Six-dimensional force / torque sensor decoupling system and method based on width neural network
Through a six-dimensional force/torque sensor decoupling system based on a widthwise neural network, data preprocessing is performed using a weight-weight calibration platform and a data acquisition card. Combined with the importance-scoring neuron pruning algorithm and the Moore-Penrose generalized inverse solution, the model parameters are optimized, which solves the problem of fixed model structure in the existing technology and realizes the precise decoupling and efficient measurement of the six-dimensional force/torque sensor.
Patent Information
- Application Number
- CN202510907109.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-02
- Publication Date
- 2025-10-10
AI Technical Summary
The existing software decoupling method for six-dimensional force/torque sensors has problems such as fixed model structure and insufficient nonlinear modeling capabilities, which leads to limited measurement accuracy.
A six-dimensional force/torque sensor decoupling system based on a width neural network is adopted. Data preprocessing is performed through a weight-hammer calibration platform and a data acquisition card. A width neural network model is established. The mapping matrix from the hidden layer to the output layer is solved using the importance scoring neuron pruning algorithm and the Moore-Penrose generalized inverse method. The model parameters are optimized to achieve precise decoupling of the six-dimensional force/torque sensor.
The measurement accuracy of the six-dimensional force/torque sensor is significantly improved, the inter-dimensional coupling error is reduced, the nonlinear fitting ability and structural adaptability of the model are enhanced, and it is suitable for precise decoupling and feedback in real-time control environments.
Smart Images

Figure CN120760915A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of six-dimensional force / torque sensor decoupling, and mainly relates to a six-dimensional force / torque sensor decoupling system and method based on a wideband neural network. Background Art
[0002] With the rapid development of mechanical, electronic, and computer technologies, robots are widely used in fields such as industrial automation and medical assistance. These applications place extremely high demands on robot control accuracy and environmental perception capabilities. Tasks such as precision assembly and complex medical surgeries require highly accurate force control technology. Force sensing technology is the foundation for robots to achieve precise force control, and six-dimensional force / torque sensors, capable of simultaneously and comprehensively measuring force and torque in three dimensions, have become a core component of robot perception systems.
[0003] However, the measurement accuracy of multi-dimensional force sensors in practical applications is often significantly affected by the mutual coupling effect between the measurement dimensions. This inter-dimensional coupling effect manifests itself as an external force acting in one measurement direction, which interferes with the measurement results in other directions, resulting in nonlinear errors and severely reducing the sensor's measurement accuracy.
[0004] Current decoupling methods are primarily categorized as hardware decoupling and software decoupling. Hardware decoupling methods are limited by high cost, long processing cycles, and the difficulty in avoiding manufacturing errors. Software decoupling, on the other hand, utilizes advanced data processing and modeling algorithms to suppress coupling errors based on experimental calibration data. This approach offers advantages such as high flexibility, low cost, and ease of integration and implementation, leading to wider application. However, existing software decoupling methods, such as the least squares method, support vector regression, backpropagation neural network, and extreme learning machine, suffer from limited modeling capabilities or fixed model structures, hindering further improvements in sensor measurement accuracy. Summary of the Invention
[0005] The present invention addresses the problems of fixed model structure and insufficient nonlinear modeling capabilities in existing six-dimensional force / torque sensor software decoupling methods. It provides a six-dimensional force / torque sensor decoupling system and method based on a wide neural network. In the data acquisition and preprocessing unit, the collected voltage signals are zero-drift corrected and normalized using a weight-weight calibration platform and a data acquisition card to form a network model training dataset. The established wide neural network six-dimensional force / torque sensor decoupling model is then trained using an importance-scoring-based neuron pruning algorithm and the Moore-Penrose generalized inverse to solve the mapping matrix from the hidden layer to the output layer to obtain the optimal model parameters. Finally, the preprocessed voltage signal is input to obtain a real-time output vector, which is then denormalized to obtain the decoupled six-dimensional force / torque vector. This method achieves precise decoupling of the six-dimensional force / torque sensor and significantly reduces inter-dimensional coupling errors.
[0006] In order to achieve the above-mentioned object, the technical solution adopted by the present invention is: a six-dimensional force / torque sensor decoupling system based on a width neural network at least includes a data acquisition and preprocessing unit, a width neural network modeling unit and a decoupling output unit;
[0007] The data acquisition and preprocessing unit is used to realize the calibration of the six-dimensional force / torque sensor and form a network model training data set after data preprocessing of the collected voltage signal; the data acquisition and preprocessing unit at least includes a weight-hammer calibration platform, a data acquisition card and a signal processing module, wherein,
[0008] The weight-hammer calibration platform applies a load of force and torque in a specific direction through the weight and transmits it to the sensor;
[0009] The data acquisition card captures the six-channel voltage signals output by the sensor in real time and sends the voltage to the signal processing module;
[0010] The signal processing module works in the host computer and is used to receive the voltage signal captured by the data acquisition card and perform data preprocessing on it; the data preprocessing at least includes zero drift correction and normalization processing;
[0011] The width neural network modeling unit: The modeling unit establishes a six-dimensional force / torque sensor decoupling model of the width neural network, which is used to construct a mapping relationship between the six-channel voltage signal output by the force sensor and the actual six-dimensional force / torque value, reduce the coupling error between dimensions, and improve measurement accuracy. The model uses a pruning algorithm based on importance scoring for parameter training and optimization. First, an expanded network model is constructed and the output weight is solved using the Moore-Penrose generalized inverse. The importance score of each neuron is calculated based on the output weight. Some neurons with high scores are retained, and the remaining neurons are pruned. Finally, the connection weights and bias matrix are reconstructed based on the retained neurons to form a final streamlined network structure, and the output weight is recalculated to complete the optimization modeling;
[0012] The decoupling output unit performs an inverse normalization process on the real-time output vector of the width neural network modeling unit to obtain a decoupled six-dimensional force / torque vector.
[0013] As an improvement of the present invention, the data acquisition card in the data acquisition and preprocessing unit collects the following data: the voltage output of the sensor at each loading point during the process of loading the weights in uniform steps from the zero state to the full-scale state and then gradually unloading and returning to the zero state.
[0014] As another improvement of the present invention, the six-dimensional force / torque sensor decoupling model of the width neural network in the width neural network modeling unit includes an input layer, a hidden layer and an output layer;
[0015] The input layer includes 6 neurons for receiving six-channel normalized voltage signals;
[0016] The output layer is obtained by fully connecting the hidden layer and is used to output the decoupled six-channel force signal;
[0017] The hidden layer is obtained by merging the feature node layer and the enhancement node layer. After fully connecting the input layer and performing nonlinear activation, the feature node layer is obtained. After fully connecting the feature node layer and performing nonlinear activation, the enhancement node layer is obtained.
[0018] In order to achieve the above object, the present invention also adopts a technical solution: a six-dimensional force / torque sensor decoupling method based on a wide neural network, comprising at least the following steps:
[0019] S1. Data acquisition: The six-dimensional force / torque sensor is mounted on a weight-weight calibration platform. Force and torque in a specific direction are applied by loading and unloading the calibration weights. The data acquisition card captures the output six-channel voltage signals in real time.
[0020] S2. Data preprocessing: The voltage signal collected in step S1 is subjected to data preprocessing. The data preprocessing includes at least zero drift correction and normalization. The zero drift correction specifically includes: using the six-channel voltage signal output by the sensor under no-load conditions as the zero drift reference, and subtracting the zero drift reference value of the corresponding channel from the voltage signal collected at each loading point. The normalization specifically includes: performing maximum value normalization on the zero drift corrected voltage data and the force / torque true value F.
[0021] S3. Wide Neural Network Model Training: The established wide neural network six-dimensional force / torque sensor decoupling model is trained. The importance score-based neuron pruning algorithm and the Moore-Penrose generalized inverse method are used to solve the mapping matrix from the hidden layer to the output layer to obtain the optimal model parameters.
[0022] S4. Decoupling result output: According to step S3, a six-dimensional force / torque sensor decoupling model of an optimal width neural network is obtained, and the voltage signal after data preprocessing is input to obtain a real-time output vector. After performing inverse normalization processing, the decoupled six-dimensional force / torque vector is obtained.
[0023] As an improvement of the present invention, the normalization process in step S2 is specifically as follows:
[0024]
[0025] in, is the normalized voltage data of the i-th channel at the j-th sampling point, V min,i and V max,i are the minimum and maximum voltage values of the i-th channel after zero drift correction in the calibration data set, is the voltage value of the i-th channel at the j-th sampling point after zero adjustment correction, is the normalized force / torque value of the i-th channel at the j-th sampling point, F i (j) is the true value of the force / torque of the i-th channel at the j-th sampling point, F min,i and F max,i are the minimum and maximum ranges of the sensor's channel i, respectively.
[0026] As another improvement of the present invention, the model training in step S3 specifically includes the following steps:
[0027] S31: Construct an extended initial width neural network model, expand the feature node layer Z and the enhanced node layer H by p times, and obtain the extended feature node layer Z p and extended enhanced node layer H p , with pM and pN neurons respectively;
[0028] S32: Randomly initialize the fully connected mapping matrix W from the input layer to the extended feature node layer e,p and the bias vector β e,p , randomly initialize the fully connected mapping matrix W from the expanded feature node layer to the expanded enhancement node layer h,p and the bias vector β h,p ;
[0029] S33: Calculate the extended feature node layer Z p and extended enhanced node layer H p , get the extended hidden layer [Z p |H p ], whose Moore-Penrose generalized inverse is [Z p |H p ] + , using the normalized force / torque reference output F norm , solve the mapping matrix from the extended hidden layer to the output layer as W o,p , specifically:
[0030] Z p =φ(XW e,p +β e,p )
[0031] H p =ξ(ZW h,p +β h,p )
[0032] W o,p =[Z p |H p ] + F norm
[0033] Among them, φ(·) is the distance from the input layer X to the extended feature node layer Z p The nonlinear activation function, ξ(·) is the function from the feature node layer Z to the extended enhancement node layer H p Non-linear activation function;
[0034] S34: Extend feature node layer Z p and extended enhanced node layer H p The importance score of each neuron in is calculated and evaluated, specifically:
[0035] S j =||W o,p (i,:)||2
[0036] Among them, S j Score the importance of the j-th neuron, W o,p (i,:) is W o,p The i-th row of the matrix, ‖·‖2 is the calculation of the vector's binorm:
[0037] S35: ranking the importance scores obtained in step S34 for the extended feature node layer Z p and the extended enhancement node layer H p , respectively retaining the M and N neurons with the highest scores, and eliminating the rest of the neurons;
[0038] S36: for the mapping matrix W e,p and W h,p , the bias vector β e,p and β h,p , connecting the neurons retained in step S35 into strong connections, retaining their corresponding data rows, and constructing the fully connected mapping matrix W e and the bias vector β e from the input layer to the feature node layer Z, and the fully connected mapping matrix W h and the bias vector β h from the feature node layer to the enhancement node layer;
[0039] S37: re-computing the feature node layer Z, the enhancement node layer H, and the output layer mapping matrix W o according to the optimal parameters obtained in step S36:
[0040] Z = φ (XW e + β e )
[0041] H = ξ (ZW h + β h )
[0042] W o = [Z | H] + F norm .
[0043] As a further improvement of the present application, the step S4 performs a reverse normalization process after decoupling the six-dimensional force / torque vector, specifically:
[0044] F i = Y i × (F max,i - F min,i ) + F min,i , i = 1, 2, …, 6
[0045] In the formula, F i is the decoupled force or torque output value of the i-th channel, and Y i is the network output value of the i-th channel.
[0046] Compared with the prior art, the present application has the beneficial effects of:
[0047] (1) The six-dimensional force / torque sensor decoupling model of the width neural network established by the method of the present invention has a flexible network model structure, which improves the nonlinear fitting ability and structural adaptability of the decoupling method and significantly improves the measurement accuracy of the six-dimensional force / torque sensor.
[0048] (2) The six-dimensional force / torque sensor decoupling model of the width neural network established by the method of the present invention has a simple model structure, a small computational burden, is easy to implement and promote, is suitable for real-time control environments, and can achieve accurate and efficient force / torque information decoupling and feedback in robot force control and complex environment perception.
[0049] (3) During the network model training process of the method of the present invention, the network parameters are optimized using a neuron pruning algorithm based on importance scoring, which effectively eliminates low-contribution nodes, thereby maintaining the simplicity and computational efficiency of the model structure and further improving the accuracy and robustness of the decoupling.
[0050] (4) After collecting data, the method of the present invention sets a data preprocessing step. The data preprocessing process adopts zero drift correction and maximum value normalization to eliminate the influence of initial drift of the sensor and unify the dimension, thereby ensuring the high quality of input data and the stability of model training.
[0051] (5) During the network model training process of the method of the present invention, the output weight matrix is efficiently solved by the Moore-Penrose generalized inverse method, which avoids the problems of gradient vanishing and long training time in traditional neural network back propagation, and realizes fast and efficient model training and online real-time decoupling. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 Schematic diagram of the structure of the decoupling model in the system of the present invention;
[0053] Figure 2 Schematic diagram of the working process of the decoupling output unit in the system of the present invention;
[0054] Figure 3 Schematic diagram of the workflow of data acquisition and data preprocessing in the method of the present invention;
[0055] Figure 4 Schematic diagram of the steps of decoupling model training in the method of the present invention;
[0056] Figure 5 Schematic diagram for comparing comprehensive indicators of different algorithms in the test examples of the present invention;
[0057] Figure 6 Schematic diagram of the comparison of inter-dimensional coupling errors of different algorithms in the test example of the present invention. DETAILED DESCRIPTION
[0058] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.
[0059] Example 1
[0060] A six-dimensional force / torque sensor decoupling system based on a width neural network includes at least a data acquisition and preprocessing unit, a width neural network modeling unit, and a decoupling output unit;
[0061] The data acquisition and preprocessing unit includes a weight-hammer calibration platform, a data acquisition card, and a signal processing module. The weight-hammer calibration platform is used to calibrate the six-dimensional force / torque sensor, applying a load F of force and torque in a specific direction through the weight and transmitting it to the sensor. The data acquisition card is used to capture the six-channel voltage signals output by the sensor in real time, collect the voltage output of the sensor at each loading point during the process of loading to full scale and gradually unloading back to zero, and send it to the signal processing module. The signal processing module works in the host computer and is used to receive the voltage signals captured by the acquisition card, perform zero drift correction and normalization processing, and form a network model training data set.
[0062] The width neural network modeling unit establishes a six-dimensional force / torque sensor decoupling model of the width neural network, and the decoupling model includes an input layer, a hidden layer and an output layer, such as Figure 1 As shown, the input layer X contains 6 neurons for receiving six-channel normalized voltage signals V norm The hidden layer of the width neural network modeling unit includes a feature node layer Z and an enhancement node layer H, which contain M and N neurons respectively. The construction steps include:
[0063] First, the input layer X is fully connected and nonlinearly activated to obtain the feature node layer Z, W e and β e are the matrix and bias vector of the fully connected mapping of this process, and φ(·) is the nonlinear activation function:
[0064] Z=φ(XW e +β e )
[0065] After fully connected mapping and nonlinear activation of the feature node layer Z, the enhanced node layers H and W are obtained. h and β h are the matrix and bias vector of the fully connected mapping of this process, and ξ(·) is the nonlinear activation function:
[0066] H=ξ(ZW h +β h )
[0067] Finally, the feature node layer Z and the enhancement node layer H are merged to obtain the complete hidden layer structure [Z|H].
[0068] The output layer Y of the width neural network modeling unit is used to output the decoupled six-channel force signal F, which is obtained by the hidden layer through full connection mapping. The mapping matrix is W o :
[0069] Y=[Z|H]W o
[0070] The decoupling output unit: Figure 2 As shown, the voltage signal after zero drift correction and normalization is input into the six-dimensional force / torque sensor decoupling model of the width neural network to obtain a real-time output vector. After performing inverse normalization processing, the decoupled six-dimensional force / torque vector is obtained.
[0071] A six-dimensional force / torque sensor decoupling method based on a widthwise neural network, using the above system, specifically includes the following steps:
[0072] Step S1, data acquisition: The six-dimensional force / torque sensor is mounted on a weight-weight calibration platform. Force and torque in a specific direction are applied by loading and unloading calibration weights. The six-channel voltage signals outputted are captured in real time by a data acquisition card.
[0073] Step S2, data preprocessing: performing data preprocessing on the voltage signal collected in step S1, wherein the data preprocessing at least includes zero drift correction and normalization processing.
[0074] like Figure 3 As shown, Figure 3 This is a flowchart of the steps for data acquisition and data preprocessing. The weight-hammer calibration method is used for data calibration. First, the six-dimensional force / torque sensor is installed on the weight-hammer calibration platform. The force and torque in a specific direction are applied by loading and unloading the calibration weights. The calibration is performed independently in the three-dimensional force and three-dimensional torque directions in the Cartesian coordinate system. During single-direction calibration, the weights are first loaded from the zero state to the full-scale state in uniform steps, and then unloaded from the full-scale state back to the zero state in uniform steps. During the calibration process, the true value F of the applied force / torque is recorded, and the six-channel voltage signal V output by the sensor is captured in real time by the data acquisition card, which receives instructions from the host computer and sends the captured voltage signal to the host computer. Signal preprocessing is performed in the host computer, including at least two steps: zero drift correction and normalization, to obtain the final training data set.
[0075] The zero drift correction step specifically involves using the six-channel voltage signals output by the sensor under no-load conditions as the zero drift reference, and subtracting the zero drift reference value of the corresponding channel from the voltage signal collected at each loading point to eliminate the influence of the zero drift error. Vij is the voltage value of the i-th channel of the j-th sampling point after zero correction, Vij is the original voltage value of the i-th channel of the j-th sampling point, offset,i V0 is the zero drift reference value of the i-th channel in the idle state,
[0076]
[0077] The normalization processing is specifically the extreme value normalization processing of the zero drift corrected voltage data and the force / torque true value F in the calibration process, Vij is the normalized voltage data of the i-th channel of the j-th sampling point, min,i and V max,i are the minimum and maximum values of the zero drift corrected voltage of the i-th channel in the calibration data set, respectively, and the processed data V norm is used for the input of the network model training process; Fij is the normalized force / torque value of the i-th channel of the j-th sampling point, i (j) Fij is the force / torque true value of the i-th channel of the j-th sampling point, min,i and F max,i are the minimum and maximum ranges of the i-th channel of the sensor, respectively, and the processed data F norm is used for the reference output of the network model training process:
[0078]
[0079] Step S3, width neural network model training: the model training is performed on the established width neural network six-dimensional force / torque sensor decoupling model, the neuron pruning algorithm based on importance score and the Moore-Penrose generalized inverse are used to solve the mapping matrix from the hidden layer to the output layer, and the optimal model parameters are obtained. The specific steps are as shown in Figure 4 .
[0080] S31: first, an expanded initial width neural network model is constructed, the feature node layer Z and the enhanced node layer H are expanded by p times to obtain the expanded feature node layer Z p and the expanded enhanced node layer H p , which have pM and pN neurons, respectively;
[0081] S32: the full connection mapping matrix W e,p and the bias vector β e,p from the input layer to the expanded feature node layer, and the full connection mapping matrix W h,p and the bias vector β h,p from the expanded feature node layer to the expanded enhanced node layer are randomly initialized;
[0082] S33: the expanded feature node layer Z pand extended enhanced node layer H p , get the extended hidden layer [Z p |H p ], whose Moore-Penrose generalized inverse is [Z p |H p ] + , further using the normalized force / torque reference output F norm Solve the mapping matrix from the expanded hidden layer to the output layer as W o,p :
[0083] Z p =φ(XW e,p +β e,p )
[0084] H p =ξ(ZW h,p +β h,p )
[0085] W o,p =[Z p |H p ] + F norm
[0086] S34: Extend feature node layer Z p and extended enhanced node layer H p The importance score of each neuron in S is calculated and evaluated. j Score the importance of the j-th neuron, W o,p (i,:) is W o,p For the i-th row of the matrix, ‖·‖2 is the vector bi-norm calculation, and the bi-norm size of the row where the neuron is located is the size of its importance score:
[0087] S j =||W o,p (i,:)||2
[0088] S35: Expand feature node layer Z according to importance score p and extended enhanced node layer H p Sort the neurons in the , retain the M and N neurons with the highest scores respectively, and remove the rest of the neurons;
[0089] S36: For the mapping matrix W e,p and W h,p , and the bias vector β e,p and β h,p In step S35, the neural network connection of the neurons is retained as a strong connection, the corresponding data rows are retained, and the fully connected mapping matrix W from the input layer to the feature node layer Z is constructed. e and the bias vector β e, and the fully connected mapping matrix W from the feature node layer to the enhanced node layer h and the bias vector β h ;
[0090] S37: Recalculate the feature node layer Z, enhanced node layer H, and output layer mapping matrix W o :
[0091] Z=φ(XW e +β e )
[0092] H=ξ(ZW h +β h )
[0093] W o =[Z|H] + F norm .
[0094] Step S4: Decoupling result output: According to step S3, the six-dimensional force / torque sensor decoupling model of the optimal width neural network is obtained, the voltage signal after data preprocessing is input, and the real-time output vector is obtained. After performing inverse normalization processing, the decoupled six-dimensional force / torque vector is obtained.
[0095] After the model training is completed, the actual voltage output of the sensor is received and zero-drift correction and normalization are performed. Then, it is passed into the width neural network to obtain the real-time output vector Y. After performing the denormalization process, the decoupled six-dimensional force / torque vector F is obtained:
[0096] F i =Y i ×(F max,i -F min,i )+F min,i , i=1,2,…,6
[0097] Where, F i is the decoupling force or torque output value of the i-th channel, Y i is the network output value of the i-th channel.
[0098] Test Case
[0099] In this test case, several existing six-dimensional force decoupling algorithms were selected to conduct index comparison experiments, including the least squares method, support vector regression method, back propagation neural network method, extreme learning machine method and parallel voltage-extreme learning machine method. At the same time, the proposed width neural network method was divided into two categories, namely the basic width neural network method without pruning training optimization, recorded as "width neural network method"; the width neural network method based on pruning optimization, recorded as "width neural network method (pruning optimization)". Decoupling tests were carried out on each algorithm based on 60 sets of complete six-dimensional force / torque sensor calibration data. The test indicators include comprehensive indicators and inter-dimensional coupling errors, as shown below. Figure 5 and Figure 6 As shown:
[0100] (1) Comprehensive indicators. These reflect the comprehensive fitting ability of the algorithm, including mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (MSE), which are defined as:
[0101]
[0102] Among them, y ij and They represent the true value of the jth direction at the i-th test point and the predicted output of the decoupling algorithm, R j Indicates the full-scale value of the sensor in the jth direction. Figure 5 The comprehensive comparison of MAE, MAPE and MSE indicators at all test points in this test case is shown. Figure 5 It can be seen that the wide neural network method after pruning training has the lowest error in all three indicators, which is 10.53%, 15.79% and 13.33% higher than the wide neural network method without pruning training, and at least 37.04%, 40.74% and 40.91% higher than other algorithms.
[0103] (2) Inter-dimensional coupling error. This reflects the decoupling accuracy of the algorithm in each measurement direction and is defined as:
[0104]
[0105] The test results are as follows Figure 6 shown. Figure 6 As can be seen in the figure, the coupling error of the pruned wide neural network method in all directions is less than 0.63%, which is lower than that of other algorithms. Specifically, it outperforms other algorithms in six directions by at least 35.7%, 40.6%, 51.9%, 56.0%, 68.4%, and 48.8%, respectively. Compared with the unpruned wide neural network method, the improvement ranges from 5.8% to 40.8%, reflecting the significant advantage of the wide neural network method in suppressing inter-dimensional coupling and the effectiveness of the pruning algorithm in improving performance.
[0106] In summary, the present method proposes a six-dimensional force / torque decoupling model based on a wide neural network. Its hidden layer consists of a feature node layer and an enhanced node layer, which provides excellent feature extraction and nonlinear modeling capabilities. Furthermore, a pruning algorithm based on importance scoring is used for parameter optimization to further improve the robustness of the model. The system and method of the present invention achieve precise decoupling of the six-dimensional force / torque sensor, significantly reducing inter-dimensional coupling errors and further improving the control accuracy and environmental adaptability of the robotic system.
[0107] It should be noted that the above content merely illustrates the technical idea of the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.
Claims
1. A six-dimensional force / torque sensor decoupling system based on a wide-width neural network, characterized by: At least includes a data acquisition and preprocessing unit, a width neural network modeling unit and a decoupling output unit; The data acquisition and preprocessing unit is used to realize the calibration of the six-dimensional force / torque sensor and form a network model training data set after data preprocessing of the collected voltage signal; the data acquisition and preprocessing unit at least includes a weight-hammer calibration platform, a data acquisition card and a signal processing module, wherein, The weight-hammer calibration platform applies a load of force and torque in a specific direction through the weight and transmits it to the sensor; The data acquisition card captures the six-channel voltage signals output by the sensor in real time and sends the voltage to the signal processing module; The signal processing module works in the host computer and is used to receive the voltage signal captured by the data acquisition card and perform data preprocessing on it; the data preprocessing at least includes zero drift correction and normalization processing; The width neural network modeling unit: The modeling unit establishes a six-dimensional force / torque sensor decoupling model of the width neural network, which is used to construct a mapping relationship between the six-channel voltage signal output by the force sensor and the actual six-dimensional force / torque value. The model uses a pruning algorithm based on importance scoring to perform parameter training and optimization, constructs an expanded network model and uses the Moore-Penrose generalized inverse to solve the output weight. The importance score of each neuron is calculated based on the output weight, and some neurons with high scores are retained, and the remaining neurons are pruned. Finally, the connection weights and bias matrices are reconstructed based on the retained neurons to form a streamlined network structure, and the output weights are recalculated to complete the optimization modeling; The decoupling output unit performs an inverse normalization process on the real-time output vector of the width neural network modeling unit to obtain a decoupled six-dimensional force / torque vector.
2. The six-dimensional force / torque sensor decoupling system based on a wide-width neural network according to claim 1, characterized in that: The data acquisition card in the data acquisition and preprocessing unit collects data such as the voltage output of the sensor at each loading point during the process of loading the weights uniformly in steps from zero state to full scale state and then gradually unloading and returning to zero state.
3. The six-dimensional force / torque sensor decoupling system based on a wide-width neural network according to claim 1, characterized in that: The six-dimensional force / torque sensor decoupling model of the width neural network in the width neural network modeling unit includes an input layer, a hidden layer and an output layer; The input layer includes 6 neurons for receiving six-channel normalized voltage signals; The output layer is obtained by fully connecting the hidden layer and is used to output the decoupled six-channel force signal; The hidden layer is obtained by merging the feature node layer and the enhancement node layer. After fully connecting the input layer and performing nonlinear activation, the feature node layer is obtained. After fully connecting the feature node layer and performing nonlinear activation, the enhancement node layer is obtained.
4. A six-dimensional force / torque sensor decoupling method based on a widthwise neural network, using the system according to claim 1, characterized in that: At least the following steps are included: S1. Data acquisition: The six-dimensional force / torque sensor is mounted on a weight-weight calibration platform. Force and torque in a specific direction are applied by loading and unloading the calibration weights. The data acquisition card captures the output six-channel voltage signals in real time. S2. Data preprocessing: The voltage signal collected in step S1 is subjected to data preprocessing. The data preprocessing includes at least zero drift correction and normalization. The zero drift correction specifically includes: using the six-channel voltage signal output by the sensor under no-load conditions as the zero drift reference, and subtracting the zero drift reference value of the corresponding channel from the voltage signal collected at each loading point. The normalization specifically includes: performing maximum value normalization on the zero drift corrected voltage data and the force / torque true value F. S3. Wide Neural Network Model Training: The established wide neural network six-dimensional force / torque sensor decoupling model is trained. The importance score-based neuron pruning algorithm and the Moore-Penrose generalized inverse method are used to solve the mapping matrix from the hidden layer to the output layer to obtain the optimal model parameters. S4. Decoupling result output: According to step S3, a six-dimensional force / torque sensor decoupling model of an optimal width neural network is obtained, and the voltage signal after data preprocessing is input to obtain a real-time output vector. After performing inverse normalization processing, the decoupled six-dimensional force / torque vector is obtained.
5. The six-dimensional force / torque sensor decoupling method based on a widthwise neural network according to claim 4, characterized in that: The normalization process in step S2 is specifically as follows: in, is the normalized voltage data of the i-th channel at the j-th sampling point, V min,i and V max,i are the minimum and maximum voltage values of the i-th channel after zero drift correction in the calibration data set, is the voltage value of the i-th channel at the j-th sampling point after zero adjustment correction, is the normalized force / torque value of the i-th channel at the j-th sampling point, F i (j) is the true value of the force / torque of the i-th channel at the j-th sampling point, F min,i and F max,i are the minimum and maximum ranges of the sensor's channel i, respectively.
6. The six-dimensional force / torque sensor decoupling method based on a wide-width neural network according to claim 4, characterized in that: The model training in step S3 specifically includes the following steps: S31: Construct an extended initial width neural network model, expand the feature node layer Z and the enhanced node layer H by p times, and obtain the extended feature node layer Z p and extended enhanced node layer H p , with pM and pN neurons respectively; S32: Randomly initialize the fully connected mapping matrix W from the input layer to the extended feature node layer e,p and the bias vector β e,p , randomly initialize the fully connected mapping matrix W from the expanded feature node layer to the expanded enhancement node layer h,p and the bias vector β h,p ; S33: Calculate the extended feature node layer Z p and extended enhanced node layer H p , get the extended hidden layer [Z p |H p ], whose Moore-Penrose generalized inverse is [Z p |H p ] + , using the normalized force / torque reference output F norm , solve the mapping matrix from the extended hidden layer to the output layer as W o,p , specifically: Z p =φ(XW e,p +β e,p ) H p =ξ(ZW h,p +b h,p ) W o,p =[Z p |H p ] + F norm Among them, φ(·) is the distance from the input layer X to the extended feature node layer Z p The nonlinear activation function, ξ(·) is the function from the feature node layer Z to the extended enhancement node layer H p Non-linear activation function; S34: Extend feature node layer Z p and extended enhanced node layer H p The importance score of each neuron in is calculated and evaluated, specifically: S j =||W o,p (i,:)||2 Among them, S j Score the importance of the jth neuron, W o,p (i,:) is W o,p The i-th row of the matrix, ||·||2 is the calculation of the vector's binorm: S35: Expand the feature node layer Z according to the importance score obtained in step S34 p and extended enhanced node layer H p Sort the neurons in the , retain the M and N neurons with the highest scores respectively, and remove the rest of the neurons; S36: For the mapping matrix W e,p and W h,p , bias vector β e,p and β h,p , connect the neural network of the neurons retained in step S35 as strong connections, retain their corresponding data rows, and form a fully connected mapping matrix W from the input layer to the feature node layer Z e and the bias vector β e , and the fully connected mapping matrix W from the feature node layer to the enhanced node layer h and the bias vector β h ; S37: Recalculate the feature node layer Z, enhanced node layer H and output layer mapping matrix W according to the optimal parameters obtained in step S36 o : Z=φ(XW e +β e ) H=ξ(ZW h +b h ) W o =[Z|H] + F norm 。 7. The six-dimensional force / torque sensor decoupling method based on a wide-width neural network according to claim 6, characterized in that: The decoupled six-dimensional force / torque vector is obtained by performing the denormalization process in step S4, specifically: F i =Y i ×(F max,i -F min,i )+F min,i ,i=1,2,…,6 Where, F i is the decoupling force or torque output value of the i-th channel, Y i is the network output value of the i-th channel.
Citation Information
Cited By
Multi-dimensional force signal processing method and system
CN121230936A
A multi-dimensional force signal processing method and system
CN121230936B
Electrochemical sensor calibration method and system based on two-stage decoupling and extreme learning
CN121955149A
