A multi-dimensional force sensor decoupling method based on PINN method and multi-task learning

Abstract

The present invention discloses a multi-dimensional force sensor decoupling method based on the PINN method and multi-task learning, and the method relies on a multi-task learning decoupling model. The present invention can improve the generalization ability of the model on the basis of the advantages of the neural network decoupling method; by jointly training the model on multiple tasks, the risk of overfitting can be reduced, the model can be made more robust, and the decoupling accuracy under various load types can be comprehensively improved. In the training process of the multi-task learning decoupling model, the theoretical decoupling matrix and the optimal decoupling matrix are obtained by using the least squares method and the PINN method, and the obtained optimal decoupling matrix is ​​used as the initial parameter of the multi-task learning neural network, which greatly improves the training speed while compensating for the loading system error and improving the decoupling accuracy. By jointly training the model on multiple tasks, the present invention can calibrate the same sensor under different working conditions, so as to realize the applicability of a single sensor to multiple working conditions.

Description

Technical Field

[0001] The present invention belongs to the field of control engineering and sensor technology, and specifically relates to a multi-dimensional force sensor decoupling method based on a PINN method and multi-task learning. Background Art

[0002] At present, multi-dimensional force sensors are widely used in humanoid robots based on force touch, automobiles, aerospace, and other fields. The research on multi-dimensional force sensors with different structural types used in different fields has never stopped. From the optimization of the sensor elastomer structure to the packaging process of the flange and the shell, the selection and pasting process of the sensitive components, to the personalized design of the decoupling algorithm and the calibration device, they are all research and development difficulties. Compared with single-axis sensors, multi-dimensional force sensors have obvious defects in accuracy due to the existence of coupling effects. In the face of new demands for the use scenarios of multi-dimensional force sensors in today's industry, aerospace and navigation fields, miniaturization / heavy duty, integrated intelligence, bionics, and dynamic response of sensors are all challenges and opportunities facing multi-dimensional force sensors.

[0003] In addition, there are several types of problems with the sensor itself: For example, the coupling problem: the force and torque in the multi-dimensional force sensor often have a coupling relationship, that is, the different dimensions affect each other. The applied torque may affect the force measurement result, and the measured force may also be disturbed by the torque. Therefore, how to accurately decouple the force and torque becomes an important challenge; in practical applications, multi-dimensional force sensors often face complex environmental conditions, such as background noise, nonlinear systems, and deformation and stiffness changes of objects. These factors may cause inaccuracy and instability of sensor data, increasing the complexity and challenge of the decoupling algorithm; after obtaining the data, the raw data generated by the multi-dimensional force sensor often contains noise and interference; in the calibration and measurement process, there is an unavoidable eccentric load phenomenon in the loading system. In the decoupling algorithm, effective data preprocessing is required, including filtering, denoising and data calibration. How to choose an appropriate data preprocessing method and maintain the accuracy of force and torque during the preprocessing process is also a problem that needs to be solved; and this series of problems will further affect each other during the decoupling process.

[0004] In summary, the high applicability decoupling of multi-dimensional force sensors has always been a major problem that has troubled the engineering community. Both physical model decoupling and mathematical model decoupling have great limitations. In recent years, decoupling methods based on machine learning have also usually treated them as black boxes for rough decoupling attempts. With the development of industry and the continuous improvement of manufacturing technology, the demand for high-precision, high-stability, and high-applicability multi-dimensional force sensor decoupling methods is also increasing urgently.

[0005] In summary, there is still a lack of a comprehensive solution that can simultaneously address challenges in various fields, thereby achieving a comprehensive solution to all aspects of multi-dimensional force sensor decoupling problems. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a multi-dimensional force sensor decoupling method based on a PINN method and multi-task learning, thereby achieving high-precision and high environmental applicability decoupling of the multi-dimensional force sensor.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A multi-dimensional force sensor decoupling method based on a PINN method and multi-task learning comprises the following steps:

[0009] 1) Obtain the electrical signal data of the multi-dimensional force sensor strain gauge by performing a uniaxial experiment on the loading platform;

[0010] 2) normalizing the data collected in step 1);

[0011] 3) Using the PINN method, combined with the uniaxial experimental data of the loading platform processed in step 2), the multi-task learning decoupling model is calibrated; in this step, the optimal decoupling matrix obtained based on the PINN method is used to calibrate some fixed parameters in the multi-task learning decoupling model, specifically replacing the optimal decoupling matrix obtained during the model training process, and then decoupling; the purpose of this step is to compensate for the system error in the experimental loading platform. The sources of error include multi-dimensional force sensor assembly error and loading platform error; the PINN method in this step needs to be retrained in different loading systems.

[0012] 4) A loading experiment of complex combined working conditions is carried out on the loading platform to obtain the electrical signal data of the strain gauge of the multi-dimensional force sensor, and the normalized electrical signal data is decoupled based on the calibrated multi-task learning decoupling model to obtain the load information.

[0013] In the above technical solution, further, the construction process of the multi-task learning decoupling model is:

[0014] Establish a mathematical model according to the sensor deformable body structure; construct a neural network decoupling structure based on the mathematical model, wherein the neural network decoupling structure is a multi-task learning shared layer; add a multi-task learning non-shared layer input module and a multi-task learning non-shared layer output module to the input part and the output part of the neural network decoupling structure respectively, and set the number of channels according to the task type;

[0015] The training process of the multi-task learning decoupling model is:

[0016] (1) Data acquisition: For multiple groups of different working conditions, the sensor load and strain gauge strain data are obtained through finite element simulation; the sensor load information and the electrical signal data of the strain gauge are obtained through experiments; wherein the experiments include uniaxial loading experiments and loading experiments under complex combined working conditions or complex environments.

[0017] (2) normalizing the data collected in step (1);

[0018] (3) Based on the mathematical model, the theoretical decoupling matrix C is solved according to the least squares method and the finite element simulation data processed in step (2) FEM ;

[0019] (4) Using the PINN method, the theoretical decoupling matrix C is calculated by combining the uniaxial loading experimental data processed in step (2) FEM Optimize and get the optimal decoupling matrix C PINN The optimal decoupling matrix can eliminate the interference of system errors in the calibration experiment and is sufficient to meet the normal decoupling requirements of multi-dimensional force sensors.

[0020] (5) Preliminary optimization of the neural network decoupling structure based on the elements in the optimal decoupling matrix;

[0021] (6) Further optimizing the neural network decoupling structure and its parameters, the multi-task learning non-shared layer input module structure and its parameters, and the multi-task learning non-shared layer output module structure and its parameters through the loading test data under the complex combination working condition or complex environment processed by step (2), and obtaining a trained multi-task learning decoupling model. The multi-task learning decoupling model can specifically improve the highly adaptable decoupling of complex combination working conditions.

[0022] Furthermore, in step 3), the multi-task learning decoupling model is calibrated by using the PINN method in combination with the uniaxial experimental data of the loading platform processed in step 2); the specific method is:

[0023] (1) Obtain the sensor load and strain gauge strain data through finite element simulation;

[0024] (2) normalizing the data collected in step (1);

[0025] (3) based on the mathematical model corresponding to the loading platform, solving the theoretical decoupling matrix according to the least squares method and the finite element simulation data obtained in step (2);

[0026] (4) using the PINN method and combining the uniaxial experimental data of the loading platform processed in step 2) to optimize the theoretical decoupling matrix to obtain the optimal decoupling matrix;

[0027] (5) Using the optimal decoupling matrix to replace the optimal decoupling matrix in the multi-task learning decoupling model.

[0028] Furthermore, the normalization process is used to convert the data into a specific range or standardized form, thereby eliminating the dimensional differences between different features.

[0029] Furthermore, in step (3), based on the mathematical model, the decoupling matrix is ​​solved according to the least square method and the data obtained in step (2). The decoupling matrix C FEM Specifically:

[0030] C FEM =F FEM X FEM T (X FEM X FEM T ) -1

[0031] in:

[0032]

[0033]

[0034] Among them, F x 、F y 、F z 、M x 、M y 、M z They are the axial force in the x direction, the axial force in the y direction, the axial force in the z direction, the moment in the x direction, the moment in the y direction, and the moment in the z direction. 1 , X 2 , X 3 , X 4 , X 5 , X 6 According to F x 、F y 、F z 、M x 、M y 、M z The electrical signal combination of the corresponding strain gauges obtained at the deformation sensitive position of the sensor deformation body when uniaxially loaded in the six strain sensitive directions, k 1 , k 2 , k 3 , k 4 , k 5 , k 6 are all constants, obtained by least squares fitting; s 1 、s 2 、s 3 ...

[0035] s 16 They are the strains at the corresponding positions of the 16 strain gauges obtained through finite element simulation.

[0036] Furthermore, in step (4), the PINN method is used to combine the uniaxial experimental data processed in step (2) to calculate the theoretical decoupling matrix C FEM Optimize and get the optimal decoupling matrix C PINN ; The formula is as follows:

[0037] C PINN =λ 1 ·λ 2 ·λ 3 ·C FEM ·D -1

[0038] D.F. Expt =λ 1 ·λ 2 ·λ 3 C FEM ·X Expt

[0039]

[0040] Among them, F Expt and X Expt are the load information and telecommunication signal information in the loading experiment respectively; C FEM is the theoretical decoupling matrix obtained in step (3); D represents the combination of the rotation matrix R, which can rotate the load direction; R is a 3×3 rotation matrix, which represents the deflection of the load; λ 1 , 2 , 3 represents the correction coefficient matrix of size 6×6, 6×n, n×m, and m×6; n and m represent the number of rows and columns of the intermediate matrix, which are the parameters to be determined. The theoretical decoupling matrix is ​​optimized using the physical constraint neural network (PINN) method. This process does not require complex loading conditions, but only single or a few uniaxial loadings.

[0041] Furthermore, the loss function of the PINN method consists of two parts:

[0042] Loss = α 1 Loss physics +α 2 Loss data

[0043] Among them, Loss is the total loss, α 1 , α 2 is the trade-off between physical constraints and Loss physicsand data fitting term Loss data The hyperparameter is used to adjust the importance of the two parts of the loss in the overall loss.

[0044] Data fitting term Loss data : It is used to ensure that the decoupling matrix can accurately approximate the observed data, which is specifically expressed as:

[0045] Loss data =||D·F Expt -λ 1 ·λ 2 ·λ 3 C FEM ·X Expt || 2

[0046] Physical Constraint Loss physics : In the common error analysis of multi-dimensional force sensors, eccentric load is the main source of error and is not easy to compensate, so the eccentric load is optimized as a compensation coefficient; since D is a combination of the rotation matrix R, it also satisfies the properties of the rotation matrix, so the orthogonality of the rotation matrix is ​​used as a physical constraint term, which is specifically expressed as:

[0047] Loss physics =||ID T D|| 2

[0048] Where I is the identity matrix.

[0049] Therefore, the present invention solves the correction coefficient matrix D, λ by the PINN method. 1 , 2 , 3 , we can get the optimal decoupling matrix C PINN .

[0050] Furthermore, during the training process of the multi-task learning model, the parameters of the multi-task learning non-shared layer input module and the multi-task learning non-shared layer output module are independent of the multi-task learning shared layer, and the parameters of each channel of the multi-task learning non-shared layer input module and the multi-task learning non-shared layer output module are constrained by their respective loss functions and by the total loss function during the training process.

[0051] The present invention also provides a multi-dimensional force sensor decoupling device based on the PINN method and multi-task learning, which includes a data acquisition module, a data normalization processing module, a PINN optimization module and a multi-task learning decoupling module.

[0052] The data acquisition module is used to collect electrical signal data of the strain gauge when the loading platform performs a loading experiment.

[0053] The data normalization processing module is used to normalize the data collected by the data acquisition module, with the purpose of converting the data into a specific range or standardized form, thereby eliminating the dimensional differences between different features. At the same time, this can also avoid the instability caused by the eigenvalue being too large or too small, thereby accelerating the convergence speed and optimization process of the model.

[0054] The PINN optimization module is used to calibrate some fixed parameters in the multi-task learning decoupling module, specifically by pre-loading the optimal decoupling matrix in the calculation loading system and replacing the optimal decoupling matrix parameters in the multi-task learning decoupling module. Because there are various possible errors in the sensor assembly process and experimental loading process, the calibrated PINN optimization module can compensate for the errors in the loading system and meet the decoupling requirements of most industrial sensors.

[0055] The multi-task learning decoupling module includes a multi-task learning non-shared layer input module, a neural network decoupling structure and a multi-task learning non-shared layer output module. The neural network decoupling structure is used to decouple the data signal processed by the multi-task learning non-shared layer input module to obtain load information. The neural network decoupling structure is a multi-task learning shared layer, which is a core neural network structure with a decoupling function and is used to implement the decoupling process. During the training of the model, the neural network decoupling structure is optimized based on the elements in the decoupling matrix, and each channel of the module shares a set of neural network model parameters. The multi-task learning decoupling module is mainly used to specifically improve the high adaptability decoupling to complex combined working conditions / environments.

[0056] The multi-task learning non-shared layer input module is used to process multiple groups of data signals under different working conditions after normalization. Compared with the traditional machine learning decoupling method, the multi-task learning non-shared layer input module can simultaneously input multiple groups of data signals under different working conditions. Each group of signals enters the multi-task learning shared layer (i.e., the neural network decoupling structure) through their respective input channels and then flows out of their respective output channels. Each channel has independent model parameters. During the training process of the model, the parameters of each channel are trained first, and then the neural network structure of the module is further optimized on this basis.

[0057] The multi-task learning non-shared layer output module is used to process and output the load information after the decoupling of the neural network decoupling structure to obtain the final load information. In the multi-task learning non-shared layer output module, the data flowing out of the multi-task learning shared layer will enter the corresponding output channel respectively, and each channel has independent model parameters. In the training process of the model, the parameters of each channel are trained first, and then the neural network structure of the module is further optimized on this basis.

[0058] Furthermore, in the multi-task learning non-shared layer input module and the multi-task learning non-shared layer output module, their functions are to deepen and expand the neural network decoupling structure forward or backward, the difference is that the parameters of these two modules are independent of the shared layer, and each channel is trained independently during the machine learning process. Ultimately, all channel data are constrained by the overall objective function while being optimized under the constraints of their respective objective functions.

[0059] This method jointly trains the model on multiple tasks, and the working conditions referred to in the model are not limited to engineering environments, loading type combinations, calibration or experimental error source combinations, etc.

[0060] The above technical solution provided by the present invention has at least the following beneficial effects:

[0061] The present invention proposes a multi-dimensional force sensor decoupling method based on the PINN method and multi-task learning. Based on the PINN method and the multi-task learning decoupling model, the system error can be compensated and the generalization ability of the model can be improved on the basis of the advantages of the neural network decoupling method. By jointly training the model on multiple tasks, the risk of overfitting can be reduced, the model can be made more robust, and the decoupling accuracy under various load types can be comprehensively improved. Compared with the existing technology, it has the following important advances:

[0062] 1) A neural network construction method for customizing a neural grid structure based on the physical deformation characteristics of the sensor and its mathematical model is proposed, which improves the construction efficiency and accuracy of the six-dimensional force sensor neural network decoupling model (i.e., multi-task learning decoupling model). The high applicability decoupling of six-dimensional force sensors has always been a major problem that has troubled the engineering community. Both physical model decoupling and mathematical model decoupling have great limitations. In recent years, decoupling methods based on machine learning have also usually treated them as black boxes for rough decoupling attempts. Therefore, the method of designing a neural network structure based on the sensor deformation characteristics and its mathematical model in this invention has important engineering value.

[0063] 2) By jointly training the model on multiple tasks, this method can calibrate the same sensor under different working conditions, making a single sensor applicable to multiple working scenarios.

[0064] 3) At the same time, this method can further improve the decoupling accuracy and enhance the robustness and generalization of the decoupling method by jointly training the model on multiple tasks and calibrating multi-condition data.

[0065] 4) By jointly training the model on multiple tasks, this method can avoid the risk of overfitting in the calibration process under a single working condition and further improve the reliability of the decoupling method.

[0066] 5) This method trains the model on multiple tasks together. Even if it is calibrated for a single working condition, it can group the data for learning, thereby speeding up the entire model training process and improving the decoupling accuracy.

[0067] 6) The least squares method and PINN method are used as part of the multi-task learning decoupling method, and pre-decoupling is performed before multi-task learning training. The obtained optimal theoretical decoupling matrix is ​​used as the initial parameters of the multi-task learning neural network, which greatly improves the training speed and the decoupling accuracy.

[0068] 7) The present invention proposes for the first time that the two optimal decoupling matrices obtained by the PINN method in the model training and decoupling processes are used to eliminate environmental error interference in the model training loading experiment and the decoupling experiment, and ensure that the parameters obtained by the subsequent multi-task learning method are all adaptive parameters of the multi-dimensional force sensor to complex combined working conditions. Environmental interference is further eliminated and decoupling accuracy is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a schematic diagram of the physical model of a simple cross-beam type six-dimensional force sensor;

[0070] Figure 2 It is a schematic diagram of the data coupling model of each channel;

[0071] Figure 3 This is a schematic diagram of the eccentric load condition of the six-dimensional force sensor;

[0072] Figure 4 It is a schematic diagram of the technical route of the method of the present invention;

[0073] Figure 5 It is a schematic diagram of the decoupling principle of the neural network model based on multi-task learning of the present invention;

[0074] Figure 6 Schematic diagram of the model training process based on multi-task learning in the present invention. DETAILED DESCRIPTION

[0075] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0076] The present invention is a multi-dimensional force sensor decoupling method based on a multi-task learning neural network, which is further described below using a cross-beam type six-dimensional force sensor model as an example:

[0077] Figure 1 , Figure 3 The sensor model and a schematic diagram of a common working condition are provided. Table 1 shows the patch scheme of the sensor model and the strain gauge serial number and its strain sensitive direction. Figure 2 , Figure 5Schematic diagram of the data coupling model of each channel and the neural network model based on multi-task learning of the present invention. Aiming at the decoupling problem of multi-dimensional force sensor, the present invention designs a multi-dimensional force sensor decoupling solution based on multi-task learning neural network. The technical route is as follows Figure 4 .

[0078] This method is based on a multi-task learning decoupling model. The model training process is as follows: Figure 6 .

[0079] 1. Finite element simulation experiment, and obtain the theoretical decoupling matrix based on finite element data

[0080] First, a uniaxial linear loading experiment is performed through finite element simulation, that is, uniaxial loading is performed on Fx, Fy, Fz, Mx, My, and Mz respectively to obtain finite element simulation experimental data. This process is to obtain a theoretical coupling model of a multi-dimensional force sensor. Among them, the loading experiment performed by the finite element simulation can also be other forms of loading.

[0081] The obtained data were normalized to obtain experimental data with a unified standardized form.

[0082] according to Figure 1 The physical model of the cross beam sensor in the data coupling model is established as follows Figure 2 .

[0083] When the forces / torques in each dimension do not interfere with each other, the force / torque expression is:

[0084]

[0085] The linear static decoupling of the six-axis force sensor is actually processed for calibration data, with the purpose of obtaining the linear relationship between the force / torque loading value of each dimension and the corresponding output voltage. Since the output channel of the six-axis force sensor not only receives input data from its own channel, but also receives coupling interference from other channels, it is expressed in the form of a matrix:

[0086]

[0087]

[0088]

[0089] Among them, F x 、F y 、F z 、M x 、M y 、M z They are the axial force in the x direction, the axial force in the y direction, the axial force in the z direction, the moment in the x direction, the moment in the y direction, and the moment in the z direction.1 , X 2 , X 3 , X 4 , X 5 , X 6 According to F x 、F y 、F z 、M x 、M y 、M z The electrical signal combination of the corresponding strain gauges obtained at the deformation sensitive position of the sensor deformation body when uniaxially loaded in six strain sensitive directions, C FEM is the theoretical decoupling matrix. 1 、s 2 、s 3 ...s 16 They are the strains at the corresponding positions of the 16 strain gauges obtained through finite element simulation, and their specific information is shown in Table 1.

[0090] Table 1 Strain gauge serial number and its strain sensitive direction diagram

[0091]

[0092] The traditional least squares decoupling method aims to calculate the coefficients in the coupling matrix C. This method models the neural network decoupling structure based on the mathematical model of the sensor deformable body structure and the decoupling matrix, that is, the mathematical model and the elements in the decoupling matrix are used as part of the neural network decoupling structure, and the neural network structure of the neural network decoupling structure is further optimized, and the module can be deepened, widened and expanded. It has been found that the obtained coupling matrix C is suitable for various sensor models.

[0093] The theoretical decoupling matrix C is obtained by calculating the least squares method and finite element simulation data. FEM .

[0094] 2. Uniaxial loading test and use PINN method to solve the optimal decoupling matrix

[0095] Then, a group or a small amount of experimental data is obtained through a uniaxial loading experiment. This process is to obtain the systematic error between the multi-dimensional force sensor and the loading system.

[0096] The obtained data are normalized to obtain experimental data with a unified standardized form, combined with the decoupling matrix C FEM , calculate the optimal decoupling matrix C by the PINN method PINN .

[0097] 3. Complex working conditions / environmental loading experiments, and training based on experimental data to obtain a multi-task learning decoupling model

[0098] Then, the experimental data under the target working condition is obtained through loading experiments. The specific working conditions can be determined according to the needs, such as uniaxial axial force loading conditions under eccentric load, uniaxial moment loading conditions, biaxial axial force combined loading conditions, biaxial moment combined loading conditions, biaxial moment axial force combined loading conditions, and triaxial random combined loading conditions, a total of six working conditions. Each working condition is loaded multiple times and data is collected (it can also be a combination of various other complex working conditions or environments).

[0099] Then, the data obtained under the six working conditions are processed through a data normalization module to obtain experimental data in a unified standardized form.

[0100] according to Figure 1 The physical model of the cross beam sensor in the paper establishes a 16x6x6 structure neural network as the core decoupling structure, and obtains a neural network decoupling structure, which can use the optimal decoupling matrix C PINN On this basis, the neural network model of the neural network decoupling structure is optimized to complete the construction of the core decoupling module based on the physical model.

[0101] A multi-task learning non-shared layer input module is added above the neural network decoupling structure. The input module includes 6 multi-task learning channels, which is the same as the number of experimental data types obtained.

[0102] Similarly, a multi-task learning non-shared layer output module is added below the neural network decoupling structure. The output module includes 6 multi-task learning channels, which is the same as the number of experimental data types obtained. Then the experimental data is used to optimize the network structure and obtain the multi-task learning decoupling model as shown in Figure 5 .

[0103] Finally, the experimental data is learned through the neural network model and calibrated and decoupled to obtain the multi-task learning neural network parameters. At this point, a trained multi-task learning decoupling model can be obtained.

[0104] Based on the multi-task learning decoupling model trained above, a decoupling experiment is conducted. Since the loading system needs to be reassembled during the decoupling experiment, the system error will also be different. Therefore, a pre-loading experiment is required to obtain new optimal decoupling parameters to eliminate the interference of system errors. Figure 4 , specifically including the following steps:

[0105] 1) Obtain the electrical signal data of the multi-dimensional force sensor strain gauge by performing a uniaxial experiment on the loading platform;

[0106] 2) normalizing the data collected in step 1);

[0107] 3) using the PINN method and combining the uniaxial experimental data of the loading platform processed in step 2) to calibrate the multi-task learning decoupling model;

[0108] 4) performing a loading experiment on the loading platform, and decoupling the normalized experimental results based on the calibrated multi-task learning decoupling model to obtain load information.

[0109] Wherein, in steps 1) and 4), the two loadings are completed on the same set of loading platforms.

[0110] The embodiment of the present invention further provides a multi-dimensional force sensor decoupling device based on the PINN method and multi-task learning, which is used to perform the above decoupling method. The device includes a data acquisition module, a data normalization processing module, a PINN optimization module and a multi-task learning decoupling module;

[0111] The data acquisition module is used to collect electrical signal data of the strain gauge when the loading platform performs a loading experiment;

[0112] The data normalization processing module is used to perform normalization processing on the data collected by the data collection module;

[0113] The PINN optimization module is used to calibrate some fixed parameters in the multi-task learning decoupling module;

[0114] The multi-task learning decoupling module includes a multi-task learning non-shared layer input module, a neural network decoupling structure and a multi-task learning non-shared layer output module; the multi-task learning non-shared layer input module is used to process multiple groups of data signals under different working conditions after normalization; the neural network decoupling structure is used to decouple the data signals processed by the multi-task learning non-shared layer input module to obtain load information; the multi-task learning non-shared layer output module is used to process and output the load information after decoupling of the neural network decoupling structure to obtain the final load information.

[0115] Another embodiment of the present invention also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned multi-dimensional force sensor decoupling method based on the PINN method and multi-task learning.

[0116] Another embodiment of the present invention further provides a computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions are used to enable a computer to execute the above-mentioned multi-dimensional force sensor decoupling method based on the PINN method and multi-task learning.

Claims

1. A multi-dimensional force sensor decoupling method based on PINN method and multi-task learning, characterized in that: The following steps are involved: 1) Obtain the electrical signal data of the multi-dimensional force sensor strain gauge by performing a uniaxial experiment on the loading platform; 2) normalizing the data collected in step 1); 3) using the PINN method and combining the uniaxial experimental data of the loading platform processed in step 2) to calibrate the multi-task learning decoupling model; 4) performing a loading experiment of a complex combined working condition on the loading platform to obtain electrical signal data of the strain gauge of the multi-dimensional force sensor, and decoupling the normalized electrical signal data based on the calibrated multi-task learning decoupling model to obtain load information; The construction process of the multi-task learning decoupling model is as follows: Establish a mathematical model according to the sensor deformable body structure; construct a neural network decoupling structure based on the mathematical model, and the neural network decoupling structure is a multi-task learning shared layer; add a multi-task learning non-shared layer input module and a multi-task learning non-shared layer output module to the input part and the output part of the neural network decoupling structure respectively, and set the number of channels according to the task type to obtain the multi-task learning decoupling model; The training process of the multi-task learning decoupling model is: (1) Data acquisition: Obtain sensor load and strain gauge strain data through finite element simulation; obtain sensor load information and strain gauge electrical signal data through loading experiments under complex combined working conditions or complex environments; (2) normalizing the data collected in step (1); (3) Based on the mathematical model, solving the theoretical decoupling matrix according to the least square method and the finite element simulation data obtained in step (2); (4) using the PINN method, the theoretical decoupling matrix is ​​optimized in combination with the uniaxial loading experimental data processed in step (2) to obtain the optimal decoupling matrix; (5) preliminarily optimizing the neural network decoupling structure based on the elements in the optimal decoupling matrix; (6) Further optimizing the neural network decoupling structure and its parameters, the multi-task learning non-shared layer input module structure and its parameters, and the multi-task learning non-shared layer output module structure and its parameters based on the loading test data under the complex combination working conditions or complex environment processed by step (2), and obtaining a trained multi-task learning decoupling model.

2. According to claim 1, a multi-dimensional force sensor decoupling method based on PINN method and multi-task learning is characterized in that: In step 3), the multi-task learning decoupling model is calibrated by using the PINN method in combination with the uniaxial experimental data of the loading platform processed in step 2); the specific method is: (3.1) Obtain the sensor load and strain gauge strain data through finite element simulation; (3.2) normalizing the data collected in step (3.1); (3.3) Based on the mathematical model corresponding to the loading platform, solve the theoretical decoupling matrix according to the least squares method and the finite element simulation data obtained in step (3.2); (3.4) using the PINN method, combined with the uniaxial experimental data of the loading platform processed in step 2), to optimize the theoretical decoupling matrix to obtain the optimal decoupling matrix; (3.5) The optimal decoupling matrix is ​​used to replace the optimal decoupling matrix in the multi-task learning decoupling model.

3. According to claim 1, a multi-dimensional force sensor decoupling method based on PINN method and multi-task learning is characterized in that: In the step (3), based on the mathematical model, the theoretical decoupling matrix C is solved according to the least squares method and the finite element simulation data obtained in step (2). FEM : C FEM =F FEM X FEM T (X FEM X FEM T ) -1 in: Among them, F x 、F y 、F z 、M x 、M y 、M z They are the axial force in the x direction, the axial force in the y direction, the axial force in the z direction, the moment in the x direction, the moment in the y direction, and the moment in the z direction. X1, X2, X3, X4, X5, and X6 are the x 、F y 、F z 、M x 、M y 、M z The electrical signal combination of the corresponding strain gauges obtained at the deformation sensitive position of the sensor deformation body when uniaxially loaded in the six strain sensitive directions, k1, k2, k3, k4, k5, k6 are all constants, obtained by least squares fitting; s1, s2, s3...s 16 They are the strains at the corresponding positions of the 16 strain gauges obtained through finite element simulation.

4. According to claim 2, a multi-dimensional force sensor decoupling method based on PINN method and multi-task learning is characterized in that: In step (4), the PINN method is used to combine the uniaxial loading experimental data processed in step (2) to calculate the theoretical decoupling matrix C FEM Optimize and get the optimal decoupling matrix C PINN ; The formula is as follows: C PINN =λ1·λ2 ·λ3·C FEM ·D -1 D·F Expt =λ1·λ2·λ3C FEM ·X Expt Among them, F Expt and X Expt are the load information and telecommunication signal information in the loading experiment respectively; D represents the combination of the rotation matrix R, which is the correction coefficient matrix used to rotate the load direction; R is a 3×3 rotation matrix, representing the deflection of the load; λ1, λ2, λ3 represent the correction coefficient matrices of sizes 6×6, 6×n, n×m, and m×6 respectively; n and m represent the number of rows and columns of the intermediate matrix, which are the parameters to be determined.

5. According to claim 1, a multi-dimensional force sensor decoupling method based on PINN method and multi-task learning is characterized in that: The total loss function in the PINN method consists of two parts: Loss=α1Loss physics +α2Loss data Loss is the total loss, α1 and α2 are the weighted physical constraint terms Loss physics and data fitting term Loss data The hyperparameters are used to adjust the importance of the two parts of the loss in the overall loss; The data fitting term is used to ensure that the decoupling matrix accurately approximates the observed data. The specific calculation formula is: Loss data =||D·F Expt -λ1·λ2·λ3C FEM ·X Expt || 2 The calculation formula of the physical constraint term is: Loss physics =||I-D T D|| 2 Where I is the identity matrix.

6. The multi-dimensional force sensor decoupling method based on PINN method and multi-task learning according to claim 1 is characterized in that: During the training process of the multi-task learning model, the parameters of the multi-task learning non-shared layer input module and the multi-task learning non-shared layer output module are independent of the multi-task learning shared layer, and the parameters of each channel of the multi-task learning non-shared layer input module and the multi-task learning non-shared layer output module are constrained by their respective loss functions and the total loss function during the training process.

7. A multi-dimensional force sensor decoupling device based on PINN method and multi-task learning, used to perform the decoupling method according to any one of claims 1 to 6, characterized in that: It includes data acquisition module, data normalization processing module, PINN optimization module and multi-task learning decoupling module; The data acquisition module is used to collect electrical signal data of the strain gauge when the loading platform performs a loading experiment; the data normalization processing module is used to perform normalization processing on the data collected by the data acquisition module; The PINN optimization module is used to calibrate some fixed parameters in the multi-task learning decoupling module; The multi-task learning decoupling module includes a multi-task learning non-shared layer input module, a neural network decoupling structure and a multi-task learning non-shared layer output module; the multi-task learning non-shared layer input module is used to process multiple groups of data signals under different working conditions after normalization; the neural network decoupling structure is used to decouple the data signals processed by the multi-task learning non-shared layer input module to obtain load information; the multi-task learning non-shared layer output module is used to process and output the load information after decoupling of the neural network decoupling structure to obtain the final load information.

8. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer instructions stored thereon, characterized in that: The computer instructions are used to enable a computer to execute the steps of the method according to any one of claims 1 to 6.

Images