Network for removing electrical cross-talk in flexible capacitive sensor arrays and method of training the same

Abstract

The application provides a network for removing electrical crosstalk in a flexible capacitive sensing array and a training method thereof, the network comprising: sequentially connected preliminary feature extraction and dimension transformation convolution layers, a backbone network and an output convolution layer; the backbone network is composed of d layers of MACB modules stacked, d is a positive integer greater than or equal to 1, and the number of layers d of the MACB module is determined according to the size of the flexible capacitive sensing array to be removed from electrical crosstalk; the MACB module sequentially comprises a crosstalk information extraction unit, a channel connection unit and a cross-channel feature fusion and dimension reduction unit; the crosstalk information extraction unit comprises four asymmetric convolution kernels of different sizes connected in parallel, compared with the limitations of the traditional solution, the network for removing electrical crosstalk in a flexible capacitive sensing array simultaneously utilizes the powerful fitting capability of a deep neural network and the spatial structure characteristics of a flexible capacitive sensing array, and better performance is obtained.

Description

Technical Field

[0001] This invention relates to the field of crosstalk removal technology, specifically to a network for removing electrical crosstalk in a flexible capacitive sensing array and its training method. Background Technology

[0002] In recent years, flexible pressure sensors have attracted widespread attention. Compared to traditional rigid MEMS (Micro-ElectroMechanical System) pressure sensors, flexible pressure sensors can conform well to irregular surfaces, making them suitable for large-area, multi-point in-situ pressure sensing. This advantage makes flexible pressure sensors promising for applications in robotics, healthcare, and wearable devices. For example, electronic skin incorporating pressure sensors can be applied to robotic hands or human prostheses to provide tactile sensation; foot pressure sensing systems can accurately detect pressure at different locations on the soles of the feet during walking, which can be used for motion analysis or diagnosis of conditions such as diabetes. Based on their operating principles, flexible pressure sensors can be categorized into capacitive, resistive, and piezoelectric types. Compared to the other two types, capacitive sensors offer advantages such as simple fabrication (sandwich structure), support for both static and dynamic sensing, and no static power consumption, thus attracting significant research attention. For large-area pressure sensing systems, there are typically dozens or even hundreds of sensor components. If each component is directly connected to the measurement circuit, the required wiring and pin count is twice the number of sensor components, posing a significant challenge to practical sensing systems. To reduce the number of interfaces, the mainstream approach is to use a crossbar array, where the upper electrodes of capacitors in the same row are connected to a common interconnect, and the lower electrodes of capacitors in the same column are also connected to a common interconnect. This reduces the number of interfaces for an n×n array from 2n... 2 Reduced to 2n.

[0003] However, the crossbar structure also leads to mechanical and electrical crosstalk. Electrical crosstalk, simply put, means that the measured value of the target capacitance is affected by all the capacitances in the array. Although electrical crosstalk is a fundamental and widespread problem in flexible capacitive sensing arrays, related research is scarce.

[0004] Yu et al. proposed a grounding method for a flexible capacitive tactile array pressure sensor system to reduce the crosstalk effect of parasitic capacitance. As shown in Figure 1(a), this method removes capacitance crosstalk by grounding the capacitors in the crossbar structure flexible capacitive sensing array that are not in the same row or column as the capacitor under test. By modifying the capacitance measurement circuit, this method removes about 50% of the capacitance crosstalk effect in the flexible capacitive sensing array. However, this method is complex to operate in practical applications of flexible capacitive sensing arrays and cannot meet the real-time measurement requirements of flexible capacitive sensing arrays. Furthermore, the capacitance removal effect is poor, failing to meet the measurement accuracy requirements.

[0005] Huang et al. constructed a simplified model to characterize the impact of crosstalk and proposed an iterative method to quickly solve for approximate values. As shown in Figure 1(b), this method establishes a relationship model between the original capacitance value of the flexible capacitive sensing array and the capacitance value to be crosstalked, modeling the measured value as the target capacitance value and the parallel connection of three series capacitors, where c r and c c c represents the sum of the capacitance values ​​in the same row and column as the target capacitance, respectively. o This is the sum of the remaining capacitances. The original value is obtained by solving a nonlinear equation. However, due to the complexity of the crosstalk model in flexible capacitive sensing arrays, this method can only solve crosstalk problems in smaller flexible capacitive sensing arrays and cannot be applied to larger flexible capacitive sensing arrays.

[0006] Traditional solutions only address one aspect—fitting the solution or improving the array structure—and these solutions gradually become ineffective as the problem size increases. Therefore, how to remove electrical crosstalk in flexible capacitive sensing arrays is an urgent problem to be solved. Summary of the Invention

[0007] To address the aforementioned problems in the prior art, this invention provides a network and its training method for removing electrical crosstalk in flexible capacitive sensing arrays. Specifically, it includes:

[0008] In a first aspect, the present invention provides a network for removing electrical crosstalk in a flexible capacitive sensing array, comprising:

[0009] The initial feature extraction and dimension transformation convolutional layer, the backbone network, and the output convolutional layer are connected sequentially.

[0010] The backbone network consists of stacked d layers of MACB modules, where d is a positive integer greater than or equal to 1. The number of MACB modules, d, is determined based on the size of the flexible capacitive sensing array from which electrical crosstalk is to be removed.

[0011] The MACB module includes, in sequence, a crosstalk information extraction unit, a channel connection unit, and a cross-channel feature fusion and dimensionality reduction unit;

[0012] The crosstalk information extraction unit consists of four asymmetric convolution kernels of different sizes connected in parallel;

[0013] The preliminary feature extraction and dimension transformation convolutional layer is used to perform preliminary feature extraction and dimension transformation on the capacitance value matrix corresponding to the flexible capacitive sensing array to be free of electrical crosstalk, so as to obtain the first capacitance value matrix.

[0014] The backbone network is used to eliminate crosstalk information in the first capacitance value matrix to obtain the second capacitance value matrix.

[0015] The output convolutional layer is used to generate the capacitance values ​​of each capacitor in the flexible capacitive sensing array to which electrical crosstalk needs to be removed, based on the second capacitance value matrix.

[0016] Crosstalk information extraction unit, used to extract row crosstalk information and column crosstalk information of the input flexible capacitive sensing array;

[0017] The channel connection unit is used to connect the outputs of different convolutional kernels in the corresponding crosstalk information extraction unit according to the channel dimension, and output the connection result.

[0018] The cross-channel feature fusion and dimensionality reduction unit is used to perform cross-channel feature fusion and dimensionality reduction processing on the connection results of the corresponding channel connection units.

[0019] Secondly, the present invention also provides a method for training a network to remove electrical crosstalk in a flexible capacitive sensing array, comprising:

[0020] For any size of flexible capacitive sensing array, target matrix data sets corresponding to flexible capacitive sensing arrays of multiple sizes are obtained through simulation. The multiple target matrix data sets are divided into training set and test set according to preset rules. The target matrix data set includes the original capacitance value matrix and the capacitance value matrix with capacitance crosstalk corresponding to each capacitor in the corresponding flexible capacitive sensing array.

[0021] Based on the size of the flexible capacitive sensing array, a network for removing electrical crosstalk in the flexible capacitive sensing array, as provided in the first aspect, is constructed. Based on a preset loss function, the constructed network is trained and tested using a training set and a test set to obtain a trained network for removing electrical crosstalk in the flexible capacitive sensing array.

[0022] Thirdly, the present invention also provides a method for removing electrical crosstalk in a flexible capacitive sensing array, comprising:

[0023] Acquire the measured values ​​of each capacitor in the target flexible capacitive sensor array from which electrical crosstalk needs to be removed;

[0024] The measured values ​​of each capacitor are input into a network for removing electrical crosstalk in the flexible capacitive sensing array, which is trained according to the method provided in the second aspect and corresponds to the size of the target flexible capacitive sensing array, to recover the true values ​​of each capacitor.

[0025] Fourthly, the present invention also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0026] Memory, used to store computer programs;

[0027] A processor, when executing a program stored in memory, implements either the second or third aspect.

[0028] The beneficial effects of this invention are:

[0029] This invention provides a network and its training method for removing electrical crosstalk in flexible capacitive sensing arrays. The network comprises a preliminary feature extraction and dimensionality transformation convolutional layer, a backbone network, and an output convolutional layer connected in sequence. The backbone network consists of stacked d-layer MACB modules, where d is a positive integer greater than or equal to 1. The number of d layers in the MACB modules is determined according to the size of the flexible capacitive sensing array to be cleaned of electrical crosstalk, thus fully utilizing the size of the flexible capacitive sensing array. This provides good adaptability to flexible capacitive sensing arrays of different sizes, is unaffected by the capacitor size, and is suitable for flexible capacitive sensing arrays of various scales. Furthermore, each MACB module sequentially includes a crosstalk information extraction unit, a channel connection unit, and a cross-channel feature fusion and dimensionality reduction unit. The crosstalk information extraction unit includes four parallel asymmetric convolutional kernels of different sizes. Since asymmetric convolutional modules are more efficient than traditional convolutional kernels, the MACB module can effectively remove electrical crosstalk from flexible capacitive sensing arrays of various sizes. Traditional convolutional modules have fewer parameters, so the network constructed from them to remove electrical crosstalk in flexible capacitive sensing arrays has a much lower number of model parameters and computational cost than traditional neural network models. It is well-suited for edge devices with limited capacity and computing power and can be easily deployed to meet the real-time acquisition and processing needs of sensing systems. In addition, this network makes full use of the structure of the flexible capacitive sensing array, integrating asymmetric convolutional blocks of different sizes in the MACB to characterize the impact of distance information on crosstalk. At the same time, it preserves the positional information of the capacitive elements in the array by keeping the feature scale in the network unchanged. Compared with the limitations of traditional solutions, the network to remove electrical crosstalk in flexible capacitive sensing arrays utilizes both the powerful fitting ability of deep neural networks and the spatial structural features of flexible capacitive sensing arrays, achieving better performance.

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0031] Figure 1(a) is a schematic diagram of the grounding method of a flexible capacitive tactile array pressure sensor system provided by the present invention;

[0032] Figure 1(b) is a schematic diagram of the relationship model between the original capacitance value and the crosstalk capacitance value of a flexible capacitive sensing array provided by the present invention.

[0033] Figure 2(a) is a schematic diagram of the architecture of a flexible capacitive sensing array with an n×n crossbar structure provided by the present invention;

[0034] Figure 2(b) is an equivalent circuit diagram of a flexible capacitive sensing array with a 3×3 crossbar structure provided by the present invention.

[0035] Figure 3 This is a schematic diagram of the architecture of a network for removing electrical crosstalk in a flexible capacitive sensing array, provided by the present invention.

[0036] Figure 4 This is a flowchart illustrating a training method for a network to remove electrical crosstalk in a flexible capacitive sensing array, as provided by the present invention.

[0037] Figure 5 This is a flowchart illustrating a method for removing electrical crosstalk in a flexible capacitive sensing array provided by the present invention.

[0038] Figure 6 A schematic diagram of experimental results provided for this invention;

[0039] Figure 7 This is a schematic diagram of another experimental result provided by the present invention;

[0040] Figure 8 This is a schematic diagram of another experimental result provided by the present invention. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0042] A flexible capacitive sensing array with an n×n crossbar structure is shown in Figure 2(a). Assume c pq (Not shown in the diagram) represents the target capacitance. The interconnecting line connecting its upper and lower electrodes is selected for connection to the measurement circuit; all other interconnecting lines are considered open circuits. It can be seen that c pqThe measured value is related to the capacitance at all other locations in the array, and the series and parallel relationships with other capacitors are not linear and are not easily derived. This manifests in the measurement results as the portion beyond the target capacitance value being the source of crosstalk. Taking a 3×3 capacitor array as an example, assuming c... 11 Figure 2(b) shows the equivalent circuit for the measurement target. It can be seen that even for a small-scale 3×3 flexible capacitive sensor array, the equivalent relationship between the measured value and the true value is quite complex. For larger-sized flexible capacitive sensor arrays, the crosstalk problem becomes even more complex. Further simulation can be used to preliminarily understand the impact of crosstalk. The capacitance in flexible capacitive sensor arrays of different sizes is set to the same value, and then the "measured value" is obtained through circuit simulation. The ratio of the measured value to the true value is given in Table 1. It can be seen that the effect of crosstalk is very significant, especially for larger flexible capacitive sensor arrays, such as a 32×32 flexible capacitive sensor array, where the ratio exceeds 16.

[0043] Table 1. Ratio of measured values ​​to true values

[0044] Measured value / Actual value 1.80 2.28 3.27 4.27 8.26 16.25

[0045] Assuming CA = [caij] is the true value matrix representation of an n×n flexible capacitive sensing array, and CT = [ctij] is the measured value matrix, then the problem of capacitive crosstalk removal can be defined as finding a function f that describes the relationship between CA and CT, i.e., CA = f(CT). Since CT is determined by CA, finding f is actually an inverse problem. An intuitive approach is to first derive the function g representing the relationship between CT and CA, i.e., CT = g(CT), and then derive its inverse function f = g -1 Based on the law of conservation of charge and the relationship between capacitance, charge, and voltage, g can be derived. Through this derivation, the inventors obtained that the expression gij for each ctij in a 4×4 flexible capacitive sensing array is in fractional form, where the numerator is the sum of 2304 products of 7 capacitors, and the denominator is the sum of 1792 products of 6 capacitors. Deriving the analytical inverse function f from this expression is nearly impossible. For larger-scale flexible capacitive sensing arrays, g becomes even more complex, and f becomes even more irreducible.

[0046] Based on the above analysis, it can be seen that the expression g is difficult to derive, and its inverse function f is even more complex. Considering the powerful representational capabilities of neural networks, and the fact that accurate CA and CT pairwise data can be easily obtained through circuit simulation for network training, this invention uses a neural network model to restore the capacitance value. For an n×n flexible capacitive sensing array, assuming c pq The target capacitance is excluding c. pq In addition, the capacitors in the array can be divided into three categories: capacitors in the same row as the target, cpj (j≠q), the capacitor c in the same column as the target iq (i≠p), capacitor c that is neither in the same row nor the same column as the target. ij (i≠p,j≠q). During the measurement process, connect c. pq The row and column lines of the electrodes are connected to the drive and sensing terminals of the measurement circuit, respectively. Therefore, c pj The upper electrode is connected to the driving terminal and can be considered as an input unit; similarly, c iq The lower electrode is connected to the sensing terminal and can be considered as the output unit; other capacitors c ij The two electrodes are connected to the input and output units, respectively. From a circuit topology perspective, capacitors of the same category are structurally equivalent. In fact, it can be analyzed that swapping two rows or two columns of the array will not affect the measurement results. Correspondingly, the sources of crosstalk can be divided into three types: row crosstalk, column crosstalk, and residual crosstalk. Corresponding to the types of crosstalk, this invention proposes a multi-scale asymmetric convolution block (MACB), which is structurally consistent with the capacitance crosstalk in a crossbar structure flexible capacitive sensing array, and structurally simulates the capacitance crosstalk in the array; at the same time, it makes full use of the structure of the flexible capacitive sensing array by setting 1×w in the MACB module. c The convolution kernel extracts line crosstalk information and h c A 1×1 convolutional kernel is used to extract column crosstalk information, thus integrating asymmetric convolutional blocks of different sizes into the MACB to characterize the impact of distance information on crosstalk. Theoretically, symmetric convolutional kernels could be used to extract information from the remaining crosstalk. However, this invention omits this part, partly to maintain the network's lightweight design, and partly because multi-layered stacked row and column convolutions can achieve similar effects to symmetric convolutions. Furthermore, since calculating 1×w... c The data required for convolution is stored in a contiguous memory region. This asymmetric convolution can improve data transmission efficiency and accelerate computation. Furthermore, based on the size of the flexible capacitive sensing array, this invention constructs a lightweight neural network model based on MACB, namely a network for removing electrical crosstalk in the flexible capacitive sensing array (CAR-Net). Based on this network, capacitance crosstalk is removed from the measured values ​​to restore the capacitance value.

[0047] Figure 3 This is a schematic diagram of the architecture of a network for removing electrical crosstalk in a flexible capacitive sensing array provided by the present invention, as shown below. Figure 3 As shown, the network includes:

[0048] The network consists of a preliminary feature extraction and dimensionality transformation convolutional layer (conv_s0), a backbone network, and an output convolutional layer (conv_s0), connected sequentially. The backbone network is composed of stacked d-layer MACB modules.

[0049] Where d is a positive integer greater than or equal to 1, and the number of layers d of the MACB module is determined according to the size of the flexible capacitive sensing array to which electrical crosstalk is to be removed.

[0050] The MACB module consists of a crosstalk information extraction unit (conv_z), a channel connection unit (concat), and a cross-channel feature fusion and dimensionality reduction unit (conv_m11). The crosstalk information extraction unit (conv_z) comprises four parallel asymmetric convolutional kernels of different sizes.

[0051] The initial feature extraction and dimensionality transformation convolutional layer `conv_s0` is used to perform initial feature extraction and dimensionality transformation on the capacitance value matrix corresponding to the flexible capacitive sensing array to which electrical crosstalk needs to be removed, obtaining the first capacitance value matrix. The backbone network is used to eliminate crosstalk information in the first capacitance value matrix, obtaining the second capacitance value matrix. The output convolutional layer `conv_s0` is used to recover the capacitance values ​​of each capacitor in the flexible capacitive sensing array to which electrical crosstalk needs to be removed based on the second capacitance value matrix.

[0052] The crosstalk information extraction unit `conv_z` extracts row and column crosstalk information from the input flexible capacitive sensing array. The channel connection unit `concat` connects the outputs of different convolutional kernels in the corresponding crosstalk information extraction unit according to the channel dimension and outputs the connection result. The cross-channel feature fusion and dimensionality reduction unit `conv_m11` performs cross-channel feature fusion and dimensionality reduction processing on the connection results of the corresponding channel connection unit.

[0053] In networks designed to remove electrical crosstalk in flexible capacitive sensing arrays, the receptive field of the eigenvectors corresponding to the original matrix data gradually increases with network depth. Therefore, it exhibits good adaptability to flexible capacitive sensing arrays of different sizes, unaffected by capacitor size, and is suitable for flexible capacitive sensing arrays of various scales. By using convolutional kernels of different sizes in the MACB block, it is possible to continuously extract information about the crosstalk influence from other locations in the flexible capacitive sensing array on the current location as the network depth increases.

[0054] The feature vectors in the network provided by this invention remain constant in size, and only the number of channels of the feature vectors is changed during network inference, thereby preserving the position information of the capacitor elements in the array, further reducing computational complexity and improving the accuracy of the recovery results.

[0055] Optional, see Figure 3The crosstalk information extraction unit conv_z includes parallel 1×3 convolution kernel conv_m13, 3×1 convolution kernel conv_m31, 1×5 convolution kernel conv_m15 and 5×1 convolution kernel conv_m51.

[0056] The number of MACB modules in the network is determined by the size of the capacitance matrix, so that the extracted feature size remains consistent with the size of the flexible capacitive sensing array throughout the inference process. This simulates the effect of capacitance crosstalk in a real flexible capacitive sensing array to the greatest extent, thereby improving the accuracy of capacitance recovery results.

[0057] Optionally, the formula for determining the layer d of the MACB module in the backbone network is:

[0058]

[0059] Where d represents the number of layers in the MACB module of the backbone network, and n represents the size of the flexible capacitive sensing array. This indicates rounding up to the nearest integer.

[0060] For example, for flexible capacitive sensing arrays of sizes 8×8, 16×16 and 32×32, it is necessary to construct networks (CAR-Net) containing 4-layer, 8-layer and 16-layer MACB blocks to remove electrical crosstalk in the flexible capacitive sensing array.

[0061] This method ensures that the receptive field of the last MACB layer can cover all locations of the input flexible capacitive sensing array.

[0062] Optional, see Figure 3 The initial feature extraction and dimensionality transformation convolutional layer conv_s0 includes a 3×3 convolutional kernel.

[0063] Optional, see Figure 3 The output convolutional layer conv_s0 consists of 3×3 convolutional kernels.

[0064] Optional, see Figure 3 The cross-channel feature fusion and dimensionality reduction unit conv_m11 includes a 1×1 convolution kernel.

[0065] Furthermore, to facilitate understanding of the present invention, a set of specific examples of network parameters for removing electrical crosstalk in flexible capacitive sensing arrays are provided, as detailed in Table 2.

[0066] Table 2. Network parameters for removing electrical crosstalk in flexible capacitive sensing arrays.

[0067]

[0068] The network for removing electrical crosstalk in flexible capacitive sensing arrays provided by this invention includes a preliminary feature extraction and dimensionality transformation convolutional layer, a backbone network, and an output convolutional layer connected in sequence. The backbone network is composed of stacked d-layer MACB modules, where d is a positive integer greater than or equal to 1. The number of d layers in the MACB modules is determined according to the size of the flexible capacitive sensing array to be cleaned of electrical crosstalk, thus fully taking into account the size of the flexible capacitive sensing array. It has good adaptability to flexible capacitive sensing arrays of different sizes, is not affected by the size of the capacitor, and is suitable for flexible capacitive sensing arrays of various sizes. Furthermore, the MACB module includes a crosstalk information extraction unit, a channel connection unit, and a cross-channel feature fusion and dimensionality reduction unit in sequence. The crosstalk information extraction unit includes four asymmetric convolutional kernels of different sizes connected in parallel. Since asymmetric convolutional modules are more efficient than traditional convolutional kernels, they are more efficient in removing electrical crosstalk. The integrator module has a smaller number of parameters, so the network constructed from it to remove electrical crosstalk in flexible capacitive sensing arrays has a much smaller model data volume and throughput than traditional neural network models. It is well adapted to edge devices with limited capacity and computing power and can be easily deployed to edge devices to meet the real-time acquisition and processing needs of sensing systems. In addition, the network makes full use of the structure of the flexible capacitive sensing array, integrating asymmetric convolutional blocks of different sizes in the MACB to characterize the impact of distance information on crosstalk. At the same time, it preserves the position information of the capacitive elements in the array by keeping the feature scale in the network unchanged. Compared with the limitations of traditional solutions, the network to remove electrical crosstalk in flexible capacitive sensing arrays utilizes the powerful fitting ability of deep neural networks and the spatial structural features of flexible capacitive sensing arrays, thus achieving better performance.

[0069] This invention also provides a method for training a network to remove electrical crosstalk in a flexible capacitive sensing array. Figure 4 This is a flowchart illustrating a training method for a network to remove electrical crosstalk in a flexible capacitive sensing array, as provided by the present invention. Figure 4 As shown, the method includes:

[0070] S401. For a flexible capacitive sensing array of any size, obtain target matrix data sets corresponding to flexible capacitive sensing arrays of multiple sizes through simulation, and divide the multiple target matrix data sets into training sets and test sets according to preset rules.

[0071] The target matrix data set includes the original capacitance value matrix and the capacitance value matrix with capacitance crosstalk for each capacitor in the corresponding flexible capacitive sensing array.

[0072] Specifically, for flexible capacitive sensing arrays of different sizes, SPICE simulation software was used to simulate the capacitor crosstalk in the flexible capacitive sensing array, generating multiple sets of original capacitance value matrices and capacitance value matrices with capacitor crosstalk as the training set and test set of the network.

[0073] There are no specific restrictions on the rules for dividing the training and test sets.

[0074] S402. Based on the size of the flexible capacitive sensing array, construct any of the networks provided by this invention for removing electrical crosstalk in the flexible capacitive sensing array, and train and test the constructed network according to the training set and test set based on the preset loss function to obtain the trained network for removing electrical crosstalk in the flexible capacitive sensing array.

[0075] Optionally, the preset expression for the loss function is:

[0076]

[0077] Where Loss represents the loss function. Indicates ca ij The corresponding recovered capacitance value, ca ij This represents the capacitance value in the i-th row and j-th column of the capacitance value matrix.

[0078] For example, for flexible capacitive sensing arrays of different sizes, networks of corresponding depths for removing electrical crosstalk in the flexible capacitive sensing arrays are constructed and trained using training datasets.

[0079] The present invention also provides a method for removing electrical crosstalk in a flexible capacitive sensing array. Figure 6 This is a flowchart illustrating a method for removing electrical crosstalk in a flexible capacitive sensing array provided by the present invention, as shown below. Figure 5 As shown, the method includes:

[0080] S501. Obtain the measured values ​​of each capacitor in the target flexible capacitive sensor array from which electrical crosstalk needs to be removed.

[0081] S502. Input the measured values ​​of each capacitor into the network for removing electrical crosstalk in the flexible capacitive sensing array, which is trained according to the method provided by the present invention and corresponds to the size of the target flexible capacitive sensing array, to recover the true values ​​of each capacitor.

[0082] To further demonstrate the beneficial effects of the present invention, a set of experimental data is also provided, as follows:

[0083] 1. Capacitor crosstalk recovery experiment

[0084] In addition to the network proposed in this invention, MLP, VGG, ResNet, GoogLeNet, and DnCNN were also trained for comparison. DnCNN was used for image denoising, a problem somewhat similar to capacitance restoration. The restoration accuracy results for flexible capacitive sensing arrays on the test set are shown in Table 3. It can be seen that for 8×8, 16×16, and 32×32 flexible capacitive sensing arrays, CAR-Net achieved capacitance restoration accuracies of 98.17%, 97.7%, and 95.74%, respectively, maintaining lower computational and parameter counts while significantly higher than other models. Although DnCNN also showed relatively high restoration accuracy, its computational and parameter counts far exceeded those of CAR-Net. For larger capacitor arrays, MLP even failed to converge. The results indicate that the proposed MACB, due to its structural correspondence with the topological relationship of the capacitors in the array, enables CAR-Net to effectively eliminate capacitance crosstalk.

[0085] Table 3 shows the restoration accuracy results of the flexible capacitive sensing array on the test set.

[0086]

[0087] In practical applications, model inference time is a key metric, especially for real-time or near-real-time applications. This invention fully considers the deployment requirements of the model in the design of CAR-Net. To verify the inference time, the model was deployed on a terminal platform using an RTX 3080Ti GPU and a Huawei Atlas 200 Development Kit (DK). The 200DK includes an Ascend 310AI processor with performance of 16 TOPS@INT8 and 8 TOPS@FP16. As shown in Table 4, the inference time of the CAR-Net model suitable for a 32×32 flexible capacitive sensor array on the RTX 3080Ti and Atlas 200DK are 0.0047s and 0.0031s, respectively. The results show that CAR-Net can easily meet the real-time requirements of deployment on edge or terminal devices.

[0088] Table 4. CAR-Net Inference Time (s)

[0089] RTX 3080 Ti 0.0012 0.0019 0.0047 Atlas 200DK* - - 0.0031

[0090] 2. Comparison of Pattern Recognition Task Results

[0091] In practical sensing systems, the acquired data is typically used for subsequent processing tasks, and removing capacitance crosstalk helps improve the performance of these tasks. To verify this, further classification experiments were conducted on the Fashion-MNIST dataset, which contains grayscale images of 10 categories. The training and test sets contain 60k and 10k sample images with a resolution of 28×28, respectively. For the images in Fashion-MNIST, the grayscale value im was first scaled to the range of 1-2 using the formula im / 255+1. The transformed value was then mapped to a 32×32 capacitor array as its true value CA. Circuit simulation was performed to obtain CT. Then, CAR-Net was used to remove capacitance crosstalk from CT to obtain the restored capacitance value CA. Using CT and CA as inputs, three different classification networks—MLP, LeNet, and ResNet18—were trained, respectively. The classification results of each model on the test set are shown below. Figure 6 As shown, the x and y axes represent the number of training generations and classification accuracy, respectively. Overall, using CAR-Net for capacitance restoration followed by classification results in higher accuracy and faster model convergence, especially for simpler models like MLP and LeNet, where the performance improvement is more significant. This indicates that in applications where precise capacitance values ​​are not required, removing capacitance crosstalk can also improve the performance of subsequent tasks or reduce the complexity of processing methods.

[0092] 3. Actual test of capacitor restoration

[0093] The accuracy of the proposed CAR-Net capacitance restoration was verified using measured data. Considering that flexible capacitive sensing arrays are typically fabricated using integrated processes, making it impossible to measure the value of individual capacitors separately, testing was conducted using flexible printed circuits (FPCs). Figure 7 As shown, FPC can construct an 8×8 flexible capacitive sensor array by soldering capacitors. Capacitors with nominal values ​​of 2.2nF and 1.5nF, with a tolerance of 10%, are used to simulate pressure and no pressure at corresponding locations. Capacitor tolerance more realistically reflects the impact of manufacturing process variations in the capacitive sensor array. The actual value of the capacitors is measured using an LCR meter, and a capacitor with a nominal value of 2.2nF is selected for soldering. Figure 7 The numbers 0-8 are represented in appropriate positions within the boxes of the FPC shown, while capacitors with a nominal value of 1.5nF are soldered in other positions. Then, actual measurements are taken on the array to obtain capacitance values ​​with crosstalk. Next, the capacitance values ​​are restored using CAR-Net trained on simulation data. As a typical example, Figure 7 This paper presents the measured and restored capacitance values ​​of the FPC array when the stress mode is digital 0. It can be seen that the stress mode can be intuitively determined from the restoration results, and the overall accuracy of the capacitance restoration is also good. According to statistics, such as... Figure 8The average restoration accuracy for all nine digit patterns shown is 97.3%, only slightly lower than the simulation results given in Table 3. The main reason is that the data used to train CAR-Net was completely randomized and not fine-tuned for the measured digit patterns. This, from another perspective, demonstrates the versatility and effectiveness of CAR-Net in practical applications.

[0094] This invention also provides an electronic device structure, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, communication interface, and memory communicate with each other via the communication bus.

[0095] Memory, used to store computer programs;

[0096] When a processor executes a program stored in memory, it implements the steps provided in the above method embodiments.

[0097] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0098] For example, the electronic device is a wearable electronic device such as a wristband or watch, or it can be a sensor on a wearable device.

[0099] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0100] For the device / electronic device embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For specific details and beneficial effects, please refer to the description of the method embodiments.

[0101] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0102] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for removing electrical crosstalk in a flexible capacitive sensing array, characterized in that, include: Acquire the measured values ​​of each capacitor in the target flexible capacitive sensor array from which electrical crosstalk needs to be removed; The measured values ​​of each capacitor are input into a network that is trained and corresponds to the size of the target flexible capacitive sensing array to remove electrical crosstalk in the flexible capacitive sensing array, so as to recover the true value of each capacitor. The network for removing electrical crosstalk in the flexible capacitive sensing array includes: The initial feature extraction and dimension transformation convolutional layer, the backbone network, and the output convolutional layer are connected sequentially. The backbone network is composed of stacked d layers of MACB modules, where d is a positive integer greater than or equal to 1. The number of layers d of the MACB modules is determined according to the size of the flexible capacitive sensing array to which electrical crosstalk is to be removed. The MACB module sequentially includes a crosstalk information extraction unit, a channel connection unit, and a cross-channel feature fusion and dimensionality reduction unit. The crosstalk information extraction unit includes four asymmetric convolution kernels of different sizes connected in parallel; The preliminary feature extraction and dimension transformation convolutional layer is used to perform preliminary feature extraction and dimension transformation on the capacitance value matrix corresponding to the flexible capacitive sensing array to which electrical crosstalk is to be removed, so as to obtain the first capacitance value matrix. The backbone network is used to eliminate crosstalk information in the first capacitance value matrix to obtain the second capacitance value matrix. The output convolutional layer is used to generate the capacitance values ​​of each capacitor in the flexible capacitive sensing array to which electrical crosstalk needs to be removed, based on the second capacitance value matrix. The crosstalk information extraction unit is used to extract the row crosstalk information and column crosstalk information of the input flexible capacitive sensing array. The channel connection unit is used to connect the outputs of different convolution kernels in the corresponding crosstalk information extraction unit according to the channel dimension, and output the connection result. The cross-channel feature fusion and dimensionality reduction unit is used to perform cross-channel feature fusion and dimensionality reduction processing on the connection results of the corresponding channel connection units.

2. The method for removing electrical crosstalk in a flexible capacitive sensing array according to claim 1, characterized in that, The crosstalk information extraction unit includes parallel 1×3 convolution kernels, 3×1 convolution kernels, 1×5 convolution kernels, and 5×1 convolution kernels.

3. The method for removing electrical crosstalk in a flexible capacitive sensing array according to claim 1, characterized in that, The formula for determining the layer d of the MACB module in the backbone network is: ， Where d represents the number of layers in the MACB module of the backbone network, and n represents the size of the flexible capacitive sensing array. This indicates rounding up to the nearest integer.

4. The method for removing electrical crosstalk in a flexible capacitive sensing array according to any one of claims 1-3, characterized in that, The preliminary feature extraction and dimension transformation convolutional layer includes a 3×3 convolutional kernel.

5. The method for removing electrical crosstalk in a flexible capacitive sensing array according to any one of claims 1-3, characterized in that, The output convolutional layer includes a 3×3 convolutional kernel.

6. The method for removing electrical crosstalk in a flexible capacitive sensing array according to any one of claims 1-3, characterized in that, The cross-channel feature fusion and dimensionality reduction unit includes a 1×1 convolution kernel.

7. A training method for a network to remove electrical crosstalk in a flexible capacitive sensing array, characterized in that, include: For a flexible capacitive sensing array of any size, target matrix data sets corresponding to multiple flexible capacitive sensing arrays of the stated size are obtained through simulation, and the multiple target matrix data sets are divided into training set and test set according to preset rules. The target matrix data set includes the original capacitance value matrix and the capacitance value matrix with capacitance crosstalk corresponding to each capacitor in the corresponding flexible capacitive sensing array. Based on the size of the flexible capacitive sensing array, a network for removing electrical crosstalk in the flexible capacitive sensing array as described in any one of claims 1-6 is constructed. The constructed network is trained and tested based on a preset loss function, according to the training set and the test set, to obtain a trained network for removing electrical crosstalk in the flexible capacitive sensing array.

8. The training method according to claim 7, characterized in that, The expression for the preset loss function is: ， express The corresponding recovered capacitance value, Represents the first value in the capacitance matrix. Line 1 The capacitance value on the column.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the training method according to any one of claims 7-8.

Images