Sensor array strong coupling positioning and suppression method based on attention mechanism

Abstract

The application discloses a sensor array strong coupling positioning and suppression method based on an attention mechanism, which comprises the following steps: collecting the amplitude, time sequence, stability characteristics and array topology characteristics of a sensor output signal, constructing a 9-dimensional standardized input matrix through cross-domain normalization; building a neural network that fuses graph convolution and multi-head attention, mining unit coupling correlation, and outputting a strong coupling positioning result and a coupling heat map; through a hardware scanning parameter mapping module, the heat map is mapped to hardware control parameters of scanning time sequence, isolation circuit and driving amplitude, so that crosstalk targeted suppression is realized; an LMS adaptive filter is used to compensate residual crosstalk, and an error is fed back in real time to form a closed-loop optimization. The application realizes intelligent suppression of the whole process of crosstalk, the crosstalk suppression rate in the whole temperature range is greater than or equal to 93%, the scanning efficiency is improved by 66.7% compared with traditional methods, and is suitable for high-precision detection scenes of various high-density flat plate sensor arrays.

Description

Technical Field

[0001] The present invention relates to the technical field of sensor detection and intelligent signal processing, and particularly to a method for strong coupling localization and suppression of a sensor array based on an attention mechanism. Background Art

[0002] With the continuous improvement of the demand for high-precision sensing in fields such as medical monitoring and industrial detection, flat panel sensor arrays have become the core devices for achieving refined detection due to their high unit density and excellent spatial resolution. However, the dense arrangement of array units leads to prominent strong interference problems such as parasitic capacitance coupling and electric field leakage crosstalk between units. Especially in high-frequency scanning scenarios, the amplitude of crosstalk signals can reach 15%-20% of the useful signals, seriously affecting the detection accuracy and becoming the key bottleneck restricting the application expansion of flat panel sensor arrays. Traditional strong coupling suppression technologies are limited to single hardware optimization or simple algorithm compensation, and there are three core bottlenecks: Insufficient positioning accuracy: Conventional hardware suppression does not consider the unit topology differences and can only roughly suppress global crosstalk. It is unable to accurately locate strong coupling unit pairs, has a poor suppression effect on local dense coupling regions, and the crosstalk residue rate still reaches 8%-10%. Poor dynamic adaptability: Traditional algorithm compensation, such as fixed coefficient filtering, is mostly targeted at static crosstalk. In the face of time-varying scenarios such as temperature drift and scanning frequency changes, the compensation coefficient cannot be adjusted in real time, resulting in a suppression effect attenuation of more than 50%. Weak system coordination: Hardware suppression and algorithm compensation are independent of each other. Hardware optimization does not combine the coupling localization results, and algorithm compensation does not feedback to guide the adjustment of hardware parameters, resulting in a disconnection between suppression and compensation, unable to achieve full-process crosstalk control, and easily leading to a decrease in scanning efficiency due to excessive suppression.

[0003] In the prior art, although the graph convolutional network (GCN) can mine the unit topology associations, it does not combine the attention mechanism to focus on strong coupling regions, resulting in low positioning efficiency; single LMS adaptive filtering can compensate for residual crosstalk, but lacks the联动优化 with hardware scanning parameters and is difficult to cope with complex time-varying coupling scenarios; traditional hardware isolation circuits are not dynamically controlled based on the coupling strength, resulting in energy waste and inaccurate suppression problems. How to construct a complete technical chain of "coupling precise positioning - hardware dynamic suppression - algorithm real-time compensation - system closed-loop optimization" to achieve full-process intelligent control of crosstalk from positioning to suppression, while balancing detection accuracy and scanning efficiency, has become the core research challenge in the field of strong coupling suppression of flat panel sensor arrays. Summary of the Invention

[0004] It should be noted that there is an unclear expression "联动优化" in the original text. I have translated it as "联动优化" as it is, and you may need to further clarify its specific meaning according to the actual situation.To address the aforementioned issues, this invention proposes a sensor array strong coupling localization and suppression method based on an attention mechanism. By combining graph convolutional networks (GCN), multi-head attention mechanisms, and LMS adaptive filtering technology, it achieves in-depth mining of sensor unit coupling correlations and accurate localization of strong coupling regions. Simultaneously, through scanning timing optimization, dynamic isolation circuit deployment, and adaptive adjustment of driving parameters, it completes the entire process control of strong coupling crosstalk from localization to hardware suppression and residual compensation.

[0005] To achieve the above objectives, the technical solution of the present invention is implemented as follows: A method for localization and suppression of strong coupling in sensor arrays based on an attention mechanism includes the following steps: Step S1: Collect the output signals of each unit of the flat panel sensor array, extract the three-dimensional signal statistical features of amplitude, time series and stability, as well as the three-dimensional topological space features of coordinates, spacing and type, perform cross-domain normalization processing on the two types of features, and construct and output the normalized multi-dimensional input matrix. Step S2: Using the normalized multi-dimensional input matrix output in Step S1 as the only input, build a neural network that integrates graph convolution topological feature extraction and multi-head attention weight calculation, mine the coupling relationship between units and quantify the coupling strength, and output the localization results and coupling heatmap of strongly coupled unit pairs. Step S3: Taking the coupled heat map and strong coupling positioning result output in step S2 as input, the preset attention weight-scan parameter mapping rule is executed through the hardware scan parameter mapping module, and the scan timing optimization parameters, dynamic isolation circuit control signal, and drive signal amplitude adaptive parameters are output in sequence to complete the dynamic optimization of the hardware scan strategy and output the hardware scan control parameter set. Step S4: Using the hardware scan control parameter set output in step S3 as a reference, an LMS adaptive filter is used to generate a reverse compensation signal for the residual crosstalk after hardware suppression, and the filter error signal is output. At the same time, the filter error signal is fed back to the neural network in step S2 and the hardware scan parameter mapping module in step S3 in real time, and the network parameters and hardware scan mapping rules are iteratively optimized to form a closed loop of localization-suppression-compensation.

[0006] Furthermore, in step S1, the extraction of the three-dimensional signal statistical features of amplitude, timing, and stability specifically involves: To analyze the dynamic output characteristics of each unit in the flat panel sensor array, a full array signal feature matrix is ​​constructed. ,in This represents the total number of array elements. The number of rows in the array. The number of columns in the array; The amplitude feature The Gaussian-weighted mean of the voltage sequence within the scan period is calculated using the following formula:

[0007] in, Scan cycle The output voltage sequence within, , , The sampling interval is... ; Number of sampling points, weighting coefficient , These are Gaussian weighting coefficients. ; The timing feature is the rise time feature. The linear fitting correction method is used for calculation, and the formula is:

[0008] in, This represents the peak amplitude of the voltage sequence. The fitting coefficients are for the first half of the rising edge. The fitting coefficients are for the latter half of the rising edge. For the voltage to reach 0.1 At that moment, For the voltage to reach 0.9 The moment; The stability feature is: Sliding window variance quantization is used, and the formula is:

[0009] Among them, window length , For the first The average voltage of each sliding window.

[0010] Furthermore, in step S1, the extraction of the coordinates, spacing, and type of the three-dimensional topological spatial features specifically involves: Based on the physical layout of the flat panel sensor array, a full array topology feature matrix is ​​constructed. ,in This represents the total number of array elements. The number of rows in the array. The number of columns in the array; The coordinate feature is the planar coordinate of the element in the array's global coordinate system. ,in The unit number is used; the spacing feature is used for the unit. With unit Weighted Euclidean distance The formula is:

[0011] Among them, the weighting coefficient , , , For unit row and column indexes, , For unit The row and column indices are used to amplify the contribution of spacing between row and column units, thus conforming to the physical law that crosstalk increases with spatial span. The type feature is a unit position type encoding. Multi-class coding is used for differentiation, and the formula is:

[0012] And through one-hot encoding Convert to vector Therefore, the topological feature vector of a single sensor unit is , For unit The average spacing, and the topological feature matrix of the entire array are The coordinates are 2-dimensional, the average spacing is 1-dimensional, and the one-hot encoding is 3-dimensional.

[0013] Furthermore, in step S1, the cross-domain normalization processing and the construction of the normalized multi-dimensional input matrix are specifically as follows: The signal feature matrix is ​​normalized using the Gaussian distribution mapping method, and the formula is:

[0014] in, for The Middle Line 1 Column elements, These correspond to amplitude, timing, and stability characteristics, respectively. The characteristic mean, The Sigmoid function ensures that the normalized values ​​fall within [0,1] and preserve the characteristic distribution trend, representing the standard deviation of the feature. Differential normalization is applied to the topological feature matrix: coordinate features are linearly normalized, the average spacing is inversely normalized, and one-hot encoding retains the original values; The normalized signal feature matrix With topological feature matrix By concatenating the columns, we obtain the final normalized multi-dimensional input matrix: .

[0015] Furthermore, in step S2, the flat panel sensor array is first abstracted as an undirected weighted topological graph G=(V,E,W), where the node set V corresponds to all array units, the edge existence threshold of the edge set E is 1.5 times the average spacing between array units, and the element calculation formula of the weight matrix W is:

[0016] in, This is the attenuation coefficient, ensuring that the weights decay rapidly as the spacing increases. For unit and The covariance of the output voltage, The Pearson correlation coefficient between the output voltages of the two units is used to quantify the degree of signal coupling, with a value range of []. [1,1], taking the absolute value reflects the coupling strength. The greater the weight, the stronger the coupling potential between the two units; Furthermore, in step S2, the graph convolutional topology feature extraction employs a three-layer graph convolutional network (GCN). The formula for calculating layer convolution is:

[0017] in, , For the first Layer learning weights For the first Layer bias, Activation function In terms of dimensional design, it was gradually increased from the initial 9 dimensions to 128 dimensions, and the final output... The carrying unit's globally coupled information; The multi-head attention weight calculation adopts a 4-head attention structure, and the single-head attention calculation formula is as follows:

[0018] Where Q, K, and V are linear mapping matrices of the high-dimensional topological feature matrix H. M represents the feature dimension of a single-head attention function, and M is a mask matrix used to mask uncoupled pairs of cells. The multi-head attention fusion output formula is:

[0019] in, For column concatenation symbols, The output weight matrix for multi-head attention; The localization results and coupling heatmap of the strongly coupled output unit pairs are specifically calculated as follows: The coupling strength of the unit pairs is calculated by fusing multi-head attention weights, signal differences, unit spacing, and unit type, using the following formula:

[0020] in, , These are the standard deviations of the output voltages of units p and q, respectively. , These are the type-specific one-hot encoded vectors for the two units; 0.5 is the weight coefficient for the spatial distance term, 0.3 is the weight coefficient for the signal stability difference term, and 0.2 is the weight coefficient for the embedding feature inner product term. A threshold is set. ,in Given the average strength of strongly coupled samples, when When the element pair (p, q) is determined to be a strongly coupled element pair, the coupling strength of the entire array is mapped to a coupling heatmap to complete the spatial positioning of the strongly coupled region.

[0021] Furthermore, in step S3, the first output of the hardware scan parameter mapping module is the scan timing optimization parameter, specifically implemented as follows: The array elements are divided into three levels based on coupling strength: strong coupling, medium coupling, and low coupling, corresponding to activation priorities from low to high. The activation interval of the elements is dynamically adjusted based on the average coupling strength of the elements, using the following formula:

[0022] in Based on the activation interval, For one of the units The average coupling strength, where For unit The number of adjacent units, To maximize the coupling strength of the entire array, ensure an extended activation interval for strongly coupled units, and maintain the basic interval for low-coupled units, the scan path planning adopts a "low-coupling region priority" approach. First, low-coupled units in the central region are activated, followed by medium and strongly coupled units. Furthermore, strongly coupled units are activated using "interval activation." Then skip adjacent strongly coupled units Activate directly This avoids the continuous activation of strongly coupled unit pairs.

[0023] The second output of the hardware scan parameter mapping module is a dynamic isolation circuit control signal, specifically implemented as follows: The driver and sampling terminals of each sensor unit are connected in series with an isolated MOSFET. The switching signal is generated by the FPGA controller based on the coupling strength output from the mapping module; an isolation threshold is set. The control logic formula is:

[0024] in, This is the MOSFET's turn-on voltage. This is the turn-off voltage of the MOSFET; after conduction, the parasitic capacitance between strongly coupled units is short-circuited to ground, enabling multi-unit collaborative isolation control of the strongly coupled unit cluster.

[0025] The third output of the hardware scanning parameter mapping module is the adaptive parameter for the amplitude of the driving signal, specifically implemented as follows: Establish a linear mapping model between coupling strength and driving amplitude, with the following formula:

[0026] in, The base driving amplitude is set, with an amplitude attenuation coefficient of 0.6. This ensures that the driving amplitude of the strongly coupled units is reduced, decreasing the electric field radiation range, while the weakly coupled units maintain the base amplitude. Amplitude output is achieved through a 12-bit DAC, achieving a resolution of [missing information]. This ensures that the amplitude adjustment accuracy matches the changes in coupling strength, thereby achieving amplitude accuracy control.

[0027] Furthermore, in step S4, an LMS adaptive filter is used to generate a reverse compensation signal for the residual crosstalk after hardware suppression, specifically as follows: The residual crosstalk is canceled by an LMS adaptive filter circuit optimized for filtering structure, with the hardware scan clock signal as the reference input. The sensor output signal is the desired input. Filter output To compensate for the signal, the filter coefficients are updated in real time using the minimum mean square error criterion, as shown in the following formula:

[0028] in for Time-based filter coefficients Step size factor As an error signal, ensure that the compensation signal and the residual crosstalk have equal amplitude and opposite phase, and set a residual crosstalk threshold. ,when The coefficient update is stopped at a certain time to balance the compensation accuracy and the amount of calculation.

[0029] Furthermore, in step S4, the LMS adaptive filter adopts a 32nd-order tap structure, and the filter output formula is:

[0030] in, Let be the weight of the k-th tap at time n. The hardware scan clock signal is used as a reference input. For delayed reference signal; The formulas for coefficient updates and convergence determination are as follows:

[0031] in, This is an error signal that reflects residual crosstalk. The desired signal output by the sensor; The convergence determination needs to avoid the randomness of a single-cycle error, and adopts the judgment logic of "the mean square value of the error meets the standard for 10 consecutive sampling cycles", the formula is as follows:

[0032] In the event of a sudden strong interference, the system will automatically increase the step size factor to 0.05 temporarily to accelerate the convergence of the coefficients. Once the error returns to within the threshold, the step size will be restored to 0.01 to ensure the effectiveness of compensation in extreme scenarios.

[0033] Furthermore, in step S4, the iterative optimization specifically involves updating the neural network parameters from step S2 using the mean square error (MSE) as the loss function, through backpropagation, as shown in the formula:

[0034] in, For the parameter set of the neural network, The gradient of the loss function with respect to the parameters is used; based on the error space distribution, the mapping rules of the hardware scan parameter mapping module in step S3 are updated to dynamically adapt to the hardware parameters. The formula for adjusting the scan interval is:

[0035] The formula for adjusting the drive amplitude is: .

[0036] Beneficial effects: (1) Significantly improved positioning accuracy and highly targeted: Through the fusion architecture of graph convolution and multi-head attention, the precise positioning of strongly coupled unit pairs is achieved. The resolution of strongly coupled region recognition is more than twice that of traditional methods, which completely solves the problem of "blind suppression" in traditional schemes. Experiments show that the coupling value of strongly coupled regions is reduced by 80%, medium coupled regions by 60%, and weak coupled regions by only 30%, achieving precise hierarchical suppression. (2) Deep hardware and software collaboration and excellent suppression effect: Through the hardware scanning parameter mapping module, the end-to-end mapping from algorithm positioning results to hardware suppression actions is achieved. The crosstalk suppression rate in the whole temperature range is stable at more than 93%, and the suppression rate reaches 95% at room temperature of 20℃. Temperature fluctuations only cause the suppression rate to decrease by 2.1%, which is far better than traditional schemes. (3) Significantly improved scanning efficiency, balancing accuracy and speed: By adopting a hierarchical activation and path optimization strategy based on coupling strength, the array scanning efficiency is improved by 66.7% compared with the traditional time-division multiplexing scanning method, avoiding the efficiency loss caused by global over-suppression in traditional schemes. (4) Strong adaptability and good long-term operational stability: Through closed-loop feedback iterative optimization of mapping rules and network parameters, it can adapt to time-varying scenarios such as temperature drift, scanning frequency changes, and device aging. When encountering sudden strong interference, it can automatically adjust the filtering step size to accelerate convergence, and has excellent robustness. (5) Wide range of applications and strong compatibility: It can be adapted to flat panel sensor arrays of different densities from 32×32 to 256×256, and is compatible with various types of flat panel sensors such as capacitive, resistive, and piezoelectric sensors. It can be widely used in multiple scenarios such as medical imaging, industrial non-destructive testing, and touch screen sensing. Attached Figure Description

[0037] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating the steps of the sensor array strong coupling localization and suppression method based on attention mechanism described in an embodiment of the present invention. Figure 2 This is a baseline distribution diagram of the array coupling strength before processing by the method of the present invention; Figure 3 This is a direct comparison diagram of the suppression effects of the method of the present invention and traditional existing technologies; Figure 4 This is a comparison chart of key performance indicators between the method of this invention and traditional existing technologies. Detailed Implementation

[0038] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0039] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0040] Example 1 See Figure 1 A method for localization and suppression of strong coupling in sensor arrays based on attention mechanisms, comprising the following steps: Step S1: Collect the output signals of each unit of the flat panel sensor array, extract the three-dimensional signal statistical features of amplitude, time series and stability, as well as the three-dimensional topological space features of coordinates, spacing and type, perform cross-domain normalization processing on the two types of features, and construct and output the normalized multi-dimensional input matrix. Step S2: Using the normalized multi-dimensional input matrix output in Step S1 as the only input, build a neural network that integrates graph convolution topological feature extraction and multi-head attention weight calculation, mine the coupling relationship between units and quantify the coupling strength, and output the localization results and coupling heatmap of strongly coupled unit pairs. Step S3: Taking the coupled heat map and strong coupling positioning result output in step S2 as input, the preset attention weight-scan parameter mapping rule is executed through the hardware scan parameter mapping module, and the scan timing optimization parameters, dynamic isolation circuit control signal, and drive signal amplitude adaptive parameters are output in sequence to complete the dynamic optimization of the hardware scan strategy and output the hardware scan control parameter set. Step S4: Using the hardware scan control parameter set output in step S3 as a reference, an LMS adaptive filter is used to generate a reverse compensation signal for the residual crosstalk after hardware suppression, and the filter error signal is output. At the same time, the filter error signal is fed back to the neural network in step S2 and the hardware scan parameter mapping module in step S3 in real time, and the network parameters and hardware scan mapping rules are iteratively optimized to form a closed loop of localization-suppression-compensation.

[0041] The first step of this embodiment is to complete the standardized representation of the original signal and array physical information through multi-dimensional feature acquisition and normalization, providing compliant input for subsequent algorithms. The second step uses the standardized features as the sole input and achieves accurate positioning and coupling strength quantification of strongly coupled units through graph convolution + multi-head attention fusion neural network, outputting the core basis for targeted suppression. The third step uses the hardware scanning parameter mapping module to directly convert the coupling positioning results output by the algorithm into hardware-executable control parameters, realizing end-to-end linkage between the algorithm and hardware. The fourth step uses LMS adaptive filtering to cancel the residual crosstalk after hardware suppression, and at the same time feeds the error signal back to the front-end neural network and hardware mapping module to iteratively optimize the parameters of the entire link and form a self-optimizing closed loop. This embodiment achieves deep collaboration between algorithm positioning and hardware suppression, breaking the bottleneck of software and hardware disconnect in traditional solutions. While ensuring the crosstalk suppression effect, it avoids the scanning efficiency loss caused by global over-suppression. The closed-loop feedback architecture can adaptively adapt to time-varying scenarios such as temperature drift, device aging, and environmental changes, and the long-term operational stability of the system is greatly improved.

[0042] In a specific example, step S1, the extraction of the three-dimensional signal statistical features of amplitude, timing, and stability, specifically involves: To analyze the dynamic output characteristics of each unit in the flat panel sensor array, a full array signal feature matrix is ​​constructed. ,in This represents the total number of array elements. The number of rows in the array. The number of columns in the array; The amplitude feature The Gaussian-weighted mean of the voltage sequence within the scan period is calculated using the following formula:

[0043] in, Scan cycle The output voltage sequence within, , , The sampling interval is... ; Number of sampling points, weighting coefficient , These are Gaussian weighting coefficients. ; The timing feature is the rise time feature. The linear fitting correction method is used for calculation, and the formula is:

[0044] in, This represents the peak amplitude of the voltage sequence. The fitting coefficients are for the first half of the rising edge. The fitting coefficients are for the latter half of the rising edge. For the voltage to reach 0.1 At that moment, For the voltage to reach 0.9 The moment; It should be noted that the rise time in this embodiment is defined as the time interval between the output voltage rising from 10% peak amplitude to 90% peak amplitude. Considering noise interference, a linear fitting correction method is used for calculation: First, the voltage sequence... Differentiating gives By searching maximum point Determine the peak amplitude Secondly, respectively in Interval fitting voltage curve and satisfy ,exist Interval fitting and satisfy .

[0045] The stability feature is: Sliding window variance quantization is used, and the formula is:

[0046] Among them, window length , For the first The average voltage of each sliding window.

[0047] This embodiment focuses on the dynamic output characteristics of each unit in the flat panel sensor array, achieving precise quantization of signal features from three core dimensions: Amplitude features are calculated using a Gaussian weighted mean, highlighting the stable output in the middle segment of the voltage sequence and suppressing interference from edge sampling noise; Timing features (rise time) are calculated using a linear fitting correction method, locking the voltage peak point through differentiation, fitting the rise time curve piecewise, and accurately calculating the rise time in the 10%~90% peak range, avoiding timing calculation deviations caused by noise; Stability features are quantized using sliding window variance, calculating local voltage fluctuations through a fixed-length window, and finally characterizing the signal stability with the global normalized variance, accurately reflecting the impact of coupling interference on signal stability. Finally, a 3D signal feature matrix for the entire array is constructed to comprehensively characterize the multi-dimensional impact of coupling interference on the sensor output signal.

[0048] In a specific example, step S1, the extraction of the coordinates, spacing, and type of the three-dimensional topological spatial features, specifically involves: Based on the physical layout of the flat panel sensor array, a full array topology feature matrix is ​​constructed. ,in This represents the total number of array elements. The number of rows in the array. The number of columns in the array; The coordinate feature is the planar coordinate of the element in the array's global coordinate system. ,in The unit number is used; the spacing feature is used for the unit. With unit Weighted Euclidean distance The formula is:

[0049] Among them, the weighting coefficient , , , For unit row and column indexes, , For unit The row and column indices are used to amplify the contribution of spacing between row and column units, thus conforming to the physical law that crosstalk increases with spatial span. It should be noted that the global coordinate system of the array is established with the top-left corner cell of the array as the origin. Horizontal to the right is The axis, vertically downwards is Shaft, assuming the physical dimensions of the unit are (Length × Width), then the coordinates of the p-th unit are row index Column index , , The coordinate offset caused by the production process is pre-calibrated using a laser rangefinder and then substituted into the model.

[0050] The type feature is a unit position type encoding. Multi-class coding is used for differentiation, and the formula is:

[0051] And through one-hot encoding Convert to vector For example, corner units correspond Therefore, the topological feature vector of a single sensor unit is , For unit The average spacing, and the topological feature matrix of the entire array are The coordinates are 2-dimensional, the average spacing is 1-dimensional, and the one-hot encoding is 3-dimensional.

[0052] This embodiment, based on the physical layout of a flat panel sensor array, accurately characterizes the coupling potential from a spatial dimension: establishing a global coordinate system for the array, calibrating the coordinate offset caused by the manufacturing process, and accurately characterizing the physical position of each unit; using weighted Euclidean distance to quantize the unit spacing, and amplifying the contribution of the spacing between units across rows and columns through weight coefficients, which perfectly matches the physical law that crosstalk increases with spatial span; using multi-class coding + one-hot coding to distinguish between three types of units: center, edge, and corner, accurately characterizing the inherent coupling characteristics differences of units at different positions, and finally constructing a 6-dimensional topological feature matrix for the entire array.

[0053] In a specific example, step S1, the cross-domain normalization processing and the construction of the normalized multi-dimensional input matrix, specifically involves: The signal feature matrix is ​​normalized using the Gaussian distribution mapping method, and the formula is:

[0054] in, for The Middle Line 1 Column elements, These correspond to amplitude, timing, and stability characteristics, respectively. The characteristic mean, The Sigmoid function ensures that the normalized values ​​fall within [0,1] and preserve the characteristic distribution trend, representing the standard deviation of the feature. Differential normalization is applied to the topological feature matrix: coordinate features are linearly normalized, the average spacing is inversely normalized, and one-hot encoding retains the original values; The normalized signal feature matrix With topological feature matrix By concatenating the columns, we obtain the final normalized multi-dimensional input matrix: .

[0055] It should be noted that differentiated normalization is used for three different topological feature types: coordinate features, average spacing, and one-hot encoding. For coordinate features... Linear normalization to [0,1] is given by the following formula: , , , The minimum and maximum values ​​of the x-coordinate of the entire array. , The minimum and maximum values ​​of the y-coordinates of the entire array. They are respectively Normalized values; for the average spacing Using reciprocal normalization, the formula is as follows: This results in units with smaller spacing and stronger coupling corresponding to larger normalized values, which aligns with the physical meaning of subsequent attention weights; for one-hot encoding The original binary encoding is preserved, and no additional normalization is required. Each row corresponds to the 9-dimensional normalized features of a single sensor unit, providing standardized input for subsequent topology association mining and attention weight calculation.

[0056] This embodiment addresses the dimensional differences (voltage vs. millimeters) between signal features and topological features by employing a differentiated normalization strategy to standardize the features: The signal feature matrix is ​​normalized using a Gaussian distribution mapping and a Sigmoid function, eliminating the dimensional differences while fully preserving the original distribution trend of the signal features; the topological feature matrix is ​​normalized using differentiated normalization: linear normalization of coordinate features, inverse normalization of the average spacing (the smaller the spacing, the larger the normalized value, aligning with the physical meaning of coupling strength), and one-hot encoding to maintain the original values, maximizing the preservation of the physical meaning of the topological features; the normalized 3D signal feature matrix and the 6D topological feature matrix are then concatenated column-wise to obtain a 9D standardized input matrix that meets the input requirements of the neural network.

[0057] In a specific example, in step S2, the flat panel sensor array is first abstracted as an undirected weighted topological graph G=(V,E,W), where the node set V corresponds to all array cells, the edge existence threshold of the edge set E is 1.5 times the average spacing between array cells, and the element calculation formula of the weight matrix W is:

[0058] in, This is the attenuation coefficient, ensuring that the weights decay rapidly as the spacing increases. For unit and The covariance of the output voltage, The Pearson correlation coefficient between the output voltages of the two units is used to quantify the degree of signal coupling, with a value range of []. [1,1], taking the absolute value reflects the coupling strength. The greater the weight, the stronger the coupling potential between the two units; It should be noted that the sensor array topology is assumed to be an undirected weighted graph. . For a set of sensor nodes, This represents the total number of array elements. Corresponding to the There are 1 sensor unit, and the feature vector of each node is the input matrix. The Okay, that is Edge set , To establish a threshold for edge existence, we take 1.5 times the average spacing between array cells, i.e. Weight matrix ,element Corresponding edges The weight.

[0059] This embodiment abstracts the flat panel sensor array into an undirected weighted topological graph structure that can be processed by a graph neural network: each sensor unit is a graph node, and the node features correspond to a 9-dimensional standardized input vector; with 1.5 times the average spacing between array units as a threshold, edge connections are established only for unit pairs with spacing within the threshold, shielding distant units with no coupling potential and reducing invalid computation; the edge weights simultaneously integrate the physical spacing attenuation factor and the signal Pearson correlation coefficient, quantifying the coupling potential between units in a dual-dimensional way, with a larger weight representing a stronger coupling potential between unit pairs, ultimately constructing the weight matrix of the topological graph.

[0060] In a specific example, in step S2, the graph convolutional topological feature extraction employs a three-layer graph convolutional network (GCN). The formula for calculating layer convolution is:

[0061] in, , For the first Layer learning weights For the first Layer bias, Activation function In terms of dimensional design, it was gradually increased from the initial 9 dimensions to 128 dimensions, and the final output... The carrying unit's globally coupled information; The multi-head attention weight calculation adopts a 4-head attention structure, and the single-head attention calculation formula is as follows:

[0062] Where Q, K, and V are linear mapping matrices of the high-dimensional topological feature matrix H. M represents the feature dimension of a single-head attention function, and M is a mask matrix used to mask uncoupled pairs of cells. The multi-head attention fusion output formula is:

[0063] in, For column concatenation symbols, The output weight matrix for multi-head attention; The localization results and coupling heatmap of the strongly coupled output unit pairs are specifically calculated as follows: The coupling strength of the unit pairs is calculated by fusing multi-head attention weights, signal differences, unit spacing, and unit type, using the following formula:

[0064] in, , These are the standard deviations of the output voltages of units p and q, respectively. , These are the type-specific one-hot encoded vectors for the two units; 0.5 is the weight coefficient for the spatial distance term, 0.3 is the weight coefficient for the signal stability difference term, and 0.2 is the weight coefficient for the embedding feature inner product term. A threshold is set. ,in Given the average strength of strongly coupled samples, when When the element pair (p, q) is determined to be a strongly coupled element pair, the coupling strength of the entire array is mapped to a coupling heatmap to complete the spatial positioning of the strongly coupled region.

[0065] This embodiment uses a topological graph model as a basis to achieve accurate identification and spatial positioning of strongly coupled unit pairs: By fusing node features with those of its neighbors using a three-layer graph convolutional network (GCN), the initial 9-dimensional features are progressively upgraded to 128 dimensions to deeply explore the global topological associations and coupling potential of array units. A 4-head attention mechanism is employed to dynamically allocate coupling weights from multiple perspectives. Uncoupled unit pairs are masked using a mask matrix to achieve automatic focusing on strongly coupled unit pairs, outputting multi-head attention fusion results. Multi-dimensional information such as attention weights, signal differences, unit spacing, and unit type is fused to quantify the coupling strength of unit pairs. Strongly coupled unit pairs are filtered through a preset threshold, and finally, the coupling strength of the entire array is mapped to a coupling heatmap to complete the spatial localization of strongly coupled regions.

[0066] In a specific example, in step S3, the first output of the hardware scan parameter mapping module is the scan timing optimization parameter, specifically implemented as follows: The array elements are divided into three levels based on coupling strength: strong coupling, medium coupling, and low coupling, corresponding to activation priorities from low to high. The activation interval of the elements is dynamically adjusted based on the average coupling strength of the elements, using the following formula:

[0067] in Based on the activation interval, For one of the units The average coupling strength, where For unit The number of adjacent units, To maximize the coupling strength of the entire array, ensure an extended activation interval for strongly coupled units, and maintain the basic interval for low-coupled units, the scan path planning adopts a "low-coupling region priority" approach. First, low-coupled units in the central region are activated, followed by medium and strongly coupled units. Furthermore, strongly coupled units are activated using "interval activation." Then skip adjacent strongly coupled units Activate directly This avoids the continuous activation of strongly coupled unit pairs.

[0068] In this embodiment, the units are classified according to coupling strength. Divided into three levels: strongly coupled units ( ), and intermediate coupling unit ( Low-coupling unit ( The activation priorities are arranged from low to high, and the activation timing is adjusted according to the unit coupling strength to avoid simultaneous activation of strongly coupled units and reduce signal superposition interference.

[0069] The second output of the hardware scan parameter mapping module is a dynamic isolation circuit control signal, specifically implemented as follows: The driver and sampling terminals of each sensor unit are connected in series with an isolated MOSFET. The switching signal is generated by the FPGA controller based on the coupling strength output from the mapping module; an isolation threshold is set. The control logic formula is:

[0070] in, This is the MOSFET's turn-on voltage. This is the turn-off voltage of the MOSFET; after conduction, the parasitic capacitance between strongly coupled units is short-circuited to ground, enabling multi-unit collaborative isolation control of the strongly coupled unit cluster.

[0071] This embodiment uses a coupling thermal map to control the switching of isolated MOSFETs, short-circuiting the parasitic capacitance between strongly coupled units to cut off the coupling path. An isolated MOSFET is connected in series between the drive and sampling terminals of each unit. The switching signal is generated by the FPGA controller based on the coupling strength. A coupling strength threshold is set. When unit and of When the FPGA outputs a high level, it turns on the isolation MOSFET. After turning on, the parasitic capacitance is short-circuited to ground, which can eliminate capacitive coupling. For strongly coupled unit clusters, such as four corner units, multi-unit collaborative control can be implemented by synchronously turning on the surrounding isolation MOSFETs to form isolation and block internal and external coupling within the cluster.

[0072] The third output of the hardware scanning parameter mapping module is the adaptive parameter for the amplitude of the driving signal, specifically implemented as follows: Establish a linear mapping model between coupling strength and driving amplitude, with the following formula:

[0073] in, The base driving amplitude is set, with an amplitude attenuation coefficient of 0.6. This ensures that the driving amplitude of the strongly coupled units is reduced, decreasing the electric field radiation range, while the weakly coupled units maintain the base amplitude. Amplitude output is achieved through a 12-bit DAC, achieving a resolution of [missing information]. This ensures that the amplitude adjustment accuracy matches the changes in coupling strength, thereby achieving amplitude accuracy control.

[0074] This embodiment uses a hardware scanning parameter mapping module to convert the coupling positioning results into three-way coordinated hardware control parameters to achieve targeted crosstalk suppression. First path: Dynamic optimization of scanning timing: The cells are divided into three levels: strong / medium / weak according to the coupling strength, and the activation priority is set from low to high accordingly. The cell activation interval is dynamically adjusted based on the coupling strength. The scanning path of "low coupling region priority and strong coupling cell interval activation" is adopted to avoid superimposed interference caused by strong coupling cells working at the same time. Second path: Dynamic isolation circuit control: Each unit is connected in series with an isolation MOSFET. The MOSFET is switched by the FPGA controller with the coupling strength as the threshold. When the MOSFET is turned on, the parasitic capacitance between strongly coupled units is short-circuited, and the capacitive coupling path is directly cut off, so as to achieve multi-unit collaborative isolation for strongly coupled clusters. Third channel: Adaptive adjustment of drive amplitude: A linear mapping model between coupling strength and drive amplitude is established. Strong coupling unit reduces drive amplitude and reduces cross coupling caused by electric field leakage. High-precision adjustment of amplitude is achieved through 12-bit DAC.

[0075] In a specific example, in step S4, an LMS adaptive filter is used to generate a reverse compensation signal for the residual crosstalk after hardware suppression, specifically as follows: The residual crosstalk is canceled by an LMS adaptive filter circuit optimized for filtering structure, with the hardware scan clock signal as the reference input. The sensor output signal is the desired input. Filter output To compensate for the signal, the filter coefficients are updated in real time using the minimum mean square error criterion, as shown in the following formula:

[0076] in for Time-based filter coefficients Step size factor As an error signal, ensure that the compensation signal and the residual crosstalk have equal amplitude and opposite phase, and set a residual crosstalk threshold. ,when The coefficient update is stopped at a certain time to balance the compensation accuracy and the amount of calculation.

[0077] This embodiment addresses the 0.5%-2% residual crosstalk remaining after hardware suppression by using an LMS adaptive filter for precise cancellation: the hardware scan clock signal (the core source of crosstalk) is used as the filter reference input, and the sensor output signal with residual crosstalk is used as the desired input; the filter weight coefficients are updated in real time using the minimum mean square error criterion to generate a compensation signal with the same amplitude but opposite phase as the residual crosstalk, thus achieving precise cancellation of the residual crosstalk; a residual crosstalk threshold is set, and coefficient updates are stopped when the error signal is below the threshold, balancing compensation accuracy and computational load.

[0078] In a specific example, in step S4, the LMS adaptive filter adopts a 32nd-order tap structure, and the filter output formula is:

[0079] in, Let be the weight of the k-th tap at time n. The hardware scan clock signal is used as a reference input. For delayed reference signal; The formulas for coefficient updates and convergence determination are as follows:

[0080] in, This is an error signal that reflects residual crosstalk. The desired signal output by the sensor; The convergence determination needs to avoid the randomness of a single-cycle error, and adopts the judgment logic of "the mean square value of the error meets the standard for 10 consecutive sampling cycles", the formula is as follows:

[0081] In the event of a sudden strong interference, the system will automatically increase the step size factor to 0.05 temporarily to accelerate the convergence of the coefficients. Once the error returns to within the threshold, the step size will be restored to 0.01 to ensure the effectiveness of compensation in extreme scenarios.

[0082] This embodiment optimizes the structure and logic of the basic LMS filter to improve compensation accuracy and anti-interference capability: a 32-order tap structure is adopted to improve the frequency coverage of the filter, adapt to residual crosstalk at different frequencies, and improve compensation accuracy; a dynamic step size strategy is adopted: in normal scenarios, a small step size of 0.01 is used to ensure steady-state accuracy after convergence; when encountering sudden strong interference, the step size is temporarily increased to 0.05 to accelerate convergence, and the normal step size is restored after the error returns to the threshold; a convergence judgment logic that requires the mean square error value to meet the standard for 10 consecutive sampling periods is adopted to avoid misjudgment caused by random noise in a single period and ensure the stability of filter convergence.

[0083] In a specific example, in step S4, the iterative optimization specifically involves updating the neural network parameters from step S2 using the mean square error (MSE) as the loss function, through backpropagation, as shown in the formula:

[0084] in, For the parameter set of the neural network, The gradient of the loss function with respect to the parameters is used; based on the error space distribution, the mapping rules of the hardware scan parameter mapping module in step S3 are updated to dynamically adapt to the hardware parameters. It should be noted that the error signal output by the adaptive filter... As feedback, the parameters of the front-end attention mechanism neural network and hardware scanning parameters are optimized in reverse to address the performance degradation caused by environmental changes during long-term system operation. The mean square error is used as the basis for this optimization. The loss function is used to calculate the gradient of the loss function with respect to the network parameters through backpropagation, and then the parameters are updated.

[0085] The formula for adjusting the scan interval is:

[0086] This embodiment, based on the spatial distribution characteristics of the error signal, uses FPGA to statistically analyze the local error mean square value of each region unit in real time, and then adjusts the hardware scanning parameters accordingly to achieve precise error optimization. Scanning interval Regarding adjustments, The large area indicates that the coupling and superposition in this area are still significant, and the activation interval of the unit needs to be extended to reduce signal interference.

[0087] The formula for adjusting the drive amplitude is: .

[0088] It should be noted that, by driving the amplitude Adjustment: For In areas with excessively large electric fields, appropriately reducing the amplitude of the drive signal can reduce electric field leakage.

[0089] This embodiment uses the filter error signal as feedback to achieve adaptive iterative optimization of the entire link parameters, forming a positive closed loop: Neural network parameter optimization: Using the mean square value of the filter error (MSE) as the loss function, the weights and bias parameters of the neural network in step S2 are updated through the backpropagation algorithm, making the subsequent coupling localization more accurate; Dynamic hardware parameter adaptation: Based on the spatial distribution of the error signal, for areas where the error exceeds the standard, the unit scanning interval is extended and the driving amplitude is reduced to further optimize the hardware suppression effect; forming a positive self-optimizing closed loop of more accurate coupling localization → better hardware suppression effect → smaller residual error → more accurate coupling localization.

[0090] Experimental Test The experimental data came from a flat-panel sensor array testing system built in the laboratory. This system includes array modules with different unit densities (32×32, 64×64, 128×128, 256×256), covering typical industrial testing application scenarios. It is also equipped with an adjustable temperature and humidity environment chamber (temperature -20℃~60℃, humidity 30%~90%) to simulate complex operating conditions in real-world scenarios. Output signal data from the sensor units under different coupling scenarios were acquired using a high-precision signal generator and a 16-bit high-speed data acquisition card.

[0091] In the experiment, we selected samples with significant strong coupling characteristics (such as corner unit clusters and edge continuous unit regions) from the actual acquired signal data. To more comprehensively verify the method's performance, we constructed simulated scenarios with various coupling strengths. These simulated coupling scenarios included parasitic capacitance coupling, electric field leakage coupling, and crosstalk coupling, with coupling strengths ranging from weak coupling (crosstalk percentage <5%) to strong coupling (crosstalk percentage >20%). These simulated coupling data, combined with the actual acquired data, constituted the dataset for this experiment.

[0092] Signal acquisition of the flat panel sensor array was performed using professional sensor array testing equipment. During acquisition, standard excitation points (spacing matched to unit size) were arranged on the array surface in a grid pattern. Time division multiplexing (TDM) scanning was used to sequentially activate each sensor unit and acquire its output signal. Simultaneously, a synchronous triggering module was introduced to ensure the timing consistency of the signal generator, data acquisition card, and environmental chamber control.

[0093] By utilizing graph convolution feature extraction technology, neighborhood correlation processing is performed on the acquired multi-dimensional signals to enhance the feature signals of strongly coupled units. Through quantitative analysis of coupling features in different regions, accurate localization of strongly coupled regions across the entire array is achieved, improving the resolution of coupling identification. (See also...) Figure 2-3 Strong coupling regions (original coupling value > 0.4) attenuate by 80% (corners turn blue, coupling value approximately 0.14-0.18), medium coupling regions (original coupling value 0.2-0.4) attenuate by 60% (edges turn light blue, coupling value approximately 0.16-0.24), and weak coupling regions (original coupling value < 0.2) attenuate by 30% (center remains low coupling).

[0094] See Figure 4 The suppression rate of this method remains high and fluctuates little across the entire temperature range, reaching approximately 95% at 20℃ and still 93% at 60℃. Temperature fluctuations only cause a 2.1% decrease in suppression rate, demonstrating excellent stability and improving array scanning efficiency by 66.7% compared to traditional methods.

[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for strong coupling localization and suppression of sensor arrays based on attention mechanisms, characterized in that, Includes the following steps: Step S1: Collect the output signals of each unit of the flat panel sensor array, extract the three-dimensional signal statistical features of amplitude, time series and stability, as well as the three-dimensional topological space features of coordinates, spacing and type, perform cross-domain normalization processing on the two types of features, and construct and output the normalized multi-dimensional input matrix. Step S2: Using the normalized multi-dimensional input matrix output in Step S1 as the only input, build a neural network that integrates graph convolution topological feature extraction and multi-head attention weight calculation, mine the coupling relationship between units and quantify the coupling strength, and output the localization results and coupling heatmap of strongly coupled unit pairs. Step S3: Taking the coupled heat map and strong coupling positioning result output in step S2 as input, the preset attention weight-scan parameter mapping rule is executed through the hardware scan parameter mapping module, and the scan timing optimization parameters, dynamic isolation circuit control signal, and drive signal amplitude adaptive parameters are output in sequence to complete the dynamic optimization of the hardware scan strategy and output the hardware scan control parameter set. Step S4: Using the hardware scan control parameter set output in step S3 as a reference, an LMS adaptive filter is used to generate a reverse compensation signal for the residual crosstalk after hardware suppression, and the filter error signal is output. At the same time, the filter error signal is fed back to the neural network in step S2 and the hardware scan parameter mapping module in step S3 in real time, and the network parameters and hardware scan mapping rules are iteratively optimized to form a closed loop of localization-suppression-compensation.

2. The sensor array strong coupling localization and suppression method based on attention mechanism according to claim 1, characterized in that, In step S1, the extraction of the three-dimensional signal statistical features of amplitude, timing, and stability specifically involves: To analyze the dynamic output characteristics of each unit in the flat panel sensor array, a full array signal feature matrix is ​​constructed. ,in This represents the total number of array elements. The number of rows in the array. The number of columns in the array; The amplitude feature The Gaussian-weighted mean of the voltage sequence within the scan period is calculated using the following formula: in, Scan cycle The output voltage sequence within, , , The sampling interval is... ; Number of sampling points, weighting coefficient , These are Gaussian weighting coefficients. ; The timing feature is the rise time feature. The linear fitting correction method is used for calculation, and the formula is: in, This represents the peak amplitude of the voltage sequence. The fitting coefficients are for the first half of the rising edge. The fitting coefficients are for the latter half of the rising edge. For the voltage to reach 0.1 At that moment, For the voltage to reach 0.9 The moment; The stability feature is: Sliding window variance quantization is used, and the formula is: Among them, window length , For the first The average voltage of each sliding window.

3. The sensor array strong coupling localization and suppression method based on attention mechanism according to claim 1, characterized in that, In step S1, the extraction of the three-dimensional topological spatial features of coordinates, spacing, and type specifically involves: Based on the physical layout of the flat panel sensor array, a full array topology feature matrix is ​​constructed. ,in This represents the total number of array elements. The number of rows in the array. The number of columns in the array; The coordinate feature is the planar coordinate of the element in the array's global coordinate system. ,in The unit number is used; the spacing feature is used for the unit. With unit Weighted Euclidean distance The formula is: Among them, the weighting coefficient , , , For unit row and column indexes, , For unit The row and column indices are used to amplify the contribution of spacing between row and column units, thus conforming to the physical law that crosstalk increases with spatial span. The type feature is a unit position type encoding. Multi-class coding is used for differentiation, and the formula is: And through one-hot encoding Convert to vector Therefore, the topological feature vector of a single sensor unit is , For unit The average spacing, and the topological feature matrix of the entire array are The coordinates are 2-dimensional, the average spacing is 1-dimensional, and the one-hot encoding is 3-dimensional.

4. The sensor array strong coupling localization and suppression method based on attention mechanism according to claim 1, characterized in that, In step S1, the cross-domain normalization processing and the construction of the normalized multi-dimensional input matrix are specifically as follows: The signal feature matrix is ​​normalized using the Gaussian distribution mapping method, and the formula is: in, for The Middle Line 1 Column elements, These correspond to amplitude, timing, and stability characteristics, respectively. The characteristic mean, The Sigmoid function ensures that the normalized values ​​fall within [0,1] and preserve the characteristic distribution trend, representing the standard deviation of the feature. Differential normalization is applied to the topological feature matrix: coordinate features are linearly normalized, the average spacing is inversely normalized, and one-hot encoding retains the original values; The normalized signal feature matrix With topological feature matrix By concatenating the columns, we obtain the final normalized multi-dimensional input matrix: 。 5. The sensor array strong coupling localization and suppression method based on attention mechanism according to claim 1, characterized in that, In step S2, the flat panel sensor array is first abstracted as an undirected weighted topological graph G=(V,E,W), where the node set V corresponds to all array cells, the edge existence threshold of the edge set E is 1.5 times the average spacing between array cells, and the element calculation formula of the weight matrix W is: in, This is the attenuation coefficient, ensuring that the weights decay rapidly as the spacing increases. For unit and The covariance of the output voltage, The Pearson correlation coefficient between the output voltages of the two units is used to quantify the degree of signal coupling, with a value range of []. [1,1], taking the absolute value reflects the coupling strength. The greater the weight, the stronger the coupling potential between the two units.

6. The sensor array strong coupling localization and suppression method based on attention mechanism according to claim 1, characterized in that, In step S2, the graph convolutional topology feature extraction uses a three-layer graph convolutional network (GCN). The formula for calculating layer convolution is: in, , For the first Layer learning weights For the first Layer bias, Activation function In terms of dimensional design, it was gradually increased from the initial 9 dimensions to 128 dimensions, and the final output... The carrying unit's globally coupled information; The multi-head attention weight calculation adopts a 4-head attention structure, and the single-head attention calculation formula is as follows: Where Q, K, and V are linear mapping matrices of the high-dimensional topological feature matrix H. M represents the feature dimension of a single-head attention function, and M is a mask matrix used to mask uncoupled pairs of units. The multi-head attention fusion output formula is: in, For column concatenation symbols, The output weight matrix for multi-head attention; The localization results and coupling heatmap of the strongly coupled output unit pairs are specifically calculated as follows: The coupling strength of the unit pairs is calculated by fusing multi-head attention weights, signal differences, unit spacing, and unit type, using the following formula: in, , These are the standard deviations of the output voltages of units p and q, respectively. , These are the type-specific one-hot encoded vectors for the two units; 0.5 is the weight coefficient for the spatial distance term, 0.3 is the weight coefficient for the signal stability difference term, and 0.2 is the weight coefficient for the embedding feature inner product term. A threshold is set. ,in Given the average strength of strongly coupled samples, when When the element pair (p, q) is determined to be a strongly coupled element pair, the coupling strength of the entire array is mapped to a coupling heatmap to complete the spatial positioning of the strongly coupled region.

7. The sensor array strong coupling localization and suppression method based on attention mechanism according to claim 1, characterized in that, In step S3, the first output of the hardware scan parameter mapping module is the scan timing optimization parameter, specifically implemented as follows: The array elements are divided into three levels based on coupling strength: strong coupling, medium coupling, and low coupling, corresponding to activation priorities from low to high. The activation interval of the elements is dynamically adjusted based on the average coupling strength of the elements, using the following formula: in Based on the activation interval, For one of the units The average coupling strength, where For unit The number of adjacent units, To maximize the coupling strength of the entire array, ensure an extended activation interval for strongly coupled units, and maintain the basic interval for low-coupled units, the scan path planning adopts a "low-coupling region priority" path. First, low-coupled units in the central region are activated, followed by medium and strongly coupled units. Furthermore, strongly coupled units are activated using "interval activation." Then skip adjacent strongly coupled units Activate directly To avoid continuous activation of strongly coupled unit pairs; The second output of the hardware scan parameter mapping module is a dynamic isolation circuit control signal, specifically implemented as follows: The driver and sampling terminals of each sensor unit are connected in series with an isolated MOSFET. The switching signal is generated by the FPGA controller based on the coupling strength output from the mapping module; an isolation threshold is set. The control logic formula is: in, This is the MOSFET's turn-on voltage. This is the turn-off voltage for the MOSFET; after conduction, the parasitic capacitance between strongly coupled units is short-circuited to ground, enabling multi-unit collaborative isolation control of the strongly coupled unit cluster; The third output of the hardware scanning parameter mapping module is the adaptive parameter for the amplitude of the driving signal, specifically implemented as follows: Establish a linear mapping model between coupling strength and driving amplitude, with the following formula: in, The base driving amplitude is set, with an amplitude attenuation coefficient of 0.

6. This ensures that the driving amplitude of the strongly coupled units is reduced, decreasing the electric field radiation range, while the weakly coupled units maintain the base amplitude. Amplitude output is achieved through a 12-bit DAC, achieving a resolution of [missing information]. This ensures that the amplitude adjustment accuracy matches the changes in coupling strength, thereby achieving amplitude accuracy control.

8. The sensor array strong coupling localization and suppression method based on attention mechanism according to claim 1, characterized in that, In step S4, an LMS adaptive filter is used to generate a reverse compensation signal for the residual crosstalk after hardware suppression, specifically as follows: The residual crosstalk is canceled by an LMS adaptive filter circuit optimized for filtering structure, with the hardware scan clock signal as the reference input. The sensor output signal is the desired input. Filter output To compensate for the signal, the filter coefficients are updated in real time using the minimum mean square error criterion, as shown in the following formula: in for Time-based filter coefficients Step size factor As an error signal, ensure that the compensation signal and the residual crosstalk have equal amplitude and opposite phase, and set a residual crosstalk threshold. ,when The coefficient update is stopped at a certain time to balance the compensation accuracy and the amount of calculation.

9. The sensor array strong coupling localization and suppression method based on attention mechanism according to claim 1, characterized in that, In step S4, the LMS adaptive filter adopts a 32nd-order tap structure, and the filter output formula is: in, Let be the weight of the k-th tap at time n. The hardware scan clock signal is used as a reference input. For delayed reference signal; The formulas for coefficient updates and convergence determination are as follows: in, This is an error signal that reflects residual crosstalk. The desired signal output by the sensor; The convergence determination needs to avoid the randomness of a single-cycle error, and adopts the judgment logic of "the mean square value of the error meets the standard for 10 consecutive sampling cycles", the formula is as follows: In the event of a sudden strong interference, the system will automatically increase the step size factor to 0.05 temporarily to accelerate the convergence of the coefficients. Once the error returns to within the threshold, the step size will be restored to 0.01 to ensure the effectiveness of compensation in extreme scenarios.

10. The sensor array strong coupling localization and suppression method based on attention mechanism according to claim 1, characterized in that, In step S4, the iterative optimization specifically involves updating the neural network parameters from step S2 using the mean square error (MSE) as the loss function, through backpropagation, as shown in the formula: in, For the parameter set of the neural network, The gradient of the loss function with respect to the parameters is used; based on the error space distribution, the mapping rules of the hardware scan parameter mapping module in step S3 are updated to dynamically adapt to the hardware parameters. The formula for adjusting the scan interval is: The formula for adjusting the drive amplitude is: 。

Images