Pressure mapping method and device for high-precision pressure sensor array

Abstract

The invention relates to the technical field of pressure sensing, and discloses a pressure mapping method and device for a high-precision pressure sensor array. The method comprises the steps that known standard pressure is applied to a high-precision pressure sensor array, and an original pressure distribution matrix of each sensing unit is collected; executing bicubic spline interpolation operation of tension coefficient compensation to obtain a continuous pressure distribution matrix; inputting a depth separable convolutional network to perform spatial feature extraction and channel feature fusion to obtain a pressure feature description matrix; carrying out confidence coefficient propagation iterative calculation to obtain a pressure confidence coefficient matrix; carrying out reliability grouping and weight coefficient calculation on the sensing units to obtain a pressure data fusion weight matrix; according to the method, the pressure mapping precision is improved, relatively high real-time performance is ensured, and the problems of data distortion, low calculation efficiency and insufficient reliability in a traditional method are solved.

Description

Technical Field

[0001] The present invention relates to the field of pressure sensing technology, and in particular to a pressure mapping method and device for a high-precision pressure sensor array. Background Art

[0002] With the rapid development of intelligent sensing technology and human-computer interaction systems, flexible high-precision pressure sensor arrays have been widely used in medical rehabilitation, industrial automation, human-computer interaction and other fields due to their excellent mechanical properties and spatial resolution. In practical applications, traditional pressure sensor arrays are often difficult to accurately reflect the continuous pressure distribution due to the nonlinear response characteristics of the sensor unit and the discreteness of the spatial distribution, which affects the measurement accuracy and reliability of the system.

[0003] Existing pressure mapping methods mainly rely on simple numerical interpolation and linear fitting algorithms, which cannot effectively deal with problems such as nonlinear response, noise interference and zero drift of sensors. At the same time, traditional methods have high computational complexity and poor real-time performance when processing large-scale sensor array data, making it difficult to meet the needs of practical applications. In addition, due to the lack of an effective data fusion mechanism, existing methods are difficult to fully utilize the spatial correlation information of the sensor array, resulting in insufficient spatial consistency and temporal stability of the results. Current pressure mapping technology still faces several key technical difficulties in practical applications: on the one hand, the calibration process of the sensor array is complex, and the influence of individual differences and environmental factors needs to be considered; on the other hand, the spatial distribution of pressure data is uneven, and there are blind spots between sensor units, requiring a reasonable interpolation algorithm for data reconstruction. At the same time, how to achieve multi-scale feature extraction and reliability evaluation of pressure data while ensuring computational efficiency is also a technical problem that needs to be solved urgently. Summary of the invention

[0004] The present invention provides a pressure mapping method and device for a high-precision pressure sensor array, which improves the pressure mapping accuracy while ensuring high real-time performance, and solves the problems of data distortion, low computational efficiency and insufficient reliability existing in traditional methods.

[0005] In a first aspect, the present invention provides a pressure mapping method for a high-precision pressure sensor array, the pressure mapping method for a high-precision pressure sensor array comprising: Apply a known standard pressure to the high-precision pressure sensor array, collect the voltage output data of each sensor unit, convert it, and obtain the original pressure distribution matrix; Performing a bicubic spline interpolation operation with tension coefficient compensation on pressure data between adjacent sensing units in the original pressure distribution matrix to obtain a continuous pressure distribution matrix; Inputting the continuous pressure distribution matrix into a deep separable convolutional network to extract spatial features and fuse channel features to obtain a pressure feature description matrix; Performing confidence propagation iterative calculation on the pressure feature description matrix through a pressure state probability model to obtain a pressure confidence matrix; Reliability grouping and weight coefficient calculation of the sensor units are performed according to the pressure confidence matrix to obtain a pressure data fusion weight matrix; The pressure data fusion weight matrix is ​​used to calculate the pressure center position and pressure gradient distribution of the pressure feature description matrix, and the target pressure mapping data is output.

[0006] In a second aspect, the present invention provides a pressure mapping device of a high-precision pressure sensor array, the pressure mapping device of the high-precision pressure sensor array comprising: The acquisition module is used to apply a known standard pressure to the high-precision pressure sensor array, collect the voltage output data of each sensor unit, convert it, and obtain the original pressure distribution matrix; An interpolation module, used for performing a bicubic spline interpolation operation of tension coefficient compensation on the pressure data between adjacent sensing units in the original pressure distribution matrix to obtain a continuous pressure distribution matrix; An extraction module, used for inputting the continuous pressure distribution matrix into a deep separable convolutional network to perform spatial feature extraction and channel feature fusion to obtain a pressure feature description matrix; An iterative module, used for performing confidence propagation iterative calculation on the pressure feature description matrix through a pressure state probability model to obtain a pressure confidence matrix; A grouping module, used to perform reliability grouping and weight coefficient calculation on the sensor units according to the pressure confidence matrix to obtain a pressure data fusion weight matrix; The output module is used to calculate the pressure center position and pressure gradient distribution of the pressure feature description matrix using the pressure data fusion weight matrix, and output target pressure mapping data.

[0007] A third aspect of the present invention provides a pressure mapping device for a high-precision pressure sensor array, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the pressure mapping device for the high-precision pressure sensor array executes the above-mentioned pressure mapping method for the high-precision pressure sensor array.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned pressure mapping method of the high-precision pressure sensor array.

[0009] In the technical solution provided by the present invention, by introducing a bicubic spline interpolation algorithm with tension coefficient compensation, the continuous pressure distribution between the sensing units is accurately reconstructed; a deep separable convolutional network is used to realize multi-scale feature extraction and fusion of pressure data, significantly reducing the computational complexity; a data reliability evaluation system is established based on the pressure state probability model and the confidence propagation mechanism; the intelligent fusion of pressure data is realized by adaptively calculating the fusion weights through a deep Q learning network; and a complete solution is constructed by combining pressure center positioning and gradient distribution analysis. While improving the pressure mapping accuracy, the present invention ensures a high degree of real-time performance, solves the problems of data distortion, low computational efficiency and insufficient reliability in traditional methods, and provides reliable technical support for the promotion and application of pressure sensor arrays in various application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 A schematic diagram of a process flow of a pressure mapping method for a high-precision pressure sensor array provided in an embodiment of the present application; Figure 2 A schematic block diagram of the structure of a pressure mapping device of a high-precision pressure sensor array provided in an embodiment of the present application; Figure 3 A schematic block diagram of the structure of a pressure mapping device of a high-precision pressure sensor array provided in an embodiment of the present application. DETAILED DESCRIPTION

[0012] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0013] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change based on actual conditions.

[0014] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0015] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0016] In conjunction with the accompanying drawings, some embodiments of the present application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0017] See also Figure 1 , Figure 1 A schematic diagram of a process flow of a pressure mapping method for a high-precision pressure sensor array provided in an embodiment of the present application, such as Figure 1 As shown, the pressure mapping method of the high-precision pressure sensor array provided in the embodiment of the present application includes steps S100 to S600.

[0018] Step S100, applying a known standard pressure to the high-precision pressure sensor array, collecting the voltage output data of each sensor unit for conversion, and obtaining an original pressure distribution matrix; It is understandable that the execution subject of the present invention may be a pressure mapping device of a high-precision pressure sensor array, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0019] Specifically, the high-precision pressure sensor array is loaded with standard pressure. By loading a known standard pressure loading sequence, it is ensured that each sensing unit can measure under predetermined pressure conditions. The standard pressure loading sequence includes a series of known pressure values, which are applied to different positions of the sensor array in turn. Each time the standard pressure is loaded, it is necessary to ensure the constancy of the pressure, and collect the voltage output data of each sensing unit through a high-speed data acquisition module. In this embodiment, the high-precision pressure sensor array is an intelligent sensing system composed of a large number of micro pressure sensing units arranged in a specific space. Its core feature is that a flexible material is used as a substrate to achieve flexible integration of the sensing unit. Each sensing unit has independent signal acquisition and conversion functions, and can convert local pressure changes into corresponding electrical signal outputs. This array structure not only ensures the overall flexibility of the sensor, but also achieves high-precision acquisition of pressure signals. Its significant advantages are reflected in three aspects: it has excellent mechanical properties and can adapt to various complex application environments; it has high spatial resolution and can accurately capture subtle changes in pressure distribution; and it uses signal processing technology to ensure the accuracy and real-time performance of data acquisition. This design enables the high-precision pressure sensor array to provide stable and reliable pressure measurement results while maintaining sensitivity. Compared with the traditional pressure sensor array, the high-precision pressure sensor array has achieved major breakthroughs in many key aspects. First, in terms of accuracy, the traditional array is limited by nonlinear response characteristics, and the measurement accuracy is generally insufficient, while the high-precision array significantly improves the measurement accuracy by introducing a special calibration mechanism and compensation algorithm. Secondly, in terms of spatial resolution, the high-precision array adopts a denser unit arrangement method, reducing the spacing between sensor units to the order of 0.1 mm, which is far better than the sparse distribution of the traditional array. In terms of signal processing technology, the traditional array mainly relies on simple numerical interpolation and linear fitting algorithms, while the high-precision array introduces advanced algorithms such as tension coefficient compensation and deep learning to achieve more accurate pressure data reconstruction. In addition, the high-precision array also has stronger environmental adaptability and data reliability. Through a perfect calibration mechanism and data fusion strategy, it effectively overcomes the defects of the traditional array that are easily affected by environmental factors. At the same time, it has also achieved a significant improvement in data processing efficiency and can support real-time pressure mapping applications.

[0020] The collected voltage data is affected by many factors, such as the dynamic range and zero drift of the sensor. The voltage output data is calibrated dynamically and compensated for zero drift. Dynamic range calibration is designed to ensure the measurement accuracy of the sensor over the entire pressure range, while zero drift compensation is to solve the errors generated by the sensor during long-term use.

[0021] In order to eliminate the high-frequency noise in the voltage signal, the voltage data is processed by wavelet denoising. Wavelet denoising can effectively remove noise while retaining important information in the signal to obtain the calibrated voltage data of each sensor unit. The calibrated voltage data of each sensor unit is reconstructed according to the spatial layout of the sensor array. The voltage data of each sensor unit is organized according to their actual positions in the array, reflecting the specific response of each sensor unit.

[0022] The reconstructed voltage data is processed by piecewise linear fitting, and the voltage response curve of each sensor unit is fitted by the piecewise linear fitting model. The piecewise linear fitting can accurately capture the relationship between voltage and pressure, and can effectively deal with nonlinear effects under different pressure ranges. Through fitting, the piecewise linear fitting coefficient of each sensor unit is obtained.

[0023] The pressure response equation of each sensor unit is constructed by least squares operation using piecewise linear fitting coefficients and standard pressure loading sequence. The least squares method is used to optimize the fitting curve to minimize the fitting error. The pressure response coefficient of each sensor unit is obtained by least squares calculation, reflecting the sensitivity and response characteristics of the sensor to different pressures. The pressure response coefficient is organized into an M×N dimensional matrix structure according to the spatial distribution of the sensor array. The matrix represents the overall response characteristics of the sensor array, that is, the pressure response coefficient matrix.

[0024] The pressure response coefficient matrix is ​​used to perform piecewise linear calculation and response curve correction on the voltage data collected in real time. By inputting the real-time collected voltage data into the pressure response equation of each sensor unit, the pressure value corresponding to each sensor unit is calculated. The pressure values ​​of the sensor units are spatially reconstructed and boundary compensated. Spatial reconstruction is to recombine the pressure values ​​at different positions into a complete pressure field distribution map according to the physical layout of the sensor array; boundary compensation is to process the pressure data in the edge area of ​​the array. The measured values ​​in these areas have certain errors and need to be corrected to improve the accuracy of the entire pressure distribution matrix. The original pressure distribution matrix of M×N dimensions is obtained, which reflects the pressure distribution of the entire sensor array.

[0025] Step S200, performing a bicubic spline interpolation operation of tension coefficient compensation on the pressure data between adjacent sensing units in the original pressure distribution matrix to obtain a continuous pressure distribution matrix; Specifically, a differential operation is performed on the pressure data between adjacent sensing units in the original pressure distribution matrix to determine the area where the pressure change rate exceeds the set threshold. The pressure change rate between adjacent sensing units is calculated. When the pressure change rate exceeds the set threshold, it indicates that there is a large fluctuation or error in the pressure distribution of these areas, and further compensation processing is required. Through differential operation, the matrix of the area to be compensated is obtained.

[0026] Based on the areas identified in the matrix of areas to be compensated, the local pressure gradient of each area to be compensated is calculated. The pressure gradient reflects the changing trend of the pressure in the area. The local tension coefficient of each area to be compensated is calculated based on the gradient value. The tension coefficient is a parameter used to describe how to smooth the pressure data during the interpolation process. The larger the local tension coefficient, the more drastic the pressure change in the area, and the stronger the smoothing required.

[0027] The local tension coefficients of each region are combined to form a spatial tension coefficient distribution, which describes the tension changes in different regions of the entire pressure field. In order to improve the smoothness of the tension coefficient, the spatial tension coefficient distribution is Gaussian smoothed to obtain a smoothed tension coefficient matrix. This matrix is ​​used for subsequent interpolation calculations to ensure that large errors will not be generated during the interpolation process due to local noise or outliers.

[0028] The original pressure distribution matrix is ​​divided into a regular grid. The grid divides the entire pressure field into several small areas so that local interpolation calculations can be performed on each small area. For each grid node, a bicubic B-spline basis function is set, which can smoothly interpolate the pressure data of the local area. Bicubic B-spline interpolation is an efficient interpolation method that can provide high interpolation accuracy and has strong data smoothing capabilities.

[0029] By multiplying the smooth tension coefficient matrix with the grid interpolation basis function, a cubic spline function with tension terms is obtained. This function can effectively compensate for the errors in the original data and smooth the pressure distribution during the interpolation process. The compensated interpolation basis function is expanded in the x and y directions. The expanded function is further optimized by local polynomial fitting to obtain more accurate coordinates of the interpolation control points.

[0030] By interpolating the coordinates of the control points, bicubic spline interpolation is performed on each point in the grid to generate a fine grid pressure data matrix. The fine grid pressure data matrix is ​​subjected to piecewise continuity constraints and derivative continuity corrections to avoid the final pressure distribution accuracy being affected by discontinuities and sudden changes in derivatives during the interpolation process. The resulting matrix has good continuity and physical rationality, and can accurately reflect the details of the pressure distribution, thereby forming a continuous pressure distribution matrix.

[0031] Step S300, inputting the continuous pressure distribution matrix into a deep separable convolutional network to extract spatial features and fuse channel features to obtain a pressure feature description matrix; Specifically, the continuous pressure distribution matrix is ​​numerically normalized to map pressure data of different magnitudes to a unified numerical range, avoiding unnecessary deviations in the deep learning model caused by data of different dimensions or scales. The normalized data is reconstructed into an input tensor suitable for input into the deep separable convolutional network. The continuous pressure data is organized in a certain format so that the deep learning model can effectively process the data.

[0032] The network input data is fed into the first depth-wise separable convolutional layer. A 3×3 convolution kernel is used to extract spatial features from each input channel. The convolution operation captures the local spatial structure and pattern features in the data. In order to improve the nonlinear expression ability of the model, the ReLU activation function is used to perform nonlinear transformation on the convolution result. At the same time, batch normalization is applied to the convolution output to ensure stability and accelerate convergence during network training. After processing, a spatial feature map is obtained, which contains spatial information of different regions in the data.

[0033] Perform 1×1 pointwise convolution on the spatial feature map. Process different channels of each pixel and fuse the feature information of all channels. Through 1×1 convolution, the channel information is mixed and compressed while maintaining the spatial structure, so that the network can more effectively extract features at different levels. Use the BatchNorm layer to normalize the data distribution of the fused features, optimize the expression of the data, and generate the channel features of the first layer.

[0034] The channel features of the first layer are input into the second depth-wise separable convolution layer. Dilated convolution is used for processing, where the dilation rate is set to 2 and the convolution kernel size is 3×3. The role of dilated convolution is to expand the receptive field without increasing the amount of calculation, and to extract multi-scale spatial features in a larger range. Through dilated convolution, the network can better capture spatial information at different scales, especially when processing pressure data with complex structures. Dilated convolution helps improve the perception ability of the model. After ReLU activation and batch normalization, multi-scale spatial features are obtained, which effectively extract different changing trends of pressure data at different scales.

[0035] The multi-scale spatial features are subjected to 1×1 convolution operations to perform channel dimension reduction processing, reduce feature redundancy, and retain the most critical feature information. The network adds a skip connection structure to fuse the reduced-dimensional features with the input features. Skip connections help prevent the loss of important information in deep networks and maintain data coherence and information flow. The fused features are processed through the BatchNorm layer to obtain the fused features of the second layer.

[0036] The fused features of the second layer are input into the third depth-wise separable convolution layer, and feature decomposition is performed using grouped convolution. Grouped convolution divides the features into several groups, and each group uses a 3×3 convolution kernel to extract local features. Grouped convolution can reduce the amount of calculation and improve the representation ability of the network. Through grouping processing, local features in the data can be analyzed more carefully. After ReLU activation and batch normalization, a grouped feature map is obtained, which contains a fine-grained description of different areas of the pressure data.

[0037] Channel rearrangement and 1×1 convolution operation are performed on the grouped feature maps. The purpose of channel rearrangement is to reorganize the information of each channel so that the features across groups can communicate with each other and enhance the complementarity between features. The 1×1 convolution operation realizes the fusion between channels, and the processing of the BatchNorm layer ensures the balanced distribution of features and the stability of network training. After processing, the feature representation of the third layer is obtained.

[0038] The feature representation of the third layer is processed through the global average pooling operation. The features obtained after pooling contain the global information of the entire pressure data. Global average pooling compresses the feature map by calculating the average value of each channel to generate a smaller but more representative feature vector. Through feature reconstruction, the compressed features are reorganized to generate a pressure feature description matrix containing spatial features and channel features.

[0039] Step S400, performing confidence propagation iterative calculation on the pressure feature description matrix through the pressure state probability model to obtain a pressure confidence matrix; Specifically, a Gaussian mixture distribution model is constructed for each spatial position in the pressure feature description matrix, and the pressure value is mapped to the probability space. Converting the original pressure data into a probabilistic form can better model uncertainty in subsequent processing. Through the Gaussian mixture distribution model, multiple possible values ​​are provided for the pressure value of each spatial position, and each possible value is weighted according to its probability to obtain the probability distribution of the initial state.

[0040] The factor graph structure is constructed based on the initial state probability distribution. The factor graph is a graph model that contains nodes and edges. The nodes represent variables, while the edges represent the dependencies between variables. In this model, each node represents the pressure state of a position in the pressure sensor array, and also includes the observation value. In order to describe the relationship between these variables, there are multiple factor nodes in the factor graph, which connect the variable nodes and the observation nodes to form a complete network structure. In this way, the factor graph can effectively represent the interdependence between different pressure points in space.

[0041] After constructing the pressure state factor graph, state variables and observation variables are assigned to each node to fully describe the state and observation information of each pressure sensing unit. The message transmission weights are initialized for the variable nodes in the pressure state factor graph, and the amount of information transmission between nodes is calculated. The initial message transmission weight determines how information is transmitted between nodes. The amount of information transmission between nodes is determined by the state probability distribution of each node. After calculation, the initial message transmission matrix is ​​obtained, which reflects the information flow between nodes in the initial state.

[0042] The initial message passing matrix is ​​input into the belief propagation network for processing. The belief propagation network calculates the marginal probability distribution of each node through forward propagation. These marginal probability distributions reflect the state information of each node and the uncertainty of the node state. In this process, the confidence value of the node is obtained by maximum a posteriori probability calculation. The maximum a posteriori probability calculation can select the most likely state from all possible states based on the initial state and observed data, combined with Bayesian reasoning theory. In this way, the optimal state estimate of each node is obtained. At the same time, iterative optimization is performed in combination with spatial correlation constraints to eliminate noise in the data and ensure that the pressure state between adjacent nodes in space remains consistent during the propagation process, so that the model can better capture the spatial distribution characteristics and improve the prediction accuracy.

[0043] Calculate the mutual information between adjacent nodes based on the updated node status. Mutual information is a measure of the dependency between two random variables, used to describe the degree of information sharing between adjacent nodes. By calculating the mutual information between adjacent nodes, we can understand how their state changes affect each other, so that we can better adjust the weight of message passing. Based on the calculation results of the mutual information, update the message passing weight to ensure that the relationship between adjacent nodes is more accurately reflected during the information propagation process, and obtain the optimized message weight.

[0044] The node states are weighted averaged using the optimized message weights to build a confidence propagation path between nodes. The state information of different nodes is effectively fused by weighted average to obtain a global confidence distribution. The global confidence distribution represents the overall confidence of the pressure state of the entire pressure sensor array, and provides a corresponding confidence value for each spatial point in the pressure field. This distribution can reflect the reliability and accuracy of pressure data at different locations.

[0045] The global confidence distribution is reconstructed into a two-dimensional matrix that displays the confidence value at each location. The values ​​in the confidence matrix are normalized and compressed into a uniform numerical range, so that the matrix can more intuitively display the pressure confidence of each spatial point.

[0046] Step S500: Reliability grouping and weight coefficient calculation of the sensor units are performed according to the pressure confidence matrix to obtain a pressure data fusion weight matrix; Specifically, the pressure confidence matrix is ​​processed, and by setting two thresholds of 0.8 and 0.6, the sensing units are divided into reliable areas, transition areas, and unreliable areas according to their confidence values. The credibility of the data is determined according to the confidence values ​​of different areas, so that different processing weights are given to different areas in the subsequent data fusion process. Areas with confidence values ​​higher than 0.8 are classified as reliable areas, indicating that the credibility of the pressure data in these areas is high; areas with confidence values ​​lower than 0.6 are classified as unreliable areas, and the pressure data in these areas are affected by noise or errors and have low credibility; areas between 0.6 and 0.8 are classified as transition areas, indicating that the credibility of the pressure data in these areas is at a medium level. Through this step, reliability grouping labels are obtained to mark the reliability of different areas.

[0047] Based on the reliability grouping labels, the sensing units are divided into three independent subsets, corresponding to the reliable area, transition area and unreliable area. For each subset, the local statistical characteristics and spatial distribution characteristics are calculated respectively. These characteristics reflect the distribution and change law of pressure data in each area. Through feature calculation, the grouping feature matrix is ​​obtained, which provides information about the spatial distribution and statistical characteristics of pressure data for each area.

[0048] The grouped feature matrix is ​​input into a state encoding network consisting of four convolutional layers for processing. Each convolutional layer uses a 3×3 convolution kernel and a LeakyReLU activation function to extract spatial features at different levels through convolution operations. The convolutional layer can capture the local spatial relationship between different areas in the sensor array, and the LeakyReLU activation function adds nonlinear characteristics, allowing the network to learn more complex spatial features. After processing through multiple convolutional layers, the state vector is obtained.

[0049] The state vector is input into the action generation network, which contains three fully connected layers to map the state vector to the initial weight distribution scheme of the sensor units. The fully connected layer generates the initial weight distribution scheme by learning the relationship between the state vector and the weight, indicating the importance of each sensor unit in different regions or different subsets.

[0050] The initial weight allocation scheme is input into the two-way Q-value evaluation network. The two-way Q-value evaluation network consists of four fully connected layers, which calculate the long-term value and short-term value of each state. The long-term value reflects the return that the sensor unit can obtain in the long term under the current weight allocation, while the short-term value reflects the immediate return at the current moment. By calculating these Q values, the network can evaluate the pros and cons of different weight schemes.

[0051] The state value function and advantage function are calculated based on the Q value sequence. The state value function represents the average return in a given state, while the advantage function measures the advantage of an action over other actions. By adding the state value function and the advantage function, the target Q value is obtained, which is used to evaluate the best weight allocation scheme in the current state.

[0052] In order to optimize the weight distribution scheme, random experience replay and temporal difference update are performed on the target Q value. Random experience replay is a technology in reinforcement learning that can randomly sample data from historical experience to prevent the model from falling into a local optimal solution. Temporal difference update is an incremental learning method that continuously updates model parameters by comparing the current Q value and actual return to achieve dynamic optimization. In this process, an optimal weight distribution scheme is determined by selecting the weight coefficient corresponding to the maximum Q value, and the weight of each unit in the sensor array is adjusted on this basis. After obtaining the optimized weight distribution, it is reconstructed according to the physical layout of the sensor array to generate a pressure data fusion weight matrix. This matrix contains the weight coefficient of each sensor unit and can be used to perform weighted averaging of pressure data in different regions to achieve pressure data fusion and optimization.

[0053] Step S600: Calculate the pressure center position and pressure gradient distribution of the pressure feature description matrix using the pressure data fusion weight matrix, and output target pressure mapping data.

[0054] Specifically, matrix multiplication is performed on the pressure feature description matrix and the pressure data fusion weight matrix, and the eigenvector of each spatial position is weighted. According to the importance of each sensing unit, its contribution to the overall pressure distribution is adjusted to obtain the initial fusion data matrix.

[0055] The first-order pressure gradient in the x- and y-directions is calculated based on the initial fused data matrix. The central difference method is used to effectively estimate the rate of change of pressure around each point, obtain the corresponding gradient operator, and obtain the pressure gradient vector field, which represents the change of the pressure field in all directions. The modulus of the pressure gradient vector field reflects the intensity of pressure change at each spatial position, while the directional decomposition can reveal the directional characteristics of the pressure change. Using polar coordinate transformation, the gradient amplitude and angle information are converted from the rectangular coordinate system to the polar coordinate system to obtain the pressure gradient distribution map. This figure can show the spatial variation law of the pressure field, especially in areas with drastic changes, and can identify the fluctuation characteristics of the pressure field.

[0056] According to the pressure gradient distribution map, the areas with drastic pressure changes are identified. These areas reflect the phenomenon of rapid changes or local concentration of pressure. The local features are enhanced by setting the gradient threshold, that is, by setting a threshold, the areas with gradients greater than the threshold are selected for local feature enhancement. In this way, the areas with drastic pressure changes are highlighted, making them occupy a more important position in subsequent processing. The enhanced pressure distribution map can more accurately reflect the characteristics of important areas in the pressure field, especially in the positioning of pressure peaks and valleys.

[0057] Perform centroid calculation on the enhanced pressure distribution to determine the center of pressure action. According to the enhanced pressure distribution map, the pressure value of each spatial position is used as a weight to calculate its weighted average position relative to the entire area, determine the center of action of the pressure field, and obtain the spatial coordinates of the pressure center position, which helps identify the main action area of ​​the pressure field and its distribution characteristics.

[0058] Based on the determined pressure center position, a polar coordinate grid is constructed to analyze the spatial distribution characteristics of the pressure field. In the polar coordinate system, the distribution of the pressure field is described by radial and tangential pressure components. The radial pressure component represents the change in pressure along the direction extending outward from the center point, while the tangential pressure component reflects the change in pressure in the direction perpendicular to the radial direction. By calculating these components, the spatial distribution characteristics of the pressure field are depicted, revealing how the pressure is distributed and changes in different directions.

[0059] The pressure spatial distribution features are fused with the original pressure data. The pressure spatial distribution features are fused with the original pressure data by weighted superposition to generate a multi-scale feature representation. The fused feature data is normalized and reconstructed to obtain the final target pressure mapping data, which reflects the spatial distribution features and important change areas of the pressure field detected by the pressure sensor array, thereby achieving high-precision pressure mapping.

[0060] In the embodiment of the present invention, by introducing a bicubic spline interpolation algorithm with tension coefficient compensation, the continuous pressure distribution between the sensing units is accurately reconstructed; a deep separable convolutional network is used to realize multi-scale feature extraction and fusion of pressure data, significantly reducing the computational complexity; a data reliability evaluation system is established based on the pressure state probability model and the confidence propagation mechanism; the intelligent fusion of pressure data is realized by adaptively calculating the fusion weights through a deep Q learning network; and a complete solution is constructed by combining pressure center positioning and gradient distribution analysis. While improving the pressure mapping accuracy, the present invention ensures a high degree of real-time performance, solves the problems of data distortion, low computational efficiency and insufficient reliability in traditional methods, and provides reliable technical support for the promotion and application of pressure sensor arrays in various application scenarios.

[0061] In a specific embodiment, the process of executing step S100 may specifically include the following steps: Standard pressure is applied to the high-precision pressure sensor array to obtain a standard pressure loading sequence, and based on the standard pressure loading sequence, a constant pressure is applied to each pressure point, and a high-speed data acquisition module is used to collect voltage output data of each sensor unit; Perform dynamic range calibration and zero drift compensation on the voltage output data, and perform wavelet denoising on the voltage output data to obtain the calibrated voltage data of each sensor unit; The calibration voltage data of each sensor unit is reconstructed according to the spatial distribution of the sensor unit, and the calibration voltage data of each sensor unit is piecewise linearly fitted to obtain a piecewise linear fitting coefficient; The least squares method is performed on the piecewise linear fitting coefficients and the standard pressure loading sequence to construct the pressure response equation of each sensing unit, obtain the unit pressure response coefficient, and organize the unit pressure response coefficient into an M×N dimensional matrix structure according to the spatial distribution of the sensor array to obtain the pressure response coefficient matrix; The pressure response coefficient matrix is ​​used to perform piecewise linear calculation and response curve correction on the real-time collected voltage data to obtain the pressure value of the sensor unit. The pressure value of the sensor unit is spatially reconstructed and boundary compensated, and an M×N dimensional original pressure distribution matrix is ​​generated according to the physical layout of the sensor array.

[0062] Specifically, a standard pressure loading scheme is designed for the pressure sensor array. The standard pressure loading sequence consists of a series of known, evenly distributed pressure points that cover the working range of the sensor array. By applying standard pressure loading to the array, it is ensured that each sensing unit can output voltage data under known pressure conditions.

[0063] For example, a set of standard pressures (such as 100 Pa, 200 Pa, 300 Pa, etc.) is applied to the sensor array through a precision air pressure source, and the voltage output data of each sensor unit is collected through a high-speed data acquisition module (such as a 24-bit ADC).

[0064] The acquired voltage output data is subjected to dynamic range calibration and zero drift compensation. Dynamic range calibration is to ensure the consistency of the sensor's response at different pressure values, especially the linearity and accuracy in the high and low pressure ranges. The calibration voltage data is compared with the known standard pressure points to adjust the sensor's output range so that it can maintain a linear response over the entire pressure range. Zero drift compensation is to eliminate the sensor output offset caused by environmental changes (such as temperature changes).

[0065] In order to compensate for the zero-point drift, a zero-pressure point (i.e., the output voltage when there is no pressure) is collected and compared with the actual output, and the corresponding correction is made according to the drift amount. Wavelet denoising is performed on the voltage output data. Wavelet denoising technology can effectively remove high-frequency noise from the original signal while retaining the effective low-frequency signal. Wavelet denoising uses the discrete wavelet transform method to decompose the signal into sub-signals of different frequency bands, and then selectively removes high-frequency noise according to the signal-to-noise ratio of different frequency bands, and reconstructs the denoised signal into clear voltage data.

[0066] In this embodiment, the high-precision pressure sensor array adopts a precise two-dimensional matrix layout structure, and its physical structure reflects the microelectronic integration technology. In terms of spatial arrangement, the sensing units are evenly distributed in the form of an M×N dimensional matrix, and each unit has a clear spatial coordinate identification (i, j). The precise spacing control between the sensing units is maintained, and the typical spacing is about 0.1 mm. This compact arrangement ensures a high spatial sampling density. In terms of structural design, flexible substrate materials are used to connect the various sensing units, which not only ensures the overall mechanical flexibility, but also improves the adaptability and durability of the sensor. The boundary area of ​​the array has been specially treated, and a boundary compensation mechanism has been adopted to effectively solve the edge effect problem. The entire array structure presents a clear hierarchy, including a sensing layer, a connection layer, and a protective layer. Each layer has been carefully designed and optimized to ensure the reliability of signal transmission and the accuracy of measurement. The calibrated voltage data is reconstructed according to the spatial distribution of the sensing units. Since the sensor array is usually a two-dimensional structure, the calibrated voltage data is spatially mapped by the position of the sensing unit in the array. For example, assuming the size of the sensor array is ,in and are the number of rows and columns of the array respectively. The calibrated voltage data is organized into a The matrix is ​​mapped to the corresponding spatial position to ensure that the voltage data of each sensor unit can accurately reflect its physical position.

[0067] Perform piecewise linear fitting on the calibration voltage data of each sensor unit, and fit the relationship between the calibration voltage data and the standard pressure into multiple linear equations. Assume that at the standard pressure point , At the sensor output voltages are , piecewise linear fitting is performed using the following formula: ; in, and Each pressure point The linear fitting coefficients on represent the slope and intercept respectively. Through this step, the pressure response equation of each sensing unit at each pressure point is obtained. In order to obtain the fitting coefficient, the least squares method is used to solve: ; When performing the least squares operation, the pressure response coefficient of each sensor unit is obtained by a standard algorithm (such as Gaussian elimination or gradient descent method). and , these coefficients reflect the response characteristics of the sensor to different pressure points. The unit pressure response coefficients are organized into an M×N dimensional matrix structure according to the spatial distribution of the sensor array. Each element in the matrix corresponds to the pressure response coefficient of a sensor unit, and each row and column of the matrix corresponds to a specific position in the array.

[0068] The voltage data collected in real time is processed by using the pressure response coefficient matrix. The pressure value of each sensor unit is obtained by performing piecewise linear calculation on the voltage data collected in real time and the pressure response coefficient matrix. Assume that the voltage value collected at a certain moment is , then through piecewise linear calculation, using the obtained response coefficient and Get the pressure value of the sensor unit : ; Through this method, the real-time voltage value of each sensor unit is converted into pressure to obtain the pressure value. In order to ensure the accuracy and integrity of the pressure data, spatial reconstruction and boundary compensation are performed. Spatial reconstruction refers to the reasonable interpolation of the pressure values ​​of different sensor units to ensure that each point in the area covered by the sensor array has valid pressure data. Boundary compensation is for the boundary part of the array. Considering that the sensor array cannot fully collect data at the edge position, compensation technology is used to process it so that the pressure data in the boundary area is as smooth as possible. Through the above steps, we can get The original pressure distribution matrix of dimensional is obtained. This matrix represents the pressure value of each sensing unit on the entire sensor array, thus achieving the goal of high-precision pressure mapping.

[0069] In a specific embodiment, the process of executing step S200 may specifically include the following steps: Perform differential operation on the pressure data of adjacent sensing units in the original pressure distribution matrix, determine the area where the pressure change rate exceeds the set threshold, and obtain the matrix of the area to be compensated; Based on the pressure gradient of the matrix of the area to be compensated, the local tension coefficient of each area to be compensated is calculated to obtain the spatial tension coefficient distribution, and the spatial tension coefficient distribution is subjected to Gaussian smoothing to obtain a smoothed tension coefficient matrix; The original pressure distribution matrix is ​​divided into regular grids, and a bicubic B-spline basis function is set for each grid node to obtain a grid interpolation basis function. The grid interpolation basis function is multiplied by the smoothing tension coefficient matrix to construct a cubic spline function with a tension term to obtain a compensation interpolation basis function. The compensation interpolation basis function is expanded in the x-direction and the y-direction respectively, and local polynomial fitting is performed to obtain the coordinates of the interpolation control points. According to the coordinates of the interpolation control points, bicubic spline interpolation calculation is performed on each point in the grid to generate a fine grid pressure data matrix. The fine grid pressure data matrix is ​​subjected to piecewise continuity constraints and derivative continuity corrections and reorganized into a continuous pressure distribution matrix.

[0070] Specifically, the pressure distribution matrix is ​​differentially calculated. Assume that the original pressure distribution matrix is ,in and Represent the rows and columns of sensor units in the array respectively. Assume the sensor array size is , each element in the matrix Indicates Row, No. The pressure value of the adjacent sensing units is calculated. The pressure change rate is determined by calculating the pressure difference between adjacent sensing units. The specific calculation formula is as follows: ; in, Representative position The pressure difference between the sensor and its adjacent sensor. According to the set pressure change threshold ,like , it is considered that the pressure change in this area is significant and needs to be compensated, forming a matrix of areas to be compensated ,in: ; The matrix identifies the areas that need to be compensated, where the area with a value of 1 indicates a dramatic pressure change, and the area with a value of 0 indicates a small change that can be ignored. Based on the pressure gradient of the area to be compensated, the local tension coefficient of each area to be compensated is calculated. The tension coefficient reflects the sensitivity of pressure changes, and the local tension strength is determined by calculating the pressure gradient. The pressure gradient is obtained by differential operation, assuming For location The pressure gradient at is calculated as: ; in, Indicates the pressure field The gradient of a point is given by and The local tension coefficient is derived from the gradient component of the direction. ; That is, the local tension coefficient is the modulus of the gradient vector. Gaussian smoothing is performed on the spatial tension coefficient to reduce the impact of high-frequency noise. Gaussian smoothing is achieved through convolution operations, and the specific formula is: ; in, is the Gaussian kernel function, defined as: ; in, is the standard deviation of the Gaussian kernel, which determines the range of smoothing, and is the size of the kernel, which is set according to the actual situation. After Gaussian smoothing, the smoothed tension coefficient matrix is ​​obtained. The original pressure distribution matrix is ​​divided into regular grids, assuming that the size of each grid is , and each grid node is set to a bicubic B-spline basis function The bicubic B-spline interpolation function is expressed as: ; in, and is the cubic B-spline basis function, is the corresponding pressure value on the grid node. By constructing the grid interpolation basis function, the pressure data is smoothly interpolated in the continuous space. Combined with the smooth tension coefficient matrix , multiply it with the grid interpolation basis function to construct a cubic spline function with tension term : ; The new interpolation function takes into account the original pressure data and introduces pressure gradient information and tension terms, which can better adjust the compensation area. Direction and The function is expanded in the direction and a local polynomial fitting is performed. The changing characteristics of the pressure field are accurately captured by local polynomial fitting. Assuming that a local quadratic polynomial is used for fitting in a certain area, the form is as follows: ; The coefficients of the polynomial are optimized by the least squares method , get the coordinates of the fitted control points. Perform bicubic spline interpolation calculation based on the interpolated control point coordinates. For each point in the grid, perform bicubic spline interpolation calculation: ; Generate fine grid pressure data matrix , which contains continuous pressure data obtained by bicubic spline interpolation. The fine grid pressure data matrix is ​​subjected to piecewise continuity constraints and derivative continuity corrections. This ensures a smooth transition of pressure values ​​between different regions and a smooth change of the gradient of the pressure field between regions. Through these constraints, a continuous pressure distribution matrix is ​​finally obtained.

[0071] In a specific embodiment, the process of executing step S300 may specifically include the following steps: The continuous pressure distribution matrix is ​​numerically normalized, and the normalized data is reconstructed into an input tensor to obtain network input data; The network input data is input into the first depth-separable convolutional layer, and spatial features are extracted for each input channel through a 3×3 convolution kernel. The ReLU activation function and batch normalization are used to obtain a spatial feature map. Perform 1×1 pointwise convolution operation on the spatial feature map to fuse the feature information of all channels, and perform data distribution normalization through the BatchNorm layer to obtain the first layer of channel features; The first layer channel features are input into the second depth separable convolution layer, and a 3×3 dilated convolution with a dilation rate of 2 is used to extract multi-scale spatial features. After ReLU activation and batch normalization, multi-scale spatial features are obtained. The multi-scale spatial features are subjected to channel dimension reduction through 1×1 convolution, and a skip connection structure is added to fuse with the input features. After being processed by the BatchNorm layer, the second layer of fusion features is obtained. The second layer fusion features are input into the third depth separable convolution layer, and the features are decomposed using grouped convolution. A 3×3 convolution kernel is used to extract local features in each group, and the grouped feature map is obtained through ReLU activation function and batch normalization. Perform channel rearrangement and 1×1 convolution operation on the group feature map, perform cross-group information interaction and fusion on the group features, and obtain the third-layer feature representation after processing by the BatchNorm layer; The third-layer feature representation is subjected to global average pooling and feature reconstruction to generate a pressure feature description matrix containing spatial features and channel features.

[0072] Specifically, the input continuous pressure distribution matrix is ​​numerically normalized. Assume there is a continuous pressure distribution matrix ,in and Represents the rows and columns in the matrix, and each element Represents the pressure value at a certain location. Scale all elements in the matrix to a specific range (for example, [0,1]). The normalization formula is as follows: ; in, and Represent the minimum and maximum values ​​in the original pressure matrix respectively. After this normalization process, Falling in the range of [0,1] helps to avoid the impact of data bias on network training. The normalized data is reconstructed into the input tensor of the network. Tensors are the basic data structure in deep learning and are used to represent multidimensional data. Assume that the pressure data matrix size is , reconstructing it into a shape of A 4D tensor of , where 1 represents the number of channels of the data. Input the network input data to the first depthwise separable convolution layer. Depthwise separable convolution is an optimized form of the standard convolution operation, which is achieved by breaking the convolution operation into two steps: depthwise convolution (convolution of each input channel separately), and then pointwise convolution (1x1 convolution to fuse the output of each channel). In the first layer, use The convolution kernel performs deep convolution and extracts spatial features for each input channel. The formula for the convolution operation is: ; in, is the output after convolution, is the weight of the convolution kernel, is the input tensor After using the ReLU activation function and batch normalization, we get the spatial feature map The formula for the ReLU activation function is as follows: ; Batch normalization can speed up the training process and improve the stability of the model. The formula for batch normalization is: ; in, and are the mean and standard deviation of the batch data, and are learnable scaling and offset parameters. Perform Pointwise convolution operation. The feature information of all channels is integrated to generate the channel features of the first layer. The formula of pointwise convolution operation is: ; in, is the number of channels, is the weight of each channel, It is The spatial feature map of the channels. After being processed by the BatchNorm layer, the channel features of the first layer are obtained. The channel features of the first layer are input to the second depth-separable convolutional layer. In this layer, the dilation rate is 2. Dilated convolution performs multi-scale spatial feature extraction. Dilated convolution captures a larger receptive field and extracts information at more scales by introducing spaced convolution kernels. The formula for dilated convolution is: ; in, is the void ratio, which determines the spacing between elements in the convolution kernel. After ReLU activation and batch normalization, the multi-scale spatial feature map of the second layer is obtained. Convolution performs channel dimension reduction and reduces the number of channels. A skip connection structure is added to fuse the input features with the features after convolution. The information in the input features is retained to avoid information loss, accelerate training and improve network performance. After being processed by the BatchNorm layer, the fused feature map of the second layer is obtained. The second layer fused feature map is input to the third depth-separable convolution layer. In this layer, grouped convolution is used for feature decomposition. Grouped convolution reduces the amount of calculation and helps to diversify the features by dividing the convolution operation into multiple groups. Each group uses The convolution kernel extracts local features, and the formula is: ; in, is the number of groups, is the convolution kernel within the group, is the first After ReLU activation and batch normalization, the group feature map of the third layer is obtained. Channel rearrangement and Convolution operation interacts and fuses group features across groups. Channel rearrangement swaps the features of each group to increase the representation capability of the network. The formula is: ; After being processed by the BatchNorm layer, the feature representation of the third layer is obtained. The feature representation of the third layer is reduced in dimension through global average pooling, and feature reconstruction is performed. The formula for global average pooling is: ; Global average pooling aggregates the spatial features of each channel into a scalar to obtain a pressure feature description matrix containing spatial features and channel features.

[0073] In a specific embodiment, the process of executing step S400 may specifically include the following steps: A Gaussian mixture distribution model is constructed for each spatial position in the pressure feature description matrix, and the pressure value is mapped to the probability space to obtain the initial state probability distribution; Based on the initial state probability distribution, a factor graph structure is constructed, and state variables and observation variables are assigned to each node to obtain a pressure state factor graph; Initialize the message transmission weights for the variable nodes in the pressure state factor graph, and calculate the amount of information transmission between nodes to obtain the initial message transmission matrix; The initial message passing matrix is ​​input into the belief propagation network, and the edge probability distribution of each node is calculated through forward propagation to obtain the node confidence value. The maximum a posteriori probability of the node confidence value is calculated, and iterative optimization is performed in combination with the spatial correlation constraint to obtain the updated node state. The mutual information between adjacent nodes is calculated according to the updated node status, and the message transmission weight is updated based on the mutual information to obtain the optimized message weight; The optimized message weights are used to perform weighted averaging of node states, build confidence propagation paths between nodes, obtain global confidence distribution, reconstruct global confidence distribution into a two-dimensional matrix form, and normalize confidence values ​​to obtain pressure confidence matrix.

[0074] Specifically, a Gaussian mixture distribution model is constructed for each spatial position in the pressure feature description matrix. Assuming that a pressure feature description matrix is ​​obtained ,in and represents the spatial position in the matrix, Represents the pressure value at each location. The Gaussian mixture model maps the pressure value at each location to a probability space, expressed as the weighted sum of multiple Gaussian distributions. The formula for the Gaussian mixture distribution is:

[0075] ; in, It is The weights of a Gaussian distribution, and They are The mean and standard deviation of a Gaussian distribution, It means the mean , standard deviation is Gaussian distribution. Through the Gaussian mixture model, the pressure value of each spatial position is mapped to the probability space to obtain the initial state probability distribution. Based on the initial state probability distribution, a factor graph structure is constructed. A factor graph is a graph model used to represent the relationship between variables. It consists of variable nodes and factor nodes. In this model, the variable node represents the pressure value of each spatial position, and the factor node represents the relationship between these pressure values ​​(such as the spatial correlation between adjacent positions). Each variable node is associated with a state variable , represents the pressure state, and each observation Corresponding to an observed variable Therefore, each edge in the factor graph connects a state variable and an observation variable. The state of each variable node is represented as: ; This shows that the state of each spatial location is determined by a Gaussian mixture model, and each observed variable Corresponding to pressure value After the factor graph is constructed, the message passing weight of each variable node is initialized. Message passing is an information transmission mechanism in the factor graph, where each node sends messages to its adjacent nodes. During initialization, an initial message passing weight is assigned to each node. , these weight values ​​represent the importance of information between adjacent nodes. The initial message passing matrix It is represented as a weight matrix between each pair of adjacent nodes: ; The initial message passing matrix is ​​input into the belief propagation network for message passing and edge probability calculation. Belief propagation is a common method for inferring variables in factor graphs. Through the belief propagation algorithm, the edge probability distribution of each node is calculated by forward propagation. The marginal probability distribution of each node is calculated by the following formula: ; in, Representation and Node The set of adjacent nodes, Is an adjacent node News, is the probability of observing the variable, is a normalization constant. The confidence value of the node is obtained by calculating the edge probability distribution. In order to optimize the state estimation of the node, the maximum a posteriori probability method is used to calculate the optimal state of each node: ; Combined with the spatial correlation constraint, in each iteration step, the information between adjacent nodes is used for iterative optimization to update the node status. The spatial correlation constraint means that the pressure values ​​of adjacent nodes should have a certain similarity, which is reflected in the smoothness of the node status. By iteratively updating the node status, the optimized state of each node is finally obtained. After the node status is updated, the mutual information between adjacent nodes is calculated. Mutual information measures the degree of information sharing between two nodes, and the calculation formula is: ; in, Is a node and nodes The joint probability of and are their marginal probabilities respectively. Based on the calculated mutual information, the weight of message passing is updated , making the message transmission between adjacent nodes more accurate. Using the optimized message transmission weight, the node states are weighted averaged to build the confidence propagation path between nodes. The global confidence distribution represents the state probability distribution of the entire pressure field, and the global confidence distribution is reconstructed into a two-dimensional matrix form. For each node of the global confidence distribution , and its confidence value is expressed as: ; The global confidence distribution is normalized so that the sum of all its elements is 1: The obtained pressure confidence matrix This is the stress state estimation result after optimization by belief propagation algorithm.

[0076] In a specific embodiment, the process of executing step S500 may specifically include the following steps: Two thresholds of 0.8 and 0.6 are set for the pressure confidence matrix, and the sensor units are divided into reliable areas, transition areas, and unreliable areas according to the confidence values ​​to obtain reliability grouping labels. Three independent sensor unit subsets are constructed based on the reliability grouping labels, and the local statistical features and spatial distribution features are calculated for each subset to obtain the grouping feature matrix. The grouped feature matrix is ​​input into a state encoding network consisting of four convolutional layers. Each convolutional layer uses a 3×3 convolution kernel and a LeakyReLU activation function to obtain a state vector through feature extraction. The state vector is mapped through the action generation network, which contains three fully connected layers. The output layer dimension is the total number of sensor units, and the initial weight distribution scheme is obtained; The initial weight distribution scheme is input into the dual-path Q-value evaluation network. Both dual-path Q-value evaluation networks contain four fully connected layers, which calculate the long-term value and short-term value respectively to obtain the Q-value sequence. The state value function and advantage function are calculated based on the Q-value sequence respectively, and the target Q-value is obtained by addition fusion. Random experience replay and temporal difference update are performed on the target Q value, and the weight coefficient corresponding to the maximum Q value is selected to obtain the optimized weight distribution. The optimized weight distribution is then reconstructed into a pressure data fusion weight matrix according to the physical layout of the sensor array.

[0077] Specifically, the pressure confidence matrix is ​​partitioned according to the set threshold. Assume that a pressure confidence matrix is ​​obtained , where each element represents the pressure confidence at that location. By setting two thresholds of 0.8 and 0.6, the matrix is ​​divided into three parts: reliable area, transition area, and unreliable area. For each location , according to its confidence value To determine the area to which it belongs: When , it belongs to the reliable area; when When , it belongs to the transition area; when Assign a reliability grouping label to each sensor unit ,in Reliable, transitional, unreliable . According to the reliability grouping labels, the sensor units are divided into three independent subsets: reliable subset, transitional subset and unreliable subset. For the sensor units in each subset, their local statistical features and spatial distribution features are calculated. These features include the mean, variance, maximum value, minimum value and other statistics of the sensor units, as well as their distribution patterns in space, such as the pressure changes in the surrounding neighborhood. For the features of each subset, a group feature matrix is ​​obtained, which contains the local statistical features and spatial distribution features of each sensor unit in the subset. Through these feature matrices, the sensor units in different regions are more accurately modeled and weighted. The group feature matrix is ​​input into a state encoding network with four convolutional layers for feature extraction. Each convolutional layer of the network uses a 3×3 convolution kernel, and the LeakyReLU activation function is used to increase the nonlinearity of the network. The role of the convolution operation is to filter the input data through the local receptive field to extract more representative spatial features. Batch normalization is used in each convolution layer to accelerate the training process and improve the stability of the model. After the feature map output by the convolutional layer is pooled, a state vector is obtained, which contains a high-level representation of the input features. The state vector is input into an action generation network. The action generation network consists of three fully connected layers, which generate an initial weight distribution scheme based on the state vector. In this network, each fully connected layer is calculated through linear transformation and activation function, and the final output weight vector represents the initial importance weight of each sensor unit in the pressure data fusion. The initial weight distribution scheme The input is sent to the dual-path Q-value evaluation network, which is used to evaluate the pros and cons of each weight allocation scheme. The dual-path Q-value evaluation network contains two parallel network paths, each of which contains four fully connected layers to calculate the long-term value and short-term value of each state. Reflects a state The cumulative reward after performing an action; short-term value It reflects the immediate reward of the current action. These two Q values ​​are added together to get the target Q value. : ; Based on the target Q value , updated through temporal difference learning. Use random experience replay to store and extract training data to reduce variance during training. For each extracted experience, update the target Q value and adjust the Q value estimate of each action according to the temporal difference algorithm. The formula for temporal difference update is: ; in, Indicates the current state. Indicates the currently selected action. It’s an instant reward. is the discount factor, Indicates the maximum Q value of the next state. Through this process, the target Q value will be continuously updated iteratively, and finally the weight coefficient corresponding to the maximum Q value will be selected as the optimized weight distribution scheme. According to the physical layout of the sensor array, it is reconstructed into a pressure data fusion weight matrix. This matrix represents the importance weight of each sensor unit in the data fusion process, and can dynamically adjust the weight according to the changes in pressure data in different areas, thereby optimizing the fusion effect of pressure data.

[0078] In a specific embodiment, the process of executing step S600 may specifically include the following steps: Perform matrix multiplication on the pressure feature description matrix and the pressure data fusion weight matrix, weight the feature vector at each spatial position, and obtain the initial fusion data matrix; The first-order pressure gradients in the x and y directions are calculated based on the initial fusion data matrix, and the gradient operator is constructed by the central difference method to obtain the pressure gradient vector field. The pressure gradient vector field is then modulus calculated and direction decomposed, and the gradient amplitude and angle information are extracted by polar coordinate transformation to obtain the pressure gradient distribution map. According to the pressure gradient distribution map, the area with drastic pressure changes is identified, and the gradient threshold is set to enhance the local features to obtain the enhanced pressure distribution. The center of mass calculation is performed on the enhanced pressure distribution, and the spatial coordinates of the pressure action center are determined by the weighted average method to obtain the pressure center position; A polar coordinate grid is constructed based on the pressure center position, and the radial and tangential pressure component distributions are calculated to obtain the pressure space distribution characteristics. The pressure space distribution characteristics are fused with the original pressure data, and a multi-scale feature representation is generated through weighted superposition to obtain fused feature data. The fused feature data is normalized and reconstructed to generate target pressure mapping data.

[0079] Specifically, matrix multiplication is performed on the pressure feature description matrix and the pressure data fusion weight matrix to weight the feature vector of each spatial position and obtain the initial fusion data matrix. Assume there is a pressure feature description matrix and a pressure data fusion weight matrix , and their dimensions are and ,in and Represent the number of rows and columns of the pressure matrix respectively. In order to combine these two matrices, we use matrix multiplication Weight the feature vector of each spatial position and initially fuse the data matrix , which is expressed as: ; in, represents the first Line The elements of the column, represents the first Line The elements of the column, is the weighted initial fusion data matrix. The result of matrix multiplication is to weight the eigenvectors of each spatial position to obtain the fused data representation. Based on the initial fusion data matrix Calculate the first-order pressure gradient. In order to calculate Direction and The pressure gradient in the direction is calculated and the gradient operator is constructed using the central difference method. Direction, the pressure gradient is calculated as: ; in, Represents the fused data matrix The pressure value in is the unit grid spacing; similarly, The pressure gradient in the direction is: ; in, yes The unit grid spacing in the direction. These two formulas calculate and The pressure gradient in the direction, the result is a pressure gradient vector field , where the gradient vector at each position contains Direction and Component of direction. Calculate the magnitude and decompose the direction of the pressure gradient vector field. For each position , the modulus of the pressure gradient Calculated by the following formula: ; At the same time, the direction of the gradient Calculated by the following formula: ; Among them, atan2 is the four-quadrant inverse tangent function, which can accurately give the angle information of the gradient direction. Through these calculations, the amplitude of the pressure gradient is obtained and direction , and then get the pressure gradient distribution map, which shows the pressure change trend and change intensity at different locations. According to the pressure gradient distribution map, the areas with drastic pressure changes are identified. In order to identify these areas, a gradient threshold is set , and the areas with drastic changes are marked by the following conditions: ; For locations that meet this condition , it is considered that the position belongs to the area with drastic pressure changes. On this basis, local feature enhancement is performed to increase the pressure value of these areas or adjust the weights to enhance their influence on the target in subsequent processing. After local feature enhancement, the enhanced pressure distribution is obtained. For the enhanced pressure distribution, the center of mass calculation is performed. The center of mass is calculated using the weighted average method, in which the pressure value of each position is used as the weight. The spatial coordinates of the center of mass Calculated by the following formula: ; ; in, It's location The pressure value, and is the corresponding spatial coordinate. Through these formulas, the center position of the pressure is calculated. A polar coordinate grid is constructed based on the pressure center position. Assume that the position of the pressure center is , for each grid point Calculate its polar coordinates relative to the pressure center. The radial component of the polar coordinates and the tangential component Respectively expressed as: ; ; Based on these polar coordinate components, the spatial distribution characteristics of pressure are obtained. These spatial distribution characteristics reflect the distribution pattern of pressure and the expansion trend of pressure in space. The pressure spatial distribution characteristics are fused with the original pressure data. By weighted superposition of the original data and spatial distribution characteristics, a multi-scale feature representation is generated. This process is expressed by the following formula: ; in, and is the weighting coefficient, is the raw pressure data, is the spatial distribution characteristic of pressure. The result of weighted superposition is a multi-scale feature representation, which reflects the characteristics of pressure data at different scales. After completing feature fusion, the fused feature data is normalized and reconstructed. Normalization is performed using the following formula: ; in, represents the normalized fusion feature data, and are the minimum and maximum values ​​of the fused feature data, respectively. Through normalization, all feature values ​​will be scaled to a uniform range, usually [0,1]. The target pressure mapping data is reconstructed.

[0080] See also Figure 2 , Figure 2 A schematic block diagram of the structure of a pressure mapping device 200 of a high-precision pressure sensor array provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the pressure mapping device 200 of the high-precision pressure sensor array includes: The acquisition module 210 is used to apply a known standard pressure to the high-precision pressure sensor array, collect the voltage output data of each sensor unit, convert it, and obtain an original pressure distribution matrix; An interpolation module 220 is used to perform a bicubic spline interpolation operation of tension coefficient compensation on the pressure data between adjacent sensing units in the original pressure distribution matrix to obtain a continuous pressure distribution matrix; An extraction module 230 is used to input the continuous pressure distribution matrix into a deep separable convolutional network to extract spatial features and fuse channel features to obtain a pressure feature description matrix; Iterative module 240, used to perform confidence propagation iterative calculation on the pressure feature description matrix through the pressure state probability model to obtain a pressure confidence matrix; A grouping module 250 is used to perform reliability grouping and weight coefficient calculation on the sensor units according to the pressure confidence matrix to obtain a pressure data fusion weight matrix; The output module 260 is used to calculate the pressure center position and pressure gradient distribution of the pressure feature description matrix using the pressure data fusion weight matrix, and output the target pressure mapping data.

[0081] Through the synergy of the above components, the continuous pressure distribution between the sensing units is accurately reconstructed by introducing a bicubic spline interpolation algorithm with tension coefficient compensation; the multi-scale feature extraction and fusion of pressure data are realized by using a deep separable convolutional network, which significantly reduces the computational complexity; a data reliability evaluation system is established based on the pressure state probability model and the confidence propagation mechanism; the intelligent fusion of pressure data is realized by adaptively calculating the fusion weights through a deep Q learning network; a complete solution is constructed by combining pressure center positioning and gradient distribution analysis. While improving the accuracy of pressure mapping, the present invention ensures a high degree of real-time performance, solves the problems of data distortion, low computational efficiency and insufficient reliability in traditional methods, and provides reliable technical support for the promotion and application of pressure sensor arrays in various application scenarios.

[0082] See also Figure 3 , Figure 3The present invention provides a schematic block diagram of the structure of a pressure mapping device 300 of a high-precision pressure sensor array, wherein the pressure mapping device 300 of a high-precision pressure sensor array includes a processor 301 and a memory 302, wherein the processor 301 and the memory 302 are connected via a device bus 303, wherein the memory 302 may include a non-volatile storage medium and an internal memory.

[0083] The non-volatile storage medium may store a computer program. The computer program includes program instructions, and when the program instructions are executed by the processor 301, the processor 301 may execute any of the above-mentioned pressure mapping methods for the high-precision pressure sensor array.

[0084] The processor 301 is used to provide computing and control capabilities to support the operation of the pressure mapping device 300 of the entire high-precision pressure sensor array.

[0085] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor 301, the processor 301 can execute any of the above-mentioned pressure mapping methods of the high-precision pressure sensor array.

[0086] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a partial structure related to the present application scheme, and does not constitute a limitation on the pressure mapping device 300 of the high-precision pressure sensor array involved in the present application scheme. The specific pressure mapping device 300 of the high-precision pressure sensor array may include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0087] It should be understood that the processor 301 may be a central processing unit (CPU), and the processor 301 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0088] It should be noted that technicians in the relevant field can clearly understand that for the convenience and conciseness of description, the specific working process of the pressure mapping device 300 of the high-precision pressure sensor array described above can refer to the corresponding process of the pressure mapping method of the aforementioned high-precision pressure sensor array, and will not be repeated here.

[0089] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by one or more processors, the one or more processors implement the pressure mapping method of the high-precision pressure sensor array provided in the embodiment of the present application.

[0090] The computer-readable storage medium may be an internal storage unit of the pressure mapping device 300 of the high-precision pressure sensor array of the aforementioned embodiment, such as a hard disk or memory of the pressure mapping device 300 of the high-precision pressure sensor array. The computer-readable storage medium may also be an external storage device of the pressure mapping device 300 of the high-precision pressure sensor array, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc., equipped with the pressure mapping device 300 of the high-precision pressure sensor array.

[0091] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0092] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0093] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A pressure mapping method for a high-precision pressure sensor array, characterized in that: include: Apply a known standard pressure to the high-precision pressure sensor array, collect the voltage output data of each sensor unit, convert it, and obtain the original pressure distribution matrix; Performing a bicubic spline interpolation operation with tension coefficient compensation on pressure data between adjacent sensing units in the original pressure distribution matrix to obtain a continuous pressure distribution matrix; Inputting the continuous pressure distribution matrix into a deep separable convolutional network to extract spatial features and fuse channel features to obtain a pressure feature description matrix; Performing confidence propagation iterative calculation on the pressure feature description matrix through a pressure state probability model to obtain a pressure confidence matrix; Reliability grouping and weight coefficient calculation of the sensor units are performed according to the pressure confidence matrix to obtain a pressure data fusion weight matrix; The pressure data fusion weight matrix is ​​used to calculate the pressure center position and pressure gradient distribution of the pressure feature description matrix, and the target pressure mapping data is output.

2. The pressure mapping method of the high-precision pressure sensor array according to claim 1, characterized in that: The method applies a known standard pressure to the high-precision pressure sensor array, collects the voltage output data of each sensor unit, and converts it to obtain an original pressure distribution matrix, including: Performing standard pressure loading on the high-precision pressure sensor array to obtain a standard pressure loading sequence, and applying a constant pressure to each pressure point based on the standard pressure loading sequence, and using a high-speed data acquisition module to collect voltage output data of each sensor unit; Performing dynamic range calibration and zero drift compensation on the voltage output data, and performing wavelet denoising on the voltage output data to obtain calibration voltage data of each sensor unit; The calibration voltage data of each sensor unit is reconstructed according to the spatial distribution of the sensor unit, and the calibration voltage data of each sensor unit is piecewise linearly fitted to obtain a piecewise linear fitting coefficient; Performing a least squares operation on the piecewise linear fitting coefficients and the standard pressure loading sequence, constructing a pressure response equation for each sensing unit to obtain a unit pressure response coefficient, and arranging the unit pressure response coefficients into an M×N dimensional matrix structure according to the spatial distribution of the sensor array to obtain the pressure response coefficient matrix; The pressure response coefficient matrix is ​​used to perform piecewise linear calculation and response curve correction on the voltage data collected in real time to obtain the pressure value of the sensing unit, and the pressure value of the sensing unit is spatially reconstructed and boundary compensated to generate an M×N dimensional original pressure distribution matrix according to the physical layout of the sensor array.

3. The pressure mapping method of the high-precision pressure sensor array according to claim 2, characterized in that: The bicubic spline interpolation operation of performing tension coefficient compensation on the pressure data between adjacent sensing units in the original pressure distribution matrix to obtain a continuous pressure distribution matrix includes: Performing differential operation on pressure data of adjacent sensing units of the original pressure distribution matrix, determining areas where the pressure change rate exceeds a set threshold, and obtaining a matrix of areas to be compensated; Based on the pressure gradient of the matrix of the area to be compensated, the local tension coefficient of each area to be compensated is calculated to obtain a spatial tension coefficient distribution, and the spatial tension coefficient distribution is subjected to Gaussian smoothing to obtain a smoothed tension coefficient matrix; The original pressure distribution matrix is ​​divided into a regular grid, a bicubic B-spline basis function is set for each grid node to obtain a grid interpolation basis function, and a product operation is performed on the grid interpolation basis function and the smoothing tension coefficient matrix to construct a cubic spline function with a tension term to obtain a compensation interpolation basis function; The compensation interpolation basis function is expanded in the x-direction and the y-direction respectively, and local polynomial fitting is performed to obtain the coordinates of the interpolation control points, and bicubic spline interpolation calculation is performed on each point in the grid according to the interpolation control point coordinates to generate a fine grid pressure data matrix; The fine grid pressure data matrix is ​​subjected to piecewise continuity constraints and derivative continuity corrections and reorganized into a continuous pressure distribution matrix.

4. The pressure mapping method of the high-precision pressure sensor array according to claim 3, characterized in that: The continuous pressure distribution matrix is ​​input into a deep separable convolutional network for spatial feature extraction and channel feature fusion to obtain a pressure feature description matrix, including: Performing numerical normalization processing on the continuous pressure distribution matrix, and reconstructing the normalized data into an input tensor to obtain network input data; Input the network input data into the first depth-wise separable convolutional layer, perform spatial feature extraction on each input channel through a 3×3 convolution kernel, use a ReLU activation function and batch normalization to obtain a spatial feature map; Performing a 1×1 pointwise convolution operation on the spatial feature map to fuse the feature information of all channels, and performing data distribution normalization processing through a BatchNorm layer to obtain the first layer of channel features; The first layer of channel features is input into the second depth-wise separable convolutional layer, and a 3×3 dilated convolution with a dilation rate of 2 is used to extract multi-scale spatial features. After ReLU activation and batch normalization, multi-scale spatial features are obtained. The multi-scale spatial features are subjected to channel dimension reduction through 1×1 convolution, and a skip connection structure is added to fuse with the input features, and processed through a BatchNorm layer to obtain a second layer of fusion features; The second layer of fusion features is input into the third depth-wise separable convolutional layer, and feature decomposition is performed using grouped convolution. A 3×3 convolution kernel is used to extract local features in each group, and a grouped feature map is obtained through ReLU activation function and batch normalization processing. Perform channel rearrangement and 1×1 convolution operation on the group feature map, perform cross-group information interaction and fusion on the group features, and obtain the third-layer feature representation after being processed by the BatchNorm layer; The third layer feature representation is subjected to global average pooling and feature reconstruction to generate a pressure feature description matrix including spatial features and channel features.

5. The pressure mapping method of the high-precision pressure sensor array according to claim 4, characterized in that: The pressure state probability model is used to perform confidence propagation iterative calculation on the pressure feature description matrix to obtain a pressure confidence matrix, including: Constructing a Gaussian mixture distribution model for each spatial position in the pressure feature description matrix, and mapping the pressure value to the probability space to obtain an initial state probability distribution; Constructing a factor graph structure based on the initial state probability distribution, assigning state variables and observation variables to each node, and obtaining a pressure state factor graph; Initializing message transfer weights for variable nodes in the pressure state factor graph, and calculating the amount of information transfer between nodes to obtain an initial message transfer matrix; Input the initial message passing matrix into the belief propagation network, calculate the edge probability distribution of each node through forward propagation, obtain the node confidence value, perform maximum a posteriori probability calculation on the node confidence value, perform iterative optimization in combination with spatial correlation constraints, and obtain the updated node state; Calculating mutual information between adjacent nodes according to the updated node status, and updating message transmission weights based on the mutual information to obtain optimized message weights; The optimized message weights are used to perform weighted averaging on the node states, build confidence propagation paths between nodes, obtain global confidence distribution, and reconstruct the global confidence distribution into a two-dimensional matrix form. The confidence values ​​are normalized to obtain a pressure confidence matrix.

6. The pressure mapping method of the high-precision pressure sensor array according to claim 5, characterized in that: The reliability grouping and weight coefficient calculation of the sensor units according to the pressure confidence matrix to obtain the pressure data fusion weight matrix includes: Two thresholds of 0.8 and 0.6 are set for the pressure confidence matrix, and the sensing units are divided into reliable areas, transition areas, and unreliable areas according to the confidence values ​​to obtain reliability grouping labels, and three independent subsets of sensing units are constructed according to the reliability grouping labels. The local statistical features and spatial distribution features of each subset are calculated to obtain a grouping feature matrix; The grouped feature matrix is ​​input into a state encoding network including four convolutional layers, each convolutional layer uses a 3×3 convolution kernel and a LeakyReLU activation function, and a state vector is obtained by feature extraction; Mapping the state vector through an action generation network, wherein the action generation network includes three fully connected layers, and the output layer dimension is the total number of sensor units, to obtain an initial weight distribution scheme; Input the initial weight allocation scheme into a dual-path Q-value evaluation network, each of which includes four fully connected layers, respectively calculating the long-term value and the short-term value to obtain a Q-value sequence, and respectively calculating the state value function and the advantage function based on the Q-value sequence, and obtaining the target Q-value by addition fusion; Random experience replay and time difference update are performed on the target Q value, the weight coefficient corresponding to the maximum Q value is selected to obtain an optimized weight distribution, and the optimized weight distribution is reconstructed into a pressure data fusion weight matrix according to the physical layout of the sensor array.

7. The pressure mapping method of the high-precision pressure sensor array according to claim 6, characterized in that: The using the pressure data fusion weight matrix to calculate the pressure center position and pressure gradient distribution of the pressure feature description matrix and outputting target pressure mapping data includes: Performing a matrix multiplication operation on the pressure feature description matrix and the pressure data fusion weight matrix, weighting the feature vector of each spatial position, and obtaining an initial fusion data matrix; Based on the initial fusion data matrix, the first-order pressure gradients in the x-direction and the y-direction are calculated, a gradient operator is constructed by a central difference method, a pressure gradient vector field is obtained, and the modulus value and direction decomposition of the pressure gradient vector field are performed, and the gradient amplitude and angle information are extracted by polar coordinate transformation to obtain a pressure gradient distribution map; Identify the area with drastic pressure changes according to the pressure gradient distribution map, set the gradient threshold to enhance the local features, obtain the enhanced pressure distribution, perform centroid calculation on the enhanced pressure distribution, determine the spatial coordinates of the pressure action center by using the weighted average method, and obtain the pressure center position; A polar coordinate grid is constructed based on the pressure center position, and the radial and tangential pressure component distributions are calculated to obtain pressure space distribution characteristics. The pressure space distribution characteristics are fused with the original pressure data, and a multi-scale feature representation is generated through weighted superposition to obtain fused feature data. The fused feature data is normalized and reconstructed to generate target pressure mapping data.

8. A pressure mapping device of a high-precision pressure sensor array, characterized in that: A method for performing pressure mapping of a high-precision pressure sensor array according to any one of claims 1 to 7, comprising: The acquisition module is used to apply a known standard pressure to the high-precision pressure sensor array, collect the voltage output data of each sensor unit, convert it, and obtain the original pressure distribution matrix; An interpolation module, used for performing a bicubic spline interpolation operation of tension coefficient compensation on the pressure data between adjacent sensing units in the original pressure distribution matrix to obtain a continuous pressure distribution matrix; An extraction module, used for inputting the continuous pressure distribution matrix into a deep separable convolutional network to perform spatial feature extraction and channel feature fusion to obtain a pressure feature description matrix; An iterative module, used for performing confidence propagation iterative calculation on the pressure feature description matrix through a pressure state probability model to obtain a pressure confidence matrix; A grouping module, used to perform reliability grouping and weight coefficient calculation on the sensor units according to the pressure confidence matrix to obtain a pressure data fusion weight matrix; The output module is used to calculate the pressure center position and pressure gradient distribution of the pressure feature description matrix using the pressure data fusion weight matrix, and output target pressure mapping data.

9. A pressure mapping device with a high-precision pressure sensor array, characterized in that: The pressure mapping device of the high-precision pressure sensor array includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instructions in the memory to enable the pressure mapping device of the high-precision pressure sensor array to execute the pressure mapping method of the high-precision pressure sensor array according to any one of claims 1 to 7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by a processor, the pressure mapping method of the high-precision pressure sensor array according to any one of claims 1 to 7 is implemented.

Images