A method and system for haptic contact location and three-dimensional force decoupling

Abstract

The application discloses a method and system for decoupling a haptic contact position and three-dimensional force, which comprises the following steps: step S1, constructing a simulation training data set; step S2, forming a calibration compensation matrix; step S3, training a region classification model and a region regression model corresponding to different plane regions respectively; step S4, acquiring a real multi-channel signal generated by actual single-point contact, correcting signal transmission characteristics and structural coupling interference in sequence by using the calibration compensation matrix, and normalizing the corrected signal; and step S5, inputting the normalized multi-channel signal into the region classification model, selecting at least one region regression model according to the region classification result, and obtaining the haptic contact position and three-dimensional force. The application does not need to fine-tune the model with a large number of labeled real samples, realizes simulation to real object migration through signal domain calibration compensation, reduces the complexity of full-plane nonlinear mapping through layered region regression, and improves the decoupling precision.

Description

Technical Field

[0001] This invention relates to the fields of tactile perception, multidimensional force sensor calibration, simulation-measurement transfer and signal decoupling technology, specifically to a method and system for decoupling tactile contact position and three-dimensional force. Background Technology

[0002] In tactile perception tasks, planar multi-dimensional force and position sensing devices can acquire local strain, resistance, voltage, or other response signals through multiple sensing nodes distributed below the contact surface, and estimate the tactile contact position and load components based on the differences between the signals of each channel. Such devices can be used in scenarios such as robot tactile interaction, human-machine input, and contact state monitoring.

[0003] Because the load transmission paths differ at different locations on the plane, multiple sensing channels are simultaneously sensitive to contact position, tangential force, and normal force, resulting in significant coupling between channel responses. If a unified regression relationship is established directly across the entire contact surface, the mapping from signal to position and three-dimensional force typically exhibits strong nonlinearity. While point-by-point calibration using a large number of real positions and load combinations can establish supervisory data, the acquisition process requires a precision motion platform, standard force sensors, and repeated loading, leading to high costs for data acquisition and batch device calibration.

[0004] Finite element simulation can calculate the structural response under different locations and load conditions in batches, but manufacturing tolerances, material parameter deviations, signal transmission differences, and sensor unit installation orientation errors can cause the actual response to deviate from the simulation response. Directly using a model trained solely on simulation data for a physical device can easily lead to amplitude domain shifts and inter-axis directional coupling.

[0005] Sensors 2026, 26, 1378 discloses a sparse strain node tactile interface device and its contact state decoding framework: It employs three non-collinear modules to form a nine-channel input, pre-trains a unified multilayer perceptron with 14,400 FEM single-point contact data points, and then fine-tunes the model parameters with prior constraints using 648 real calibration samples to output two-dimensional contact position and three-dimensional force. This approach can reduce the difference between the simulation and the real-world domain; however, when deployed across devices, it is still necessary to collect real training samples with position and load labels and readjust the model weights. Furthermore, the unified regression model needs to fit a nonlinear mapping from multi-channel signals to five-dimensional contact states across the entire plane.

[0006] Therefore, there is a need for a tactile contact position and three-dimensional force decoupling scheme that does not rely on real contact samples to fine-tune model parameters, completes the simulation to physical object transfer only by loading a small number of standards in the signal domain, and reduces the complexity of full-plane nonlinear regression by partitioning modeling. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, one of the objectives of this invention is to provide a method for decoupling tactile contact position from three-dimensional force. This method does not require a large number of labeled real samples for model fine-tuning, but achieves simulation-to-physical transfer through signal domain calibration compensation. Furthermore, it reduces the complexity of full-plane nonlinear mapping and improves decoupling accuracy through hierarchical region regression.

[0008] To achieve one of the objectives of this invention, the following solution is adopted: A method for decoupling tactile contact position from three-dimensional force includes the following steps: Step S1: Establish a finite element model of the planar multidimensional force position sensing device, apply loads at multiple sampling positions on the planar contact surface, generate multi-channel simulation response through finite element solution and strain extraction, and use the two-dimensional position coordinates of each sampling position and the three-dimensional force components of the load as labels to construct a simulation training dataset. Step S2: Under the same standard load conditions, obtain the simulated response and the actual response respectively, determine the signal transmission characteristic correction parameters based on the difference in channel amplitude between the two, and determine the structural coupling interference correction parameters based on the measured response direction of the local triaxial sensing node and the target direction. The calibration compensation matrix is ​​formed by the signal transmission characteristic correction parameters and the structural coupling interference correction parameters. Step S3: Train a region classification model and a region regression model corresponding to different planar regions using only the simulation training dataset. The region classification model outputs region probabilities, and the region regression model outputs two-dimensional position coordinates and three-dimensional force components. The real response is not used to fine-tune the parameters of the region classification model and the region regression model. Step S4: Obtain the actual multi-channel signal generated by the actual single-point contact, and use the calibration compensation matrix to perform signal transmission characteristic correction and structural coupling interference correction in sequence, and normalize the corrected signal. Step S5: Input the normalized multi-channel signal into the region classification model, select at least one region regression model according to the region classification result, and when multiple region regression models are selected, fuse the regression outputs to obtain the tactile contact position and three-dimensional force.

[0009] Further, in step S1, sampling positions are set for the planar contact surface in both radial and circumferential directions, and the planar contact surface is divided into multiple planar regions according to the circumferential angle; The simulation training dataset includes normal force samples with only normal force applied and three-dimensional force samples with both tangential and normal force components applied simultaneously.

[0010] Furthermore, the planar multidimensional force position sensing device includes three local triaxial sensing nodes; The center points of the three local triaxial sensing nodes are located on the same node reference plane and are not collinear. The three center points together determine the node reference plane and constitute a three-node planar sensing structure. Each of the aforementioned local triaxial sensing nodes outputs sensing signals in three directions, and the three aforementioned local triaxial sensing nodes together form a nine-channel sensing signal; The tactile contact position is represented by a two-dimensional coordinate system established in the reference plane of the node.

[0011] Further, in step S2, the signal transmission characteristic correction parameters include correction coefficients corresponding to nine signal channels respectively. Each correction coefficient is determined by the ratio of the simulated response to the actual response of the corresponding channel under the same standard load conditions. The nine correction coefficients constitute a diagonal signal transmission characteristic correction matrix. The structural coupling interference correction parameters include rotation correction matrices corresponding to the three local triaxial sensing nodes, and each rotation correction matrix is ​​determined by the rotation relationship between the measured response direction of the corresponding local triaxial sensing node under standard load and the target direction.

[0012] Furthermore, the planar contact surface is divided into twelve sector-shaped regions; The region classification model is a one-dimensional residual convolutional network trained solely on the simulation training dataset. It includes a one-dimensional convolutional layer, multiple residual blocks, a global average pooling layer, and a twelve-class linear layer, which outputs the region probabilities of twelve sector regions based on the nine-dimensional input vector.

[0013] Furthermore, each of the aforementioned sector regions corresponds to a region regression model trained using only the simulation samples of that sector region; Each of the aforementioned region regression models is a multilayer perceptron, comprising a nine-dimensional input layer, three hidden layers, and a five-dimensional output layer. The five-dimensional output layer sequentially outputs two components of two-dimensional position coordinates and three components of three-dimensional force.

[0014] Further, in step S5, when the maximum region probability is not lower than the preset confidence threshold, a region regression model corresponding to the maximum region probability is selected, and the five-dimensional output of the region regression model is output as the tactile contact position and the three-dimensional force. When the maximum region probability is lower than the preset confidence threshold or the tactile contact position is near the boundary of an adjacent sector region, the two region regression models with the highest region probabilities are selected, and the five-dimensional outputs of the two region regression models are weighted and fused according to the corresponding region probabilities to obtain the tactile contact position and the three-dimensional force.

[0015] Further, in step S3, the simulation training dataset is divided into a training set and a test set, the multi-channel simulation sensing signals are normalized to the target interval, and the mean square error loss function is used to train the regression model for each region. The region classification model and each of the region regression models are trained separately, and each of the region regression models is trained using at least the simulation samples of its corresponding planar region.

[0016] Furthermore, in step S2, under the same standard load conditions, the signal transmission characteristic correction parameter is used to adjust the amplitude of the real channel response to a range that matches the simulated channel response; The structural coupling interference correction parameters are used to correct the direction of the measured three-dimensional response vector of each local triaxial sensing node to the target direction, so as to eliminate interaxial coupling caused by installation direction and structural errors. The real multi-channel signal after being corrected by the calibration compensation matrix is ​​in the same amplitude and direction representation space as the multi-channel simulation response in the simulation training dataset.

[0017] The second objective of this invention is to provide a tactile contact position and three-dimensional force decoupling system that does not require a large number of labeled real samples for model fine-tuning, achieves simulation-to-physical transfer through signal domain calibration compensation, and reduces the complexity of full-plane nonlinear mapping through hierarchical region regression, thereby improving decoupling accuracy.

[0018] To achieve the second objective of this invention, the following solution is adopted: A tactile contact position and three-dimensional force decoupling system includes: a simulation data generation module, used to establish a finite element model of a planar multi-dimensional force position sensing device, apply loads at multiple sampling positions on the planar contact surface, generate multi-channel simulation responses through finite element solution and strain extraction, and construct a simulation training dataset by using the two-dimensional position coordinates of each sampling position and the three-dimensional force components of the load as labels; The calibration parameter determination module is used to obtain the simulated response and the actual response under the same standard load conditions, determine the signal transmission characteristic correction parameters based on the difference in channel amplitude between the two, and determine the structural coupling interference correction parameters based on the measured response direction of the local triaxial sensing node and the target direction. The calibration compensation matrix is ​​formed by the signal transmission characteristic correction parameters and the structural coupling interference correction parameters. The model training module is used to train a region classification model and a region regression model corresponding to different planar regions using only the simulation training dataset. The region classification model outputs region probabilities, and the region regression model outputs two-dimensional position coordinates and three-dimensional force components. The real response is not used to fine-tune the parameters of the region classification model and the region regression model. The signal correction module is used to acquire the real multi-channel signal generated by the actual single-point contact, and to perform signal transmission characteristic correction and structural coupling interference correction in sequence using the calibration compensation matrix, and to normalize the corrected signal. The hierarchical decoupling module is used to input the normalized multi-channel signal into the region classification model, select at least one region regression model according to the region classification result, and fuse the regression outputs when multiple region regression models are selected to obtain the tactile contact position and three-dimensional force.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention eliminates the need for fine-tuning model parameters using real contact samples, significantly reducing data acquisition and calibration costs. This invention generates multi-channel simulation responses corresponding to different contact positions and load combinations in batches through finite element simulation to construct a simulation training dataset. This dataset is then used only to train region classification and region regression models; real responses do not participate in model parameter fine-tuning. Compared to existing methods, this invention eliminates the need to collect hundreds or even thousands of real training samples with precise location and load labels, and also eliminates the need to readjust model weights during cross-device deployment. Real standard load responses are used only to determine the calibration compensation matrix and do not enter the model parameter update process, thereby significantly reducing the time and economic costs of batch device calibration.

[0020] 2. This invention achieves simulation-to-physical transfer through signal domain calibration compensation, effectively eliminating the difference between the simulation and physical domains. Addressing the issue that manufacturing tolerances, material parameter deviations, signal transmission differences, and sensor unit installation orientation errors can cause the physical response to deviate from the simulation response, this invention determines signal transmission characteristic correction parameters and structural coupling interference correction parameters under the same standard load conditions to form a calibration compensation matrix. This matrix is ​​then used to sequentially correct the signal transmission characteristics and structural coupling interference of the real multi-channel signal, ensuring that the corrected real multi-channel signal and the multi-channel simulation response in the simulation training dataset are in the same amplitude and direction representation space. This method completes the simulation-to-physical transfer in the signal domain before model input, rather than updating model parameters through backpropagation of real samples, resulting in high calibration efficiency and strong versatility.

[0021] 3. This invention reduces the complexity of full-plane nonlinear regression and improves regression accuracy through partitioned modeling. Addressing the issue of strong nonlinearity in the mapping of signals to position and three-dimensional force when establishing a unified regression relationship across the entire contact surface, this invention decomposes the full-plane five-dimensional nonlinear mapping problem into two sub-tasks: coarse regional classification and local regression, using a region classification model and region regression models corresponding to different planar regions. The region classification model outputs region probabilities for coarse localization, while the region regression model performs local regression of tactile contact position and three-dimensional force within its respective smaller planar region. This effectively reduces the difficulty of fitting complex nonlinear mappings across the entire plane using a single regression model, thus improving regression accuracy.

[0022] 4. The confidence-based routing fusion strategy of this invention ensures the continuity and reliability of contact state estimation. This invention selects at least one regional regression model based on the regional classification results: when the maximum regional probability is not lower than a preset confidence threshold, the regional regression model corresponding to the maximum regional probability is selected and the result is directly output, resulting in high inference efficiency; when the maximum regional probability is lower than the confidence threshold or the contact location is near the boundary of adjacent planar regions, the two regional regression models with the highest regional probabilities are selected and weighted and fused according to their corresponding probabilities, effectively avoiding output discontinuity caused by hard expert switching and ensuring a smooth transition of contact state estimation near the boundary of adjacent regions.

[0023] 5. This invention requires only a small amount of standard load to complete the physical signal domain adaptation, without the need to collect a large number of labeled real training samples or to retrain the model across devices. For multiple mass-produced sensing devices of the same type, different devices only need to determine their own calibration compensation matrices to share the same set of simulation training data and model weights, without the need to recalibrate each device across the entire range point by point. It has good scalability and consistency, and can be widely used in scenarios such as robot tactile interaction, human-machine input, and contact status monitoring. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method for decoupling tactile contact position and three-dimensional force in an embodiment of the present invention; Figure 2 This is a flowchart of the simulation training dataset generation process in an embodiment of the present invention, which shows the establishment of a three-node planar sensing structure and finite element model, setting of planar sampling positions and region labels, application of loads, solving and strain extraction, generation of multi-channel simulation signals and label pairing; Figure 3 This is a flowchart of the calibration compensation matrix processing in an embodiment of the present invention; Figure 4 This is a diagram illustrating the hierarchical decoupling structure of region classification and region regression in an embodiment of the present invention. Figure 5This is a schematic diagram of the generation of simulation training dataset in an embodiment of the present invention, wherein (a) shows a three-node planar sensing structure in which the node reference plane is jointly determined by non-collinear nodes S1, S2 and S3, as well as planar sampling and load setting; (b) shows the strain gauge distribution of the three local triaxial sensing nodes; (c) shows the finite element model, mesh solution and strain extraction; and (d) shows the dataset consisting of nine-channel simulation signals paired with X, Y, Fx, Fy and Fz labels and 12852 samples. Figure 6 The network structure diagram in this embodiment of the invention is as follows: after the real nine-channel signal is corrected and normalized by STC and SCIC, it is sequentially passed through a regional classification model that outputs the probability of twelve fan-shaped regions for coarse classification, Top-1 or Top-2 expert routing, a twelve-region regression model for local regression, and a five-dimensional output. Figure 7 This is a block diagram of the tactile contact position and three-dimensional force decoupling system in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0026] Example 1 like Figure 1-6 As shown, this embodiment of the invention provides a method for decoupling tactile contact position from three-dimensional force, which can reduce the collection of labeled real samples and repeated model training, and can be used for robot tactile interaction, human-machine input and contact state monitoring.

[0027] like Figure 1 As shown, the tactile contact position and three-dimensional force decoupling method of this invention includes the following steps: Step S1: Establish a finite element model of the planar multidimensional force position sensing device, apply loads at multiple sampling positions on the planar contact surface, generate multi-channel simulation responses through finite element solution and strain extraction, and use the two-dimensional position coordinates of each sampling position and the three-dimensional force components of the load as labels to construct a simulation training dataset.

[0028] Step S2: Under the same standard load conditions, obtain the simulated response and the actual response respectively. Determine the signal transmission characteristic correction parameters based on the difference in channel amplitude between the two, and determine the structural coupling interference correction parameters based on the measured response direction of the local triaxial sensing node and the target direction. The calibration compensation matrix is ​​formed by the signal transmission characteristic correction parameters and the structural coupling interference correction parameters.

[0029] Step S3: Train a region classification model and a region regression model corresponding to different planar regions using only the simulation training dataset. The region classification model outputs region probabilities, and the region regression model outputs two-dimensional position coordinates and three-dimensional force components. The real response is not used to fine-tune the parameters of the region classification model and the region regression model.

[0030] Step S4: Obtain the actual multi-channel signal generated by the actual single-point contact, and use the calibration compensation matrix to perform signal transmission characteristic correction and structural coupling interference correction in sequence, and normalize the corrected signal.

[0031] Step S5: Input the normalized multi-channel signal into the region classification model, select at least one region regression model according to the region classification result, and when multiple region regression models are selected, fuse the regression outputs to obtain the tactile contact positions X, Y and three-dimensional forces Fx, Fy and Fz.

[0032] Further, in step S1, sampling positions are set for the planar contact surface according to the radial and circumferential directions, and the planar contact surface is divided into multiple planar regions according to the circumferential angle; the simulation training dataset includes normal force samples with only normal force applied and three-dimensional force samples with both tangential force components and normal force components applied simultaneously.

[0033] Furthermore, the planar contact surface is divided into twelve sector regions; the simulation training dataset includes 12,852 samples, including 7,140 normal force samples and 5,712 three-dimensional force samples; Furthermore, in the three-dimensional force sample, the force components Fx and Fy have values ​​ranging from -5 N to 5 N, and the force component Fz has values ​​ranging from 0 N to 5 N.

[0034] Furthermore, the planar multidimensional force position sensing device includes three local triaxial sensing nodes S1, S2, and S3; the center points of the three local triaxial sensing nodes are located on the same node reference plane and are not collinear, and the three center points jointly determine the node reference plane and constitute a three-node planar sensing structure; each local triaxial sensing node outputs sensing signals in three directions to form a nine-channel sensing signal; the tactile contact positions X and Y are represented by a two-dimensional coordinate system established in the node reference plane; in step S1, the strain components of the corresponding positions of each local triaxial sensing node are extracted from the finite element solution results, and the nine-channel simulated sensing signal is generated according to the conversion relationship between the strain components and the response of the sensing element.

[0035] Furthermore, in step S2, for the first... Each signal channel, according to Determine the signal transmission characteristic correction coefficients, and according to The true channel response after signal transmission characteristic correction is obtained; where, Indicates the first under the standard load condition Each simulated channel response, Indicates the first under the same standard load conditions A real channel response, Indicates the first Signal transmission characteristic correction coefficient for each signal channel Indicates the first after signal transmission characteristic correction Each signal channel has a real channel response; the signal transmission characteristic correction coefficients of each signal channel constitute a diagonal signal transmission characteristic correction matrix.

[0036] Furthermore, in step S2, for the first... Let there be a local triaxial sensing node, and its measured three-dimensional response vector under standard load be denoted as... The direction of the target unit is denoted as Calculation unit measured direction ,according to Calculate the rotation angle ,according to Calculate the unit axis of rotation And construct the antisymmetric matrix ,according to Calculate the structural coupling interference correction matrix In step S4, according to For the first The three-dimensional response vectors of each local triaxial sensing node are used to perform structural coupling interference correction; where... Represents the identity matrix. This represents the three-dimensional response vector after orientation correction.

[0037] Furthermore, the region classification model is a one-dimensional residual convolutional network trained solely using the simulation training dataset. It includes a one-dimensional convolutional layer with a kernel width of 3, three residual blocks with channel numbers of 32, 64, and 128 respectively, a global average pooling layer, and a twelve-class linear layer, which are used to output the region probabilities of twelve fan-shaped regions based on the nine-dimensional input vector.

[0038] Furthermore, the planar contact surface is divided into twelve sector regions, each sector region corresponding to a region regression model trained using only simulation samples from that sector region; each region regression model is a 9-128-128-128-5 multilayer perceptron, including a nine-dimensional input layer, three hidden layers with 128 neurons each and activated by ReLU, and a five-dimensional output layer. The hidden layers are set with a dropout ratio of 0.1, and the five-dimensional output layer outputs X, Y, Fx, Fy, and Fz in sequence.

[0039] Further, in step S5, when the maximum region probability is not lower than the preset confidence threshold, a region regression model corresponding to the maximum region probability is selected; when the maximum region probability is lower than the preset confidence threshold or the tactile contact position is located near the boundary of an adjacent planar region, the two region regression models with the highest region probabilities are selected, and the regression outputs of the two region regression models are weighted and fused according to the corresponding region probabilities.

[0040] Further, in step S3, the simulation training dataset is divided into a training set and a test set, the multi-channel simulation sensing signal is normalized to [-1,1], and the mean square error loss function is used to train the regression model for each region; the region classification model and the region regression model are trained respectively, and each region regression model is trained using at least the simulation samples of its corresponding planar region.

[0041] The following is a further detailed explanation of the tactile contact position and three-dimensional force decoupling method of the present invention.

[0042] The tactile contact position and three-dimensional force decoupling method of this invention uses three non-collinear local triaxial sensing nodes to jointly determine the node reference plane, and uses finite element simulation data, calibration compensation matrix, and region classification-local regression hierarchical network to jointly decouple the two-dimensional position and three-dimensional force of a single-point tactile contact within the node reference plane. The tactile contact position referred to in this invention is the two-dimensional position of a single-point tactile contact within the node reference plane, denoted by X and Y.

[0043] The tactile contact position and three-dimensional force decoupling method of the present invention can avoid secondary training of model weights through real contact samples with position and load labels in the process of joint perception of two-dimensional position and three-dimensional force of single-point tactile contact. It can compensate for the channel amplitude difference and structural direction coupling between the simulation response and the real response by using a small amount of standard loading, and improve the stability of joint regression of tactile contact position and three-dimensional force in different planar regions.

[0044] This invention sets up three local triaxial sensing nodes non-collinearly, with the center points of the three nodes jointly determining the node reference plane, thus forming a multi-channel planar sensing structure. A finite element model corresponding to the planar sensing structure is established, and multiple sampling positions are set on the planar contact surface and different loads are applied. Multi-channel simulation responses are obtained through finite element solution and strain extraction, and paired with X, Y, Fx, Fy, and Fz labels to form a simulation training dataset.

[0045] This invention obtains simulated and real responses under the same standard load, and determines the Signal Transmission Characteristic (STC) parameters based on the channel amplitude differences between the two. A Structural Coupling Interference Correction (SCIC) rotation matrix is ​​constructed based on the measured response directions and target directions of each local triaxial sensing node. The STC parameters and the SCIC rotation matrix together constitute the Calibration Compensation Matrix (CCM), which is used to sequentially correct the amplitude differences and directional coupling of the real multi-channel signals before model input.

[0046] This invention utilizes only a simulation training dataset to train a region classification model and region regression models corresponding to different planar regions. During actual inference, CCM correction and normalization are performed on the real multi-channel signals. At least one region regression model is selected based on the region classification results to obtain the tactile contact positions X and Y and the three-dimensional forces Fx, Fy, and Fz. In one specific embodiment, the planar contact surface is divided into twelve sector regions, using Top-1 or Top-2 expert routing, and the outputs of the two region regression models are weighted and fused during Top-2 routing. The real standard load response is only used to determine the CCM and is not included in the parameter fine-tuning process of the region classification and region regression models.

[0047] The tactile contact position and three-dimensional force decoupling method of the present invention has the following advantages: This invention generates multi-channel simulation responses corresponding to different contact positions and load combinations in batches using finite element models, and trains regional classification and regional regression models using only simulation data; the real standard load response is only used to determine the CCM and not for model weight fine-tuning, thereby reducing the collection of real training samples with position and load labels and the repeated training of models across devices.

[0048] This invention compensates for signal transmission and manufacturing differences through STC channel-by-channel proportional correction and compensates for inter-axis coupling caused by local triaxial sensing node installation orientation and structural errors through SCIC rotation correction, so that the real multi-channel signal is converted to an amplitude and direction representation space consistent with the simulation training data.

[0049] This invention employs a region classification model for coarse classification and a region regression model for local regression of tactile contact positions and three-dimensional forces within corresponding planar regions, reducing the nonlinear fitting difficulty of a single full-plane regression model. In the specific implementation of the twelve-sector region, Top-1 routing reduces inference redundancy, and Top-2 weighted fusion near the region boundaries reduces output discontinuities caused by hard expert switching. This invention can be used for robot tactile interaction, human-machine input, and contact state monitoring.

[0050] In some embodiments, the specific content of the simulation training dataset is as follows: like Figure 5 (a) and Figure 5 As shown in (b), a finite element model is established with a planar contact surface, three local triaxial sensing nodes, and a supporting structure. In a specific implementation, the planar contact surface is a circular rigid plate with a radius of 30 mm. The three local triaxial sensing nodes are set at approximately 120 degrees intervals along the circumference and are denoted as S1, S2, and S3, respectively. The center points of the three nodes are located on the same node reference plane and are not collinear. The three center points jointly determine the node reference plane and constitute a three-node planar sensing structure. The planar contact surface is parallel to the node reference plane. Each node is equipped with seven strain elements R1 to R7, which are combined through bridge circuits to form sensing channels in three directions. The three nodes together form a nine-channel response.

[0051] like Figure 2 and Figure 5 As shown in (a), 15 concentric sampling rings were set on the planar contact surface with a radius of 30 mm, and the radial spacing between adjacent sampling rings was 2 mm. The number of circumferential sampling points in each ring increased from 24 to 360 from the inside out, so that the outer perimeter had a denser circumferential coverage. The planar contact surface was divided into twelve sector regions according to the circumferential angle. Normal loads or three-dimensional loads containing both tangential and normal components were applied in batches at each sampling position, and the two-dimensional positions X and Y and the three-dimensional forces Fx, Fy, and Fz were recorded.

[0052] like Figure 5 As shown in (c), the finite element analysis employs a static general solution procedure, using hybrid tetrahedral elements to mesh the local structure, and only outputs the strain field required for subsequent signal calculations to reduce computational load. One specific implementation uses C3D10H ten-node hybrid tetrahedral elements; the material parameters of the upper contact plate, adhesive layer, flexible circuit, polyimide layer, and strain element are set according to the materials of their respective components.

[0053] like Figure 5 (b) to Figure 5As shown in (d), after solving, the integral strain of the region containing R1 to R7 in each local triaxial sensing node is extracted. The integral strain of the same strain element is averaged, and the output in three directions is calculated by combining the sensitivity in the length direction, the width direction, and the bridge combination relationship. Among them, R1 and R3 correspond to one direction, R2 and R4 correspond to another direction, and R5, R6, and R7 correspond to the third direction. The three outputs of S1 are denoted as CH1 to CH3, the three outputs of S2 are denoted as CH4 to CH6, and the three outputs of S3 are denoted as CH7 to CH9. They are spliced ​​together to form a nine-channel simulation signal.

[0054] The simulation coordinate system and the sensor output coordinate system may be different. In this embodiment, a two-dimensional position coordinate system is established using a node reference plane jointly determined by three non-collinear nodes, and X and Y represent the tactile contact position. The Fx component in the simulation data is mapped to the sensor coordinate system Fx, the Fz component is mapped to the sensor coordinate system Fy, and the absolute value of the Fy component is mapped to the sensor coordinate system Fz; the simulation x and z are mapped to the positions X and Y in the node reference plane. After the coordinate transformation is completed, the nine channels CH1 to CH9 are paired sample by sample with the X, Y, Fx, Fy, and Fz labels.

[0055] The final simulation training dataset includes 12,852 single-point samples, of which 7,140 are pure normal force samples and 5,712 are three-dimensional force samples. The pure normal force samples include random normal forces ranging from 0 to 5 N, as well as constant normal forces of 1 N, 2 N, and 4 N. The three-dimensional force samples have Fx and Fy values ​​ranging from approximately -5 N to 5 N, and Fz values ​​ranging from approximately 0 N to 5 N. These 12,852 samples constitute the final simulation dataset used for training and testing the region classification model and twelve region regression models.

[0056] In some embodiments, the determination of the calibration compensation matrix is ​​as follows: like Figure 3 As shown, a planar multidimensional force position sensing device is mounted on a calibration platform with three-axis translation and rotation capabilities, and the applied load is monitored using a standard force sensor. Standard loads are applied to S1, S2, and S3 along their local X, Y, and Z directions, respectively; in one embodiment, a 1 N standard load is used, and the response of the stable segment under repeated loading is averaged. The STC and SCIC parameters are both determined by the above standard load response.

[0057] For each signal channel, the simulated channel response and the actual channel response are obtained under the same standard load, and the STC coefficients are calculated according to formula (1). The nine STC coefficients are used as diagonal elements to form an STC correction matrix. After the actual nine-channel signals are multiplied by the matrix, the linear amplitude of each channel is adjusted to match the simulated response.

[0058] For SCIC correction, under ideal installation conditions, when a normal standard load is applied to the planar contact surface, the response direction of each local triaxial sensing node should be consistent with the target normal direction. The tilt caused by actual assembly causes the unit measured direction to deviate from the target direction. The amplitude, unit direction, rotation angle, rotation axis and antisymmetric matrix are calculated according to formulas (2) to (5), and then the SCIC rotation matrix of each local triaxial sensing node is constructed according to formula (6). Finally, the local three-dimensional response vector is corrected according to formula (7).

[0059] (1) In the formula, Number the signal channel. and These are the simulated and actual responses obtained under the same standard load, respectively. For signal transmission characteristic correction coefficients, This is the corrected true response. (The remaining text appears to be a fragment and doesn't form a coherent sentence.) Placed in a diagonal matrix, it can perform channel-by-channel amplitude correction on multi-channel real signals.

[0060] (2) In the formula, Numbering of local triaxial sensing nodes. Let it be the measured three-dimensional response vector under standard load. For vector magnitude, The measured direction is the unit.

[0061] (3) In the formula, The rotation angle is... For the direction of the target unit, For the unit of measured direction, This indicates transpose.

[0062] (4) In the formula, The unnormalized rotation axis is obtained by the cross product of the measured direction and the target direction. The normalized unit rotation axis.

[0063] (5) In the formula, , and They are the unit rotation axes The three components; the antisymmetric matrix multiplied by any vector is equivalent to The cross product with the vector.

[0064] (6) In the formula, For the first Structural coupling interference correction matrix for a local triaxial sensing node It is the identity matrix. and These are antisymmetric matrices for the rotation angle and rotation axis, respectively.

[0065] (7) First relation representation The unit measured direction is mapped to the target direction. The second relationship means that the same rotation matrix is ​​applied to the actual local three-dimensional response vector to obtain the response after direction correction.

[0066] When using it online, such as Figure 6 As shown on the left, STC correction is first performed on the real nine-channel signal, then every three channels are combined into a local three-dimensional response vector and the corresponding SCIC correction is performed. Subsequently, the normalization parameters of the simulation training data are used for normalization to obtain a nine-dimensional input vector shared by the regional classification model and the regional regression model.

[0067] In this embodiment, the actual measurements only include the standard load response used to determine the signal transmission characteristic correction coefficients and the structural coupling interference correction matrix; it does not require the collection of real training samples labeled X, Y, Fx, Fy, and Fz at each location on the plane. The weights of the region classification model and the region regression model are determined on the simulation training data and remain unchanged when deployed to the physical device. Thus, physical adaptation occurs in the signal domain before the model input, rather than updating the model parameters through backpropagation of real samples.

[0068] In this embodiment, the specific details of training the hierarchical decoupling model are as follows: like Figure 4 and Figure 6 As shown, the hierarchical decoupling model includes a region classification model that outputs probabilities for twelve sector regions, an expert routing unit, and twelve region regression models. Model parameters are determined solely by simulation training data; 12,852 simulation samples are divided into training and testing parts at 90% and 10% respectively, and the nine-channel input is normalized to [-1, 1]. The classification training data consists of nine-channel signals and twelve sector region labels; the r-th region regression model is trained using only the nine-channel signal of the r-th sector region and the X, Y, Fx, Fy, and Fz labels.

[0069] The region classification model is used for coarse classification and employs a one-dimensional residual convolutional network. For example... Figure 6As shown in the upper part, the nine-dimensional input vector first passes through a one-dimensional convolutional layer with a kernel width of 3 to extract local cross-channel relationships, and then passes through three residual blocks with 32, 64 and 128 channels in sequence. After global average pooling, the twelve-class linear layer outputs the regional probabilities of twelve fan-shaped regions.

[0070] Each sector is configured with a regional regression model for local regression, and the twelve regional regression models constitute a local expert model group. For example... Figure 6 As shown in the middle, each local expert model uses a 9-128-128-128-5 multilayer perceptron: after the nine-dimensional input layer, three fully connected hidden layers, each containing 128 neurons and activated by ReLU, are set. The hidden layer can be set with a dropout ratio of 0.1. The five-dimensional output layer outputs X, Y, Fx, Fy and Fz in sequence.

[0071] The model can be trained using the Adam optimizer and mean squared error loss function. In one embodiment, the learning rate is 1×10^-4, and the training time does not exceed 300 epochs. The region classification model and the region regression model can be trained separately. During the training phase, a temperature-sharpened Top-2 soft route can also be used based on the region probability to maintain a certain continuity between experts in adjacent regions on the boundary samples.

[0072] In this embodiment, the specific details of the actual decoupling are as follows: like Figure 6 As shown, the actual nine-channel signal of the physical device is acquired, and STC correction, SCIC correction, and normalization are performed sequentially. The resulting nine-dimensional input vector is used as input to both the region classification model and the routed local expert model. If the maximum region probability reaches a preset confidence threshold, the local expert model corresponding to the maximum probability is called in a Top-1 manner. If the maximum region probability is lower than the threshold or the tactile contact position is near the boundary of an adjacent sector region, the two local expert models with the highest probabilities are called in a Top-2 manner. The two five-dimensional output vectors are then weighted and fused using the normalized region probabilities as weights.

[0073] The hierarchical decoupling model is as follows: Figure 6 The method outputs the tactile contact positions X and Y, and the three-dimensional forces Fx, Fy, and Fz in the indicated order. This method is designed for single-point contact and does not output gesture categories or multi-point contact states. A test example achieved a mean absolute error of approximately 1.83 mm for position and approximately 0.14 N for force within a training range of 0 to 5 N, with a model inference latency of approximately 52 ms; the CCM (Continuous Computational Correction) reduced the overall error by approximately 32% compared to uncalibrated input. These values ​​correspond to specific sensing devices, models, and test conditions and do not limit the scope of protection of this invention.

[0074] In this embodiment, the specific details of the decoupling system are as follows: The system comprises a multi-channel planar sensing structure consisting of three non-collinear local triaxial sensing nodes, a simulation data generation module, a calibration parameter determination module, a model training module, a signal correction module, and a hierarchical decoupling module. The simulation data generation module can be deployed on a computer running finite element software and batch processing scripts; the calibration parameter determination module receives the simulation and actual responses under standard loading and saves the signal transmission characteristic correction matrix and the structural coupling interference correction matrix for each local triaxial sensing node; the model training module trains and saves the region classification model and multiple region regression models; the signal correction module is connected to the sensor signal acquisition module; and the hierarchical decoupling module runs on the processor and outputs the tactile contact positions X and Y and the three-dimensional forces Fx, Fy, and Fz.

[0075] For multiple identical planar multidimensional force and position sensing devices manufactured in batches, they can share the same finite element simulation training dataset and the same hierarchical decoupling model. Each physical device determines its own signal transmission characteristic correction coefficient and structural coupling interference correction matrix only through a small number of standard loads. The system can be deployed on robot end effects or tactile interaction interfaces for robot tactile interaction, human-machine input, and contact status monitoring.

[0076] Example 2 like Figure 7 As shown, this embodiment of the invention also provides a tactile contact position and three-dimensional force decoupling system, comprising: The three-node planar sensing structure includes three non-collinear local triaxial sensing nodes. The center points of the three nodes jointly determine the node reference plane and form a multi-channel planar sensing structure. The simulation data generation module is used to establish a finite element model of the planar multidimensional force position sensing device. Loads are applied at multiple sampling positions on the planar contact surface. Multi-channel simulation responses are generated through finite element solution and strain extraction. The two-dimensional position coordinates of each sampling position and the three-dimensional force components of the load are used as labels to construct a simulation training dataset. The calibration parameter determination module is used to obtain the simulated response and the actual response under the same standard load conditions, determine the signal transmission characteristic correction parameters based on the difference in channel amplitude between the two, and determine the structural coupling interference correction parameters based on the measured response direction of the local triaxial sensing node and the target direction. The calibration compensation matrix is ​​formed by the signal transmission characteristic correction parameters and the structural coupling interference correction parameters. The model training module is used to train a region classification model and a region regression model corresponding to different planar regions using only the simulation training dataset. The region classification model outputs region probabilities, and the region regression model outputs two-dimensional position coordinates and three-dimensional force components. The real response is not used to fine-tune the parameters of the region classification model and the region regression model. The signal correction module is used to acquire the real multi-channel signal generated by the actual single-point contact, and to perform signal transmission characteristic correction and structural coupling interference correction in sequence using the calibration compensation matrix, and to normalize the corrected signal. The hierarchical decoupling module is used to input the normalized multi-channel signal into the region classification model, select at least one region regression model according to the region classification result, and fuse the regression outputs when multiple region regression models are selected to obtain the tactile contact positions X, Y and three-dimensional forces Fx, Fy and Fz.

[0077] Each module sequentially performs simulation data generation, calibration compensation matrix determination, simulation data-only model training, real multi-channel signal correction, and region regression and five-dimensional decoupling output based on region classification results.

[0078] Furthermore, each local triaxial sensing node outputs sensing signals in three directions, and the three local triaxial sensing nodes together form a nine-channel sensing signal; the calibration parameter determination module is used to determine the signal transmission characteristic correction coefficients of the nine signal channels and the structural coupling interference correction matrix of the three local triaxial sensing nodes; the planar contact surface is divided into twelve sector regions, the region classification model outputs twelve region probabilities, and the hierarchical decoupling module performs Top-1 selection or Top-2 selection and weighted fusion according to the region probabilities.

[0079] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for decoupling tactile contact position from three-dimensional force, characterized in that, Includes the following steps: Step S1: Establish a finite element model of the planar multidimensional force position sensing device, apply loads at multiple sampling positions on the planar contact surface, generate multi-channel simulation response through finite element solution and strain extraction, and use the two-dimensional position coordinates of each sampling position and the three-dimensional force components of the load as labels to construct a simulation training dataset. Step S2: Under the same standard load conditions, obtain the simulated response and the actual response respectively, determine the signal transmission characteristic correction parameters based on the difference in channel amplitude between the two, and determine the structural coupling interference correction parameters based on the measured response direction of the local triaxial sensing node and the target direction. The calibration compensation matrix is ​​formed by the signal transmission characteristic correction parameters and the structural coupling interference correction parameters. Step S3: Train a region classification model and a region regression model corresponding to different planar regions using only the simulation training dataset. The region classification model outputs region probabilities, and the region regression model outputs two-dimensional position coordinates and three-dimensional force components. The real response is not used to fine-tune the parameters of the region classification model and the region regression model. Step S4: Obtain the actual multi-channel signal generated by the actual single-point contact, and use the calibration compensation matrix to perform signal transmission characteristic correction and structural coupling interference correction in sequence, and normalize the corrected signal. Step S5: Input the normalized multi-channel signal into the region classification model, select at least one region regression model according to the region classification result, and when multiple region regression models are selected, fuse the regression outputs to obtain the tactile contact position and three-dimensional force.

2. The method for decoupling tactile contact position and three-dimensional force according to claim 1, characterized in that, In step S1, sampling positions are set for the planar contact surface in both radial and circumferential directions, and the planar contact surface is divided into multiple planar regions according to the circumferential angle. The simulation training dataset includes normal force samples with only normal force applied and three-dimensional force samples with both tangential and normal force components applied simultaneously.

3. The method for decoupling tactile contact position and three-dimensional force according to claim 1, characterized in that, The planar multidimensional force position sensing device includes three local triaxial sensing nodes; The center points of the three local triaxial sensing nodes are located on the same node reference plane and are not collinear. The three center points together determine the node reference plane and constitute a three-node planar sensing structure. Each of the aforementioned local triaxial sensing nodes outputs sensing signals in three directions, and the three aforementioned local triaxial sensing nodes together form a nine-channel sensing signal; The tactile contact position is represented by a two-dimensional coordinate system established in the reference plane of the node.

4. The method for decoupling tactile contact position and three-dimensional force according to claim 3, characterized in that, In step S2, the signal transmission characteristic correction parameters include correction coefficients corresponding to nine signal channels respectively. Each correction coefficient is determined by the ratio of the simulated response to the actual response of the corresponding channel under the same standard load conditions. The nine correction coefficients constitute a diagonal signal transmission characteristic correction matrix. The structural coupling interference correction parameters include rotation correction matrices corresponding to the three local triaxial sensing nodes, and each rotation correction matrix is ​​determined by the rotation relationship between the measured response direction of the corresponding local triaxial sensing node under standard load and the target direction.

5. The method for decoupling tactile contact position and three-dimensional force according to claim 3, characterized in that, The planar contact surface is divided into twelve sector-shaped regions; The region classification model is a one-dimensional residual convolutional network trained solely on the simulation training dataset. It includes a one-dimensional convolutional layer, multiple residual blocks, a global average pooling layer, and a twelve-class linear layer, which outputs the region probabilities of twelve sector regions based on the nine-dimensional input vector.

6. The method for decoupling tactile contact position and three-dimensional force according to claim 5, characterized in that, Each of the aforementioned sector regions corresponds to a region regression model trained using only simulation samples from that sector region; Each of the aforementioned region regression models is a multilayer perceptron, comprising a nine-dimensional input layer, three hidden layers, and a five-dimensional output layer. The five-dimensional output layer sequentially outputs two components of two-dimensional position coordinates and three components of three-dimensional force.

7. The method for decoupling tactile contact position and three-dimensional force according to claim 5, characterized in that, In step S5, when the maximum region probability is not lower than the preset confidence threshold, a region regression model corresponding to the maximum region probability is selected, and the five-dimensional output of the region regression model is output as the tactile contact position and the three-dimensional force. When the maximum region probability is lower than the preset confidence threshold or the tactile contact position is near the boundary of an adjacent sector region, the two region regression models with the highest region probabilities are selected, and the five-dimensional outputs of the two region regression models are weighted and fused according to the corresponding region probabilities to obtain the tactile contact position and the three-dimensional force.

8. The method for decoupling tactile contact position and three-dimensional force according to claim 1, characterized in that, In step S3, the simulation training dataset is divided into a training set and a test set, the multi-channel simulation sensing signals are normalized to the target interval, and the mean square error loss function is used to train the regression model for each region. The region classification model and each of the region regression models are trained separately, and each of the region regression models is trained using at least the simulation samples of its corresponding planar region.

9. The method for decoupling tactile contact position and three-dimensional force according to claim 1, characterized in that, In step S2, under the same standard load conditions, the signal transmission characteristic correction parameter is used to adjust the amplitude of the real channel response to a range that matches the simulated channel response. The structural coupling interference correction parameters are used to correct the direction of the measured three-dimensional response vector of each local triaxial sensing node to the target direction, so as to eliminate interaxial coupling caused by installation direction and structural errors. The real multi-channel signal after being corrected by the calibration compensation matrix is ​​in the same amplitude and direction representation space as the multi-channel simulation response in the simulation training dataset.

10. A tactile contact position and three-dimensional force decoupling system, characterized in that, include: The simulation data generation module is used to establish a finite element model of the planar multidimensional force position sensing device. Loads are applied at multiple sampling positions on the planar contact surface. Multi-channel simulation responses are generated through finite element solution and strain extraction. The two-dimensional position coordinates of each sampling position and the three-dimensional force components of the load are used as labels to construct a simulation training dataset. The calibration parameter determination module is used to obtain the simulated response and the actual response under the same standard load conditions, determine the signal transmission characteristic correction parameters based on the difference in channel amplitude between the two, and determine the structural coupling interference correction parameters based on the measured response direction of the local triaxial sensing node and the target direction. The calibration compensation matrix is ​​formed by the signal transmission characteristic correction parameters and the structural coupling interference correction parameters. The model training module is used to train a region classification model and a region regression model corresponding to different planar regions using only the simulation training dataset. The region classification model outputs region probabilities, and the region regression model outputs two-dimensional position coordinates and three-dimensional force components. The real response is not used to fine-tune the parameters of the region classification model and the region regression model. The signal correction module is used to acquire the real multi-channel signal generated by the actual single-point contact, and to perform signal transmission characteristic correction and structural coupling interference correction in sequence using the calibration compensation matrix, and to normalize the corrected signal. The hierarchical decoupling module is used to input the normalized multi-channel signal into the region classification model, select at least one region regression model according to the region classification result, and fuse the regression outputs when multiple region regression models are selected to obtain the tactile contact position and three-dimensional force.

Images