A multi-dimensional force sensor decoupling method based on transfer learning

Abstract

The present invention belongs to the field of sensor technology, and specifically relates to a multi-dimensional force sensor decoupling method based on transfer learning. The present invention includes: constructing a calibration data set through static calibration; constructing a pre-training data set through numerical analysis and simulation methods; using the pre-training data set as the source data set, and pre-training using a neural network based on a transfer learning mechanism to obtain a pre-training model and all network parameters; fixing the shallow network parameters, and using the calibration data set as the target data set to fine-tune the last several layers of the network, obtaining new network parameters for the last several layers of the network, and obtaining a multi-dimensional force sensor decoupling model; using the multi-dimensional force sensor decoupling model to decouple the data collected by the multi-dimensional force sensor. The present invention accelerates the training speed of the neural network, avoids the neural network from falling into a local optimum during learning, and enables the decoupling model to have good convergence and robustness, while reducing the dependence on calibration data and reducing the calibration cost of the sensor.

Description

Technical Field

[0001] The present invention belongs to the technical field of sensors, and in particular relates to a multi-dimensional force sensor decoupling method based on transfer learning. Background Art

[0002] With the development of the times, single-dimensional force sensors can no longer meet the measurement requirements of daily production and research, so multi-dimensional force sensors are increasingly favored by engineering applications. Multi-dimensional force sensors refer to force sensors that can simultaneously measure force and torque components in more than two directions. In the Cartesian coordinate system, force and torque can be decomposed into three components respectively. The most complete form of multi-dimensional force is a six-dimensional force or six-dimensional torque sensor. Sensors that can simultaneously measure three force components and three torque components are widely used in large-scale engineering machinery and the mechanical arms of intelligent robots.

[0003] However, due to the integrated elastic structure of the multi-dimensional force sensor and the errors in the manufacturing process, there is inevitable coupling in the conversion channels of the sensor in different dimensional directions, that is, inter-dimensional coupling, which seriously affects the measurement accuracy of the sensor and does not meet the actual measurement requirements. Therefore, it is very important to decouple the sensor. The traditional decoupling method is to linearly decouple the sensor through the single-axis calibration data under the decoupling principle. However, the sensor is affected by many factors such as structural design and manufacturing process, which will cause problems such as incomplete decoupling and ill-conditioned decoupling matrix, and also faces interference from other accidental errors. Machine learning does not need to understand the internal coupling principle of the sensor measurement system. It only needs to be trained by the voltage signal data set corresponding to the six-axis load to build the coupling relationship between the multi-dimensional loads to achieve decoupling. The neural network algorithm has become a popular nonlinear decoupling algorithm because of its strong generalization ability. However, this iterative algorithm based on gradient descent is prone to fall into the local optimal solution during the learning process. In addition, training machine learning models usually requires a large amount of data to achieve good decoupling performance, but the method of obtaining a large amount of calibration data through sensor calibration experiments is costly. Summary of the invention

[0004] The purpose of the present invention is to solve the problems of the prior art and propose a multi-dimensional force sensor decoupling method based on transfer learning, aiming to provide the best initial weights and biases for the neural network, accelerate the training speed of the neural network, avoid the neural network from falling into local optimality during learning, make the decoupling model have good convergence and robustness, and at the same time reduce the dependence on calibration data and reduce the calibration cost of the sensor.

[0005] In order to achieve the above object, the technical solution provided by the present invention is:

[0006] A multi-dimensional force sensor decoupling method based on transfer learning, comprising:

[0007] Constructing calibration dataset X by statically calibrating the multi-dimensional force sensor experiment ;

[0008] Constructing pre-training dataset X through numerical analysis and simulation methods pre ;

[0009] For the pre-training dataset X pre Perform normalization to obtain the normalized pre-training dataset X pre The source data set is pre-trained using a neural network based on a transfer learning mechanism to obtain a pre-trained model and all network parameters;

[0010] Fix the shallow network parameters and use the normalized calibration data set X experiment Fine-tune the last, two, ... N network layers respectively, calculate the decoupling accuracy of multiple models obtained after fine-tuning, and select the model with the highest decoupling accuracy as the multi-dimensional force sensor decoupling model, wherein the shallow layer refers to the network layer other than the retrained network layer;

[0011] The multi-dimensional force sensor decoupling model is used to decouple the data collected by the multi-dimensional force sensor.

[0012] Furthermore, the calibration data set X is constructed by statically calibrating the multi-dimensional force sensor. experiment ,include:

[0013] The multi-dimensional force sensor and the loading cap are installed on the calibration bracket. According to the measuring range of the multi-dimensional force sensor in each dimension, each dimension is divided into several equally spaced measuring points. The measuring points in each dimension are quantitatively increased to the full range, and then quantitatively reduced to the minimum load. The load is repeated several times to obtain the average value. The data acquisition card collects the voltage signals of each dimension. After A / D conversion, the multi-component data is uploaded to the host computer, and the corresponding multi-dimensional voltage signals are recorded to obtain the calibration data set X. experiment =[F experiment U experiment ], where F experiment The actual load on each dimension of the multi-dimensional force sensor is included in U experiment It contains the voltage signals generated by each dimension of the multi-dimensional force sensor each time loading occurs.

[0014] Furthermore, the pre-training data set X is constructed by numerical analysis and simulation methods. pre ,include:

[0015] The multi-dimensional force sensor is decoupled and calculated by the least square method to obtain the calibration matrix;

[0016] Using Latin Hypercube Sampling to Obtain Multidimensional Force Data F LS, the corresponding multi-dimensional voltage signals U are calculated by inversely solving the calibration matrix LS , build a pre-training dataset X pre =X LS =[F LS U LS ].

[0017] Furthermore, the pre-training data set X is constructed by numerical analysis and simulation methods. pre ,include:

[0018] A multi-dimensional force sensor model is established in the simulation software, and the load F is randomly generated within the range of the multi-dimensional force sensor through the joint simulation of the programming language and the simulation software. FEM And load, submit calculation, and extract the corresponding multi-dimensional voltage signal U of the multi-dimensional force sensor in sequence FEM , build a pre-training dataset X pre =X FEM =[F FEM U FEM ].

[0019] Furthermore, the normalized pre-training data set X pre For the source data set, a neural network based on the transfer learning mechanism is used for pre-training, including:

[0020] A fully connected BP neural network model with n+2 fully connected layers is used to normalize the pre-trained dataset X pre Training is performed, where the first layer is used as the input layer, the last layer is used as the output layer, and the other layers are hidden layers. n is the number of hidden layers. The output of the i-th layer of the fully connected layer is expressed by the following formula:

[0021] x (i) =a (i) (ω (i) x (i-1) +b (i) )

[0022] Among them, i∈[2,n+2], L (i) represents the number of neurons in the i-th layer of the fully connected layer, x (i-1) ∈R L(i-1) represents the input vector of the i-th layer of the fully connected layer, x (i) ∈R L(i) represents the output vector of the i-th layer of the fully connected layer, ω (i) ∈R L(i)×L(i-1) represents the weight vector of the i-th layer of the fully connected layer, a (i) represents the activation function of the i-th layer of the fully connected layer, b (i) ∈R L(i)represents the bias term of the i-th layer of the fully connected layer; the input of the first hidden layer is x (1) =U pre , U pre is the voltage signal contained in the pre-training data set, and the output of the output layer is x (n+2) =F pre , F pre is the payload contained in the pre-training dataset.

[0023] Compared with the prior art, the present invention has the following significant advantages: first, a large amount of data is generated through calibration matrix or finite element simulation to pre-train the neural network model, and then the neural network model parameters are fine-tuned with high-fidelity calibration experimental data through transfer learning strategy, so that the decoupled model has good convergence and robustness. By freezing the shallow network parameters, only the parameters of the calibration layer are fine-tuned to achieve the purpose of enhancing the generalization ability and training efficiency of the deep neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A flow chart of a multi-dimensional force sensor decoupling method based on transfer learning according to the present invention;

[0025] Figure 2 A schematic diagram of a calibrated support platform of the present invention;

[0026] Figure 3 This is a calibration curve diagram of static calibration in Example 1 of the present invention;

[0027] Figure 4 This is a schematic diagram of transfer learning in Example 1 of the present invention;

[0028] Figure 5 It is the box plot of decoupling error of each F / M component under different decoupling methods;

[0029] Figure 6 Schematic diagram of the error rate of BPNN and BP_TL decoupling methods in different directions. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0031] Embodiment 1: Figure 1 As shown, the present invention provides a multi-dimensional force sensor decoupling method based on transfer learning, and the specific steps are as follows:

[0032] Step 1: Static calibration of the multi-dimensional force sensor and constructing the calibration data set X experiment .

[0033] Install the multi-axis force sensor and loading cap on the calibration bracket (such as Figure 2 As shown in the figure, according to the measuring range of the multi-dimensional force sensor in each dimension, each dimension is divided into several equally spaced measuring points, and the F z The load is gradually increased to the full range according to the divided measurement points in each dimension, and then gradually reduced to the minimum load, and repeated 3 times. The range of the multi-dimensional force sensor used in this embodiment is as follows: F x 、F y 、F z Direction: 80N, M x 、M y 、M z direction is 8800N·mm. During the calibration process, F z The direction is divided into 16 measurement points, and only the positive direction is calibrated; the other five directions are divided into 32 measurement points, and the positive and negative directions are calibrated. The specific calibration process is as follows: z Starting from zero, the load at each measuring point increases by 5N. After reaching 80N, the load at each measuring point decreases by 5N to zero. x 、F y Direction: Starting from -80N, each measuring point increases the load by 5N, and after reaching 80N, each measuring point decreases the load by 5N to -80N; M x 、M y 、M z Starting from -8800N·mm, the load at each measuring point increases by 550N·mm. After reaching 8800N·mm, the load at each measuring point decreases by 550N·mm to -8800N·mm. This process is repeated three times, and the final result is the average of the three experiments. After the load is applied, the elastic body of the multi-dimensional force sensor is deformed, and the corresponding differential voltage signal is generated through the Wheatstone bridge and signal amplification circuit, which corresponds to the voltage signal of the six-axis load. The data acquisition card collects the voltage signal of each dimension, uploads the six-component data to the host computer after A / D conversion, records the corresponding six-dimensional voltage signal, and obtains the calibration data set X experiment =[F experiment U experiment ], where F experiment The actual load F applied to each dimension of the sensor each time is included x ,F y ,F z ,M x ,M y ,M z , F x is the magnitude of the force applied in the x direction, M x It refers to the torque generated on the x-axis, F yis the magnitude of the force applied in the y direction, M y It refers to the magnitude of the moment generated on the y-axis, F z refers to the magnitude of the force applied in the z direction, M z It refers to the magnitude of the moment generated on the z-axis. experiment The voltage signals U1, U2, U3, U4, U5, and U6 generated by each dimension of the sensor each time are loaded are included. The results of static calibration are as follows: Figure 3 As shown. Figure 3 It can be seen that when a load is applied in one dimension, it will affect the output voltage of other dimensions. Therefore, there is a certain degree of inter-dimensional coupling between the dimensions of the multi-dimensional force sensor, which seriously affects the measurement accuracy of the sensor and requires decoupling of the multi-dimensional force sensor.

[0034] Step 2: Get the calibration matrix by the least squares method and construct the pre-training dataset X using Latin hypercube sampling LS .

[0035] Assuming that the input load and output voltage signal of each dimension are linearly related, and the output of one dimension is the superposition of the output generated by the input load of the current dimension and the output voltage signal generated by the input load of other dimensions, the least squares method can be used to decouple the sensor. In this embodiment, the six-dimensional voltage signal obtained by static calibration is a 6×n matrix, and the six-dimensional calibration force is a 6×n matrix, so a 6×6 transformation matrix is ​​required to convert the voltage signal into a six-dimensional calibration force. The relationship between the output voltage and the six-dimensional calibration force can be expressed by the following formula:

[0036] F experiment =G LS U experiment

[0037] Among them, G LS is the decoupling matrix.

[0038] According to the data obtained by static calibration in this embodiment, the decoupling matrix G can be solved as LS :

[0039]

[0040] According to the decoupling matrix, we can get the calibration matrix G c , which can be expressed as follows:

[0041]

[0042] Using Latin hypercube sampling to sample a large amount of six-dimensional force data in six-dimensional force space F LS (10,000 groups in this embodiment), by the calibration matrix G cThe corresponding six-dimensional voltage signal U can be calculated in reverse LS , construct a pre-training dataset X pre =X LS =[F LS U LS ].

[0043] The least square method is a linear decoupling method for decoupling the force sensor, and the decoupling matrix can be used to obtain the force / torque value from the voltage value. However, in order to construct a pre-training set, the present invention first uses Latin hypercube sampling to obtain a large number of force / torque values, then obtains the voltage value from the calibration matrix, and obtains the voltage from the force / torque inversely.

[0044] Step 3: Use the pre-training dataset X LS The source data set is used for pre-training using a neural network based on a transfer learning mechanism to obtain a pre-trained model.

[0045] Before machine learning, you need to pre-train the dataset X LS Perform data set division and normalization. First, LS The training set and test set are divided, and 75% of the data are selected for training the pre-trained model, and 25% are used to test the decoupling accuracy of the pre-trained model. Because the calibration force of the multi-dimensional force sensor not only has force but also torque, there are certain numerical differences, which may affect the training effect and efficiency of the neural network. In order to eliminate the influence of different dimensions on data analysis and improve the training efficiency of the neural network, the data needs to be normalized. The normalization formula is as follows:

[0046]

[0047] Among them, x is the value of the sample data, x min is the minimum value in the sample data, x max is the maximum value in the sample data. The sample data includes the pre-training data set X LS The data contained in .

[0048] The machine learning strategy adopted in this embodiment is to use a fully connected BP neural network model including n+2 fully connected layers (5 layers in this embodiment) to normalize the pre-trained data set X LS For training, the first layer is used as the input layer, the last layer is used as the output layer, and the other layers are hidden layers. For the i-th layer of the fully connected layer, assume that the number of neurons is L (i) , the input vector is x (i-1) ∈R L(i-1) , the output vector is x (i) ∈R L(i) , the weight vector is ω (i) ∈R L(i)×L(i-1), the activation function is a (i) , the bias term is b (i) ∈R L(i) , then the output of this layer can be expressed as follows:

[0049] x (i) =a (i) (ω (i) x (i-1) +b (i) )

[0050] In the overall neural network structure, the input is x (1) =U LS , that is U pre , the output is x (n+2) =F LS , that is, F pre The activation function a used in the present invention (i) is the linear rectifier function ReLU, which realizes the fitting of neural network to nonlinear data. In the pre-training process, the weight parameter ω is obtained by solving an unconstrained optimization problem (i) and the bias parameter b (i) , thus obtaining a pre-trained model.

[0051] This embodiment is described by taking the fully connected BP neural network model as an example, but is not limited to the fully connected BP neural network model. All neural network models based on the transfer learning mechanism are within the protection scope of the present invention.

[0052] Step 4: Calibrate the dataset X experiment The target dataset is used, and a multi-dimensional force sensor decoupling model (BP_TL) is obtained through transfer learning.

[0053] Transfer learning mainly includes two stages: 1. First, in the source domain (pre-training dataset X LS ) to pre-train the fully connected BP neural network model to learn all the network parameters (i.e. Step 3). 2. Fix the shallow network parameters and use the target domain (calibration dataset X experiment ) Fine-tune the last layer, two layers, ... N layers of the pre-trained model to obtain the new network parameters of the fine-tuned fully connected layer. The shallow layer here refers to the network layer other than the fine-tuned fully connected layer. Calculate the decoupling accuracy of multiple models obtained after fine-tuning, and select the model with the highest decoupling accuracy as the multi-dimensional force sensor decoupling model (BP_TL). Because different sensors, different calibration operations, different pre-training sets, and different pre-training methods will lead to different pre-training models, the network layers that need to be adjusted for transfer learning are also different.

[0054] Before performing transfer learning, it is necessary to calibrate the dataset X experimentDivide and normalize the data set according to the method in step 3, and then use the data set to fine-tune the deep parameters of the pre-trained model to establish a multi-dimensional force sensor decoupling model. Figure 4 As shown in Figure 1, the core of transfer learning lies in the transfer and freezing of parameters. Considering that there is a significant similarity in data between the source domain and the target domain, the parameters of the shallow network in the neural network have strong transferability. Therefore, the weight parameters ω of the network before the mth layer can be transferred to (i) and the bias parameter b (i) Transfer and freeze, and only fine-tune the parameters of the calibration layer (the last layer in this embodiment) to achieve the purpose of enhancing the generalization ability and training efficiency of the neural network. Following the transfer learning method, the parameters of the neural network are expressed as:

[0055]

[0056] in, and are the neural network parameters pre-trained based on the source dataset, and and These are the neural network parameters after fine-tuning based on the target dataset.

[0057] Step 5: Multi-dimensional force sensor decoupling model testing based on transfer learning.

[0058] In order to verify the superiority of the method proposed in this invention, the calibration data set X experiment Decoupling experiments based on least squares (LS) and BP neural network (BPNN) were carried out respectively.

[0059] Table 1 shows the type I error and type II error of the six-axis force sensor under three decoupling algorithms. Type I error reflects the degree of deviation between the measured value in the main direction and the actual load value, also known as the nonlinear error in the main direction. Among them, the type I error in the i direction is expressed as follows:

[0060]

[0061] Where: F i(P) Indicates the force measured in direction i; F i(L) Indicates the actual force applied in direction i; F i(FS) Indicates the full-scale value in the i direction.

[0062] Table 1

[0063]

[0064] Type II error means that the load applied in one direction causes additional interference force in other directions, which is also called coupling error. The coupling error caused by the load applied in direction j and direction i is expressed as follows:

[0065]

[0066] Where: F ij It means the force measured in the i direction when the force applied in the i direction is zero and the force applied in the j direction; F i(FS) Indicates the full-scale value in the i direction.

[0067] It can be clearly seen from Table 1 that the maximum type I errors of LS, BPNN, and BP_TL are 2.151%, 1.291%, and 0.691%, respectively, and the type II errors are 5.850%, 1.249%, and 0.551%, respectively. The decoupling performance of the multi-dimensional force sensor decoupling method based on transfer learning proposed in the present invention is significantly better than the other two decoupling methods, and the decoupling performance is nearly doubled compared to the better BPNN.

[0068] Use the decoupling error in Table 1 to draw a box plot. Figure 5 As shown, the three box plots show the decoupling error of each F / M component under different decoupling methods. The horizontal lines of the box plots represent the minimum, 25%, 50%, 75% and maximum quantiles of the error distribution. After decoupling using the BP_TL algorithm, the average and median errors in the six directions are controlled within 0.7%, and the error data of the decoupling algorithm is more concentrated. Therefore, the decoupling performance of the BP_TL algorithm proposed in the present invention is better than that of LS and BPNN, and can effectively improve the measurement accuracy of the six-dimensional force sensor.

[0069] In order to further verify the effectiveness of the proposed BP_TL decoupling algorithm, the error rates of the BPNN and BP_TL decoupling methods are shown in Figure 6 54 test samples (30 in the z direction) were selected to test the decoupling method, and type i and type ii errors were marked with different color fonts. Figure 6 It can be concluded that, except for the type I error of Fy-Fy, the two methods perform similarly, while the comprehensive performance of the remaining errors is that of BP_TL, which indicates that BP_TL has better robustness.

[0070] Embodiment 2:

[0071] The present invention provides a multi-dimensional force sensor decoupling method based on transfer learning, and the specific steps are as follows:

[0072] Step 1: Static calibration of the multi-dimensional force sensor and constructing the calibration data set X experiment .

[0073] Install the multi-axis force sensor and loading cap on the calibration bracket (such as Figure 2As shown in the figure, according to the measuring range of the multi-dimensional force sensor in each dimension, each dimension is divided into several equally spaced measuring points, and the F z The load is gradually increased to the full range according to the divided measurement points in each dimension, and then gradually reduced to the minimum load, and repeated 3 times. The range of the multi-dimensional force sensor used in this embodiment is as follows: F x 、F y 、F z Direction: 80N, M x 、M y 、M z direction is 8800N·mm. During the calibration process, F z The direction is divided into 16 measurement points, and only the positive direction is calibrated; the other five directions are divided into 32 measurement points, and the positive and negative directions are calibrated. The specific calibration process is as follows: z Starting from zero, the load at each measuring point increases by 5N. After reaching 80N, the load at each measuring point decreases by 5N to zero. x 、F y Direction: Starting from -80N, each measuring point increases the load by 5N, and after reaching 80N, each measuring point decreases the load by 5N to -80N; M x 、M y 、M z Starting from -8800N·mm, the load at each measuring point increases by 550N·mm. After reaching 8800N·mm, the load at each measuring point decreases by 550N·mm to -8800N·mm. This process is repeated three times, and the final result is the average of the three experiments. After the load is applied, the elastic body of the multi-dimensional force sensor is deformed, and the corresponding differential voltage signal is generated through the Wheatstone bridge and signal amplification circuit, which corresponds to the voltage signal of the six-axis load. The data acquisition card collects the voltage signal of each dimension, uploads the six-component data to the host computer after A / D conversion, records the corresponding six-dimensional voltage signal, and obtains the calibration data set X experiment =[F experiment U experiment ], where F experiment The actual load F applied to each dimension of the sensor each time is included x ,F y ,F z ,M x ,M y ,M z , F x is the magnitude of the force applied in the x direction, M x It refers to the torque generated on the x-axis, F y is the magnitude of the force applied in the y direction, M yrefers to the magnitude of the moment generated on the y-axis, F refers to the magnitude of the force applied in the z-direction, and M z It refers to the magnitude of the moment generated on the z-axis. experiment It contains the voltage signals U1, U2, U3, U4, U5, U6 generated by each dimension of the sensor each time it is loaded. Figure 3 It can be seen that when a load is applied in one dimension, it will affect the output voltage of other dimensions. Therefore, there is a certain degree of inter-dimensional coupling between the dimensions of the multi-dimensional force sensor, which seriously affects the measurement accuracy of the sensor and requires decoupling of the multi-dimensional force sensor.

[0074] Step 2: Use finite element simulation to construct a pre-training dataset X FEM .

[0075] A multi-dimensional force sensor model is established in a simulation software (such as ANSYS or Abaqus, ANSYS is used in this embodiment), and a large number of loads F are randomly generated within the range of the multi-dimensional force sensor through the joint simulation of a programming language (Python is used in this embodiment) and the simulation software. FEM And load, submit calculation, and extract the corresponding six-dimensional voltage signal U of the multi-dimensional force sensor in sequence FEM , build a pre-training dataset X pre =X FEM =[F FEM U FEM ].

[0076] For rigid multi-dimensional force sensors, the effects of the decoupling matrix and finite element simulation methods are similar; while for flexible multi-dimensional force sensors, the finite element simulation method is better because the input load and output voltage signal are in a nonlinear relationship.

[0077] Step 3: Use the pre-training dataset X FEM The source data set is used for pre-training using a neural network based on a transfer learning mechanism to obtain a pre-trained model.

[0078] Before machine learning, you need to pre-train the dataset X FEM Perform data set division and normalization. First, FEM The training set and test set are divided, and 80% of the data are selected for training the pre-trained model, and 20% are used to test the decoupling accuracy of the pre-trained model. Because the calibration force of the multi-dimensional force sensor not only has force but also torque, there are certain numerical differences, which may affect the training effect and efficiency of the neural network. In order to eliminate the influence of different dimensions on data analysis and improve the training efficiency of the neural network, the data needs to be normalized. The normalization formula is as follows:

[0079]

[0080] Among them, x is the value of the sample data, x min is the minimum value in the sample data, x max is the maximum value in the sample data. The sample data includes the pre-training data set X FEM The data contained in .

[0081] The machine learning strategy adopted in this embodiment is to use a fully connected BP neural network model including n+2 fully connected layers (5 layers in this embodiment) to normalize the pre-trained data set X FEM For training, the first layer is used as the input layer, the last layer is used as the output layer, and the other layers are hidden layers. For the i-th layer of the fully connected layer, assume that the number of neurons is L (i) , the input vector is x (i-1) ∈R L(i-1) , the output vector is x (i) ∈R L(i) , the weight vector is ω (i) ∈R L(i)×L(i-1) , the activation function is a (i) , the bias term is b (i) ∈R L(i) , then the output of this layer can be expressed as follows:

[0082] x (i) =a (i) (ω (i) x (i-1) +b (i) )

[0083] In the overall neural network structure, the input is x (1) =U LS , that is U pre , the output is x (n+2) =X FEM , that is, F pre The activation function a used in the present invention (i) is the linear rectifier function ReLU, which realizes the fitting of neural network to nonlinear data. In the pre-training process, the weight parameter ω is obtained by solving an unconstrained optimization problem (i) and the bias parameter b (i) , thus obtaining a pre-trained model.

[0084] This embodiment is described by taking the fully connected BP neural network model as an example, but is not limited to the fully connected BP neural network model. All neural network models based on the transfer learning mechanism are within the protection scope of the present invention.

[0085] Step 4: Calibrate the dataset X experimentThe target dataset is used, and a multi-dimensional force sensor decoupling model (BP_TL) is obtained through transfer learning.

[0086] Transfer learning mainly includes two stages: 1. First, in the source domain (pre-training dataset X LS ) to pre-train the fully connected BP neural network model to learn all the network parameters (i.e. Step 3). 2. Fix the shallow network parameters and use the target domain (calibration dataset X experiment ) Fine-tune the last layer, two layers, ... N layers of the pre-trained model to obtain the new network parameters of the fine-tuned fully connected layer. The shallow layer here refers to the network layer other than the fine-tuned fully connected layer. Calculate the decoupling accuracy of multiple models obtained after fine-tuning, and select the model with the highest decoupling accuracy as the multi-dimensional force sensor decoupling model (BP_TL). Because different sensors, different calibration operations, different pre-training sets, and different pre-training methods will lead to different pre-training models, the network layers that need to be adjusted for transfer learning are also different.

[0087] Before performing transfer learning, it is necessary to calibrate the dataset X experiment Divide and normalize the data set according to the method in step 3, and then use the data set to fine-tune the deep parameters of the pre-trained model to establish a multi-dimensional force sensor decoupling model. Figure 4 As shown in Figure 1, the core of transfer learning lies in the transfer and freezing of parameters. Considering that there is a significant similarity in data between the source domain and the target domain, the parameters of the shallow network in the neural network have strong transferability. Therefore, the weight parameters ω of the network before the mth layer can be transferred to (i) and the bias parameter b (i) Transfer and freeze, and only fine-tune the parameters of the calibration layer (the last layer in this embodiment) to achieve the purpose of enhancing the generalization ability and training efficiency of the neural network. Following the transfer learning method, the parameters of the neural network are expressed as:

[0088]

[0089] in, and are the neural network parameters pre-trained based on the source dataset, and and These are the neural network parameters after fine-tuning based on the target dataset.

[0090] The above-mentioned embodiments only express one or several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A multi-dimensional force sensor decoupling method based on transfer learning, characterized in that: The multi-dimensional force sensor decoupling method based on transfer learning includes: Constructing a calibration dataset by statically calibrating a multi-dimensional force sensor ; Constructing pre-training datasets through numerical analysis and simulation methods ; For pre-training dataset Perform normalization to normalize the pre-training data set The source data set is pre-trained using a neural network based on a transfer learning mechanism to obtain a pre-trained model and all network parameters; Fix the shallow network parameters and use the normalized calibration data set Fine-tune the last, two, ... N layers of the network layer respectively, calculate the decoupling accuracy of multiple models obtained after fine-tuning, and select the model with the highest decoupling accuracy as the multi-dimensional force sensor decoupling model, wherein the shallow layer refers to the network layer other than the retrained network layer; The multi-dimensional force sensor decoupling model is used to decouple the data collected by the multi-dimensional force sensor; in, The calibration data set is constructed by statically calibrating the multi-dimensional force sensor. ,include: The multi-dimensional force sensor and the loading cap are installed on the calibration bracket. According to the measuring range of the multi-dimensional force sensor in each dimension, each dimension is divided into several equally spaced measuring points. The measuring points in each dimension are quantitatively increased to the full range, and then quantitatively reduced to the minimum load. The load is repeated several times to obtain the average value. The data acquisition card collects the voltage signals of each dimension. After A / D conversion, the multi-component data is uploaded to the host computer, and the corresponding multi-dimensional voltage signals are recorded to obtain the calibration data set. ,in Contains the actual load on each dimension of the multi-dimensional force sensor each time F x , F y , F z , M x , M y , M z , Contains the voltage signal generated by each dimension of the multi-dimensional force sensor at each loading U 1. U 2. U 3. U 4. U 5. U 6; in, The pre-training dataset after normalization For the source data set, a neural network based on the transfer learning mechanism is used for pre-training, including: Use a The fully connected BP neural network model with layers of fully connected layers is used for the normalized pre-training data set For training, the first layer is used as the input layer, the last layer is used as the output layer, and the other layers are hidden layers. is the number of hidden layers, the number of fully connected layers The output of the layer is expressed as follows: ； in, , represents the first The number of neurons in the layer, represents the first The input vector of the layer, represents the first The output vector of the layer, represents the first The weight vector of the layer, represents the first The activation function of the layer, represents the first The bias term of the layer; the input of the first hidden layer is , is the voltage signal contained in the pre-training data set, and the output of the output layer is , is the payload contained in the pre-training dataset.

2. The multi-dimensional force sensor decoupling method based on transfer learning according to claim 1 is characterized in that: The pre-training data set is constructed by numerical analysis and simulation methods. ,include: The multi-dimensional force sensor is decoupled and calculated by the least square method to obtain the calibration matrix; Acquiring Multidimensional Force Data Using Latin Hypercube Sampling , the corresponding voltage signals of multiple dimensions are calculated by inversely solving the calibration matrix , build a pre-training dataset .

3. The multi-dimensional force sensor decoupling method based on transfer learning according to claim 1 is characterized in that: The pre-training data set is constructed by numerical analysis and simulation methods. ,include: A multi-dimensional force sensor model is established in the simulation software, and loads are randomly generated within the range of the multi-dimensional force sensor through joint simulation of programming language and simulation software. And load, submit calculation, and extract the corresponding multi-dimensional voltage signal of the multi-dimensional force sensor in sequence , build a pre-training dataset .

Images