Multi-region collaborative control method and device for high-precision pressure sensing array

Abstract

The present invention relates to a multi-region collaborative control method and device for a high-precision pressure sensor array. The method: based on a first collaborative sampling parameter, data is collected from multiple sensor units in a pressure sensor array to obtain a pre-processed pressure data set and a sensor unit spatial position matrix; a spatial weighted pressure covariance matrix is ​​constructed; eigenvalue decomposition and information entropy calculation are performed, and the number of target pressure points is determined according to the minimum description length criterion and the information gain ratio; a pressure feature subspace and a local perception domain are constructed; the pre-processed pressure data set is projected onto the pressure feature subspace to generate a sensor unit state data set; a multi-layer attention encoder is input for processing, and the first collaborative sampling parameter of the pressure sensor array is adjusted to obtain a second collaborative sampling parameter. The implementation of the present invention improves the system's response sensitivity to local pressure changes, thereby realizing multi-region collaborative control of pressure sensing.

Description

Technical Field

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

[0002] Traditional pressure sensor array control methods mainly rely on large-scale sampling data for statistical analysis. However, in actual medical application scenarios, due to factors such as surgical timeliness and equipment power consumption, it is often impossible to obtain sufficient sampling data.

[0003] Under high-dimensional and small sample conditions, traditional pressure sensor array control methods face two major challenges: on the one hand, limited sampling data makes it difficult to accurately characterize the spatiotemporal characteristics of pressure distribution, resulting in reduced measurement accuracy; on the other hand, the high-dimensional data generated by large-scale sensor arrays increases the computational complexity of signal processing, affecting the real-time response capability of the system. In addition, existing pressure sensor array control methods generally adopt fixed sampling strategies and cannot dynamically adjust sampling parameters according to pressure changes in different regions, which not only wastes sampling resources, but also limits the system's ability to respond quickly to local pressure changes. Summary of the invention

[0004] The main purpose of the present invention is to provide a multi-region collaborative control method and device for a high-precision pressure sensor array. The present invention improves the system's response sensitivity to local pressure changes, thereby realizing multi-region collaborative control of pressure sensors.

[0005] To achieve the above object, the present invention provides a multi-region collaborative control method for a high-precision pressure sensor array, comprising the following steps:

[0006] Based on the first collaborative sampling parameter, data is collected from multiple sensing units in the pressure sensor array to obtain pressure data, and wavelet transform noise reduction processing is performed on the pressure data to obtain a preprocessed pressure data set and a sensing unit spatial position matrix;

[0007] Calculating the pressure signal covariance of the sensing unit according to the preprocessed pressure data set and constructing a spatially weighted pressure covariance matrix;

[0008] Performing eigenvalue decomposition and information entropy calculation on the spatial weighted pressure covariance matrix, and determining the number of target pressure points according to a minimum description length criterion and an information gain ratio;

[0009] Based on the number of target pressure points and the spatial weighted pressure covariance matrix, construct a pressure feature subspace through singular value decomposition, and construct a local perception domain according to the pressure feature subspace;

[0010] Projecting the preprocessed pressure data set to the pressure feature subspace, performing weighted calculation on the sensor unit feature vectors in the perception domain through a hypernetwork structure, and generating a sensor unit state data set;

[0011] The sensing unit state data set is input into a multi-layer attention encoder for processing to generate a pressure control gain matrix, and the first collaborative sampling parameter of the pressure sensor array is adjusted to obtain a second collaborative sampling parameter.

[0012] The present invention also provides a multi-region collaborative control device for a high-precision pressure sensing array, comprising:

[0013] an acquisition module, configured to acquire data from a plurality of sensing units in the pressure sensor array based on a first collaborative sampling parameter to obtain pressure data, and to perform wavelet transform noise reduction processing on the pressure data to obtain a preprocessed pressure data set and a sensing unit spatial position matrix;

[0014] A calculation module, used to calculate the pressure signal covariance of the sensing unit according to the preprocessed pressure data set, and construct a spatial weighted pressure covariance matrix;

[0015] A decomposition module, used for performing eigenvalue decomposition and information entropy calculation on the spatial weighted pressure covariance matrix, and determining the number of target pressure points according to a minimum description length criterion and an information gain ratio;

[0016] A construction module, configured to construct a pressure feature subspace through singular value decomposition based on the number of target pressure points and the spatial weighted pressure covariance matrix, and to construct a local perception domain according to the pressure feature subspace;

[0017] A projection module, used to project the preprocessed pressure data set to the pressure feature subspace, perform weighted calculation on the sensor unit feature vectors in the perception domain through a hypernetwork structure, and generate a sensor unit state data set;

[0018] The adjustment module is used to input the sensor unit state data set into the multi-layer attention encoder for processing, generate a pressure control gain matrix, and adjust the first collaborative sampling parameter of the pressure sensor array to obtain a second collaborative sampling parameter.

[0019] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0020] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0021] In summary, the technical solution provided by the present invention has achieved a number of technical innovations in the sampling control, signal processing and parameter optimization of the sensor array by introducing a pressure covariance matrix construction technology based on second-order statistics, a method for determining the number of key pressure points based on the minimum description length criterion, a feature extraction mechanism based on a hypernetwork structure, and a dynamic parameter adjustment strategy of a multi-layer attention encoder. The present invention not only effectively solves the problem of signal feature extraction under high-dimensional small sample conditions, but also significantly improves the system's response sensitivity to local pressure changes through a multi-region collaborative control mechanism. In addition, the method adopts an adaptive sampling strategy to dynamically adjust the sampling parameters according to the pressure distribution characteristics, which reduces the system power consumption while ensuring the measurement accuracy, thereby realizing multi-region collaborative control of pressure sensing. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 1 is a schematic diagram of the steps of a multi-region collaborative control method of a high-precision pressure sensor array in one embodiment of the present invention;

[0023] Figure 2 It is a structural block diagram of a multi-region cooperative control device of a high-precision pressure sensing array in one embodiment of the present invention;

[0024] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0025] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0027] Reference Figure 1 This embodiment provides a multi-region collaborative control method for a high-precision pressure sensing array, comprising the following steps:

[0028] S1, collecting data from multiple sensing units in the pressure sensor array based on a first collaborative sampling parameter to obtain pressure data, and performing wavelet transform noise reduction processing on the pressure data to obtain a preprocessed pressure data set and a sensing unit spatial position matrix;

[0029] The pressure sensor array in this application adopts an N×M matrix layout structure. Each sensing unit has a unique identification number (i, j), where i∈[1, N] represents the row number and j∈[1, M] represents the column number. The physical layout of this array in three-dimensional space has specific spatial distribution characteristics, enabling multi-point collaborative acquisition of pressure signals. Compared with traditional pressure sensor arrays, the array layout in this application fully considers the spatial correlation between sensing units. By constructing a three-dimensional rectangular coordinate system, the physical positions of each sensing unit are accurately mapped, realizing the standardized characterization of the spatial positions of sensing units. This layout structure not only facilitates the construction of local sensing fields and feature extraction but also effectively supports the implementation of multi-region collaborative control strategies. A pressure sensor is a basic sensing unit that can convert pressure signals into electrical signals and has specific measurement ranges, sensitivities, and response characteristics. A pressure sensing array is a sensing system composed of multiple pressure sensors arranged in a specific geometric pattern, capable of simultaneously acquiring pressure information at multiple spatial positions. In this application, each pressure sensor is integrated into the array structure as an independent sensing unit, and position positioning is achieved through a unified numbering system and coordinate mapping. This integrated design enables the system to obtain more comprehensive pressure distribution information and improve the overall measurement accuracy and reliability through the collaborative action between sensing units. A pressure sensor and a pressure sensing array are in a relationship of component and system. The former is a basic functional unit, and the latter is an integrated measurement system.

[0030] Among them, parameter settings are performed based on the collaborative sampling frequency and collaborative sampling time window in the first collaborative sampling parameter. These parameters control the sampling frequency and time range, ensuring that data acquisition is synchronously performed on different sensing units of the sensor array within a predetermined time window. The obtained sampling control signal guides the pressure sensor array to perform data acquisition. The sensor array obtains the pressure values of the sensing units at preset time points according to the set sampling frequency and time window, forming a pressure data sample set, which contains the pressure measurement data of different sensing units during collaborative sampling.

[0031] Perform multi-scale decomposition on the pressure data sample set according to the sampling control signal. The signal is decomposed into components of multiple scales through wavelet transform to extract different frequency components. During the decomposition process, the obtained wavelet decomposition coefficients contain the low-frequency part and the high-frequency part of the signal, and the high-frequency part contains noise information. Perform adaptive threshold signal truncation processing on the high-frequency part. By setting a threshold, the high-frequency noise signal is suppressed or removed to obtain the denoised wavelet coefficients. Based on the denoised wavelet coefficients, the signal is reconstructed to restore the core features of the original signal, forming a preprocessed pressure data set.

[0032] A three-dimensional rectangular coordinate system is constructed according to the layout structure of the pressure sensor array. The physical position of each sensor unit in the sensor array is mapped to the coordinate system to obtain the spatial position information of each sensor unit. Through this step, an initial coordinate set is obtained, which contains the spatial position coordinates of all sensor units. The x, y, and z coordinate values ​​in the initial coordinate set are normalized and mapped, and these coordinate values ​​are converted into a unified range according to a certain ratio. The normalization operation helps to reduce the calculation error caused by the difference in sensor position, so that the coordinates of different sensor units have the same weight in the calculation. Through this process, a normalized coordinate set is obtained. The normalized coordinate set is organized into a three-dimensional matrix structure to obtain a sensor unit spatial position matrix. This matrix not only contains the spatial coordinate information of each sensor unit, but also reflects the spatial relationship between the various sensor units in the sensor array.

[0033] S2, calculating the pressure signal covariance of the sensing unit based on the preprocessed pressure data set and constructing a spatially weighted pressure covariance matrix;

[0034] Specifically, the pressure data of any two sensor units in the preprocessed pressure data set are averaged to obtain a pressure data mean matrix, which reflects the mean pressure change between each pair of sensor units and captures the basic pressure distribution between sensors. Based on the pressure data mean matrix, the second-order statistics of the pressure data of any two sensor units are calculated to calculate the covariance between the two sensor units. The covariance reflects the linear correlation of the pressure signals of the two sensor units and can reveal the strength and direction of the relationship between them, thus obtaining a basic covariance matrix, which represents the relationship between the pressure signals of different sensor units in the entire sensor array.

[0035] At the same time, considering that the sensor units in the sensor array have spatial distribution characteristics, the simple covariance matrix cannot fully reflect the spatial influence between different sensor units, so it is necessary to introduce spatial weights. According to the coordinate information in the sensor unit spatial position matrix, the spatial distance between any two sensor units is calculated to obtain a spatial distance matrix. This matrix reflects the distribution of sensor units in physical space. Based on the spatial distance matrix, the spatial distance is exponentially attenuated, so that the influence weight between sensor units with a longer distance gradually decreases, and a spatial weight matrix is ​​obtained. This matrix assigns weights to the spatial relationship between each pair of sensor units, so that the covariance influence between sensor units with a closer distance in space is more significant.

[0036] The basic covariance matrix and the spatial weight matrix are element-wise multiplied to obtain weighted covariance values, which more accurately reflect the spatial relationship between the sensor units and the correlation of the pressure signals. According to the dimensional information of the sensor array, the weighted covariance values ​​are reconstructed into a matrix to obtain the initial weighted pressure covariance matrix. This matrix contains the weighted covariance relationship between the various sensor units in the sensor array in space and pressure signals. The initial weighted pressure covariance matrix is ​​symmetrized to obtain a symmetric covariance matrix. The symmetric covariance matrix is ​​positively corrected, that is, the matrix is ​​mathematically positively definite, so that it has a good eigenvalue structure and can be used for subsequent algorithm calculations. Through positive definiteness correction, the unstable factors caused by numerical errors or model assumptions are eliminated to obtain the spatial weighted pressure covariance matrix.

[0037] S3, performing eigenvalue decomposition and information entropy calculation on the spatial weighted pressure covariance matrix, and determining the number of target pressure points according to the minimum description length criterion and the information gain ratio;

[0038] It should be noted that the spatial weighted pressure covariance matrix is ​​subjected to eigenvalue decomposition to obtain an eigenvalue sequence. Eigenvalue decomposition can decompose the covariance matrix into a set of eigenvalues ​​and eigenvectors. The eigenvalues ​​reflect the main direction and amplitude of changes in the data. Based on the eigenvalue sequence, the sum is calculated and each eigenvalue in the eigenvalue sequence is normalized to obtain a normalized eigenvalue sequence. Through normalization, the absolute size of the eigenvalue is eliminated, and the relative importance of each eigenvalue is more highlighted.

[0039] Based on the normalized eigenvalue sequence, information entropy is calculated. Information entropy is an indicator to measure the uncertainty and complexity of data. Calculating information entropy can help the system evaluate the contribution of different eigenvalues ​​to data changes. On this basis, the obtained information entropy index is processed according to the minimum description length criterion. The minimum description length criterion is a method to describe data by selecting the most concise model, which can help the system select the appropriate number of feature points to avoid overfitting or underfitting problems. In this step, the description length calculation sequence is obtained, which describes the simplicity of the model under different numbers of eigenvalues. The optimal number of feature points is determined by optimizing this sequence.

[0040] The information gain ratio sequence is obtained by processing the ratio of the cumulative sum to the total sum of the eigenvalue sequence. The information gain ratio measures the importance of each eigenvalue's contribution to the model and identifies which eigenvalues ​​dominate the overall model. By processing the information gain ratio, a feature set with a high amount of information is determined. The minimum value principle is used to search the description length calculation sequence to find the optimal number of pressure points, that is, the number of first pressure points. The information gain ratio sequence is judged based on the preset threshold to obtain the second number of pressure points. By setting a threshold, the eigenvalues ​​with information gain ratios higher than the threshold are screened out, and the pressure points with a greater impact on the model are identified. The weighted average operation of the first number of pressure points and the second number of pressure points is performed to obtain the target number of pressure points.

[0041] S4, based on the number of target pressure points and the spatial weighted pressure covariance matrix, construct the pressure feature subspace through singular value decomposition, and construct the local perception domain according to the pressure feature subspace;

[0042] Specifically, the spatial weighted pressure covariance matrix is ​​matrix decomposed, and the orthogonal matrix, singular value diagonal matrix and another orthogonal matrix of the matrix are obtained through singular value decomposition. The singular values ​​in the singular value diagonal matrix reflect the principal components in the data, and the orthogonal matrix contains the corresponding eigenvectors, reflecting the main direction of the data. On this basis, according to the number of target pressure points, the column vector of the orthogonal matrix is ​​intercepted, and the first few most important eigenvectors are extracted from it to form a new basis vector matrix. The basis vector matrix and its transpose are multiplied to obtain a new matrix in this way, and then the unit matrix is ​​subtracted from it to obtain the regularization term matrix. The role of the regularization term matrix is ​​to adjust the correlation of the eigenvectors, remove the redundant components, and improve the efficiency and accuracy of feature representation.

[0043] Based on the basis vector matrix and the regularization term matrix, a pressure feature subspace is constructed, which contains the most important directions and features in the pressure change. According to the pressure feature subspace, a local perception domain is constructed for each sensor unit. For each sensor unit, a perception domain containing the sensor units in the 3×3 area around it is constructed according to the neighborhood range of its position in the array. The perception domain reflects the influence range of the sensor unit in its local space, which can help capture the interaction and spatial correlation of the pressure signals of adjacent sensor units. The pressure features of each sensor unit in the perception domain set are normalized. The deviation caused by differences in measurement range, sensitivity, etc. between different sensor units is eliminated to ensure that the pressure features of all sensor units in the perception domain have the same scale, and a local feature vector set is obtained.

[0044] The local feature vector set is combined according to the physical location information of each sensor unit to obtain a multi-scale feature tensor that reflects the multi-level features of each area in the sensor array. The multi-scale feature tensor is subjected to spatial convolution operation, and the information of different spatial scales is combined through convolution operation to enhance the responsiveness and accuracy of the local perception domain. The convolution operation helps to extract the associated features in different spatial ranges, better capture the complex spatial and pressure change patterns in the pressure sensor array, and obtain the local perception domain.

[0045] S5, projecting the preprocessed pressure data set into the pressure feature subspace, performing weighted calculation on the sensor unit feature vectors in the perception domain through the hypernetwork structure, and generating a sensor unit state data set;

[0046] Among them, the preprocessed pressure data set is projected into the pressure feature subspace, and the representation of each sensor unit in the feature space is extracted. The pressure data is transformed by the most significant direction in the pressure feature subspace to obtain the projection coordinates of each sensor unit, reflecting the position of the sensor unit in the feature space. For each sensor unit, during the projection process, its projection coordinates in the 32-dimensional feature space are extracted, and combined with the relative position encoding of each sensor unit relative to the center of its local perception domain, this encoding helps to reflect the spatial relationship of the sensor unit in the local area, ensuring that the spatial position information in the perception domain can be effectively utilized to form the initial feature data.

[0047] Pressure data is extracted from the initial feature data in the local perception domain to calculate the real-time pressure data of each sensor unit and the pressure change rate at adjacent moments. Real-time pressure data can reflect the current working state of the sensor unit, while the pressure change rate helps to capture the dynamic characteristics of pressure changes. Through this step, the feature vector of each sensor unit is obtained. A three-layer fully connected neural network, namely a hypernetwork structure, is constructed based on the dimension of the sensor unit feature vector. The first hidden layer of the neural network contains 64 neurons, the second hidden layer contains 32 neurons, and the output layer also contains 32 neurons. This structural design extracts the complex features of each sensor unit through layer-by-layer processing to ensure that the deep information and nonlinear relationships in the pressure data can be captured. By inputting the feature vector of each sensor unit into the hypernetwork structure, the network obtains the hidden layer features of each sensor unit through nonlinear activation functions and mapping transformations of the fully connected layer. These hidden layer features reveal the potential characteristics of the sensor unit and reflect its dynamic changes in the local environment.

[0048] The attention mechanism is calculated based on the hidden layer features of the sensor unit to obtain the attention distribution coefficient, which measures the relative importance of each sensor unit in the local perception domain. In this way, the system can automatically adjust the attention to different sensor units, strengthen the response to key areas, and weaken the dependence on unimportant areas. The attention distribution coefficient is normalized using the Softmax function and converted into standardized attention weights. These weights can ensure that the features of each sensor unit maintain a reasonable proportional relationship when weighted combination. The state representation vector of each sensor unit is generated by weighted combination of the standardized attention weights and the feature vector of each sensor unit. The state representation vectors of all sensor units are organized and rearranged according to the layout structure of the sensor array to ensure that the state of each sensor unit can be reasonably spatially laid out in the array, which is convenient for subsequent processing and control. At the same time, the state representation vectors of all sensor units are spliced ​​and integrated to form a complete sensor unit state data set.

[0049] S6, input the sensor unit state data set into the multi-layer attention encoder for processing, generate a pressure control gain matrix, and adjust the first collaborative sampling parameter of the pressure sensor array to obtain the second collaborative sampling parameter.

[0050] Specifically, the sensor unit state data set is time-series expanded, and the state information of the sensor is arranged in a time series to capture the dynamic characteristics of the pressure data changing over time and obtain the state sequence data. The state sequence data is input into the first-layer attention module of the multi-layer attention encoder, and the query matrix Q, key matrix K and value matrix V are calculated in parallel through 8 attention heads. These matrices are used to capture the correlation between different sensor units. During the calculation process, the scaling operation is performed according to the product result of the query matrix and the key matrix to obtain the multi-head attention features of the first layer, which represent the potential correlation between the states of the sensor units at different time points.

[0051] The first layer of multi-head attention features are input into a two-layer feedforward neural network with a dimension of 256 for nonlinear transformation, and more complex feature representations are extracted through the inter-layer activation function of the network. In this process, the LayerNorm normalization layer and residual connection are combined to ensure the training stability of the model and effectively avoid the situation of gradient disappearance or explosion. Through this step, the output features of the first layer encoding are obtained, which contain the basic state information of the sensor unit and integrate multi-level time dependencies. According to the output features of the first layer encoding, a 3×3 local perception window is constructed for each sensor unit, and the spatial attention weights between the sensor units in the window are calculated. The local features are aggregated through these weights to obtain the spatial enhanced feature tensor. Through the spatial attention mechanism, the relationship between the sensor units is enhanced, especially the connection between adjacent sensor units, making the local perception more accurate and comprehensive.

[0052] The spatially enhanced feature tensor is sequentially input into the second, third, and fourth attention modules of the multi-layer attention encoder. Each layer adopts the same structure, that is, it contains 8 attention heads and a 256-dimensional feedforward neural network. At each layer, after residual connection and LayerNorm normalization, the output features of each layer are obtained. These features enhance the sensitivity of the system to pressure changes at different spatial and temporal scales through multi-layer encoding. The role of each layer of attention module is to continuously adjust and optimize the weights between sensor units to ensure that pressure changes in different regions can be effectively captured and responded to. After completing multi-layer encoding, the multi-level encoding feature sequence is adaptively weighted fused. According to the importance of each layer of encoding features and the historical state information of the sensor unit, the temporal correlation weight is calculated, thereby combining the information of time and space to generate a spatiotemporal fusion feature matrix.

[0053] The spatiotemporal fusion feature matrix is ​​input into a three-layer fully connected neural network for dimensionality reduction mapping, and the pressure control gain matrix is ​​obtained by layer-by-layer processing. The structure of the network includes three fully connected layers, and the number of neurons is 128, 64, and 32, respectively. Through the nonlinear transformation of these fully connected layers, the compressed key information is extracted from the high-dimensional features to obtain the gain coefficient corresponding to each sensor unit, which is used to adjust the accuracy and responsiveness of the pressure control. According to the gain coefficient of each sensor unit in the pressure control gain matrix, the collaborative sampling frequency and collaborative sampling time window in the first collaborative sampling parameter are weighted and adjusted to obtain the second collaborative sampling parameter. The sampling frequency and time window are adjusted according to the pressure change requirements in different areas to ensure that the system can respond more sensitively and efficiently in different pressure areas.

[0054] This application introduces the pressure covariance matrix construction technology based on second-order statistics, the method for determining the number of key pressure points based on the minimum description length criterion, the feature extraction mechanism based on the hypernetwork structure, and the dynamic parameter adjustment strategy of the multi-layer attention encoder. In combination with sampling control, signal processing and parameter optimization, it not only effectively solves the problem of signal feature extraction under high-dimensional small sample conditions, but also significantly improves the system's response sensitivity to local pressure changes through a multi-region collaborative control mechanism. In addition, the method adopts an adaptive sampling strategy to dynamically adjust the sampling parameters according to the pressure distribution characteristics, while ensuring the measurement accuracy. While reducing the system power consumption, it realizes the multi-region collaborative control of pressure sensing. This application solves the problem of difficult signal feature extraction under high-dimensional small samples through three levels: at the data acquisition level, the pressure covariance matrix construction technology based on second-order statistics is adopted, and the spatial correlation between the sensing units is fully utilized through spatial weighting, which effectively alleviates the difficulty of feature extraction caused by sample sparsity. At the feature extraction level, the number of key pressure points is determined by combining the minimum description length criterion and the information gain ratio, which realizes the dimensionality reduction and feature compression of high-dimensional data and improves the efficiency of feature extraction. At the feature expression level, through the cooperation of the hypernetwork structure and the multi-layer attention encoder, multi-scale expression and dynamic weight allocation of pressure features are achieved, overcoming the computational complexity problem in high-dimensional data processing, enabling the system to achieve stable and reliable feature extraction under small sample conditions.

[0055] In one example, data is collected from multiple sensing units in a pressure sensor array based on a first collaborative sampling parameter to obtain pressure data, and wavelet transform noise reduction processing is performed on the pressure data to obtain a preprocessed pressure data set and a sensing unit spatial position matrix, including:

[0056] Parameters are set based on the collaborative sampling frequency and the collaborative sampling time window in the first collaborative sampling parameter to obtain a sampling control signal, and data is collected on the pressure sensor array according to the sampling control signal to obtain the pressure value of the sensing unit at a preset time point to obtain a pressure data sample set;

[0057] According to the sampling control signal, the pressure data sample set is decomposed in multiple scales to obtain the wavelet decomposition coefficients, and the high-frequency part of the wavelet decomposition coefficients is subjected to adaptive threshold signal truncation processing to obtain the denoised wavelet coefficients;

[0058] Signal reconstruction is performed based on the denoised wavelet coefficients to obtain a preprocessed pressure data set;

[0059] A three-dimensional rectangular coordinate system is constructed according to the layout structure of the pressure sensor array, and the physical position of each sensor unit is mapped into the coordinate system to obtain an initial coordinate set;

[0060] The x, y, and z coordinate values ​​in the initial coordinate set are respectively normalized and mapped to obtain a normalized coordinate set, and the normalized coordinate set is organized into a three-dimensional matrix structure to obtain a sensing unit spatial position matrix.

[0061] In this example, the parameters are set based on the collaborative sampling frequency and collaborative sampling time window in the first collaborative sampling parameter, which determine the timing and accuracy of sensor data collection. Assume that the collaborative sampling frequency in the first collaborative sampling parameter is known. and co-sampling time windows ,in is the number of data sampling per unit time, is the time window for each sampling. Based on these parameters, a sampling control signal is constructed to control the moment of the sampling process. The sampling control signal is set to , which is a function of time and is used to indicate the start and end of each sampling period. The sampling control signal is expressed by the following formula:

[0062] ;

[0063] in, is the index of the sampling period, is the current time, is the length of the time window, A value of 1 indicates that sampling is performed during the sampling period, and a value of 0 indicates that sampling is not performed at the current moment. , collect data from the pressure sensor array. At the preset time point, according to the sampling control signal To trigger data collection and obtain the pressure value of each sensor unit. satisfy , then the pressure value of each sensing unit in the pressure sensor array is read at this moment ,in Indicates The sensor unit at the time The collected pressure data is organized into a pressure data sample set. ,in represents the index of the sensor unit, is the total number of sensor units in the sensor array, is a sequence of sampling time points. Get a pressure data sample set , contains the pressure data of the sensor array at multiple times. For the pressure data sample set , multi-scale decomposition is performed to extract information of different scales in the signal, thereby suppressing high-frequency noise. Multi-scale decomposition is achieved through wavelet transform. Assume that the pressure signal is a continuous pressure signal. For each sensing unit, the pressure signal , and perform wavelet decomposition. The result of wavelet transform is a series of wavelet coefficients, which are usually divided into different scales. These coefficients reflect the changes of the signal on different time scales. Assume that the decomposition coefficients obtained by wavelet transform are ,in is the scale parameter, is the time position parameter. The formula of wavelet transform is:

[0064] ;

[0065] in, is the wavelet basis function, is in scale and time location After wavelet transform, a matrix composed of wavelet coefficients is obtained. In order to remove the noise in the signal, the high frequency part of the wavelet coefficients is subjected to adaptive threshold signal truncation. Set a threshold , for each wavelet coefficient , if its absolute value is less than the threshold, it is considered to be noise and should be truncated to 0; otherwise, the coefficient is retained. The formula is expressed as:

[0066] ;

[0067] Through this step, high-frequency noise is effectively removed to obtain the denoised wavelet coefficients. Signal reconstruction is performed. The reconstructed signal is obtained by inversely applying the denoised coefficients to the wavelet basis function. The reconstruction formula is:

[0068] ;

[0069] Reconstructed signal This is the denoised pressure signal, which is used as the preprocessed pressure data set According to the layout structure of the pressure sensor array, a three-dimensional rectangular coordinate system is constructed to describe the spatial position of the sensor unit. The layout of the sensor array is set as a set of discrete points in three-dimensional space, representing the physical position of the sensor unit. Assume that each sensor unit The physical coordinates are , these coordinates are obtained by measuring the installation position of the sensor. After constructing the three-dimensional coordinate system, the physical position of each sensor unit is mapped to the coordinate system to obtain the initial coordinate set ,in is the index of the sensor unit. , and perform a normalization mapping on the coordinate values. Assume that , , , , and are respectively , and the minimum and maximum values of the coordinates. The normalization formula is:

[0070] ;

[0071] The normalized coordinate is within the range of [0, 1], ensuring that the spatial positions of all sensing units are within a unified scale range. Organize the set of normalized coordinates into a three-dimensional matrix structure, representing the spatial positions of each sensing unit in the sensor array. This matrix is represented as:

[0072] ;

[0073] The spatial position matrix C of the sensing unit is the final result describing the spatial positions of all sensing units in the sensor array.

[0074] In an example, according to the preprocessed pressure dataset, calculate the pressure signal covariance of the sensing units and construct a spatially weighted pressure covariance matrix, including:

[0075] Calculate the mean of the pressure data of any two sensing units in the preprocessed pressure dataset to obtain a pressure data mean matrix;

[0076] Based on the pressure data mean matrix, calculate the second-order statistics of the pressure data of any two sensing units to obtain a basic covariance matrix;

[0077] According to the coordinate information in the spatial position matrix of the sensing unit, calculate the spatial distance between any two sensing units to obtain a spatial distance matrix, and based on the spatial distance matrix, perform an exponential decay transformation on the spatial distance to obtain a spatial weight matrix;

[0078] Perform an element-wise product operation on the basic covariance matrix and the spatial weight matrix to obtain weighted covariance values, and according to the dimension of the sensor array, perform matrix reconstruction on the weighted covariance values to obtain an initial weighted pressure covariance matrix;

[0079] Perform a symmetrization process on the initial weighted pressure covariance matrix to obtain a symmetric covariance matrix, and perform a positive-definiteness correction on the symmetric covariance matrix to obtain a spatially weighted pressure covariance matrix.

[0080] In this example, the pressure dataset is preprocessed Include The pressure data of the sensor unit at multiple times, Indicates The sensor unit at time The pressure value of each sensor unit is calculated by averaging the pressure signal. Assume that the mean value of the pressure signal of each sensor unit at all sampling times is ,in is the total number of sampling moments, Indicates On this basis, a pressure data mean matrix is ​​constructed , each row in the matrix represents the mean pressure data of a sensor unit, that is:

[0081] ;

[0082] For any two sensor units and , and obtain the pressure signal difference between these sensing units by calculating their pressure data mean. Calculate the second-order statistics of the pressure data, namely the covariance. Covariance measures the linear relationship between two variables. Assume and The pressure data of each sensor unit are and , then their covariance Defined as:

[0083] ;

[0084] Through this formula, the covariance between any two sensor units is calculated to construct the basic covariance matrix , each item in this matrix Representative The sensor unit and The basic covariance matrix is ​​a quantitative expression of the signal dependency between all sensor units in the sensor array. The spatial distance between them is calculated based on the spatial position information of the sensor units. Assume that the position matrix of the sensor array is ,in It is The position of the sensor units in three-dimensional space. In order to calculate the spatial distance between any two sensor units , using the Euclidean distance formula:

[0085] ;

[0086] The obtained spatial distance matrix is a symmetric matrix in which each Indicates and In order to quantify the impact of spatial distance on covariance, based on the spatial distance matrix Perform exponential decay transformation to obtain the spatial weight matrix . Each entry of this matrix The spatial distance The weight value calculated by the exponential decay function uses the following formula:

[0087] ;

[0088] in is an attenuation factor that controls the degree of influence of spatial distance on weight. The weight between sensor units with smaller spatial distance is larger, and vice versa. Through this step, the spatial weight matrix is ​​constructed , expressing the spatial correlation between sensors. The basic covariance matrix and the spatial weight matrix Perform element-wise multiplication to obtain the weighted covariance matrix . Through the following formula:

[0089] ;

[0090] in Represents the element-wise product of matrices. Weighted covariance matrix Combining the covariance of sensor data and their spatial relationship, it reflects the weighted similarity between each sensor unit in the sensor array. The weighted covariance matrix is ​​symmetric so that each pair of non-diagonal elements in the matrix is ​​equal. , and its symmetry operation is realized by the following formula:

[0091] ;

[0092] in, is the transposed matrix of the weighted covariance matrix. The symmetry of the matrix is ​​ensured by symmetrization. The positive definiteness correction is performed on the symmetric covariance matrix to ensure that the matrix does not have negative eigenvalues ​​in numerical calculations. The positive definiteness correction method is to add a small constant to the diagonal elements, ensuring that all eigenvalues ​​are positive. Specifically, the correction operation is expressed as:

[0093] ;

[0094] in is the identity matrix, is a very small constant (e.g. ), ensuring the positive definiteness of the matrix. The obtained spatial weighted pressure covariance matrix is is a positive definite symmetric matrix that contains the weighted covariance information of the pressure signals between the sensing units and takes into account their spatial relationship.

[0095] In one example, eigenvalue decomposition and information entropy calculation are performed on the spatial weighted pressure covariance matrix, and the number of target pressure points is determined according to the minimum description length criterion and the information gain ratio, including:

[0096] Perform eigenvalue decomposition on the spatial weighted pressure covariance matrix to obtain an eigenvalue sequence;

[0097] The eigenvalue sequence is summed up, and each eigenvalue in the eigenvalue sequence is normalized to obtain a normalized eigenvalue sequence;

[0098] The information entropy is calculated based on the normalized eigenvalue sequence to obtain the information entropy index, and the information entropy index is processed according to the minimum description length criterion to obtain the description length calculation sequence;

[0099] The characteristic value sequence is processed according to the ratio of the cumulative sum to the total sum to obtain the information gain ratio sequence, and the description length calculation sequence is searched according to the minimum value principle to obtain the number of the first pressure points;

[0100] The information gain ratio sequence is judged based on a preset threshold value to obtain the second number of pressure points, and the first number of pressure points and the second number of pressure points are weighted averaged to obtain the target number of pressure points.

[0101] In this example, the spatially weighted pressure covariance matrix Perform eigenvalue decomposition to obtain its eigenvalues ​​and eigenvectors. Perform eigenvalue decomposition to obtain the eigenvalue sequence ,in is the number of sensing units. Each eigenvalue Indicates the variance contributed by a specific dimension in the sensor unit data. Eigenvalue decomposition is performed in the following way:

[0102] ;

[0103] in, is the eigenvector matrix, is a diagonal matrix whose diagonal elements are eigenvalues . The eigenvalue sequence Normalize each eigenvalue by the sum of the eigenvalue sequence, that is:

[0104] ;

[0105] in, is the normalized eigenvalue, representing the contribution ratio of each eigenvalue to the entire eigenvalue sequence. The sum of is equal to 1, that is:

[0106] ;

[0107] Based on the normalized eigenvalue sequence , calculate information entropy. Information entropy is used to measure the unpredictability or randomness of a set of data. Information entropy Calculated by the following formula:

[0108] ;

[0109] in, is the normalized eigenvalue, The entropy value of the entire eigenvalue sequence reflects the uncertainty contributed by these eigenvalues ​​at different pressure points. The larger the information entropy, the more dispersed the eigenvalue sequence is, indicating that the system has more pressure points. The information entropy is processed based on the minimum description length criterion. The minimum description length criterion is used to select an optimal model from multiple candidate models by balancing the complexity and fit of the model. The optimal number of pressure points is selected by calculating the description length of each information entropy indicator. Description length Calculated by the following formula:

[0110] ;

[0111] in, is the information entropy, is a weight factor, and complexity is a measure of model complexity. By minimizing the description length , and get a set of optimal pressure points. The eigenvalue sequence is processed according to the ratio of the cumulative sum to the total sum to get the information gain ratio sequence. The information gain ratio is used to measure the information gain that can be brought when selecting a eigenvalue. It is expressed as:

[0112] ;

[0113] in, Before the Representative The ratio of the cumulative contribution of the eigenvalues ​​to the total contribution. In this way, the importance of each eigenvalue to the system information is quantified. According to the minimum principle, the position corresponding to the minimum value is searched in the information gain ratio sequence to obtain the first number of pressure points. By selecting the point corresponding to the eigenvalue with the lowest information gain ratio, it is determined how many pressure points are needed to describe the system. The information gain ratio sequence is judged according to the preset threshold to obtain the second number of pressure points. This judgment is based on which positions in the information gain ratio sequence have large changes in eigenvalues, which can significantly improve the accuracy of the pressure model. The selection of the second number of pressure points depends on the analysis of the information gain ratio sequence to screen out points that have significant changes and meet physical constraints. The weighted average of the first number of pressure points and the second number of pressure points is taken to obtain the final target number of pressure points. The formula for weighted average is:

[0114] ;

[0115] in, and are the number of the first and second pressure points, and is a weight factor, which is set by experience or preset standards. The number of target pressure points obtained It is a comprehensive result based on multiple statistical indicators.

[0116] In one example, based on the number of target pressure points and the spatial weighted pressure covariance matrix, a pressure feature subspace is constructed by singular value decomposition, and a local perception domain is constructed according to the pressure feature subspace, including:

[0117] The spatial weighted pressure covariance matrix is ​​decomposed to obtain an orthogonal matrix, a singular value diagonal matrix and an orthogonal matrix, and the column vector of the orthogonal matrix is ​​intercepted according to the number of target pressure points to obtain a basis vector matrix;

[0118] Perform product operation on the basis vector matrix and the transpose of the basis vector matrix, and subtract the identity matrix to obtain the regularization term matrix;

[0119] A pressure feature subspace is constructed based on a basis vector matrix and a regularization term matrix, and a perception domain including adjacent sensor units within a 3×3 range is constructed for each sensor unit according to the pressure feature subspace to obtain a perception domain set;

[0120] Normalize the pressure features of each sensing unit in the perception domain set to obtain a local feature vector set;

[0121] The local feature vector set is combined according to the physical position information of the sensing unit to obtain a multi-scale feature tensor, and a spatial convolution operation is performed on the multi-scale feature tensor to obtain a local perception field.

[0122] In this example, based on the spatially weighted pressure covariance matrix , matrix decomposition is performed. The pressure covariance matrix is converted into a form that is easy to analyze. Through singular value decomposition, is decomposed into the product form of three matrices:

[0123] ;

[0124] where, and are orthogonal matrices, containing the left and right singular vectors of the matrix; is a diagonal matrix, and its diagonal elements are the singular values of the matrix, representing the main information components in the pressure covariance matrix. Each singular value corresponds to the intensity of a characteristic direction, and larger singular values correspond to important characteristic directions in the data. The number of target pressure points determines the number of basis vectors selected from the decomposition. To obtain the basis vector matrix, the first columns are intercepted from the orthogonal matrix to obtain the basis vector matrix :

[0125] ;

[0126] Construct the regularization term matrix. The product of the basis vector matrix and its transpose helps to evaluate the inner product relationship between the basis vectors, regularize the calculated matrix to prevent overfitting of the data or excessive weights in subsequent calculations. Perform the matrix product operation:

[0127] ;

[0128] By subtracting this result from the identity matrix , the regularization term matrix

[0129] is obtained;

[0130] where, is the identity matrix, is a symmetric matrix, representing the deviation between the basis vector matrix and the identity matrix. Based on the basis vector matrix and the regularization term matrix , construct the pressure eigen-subspace. Extract the most representative dimensions from the high-dimensional pressure data, and these dimensions can most effectively represent the state of the pressure sensor array. Define the pressure eigen-subspace in the following way:

[0131] ;

[0132] Among them, is the matrix of the pressure feature subspace, representing the pressure data in the direction of the regularized basis vectors. A local perception field is constructed based on the characteristics of the pressure feature subspace. The local perception field refers to selecting a region containing adjacent sensing units based on the spatial position of each sensing unit in order to extract the pressure features within this region. A 3×3 perception window is selected to define the perception field of each sensing unit. Assume that each sensing unit in the sensor array has coordinates in three-dimensional space as , then the perception field of each sensing unit includes other sensing units in the surrounding 3×3 region, expressed as:

[0133] ;

[0134] Among them, is the maximum distance of the perception field, represents the Euclidean distance metric. In this way, the perception field set is obtained, and each perception field corresponds to the neighborhood of a sensing unit. For each perception field , the pressure features it contains are normalized to remove the dimensional differences caused by different positions and sizes of the sensing units, ensuring that all features are compared under the same standard. The normalization process is achieved by subtracting the mean and dividing by the standard deviation of the pressure features within each perception field. The formula is:

[0135] ;

[0136] Among them, is the pressure feature of the th sensing unit in the perception field , and are the mean and standard deviation of the pressure features of all sensing units in the perception field respectively, is the normalized pressure feature. The set of all normalized pressure feature vectors is combined to form the local feature vector set , and it is combined according to the physical position information of the sensing units to form the multi-scale feature tensor . A spatial convolution operation is performed on the multi-scale feature tensor . The convolution operation can effectively capture the spatial relationship between local features and extract the key information within each perception field. Assume that the convolution kernel is , then the convolution operation is performed through the following formula:

[0137] ;

[0138] Among them, * represents the convolution operation, It is the output feature after convolution, which represents the enhanced feature of the local perception domain.

[0139] In one example, the preprocessed pressure data set is projected into the pressure feature subspace, and the sensor unit feature vectors in the perception domain are weighted calculated through the hypernetwork structure to generate a sensor unit state data set, including:

[0140] The preprocessed pressure data set is projected into the pressure feature subspace, the projection coordinates of each sensor unit in the 32-dimensional feature space are extracted, and the initial feature data is obtained by combining the relative position encoding relative to the center of the local perception domain;

[0141] Extract pressure data from the initial feature data in the local perception domain, calculate the real-time pressure data of each sensor unit and the pressure change rate at adjacent moments, and obtain the sensor unit feature vector;

[0142] A three-layer fully connected neural network is constructed according to the dimension of the sensor unit feature vector to obtain a super network structure. The first hidden layer of the super network structure contains 64 neurons, the second hidden layer contains 32 neurons, and the output layer contains 32 neurons.

[0143] The sensor unit feature vector is input into the hypernetwork structure, and the corresponding sensor unit hidden layer features are obtained through nonlinear activation function and full connection layer mapping transformation. The attention mechanism is calculated based on the sensor unit hidden layer features to obtain the attention distribution coefficient.

[0144] The attention distribution coefficient is normalized by the Softmax function to obtain the standardized attention weight. According to the standardized attention weight, the features of the sensor units in the local perception domain are weighted and combined to generate the state representation vector of each sensor unit.

[0145] The state representation vectors of all sensor units are organized and rearranged according to the array layout structure, and are spliced ​​and integrated to obtain a sensor unit state data set.

[0146] In this example, the preprocessed pressure dataset is projected into the pressure feature subspace, and the projection coordinates of each sensor unit in the 32-dimensional feature space are extracted. Assume that the preprocessed pressure dataset , which contains the pressure data of each sensor unit at different time points. A feature subspace constructed by singular value decomposition or principal component analysis is used. Assume that the pressure feature subspace matrix is , whose dimensions are (where 32 is the dimension of the target feature space), the pressure data of each sensor unit The process of projection to the feature subspace is expressed as:

[0147] ;

[0148] in, It is the sensing unit The projection coordinates in the 32-dimensional feature space, is the characteristic subspace matrix, It is the sensing unit The original pressure data. Combined with the relative position encoding of the sensor unit relative to the center of the local perception field, this means that each sensor unit not only has a feature vector, but also needs to consider its position relationship in the array. Assuming that the sensor unit The physical location is , and the center of the local perception domain is , then the relative position encoding It is expressed as:

[0149] ;

[0150] in, is the maximum distance of the local perception domain (usually the farthest distance between sensing units). Combined to get the initial feature data

[0151] ;

[0152] Extract pressure data from the initial feature data in the local perception domain. Calculate the real-time pressure data of each sensor unit and the pressure change rate at adjacent moments ,in It is the sensing unit At the moment The pressure value of is the difference between the pressure value and the pressure value at the previous moment, expressed as:

[0153] ;

[0154] Through this step, the feature vector of each sensor unit is obtained , the feature vector contains real-time pressure data and pressure change rate. That is:

[0155] ;

[0156] According to the dimension of each sensor unit’s feature vector, a three-layer fully connected neural network is constructed. Assume that the feature vector of each sensor unit is The dimension is 2 (i.e. and ), a three-layer fully connected neural network is constructed to process these features. The first hidden layer contains 64 neurons, the second hidden layer contains 32 neurons, and the output layer contains 32 neurons. The calculation of each layer is expressed as:

[0157] ;

[0158] ;

[0159] ;

[0160] in, is the weight matrix of each layer, is the bias term, is the activation function (such as ReLU or Sigmoid). The output layer of the network Represents the hidden features of each sensor unit. Through these hidden features, the attention mechanism is used to weight the correlation between different sensor units. The attention mechanism calculates the weight of each sensor unit according to its contribution to the overall feature. Assume that the calculated attention distribution coefficient is , then the attention distribution coefficient of each sensor unit is calculated by the following formula:

[0161] ;

[0162] in, It is The Softmax function ensures that the sum of the attention distribution coefficients of all sensor units is 1. After normalizing the attention distribution coefficients through the Softmax function, the standardized attention weight is obtained. :

[0163] ;

[0164] According to the standardized attention weight, the features of the sensor units in the local perception domain are weighted and combined to generate the state representation vector of each sensor unit. :

[0165] ;

[0166] The state representation vector of all sensor units Rearrange and splice according to the physical layout structure of the sensor array to form the final sensor unit state data set :

[0167] ;

[0168] in, It is a matrix containing the status information of all sensor units.

[0169] In one example, a sensor unit state data set is input into a multi-layer attention encoder for processing, a pressure control gain matrix is ​​generated, and a first collaborative sampling parameter of a pressure sensor array is adjusted to obtain a second collaborative sampling parameter, including:

[0170] Perform time series expansion on the sensor unit state data set to obtain state sequence data;

[0171] The state sequence data is input into the first-layer attention module of the multi-layer attention encoder. The query matrix Q, key matrix K and value matrix V with a dimension of 64 are calculated through 8 attention heads respectively, and scaled according to the product result of the query matrix Q and the key matrix K to obtain the first-layer multi-head attention feature;

[0172] The first layer of multi-head attention features are input into a two-layer feedforward neural network with a dimension of 256 for nonlinear transformation, and combined with the LayerNorm normalization layer and residual connection to obtain the first layer of encoded output features;

[0173] According to the first layer encoding output features, a 3×3 local perception window is constructed for adjacent sensor units, the spatial attention weights between sensor units in each local perception window are calculated, and local features are aggregated to obtain a spatial enhanced feature tensor;

[0174] The spatially enhanced feature tensor is sequentially input into the second-layer attention module, the third-layer attention module, and the fourth-layer attention module of the multi-layer attention encoder, and the output features of each layer are residually connected and LayerNorm normalized to obtain a multi-level encoding feature sequence, where the second-layer attention module, the third-layer attention module, and the fourth-layer attention module have the same structure, and each layer of the attention module contains 8 attention heads and a 256-dimensional feedforward network;

[0175] Adaptively weighted fusion is performed on multi-level coding feature sequences, and the time-series correlation weights are calculated based on the historical state information of the sensor unit to obtain a spatiotemporal fusion feature matrix;

[0176] The spatiotemporal fusion feature matrix is ​​input into a three-layer fully connected network for dimensionality reduction mapping to obtain a pressure control gain matrix. The number of neurons in the three-layer fully connected network is 128, 64, and 32 respectively.

[0177] Based on the gain coefficient corresponding to each sensing unit in the pressure control gain matrix, the collaborative sampling frequency and the collaborative sampling time window in the first collaborative sampling parameter are weighted adjusted to obtain the second collaborative sampling parameter.

[0178] In this example, the sensor unit status data set is time-expanded. Assume there is a sensor unit status data set ,in represents the number of time steps, and It is The sensor unit status data at the moment. The data set is expanded in time series, and the data at each moment is arranged in chronological order to obtain the state sequence data. . Input the state sequence data into the first layer attention module of the multi-layer attention encoder. Calculate the query, key, and value matrix for the input state sequence data. Assume that the input is , mapping it to three matrices: query matrix , the bond matrix , and the value matrix The dimensions of these matrices are , where 64 is the feature dimension used by the system at this layer. The process of calculating the query, key, and value matrix is ​​expressed as:

[0179] ;

[0180] in, ,and is a trainable weight matrix used to generate queries, keys, and values. The dot product of the query matrix and the key matrix is ​​calculated and scaled to calculate the attention weight matrix :

[0181] ;

[0182] in, is the dimension of the key vector, which is equal to 64. Through the above calculation, we get the attention matrix , the matrix is ​​normalized by the softmax function to obtain the attention weight matrix :

[0183] ;

[0184] By combining the attention weight matrix with the value matrix Multiply them together to get the multi-head attention features of the first layer:

[0185] ;

[0186] The output of this layer Input feedforward neural network. Assume that the feedforward neural network consists of two layers, each with 256 neurons. Input features Nonlinear transformation is performed through two layers of fully connected networks:

[0187] ;

[0188] ;

[0189] in, is the weight matrix, is the bias term, is the activation function (usually ReLU is used). Add layer normalization and residual connection:

[0190] ;

[0191] The encoded output features of the first layer are . Output features according to the encoding of the first layer , construct a 3×3 local perception window, and calculate the spatial attention weights between the sensor units within the local perception window. The local perception window is a spatial enhancement operation that allows the model to focus on the relationship between the sensor units in the local area. The calculation of the spatial attention weight is similar to the aforementioned attention mechanism, which calculates the similarity within the local area. Assume that The characteristics of the sensor unit within the local perception window are: , then the spatial attention weight It is expressed as:

[0192] ;

[0193] Perform weighted aggregation on local features to obtain the spatial enhanced feature tensor:

[0194] ;

[0195] The spatially enhanced feature tensor is fed into the second, third, and fourth layers of the multi-layer attention encoder in sequence. Each layer has the same structure, consisting of 8 attention heads and a 256-dimensional feedforward network. The output features are processed by residual connections and layer normalization. The output of this process is a multi-level encoded feature sequence , which contains deep spatiotemporal information and features. Adaptive weighted fusion is performed. The time-series correlation weights are calculated based on the historical state information of the sensor unit to obtain the spatiotemporal fusion feature matrix , by weighted fusion of the features at each moment, the final spatiotemporal feature representation is obtained:

[0196] ;

[0197] in, is the time series correlation weight at each moment. The spatiotemporal fusion feature matrix is ​​input into the three-layer fully connected network for dimensionality reduction mapping to obtain the pressure control gain matrix The number of neurons in the three-layer fully connected network is 128, 64, and 32 respectively, and the calculation of each layer is expressed as:

[0198] ;

[0199] ;

[0200] ;

[0201] What you get is the pressure control gain matrix. Based on the gain coefficient corresponding to each sensing unit, the collaborative sampling frequency and the collaborative sampling time window in the first collaborative sampling parameter are weighted adjusted to obtain the second collaborative sampling parameter.

[0202] Reference Figure 2 , this embodiment provides a multi-region collaborative control device of a high-precision pressure sensing array, comprising:

[0203] Acquisition module 1, used to collect data from multiple sensing units in the pressure sensor array based on the first collaborative sampling parameter to obtain pressure data, and perform wavelet transform noise reduction processing on the pressure data to obtain a preprocessed pressure data set and a sensing unit spatial position matrix;

[0204] Calculation module 2, used to calculate the pressure signal covariance of the sensing unit according to the preprocessed pressure data set, and construct a spatial weighted pressure covariance matrix;

[0205] Decomposition module 3, used for performing eigenvalue decomposition and information entropy calculation on the spatial weighted pressure covariance matrix, and determining the number of target pressure points according to the minimum description length criterion and the information gain ratio;

[0206] A construction module 4 is used to construct a pressure feature subspace through singular value decomposition based on the number of target pressure points and the spatial weighted pressure covariance matrix, and to construct a local perception domain according to the pressure feature subspace;

[0207] Projection module 5, used to project the preprocessed pressure data set to the pressure feature subspace, perform weighted calculation on the sensor unit feature vectors in the perception domain through the hypernetwork structure, and generate a sensor unit state data set;

[0208] The adjustment module 6 is used to input the sensor unit state data set into the multi-layer attention encoder for processing, generate a pressure control gain matrix, and adjust the first collaborative sampling parameter of the pressure sensor array to obtain the second collaborative sampling parameter.

[0209] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0210] Reference Figure 3In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0211] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0212] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0213] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0214] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0215] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A multi-region collaborative control method for a high-precision pressure sensor array, characterized in that: The following steps are involved: Based on the first collaborative sampling parameter, data is collected from multiple sensing units in the pressure sensor array to obtain pressure data, and wavelet transform noise reduction processing is performed on the pressure data to obtain a preprocessed pressure data set and a sensing unit spatial position matrix; According to the preprocessed pressure data set, the pressure signal covariance of the sensor unit is calculated to construct a spatial weighted pressure covariance matrix; specifically comprising: performing mean calculation on the pressure data of any two sensor units in the preprocessed pressure data set to obtain a pressure data mean matrix; performing second-order statistics calculation on the pressure data of any two sensor units based on the pressure data mean matrix to obtain a basic covariance matrix; performing spatial distance calculation on any two sensor units based on the coordinate information in the spatial position matrix of the sensor units to obtain a spatial distance matrix, and performing exponential decay transformation on the spatial distance based on the spatial distance matrix to obtain a spatial weight matrix; performing element-wise product operation on the basic covariance matrix and the spatial weight matrix to obtain weighted covariance values, and performing matrix reconstruction on the weighted covariance values ​​according to the sensor array dimension to obtain an initial weighted pressure covariance matrix; performing symmetry processing on the initial weighted pressure covariance matrix to obtain a symmetric covariance matrix, and performing positive definiteness correction on the symmetric covariance matrix to obtain a spatial weighted pressure covariance matrix; Performing eigenvalue decomposition and information entropy calculation on the spatial weighted pressure covariance matrix, and determining the number of target pressure points according to a minimum description length criterion and an information gain ratio; Based on the number of target pressure points and the spatial weighted pressure covariance matrix, construct a pressure feature subspace through singular value decomposition, and construct a local perception domain according to the pressure feature subspace; Projecting the preprocessed pressure data set to the pressure feature subspace, performing weighted calculation on the sensor unit feature vectors in the perception domain through a hypernetwork structure, and generating a sensor unit state data set; The sensing unit state data set is input into a multi-layer attention encoder for processing to generate a pressure control gain matrix, and the first collaborative sampling parameter of the pressure sensor array is adjusted to obtain a second collaborative sampling parameter.

2. The multi-region coordinated control method of the high-precision pressure sensor array according to claim 1 is characterized in that: The method collects data from a plurality of sensing units in the pressure sensor array based on the first collaborative sampling parameter to obtain pressure data, and performs wavelet transform noise reduction processing on the pressure data to obtain a preprocessed pressure data set and a sensing unit spatial position matrix, including: Parameters are set based on the collaborative sampling frequency and the collaborative sampling time window in the first collaborative sampling parameter to obtain a sampling control signal, and data is collected on the pressure sensor array according to the sampling control signal to obtain the pressure value of the sensing unit at a preset time point to obtain a pressure data sample set; Performing multi-scale decomposition on the pressure data sample set according to the sampling control signal to obtain wavelet decomposition coefficients, and performing adaptive threshold signal truncation processing on the high-frequency part of the wavelet decomposition coefficients to obtain denoised wavelet coefficients; Reconstruct the signal based on the denoised wavelet coefficients to obtain a preprocessed pressure data set; Constructing a three-dimensional rectangular coordinate system according to the layout structure of the pressure sensor array, mapping the physical position of each sensor unit into the coordinate system, and obtaining an initial coordinate set; The x, y, and z coordinate values ​​in the initial coordinate set are respectively normalized and mapped to obtain a normalized coordinate set, and the normalized coordinate set is organized into a three-dimensional matrix structure to obtain a sensing unit spatial position matrix.

3. The multi-region coordinated control method of the high-precision pressure sensor array according to claim 1 is characterized in that: The performing eigenvalue decomposition and information entropy calculation on the spatial weighted pressure covariance matrix, and determining the number of target pressure points according to a minimum description length criterion and an information gain ratio, includes: Performing an eigenvalue decomposition operation on the spatial weighted pressure covariance matrix to obtain an eigenvalue sequence; Calculating the sum of the eigenvalue sequence and normalizing each eigenvalue in the eigenvalue sequence to obtain a normalized eigenvalue sequence; Performing information entropy calculation based on the normalized eigenvalue sequence to obtain an information entropy index, and processing the information entropy index according to a minimum description length criterion to obtain a description length calculation sequence; The characteristic value sequence is processed according to the ratio of the cumulative sum to the total sum to obtain an information gain ratio sequence, and the description length calculation sequence is searched according to the minimum value principle to obtain the number of first pressure points; A judgment operation is performed on the information gain ratio sequence based on a preset threshold to obtain the second number of pressure points, and a weighted average operation is performed on the first number of pressure points and the second number of pressure points to obtain the target number of pressure points.

4. The multi-region coordinated control method of the high-precision pressure sensor array according to claim 3 is characterized in that: The step of constructing a pressure feature subspace based on the number of target pressure points and the spatial weighted pressure covariance matrix by singular value decomposition, and constructing a local perception domain according to the pressure feature subspace includes: Performing matrix decomposition on the spatial weighted pressure covariance matrix to obtain an orthogonal matrix, a singular value diagonal matrix and an orthogonal matrix, and performing column vector interception on the orthogonal matrix according to the number of target pressure points to obtain a basis vector matrix; Performing a product operation on the basis vector matrix and the transpose of the basis vector matrix, and subtracting the identity matrix to obtain a regularization term matrix; Constructing a pressure feature subspace based on the basis vector matrix and the regularization term matrix, and constructing a perception domain including adjacent sensor units within a 3×3 range for each sensor unit according to the pressure feature subspace to obtain a perception domain set; Normalizing the pressure characteristics of each sensing unit in the perception domain set to obtain a local feature vector set; The local feature vector set is combined according to the physical position information of the sensing unit to obtain a multi-scale feature tensor, and a spatial convolution operation is performed on the multi-scale feature tensor to obtain a local perception field.

5. The multi-region coordinated control method of the high-precision pressure sensor array according to claim 4 is characterized in that: The projecting the preprocessed pressure data set to the pressure feature subspace, performing weighted calculation on the sensor unit feature vectors in the perception domain through a hypernetwork structure, and generating a sensor unit state data set includes: Projecting the preprocessed pressure data set to the pressure feature subspace, extracting the projection coordinates of each sensing unit in the 32-dimensional feature space, and combining the relative position encoding relative to the center of the local perception domain to obtain initial feature data; Extracting pressure data from the initial feature data in the local perception domain, respectively calculating the real-time pressure data of each sensing unit and the pressure change rate at adjacent moments, and obtaining a sensing unit feature vector; Constructing a three-layer fully connected neural network according to the dimension of the sensor unit feature vector to obtain a super network structure, wherein the first hidden layer of the super network structure includes 64 neurons, the second hidden layer includes 32 neurons, and the output layer includes 32 neurons; Input the sensor unit feature vector into the super network structure, obtain the corresponding sensor unit hidden layer features through nonlinear activation function and full connection layer mapping transformation, and perform attention mechanism calculation based on the sensor unit hidden layer features to obtain the attention distribution coefficient; The attention distribution coefficient is normalized by a Softmax function to obtain a standardized attention weight, and the features of the sensor units in the local perception domain are weighted combined according to the standardized attention weight to generate a state representation vector for each sensor unit; The state representation vectors of all sensor units are organized and rearranged according to the array layout structure, and are spliced ​​and integrated to obtain a sensor unit state data set.

6. The multi-region coordinated control method of the high-precision pressure sensor array according to claim 5, characterized in that: The step of inputting the sensing unit state data set into a multi-layer attention encoder for processing, generating a pressure control gain matrix, and adjusting the first collaborative sampling parameter of the pressure sensor array to obtain a second collaborative sampling parameter includes: Performing time series expansion on the sensor unit state data set to obtain state sequence data; Input the state sequence data into the first layer attention module of the multi-layer attention encoder, calculate the query matrix Q, key matrix K and value matrix V with a dimension of 64 respectively through 8 attention heads, and scale them according to the product result of the query matrix Q and the key matrix K to obtain the first layer multi-head attention feature; The first layer of multi-head attention features are input into a two-layer feedforward neural network with a dimension of 256 for nonlinear transformation, and combined with the LayerNorm normalization layer and residual connection to obtain the first layer of encoded output features; According to the first layer encoding output features, a 3×3 local perception window is constructed for adjacent sensor units, the spatial attention weights between sensor units in each local perception window are calculated, and local feature aggregation is performed to obtain a spatial enhanced feature tensor; The spatial enhanced feature tensor is sequentially input into the second-layer attention module, the third-layer attention module and the fourth-layer attention module of the multi-layer attention encoder, and the output features of each layer are residually connected and LayerNorm normalized to obtain a multi-level encoding feature sequence, wherein the second-layer attention module, the third-layer attention module and the fourth-layer attention module have the same structure, and each layer of the attention module includes 8 attention heads and a 256-dimensional feedforward network; Adaptively weighting and fusing the multi-level coding feature sequences, and calculating the time series correlation weights based on the historical state information of the sensing unit to obtain a spatiotemporal fusion feature matrix; Inputting the spatiotemporal fusion feature matrix into a three-layer fully connected network for dimensionality reduction mapping to obtain a pressure control gain matrix, wherein the number of neurons in the three-layer fully connected network is 128, 64, and 32 respectively; Based on the gain coefficient corresponding to each sensing unit in the pressure control gain matrix, the collaborative sampling frequency and the collaborative sampling time window in the first collaborative sampling parameter are weightedly adjusted to obtain the second collaborative sampling parameter.

7. A multi-region collaborative control device for a high-precision pressure sensing array, characterized in that: The method for implementing the multi-region coordinated control method of the high-precision pressure sensing array according to any one of claims 1 to 6, wherein the multi-region coordinated control device of the high-precision pressure sensing array comprises: an acquisition module, configured to acquire data from a plurality of sensing units in the pressure sensor array based on a first collaborative sampling parameter to obtain pressure data, and to perform wavelet transform noise reduction processing on the pressure data to obtain a preprocessed pressure data set and a sensing unit spatial position matrix; A calculation module, used to calculate the pressure signal covariance of the sensing unit according to the preprocessed pressure data set, and construct a spatial weighted pressure covariance matrix; A decomposition module, used for performing eigenvalue decomposition and information entropy calculation on the spatial weighted pressure covariance matrix, and determining the number of target pressure points according to a minimum description length criterion and an information gain ratio; A construction module, configured to construct a pressure feature subspace through singular value decomposition based on the number of target pressure points and the spatial weighted pressure covariance matrix, and to construct a local perception domain according to the pressure feature subspace; A projection module, used to project the preprocessed pressure data set to the pressure feature subspace, perform weighted calculation on the sensor unit feature vectors in the perception domain through a hypernetwork structure, and generate a sensor unit state data set; The adjustment module is used to input the sensor unit state data set into the multi-layer attention encoder for processing, generate a pressure control gain matrix, and adjust the first collaborative sampling parameter of the pressure sensor array to obtain a second collaborative sampling parameter.

8. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the multi-region collaborative control method of the high-precision pressure sensing array according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the multi-region collaborative control method of a high-precision pressure sensing array according to any one of claims 1 to 6 are implemented.

Images