK, MECHANICAL POWER, MECHANICAL EFFICIENCY, OR FLUID PRESSURE
- G01L25/00—Testing or calibrating of apparatus for measuring force, torque, work, mechanical power, or mechanical efficiency
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Force Measurement Appropriate To Specific Purposes (AREA)
- Aerodynamic Tests, Hydrodynamic Tests, Wind Tunnels, And Water Tanks (AREA)
Abstract
The invention provides a multi-dimensional force sensor static calibration data processing method based on machine learning, which comprises the following steps: s1, constructing a learning model, a data deviation model and a training model for evaluating the accuracy of the function relation in the learning model; s2, inputting a calibration voltage value set to the learning model to generate a simulation output quantity, wherein the simulation output quantity comprises at least one of output force and output torque; s3, importing the simulation output quantity generated by the learning model into a data deviation model, and comparing the simulation output quantity with corresponding calibration quantity through the data deviation model to obtain data deviation; s4, inputting the data deviation into a training model, and generating parameter updating quantity through the training model; s5, introducing the parameter updating quantity into a learning model to adjust the function relation of the learning model; and repeating the steps S2, S5 until the functional relation in the learning model reaches a preset fitting degree. The invention effectively improves the efficiency and the precision of static calibration.
Description
Technical Field
The invention relates to a static calibration data processing method of a multi-dimensional force sensor based on machine learning.
Background
The simulated wind field is the most basic test device for carrying out aerodynamic research and aircraft development. The force measurement test is the most basic experimental item in the simulation wind field experiment. The multidimensional force sensor is the most important measuring device in the force measurement test and is used for measuring the size, the direction and the acting point of aerodynamic load (force and moment) acting on a model. The multi-dimensional force sensor needs static calibration before a simulated wind field test, and the multi-dimensional force sensor calibration is to simulate the actual working state of the multi-dimensional force sensor to calibrate the multi-dimensional force sensor, obtain the measurement precision of the multi-dimensional force sensor, identify the performance of the multi-dimensional force sensor and provide necessary technical basis for later-stage test data of the multi-dimensional force sensor.
The multidimensional force sensor consists of an elastic element, a strain gauge and a measuring circuit (measuring bridge). In simulating wind field dynamometric experiments, the multi-dimensional force sensor bears the load acting on the model and transmits it to the support system. The elastic element deforms under the action of aerodynamic load, and the strain of the elastic element is in direct proportion to the magnitude of external force. The strain gauge adhered to the surface of the elastic element is also deformed simultaneously, so that the resistance value of the strain gauge is changed, the resistance increment is increased by a Wheatstone full-bridge measuring circuit formed by the strain gauges, the resistance increment is converted into voltage increment, and the voltage increment is in direct proportion to the aerodynamic load value born by the multi-dimensional force sensor. After the voltage signal is subjected to A/D conversion, the voltage signal is input to a computer for processing, and then the force and the moment acting on the model can be obtained.
The calibration of the multi-dimensional force sensor is the most important link in the development process of the multi-dimensional force sensor, and the calibration device and the calibration method have important influences on the calibration efficiency of the multi-dimensional force sensor and the measurement accuracy in future use. For a multi-dimensional force sensor with a newly designed or newly adhered strain gauge, static calibration is particularly important, but factors such as ambient temperature, a heating device, an experiment chamber, a loading system, a data acquisition system and the like need to be considered in the static calibration process, the process is complex, and calibration experience with abundant experience is needed, so that the efficiency needs to be further improved.
In view of the above, the present inventors have specially designed a method for processing static calibration data of a multi-dimensional force sensor based on machine learning, and have resulted in the present disclosure.
Disclosure of Invention
In order to solve the problems, the technical scheme of the invention is as follows:
the multi-dimensional force sensor static calibration data processing method based on machine learning comprises the following steps:
s1, constructing a learning model, a data deviation model and a training model for evaluating the accuracy of the function relation in the learning model;
s2, inputting a calibration voltage value set to the learning model to generate a simulation output quantity, wherein the simulation output quantity comprises at least one of output force and output torque;
s3, importing the simulation output quantity generated by the learning model into a data deviation model, and comparing the simulation output quantity with corresponding calibration quantity through the data deviation model to obtain data deviation;
s4, inputting the data deviation into a training model, and generating parameter updating quantity through the training model;
s5, introducing the parameter updating quantity into the learning model to adjust the function relation of the learning model;
and S6, repeating S2-S5 until the functional relation in the learning model reaches a preset fitting degree.
In some embodiments, constructing the learning model comprises: a learning model is built on the basis of an input layer, a hidden layer and an output layer of a neural network, the input layer is used for inputting a calibration voltage value set, the hidden layer is used for processing the calibration voltage value set, and corresponding output force and output torque are generated to the output layer.
In some embodiments, constructing the learning model based on the input layer, the hidden layer, and the output layer of the neural network comprises: setting the calibration voltage value set as a calibration voltage value set with six components and leading the calibration voltage value set into the input layer; the mapping of the calibration voltage value to the output force and the output torque is realized through the hidden layer.
In some embodiments, mapping the calibration voltage values to the output force and output torque through the hidden layer comprises: and fitting the data by using a linear function and a nonlinear activation function, wherein the activation function is a ReLU function.
In some embodiments, comparing the simulated output quantities to the calibration quantities by the data bias model comprises: the data bias is generated by calculating the sum of the squares of the differences between the simulated output quantities and the calibration quantities using a loss function method.
In some embodiments, generating the parameter update quantities by training the model comprises: using a back propagation method, a parameter update is generated based on data deviations from the output force and the output torque back to the calibration voltage value.
According to the static calibration method, the correspondence between the calibration voltage value and the output force and the torque is carried out through a method based on the neural network, the deviation is corrected through the data deviation model and the training model, and the efficiency and the precision of static calibration are effectively improved.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention and not to limit the invention.
Wherein:
FIG. 1 is a schematic overall flow chart provided by an embodiment of the present invention;
FIG. 2 is a block diagram of a structure provided by an embodiment of the present invention;
FIG. 3 is a block diagram of an overall learning model provided by an embodiment of the invention;
FIG. 4 is a schematic diagram of a loss function provided by an embodiment of the present invention;
FIG. 5 is a schematic diagram of the forward propagation principle provided by an embodiment of the present invention;
FIG. 6 is a schematic diagram of the back propagation principle provided by an embodiment of the present invention;
FIG. 7 is a schematic diagram of a cross-validation method provided by an embodiment of the invention;
fig. 8 is a diagram illustrating a calibration result according to an embodiment of the present invention.
Detailed Description
In order to make the technical problems, technical solutions and advantageous effects to be solved by the present invention clearer and clearer, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Referring to fig. 1 to 8, a method for processing static calibration data of a multi-dimensional force sensor based on machine learning according to a preferred embodiment of the present invention includes:
s1: constructing a learning model for describing a functional relation between input and output, a data deviation model and a training model for evaluating the accuracy of the functional relation in the learning model;
s2: inputting a calibration voltage value set to the learning model to generate a simulation output quantity, wherein the simulation output quantity comprises at least one of output force and output torque, and the calibration voltage set is one of three-component or four-component or five-component or six-component voltage sets and is used for inducing aerodynamic force and torque of air flow acting on the model in the simulation wind field;
s3: importing the simulation output quantity generated by the learning model into a data deviation model, and respectively comparing the simulation output quantity with corresponding calibration quantity through the data deviation model to obtain data deviation;
specifically, the learning model is y ═ b + w · XcpWherein X iscpFor input, y is output, b and w are weighting factors, in this embodiment, the calibration voltage value is input X of the testcpThe calibration quantity is a result value of the test, a preset functional relationship exists between the calibration voltage value and the calibration quantity is real and effective data, and the simulation output quantity is y used for approaching the calibration quantity so as to find out the optimal functional relationship between the calibration voltage value and the calibration quantity.
S4: inputting the data deviation into a training model, and generating a parameter updating amount through the training model;
s5: importing the parameter updating quantity into a learning model to enable the learning model to carry out adjustment on the functional relation;
s6: and repeating S2-S5 until the functional relation in the learning model reaches a preset fitting degree.
In the present embodiment, the calibration process of the multi-dimensional force sensor is a multi-dimensional non-linear regression problem, i.e. a non-linear fitting problem. In the process of deep learning of the learning model, the simulation output quantity generated according to the calibration voltage is led into a data deviation model, further, the data deviation model compares the relation between the simulation output quantity and the calibration quantity to calculate deviation, then the deviation is led into a training model, the training model calculates forward loss according to the deviation and then calculates backward error, and finally the backward error is fed back to the learning model for parameter revision so as to finish a cycle; with the increase of the number of tests, the obtained functional relationship has higher precision, and when the data deviation is smaller than a preset threshold value, such as 0.2%, the data can be normally used, and the specific data deviation algorithm please refer to the following text.
In some embodiments, constructing the learning model comprises: constructing a learning model based on an input layer, a hidden layer and an output layer of a neural network, wherein the input layer is used for inputting a calibration voltage value set, and the hidden layer is used for processing the calibration voltage value set and generating corresponding output force and output torque to the output layer; the method for constructing the learning model based on the input layer, the hidden layer and the output layer of the neural network comprises the following steps: setting the calibration voltage value set as a calibration voltage value set with six components and leading the calibration voltage value set into the input layer; mapping from the calibration voltage value to the output force and the output torque is realized through the hidden layer; the mapping of the calibration voltage values to the output force and the output torque through the hidden layer comprises the following steps: and fitting the data by using a linear function and a nonlinear activation function, wherein the activation function is a ReLU function.
In this embodiment, referring to fig. 3, fig. 3 is divided into three regions, where an input region corresponds to an input layer for importing a calibration voltage set, an output region corresponds to an output layer, and a sequential corresponds to a hidden layer; specifically, the present embodiment employs a six-component multidimensional force sensor, and thus the calibration voltage includes a voltage value set based on six-component force or moment, that is, the input layer is provided with 6 neurons, and accordingly, the output layer has 6 neurons, the hidden layer has more than 100 neurons, specifically, the number of neurons in the hidden layer in the present embodiment is 500, the number of hidden layers depends on the precision, it should be noted that the neurons are composed of two parts, the first part (e) is the sum of the products of the input value and the weight coefficient, the second part f (e) is the output of an activation function (ReLU function), and y ═ f (e) is the output of a certain neuron.
In some embodiments, comparing the simulated output quantities to the calibration quantities by the data bias model comprises: the data bias is generated by calculating the sum of the squares of the differences between the simulated output quantities and the calibration quantities using a loss function method.
In this embodiment, the loss function includes two partsAs shown in FIG. 4, the first part is to calculate the sum of the squares of the differences between the simulated output quantity and the calibration quantity, and the specific formula is
Wherein
In order to calibrate the amount of the measurement,
to simulate the output; the second part is to calculate the data deviation, and the concrete formula is
Wherein
The function is used to minimize a set of values for the weighting coefficients.
In some embodiments, generating the parameter update quantities by training the model comprises: using a back propagation method, a parameter update is generated based on data deviations from the output force and the output torque back to the calibration voltage value.
In this embodiment, referring to FIG. 5, first pass x1、x2The input forward propagation is calculated to obtain a simulated output y, 6 neurons are arranged between the input and the simulated output in fig. 5, and the input of the neuron behind is based on the weighted output of the neuron in front until the simulated output y is obtained, and then the simulated output y is compared with the calibration quantity z; secondly, referring to fig. 6, the total error δ is obtained as z-y by using back propagation, then the deviation of each neuron is counted in sequence, the derivative of each neuron (weight) is calculated, and the back propagation is started to modify the weight;
in actual experiments, taking the calibration data 1056 as an example, the data is divided into 4 groups, three of which are used for cross-validation, and the other group evaluates the error of the training model in the simulated wind field, and further, as shown in fig. 7, the data in the cross-validation is divided into three groups, two of which are used as training data, and one of which evaluates the loss. The combination form has three types, and finally, the model with the minimum average loss is selected.
Finally, referring to fig. 8, taking 60 sets of test data as an example, the loss is continuously reduced through the cyclic update, so as to find the optimal functional relationship between the calibration voltage value and the simulation output quantity, i.e. the closest functional relationship between the calibration voltage value and the calibration quantity.
According to the static calibration method, the correspondence between the calibration voltage value and the output force and the torque is carried out through a method based on the neural network, the deviation is corrected through the data deviation model and the training model, and the efficiency and the precision of static calibration are effectively improved.
The invention has been described above with reference to the accompanying drawings, it is obvious that the invention is not limited to the specific implementation in the above-described manner, and it is within the scope of the invention to apply the inventive concept and solution to other applications without substantial modification.
Claims (6)
1. The multi-dimensional force sensor static calibration data processing method based on machine learning is characterized by comprising the following steps:
s1, constructing a learning model, a data deviation model and a training model for evaluating the accuracy of the function relation in the learning model;
s2, inputting a calibration voltage value set to the learning model to generate a simulation output quantity, wherein the simulation output quantity comprises at least one of output force and output torque;
s3, importing the simulation output quantity generated by the learning model into a data deviation model, and comparing the simulation output quantity with corresponding calibration quantity through the data deviation model to obtain data deviation;
s4, inputting the data deviation into a training model, and generating parameter updating quantity through the training model;
s5, introducing the parameter updating quantity into a learning model to adjust the function relation of the learning model;
and S6, repeating S2-S5 until the functional relation in the learning model reaches a preset fitting degree.
2. The machine-learning-based multi-dimensional force sensor static calibration data processing method of claim 1, wherein the building a learning model comprises: a learning model is constructed on the basis of an input layer, a hidden layer and an output layer of a neural network, wherein the input layer is used for inputting a calibration voltage value set, and the hidden layer is used for processing the calibration voltage value set and generating corresponding output force and output torque to the output layer.
3. The method of processing machine learning based multi-dimensional force sensor static calibration data according to claim 2, wherein the building of the learning model based on the input layer, the hidden layer and the output layer of the neural network comprises: setting the calibration voltage value set as a calibration voltage value set with six components and leading the calibration voltage value set into an input layer; the mapping of the calibration voltage value to the output force and the output torque is realized through the hidden layer.
4. The machine-learning-based multi-dimensional force sensor static calibration data processing method of claim 3, wherein the mapping of calibration voltage values to output forces and output torques by hidden layers comprises: the data is fitted using a linear function and a non-linear activation function, the activation function being a ReLU function.
5. The machine-learning based multi-dimensional force sensor static calibration data processing method of claim 1, wherein the comparing the simulated output quantities with the calibration quantities by the data bias model respectively comprises: the data bias is generated by calculating the sum of the squares of the differences between the simulated output quantities and the calibration quantities using a loss function method.
6. The machine-learning based multi-dimensional force sensor static calibration data processing method of claim 5, wherein the generating parameter updates through the training model comprises: using a back propagation method, a parameter update is generated based on data deviations from the output force and the output torque back to the calibration voltage value.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110287377.XA CN113188715A (en) | 2021-03-17 | 2021-03-17 | Multi-dimensional force sensor static calibration data processing method based on machine learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110287377.XA CN113188715A (en) | 2021-03-17 | 2021-03-17 | Multi-dimensional force sensor static calibration data processing method based on machine learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113188715A true CN113188715A (en) | 2021-07-30 |
Family
ID=76973350
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110287377.XA Pending CN113188715A (en) | 2021-03-17 | 2021-03-17 | Multi-dimensional force sensor static calibration data processing method based on machine learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113188715A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114705356A (en) * | 2022-04-19 | 2022-07-05 | 上海工业自动化仪表研究院有限公司 | Self-calibration method of resistance strain gauge type force transducer |
| CN114894379A (en) * | 2022-05-26 | 2022-08-12 | 湖南大学 | Calibration device and calibration method for fingertip type touch sensor of manipulator |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108351986A (en) * | 2015-10-30 | 2018-07-31 | 株式会社摩如富 | Learning system, learning device, learning method, learning program, training data generating means, training data generation method, training data generate program, terminal installation and threshold value change device |
| CN111256905A (en) * | 2020-02-14 | 2020-06-09 | 金陵科技学院 | Decoupling method of multi-dimensional force sensor in noise environment |
| CN111272334A (en) * | 2020-02-18 | 2020-06-12 | 金陵科技学院 | Particle swarm optimization BP neural network-based multidimensional force sensor calibration decoupling method |
| CN111380649A (en) * | 2020-03-30 | 2020-07-07 | 金陵科技学院 | Vehicle multi-dimensional force sensor decoupling method based on PSO (particle swarm optimization) optimization LSSVM (least squares support vector machine) |
| CN111637992A (en) * | 2020-05-19 | 2020-09-08 | 合肥沃智信息科技有限公司 | Error compensation method of high-precision pressure sensor |
-
2021
- 2021-03-17 CN CN202110287377.XA patent/CN113188715A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108351986A (en) * | 2015-10-30 | 2018-07-31 | 株式会社摩如富 | Learning system, learning device, learning method, learning program, training data generating means, training data generation method, training data generate program, terminal installation and threshold value change device |
| CN111256905A (en) * | 2020-02-14 | 2020-06-09 | 金陵科技学院 | Decoupling method of multi-dimensional force sensor in noise environment |
| CN111272334A (en) * | 2020-02-18 | 2020-06-12 | 金陵科技学院 | Particle swarm optimization BP neural network-based multidimensional force sensor calibration decoupling method |
| CN111380649A (en) * | 2020-03-30 | 2020-07-07 | 金陵科技学院 | Vehicle multi-dimensional force sensor decoupling method based on PSO (particle swarm optimization) optimization LSSVM (least squares support vector machine) |
| CN111637992A (en) * | 2020-05-19 | 2020-09-08 | 合肥沃智信息科技有限公司 | Error compensation method of high-precision pressure sensor |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114705356A (en) * | 2022-04-19 | 2022-07-05 | 上海工业自动化仪表研究院有限公司 | Self-calibration method of resistance strain gauge type force transducer |
| CN114705356B (en) * | 2022-04-19 | 2023-11-14 | 上海工业自动化仪表研究院有限公司 | Self-calibration method of resistance strain gauge force transducer |
| CN114894379A (en) * | 2022-05-26 | 2022-08-12 | 湖南大学 | Calibration device and calibration method for fingertip type touch sensor of manipulator |
| CN114894379B (en) * | 2022-05-26 | 2023-03-07 | 湖南大学 | Calibration device and calibration method for fingertip type touch sensor of manipulator |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113188715A (en) | Multi-dimensional force sensor static calibration data processing method based on machine learning | |
| CN108073075A (en) | Silicon micro accerometer temperature-compensation method, system based on GA Optimized BP Neural Networks | |
| Cook | Influence assessment | |
| Liang et al. | Novel decoupling algorithm based on parallel voltage extreme learning machine (PV-ELM) for six-axis F/M sensors | |
| CN104915509B (en) | Large deformation flexible body dynamic stress metrical information conversion method based on neutral net | |
| CN113268935B (en) | Decoupling analysis method based on array type flexible touch sensor | |
| CN111638034B (en) | Strain balance temperature gradient error compensation method and system based on deep learning | |
| CN116305522A (en) | Digital twin-based aircraft structure reliability simulation test model calibration method, device and medium | |
| WO2021059787A1 (en) | Control device, learning device, control method, learning method, and program | |
| CN107367334A (en) | A kind of non-linear compensation method for RTD measurement | |
| CN112711902A (en) | Power grid voltage calculation method based on Monte Carlo sampling and deep learning | |
| CN113343559B (en) | Reliability analysis method for response surface of iterative reweighted least square method extreme learning machine | |
| CN104899642B (en) | Large deformation flexible body dynamic stress compensation method based on hybrid production style | |
| CN111256905B (en) | Decoupling method of multi-dimensional force sensor in noise environment | |
| CN109579967A (en) | Intelligent Dynamic weighing method and system | |
| CN111598215A (en) | Temperature compensation method and system based on neural network | |
| CN109885896A (en) | A kind of nonlinear organization correction method for finite element model based on multiple change differential sensitivity | |
| CN114791334A (en) | Calibration simplification method for pressure sensor | |
| CN107587955A (en) | The scaling method of rocket engine thrust offset based on deep belief network | |
| CN105160069B (en) | A kind of mechanical parameter flexible measurement method for optimization algorithm of being imparted knowledge to students based on improved close-coupled | |
| Chuan et al. | The application of RBF neural network in the compensation for temperature drift of the silicon pressure sensor | |
| CN107607182B (en) | truck weighing system and weighing method | |
| CN104765929B (en) | A kind of nonlinear state decoupling observation procedure based on Multiple Time Scales recurrent neural network | |
| Zhang et al. | Nonlinear dynamic and effect parameters analysis for extension pneumatic muscle | |
| Xie et al. | Hybrid neural network models of transducers |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20210730 |