Six-dimensional force sensor calibration method and device

Abstract

The invention relates to the technical field of sensor calibration, and discloses a six-dimensional force sensor calibration method and device, and the method comprises the steps: constructing a conditional generative adversarial network model based on an extreme learning machine, and the model comprises a generator and a discriminator; determining structural parameters of the generator through an extreme learning machine pre-experiment; the voltage signal of the sensor and the generated force value obtained by the generator form a generated data pair, and the voltage signal of the sensor and the corresponding standard force value applied by the calibration instrument form a real data pair; the discriminator and the generator are alternately trained, training of the conditional generative adversarial network model is achieved, and structural parameters of the discriminator are selected through performance of the test set in the training process: voltage signals collected in real time are input into the trained conditional generative adversarial network model, and a calibrated force value is output. According to the invention, timely calibration of the six-dimensional force sensor in a dynamic working environment is ensured, so that measurement errors caused by external interference are effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of sensor calibration, and particularly relates to a six-dimensional force sensor calibration method and device. Background Art

[0002] Six-dimensional force sensors play an increasingly important role in modern robotics, automated production lines, medical equipment and other fields. Their precise force and torque measurement capabilities make them key components in various applications. However, due to environmental changes, material fatigue and long-term use, the performance of the sensors may drift. Therefore, regular calibration is necessary. Traditional calibration methods such as the least squares method and extreme learning machine are effective, but may have limitations in terms of model generalization ability, non-linear fitting, etc.

[0003] The least squares method is a global optimization algorithm that optimizes parameters by calculating the minimum of the sum of squared errors. If there are outliers in the experimental data, these outliers will have a significant impact on the calculation results, leading to the calibration results deviating from the true values. Especially in the calibration process of six-dimensional force sensors, due to the involvement of multiple measurement directions, outliers may more significantly affect the calibration of a certain dimension. The least squares method usually assumes that the relationship between the model and the data is linear. However, the output of an actual six-dimensional force sensor is often non-linear, especially under high loads or different loading conditions. For non-linear relationships, simple linear least squares methods may not be able to accurately fit the data, resulting in a decrease in calibration accuracy. Although this can be improved by transforming the data or using non-linear fitting methods, this will complicate the problem. The amount of data required by the least squares method is usually large, especially for a multi-dimensional sensor system. To obtain reliable calibration results, a wide enough input range and various working conditions need to be covered. This may be difficult to achieve in actual operation, especially during the experiment. If the conditions are limited or the experimental site is restricted, sufficient training data may not be obtained.

[0004] Extreme Learning Machine (ELM) is a novel machine learning algorithm. It randomly selects the parameters of the hidden layer and optimizes the output weights by the least squares method. Its structure is the same as that of a single-hidden-layer feedforward neural network. A key feature of Extreme Learning Machine (ELM) is that the weights (the weights connecting the input layer and the hidden layer) and biases of the hidden layer are randomly generated and remain unchanged during the training process. This is different from traditional neural networks, which usually adjust these weights during the training process. Since directly solving the output weights avoids iterative algorithms such as gradient descent, the ELM algorithm has a faster training speed compared to traditional neural networks, reducing the complex operations during the training process, and is especially suitable for processing large-scale data. In ELM, activation functions of the hidden layer can be selected such as sigmoid, ReLU, etc. ELM is insensitive to the distribution of input data and can better adapt to various data patterns. It has good generalization ability in many cases, that is, its performance on unseen data is often excellent. And the model is relatively simple and does not require complex hyperparameter tuning, allowing users to build and apply the model faster. However, due to the randomness of its initialization - the weights and biases of the hidden layer neurons are randomly generated, different random initializations may lead to differences in model performance, especially in the case of a small dataset or high complexity. At this time, the model may not be stable enough to ensure the optimal result can be obtained every time during training. In addition, ELM cannot naturally perform incremental learning during the training process. If real-time updates of new data are required, usually the entire model needs to be retrained instead of being updated step by step, which is not convenient when dealing with streaming data. Since the number of hidden layer nodes (the number of neurons) and the selection of activation functions have a greater impact on the final performance of the model, improper selection may lead to overfitting or underfitting of the model. Although ELM usually has strong robustness, in the case of a large amount of noisy data, random weight initialization may lead to higher errors.

[0005] Generative Adversarial Network (GAN) is a deep learning model proposed by Ian Goodfellow et al. in 2014. It consists of two main parts: the Generator and the Discriminator. The goal of the Generator is to generate data as realistic as possible, while the goal of the Discriminator is to distinguish between real data and generated data. The two are in a relationship of game and confrontation, and they continuously evolve during the confrontation process, each making the other party transform into a more perfect state, enabling the Generator to learn to generate more real data and the Discriminator to become a better discriminator.

[0006] As an emerging deep learning technology, Conditional Generative Adversarial Network (cGAN) provides a new idea for online calibration with its powerful generation ability and excellent feature learning ability. Summary of the Invention

[0007] To solve the above technical problems, the present invention provides a six - dimensional force sensor calibration method based on a conditional generative adversarial network (cGAN).

[0008] To solve the above technical problems, the present invention adopts the following technical solutions:

[0009] A six - dimensional force sensor calibration method, comprising:

[0010] Construct a conditional generative adversarial network model based on an extreme learning machine, including a generator and a discriminator;

[0011] Determine the structural parameter M of the generator through pre - experiments of the extreme learning machine;

[0012] Combine the voltage signal U of the sensor real and the generated force value F obtained by the generator based on U real to form a generated data pair, and combine the voltage signal U of the sensor gen and the corresponding standard force value F applied by the calibration instrument real to form a real data pair; standard

[0013] Alternately train the discriminator and the generator to realize the training of the conditional generative adversarial network model. During the training process, select the structural parameter N of the discriminator according to the performance of the test set: by minimizing the discriminator loss, make the discriminator learn to distinguish between generated data pairs and real data pairs; the discriminator loss includes the probability loss of judging real data pairs as true and the probability loss of judging generated data pairs as false; by minimizing the generator loss, update the parameters of the generator; the generator loss includes the adversarial loss based on the generated data pair and the reconstruction loss based on the generated force value F gen and the standard force value F standard ;

[0014] Input the real - time collected voltage signal into the trained conditional generative adversarial network model, and output the calibrated force value.

[0015] In one embodiment, the determining the structural parameter M of the generator through pre - experiments of the extreme learning machine specifically includes:

[0016] The generator adopts the same optimal structure as the extreme learning machine, with 6 neurons in each of the input layer and the output layer, corresponding to the six - dimensional voltage vector and the force vector respectively; the hidden layer is fully connected with M neurons, which is determined by pre - experiments of the extreme learning machine, that is, select the optimal number of neurons M according to the extreme learning machine method; the activation function is selected as the hyperbolic tangent function.

[0017] ​In one embodiment, the input of the discriminator is a 12-dimensional concatenated vector formed by concatenating a six-dimensional voltage vector and a six-dimensional force vector. After passing through a fully connected layer activated by ReLU, it is connected to a hidden layer with N neurons, and finally outputs a probability within the range of 0 to 1 through a Sigmoid activation function.

[0018] In one embodiment, the structural parameter N of the discriminator is selected based on the performance on the test set during the training process, specifically including:

[0019] For each candidate N value, perform the training of the conditional generative adversarial network model, and alternately optimize the parameters of the generator G and the discriminator D;

[0020] After each completion of the training of the conditional generative adversarial network model, use the test set to calculate the generator loss, and record the N value and the corresponding loss;

[0021] Select the N value that minimizes the generator loss of the test set as the final structural parameter of the discriminator.

[0022] In one embodiment, the discriminator loss includes the probability loss of judging the real data pair as true and the probability loss of judging the generated data pair as false, specifically including:

[0023]

[0024]

[0025] L D is the discriminator loss, is the probability loss of the discriminator judging the real data pair as true, represents the probability loss of judging the generated data pair as false, and E represents calculating the expectation.

[0026] In one embodiment, the parameters of the generator are updated by minimizing the generator loss; the generator loss includes the adversarial loss based on the generated data pair and the reconstruction loss based on the generated force value F gen and the standard force value F standard The specific details are as follows:

[0027]

[0028] Both λ1 and λ2 are weight coefficients, where λ1 = 0.2 and λ2 = 0.8. is the adversarial loss, and the goal is to maximize the confidence of the discriminator D in the generated data pair; is the reconstruction loss, and the difference between the generated force value F gen and the standard force value F standard is calculated through the mean square error to constrain the deviation between the generated force value and the standard force value.

[0029] A six - dimensional force sensor calibration device, comprising:

[0030] A model construction module that constructs a conditional generative adversarial network model based on an extreme learning machine, including a generator and a discriminator; determines the structural parameter M of the generator through pre - experiments of the extreme learning machine;

[0031] A data construction module that combines the voltage signal U of the sensor real and the generated force value F obtained by the generator based on U real to form a generated data pair, and combines the voltage signal U of the sensor gen and the corresponding standard force value F applied by the calibration instrument real to form a real data pair; standard

[0032] A training module that alternately trains the discriminator and the generator to implement the training of the conditional generative adversarial network model. During the training process, selects the structural parameter N of the discriminator through the performance of the test set: by minimizing the discriminator loss, enables the discriminator to learn to distinguish between generated data pairs and real data pairs; the discriminator loss includes the probability loss of judging real data pairs as true and the probability loss of judging generated data pairs as false; updates the parameters of the generator by minimizing the generator loss; the generator loss includes the adversarial loss based on the generated data pair and the reconstruction loss based on the generated force value F gen and the standard force value F standard ;

[0033] An online calibration module that inputs the real - time collected voltage signal into the trained conditional generative adversarial network model and outputs the calibrated force value.

[0034] In one embodiment, the determination of the structural parameter M of the generator through pre - experiments of the extreme learning machine specifically includes:

[0035] The generator adopts the same optimal structure as the extreme learning machine, with 6 neurons in each of the input layer and the output layer, corresponding to the six - dimensional voltage vector and the force vector respectively; the hidden layer is fully connected with M neurons, determined by pre - experiments of the extreme learning machine; the activation function selects the hyperbolic tangent function.

[0036] In one embodiment, the input of the discriminator is a 12 - dimensional concatenated vector formed by concatenating a six - dimensional voltage vector and a six - dimensional force vector. After passing through a fully - connected layer activated by ReLU, it is connected to a hidden layer with N neurons, and finally outputs a probability in the range of 0 to 1 through a Sigmoid activation function.

[0037] In one embodiment, the selection of the structural parameter N of the discriminator through the performance of the test set during the training process specifically includes: ​

[0038] For each candidate N value, perform conditional generative adversarial network model training, and alternately optimize the parameters of the generator G and the discriminator D;

[0039] After each completion of the training of the conditional generative adversarial network model, calculate the generator loss using the test set, and record the N value and the corresponding loss;

[0040] Select the N value that minimizes the generator loss of the test set as the final structural parameter of the discriminator.

[0041] The device of the present invention corresponds to the method, and the specific embodiment solutions applicable to the method are equally applicable to the device.

[0042] Compared with the prior art, the beneficial technical effects of the present invention are:

[0043] Compared with the least squares method and the extreme learning machine, ELM-Dc 2 The GAN model has the following advantages: First, the adversarial training process of the conditional generative adversarial network (cGAN) endows the generator G with stronger robustness. Through the adversarial optimization between the generator G and the discriminator D, the conditional generative adversarial network (cGAN) can effectively cope with the noise and outliers in the input data, thereby improving the stability and accuracy of the model in the six-dimensional force sensor calibration. Second, the M and N parameter optimization methods proposed in the present invention further improve the calibration accuracy. By optimizing these parameters, the ELM-Dc 2 The GAN model can more accurately capture the non-linear relationships in the data, improving the error control in the calibration process. In addition, the advantage of setting the hybrid loss function enables the model to take into account multiple objectives during the optimization process, ensuring that the generator G can not only maintain high quality when generating data, but also reduce the errors in the calibration process. Finally, the ELM-Dc 2 The prediction speed of the GAN model is one order of magnitude faster than that of the ELM, and it is more suitable for online calibration. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of the six-dimensional force sensor conditional generative adversarial network (cGAN) model in the embodiment of the present invention;

[0045] Figure 2 Schematic diagram of the network structure of the generator in the embodiment of the present invention;

[0046] Figure 3 Schematic diagram of the network structure of the discriminator in the embodiment of the present invention;

[0047] Figure 4 Schematic diagram of the relationship between the parameters of the discriminator and the test set loss in the embodiment of the present invention;

[0048] Figure 5 Schematic diagram of the relationship between the parameters of the discriminator and the average prediction time in the embodiments of the present invention;

[0049] Figure 6 Schematic diagram of the relationship between the number of neurons in the extreme learning machine and the loss of the test set in the embodiments of the present invention;

[0050] Figure 7 Schematic diagram of the relationship between the number of neurons in the extreme learning machine and the average prediction time in the embodiments of the present invention. Specific embodiments

[0051] A preferred embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] 1. Six-dimensional force / torque sensor calibration model of conditional generative adversarial network (cGAN):

[0053] Generative Adversarial Network (GAN) is a deep learning model proposed by Ian Goodfellow et al. in 2014. It consists of two main parts: a generator and a discriminator. The goal of the generator is to generate as realistic data as possible, while the goal of the discriminator is to distinguish between real data and generated data. The two are in a relationship of game and confrontation, and they continuously evolve during the confrontation process, each making the other party transform into a more perfect state, enabling the generator to learn to generate more real data and the discriminator to become a better discriminator. The loss function L GAN (G, D) of traditional GAN is:

[0054] L GAN (G, D) = E X [logD(x)] + E z [log(1 - D(G(z)))];

[0055] Among them, E[·] represents the operation of taking the expectation.

[0056] The mapping relationship of the generator can be expressed as:

[0057] G: z → x, that is, G(z) → x;

[0058] Among them, z is a random vector from the noise distribution. G(z) is the sample generated by the generator.

[0059] The mapping relationship of the discriminator is:

[0060] D: x → [0, 1], D(x) → 1, D(G(z)) → 0;

[0061] D(x) represents the probability that the sample x comes from the real data distribution, and D(G(z)) represents the probability that the sample comes from the generated data distribution.

[0062] The goal of the generator is to maximize the probability that the discriminator wrongly believes that the generated data comes from the real distribution and minimize this loss function. The goal of the discriminator is to minimize the error probability of discriminating real data and the error probability of discriminating generated data and maximize this loss function.

[0063] Supervised learning refers to a machine learning method of training a model through a labeled data set. By introducing the method of supervised learning into GAN and using the label items of the labeled data set as the input for model training, researchers Mehdi and Simon proposed the concept of conditional generative adversarial network (cGAN) based on GAN. The loss function L cGSN (G, D) is as follows:

[0064] L cGAN (G, D) = E X,Y [logD(x|y)] + E Z,Y [log(1 - D(X, G(z|y)))];

[0065] The special thing is that x represents the condition input to the generator or discriminator.

[0066] The mapping relationship of the generator can be expressed as: G:(z|y)→x, that is, G(z|y)→x, where z is a random vector from the noise distribution. x = G(z) is the sample generated by the generator. The mapping relationship of the discriminator is:

[0067] D:(x, y)→[0, 1], D(x|y)→1, D(x, G(z|y)→0;

[0068] D(x, y) represents the probability that the sample x comes from the real data distribution, and D(x, G(z|y) represents the probability that the sample comes from the generated data distribution.

[0069] The present invention establishes a conditional generative adversarial network (cGAN) model for a six-axis force sensor, as Figure 1 shown. In view of the deterministic mapping requirement for the calibration task of the six-axis force sensor, the present invention makes the following key improvements to the traditional conditional generative adversarial network (cGAN) framework.

[0070] (1) Deterministic conditional generation mechanism:

[0071] The present invention abandons the random noise input used in the traditional conditional generative adversarial network (cGAN) and uses the original voltage signal of the sensor As a conditional input to the generator G, the calibration process's deterministic mapping is ensured by directly utilizing the raw sensor voltage signal: F gen = G(U real ).

[0072] In practical applications, the raw voltage signal already contains various interference noises, which are reflected in the voltage signal. Therefore, there is no need to additionally introduce random noise. In contrast, adding random noise may not only fail to provide useful information but may also increase the complexity of the non-linear mapping. The non-linear characteristics caused by the noise contained in the raw voltage signal can be effectively identified and processed through the capabilities of the non-linear model, thereby further improving the robustness and accuracy of the model. Therefore, using the raw voltage signal as the input can ensure a more stable and efficient online calibration process without introducing additional noise.

[0073] (2) Data pair discrimination mechanism:

[0074] The generator predicts the corresponding force value F real based on U gen , and this force value and the raw voltage input U real are jointly spliced to form the generated data pair (F gen , U real ) for simulating the real data distribution. The real data pair consists of the force F standard applied by the calibration instrument and its corresponding voltage input U real (F standard , U real ), representing the ideal input-output relationship.

[0075] These two data pairs share the same voltage input, aiming to ensure the consistency of the voltage signal during the model training process. Due to the consistent voltage input, the discriminator only needs to judge whether the input force value F gen or F standard matches the raw voltage input U real , thus simplifying the training process, making it easier to distinguish between generated data and real data, and improving the discriminator's recognition ability of the relationship between the force value and the voltage input.

[0076] (3) Design of the hybrid loss function:

[0077] During the training process, the discriminator D learns to distinguish these two types of data by minimizing its loss function L D : For the real data pair (F standard , U real ), the discriminator D expects the output probability P real to be close to 1, and the probability loss for giving these data as real is For the generated data pair (F gen , Ureal ) The discriminator D expects the output probability P fake to be close to 0, giving the probability loss for these data being real By calculating these two loss functions, the discriminator can more accurately measure the difference between the two types of data and effectively distinguish real and generated data.

[0078]

[0079] The generator G updates its parameters by minimizing its loss function L G where λ1 = 0.2 and λ2 = 0.8. is the adversarial loss, which is the main part of the generator loss. The goal is to maximize the confidence of the discriminator D in the generated data pairs, that is, it is hoped that the discriminator D can regard (F gen , U real ) as real data, forcing the generated data distribution to approximate the real physical response and solving the problem of insufficient modeling of non - linear coupling by traditional linear methods; is the reconstruction loss. The generated force F gen is calculated with the standard force F standard through the mean square error (MSE) to calculate the difference between them, restricting the deviation of the generated force value from the measured value of the calibration instrument to ensure that the generated force value approaches the real output.

[0080]

[0081] 2. Six - dimensional force sensor calibration algorithm based on ELM - Dc 2 GAN:

[0082] In the ELM - Dc 2 GAN (ELM - Optimized Dual - Constraint cGAN) calibration algorithm proposed by the present invention, ELM represents pre - experimental optimization of the model through the extreme learning machine (ELM) to determine the structural parameter M of the generator G. After the generator parameters are determined, the structural parameter N of the discriminator D is selected through the performance of the test set. Here, c not only represents the use of the conditional generative adversarial network (cGAN) method, but also symbolizes the dual - parameter optimization process of the generator and the discriminator. The specific parameter selection method is as follows:

[0083] (1) As Figure 2 shown, the generator G adopts the same optimal structure as the extreme learning machine (ELM): 6 neurons in each of the input / output layers (corresponding to the six - dimensional voltage and force vectors), and the hidden layer is fully connected with M neurons (determined by ELM pre - experiment), M = m = 39, and the activation function is selected as Tanh. The determination process of M is as follows:

[0084] The hidden layer of the extreme learning machine has m neurons, and the value of m can be adjusted according to the task requirements. The optimal number of hidden layer neurons is selected based on the performance of ELM on the test set. For the corresponding number of neurons, the changes in the loss on the test set and the average prediction time of each data are as follows Figure 6 and Figure 7 shown. By synthesizing the performance of both, it can be seen that when m = 39, the mean square error (MSE) is 0.3682, and the performance of the model is optimal.

[0085] The discriminator D takes a 12-dimensional concatenated vector (voltage vector U and six-dimensional force vector F) as input. After passing through a fully connected layer activated by ReLU, it is connected to a hidden layer with N neurons (dynamically adjustable). Finally, it outputs a probability in the range of 0 to 1 through the Sigmoid activation function.

[0086] (2) As Figure 3 shown, in the conditional generative adversarial network (cGAN), the discriminative ability of the discriminator should match that of the generator to ensure the effective training of the model. Therefore, the number of neurons N in the fully connected layer of the discriminator D has a significant impact on the training effect. To avoid overfitting or underfitting of the discriminator, thereby reducing the generalization ability of the model, after the model is trained, the test set data is used for systematic evaluation to determine the optimal N value, as Figure 4 and Figure 5 shown. By comparing the training results under different N values, it is found that when N = 3, the maximum, minimum, and average values of the generator loss L G on the validation set are all shown to be the smallest overall. This indicates that under this structure, the discriminator can more effectively balance the adversarial relationship between the generator and the discriminator, thereby improving the stability and robustness of the model. Therefore, N = 3 is selected as the optimal structure of the discriminator to optimize the training effect of the model and enhance its generalization ability.

[0087] During the actual training process, G and D alternately learn and update parameters. The pseudocode of the ELM-Dc 2 GAN model is shown in Table 1.

[0088] Table 1 Pseudocode implementation of the ELM-Dc 2 GAN model

[0089]

[0090]

[0091] The parameters of this algorithm are shown in Table 2.

[0092] Table 2 Parameters of the ELM-Dc 2 GAN model

[0093]

[0094]

[0095] 3. Calculation Indexes of Online Calibration Algorithm

[0096] The errors of the six - axis force sensor are mainly divided into two categories: systematic errors and random errors. Systematic errors stem from the characteristics of the sensor itself or changes in external environmental conditions, mainly coupling errors. Such errors usually occur between different measurement dimensions of the sensor. Due to unreasonable sensor structure design or the interaction between sensor components, a force applied in one direction may affect the strain measurement in other directions. Specifically, the design layout of the sensor, the installation position of the strain gauges, and the physical properties of the materials will all affect the measurement accuracy. For example, if the strain gauges are not accurately attached to the strain - concentrated area, or the bonding quality is poor due to improper surface treatment, it may lead to deviations in the measurement results. In addition, environmental conditions such as temperature changes, humidity, and changes in the physical properties of materials after long - term stress will also affect the performance of the sensor. Random errors are measurement errors caused by external environmental factors or random fluctuations within the system, usually manifested as fluctuations in measurement data. Their sources can be noise in electrical signals, including power supply noise and external electromagnetic interference, etc. In addition, factors such as vibration, shock, and temperature fluctuations in the environment where the sensor is located will also cause fluctuations in the measurement results. Incorrect operations by operators, such as inaccurate readings or incorrect equipment configurations, may also be one of the reasons for random errors. Therefore, precise operation and stable environmental conditions are crucial for reducing random errors.

[0097] The main coupling errors of the six - axis force sensor usually exhibit non - linear characteristics. Therefore, it is necessary to adopt a non - linear fitting algorithm in the calibration calculation. The non - linear calibration algorithm adopted in the present invention can more accurately reflect the performance of the sensor in actual applications. However, traditional performance indicators such as linearity, hysteresis, and repeatability are mainly evaluated based on the linear relationship fitted by the least - squares method, which makes them unable to effectively reflect the complexity of non - linear coupling errors and their impact on measurement results. Therefore, in the present invention, the method of selecting type - I error and type - II error is used to calculate non - linear indicators, which can comprehensively characterize the influence of systematic errors and random errors. In order to more comprehensively evaluate the working performance of the sensor in the non - linear field. In this way, the influence of non - linear coupling errors on the performance of the six - axis force sensor can be understood and quantified more clearly.

[0098] One type of error reflects the deviation between the measured value of a certain dimension of a multi-dimensional force sensor and the actually applied value, which affects the accuracy of the sensor and the reliability of the readings. The occurrence of this type of error is often caused by the limitations of the device itself, the non-linear response of the sensor, or environmental factors (such as temperature, humidity), etc. This indicates that this error has a certain regularity, and all measurements will deviate from the true value in a certain direction. In practical applications, it is necessary to control and correct these errors to ensure that the measurement results of the force sensor are close to the actual value. One type of error characterizes the systematic deviation of the i-th component (such as F x 、M y etc.) at its measurement point, reflecting the average deviation between the measured value and the theoretical value of this component under ideal one-dimensional loading conditions.

[0099] Calculate the average value of the test results of the i-th component at the k-th measurement point, and the systematic error of this component at the k-th measurement point is obtained as:

[0100]

[0101] In the formula: F ik (real) represents the actually applied force or torque value of the i-th component at the k-th test point, that is, the standard value measured by the one-dimensional force sensor; F ik (measured) represents the average value of the measured force or torque of the i-th component at the k-th test point, that is, the value measured by the six-dimensional force sensor; F i (full range) represents the full range of the six-dimensional force sensor in the i-th component.

[0102] Calculate the one type of error of the i-th component of the system using Bessel's formula as:

[0103]

[0104] In the formula: n is the number of test points, and k is the ordinal number of the test point.

[0105] The one type of comprehensive error of the system is:

[0106]

[0107] In the formula: i is the ordinal number of the test component.

[0108] Two types of errors reflect the magnitude of the interference on a certain dimension when no force / torque is applied to that dimension due to the application of force / torque on other dimensions, and are also called coupling errors. This type of error is usually the result of the interaction of various factors (including structural factors and environmental factors) in a complex system. Two types of errors characterize the coupling interference caused by the loading of other components (j≠i) at the measurement point of the i-th component, reflecting the multi-dimensional mechanical coupling characteristics.

[0109] Calculate the coupling error of the i-th component under the force or moment of the j-th (j≠i) component at the k-th measurement point:

[0110]

[0111] Where: F ijk (measured) is the average coupling error of the i-th component under the force or moment of the j-th component at the k-th measurement point.

[0112] Use Bessel's formula to calculate the coupling error D of the i-th component under the force or moment of the j-th component ij and the secondary error D of the i-th component i :

[0113]

[0114] Where: n is the number of test points; k is the test point ordinal number.

[0115]

[0116] The secondary comprehensive error of the system is:

[0117]

[0118] 4. Comparative Analysis of Calibration Algorithms

[0119] To verify the performance of the ELM-Dc 2 GAN model, the least squares method (OLS) and the extreme learning machine (ELM) method are compared with it, and the validation set data is used for verification. According to the given situation, the mean square error matrix is calculated as a 6th-order square matrix. The calibration error matrices, primary errors, and secondary errors of these three methods are shown in Table 3.

[0120] Table 3 Comparison of Calibration Algorithm Errors

[0121]

[0122] Nonlinear algorithms have more powerful processing capabilities than linear algorithms, especially suitable for processing data with nonlinear relationships. Through the data set, nonlinear algorithms can capture the complex relationships between data more comprehensively, especially having a better fitting effect on those nonlinear and asymmetric error dimensions. This advantage of being able to fit asymmetric features makes nonlinear methods perform better when dealing with complex data sets.

[0123] As can be seen from Table 3, on the premise of using the data obtained from the calibration experiment for calibration calculation, the primary error of the ELM method is slightly better than that of the least squares method, and the secondary error is slightly worse, while the ELM-Dc 2Compared with the least squares method, the error of one type of the GAN algorithm model is reduced by 40.91% to 0.182264%, and the error of the second type is reduced by 59.91% to 0.129102%. ELM-Dc 2 The G model structure used for prediction in GAN is the same as that of ELM, but the performance has been greatly improved. This is because in complex non-linear data relationships, ELM-Dc 2 The generation ability of GAN enables the model to flexibly adapt to different data distributions and structures, especially when the data volume is limited or the data distribution is uneven. ELM-Dc 2 The adaptive ability shown by GAN significantly improves its prediction effect, and the prediction time is one order of magnitude faster than the ELM method.

[0124] By comparing the error results of the least squares method in laboratory calibration and robotic arm calibration, the present invention comprehensively demonstrates the importance of online calibration in practical applications. Although the traditional least squares method has excellent performance in mathematics, in actual operation, especially in complex environments, it is easily affected by noise and non-linear factors, resulting in limitations in the accuracy of its calibration results. The online calibration method can realize real-time data acquisition and processing, ensure timely calibration of the six-axis force sensor in a dynamic working environment, and thus effectively reduce the measurement error caused by external interference.

[0125] In addition, the ELM-Dc proposed in the present invention 2 The optimization effect of the GAN model is compared and analyzed in detail with the classical linear calibration method (least squares method) and non-linear method (extreme learning machine ELM). Experiments prove that the ELM-Dc proposed in the present invention 2 The GAN algorithm shows superior performance in dealing with complex non-linear data features. Compared with traditional methods, ELM-Dc 2 GAN can more accurately fit the input-output relationship, thus significantly reducing the error of the six-axis force sensor during actual use and improving the measurement accuracy.

[0126] By comprehensively comparing the error results of these methods, it can be clearly recognized that choosing an appropriate calibration technology is important for improving the working performance of the six-axis force sensor. Online calibration not only overcomes the deficiencies of traditional calibration methods in consistency and timeliness, but also can perform adaptive adjustment combined with real-time feedback, improving the flexibility and reliability of the entire calibration process. Therefore, in modern high-precision measurement applications, using the online calibration method of the ELM-Dc 2 GAN non-linear algorithm will significantly enhance the overall performance of the six-axis force sensor and its related systems, providing a new solution for its application in complex scenarios.

[0127] It should be understood that although the steps in the flowchart of the accompanying drawings of the specification are shown sequentially according to the indication of the arrows, these steps are not necessarily executed sequentially in the order indicated by the arrows. Unless there is a clear description in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings of the specification may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0128] Based on the description of the above method embodiments, the present disclosure also provides an apparatus. The apparatus may be a system (including a distributed system), software (application), module, component, server, client, etc. that uses the method described in the embodiments of this specification and combines the necessary implementation hardware. Based on the same inventive concept, the apparatus in one or more embodiments provided by the embodiments of the present disclosure is as described in the following embodiments. Since the implementation solutions for the apparatus to solve problems are similar to those of the method, the implementation of the specific apparatus in the embodiments of this specification may refer to the implementation of the foregoing method, and repeated parts will not be described again. The term "module" or "modular" used hereinafter may be a combination of software and / or hardware capable of implementing a predetermined function. Although the systems described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0129] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as falling within the scope described in this specification.

[0130] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, it is intended to embrace all changes within the meaning and scope of the equivalent elements of the claims in the present invention, and any reference signs in the claims should not be regarded as limiting the claimed rights.

[0131] In addition, it should be understood that although this specification is described in terms of embodiments, not every embodiment contains only one independent technical solution. This narrative style of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A six-dimensional force sensor calibration method, characterized in that: include: Construct a conditional generative adversarial network model based on extreme learning machine, including generator and discriminator; Determine the structural parameter M of the generator through the extreme learning machine pre-experiment; The voltage signal U real and generator based on U real The resulting force value F gen The sensor voltage signal U real and the corresponding standard force F applied by the calibration instrument standard Form real data pairs; The discriminator and generator are trained alternately to realize the training of the conditional generative adversarial network model. During the training process, the structural parameter N of the discriminator is selected according to the performance of the test set: by minimizing the discriminator loss, the discriminator learns to distinguish between generated data pairs and real data pairs; the discriminator loss includes the probability loss of judging the real data pairs as true, and the probability loss of judging the generated data pairs as false; by minimizing the generator loss, the parameters of the generator are updated; the generator loss includes the adversarial loss based on the generated data pairs, and the loss based on the generated force value F gen And standard force value F standard The reconstruction loss of The real-time collected voltage signal is input into the trained conditional generative adversarial network model to output the calibrated force value.

2. A six-dimensional force sensor calibration method according to claim 1, characterized in that: The structural parameter M of the generator is determined by the extreme learning machine pre-experiment, specifically including: The generator adopts the same optimal structure as the extreme learning machine, with 6 neurons in the input layer and the output layer, corresponding to the six-dimensional voltage vector and force vector respectively; the hidden layer is M neurons fully connected, determined by the extreme learning machine preliminary experiment; the activation function uses the hyperbolic tangent function.

3. A six-dimensional force sensor calibration method according to claim 1, characterized in that: The input of the discriminator is a 12-dimensional concatenated vector composed of a six-dimensional voltage vector and a six-dimensional force vector. After passing through a fully connected layer activated by ReLU, it is connected to a hidden layer with N neurons, and finally outputs a probability in the range of 0 to 1 through a Sigmoid activation function.

4. A six-dimensional force sensor calibration method according to claim 1, characterized in that: The structural parameter N of the discriminator is selected according to the performance of the test set during the training process, specifically including: For each candidate N value, perform conditional generative adversarial network model training, alternately optimizing the parameters of the generator G and the discriminator D; After each training of the conditional generative adversarial network model, the generator loss is calculated using the test set, and the N value and the corresponding loss are recorded; The N value that minimizes the generator loss of the test set is selected as the final structural parameter of the discriminator.

5. A six-dimensional force sensor calibration method according to claim 1, characterized in that: The discriminator loss includes the probability loss of judging the real data pair as true, and the probability loss of judging the generated data pair as false, specifically including: L D is the discriminator loss, is the probability loss of the discriminator judging the real data pair as true, represents the probability loss of judging the generated data pair as false, and F represents the calculation expectation.

6. A six-dimensional force sensor calibration method according to claim 1, characterized in that: The generator parameters are updated by minimizing the generator loss; the generator loss includes the adversarial loss based on the generated data pair and the generated force value F gen And standard force value F standard The reconstruction loss includes: λ1,λ2 are weight coefficients, where λ1=0.2,λ2=0.8, For adversarial loss, the goal is to maximize the confidence of the discriminator D on the generated data pairs; For the reconstruction loss, the force value F is generated by the mean square error calculation gen With standard force value F standard The difference between them constrains the deviation of the generated force value from the standard force value.

7. A six-dimensional force sensor calibration device, characterized in that: include: Model building module, which builds a conditional generative adversarial network model based on extreme learning machine, including generator and discriminator; Determine the structural parameter M of the generator through the extreme learning machine pre-experiment; The data building module converts the voltage signal U real and generator based on U real The resulting force value F gen The sensor voltage signal U real and the corresponding standard force F applied by the calibration instrument standard Form real data pairs; The training module alternately trains the discriminator and the generator to realize the training of the conditional generative adversarial network model. During the training process, the structural parameter N of the discriminator is selected according to the performance of the test set: by minimizing the discriminator loss, the discriminator learns to distinguish between generated data pairs and real data pairs; the discriminator loss includes the probability loss of judging the real data pairs as true, and the probability loss of judging the generated data pairs as false; by minimizing the generator loss, the parameters of the generator are updated; the generator loss includes the adversarial loss based on the generated data pairs, and the loss based on the generated force value F gen And standard force value F standard The reconstruction loss of The online calibration module inputs the real-time collected voltage signal into the trained conditional generative adversarial network model and outputs the calibrated force value.

8. A six-dimensional force sensor calibration device according to claim 7, characterized in that: The structural parameter M of the generator is determined by the extreme learning machine pre-experiment, specifically including: The generator adopts the same optimal structure as the extreme learning machine, with 6 neurons in the input layer and the output layer, corresponding to the six-dimensional voltage vector and force vector respectively; the hidden layer is M neurons fully connected, determined by the extreme learning machine preliminary experiment; the activation function uses the hyperbolic tangent function.

9. A six-dimensional force sensor calibration device according to claim 7, characterized in that: The input of the discriminator is a 12-dimensional concatenated vector composed of a six-dimensional voltage vector and a six-dimensional force vector. After passing through a fully connected layer activated by ReLU, it is connected to a hidden layer with N neurons, and finally outputs a probability in the range of 0 to 1 through a Sigmoid activation function.

10. A six-dimensional force sensor calibration device according to claim 7, characterized in that: The structural parameter N of the discriminator is selected according to the performance of the test set during the training process, specifically including: For each candidate N value, perform conditional generative adversarial network model training, alternately optimizing the parameters of the generator G and the discriminator D; After each training of the conditional generative adversarial network model, the generator loss is calculated using the test set, and the N value and the corresponding loss are recorded; The N value that minimizes the generator loss of the test set is selected as the final structural parameter of the discriminator.

Images