A flexible capacitive two-dimensional force sensor decoupling method based on physical information neural network

Abstract

The application relates to the technical field of sensor decoupling and discloses a flexible capacitive two-dimensional force sensor decoupling method based on a physical information neural network, which comprises the following steps: using a standard force loading device to respectively apply known pressure and shear force, calibrating a flexible two-dimensional force sensor, collecting capacitive signals output by a plurality of sensing units in the flexible two-dimensional force sensor, and establishing a calibration data set; a physical information neural network PINN model is constructed, and a mechanical constraint condition and a physical model of a sensor structure are introduced as physical constraint terms in the network training process; the physical information neural network PINN model is trained by using a training set, so that the network learns the mapping relationship between the capacitive signals and two-dimensional forces; the capacitive signals collected in an actual measurement process are input into the trained physical information neural network PINN model, and the size and direction of two-dimensional forces borne by the flexible two-dimensional force sensor are output. The application effectively reduces the dependence on a large amount of accurate annotation data.

Description

Technical Field

[0001] This invention relates to the field of sensor decoupling technology, and more specifically, to a decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network. Background Technology

[0002] In fields such as soft robotics, flexible wearable devices, and biomedicine, there is an increasing demand for flexible force sensors capable of conforming to irregular surfaces and sensing multidimensional contact forces. Capacitive sensors, due to their advantages of high sensitivity, fast response, and low power consumption, have become one of the main solutions for flexible force monitoring.

[0003] Currently, the mainstream design for flexible capacitive sensors used to measure two-dimensional forces (pressure and shear force) typically involves arranging two sets of orthogonal electrode arrays on the upper and lower sides of a flexible dielectric layer, respectively for sensing forces in the X and Y directions (such as normal pressure or tangential shear force). However, due to limitations in the electrode arrangement, the measurement accuracy of the two-dimensional force direction is difficult to guarantee.

[0004] Due to their complex structure, multidimensional force sensors inevitably exhibit coupling phenomena. Decoupling methods for multidimensional forces include linear decoupling and nonlinear decoupling. Among these, using neural networks can significantly improve decoupling accuracy. However, traditional neural networks rely on a large amount of accurate data, which greatly increases the difficulty of calibration and decoupling.

[0005] Therefore, it is necessary to design a decoupling method for flexible capacitive two-dimensional force sensors based on physical information neural networks to solve the problems existing in the current technology. Summary of the Invention

[0006] In view of this, the present invention proposes a decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network, aiming to solve the problem of dependence on a large amount of accurate data in the current technology of traditional neural network decoupling.

[0007] This invention proposes a decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network, comprising: A flexible two-dimensional force sensor was calibrated by applying known pressure and shear force using a standard force loading device, collecting capacitance signals output by multiple sensing units in the flexible two-dimensional force sensor, and establishing a calibration dataset. The two-dimensional force parameters obtained during the calibration process are correlated with their corresponding capacitance signals to construct a sample dataset, which is then divided into a training set and a test set. A physical information neural network (PINN) model is constructed, and mechanical constraints and sensor structure physical models are introduced as physical constraint terms during network training. The physical information neural network PINN model is trained using a training set, enabling the network to learn the mapping relationship between capacitance signals and two-dimensional forces. The capacitance signal collected during the actual measurement process is input into the trained physical information neural network PINN model, and the magnitude and direction of the two-dimensional force on the flexible two-dimensional force sensor are output to achieve decoupling of pressure and shear force.

[0008] Furthermore, the flexible two-dimensional force sensor includes: Top cover, support column, hemispherical bump structure, upper electrode, buffer pad, dielectric material, base, lower electrode, circuit, flexible flat cable and data acquisition chip.

[0009] Furthermore, the hemispherical protrusion structure is disposed on the force-bearing side surface of the top cover to receive external forces; the upper electrode is disposed on the inner side of the top cover and integrally formed with the top cover; the upper electrode and the lower electrode disposed on the base are arranged vertically correspondingly; the dielectric material is sandwiched between the upper electrode and the lower electrode to generate capacitance changes by compression or shear deformation under force; the top cover is connected to the base through the support column and maintains a preset gap; a buffer pad is disposed between the dielectric material and the lower electrode; the lower electrode is electrically connected to one end of the flexible flat cable through a circuit disposed on the base; the other end of the flexible flat cable is connected to the data acquisition chip.

[0010] Furthermore, the flexible two-dimensional force sensor is calibrated by applying known pressure and shear force using a standard force loading device. This involves collecting capacitance signals output from multiple sensing units within the flexible two-dimensional force sensor and establishing a calibration dataset, including: The flexible two-dimensional force sensor is mounted on the test bench of the standard force loading device; Adjust the position of the standard force loading device so that the force application end of the standard force loading device contacts the hemispherical protrusion structure; Set the vertical pressure amplitude sequence and apply the pressure sequentially; Set the shear force amplitude sequence and direction, and apply the shear force sequentially; Acquire the capacitance signal corresponding to each pressure and shear force combination, and pair each applied pressure and shear force with the corresponding capacitance signal to form a calibration dataset.

[0011] Furthermore, when setting a sequence of vertical pressure amplitudes and applying pressure sequentially, the process includes: The thickness and compressive modulus of the dielectric material, the force response speed of the top cover, the applied pressure amplitude, and the force application speed of the standard force loading device are collected. A pressure application interval feature vector is constructed based on the thickness, compressive modulus, and applied pressure amplitude. The pressure application interval feature vector is compared with the historical pressure application interval group, and the pressure base application interval for each group of pressures is determined based on the comparison results. A velocity influence index is constructed based on the force response speed and the force application speed. The optimization factor for the pressure base application interval is determined based on the velocity influence index, and the pressure application interval is obtained. Pressure is applied sequentially at the specified pressure application intervals.

[0012] Furthermore, when determining the pressure base application interval for each group of pressures based on the comparison results, the following steps are included: When there is a historical pressure application interval feature vector in the historical pressure application interval group that is the same as the pressure application interval feature vector, the historical pressure base application interval corresponding to the historical pressure application interval feature vector shall be used as the pressure base application interval; When there is no historical pressure application interval feature vector in the historical pressure application interval group that is the same as the pressure application interval feature vector, the pressure base application interval is determined according to the pressure application interval feature vector.

[0013] Further, when determining the pressure base application interval based on the pressure application interval feature vector, the process includes: Calculate the Euclidean distance between the current pressure application interval feature vector and the feature vector of each of the historical pressure application intervals in the historical pressure application interval group; The calculated Euclidean distances are sorted in ascending order, and the first preset number of historical pressure application interval feature vectors are selected as similarity feature vectors. Extract the historical pressure base application intervals corresponding to the similar feature vectors, calculate their average value, and use the average value as the pressure base application interval.

[0014] Further, determining the optimization factor for the pressure base application interval based on the velocity influence index, and obtaining the pressure application interval, includes: The speed influence index is compared with the first speed influence index and the second speed influence index, and the optimization factor is determined based on the comparison result; wherein the first speed influence index is smaller than the second speed influence index. When the speed influence index is less than or equal to the first speed influence index, the optimization factor is determined to be the first optimization factor; When the speed influence index is greater than the first speed influence index and less than or equal to the second speed influence index, the optimization factor is determined to be the second optimization factor. When the speed influence index is greater than the second speed influence index, the optimization factor is determined to be the third optimization factor; Multiply the pressure base application interval by the optimization factor to obtain the final pressure application interval.

[0015] Furthermore, when constructing the Physical Information Neural Network (PINN) model and introducing mechanical constraints and sensor structure physical models as physical constraint terms during network training, the following are included: The range of compression and shear deformation of the dielectric material of the flexible two-dimensional force sensor under stress is set, and it is input as a mechanical constraint into the physical information neural network PINN model. The geometric dimensions, layout, and interconnection of the top cover, support column, hemispherical protrusion, buffer pad, dielectric material, base, and upper and lower electrodes are set as the physical model of the sensor structure, and the relative positions and force transmission relationships of each structural unit are maintained during network training. The mechanical constraints and the physical model of the sensor structure are jointly input into the physical information neural network PINN model to constrain the mapping relationship between the two-dimensional force and capacitance signal predicted by the network.

[0016] Furthermore, training the physical information neural network PINN model using the training set, so that the network learns the mapping relationship between capacitance signals and two-dimensional forces, includes: Each set of capacitance signals and corresponding two-dimensional forces in the training set are used as input and output samples and input into the physical information neural network PINN model. During network training, the mechanical constraints and the physical model of the sensor structure are used as physical constraints and added to the loss function to restrict the network prediction results to meet the actual mechanical properties and structural constraints of the sensor. The backpropagation algorithm is used to adjust the network parameters so that the error between the network output two-dimensional force prediction value and the actual two-dimensional force corresponding to the training set is minimized, while satisfying the physical constraints. The prediction error and physical constraint violation of the network during the training process are iteratively monitored, and the training is completed according to the preset number of training rounds or error convergence criteria. Output the trained physical information neural network PINN model.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively reduces the reliance on large amounts of precisely labeled data by introducing a Physical Information Neural Network (PINN) into the decoupling process of a flexible capacitive two-dimensional force sensor. Traditional neural network decoupling methods often require massive amounts of training samples to ensure accuracy. However, this invention introduces mechanical constraints (such as the compression and shear deformation range of the dielectric material) and the physical model of the sensor structure (including the geometry, layout, and interconnections of the top cover, support pillars, hemispherical protrusions, buffer pads, dielectric material, base, and upper and lower electrodes) as physical constraints into the PINN model. This integrates prior physical knowledge into the network training, enabling the model to achieve high-precision force decoupling even on a limited calibration dataset. This method not only reduces the workload of data acquisition and labeling, lowering experimental costs and time consumption, but also effectively improves the model's generalization ability and prediction accuracy for unseen data. Especially under complex force conditions, it more realistically reflects the actual mechanical behavior and signal transmission laws of the sensor, thereby improving the accuracy and reliability of two-dimensional force measurement. Furthermore, during the sensor calibration process, this invention determines the optimized pressure application interval by constructing a pressure application interval feature vector, comparing historical data, and combining the force response speed and pressure application speed. This further ensures the accuracy and effectiveness of the calibration dataset and lays a solid foundation for subsequent model training. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the flexible two-dimensional force sensor provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a lateral explosion of a flexible two-dimensional force sensor provided in an embodiment of the present invention.

[0019] In the diagram: 1. Top cover; 2. Support column; 3. Hemispherical bump structure; 4. Upper electrode; 5. Buffer pad; 6. Dielectric material; 7. Base; 8. Lower electrode; 9. Circuit; 10. Flexible flat cable; 11. Data acquisition chip. Detailed Implementation

[0020] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] See Figure 1-3 As shown, in some embodiments of this application, this embodiment provides a decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network, including: S100: Using a standard force loading device, apply known pressure and shear force respectively to calibrate the flexible two-dimensional force sensor, collect the capacitance signals output by multiple sensing units in the flexible two-dimensional force sensor, and establish a calibration dataset. S200: Correlate the two-dimensional force parameters obtained during the calibration process with their corresponding capacitance signals to construct a sample dataset, and divide the sample dataset into a training set and a test set; S300: Construct a physical information neural network (PINN) model, and introduce mechanical constraints and sensor structure physical models as physical constraint terms during network training; S400: The physical information neural network PINN model is trained using the training set, so that the network learns the mapping relationship between capacitance signal and two-dimensional force; S500: Inputs the capacitance signal collected during the actual measurement process into the trained physical information neural network PINN model, and outputs the magnitude and direction of the two-dimensional force on the flexible two-dimensional force sensor to achieve decoupling of pressure and shear force.

[0022] Understandably, this embodiment effectively reduces the reliance on large amounts of precisely labeled data by introducing a Physics-Informed Neural Network (PINN) into the decoupling process of the flexible capacitive two-dimensional force sensor. Traditional neural network decoupling methods often require massive amounts of training samples to ensure accuracy. However, this invention introduces mechanical constraints (such as the compression and shear deformation range of dielectric material 6) and the physical model of the sensor structure (including the geometric dimensions, layout, and interconnections of the top cover 1, support column 2, hemispherical protrusion, buffer pad 5, dielectric material 6, base 7, and upper and lower electrodes 8) as physical constraints into the PINN model. This integrates prior physical knowledge into the network training, enabling the model to achieve high-precision force decoupling on a limited calibration dataset. This method not only reduces the workload of data acquisition and labeling, lowers experimental costs and time consumption, but also effectively improves the model's generalization ability and prediction accuracy for unseen data. Especially under complex force conditions, it can more realistically reflect the actual mechanical behavior and signal transmission laws of the sensor, thereby improving the accuracy and reliability of two-dimensional force measurement. Furthermore, in the sensor calibration process, this embodiment constructs a pressure application interval feature vector, compares historical data, and determines an optimized pressure application interval by combining the force response speed and pressure application speed. This further ensures the accuracy and effectiveness of the calibration dataset and lays a solid foundation for subsequent model training.

[0023] Specifically, flexible two-dimensional force sensors include: 1. Top cover, 2. Support column, 3. Hemispherical protrusion structure, 4. Upper electrode, 5. Buffer pad, 6. Dielectric material, 7. Base, 8. Lower electrode, 9. Circuit, 10. Flexible flat cable, and 11. Data acquisition chip.

[0024] Specifically, the hemispherical bump structure 3 is disposed on the force-bearing side surface of the top cover 1 to receive external forces; the upper electrode 4 is disposed on the inner side of the top cover 1 and is integrally formed with the top cover 1; the upper electrode 4 and the lower electrode 8 disposed on the base 7 are arranged vertically in correspondence; the dielectric material 6 is sandwiched between the upper electrode 4 and the lower electrode 8 to generate capacitance changes by compression or shear deformation when subjected to force; the top cover 1 is connected to the base 7 through the support column 2 and maintains a preset gap; a buffer pad 5 is disposed between the dielectric material 6 and the lower electrode 8; the lower electrode 8 is electrically connected to one end of the flexible flat cable 10 through the circuit 9 disposed on the base 7; the other end of the flexible flat cable 10 is connected to the data acquisition chip 11.

[0025] Understandably, the flexible two-dimensional force sensor includes a top cover 1, a base 7, an upper electrode 4, a lower electrode 8, and a circuit 9. The top cover 1 and base 7 are formed by curing polydimethylsiloxane material through a mold. The hemispherical protrusion structure 3 is arranged in a ring to sense the magnitude and direction of external forces (compression and shear forces). The overall size of the sensor unit is 20*20*2.9mm. The upper electrode 4, lower electrode 8, and circuit 9 are all made of gallium indium alloy ink through 3D printing. The ionogel, serving as the dielectric material 6 for the capacitor, comprises eight pieces arranged in a ring inside the sensor. The presence of the buffer pad 5 effectively adjusts the sensor's load-bearing capacity. The dielectric material 6 is an ionogel with excellent flexibility and biocompatibility. Preferably, the dielectric material 6 should cover each electrode area, with a thickness not exceeding 0.5mm and a range of 1KPa.

[0026] Understandably, the data acquisition chip 11 is equipped with a casing and a battery. The chip's capacitance scanning voltage is 1.2V, employing multi-frequency scanning with a frequency between 1.0MHz and 3.0MHz; the data transmission rate is no less than 1Mbps, and the communication distance is no less than 5m. When the sensor is subjected to pressure from the vertical direction, the top cover 1 uniformly presses down on the eight sensing units, causing all eight sensing units to produce equal displacement changes, i.e., equal capacitance changes. When the sensor's hemispherical protrusion structure 3 is subjected to tangential force, the top cover 1 rotates slightly around the support column 2, causing the longitudinal displacement of the sensing unit on one side of the tangential force to be significantly greater than that on the other side. This results in the eight sensing units outputting capacitance signal values ​​of different magnitudes. Similarly, under different load conditions, the sensing units will produce different longitudinal displacements, outputting capacitance signals of different magnitudes. In this way, all eight sensing points can independently input and output signals, enabling the sensing of capacitance changes in each sensing unit, and thus sensing the magnitude and direction of pressure and shear force.

[0027] It is understandable that the signal acquisition of each sensing point of the flexible two-dimensional force sensor is independent and does not affect each other. The data acquisition chip 11 processes the collected electrical signals and then transmits them to the host. The host then packages all the data and sends it to the terminal device to complete the real-time measurement.

[0028] Specifically, when calibrating a flexible two-dimensional force sensor by applying known pressure and shear force using a standard force loading device, collecting capacitance signals output from multiple sensing units in the flexible two-dimensional force sensor, and establishing a calibration dataset, the process includes: The flexible two-dimensional force sensor was installed on the test bench of the standard force loading equipment; Adjust the position of the standard force loading device so that the force application end of the standard force loading device contacts the hemispherical protrusion structure 3; Set the vertical pressure amplitude sequence and apply the pressure sequentially; Set the shear force amplitude sequence and direction, and apply the shear force sequentially; Acquire the capacitance signal corresponding to each pressure and shear force combination, and pair each applied pressure and shear force with the corresponding capacitance signal to form a calibration dataset.

[0029] Understandably, the construction and training of the Physical Information Neural Network (PINN) is a core step in achieving sensor decoupling. This network incorporates the physical model information of the sensors into the loss function of the neural network to improve decoupling accuracy and generalization ability. Specifically, the input layer of PINN receives capacitance signals from eight sensing units. After nonlinear mapping in the hidden layers, these signals yield the corresponding pressure magnitude, shear force magnitude, and shear force direction angle in the output layer. During network training, in addition to supervised learning using samples from the calibration dataset (i.e., minimizing the error between predicted and actual forces), physical constraints based on the sensor's mechanical equilibrium principle and capacitance change law are introduced. For example, when the sensor is subjected to pure pressure, theoretically, the capacitance changes of the eight sensing units should be symmetrical. This prior knowledge is transformed into a regularization term in the loss function, guiding the network to learn mapping relationships that conform to physical laws. In this way, PINN can effectively uncover the complex nonlinear relationship between capacitance signals and multidimensional forces, maintaining high decoupling accuracy even under conditions not covered by the calibration dataset.

[0030] Specifically, when setting a sequence of vertical pressure amplitudes and applying pressure sequentially, the process includes: The thickness and compressive modulus of dielectric material 6, the force response speed of top cover 1, the applied pressure amplitude, and the force application speed of standard force loading device are collected. Construct a pressure application interval feature vector based on thickness, compressive modulus, and applied pressure amplitude; The pressure application interval feature vector is compared with the historical pressure application interval group, and the pressure base application interval for each group of pressures is determined based on the comparison results. A velocity influence index is constructed based on the force response speed and the force application speed. The optimization factor for the pressure base application interval is determined based on the velocity influence index, and the pressure application interval is obtained. Apply pressure sequentially at pressure application intervals.

[0031] It is understandable that the historical pressure application interval group refers to a collection of pressure application interval data that has been verified or optimized in practice, recorded and accumulated under different combinations of conditions such as the characteristics of different dielectric materials (e.g., thickness, compressive modulus), different applied pressure amplitude ranges, different force application speeds of standard force loading devices, and the force response speed of the sensor itself, when calibrating similar flexible capacitive two-dimensional force sensors in the past.

[0032] The process of constructing the speed influence index can be understood as follows: First, determine the quantitative indicators of force response speed and force application speed, such as force response speed (the time required for the sensor to stabilize from the application of force to the stability of the capacitive signal, denoted by tresponse, in milliseconds) and force application speed (the change in force applied by a standard force loading device per unit time, denoted by vapply, in N / s). Next, normalize these two indicators to eliminate the influence of dimensions. Then, set weighting coefficients based on the sensor's dynamic characteristics. If the sensor is more sensitive to the force application speed, assign a larger weight to vnorm (wv); if it is sensitive to the force response speed, assign a larger weight to tnorm (wt), and wt + wv = 1. Finally, calculate the speed influence index using the formula speed-influence-index = wt * tnorm + wv * vnorm. This index ranges from 0 to 1; the larger the index, the greater the influence of the speed factor on the stability of the sensor response during force application, requiring greater optimization and adjustment of the pressure base application interval.

[0033] Specifically, when determining the pressure base application interval for each group of pressures based on the comparison results, the following is included: When there is a historical pressure application interval feature vector in the historical pressure application interval group that is the same as the pressure application interval feature vector, the historical pressure base application interval corresponding to the historical pressure application interval feature vector shall be taken as the pressure base application interval. When there is no historical pressure application interval feature vector in the historical pressure application interval group that is the same as the pressure application interval feature vector, the pressure base application interval is determined according to the pressure application interval feature vector.

[0034] Specifically, determining the pressure base application interval based on the pressure application interval feature vector includes: Calculate the Euclidean distance between the current pressure application interval feature vector and the feature vector of each historical pressure application interval in the historical pressure application interval group; The calculated Euclidean distances are sorted in ascending order, and the first preset number of historical pressure application interval feature vectors are selected as similarity feature vectors. Extract the historical pressure base application intervals corresponding to similar feature vectors, calculate their average value, and use this average value as the pressure base application interval.

[0035] Understandably, the selection of the preset number should comprehensively consider the amount of historical data and the similarity of the feature vectors. If the historical data sample is sufficient and the distribution of each feature vector is relatively discrete, the preset number can be appropriately increased to ensure that similar feature vectors can comprehensively reflect the pressure base application interval pattern under different conditions. If the historical data sample is limited or some feature vectors are relatively concentrated, the preset number should be reduced to avoid introducing too much historical data with large differences that may affect the representativeness of the average value. For example, when the historical pressure application interval group contains more than 100 sets of valid data, the preset number can be set to 5-8 sets; if the data volume is less than 50 sets, the preset number can be adjusted to 3-5 sets, thereby improving the efficiency and reliability of determining the pressure base application interval while ensuring the accuracy of the calculation.

[0036] Specifically, when determining the optimization factor for the pressure application interval based on the velocity influence index and obtaining the pressure application interval, the following steps are included: The speed influence index is compared with the first speed influence index and the second speed influence index, and the optimization factor is determined based on the comparison results; wherein, the first speed influence index is smaller than the second speed influence index. When the speed influence index is less than or equal to the first speed influence index, the optimization factor is determined as the first optimization factor; When the speed influence index is greater than the first speed influence index and less than or equal to the second speed influence index, the optimization factor is determined as the second optimization factor. When the speed influence index is greater than the second speed influence index, the optimization factor is determined as the third optimization factor; Multiply the pressure base application interval by the optimization factor to obtain the final pressure application interval.

[0037] Understandably, the first and second velocity influence indices are critical values ​​used to divide the velocity influence index range, set based on the sensor's dynamic response characteristics and the requirements of actual application scenarios. For example, through statistical analysis of a large amount of experimental data, the stability of the sensor's capacitive signal under different velocity influence indices is used as the criterion: when the velocity influence index is less than 0.3, the sensor response is less affected by velocity interference, so the first velocity influence index is set to 0.3, the first optimization factor is set to 1.0, and the pressure application interval is not adjusted; when the velocity influence index is between 0.3 and 0.7, velocity has a certain impact on the stability of the sensor response, so the second optimization factor is set to 1.2, and the pressure application interval is appropriately extended to ensure signal stability; when the velocity influence index is greater than 0.7, velocity interference is significant, so the third optimization factor is set to 1.5, and the pressure application interval is increased to ensure accurate and reliable capacitive signals. The specific values ​​of these critical values ​​and optimization factors can be dynamically adjusted and optimized according to the sensor model, dielectric material characteristics, and the performance of the standard force loading equipment.

[0038] Understandably, the process of determining the shear force application interval is similar to that of determining the pressure application interval. It can be determined based on parameters such as the shear modulus of the dielectric material 6, the stiffness of the buffer pad 5, the sensor's response speed to the shear force, the amplitude and direction of the applied shear force, and the force application rate of the loading device. The specific steps are as follows: First, collect the shear modulus of the dielectric material 6, the stiffness of the buffer pad 5, the sensor's response speed to the shear force (i.e., the time required for the sensor to stabilize from the application of shear force, denoted by tshear_response in milliseconds), the amplitude of the applied shear force, the shear force direction angle, and the shear force application rate of the loading device (in N / s); then, construct a shear force application interval feature vector based on the shear modulus, the stiffness of the buffer pad 5, and the amplitude of the applied shear force; finally, combine this feature vector with historical shear force application intervals. (This refers to the set of effective shear force application interval data accumulated in the past when calibrating similar sensors under different conditions such as shear characteristics of dielectric materials, parameters of buffer pads, shear force amplitude range, application speed and sensor response speed.) If the same historical feature vector exists, its corresponding historical shear force application interval is directly used; if not, the Euclidean distance between the current feature vector and each historical feature vector is calculated, sorted, and the first preset number of similar feature vectors (determined according to the amount and dispersion of historical data, such as 58 groups when historical data is sufficient, and 35 groups when data is limited) are selected, and the average value of their corresponding historical shear force application intervals is taken as the current shear force application interval. Subsequently, a shear speed influence index is constructed. The determination process for this index is similar to that of the pressure speed influence index: first, the shear response speed (tshear_response) and the shear application speed are normalized (the normalization formulas refer to tnorm and vnorm in the pressure speed influence index calculation). Then, weighting coefficients are set based on the sensor's sensitivity to the shear application speed and its own response speed (the sum of the two is 1). The index is calculated using the formula shear_speed_influence_index = wt_shear * tnorm_shear + wv_shear * vnorm_shear, with a range between 0 and 1. Finally, the shear speed influence index is compared with preset first shear speed influence indices (e.g., 0.3) and second shear speed influence indices (e.g., 0.7), and corresponding first shear optimization factors (e.g., 1.0), second shear optimization factors (e.g., 1.2), or third shear optimization factors (e.g., 1.5) are determined. Multiplying the basic shear application interval by this optimization factor yields the final shear application interval, at which shear force is applied sequentially.

[0039] Specifically, when constructing the Physical Information Neural Network (PINN) model, and introducing mechanical constraints and sensor structure physical models as physical constraint terms during network training, the following are included: The range of compression and shear deformation of the dielectric material 6 of the flexible two-dimensional force sensor under stress is defined and input as mechanical constraint conditions into the physical information neural network PINN model. The geometric dimensions, layout, and interconnection of the top cover 1, support column 2, hemispherical protrusion, buffer pad 5, dielectric material 6, base 7, and upper and lower electrodes 8 are set as the physical model of the sensor structure, and the relative positions and force transmission relationships of each structural unit are maintained during network training. The mechanical constraints and the physical model of the sensor structure are jointly input into the PINN physical information neural network model to constrain the mapping relationship between the two-dimensional force and capacitance signals predicted by the network.

[0040] Understandably, the combined input of mechanical constraints and the physical model of the sensor structure essentially transforms the sensor's "hardware characteristics" into "prior knowledge" for the network to learn. For example, the compression and shear deformation range of dielectric material 6 directly limits the possible range of capacitance signal changes. If the deformation corresponding to the force value predicted by the network exceeds this range, even if similar samples exist in the calibration data, the physical constraint term will significantly increase the loss function value, forcing the network to correct the prediction results. The geometric dimensions and layout of each structural unit of the sensor (such as the height of support column 2, the radius of curvature of the hemispherical protrusion, the area and spacing of the electrodes, etc.) determine the force transmission path and the spatial distribution characteristics of capacitance changes. During training, the network must follow the physical laws inherent in these structural parameters—for example, when shear force is applied in a certain direction, the capacitance change of the sensing unit near the force application end should have a specific difference pattern compared to the end far away. This pattern is predefined by the structural physical model to avoid the network learning false mapping relationships that do not conform to the actual structural response. In this way, the PINN model not only leverages the advantages of data-driven approaches but also incorporates the inherent physical nature of sensors, thus maintaining the physical consistency and reliability of the decoupling results even when faced with complex loading conditions or data noise.

[0041] Specifically, training the PINN physical information neural network model using a training set, so that the network learns the mapping relationship between capacitance signals and two-dimensional forces, includes: Each set of capacitance signals and corresponding two-dimensional forces in the training set are used as input and output samples, which are then input into the PINN physical information neural network model. During network training, mechanical constraints and the physical model of the sensor structure are used as physical constraints and added to the loss function to ensure that the network prediction results meet the actual mechanical properties and structural constraints of the sensor. The backpropagation algorithm is used to adjust the network parameters so that the error between the network output two-dimensional force prediction value and the actual two-dimensional force corresponding to the training set is minimized, while satisfying the physical constraints. The prediction error and physical constraint violation of the network during the training process are iteratively monitored, and the training is completed according to the preset number of training rounds or error convergence criteria. Output the trained physical information neural network PINN model.

[0042] Understandably, error monitoring during training not only focuses on the numerical deviation between predicted force and actual force (such as root mean square error RMSE), but also assesses the degree of violation of physical constraints. For example, when the pressure value predicted by the network corresponds to a compression of dielectric material 6 that exceeds the set mechanical constraint range, the ratio of this excess value to the maximum allowable deformation is calculated as a constraint violation index and weighted into the overall loss function. This dual monitoring mechanism ensures that the network, while learning data patterns, never deviates from the physical nature of the sensor. Furthermore, during training iterations, a dynamic learning rate strategy can be adopted. During the rapid error reduction phase, the learning rate can be appropriately increased to accelerate convergence. When the error enters a plateau or the degree of physical constraint violation suddenly increases, the learning rate is reduced for fine-tuning to balance training efficiency and model stability. Finally, when the prediction error of several consecutive training rounds is below a preset threshold and the degree of physical constraint violation approaches zero, the model training is considered complete. At this point, the PINN model has the ability to accurately map the input capacitance signal into the magnitude and direction angle of pressure and shear force, and its decoupling results conform to the actual mechanical response law of the sensor.

[0043] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0044] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0045] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0046] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network, characterized in that, include: A flexible two-dimensional force sensor was calibrated by applying known pressure and shear force using a standard force loading device, collecting capacitance signals output by multiple sensing units in the flexible two-dimensional force sensor, and establishing a calibration dataset. The two-dimensional force parameters obtained during the calibration process are correlated with their corresponding capacitance signals to construct a sample dataset, which is then divided into a training set and a test set. A physical information neural network (PINN) model is constructed, and mechanical constraints and sensor structure physical models are introduced as physical constraint terms during network training. The physical information neural network PINN model is trained using a training set, enabling the network to learn the mapping relationship between capacitance signals and two-dimensional forces. The capacitance signal collected during the actual measurement process is input into the trained physical information neural network PINN model, and the magnitude and direction of the two-dimensional force on the flexible two-dimensional force sensor are output to achieve decoupling of pressure and shear force.

2. The decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network according to claim 1, characterized in that, The flexible two-dimensional force sensor includes: Top cover, support column, hemispherical bump structure, upper electrode, buffer pad, dielectric material, base, lower electrode, circuit, flexible flat cable and data acquisition chip.

3. The decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network according to claim 2, characterized in that, The hemispherical protrusion structure is disposed on the force-bearing side surface of the top cover to receive external forces; the upper electrode is disposed on the inner side of the top cover and is integrally formed with the top cover; the upper electrode and the lower electrode disposed on the base are arranged vertically correspondingly; the dielectric material is sandwiched between the upper electrode and the lower electrode to generate capacitance changes by compression or shear deformation under force; the top cover is connected to the base through the support column and maintains a preset gap; a buffer pad is disposed between the dielectric material and the lower electrode; the lower electrode is electrically connected to one end of the flexible flat cable through a circuit disposed on the base; the other end of the flexible flat cable is connected to the data acquisition chip.

4. The decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network according to claim 3, characterized in that, The flexible two-dimensional force sensor is calibrated by applying known pressure and shear force using a standard force loading device. This involves acquiring capacitance signals output from multiple sensing units within the sensor and establishing a calibration dataset, including: The flexible two-dimensional force sensor is mounted on the test bench of the standard force loading device; Adjust the position of the standard force loading device so that the force application end of the standard force loading device contacts the hemispherical protrusion structure; Set the vertical pressure amplitude sequence and apply the pressure sequentially; Set the shear force amplitude sequence and direction, and apply the shear force sequentially; Acquire the capacitance signal corresponding to each pressure and shear force combination, and pair each applied pressure and shear force with the corresponding capacitance signal to form a calibration dataset.

5. The decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network according to claim 4, characterized in that, When setting a sequence of vertical pressure amplitudes and applying pressure sequentially, the following steps are included: The thickness and compressive modulus of the dielectric material, the force response speed of the top cover, the applied pressure amplitude, and the force application speed of the standard force loading device are collected. A pressure application interval feature vector is constructed based on the thickness, compressive modulus, and applied pressure amplitude. The pressure application interval feature vector is compared with the historical pressure application interval group, and the pressure base application interval for each group of pressures is determined based on the comparison results. A velocity influence index is constructed based on the force response speed and the force application speed. The optimization factor for the pressure base application interval is determined based on the velocity influence index, and the pressure application interval is obtained. Pressure is applied sequentially at the specified pressure application intervals.

6. The decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network according to claim 5, characterized in that, When determining the pressure base application interval for each group of pressures based on the comparison results, the following are included: When there is a historical pressure application interval feature vector in the historical pressure application interval group that is the same as the pressure application interval feature vector, the historical pressure base application interval corresponding to the historical pressure application interval feature vector shall be used as the pressure base application interval; When there is no historical pressure application interval feature vector in the historical pressure application interval group that is the same as the pressure application interval feature vector, the pressure base application interval is determined according to the pressure application interval feature vector.

7. The decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network according to claim 6, characterized in that, Determining the pressure base application interval based on the pressure application interval feature vector includes: Calculate the Euclidean distance between the current pressure application interval feature vector and the feature vector of each historical pressure application interval in the historical pressure application interval group; The calculated Euclidean distances are sorted in ascending order, and the first preset number of historical pressure application interval feature vectors are selected as similarity feature vectors. Extract the historical pressure base application intervals corresponding to the similar feature vectors, calculate their average value, and use the average value as the pressure base application interval.

8. The decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network according to claim 7, characterized in that, When determining the optimization factor for the pressure application interval based on the velocity influence index, and obtaining the pressure application interval, the following steps are included: The speed influence index is compared with the first speed influence index and the second speed influence index, and the optimization factor is determined based on the comparison result; wherein the first speed influence index is smaller than the second speed influence index. When the speed influence index is less than or equal to the first speed influence index, the optimization factor is determined to be the first optimization factor; When the speed influence index is greater than the first speed influence index and less than or equal to the second speed influence index, the optimization factor is determined to be the second optimization factor. When the speed influence index is greater than the second speed influence index, the optimization factor is determined to be the third optimization factor; Multiply the pressure base application interval by the optimization factor to obtain the final pressure application interval.

9. The decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network according to claim 8, characterized in that, When constructing a Physical Information Neural Network (PINN) model, and introducing mechanical constraints and sensor structural physical models as physical constraint terms during network training, the following are included: The range of compression and shear deformation of the dielectric material of the flexible two-dimensional force sensor under stress is set, and it is input as a mechanical constraint into the physical information neural network PINN model. The geometric dimensions, layout, and interconnection of the top cover, support column, hemispherical protrusion, buffer pad, dielectric material, base, and upper and lower electrodes are set as the physical model of the sensor structure, and the relative positions and force transmission relationships of each structural unit are maintained during network training. The mechanical constraints and the physical model of the sensor structure are jointly input into the physical information neural network PINN model to constrain the mapping relationship between the two-dimensional force and capacitance signal predicted by the network.

10. The decoupling method for a flexible capacitive two-dimensional force sensor based on a physical information neural network according to claim 9, characterized in that, Training the physical information neural network PINN model using a training set, so that the network learns the mapping relationship between capacitance signals and two-dimensional forces, includes: Each set of capacitance signals and corresponding two-dimensional forces in the training set are used as input and output samples and input into the physical information neural network PINN model. During network training, the mechanical constraints and the physical model of the sensor structure are used as physical constraints and added to the loss function to restrict the network prediction results to meet the actual mechanical properties and structural constraints of the sensor. The backpropagation algorithm is used to adjust the network parameters so that the error between the network output two-dimensional force prediction value and the actual two-dimensional force corresponding to the training set is minimized, while satisfying the physical constraints. The prediction error and physical constraint violation of the network during the training process are iteratively monitored, and the training is completed according to the preset number of training rounds or error convergence criteria. Output the trained physical information neural network PINN model.

Images