Six-dimensional force / torque sensor calibration data enhancement method and system

Abstract

The invention discloses a calibration data enhancement method for a six-dimensional force / torque sensor, and the method comprises the steps: carrying out the calibration experiment of the six-dimensional force / torque sensor through a counterweight type loading method, and obtaining the output data of the sensor in real time; a data vector composed of output data of each loading point sensor and an actual loading force value forms training data and test data; self-adaptive interpolation is carried out according to the local linear error of each loading point, and data is expanded; and finally, merging the training data and the extended data obtained by interpolation, inputting the merged data into a six-dimensional force / torque decoupling model for weighted training, fixing model parameters to obtain a final decoupling model, and checking the decoupling effect of the model by adopting test data so as to realize data enhancement. According to the method, on the premise that the experiment complexity is not increased, the space density and feature expression ability of the training data set are enhanced, the generalization ability of the nonlinear decoupling model in an unknown loading state is improved, and the measurement accuracy and stability of the six-dimensional force / torque sensor are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of six-dimensional force / torque sensor data calibration, and mainly relates to a six-dimensional force / torque sensor calibration data enhancement method and system. BACKGROUND

[0002] As a high-precision force measuring device, the six-dimensional force / torque sensor can simultaneously perceive the force components in three directions and the torque components in three directions in space, and is widely used in robot end force control, intelligent manufacturing, aerospace assembly, biomedical equipment, virtual reality interaction and other fields. With the development of related technologies, the requirements for the precision, robustness and decoupling ability of the six-dimensional force / torque sensor are continuously improved. At present, one of the key problems restricting the precision improvement of the six-dimensional force / torque sensor is the inter-dimensional coupling error, that is, when the sensor is loaded in a certain direction, non-zero output signals appear in other directions. In order to solve the coupling error problem, researchers generally use decoupling algorithms to reconstruct the sensor signals.

[0003] The existing decoupling methods mainly include two types: one is hardware decoupling achieved through structural design, and the other is software decoupling based on experimental data driving. Among them, the software decoupling method is widely adopted due to its flexibility, low cost, easy deployment and other advantages. Representative decoupling algorithms include least squares, support vector regression, back propagation neural network and extreme learning machine, etc. These algorithms rely on high-quality calibration data sets to establish the mapping relationship from sensor voltage output to actual force / torque input.

[0004] The current mainstream calibration method is the weight hammer loading method, which applies standard weights to the sensor in different directions through a pulley rope set to realize step-by-step loading and then collects the sensor output. Although this method has high precision, it is subject to the discreteness of the weight and the complexity of manual loading, resulting in sparse calibration data, which is difficult to meet the requirements of data density and diversity for nonlinear decoupling algorithms, and is prone to overfitting. SUMMARY

[0005] The present application is just for the problems of less calibration data quantity, sparse distribution, easy to cause over-fitting phenomenon of decoupling model and so on in the calibration process of traditional six-dimensional force / torque sensor, and provides a six-dimensional force / torque sensor calibration data enhancement method, first, the six-dimensional force / torque sensor is calibrated by using the weight hammer loading method, the output data of the sensor is obtained in real time, the data vector composed of the output data of the sensor of each loading point and the actual loading force value constitutes the training data and the test data, then the local linear error of each loading point is adaptively interpolated to expand the data, finally, the training data and the expanded data obtained by interpolation are combined and input into the six-dimensional force / torque decoupling model for weighted training, the model parameters are fixed, the optimal decoupling model is obtained, the verification data is used to test the decoupling effect of the model, so that the data enhancement is realized. The method of the present application enhances the spatial density and feature expression ability of the training data set without increasing the experimental complexity, improves the generalization ability of the nonlinear decoupling model under unknown loading state, and improves the measurement accuracy and stability of the six-dimensional force / torque sensor.

[0006] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: a six-dimensional force / torque sensor calibration data enhancement method, comprising the following steps:

[0007] S1, data acquisition: the six-dimensional force / torque sensor is calibrated by using the weight hammer loading method, and the output data of the sensor is obtained in real time; the calibration test is divided into high-density calibration test group and low-density calibration test group according to the step size, in the high-density calibration test group, the data vector composed of the output data of the sensor of each loading point and the actual loading force value constitutes the training data; in the low-density calibration test group, the data vector composed of the output data of the sensor of each loading point and the actual loading force value constitutes the test data;

[0008] S2, adaptive interpolation based on error: the training data obtained in step S1 is adaptively interpolated based on error, the training data obtained in step S1 is linearly fitted to obtain the linear prediction output of each loading point in the training data, the linear error analysis of the training data is carried out according to the linear prediction output, the error box number of each loading point is obtained, and finally the expanded data is obtained by cubic spline interpolation according to the error box number;

[0009] S3, training and evaluation: the training data obtained in step S1 and the expanded data obtained after step S2 are combined, input into the six-dimensional force / torque decoupling model for weighted training, the model parameters are fixed, the optimal decoupling model is obtained; the test data is input into the optimal decoupling model, the difference between the model output and the actual loading force value is compared, the measurement accuracy of the sensor is obtained, and the data enhancement is realized.

[0010] As an improvement of the present application, the calibration step length of the high-density calibration experiment group is not higher than one-tenth of the range, the calibration step length of the low-density calibration experiment group is not lower than one-eighth of the range, and the calibration step length of the low-density calibration experiment group is greater than that of the high-density calibration experiment group; both experiment groups use full-process calibration method to load and unload the weights; the output data of the sensor is six-dimensional voltage analog quantity, and the actual loading force value is the actual loading six-dimensional force / torque value in each direction during the calibration process.

[0011] As an improvement of the present application, the weight loading process of the calibration test is specifically as follows: the weights are loaded from zero state to the upper limit of the sensor range on the positive loading end of the calibration platform, and then unloaded to zero state; the weights are loaded from zero state to the lower limit of the sensor range on the negative loading end of the calibration platform, and then unloaded to zero state.

[0012] As another improvement of the present application, the linear error analysis in step S2 specifically includes the following steps:

[0013] S21: calculating the linear fitting error at each loading point, the error of the i-th loading point is e i :

[0014]

[0015] wherein y i is the output vector of the i-th loading point, is the actual loading force value, is the linear prediction output of the i-th loading point;

[0016] S22: sorting the linear error set of the loading points in ascending order to obtain the sorted linear error sequence wherein N is the total number of loading points in the training data;

[0017] S23: defining the percentile function P -1 (q):

[0018] P -1 (q)=E[qN]

[0019] wherein q is the percentile probability between 0 and 1, [·] represents the down rounding operation, and E[qN] represents the [qN]-th number in the linear error sequence E;

[0020] S24: performing percentile segmentation on the linear error sequence E obtained in step S22 to obtain n error bins, the number of bins n can be adjusted according to actual conditions, and the error interval range corresponding to the j-th bin is:

[0021]

[0022] wherein B j-1 , j respectively represent the lower and upper bounds of the error of the jth box;

[0023] S25: Each loading point is classified into the corresponding box according to the size of its linear error, and the corresponding error box number is obtained.

[0024] As another improvement of the present application, the cubic spline interpolation in step S2 specifically includes the following steps:

[0025] S26: An adaptive interpolation density is designed according to the error box number of each calibration data point, and an interpolation density multiple M is defined i Specifically:

[0026]

[0027] wherein a is the interpolation multiple, b j is the box number of data point i;

[0028] S27: The interpolation range of each loading point is determined to be half of the distance between the current point and its adjacent points, and the interpolation range at the end point is reduced by half. For the ith loading point, the interpolation range is defined by the input vector of the loading point:

[0029]

[0030] wherein xi i represents the input vector of the ith loading point in the training data, i.e., the sensor output data;

[0031] S28: A cubic spline function S(x) from the input vector to the output vector is constructed on the training data, which is a cubic polynomial in each interpolation interval, and satisfies the conditions that the function value is continuous at the interpolation nodes, the function value is consistent with the actual output, the first and second derivatives are continuous at the nodes, and the natural boundary condition is satisfied at the boundary; Specifically:

[0032] S(x i )=y i ,S(x i+1 )=y i+1

[0033]

[0034] S″(x1)=S″(x N )=0

[0035] wherein and represent the first left derivative and the first right derivative of the cubic spline function at the ith loading point, and denote the second order left and right derivatives of the cubic spline function at the i-th loading point, S''(xi) and S''(xi+1) N denote the second order derivative values of the cubic spline function at the first and last loading points;

[0036] S29: At each loading point, according to the interpolation range and the interpolation multiple M i new points are uniformly inserted, for the i-th loading point, M i -1 new points, wherein the input vector x interp (k) and the output vector y interp (k) of the k-th new point are:

[0037]

[0038] y interp (k) = S(x interp (k))

[0039] where S(x interp (k)) is the value of the cubic spline function obtained in step S28 at x interp (k), and all the new points form an extended data set.

[0040] As a further improvement of the present application, the data set D used for the weighted training of the six-dimensional force / torque decoupling model in step S3 is specifically:

[0041]

[0042] where N total is the total number of data points, x i , y i and w i are the input vector, the output vector and the weight of the i-th data point, w real is the training data weight, and w aug ∈(0,1) is the weight of the extended data.

[0043] As a further improvement of the present application, the six-dimensional force / torque decoupling model in step S3, when the input is a support vector machine-based decoupling model, is trained by adjusting the optimization objective function:

[0044]

[0045] where w and b are model parameters, ξ, ξ * are relaxation variables, and C is a penalty coefficient.

[0046] When the decoupling model based on the back propagation neural network method is input, weighted training is achieved by modifying the loss function, and the loss function Loss is defined using the weighted mean square error:

[0047]

[0048] in is the predicted output vector of the current model;

[0049] When the decoupling model based on the extreme learning machine method is input, the weighted operation is introduced in the step of solving the weight parameters of the output layer to construct the weight matrix W d , randomly initialize the input weight matrix W in and bias b, calculate the hidden layer output matrix H, and finally use the weighted least squares method to calculate the output layer weight W out :

[0050] W d =diag(w1,w2,…,w N )

[0051] H=φ(XW in +b)

[0052] W out =(H T W d H) -1 H T W d Y

[0053] where φ(·) represents the hidden layer activation function, diag(w1,w2,…,w N ) represents the diagonal matrix composed of the weights of the 1st to Nth data points, X is the input matrix of the training dataset, and Y is the reference output matrix of the training dataset.

[0054] In order to achieve the above-mentioned purpose, the present invention also adopts the following technical solution: a six-dimensional force / torque sensor calibration data enhancement system, comprising a computer program, which implements the steps of any of the above-mentioned methods when executed by a processor.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] (1) The method of the present invention performs adaptive interpolation based on the linear error of the original data, can dynamically adjust the interpolation density according to the local nonlinear characteristics, optimize the spatial distribution of the data, and significantly improve the richness and effectiveness of the training data set.

[0057] (2) The number of bins in the error analysis process and the interpolation multiple in the adaptive interpolation process of the method can be adjusted according to the specific error degree of the data, have high flexibility, and adapt to different task requirements.

[0058] (3) The method adopts a weighted training mechanism, ensures that the training process is dominated by real data, fully utilizes interpolation data to enrich model training, and effectively ensures the stability of actual application.

[0059] (4) The method is suitable for mainstream six-dimensional force / torque sensor software decoupling algorithms, can improve the overfitting problem of nonlinear decoupling models, and improves the robustness of the model to unknown data.

[0060] (5) The method can greatly expand the training data set through data enhancement without increasing the complexity and cost of calibration experiments, is easy to implement, has good practical promotion value and wide application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 is a schematic diagram of the sensor calibration process in step S1 of the method of the application;

[0062] Figure 2 is a schematic diagram of the full-process calibration method used in step S1 of the method of the application;

[0063] Figure 3 is a schematic diagram of the error-based adaptive interpolation algorithm in step S2 of the method of the application;

[0064] Figure 4 is a schematic diagram of the training and evaluation process in step S3 of the method of the application. DETAILED DESCRIPTION

[0065] The application will be further illustrated below in conjunction with the drawings and specific embodiments, and it should be understood that the following specific embodiments are only used to illustrate the application and not to limit the scope of the application.

[0066] Example 1

[0067] A six-dimensional force / torque sensor calibration data enhancement method specifically includes the following steps:

[0068] Step S1, data acquisition: the six-dimensional force / torque sensor is calibrated by the weight hammer loading method, and the output data of the sensor is acquired in real time.

[0069] As Figure 1As shown in the sensor calibration process schematic diagram, a weight hammer loading method is used to carry out calibration experiments on a six-dimensional force / torque sensor, and is independently carried out in the three-dimensional force and three-dimensional torque directions of the X axis, Y axis and Z axis in the Cartesian coordinate system. The output data of the sensor are recorded in real time during the process, including the following steps:

[0070] S11: A calibration platform is built using a weight hammer loading method, and a six-dimensional force / torque sensor is installed.

[0071] S12: Carrying out calibration experiments, including a high-density calibration experiment group and a low-density calibration experiment group, wherein the high-density calibration experiment group specifically is: using the full-process calibration method to load and unload the weight, and the calibration step is small and repeated multiple times; the low-density calibration experiment group specifically is: using the full-process calibration method to load and unload the weight, and the calibration step is larger than that of the high-density calibration experiment group and is at least greater than the calibration step of the high-density calibration experiment group; the calibration step is the weight quality of single loading / unloading in the full-process calibration method.

[0072] S13: The experimental data collected in step S12 are classified and arranged, the experimental data from the high-density calibration experiment group constitute training data, which are used for further processing in the subsequent error-based adaptive interpolation link, and the experimental data from the low-density calibration experiment group constitute test data, which are used for error testing in the training and evaluation link;

[0073] The full-process calibration method is as shown in Figure 2 The specific process is: first, the weight is loaded from zero state to the upper limit of the sensor range on the positive loading end of the calibration platform, then unloaded to zero state, then the weight is loaded to the lower limit of the sensor range on the negative loading end of the calibration platform, and then unloaded to zero state, during which the sensor output data are collected in real time.

[0074] The training data and the verification data both consist of multiple loading points, each loading point is a data vector composed of the data output of the sensor and the actual loading force value, wherein the data output of the sensor is a six-dimensional voltage analog quantity; the actual loading force value is the six-dimensional force / torque value actually loaded in each direction during the calibration process.

[0075] Step S2, error-based adaptive interpolation: error-based adaptive interpolation is performed on the training data obtained in step S1.

[0076] Figure 3 An error-based adaptive interpolation algorithm schematic diagram is shown in

[0077] First, linear fitting is performed, taking the sensor data output of each loading point in the training data as the input sample matrix X, and the actual loading force value as the output sample matrix Y, N is the number of loading points, x i , y i are the input vector and output vector of the ith loading point respectively, and the coefficient matrix W is solved by using linear least squares method:

[0078] W=(X T X) -1 X T Y

[0079]

[0080] Second, linear error analysis is performed, including the following steps:

[0081] S1: Calculate the linear fitting error of each loading point, and the error of the ith loading point is e i :

[0082]

[0083] S2: Sort the linear error set of the loading points in ascending order to obtain the sorted linear error sequence

[0084] S3: Define the percentile function P -1 (q), where q is the percentile probability between 0 and 1, and [·] represents the down rounding operation.

[0085] P -1 (q)=E[qN]

[0086] S4: Percentile partitioning is performed on the linear error sequence E to obtain n error bins, and the error interval range corresponding to the jth bin is:

[0087]

[0088] Where B j-1 ,B j are the lower and upper bounds of the jth bin, respectively.

[0089] S5: Each loading point is assigned to the corresponding bin according to the size of its linear error, and the corresponding error bin number is obtained.

[0090] Finally, cubic spline interpolation is performed, including the following steps:

[0091] S1: Design adaptive interpolation density according to the error bin number of each calibration data point, and the interpolation multiple is α, and the bin number of data point i is b j , define the interpolation density multiple M​i :

[0092]

[0093] S2: Determine the interpolation range at each loading point as half of the distance between the current point and its adjacent points, and reduce the interpolation range by half at the end point. For the i-th loading point, define the interpolation range with the input vector of the loading point:

[0094]

[0095] S3: Construct a cubic spline function S(x) from the input vector to the output vector on the training data, which is a cubic polynomial in each interpolation interval, and satisfies the continuous function value at the interpolation nodes, and the function value is consistent with the actual output, while the first and second derivatives are continuous at each node, and the natural boundary condition is satisfied at the boundary; Specifically:

[0096] S(x i )=y i ,S(x i+1 )=y i+1

[0097]

[0098] S″(x1)=S″(x N )=0

[0099] where and represent the first left derivative and the first right derivative of the cubic spline function at the i-th loading point, and represent the second left derivative and the second right derivative of the cubic spline function at the i-th loading point, S″(x1) and S″(xN) represent the second derivative values of the cubic spline function at the first loading point and the last loading point;

[0100] S4: At each loading point, according to the interpolation range and the interpolation multiple M i , uniformly insert new points, for the i-th loading point, insert M i -1 new points, where the input vector x interp (k) and the output vector y interp (k) of the k-th new point are:

[0101]

[0102] y interp (k)=S(x interp (k))

[0103] where S(x interp(k) is the value of the cubic spline function obtained in step S28 at x interp (k), and all the new points obtained constitute the augmented data.

[0104] Step S3, training and evaluation: as Figure 4 the schematic diagram of the training and evaluation process, first, the training data and the augmented data obtained by interpolation are combined and sent into the specific six-dimensional force / torque decoupling model for weighted training, the model parameters are fixed, and then the test data is used to test the decoupling effect of the model;

[0105] In the weighted training of the model, different training weights are given to the training data and the augmented data, wherein the training data weight w real =1, the weight of the augmented data w aug ∈(0, 1) and can be adjusted according to the situation, and a complete training data set D is constructed:

[0106]

[0107] Where N total is the total number of data points, x i , y i and w i are the input vector, output vector and weight of the ith data point, respectively.

[0108] For different types of nonlinear decoupling algorithms, the corresponding weighted training method is constructed:

[0109] For the decoupling model based on the support vector machine method, the weighted training is carried out by adjusting the optimization objective function:

[0110]

[0111] Where w and b are model parameters, ξ, ξ * are relaxation variables, and C is a penalty coefficient.

[0112] For the decoupling model based on the back propagation neural network method, the weighted training is realized by modifying the loss function, and the weighted mean square error is used to define the loss function Loss:

[0113]

[0114] Where is the predicted output vector of the current model.

[0115] For the decoupling model based on the extreme learning machine method, the weighted operation is introduced in the output layer weight parameter solving step, first, the weight matrix W d is constructed, and then the input weight matrix W in is randomly initialized.and bias b, calculate the hidden layer output matrix H, and finally calculate the output layer weight W by using the weighted least square method out :

[0116] W d = diag(w1,w2,…,w N )

[0117] H = φ(XW in +b)

[0118] W out = (H T W d H) -1 H T W d Y

[0119] where φ(·) represents the hidden layer activation function, diag(w1,w2,…,w N ) represents a diagonal matrix composed of the first to the Nth data point weights, X is the input matrix of the training data set, and Y is the reference output matrix of the training data set.

[0120] Solve the model parameters by gradient descent or Moore-Penrose generalized inverse, and obtain the final decoupling model.

[0121] Finally, run the decoupling model on the test data, compare the differences between the model output and the actual load value in terms of root mean square error, inter-dimensional coupling error, etc., and obtain the measurement accuracy of the sensor to complete data enhancement.

[0122] Test example

[0123] In this test example, three mainstream six-dimensional force decoupling algorithms are selected for index comparison experiments, including support vector regression method, back propagation neural network method and extreme learning machine method. The calibration experiment includes two groups of high-density calibration experiment groups and eight groups of low-density calibration experiment groups. The step size of the weight loading / unloading process of the high-density calibration experiment group is half of that of the low-density calibration experiment group, and two groups of test data and eight groups of training data are obtained accordingly. The training data is subjected to data enhancement to obtain expanded data. The expanded data and the training data are used for model training, and the test data are used for comparison of the decoupling effect of the model. The test indexes include the root mean square error and the inter-dimensional coupling error of each direction.

[0124] (1) Root mean square error (MSE), defined as:

[0125]

[0126] where y ij and represent the true value and the decoupling algorithm prediction output in the jth direction at the ith test point, respectively.

[0127] The experimental results are shown in Table 1, and ''before enhancement'' means that the model is trained by using the original eight groups of training data, and ''after enhancement'' means that the model is trained by using the original eight groups of training data and the expanded data:

[0128] Table 1

[0129]

[0130] As can be seen from Table 1, after the data enhancement method is processed, the root mean square errors of each decoupling model are significantly reduced, wherein the back propagation neural network method is improved by 62.6%-82.9% in the three-dimensional force direction, and by 10.7%-38.5% in the moment direction; the support vector regression method is improved by more than 86% in the Mx and My directions, significantly alleviating the overfitting problem, and also has a reduction of 20%-47% in the force direction and the Mz direction; the extreme learning machine method achieves an error reduction of more than 64.0% in all directions, with an average improvement of about 80.8%.

[0131] (2) Inter-dimensional coupling error: reflects the decoupling accuracy of the algorithm in each measurement direction, defined as:

[0132]

[0133] The test results are shown in Table 2:

[0134] Table 2

[0135]

[0136]

[0137] As can be seen from Table 2, after the data enhancement method is processed, the inter-dimensional coupling errors of each decoupling model are significantly reduced, and the maximum improvement of the back propagation neural network method reaches 84%. The support vector regression method also significantly reduces the serious overfitting in the moment Mx and My directions after data enhancement, with a maximum improvement of more than 86%. The extreme learning machine method also improves significantly, with an overall accuracy improvement of more than 64.1% in each direction, and a maximum improvement of more than 80%.

[0138] In summary, the method of the present application improves the generalization ability of the nonlinear decoupling model under unknown loading conditions, and improves the measurement accuracy and stability of the six-dimensional force / torque sensor.

[0139] It should be noted that the above content only illustrates the technical idea of the present application, and cannot limit the protection scope of the present application. For ordinary skilled persons in the technical field, without departing from the principles of the present application, a number of improvements and refinements can be made, which fall within the protection scope of the claims of the present application.

Claims

1. A six-dimensional force / torque sensor calibration data enhancement method, characterized in that: The steps include: S1. Data Acquisition: A calibration experiment is performed on the six-dimensional force / torque sensor using a weight-weight loading method to obtain the sensor output data in real time. The calibration experiment is divided into a high-density calibration experiment group and a low-density calibration experiment group based on the step size. In the high-density calibration experiment group, the data vector consisting of the output data of the sensor at each loading point and the actual loading force value constitutes the training data; in the low-density calibration experiment group, the data vector consisting of the output data of the sensor at each loading point and the actual loading force value constitutes the test data. S2. Error-based adaptive interpolation: performing linear fitting on the training data obtained in step S1 to obtain a linear prediction output for each loading point in the training data, performing linear error analysis on the training data based on the linear prediction output to obtain an error bin number for each loading point, and finally performing cubic spline interpolation based on the error bin number to obtain expanded data; S3. Training and evaluation: The training data obtained in step S1 and the expanded data obtained after interpolation in step S2 are combined, and input into the six-dimensional force / torque decoupling model for weighted training. The model parameters are fixed to obtain the optimal decoupling model; the test data is input into the optimal decoupling model, and the difference between the model output and the actual loading force value is compared to obtain the measurement accuracy of the sensor, thereby confirming the realization of data enhancement.

2. The method for enhancing calibration data of a six-dimensional force / torque sensor according to claim 1, wherein: The calibration step length of the high-density calibration experimental group is not higher than one tenth of the range, and the calibration step length of the low-density calibration experimental group is not lower than one eighth of the range, and the calibration step length of the low-density calibration experimental group is larger than that of the high-density calibration experimental group; both experimental groups adopt the full-process calibration method to load and unload the weights; the output data of the sensor is a six-dimensional voltage analog quantity, and the actual loading force value is the actual loading six-dimensional force / torque value during the calibration process in each direction.

3. The six-dimensional force / torque sensor calibration data enhancement method according to claim 2, characterized in that: The weight loading process of the calibration test is specifically as follows: at the positive loading end of the calibration platform, the weight is gradually added from the zero state to the upper limit of the sensor range, and then unloaded to the zero state; at the negative loading end of the calibration platform, the weight is gradually added to the lower limit of the sensor range, and then unloaded to the zero state.

4. The method for enhancing calibration data of a six-dimensional force / torque sensor according to claim 1, wherein: The linear error analysis in step S2 specifically includes the following steps: S21: Calculate the linear fitting error at each loading point, the linear error e of the i-th loading point i for: Among them, y i is the output vector of the i-th loading point, which is the actual loading force value. is the linear prediction output of the i-th loading point; S22: Linear error set for loading points Sort in ascending order to obtain the sorted linear error sequence Where N is the total number of loading points in the training data; S23: Define percentile function P -1 (q): P -1 (q)=E[qN] Where q is the percentile probability between 0 and 1, [·] represents the rounding down operation, and E[qN] represents the [qN]th number in the linear error sequence E; S24: Perform percentile segmentation on the linear error sequence E obtained in step S22 to obtain n error boxes. The number of boxes n can be adjusted according to actual conditions. The error interval range corresponding to the j-th box is: Among them B j-1 ,B j They represent the lower and upper bounds of the error of the j-th box respectively; S25: Each loading point is classified into a corresponding box according to the size of its own linear error, and the corresponding error box number is obtained.

5. The method for enhancing calibration data of a six-dimensional force / torque sensor according to claim 4, wherein: The cubic spline interpolation in step S2 specifically includes the following steps: S26: Design an adaptive interpolation density based on the error bin number of each calibration data point and define the interpolation density multiple M i Specifically: Among them, α is the interpolation multiple, b j is the box number of data point i; S27: Determine the interpolation range at each loading point as half the distance between the current point and its adjacent point. The interpolation range at the endpoint is reduced by half. For the i-th loading point, the input vector of the loading point is used to define the interpolation range: where x i represents the input vector of the i-th loading point in the training data, i.e., the sensor output data; S28: Construct a cubic spline function S(x) from the input vector to the output vector on the training data. The function is a cubic polynomial in each interpolation interval, and satisfies the continuity of the function value at the interpolation node, and the function value is consistent with the actual output. At the same time, the first-order derivative and the second-order derivative are continuous at each node, and the boundary satisfies the natural boundary condition. Specifically: S(x i )=y i ,S(x i+1 )=y i+1 S(x1)=S(x N )=0 in and represents the first-order left derivative and first-order right derivative of the cubic spline function at the i-th loading point, and represents the second-order left derivative and second-order right derivative of the cubic spline function at the i-th loading point, S″(x1) and S″(x N ) represents the second-order derivative value of the cubic spline function at the first loading point and the last loading point; S29: At each loading point, according to the interpolation range and interpolation multiple M i Insert new points uniformly. For the i-th loading point, interpolate M i -1 new point, where the input vector x of the kth new point interp (k) and the output vector y interp (k) is: y interp (k)=S(x interp (k)) Where S(x interp (k)) is the cubic spline function obtained in step S28 at x interp The value at (k) is obtained, and all the new points obtained constitute the expanded data.

6. The method for enhancing calibration data of a six-dimensional force / torque sensor according to claim 1, wherein: The data set D for weighted training of the six-dimensional force / torque decoupling model in step S3 is specifically: where N total is the total number of data points, x i 、y i and w i are the input vector, output vector and weight of the i-th data point, w real is the training data weight, w aug ∈(0, 1) is the weight of the augmented data.

7. The method for enhancing calibration data of a six-dimensional force / torque sensor according to claim 6, wherein: The six-dimensional force / torque decoupling model in step S3, when inputting a decoupling model based on the support vector machine method, performs weighted training by adjusting the optimization objective function: Where w and b are model parameters, ξ, * is the slack variable, C is the penalty coefficient; When the decoupling model based on the back propagation neural network method is input, weighted training is achieved by modifying the loss function, and the loss function Loss is defined using the weighted mean square error: in is the predicted output vector of the current model; When the decoupling model based on the extreme learning machine method is input, the weighted operation is introduced in the step of solving the weight parameters of the output layer to construct the weight matrix W d , randomly initialize the input weight matrix W in and bias b, calculate the hidden layer output matrix H, and finally use the weighted least squares method to calculate the output layer weight W out : IN d =diag(w1,w2,...,w N ) H=φ(XW in +b) W out =(H T W d H) -1 H T W d Y where φ(·) represents the hidden layer activation function, diag(w1,w2,...,w N ) represents the diagonal matrix composed of the weights of the 1st to Nth data points, X is the input matrix of the training dataset, and Y is the reference output matrix of the training dataset.

8. A six-dimensional force / torque sensor calibration data enhancement system, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images