Three-dimensional force decoupling method and system for flexible sensor based on lightweight hybrid algorithm

Abstract

The present application belongs to the technical field of three-dimensional force decoupling of flexible sensors, and relates to a three-dimensional force decoupling method and system of a flexible sensor based on a lightweight hybrid algorithm. The method comprises collecting multiple original signals output by each electrode in a target sensing unit of the flexible sensor and converting them into multiple digital signals; and performing baseline correction based on adaptive baseline adjustment, neighborhood crosstalk suppression based on a neighborhood signal correlation analysis model, and filtering processing based on an adaptive filtering architecture, respectively, to obtain cleaned multiple signals; extracting core features of the cleaned multiple signals, calculating a three-dimensional force rough value according to the core features, determining a current correction mode according to a real-time resultant force amplitude of the target sensing unit at the current moment, correcting the three-dimensional force rough value, and outputting the final three-dimensional force after decoupling. The problem that the accuracy, stability and algorithm power consumption cannot be balanced when decoupling the end-side three-dimensional force of an array-type flexible sensor is solved.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional force decoupling technology for flexible sensors, and relates to a method and system for three-dimensional force decoupling of flexible sensors based on a lightweight hybrid algorithm. Background Technology

[0002] Flexible tactile sensors, with their excellent flexibility and surface conformability, have become core components in the field of intelligent sensing, with their three-dimensional force detection capability being a key performance indicator. Existing three-dimensional force decoupling technologies mainly fall into two categories: one is hardware self-decoupling technology, which relies on special physical structures or composite sensitive materials to reduce algorithm dependence, but this is cumbersome in manufacturing, costly, and suffers from poor consistency; the other is algorithm decoupling technology, which, based on traditional mechanical modeling methods, requires calibrating massive amounts of parameters and is greatly affected by power frequency noise, while heavy-duty deep learning models have extremely high computational requirements and cannot run in real-time on resource-constrained low-cost embedded devices such as MCUs. Furthermore, existing algorithms are susceptible to baseline drift caused by sensor temperature drift and interference from physical crosstalk between channels, resulting in insufficient decoupling accuracy and stability of the final tactile information. Summary of the Invention

[0003] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a three-dimensional force decoupling method for flexible sensors based on a lightweight hybrid algorithm.

[0004] The objective of this invention can be achieved through the following technical solution: a three-dimensional force decoupling method for flexible sensors based on a lightweight hybrid algorithm, comprising: The system collects multiple raw signals output from each electrode within the target sensing unit of the flexible sensor and converts them into multiple digital signals. The multiple digital signals are subjected to baseline correction based on adaptive baseline adjustment, neighborhood crosstalk suppression based on neighborhood signal correlation analysis model, and filtering processing based on adaptive filtering architecture to filter the multiple digital signals and obtain cleaned multiple signals. The core features of the cleaned multi-channel signals are extracted, and the three-dimensional force rough value is calculated based on the core features. The core features include mean features, difference features, and gradient features. The current correction mode is determined based on the real-time resultant force amplitude of the target sensing unit at the current moment; Based on the determined correction mode, the coarsely calculated three-dimensional force value is corrected, and the corrected force value is smoothed by sliding mean filtering to output the decoupled final three-dimensional force.

[0005] As an optional embodiment of the present invention, baseline correction based on adaptively adjusted baseline includes: Calculate the rate of change of variance of the multiple digital signals within the current sliding window; The window length of the sliding window is dynamically adjusted based on the comparison between the variance change rate and the preset change rate threshold. Based on the number of sampling points in the adjusted sliding window, calculate the current window mean and window standard deviation to determine the current dynamic threshold range; In response to the current sampling point falling within the dynamic threshold range, the baseline is updated using an exponentially weighted moving average. The drift error is corrected by subtracting the updated baseline from the current multi-channel digital signal.

[0006] As an optional embodiment of the present invention, neighborhood crosstalk suppression based on a neighborhood signal correlation analysis model includes: Centered on the target sensing unit, acquire the signals of the target sensing unit and its eight neighboring sensing units within the sliding window period; Calculate the Pearson correlation coefficient between the target sensing unit and each adjacent sensing unit, and normalize the influence weight of the adjacent sensing units based on the Pearson correlation coefficient to construct the neighborhood influence weight matrix. The crosstalk component estimate is obtained by weighting and summing the weights of each neighborhood influence in the neighborhood influence weight matrix with the multi-channel digital signals of the corresponding adjacent sensing units. The crosstalk component is removed by subtracting the crosstalk component estimate from the signal of the target sensing unit.

[0007] As an optional embodiment of the present invention, the filtering process based on the adaptive filtering architecture includes: The multi-channel digital signals are input to the front-end power frequency notch filter to filter out power frequency interference. The dead zone threshold is set based on the noise standard deviation of the multiple digital signals; The error term of the subsequent minimum mean square adaptive filter is ternarily quantized using the dead zone threshold. In response to the absolute value of the error term before quantization being greater than the dead zone threshold, the weight coefficients of the adaptive filter are updated using the error term after ternary quantization. In response to the absolute value of the error term before quantization being less than or equal to the dead zone threshold, the current adaptive filter weight coefficient update is frozen.

[0008] As an optional embodiment of the present invention, determining the current correction mode based on the real-time resultant force amplitude of the target sensing unit at the current moment includes: The real-time resultant force amplitude is calculated using the aforementioned three-dimensional force coarse calculation value; In response to the real-time resultant force amplitude of the target sensing unit being less than a preset force amplitude threshold, the first correction mode is determined; In response to the real-time resultant force amplitude of the target sensing unit being greater than or equal to the preset force amplitude threshold, a second correction mode is determined; The decoupling accuracy of the first correction mode is greater than that of the second correction mode, and the single-frame decoupling time in the first correction mode is greater than that in the second correction mode.

[0009] As an optional embodiment of the present invention, the three-dimensional force coarse calculation value is corrected based on a determined correction mode, including: In the first correction mode, the core features and the coarse calculation value of the three-dimensional force are input into the trained first fully connected neural network, and the corresponding compensation amount of the three-dimensional force is output. The coarse calculation value of the three-dimensional force is added to the compensation amount to obtain the decoupled three-dimensional force of the target sensing unit. In the second correction mode, the coarse three-dimensional force value is linearly corrected according to a preset linear correction coefficient to obtain the decoupled three-dimensional force of the target sensing unit.

[0010] As an optional embodiment of the present invention, when acquiring the multiple raw signals, an intermittent adaptive sampling step is further included: Periodically calculate the peak-to-peak value of the original signals of all sensing units within the sliding window; In response to the fact that the peak-to-peak value of all sensing units is less than the preset silence threshold and the duration is greater than the preset standby threshold, the signal is acquired at the first sampling frequency. When the signal is acquired at the first sampling frequency, in response to the mean feature of the multiple raw signals of any sensing unit deviating from the baseline by more than a preset wake-up threshold, the sampling frequency is switched to the second sampling frequency to acquire the signal. Specifically, a hysteresis protection mechanism is used to switch between the first sampling frequency and the second sampling frequency. The preset switching threshold for switching back from the second sampling frequency to the first sampling frequency is less than the preset wake-up threshold for switching from the first sampling frequency to the second sampling frequency.

[0011] As an optional embodiment of the present invention, the filtering of the multiple digital signals, the calculation of the three-dimensional force rough value, and the correction of the three-dimensional force rough value are all performed using a fixed-point number format for signed integer calculation.

[0012] This invention also proposes a three-dimensional force decoupling system for flexible sensors based on a lightweight hybrid algorithm, comprising: The first unit is used to collect multiple raw signals output by each electrode in the target sensing unit of the flexible sensor and convert them into multiple digital signals. The second unit is used to perform baseline correction based on adaptive baseline adjustment, neighborhood crosstalk suppression based on neighborhood signal correlation analysis model, and filtering processing based on adaptive filtering architecture on the multiple digital signals respectively, so as to filter the multiple digital signals and obtain cleaned multiple signals. The third unit is used to extract the core features of the cleaned multi-channel signals and calculate the three-dimensional force rough value based on the core features, wherein the core features include mean features, difference features and gradient features. The fourth unit is used to determine the current correction mode based on the real-time resultant force amplitude of the target sensing unit at the current moment. The fifth unit is used to correct the coarsely calculated three-dimensional force value based on a determined correction mode, and to perform sliding mean smoothing filtering on the corrected force value to output the decoupled final three-dimensional force.

[0013] The present invention also provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement the aforementioned three-dimensional force decoupling method for flexible sensors based on a lightweight hybrid algorithm when executing executable instructions.

[0014] Compared with the prior art, the present invention has the following beneficial technical effects: 1) By organically combining signal cleaning and preprocessing, coarse calculation and extraction with dynamic correction mode, the problem of unbalanced accuracy, stability and computing power consumption when array-type flexible sensors are decoupled at the end side three-dimensional force is solved.

[0015] 2) In the preprocessing stage, adaptive correction of temperature drift, suppression of inter-channel crosstalk, and filtering of power frequency noise ensure the signal-to-noise ratio of the physical signal from the source. In the decoupling stage, real-time adaptive switching between "high precision - heavy computing power" and "low computing power - high efficiency" correction modes achieves a synergistic effect of saving computing resources under high force values ​​and ensuring decoupling accuracy under minute torque changes. This ensures excellent real-time decoupling performance and anti-interference capabilities with extremely low computing power consumption, making it ideal for low-cost, resource-constrained embedded deployments. Attached Figure Description

[0016] Figure 1 This is a flowchart of a three-dimensional force decoupling method for flexible sensors based on a lightweight hybrid algorithm according to an embodiment of the present invention; Figure 2 This is a hardware connection block diagram of the three-dimensional force decoupling of the flexible sensor according to an embodiment of the present invention; Figure 3 This is a three-dimensional force waveform diagram of the flexible sensor before three-dimensional force decoupling according to an embodiment of the present invention; Figure 4 This is a three-dimensional force waveform diagram of the flexible sensor after three-dimensional force decoupling according to an embodiment of the present invention; Figure 5 This is a block diagram of a three-dimensional force decoupling system for a flexible sensor based on a lightweight hybrid algorithm, according to an embodiment of the present invention. Detailed Implementation

[0017] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings to further illustrate the technical solutions of the present invention. However, the present invention is not limited to these embodiments.

[0018] Example 1

[0019] Based on the technical problems highlighted in the background, this embodiment proposes a three-dimensional force decoupling method for flexible sensors based on a lightweight hybrid algorithm. This method is particularly suitable for low-cost flexible piezoresistive, capacitive, and piezoelectric sensor arrays without complex biomimetic microstructures or special electrode layouts. Figure 1 As shown, the method includes: S1: Collect multiple raw signals output by each electrode in the target sensing unit of the flexible sensor and convert them into multiple digital signals; S2, perform baseline correction based on adaptive baseline adjustment, neighborhood crosstalk suppression based on neighborhood signal correlation analysis model, and filtering processing based on adaptive filtering architecture on the multiple digital signals respectively to filter the multiple digital signals and obtain the cleaned multiple signals. S3, extract the core features of the cleaned multi-channel signals, and calculate the three-dimensional force rough value based on the core features, wherein the core features include mean features, difference features and gradient features; S4, determine the current correction mode based on the real-time resultant force amplitude of the target sensing unit at the current moment; S5. Based on the determined correction mode, the coarsely calculated three-dimensional force value is corrected, and the corrected force value is smoothed by sliding mean filtering to output the decoupled final three-dimensional force.

[0020] The flexible sensor in this embodiment is specifically a flexible tactile sensor, such as... Figure 2 As shown, the flexible tactile sensor array adopts a cross-array structure, consisting of several rows of input axes and several columns of output axes intersecting, with each intersection integrating a force-sensitive sensing unit. The input axes are connected to the channel selection path, and a constant current drive voltage is provided by a DC excitation signal; the output axes are connected to a multi-channel analog switch array, realizing point-by-point scanning of row and column signals under the control of the MCU. The selected force-sensitive sensing unit is the target sensing unit. The multiple original signals output by each electrode in this force-sensitive sensing unit are analog signals, which are differentially amplified by an operational amplifier (OPA) to convert the weak force-sensitive signal into a voltage range suitable for ADC sampling. Subsequently, the STM32 MCU performs analog-to-digital conversion through the ADC to obtain multiple digital force-sensitive signals.

[0021] Multi-channel digital force-sensitive signals are susceptible to environmental high-frequency noise, crosstalk between array units, and baseline drift caused by temperature drift and material viscoelasticity. Directly using these signals for decoupling would significantly reduce accuracy and stability. This embodiment employs multi-level adaptive preprocessing to eliminate interference signals and solidify the foundation of decoupling data. Multi-level adaptive preprocessing includes baseline correction through adaptive baseline adjustment, neighborhood crosstalk suppression based on a neighborhood signal correlation analysis model, and filtering based on an adaptive filtering architecture. This breaks through the limitations of traditional preprocessing's single noise reduction approach, forming a lightweight, end-to-end preprocessing system.

[0022] The core features of the multi-channel digital force-sensitive signals after multi-level adaptive preprocessing are extracted, including mean features, difference features, and gradient features. Based on the core features, the three-dimensional force coarse value is calculated. Then, a hybrid decoupling architecture of simplified mechanical coarse decoupling + lightweight neural network fine correction is adopted. The simplified mechanical coarse decoupling correction mode or the lightweight neural network fine correction mode is determined according to the real-time resultant force amplitude of the target sensing unit at the current moment to correct the three-dimensional force coarse value. Then, the sliding mean smoothing filter is applied to obtain the final three-dimensional force after decoupling. The purpose of the sliding mean smoothing filter is to remove instantaneous noise spikes, optimize the force output curve, avoid abrupt changes in the decoupling result, and improve the stability of the three-dimensional force output, which is suitable for scenarios with strict stability requirements such as robot precision grasping and medical flexible palpation.

[0023] Preferably, baseline correction based on adaptively adjusted baseline includes: Calculate the rate of change of variance of the multiple digital signals within the current sliding window; The window length of the sliding window is dynamically adjusted based on the comparison between the variance change rate and the preset change rate threshold. Based on the number of sampling points in the adjusted sliding window, calculate the current window mean and window standard deviation to determine the current dynamic threshold range; In response to the current sampling point falling within the dynamic threshold range, the baseline is updated using an exponentially weighted moving average. The drift error is corrected by subtracting the updated baseline from the current multi-channel digital signal.

[0024] To address the baseline drift issue of flexible sensors during long-term operation, this embodiment employs a sliding window mean fitting combined with a dynamic threshold compensation algorithm to dynamically adjust the baseline in real time. Specifically, the variance change rate of multiple digital signals within the sliding window is monitored in real time. Based on the variance change rate, it is determined whether the sensor is in a rapid change phase or a steady-state phase, thereby adjusting the window length of the sliding window. The variance change rate is compared with a preset change rate threshold. If the variance change rate is greater than the preset change rate threshold, it is determined that the sensor response has entered a rapid change phase, and the window length is automatically shortened. (This embodiment defaults to 10 sampling points) to improve response speed. If the variance change rate is less than or equal to a preset change rate threshold, the sensor is determined to be in a steady state, and the window length is automatically extended to... (This embodiment defaults to 50 sampling points) to enhance the filtering and smoothing effect. The preset rate of change threshold is determined according to the following mapping relationship:

[0025] To preset the rate of change threshold, As the baseline threshold, The sensor response sensitivity, measured in mV / N, is determined during the calibration phase. Different sensitive materials have different response sensitivities; the control system automatically measures and stores this value during the initial calibration to achieve plug-and-play adaptive adaptation.

[0026] The dynamic threshold compensation algorithm specifically avoids misclassifying slow, hysteretic changes as valid signals by using a dynamic threshold and participates in the baseline update rule. Specifically, it adjusts the number of sampling points N in the adjusted sliding window and the current sampling point... According to the formula Calculate the current window mean Further based on the formula Calculate the standard deviation of the window .

[0027] The dynamic threshold range is calculated using the following formula. :

[0028] Where k is an adaptive coefficient, which is set to 2.5 by default in this embodiment. It can be automatically adjusted according to the hysteresis characteristics of the sensor material. Specifically, when M consecutive sampling points deviate from the current baseline by more than a preset offset, material hysteresis deformation is determined, and the value of k is dynamically increased (e.g., k = k + 0.5) to widen the threshold range. Furthermore, if the current sampling point Falling within the dynamic threshold range Within this period, the baseline is updated using an exponentially weighted moving average (EWMA).

[0029] in Adjustable, used to balance baseline tracking speed and anti-interference capability. Indicates the current baseline. This represents the updated baseline. The drift error is corrected by subtracting the updated baseline from the current multi-channel digitized force-sensitive signal.

[0030] By dynamically adjusting the sliding window length through the variance change rate, the baseline update frequency can be made adaptive to the sensor's stress state, preventing the misinterpretation of actual external force changes as temperature drift. Dynamically updating the baseline using an exponentially weighted moving average accurately tracks and eliminates low-frequency zero-point drift, improving the sensor's zero-point stability under complex thermal response environments.

[0031] Preferably, neighborhood crosstalk suppression based on a neighborhood signal correlation analysis model includes: Centered on the target sensing unit, acquire the signals of the target sensing unit and its eight neighboring sensing units within the sliding window period; Calculate the Pearson correlation coefficient between the target sensing unit and each adjacent sensing unit, and normalize the influence weight of the adjacent sensing units based on the Pearson correlation coefficient to construct the neighborhood influence weight matrix. The crosstalk component estimate is obtained by weighting and summing the weights of each neighborhood influence in the neighborhood influence weight matrix with the multi-channel digital signals of the corresponding adjacent sensing units. The crosstalk component is removed by subtracting the crosstalk component estimate from the signal of the target sensing unit.

[0032] To address the crosstalk problem between units in array-type flexible sensors, this embodiment constructs a neighborhood influence weight matrix and uses a weighted filtering algorithm to accurately remove crosstalk components. Specifically, let the target sensing unit be C(i,j), and its corresponding digital force-sensitive signal be S(i,j); the set of adjacent sensing units within the eight neighborhoods of the target sensing unit C(i,j) is denoted as... , can be represented as:

[0033] Correspondingly, the digital force-sensitive signal output by each adjacent sensing unit in the eight-neighbor domain is denoted as S(m,n).

[0034] Define the neighborhood correlation coefficient matrix R, and calculate the Pearson correlation coefficient R(m,n) between the target sensing unit C(i,j) and each of its neighboring sensing units C(m,n):

[0035] In the formula This represents the covariance between the target sensing unit signal and the signals of adjacent sensing units within the current sliding window; and These represent the standard deviations of the target sensing unit signal and the adjacent sensing unit signal within the current sliding window, respectively.

[0036] Furthermore, the influence weights of each neighborhood are calculated by normalizing the absolute value of the Pearson correlation coefficient. :

[0037] according to Construct the corresponding 3x3 dimension neighborhood influence weight matrix Taking the target sensing unit as C(2,2) as an example, the corresponding neighborhood influence weight matrix Represented as:

[0038] Since the signal contribution of the target sensing unit itself is not taken into account when calculating crosstalk components, the weight of the target sensing unit itself is set... The remaining elements Normalized influence weights corresponding to each adjacent sensing unit C(m, n) .

[0039] Then, the non-zero neighborhood influence weights in the neighborhood influence weight matrix are weighted and summed with the multi-channel digital signals S(m,n) of the corresponding adjacent sensing units C(m,n) to obtain the estimated crosstalk component for the target sensing unit C(i,j):

[0040] Finally, the crosstalk component estimate is subtracted from the signal of the target sensing unit, and the true digital force-sensitive signal after removing crosstalk is expressed as:

[0041] It should be noted that in the above calculation process, the signal of the target sensing unit participating in the crosstalk component removal calculation can be an analog signal without A / D conversion or a digitized digital signal before stripping. This embodiment does not impose specific restrictions on this, and both can effectively remove crosstalk components using the neighborhood crosstalk suppression method based on the neighborhood signal correlation analysis model described above. The Pearson correlation coefficient within the sliding window reflects the strength of mechanical and electrical coupling between channels, and the suppression matrix is ​​dynamically constructed through normalized weights, eliminating the limitations of fixed-weight filtering. Accurate estimation and stripping of the physical crosstalk components of adjacent units in the spatial domain significantly reduces inter-channel mis-touch and crosstalk deviations caused by localized pressure on the sensing array.

[0042] Preferably, the filtering process based on the adaptive filtering architecture includes: The multi-channel digital signals are input to the front-end power frequency notch filter to filter out power frequency interference. The dead zone threshold is set based on the noise standard deviation of the multiple digital signals; The error term of the subsequent minimum mean square adaptive filter is ternarily quantized using the dead zone threshold. In response to the absolute value of the error term before quantization being greater than the dead zone threshold, the weight coefficients of the adaptive filter are updated using the error term after ternary quantization. In response to the absolute value of the error term before quantization being less than or equal to the dead zone threshold, the current adaptive filter weight coefficient update is frozen.

[0043] The adaptive filtering architecture in this embodiment is a hybrid filtering structure combining Infinite Impulse Response (IIR) and Finite Impulse Response (FIR). Specifically, the front-end uses a second-order IIR notch filter to specifically filter out 50Hz power frequency interference. The system transfer function of the second-order IIR notch filter... for:

[0044] in The normalized chemical frequency center angular frequency, Indicates the sampling frequency of the signal. The convergence factor r is used to control the bandwidth of the notch filter. The closer r is to 1, the narrower the notch filter bandwidth and the better the frequency selectivity, but the longer the transient response. In this embodiment, r is preferably set to 0.95. After multiple digital signals pass through the front-end power frequency notch filter, the output is the digital signal after power frequency filtering.

[0045] For the subsequent least mean square (LMS) adaptive filter, the traditional LMS filter requires the following weight vector update formula for each iteration:

[0046] In the formula, Let be the weight vector of the filter. To update the step size, This is the error term before quantization. The input vector of the filter is used, and the process involves a large number of floating-point multiplication operations.

[0047] To enable real-time operation on resource-constrained target MCUs (such as microcontrollers without floating-point units), this embodiment introduces a dead-zone quantization mechanism to eliminate complex multiplication calculations.

[0048] First, calculate the noise standard deviation of the current digitized force-sensitive signal. In specific calculations, a preset number of digital force-sensitive signal samples can be collected during system initialization or in silent standby mode, and their standard deviation can be calculated. This standard deviation can then be used as the noise standard deviation. Next, a dead zone threshold is set based on the noise standard deviation. :

[0049] In the formula, The preset multiplier factor is preferably set to 1.5 in this embodiment.

[0050] Next, we will use the dead zone threshold. For the error term Perform ternary quantization, specifically:

[0051]

[0052]

[0053] Original item: Simplified to: sign(x), which performs binary quantization on each component of the input vector:

[0054]

[0055] The weight coefficients are then updated as follows:

[0056] in, The preferred update step size is an integer power of 2. In this case, the multiplication calculation in the weight vector update formula is completely reduced to a data shift operation, thus eliminating floating-point multiplication and significantly reducing the processing power of the embedded processor. Furthermore, the update step size is set only if the absolute value of the original error before quantization is greater than a preset dead zone threshold. The weight coefficient update step will only be triggered when the absolute value of the original error before quantization is less than or equal to the preset dead zone threshold. If this happens, the current adaptive filter weight coefficient update will be automatically frozen, maintaining the original weight values. It should be noted that the error term... Trivalued quantization and original terms The binary quantizations are independent of each other, and the two quantizations do not interfere with each other. This freezing strategy can effectively avoid the weight coefficient drift caused by the continuous accumulation of noise in the signal silent or pure noise interference segments, and can reduce more than 90% of invalid update operations, greatly alleviating the computational pressure on the kernel.

[0057] Preferably, determining the current correction mode based on the real-time resultant force amplitude of the target sensing unit at the current moment includes: The real-time resultant force amplitude is calculated using the aforementioned three-dimensional force coarse calculation value; In response to the real-time resultant force amplitude of the target sensing unit being less than a preset force amplitude threshold, the first correction mode is determined; In response to the real-time resultant force amplitude of the target sensing unit being greater than or equal to the preset force amplitude threshold, a second correction mode is determined; The decoupling accuracy of the first correction mode is greater than that of the second correction mode, and the single-frame decoupling time in the first correction mode is greater than that in the second correction mode.

[0058] In this embodiment, after calculating the three-dimensional force rough value based on the core features of the multi-channel signals after cleaning the target sensing unit, the real-time resultant force amplitude is further calculated. Based on the comparison between the real-time resultant force amplitude and the preset force amplitude threshold, a small force detection scenario and a large force detection scenario are divided, each corresponding to a different correction mode. Specifically, the three-dimensional force rough value is determined based on a mapping relationship. This mapping relationship is established for conventional flexible piezoresistive / capacitive sensing units without special microstructures, establishing a simplified "stress-conductivity / capacitance" mapping relationship. Higher-order nonlinear interference terms are ignored, enabling the extraction of the core features of the digital force-sensitive signal output by the multi-electrode of the target sensing unit, and quickly calculating the three-dimensional force rough value. The specific calculation process is as follows: Let the multiple signals (four in this embodiment) after cleaning the same target sensing unit be S1, S2, S3, and S4. Extract the core features, including the mean feature M and the difference feature... and Gradient features and The feature extraction formulas are as follows:

[0059]

[0060]

[0061]

[0062]

[0063] The mean feature M represents the mean characteristic of the normal force of the target sensing unit, and the difference feature... This represents the characteristic of the tangential force difference in the X direction of the target sensing unit. This represents the tangential force difference characteristic and gradient characteristic of the target sensing unit in the Y direction. The gradient feature represents the gradient characteristics in the X-direction after nonlinear correction. This represents the gradient characteristics in the Y-direction after nonlinear correction.

[0064] The three-dimensional force coarse decoupling formula based on the above characteristics is as follows:

[0065]

[0066]

[0067] in, To roughly calculate the tangential force in the X direction, To roughly calculate the tangential force in the Y direction, For coarse calculation of axial force, , , The coefficients are linear. , , The bias is obtained through pre-calibration. Using the above method to calculate the coarse three-dimensional force values, there is no need to accurately calibrate all mechanical parameters. Only 10-20 sets of samples are needed to complete the preliminary modeling, which greatly reduces the calibration time and cost.

[0068] The formula for calculating the real-time resultant force amplitude using the three-dimensional force rough calculation is as follows:

[0069] Real-time resultant force amplitude The calibration mode is determined by comparing the preset force amplitude threshold with the actual force amplitude threshold. The preset force amplitude threshold is determined as a percentage of the sensor's full-scale range, preferably 15%. For example, if the full-scale range is 10N, the preset force amplitude threshold is 1.5N. When the real-time resultant force amplitude is less than 1.5N, it is determined to be a low-force detection scenario, and the first calibration mode is used to correct the coarsely calculated three-dimensional force value. When the real-time resultant force amplitude is greater than or equal to 1.5N, it is determined to be a high-force detection scenario, and the second calibration mode is used to correct the coarsely calculated three-dimensional force value. The first calibration mode used in the low-force detection scenario is a high-precision calibration mode, characterized by ensuring the decoupling accuracy of weak force signals. The second calibration mode used in the high-force detection scenario is a fast decoupling mode, characterized by improving the calculation response speed. Using two calibration modes balances decoupling accuracy and real-time performance under different working conditions, solves the problems of distortion in low-force scenarios and delay in high-force scenarios, and reasonably optimizes the processor's power consumption.

[0070] Preferably, the three-dimensional force coarse calculation value is corrected based on a determined correction mode, including: In the first correction mode, the core features and the coarse calculation value of the three-dimensional force are input into the trained first fully connected neural network, and the corresponding compensation amount of the three-dimensional force is output. The coarse calculation value of the three-dimensional force is added to the compensation amount to obtain the decoupled three-dimensional force of the target sensing unit. In the second correction mode, the coarse three-dimensional force value is linearly corrected according to a preset linear correction coefficient to obtain the decoupled three-dimensional force of the target sensing unit.

[0071] In the first correction mode, a complete neural network fine correction process is activated to address the nonlinear error remaining from the coarse decoupling. The neural network is a pre-trained first fully connected neural network, which calculates three-dimensional force through the entire chain of the input layer, the first hidden layer, the second hidden layer, and the output layer. , , Corresponding compensation amount , , Simultaneously, an additional adaptive filtering and noise reduction process is performed on the cleaned multi-channel signals. The input layer has 8 nodes, specifically the mean feature M, the difference feature... and Gradient features and Rough calculation of X-direction tangential force Rough calculation of tangential force in the Y direction Coarse algorithm direction force The first hidden layer has 20 nodes and uses ReLU (Rectified Linear Unit) as the activation function. The second hidden layer has 8 nodes and also uses ReLU as the activation function. This dual-hidden-layer design better captures the decoupling behavior. The output layer has 3 nodes, specifically for compensation. , , .

[0072] To facilitate the training and real-time computation of the first fully connected neural network, this embodiment establishes an embedded edge hardware verification system. This system uses a commercially available polydimethylsiloxane (PDMS)-based flexible piezoresistive sensing array (containing four electrode sensing units and without special surface microstructures) as the sensing element, paired with a low-cost microcontroller (specifically an STM32F103C8T6 chip) as the edge algorithm controller. After preprocessing, hybrid decoupling, and low-power optimization, the algorithm is integrated and programmed into the microcontroller to acquire sensor electrode signals in real time and complete the computation. , , Real-time calculation of three-dimensional forces.

[0073] In this embodiment, by collecting 15 sets of three-dimensional force calibration samples with different amplitudes and directions, the algorithm parameters were debugged and the lightweight training of the first fully connected neural network was completed. The specific process is as follows: (1) Sample characteristic distribution: The specific distribution of the 15 sets of calibration samples is as follows: normal force Five groups of directional (vertical pressing) force values, each corresponding to 10%, 25%, 50%, 75%, and 100% of full scale; tangential force. Five groups of forces (horizontal X-axis sliding) are applied, with each group corresponding to 10%, 25%, 50%, 75%, and 100% of the full scale; tangential force... Five groups of directions (horizontal Y-axis sliding) are used, and the applied force amplitude of each group corresponds to 10%, 25%, 50%, 75% and 100% of the full scale.

[0074] (2) Sample size and dataset partitioning: At a sampling frequency of 1 kHz, each set of calibration samples was continuously sampled for 0.5 seconds to obtain 500 valid feature sampling points. A total of 7500 sample points were constructed from 15 sets of calibration samples. When constructing the neural network offline, the dataset was randomly divided into a training set containing 6000 sampling points and a validation set containing 1500 sampling points in an 8:2 ratio. The validation set was used to monitor overfitting during training iterations.

[0075] (3) Online data augmentation strategy: To address the risk of overfitting due to the small size of the calibration sample, this embodiment introduces an online data augmentation mechanism to expand the feature signals during network training. The online data augmentation mechanism includes: (a) injecting standard deviation Additive Gaussian noise, in which (a) representing the original signal amplitude; (b) introducing a random temporal shift within ±10 sampling points on the time axis; (c) introducing a random amplitude scaling factor of 0.95 to 1.05. Through online augmentation, the equivalent training sample size is increased by approximately 10 times, improving the model's generalization ability and drift robustness.

[0076] (4) Training and convergence status: After online enhancement and expansion, the first fully connected neural network (with a total of only about 450 parameters for network weights and biases) constructed in this embodiment converged stably after 1000 rounds of iterative training. The training loss and validation loss curves decreased smoothly and tended to fit each other, and there was no obvious overfitting phenomenon.

[0077] In this embodiment, the single-frame calculation time in this correction mode is about 8-10ms, and the decoupling accuracy is less than or equal to 2%, which meets the requirements for decoupling accuracy in low-force detection scenarios.

[0078] After obtaining the compensation amount, add the rough calculation value of the three-dimensional force to the corresponding compensation amount to obtain the final decoupled three-dimensional force:

[0079]

[0080]

[0081] Actual measurements show that the average error of decoupled three-dimensional force is ≤2.8%, and the single-frame decoupling response delay is ≤10ms, meeting the requirements for real-time tactile perception.

[0082] In the second correction mode, in order to maximize computing power in high-force detection scenarios where sensor deformation is large and nonlinear coupling is relatively smooth, the system disables the inference of the first and last multi-layer neural networks and directly performs linear correction on the coarse three-dimensional force value according to the preset linear correction coefficient to obtain the decoupled three-dimensional force of the target sensing unit. The correction formula is as follows:

[0083]

[0084]

[0085] in , The preset linear correction coefficient can be obtained in advance through calibration.

[0086] In this embodiment, the computational overhead of the neural network is eliminated in this correction mode, resulting in a single-frame decoupling time of approximately 3-5ms and a decoupling accuracy within 5%, which fully meets the priority requirements of computation and response speed in high-intensity detection scenarios. The decoupling accuracy of the first correction mode is greater than that of the second correction mode, but the single-frame decoupling time in the first correction mode is greater than that in the second correction mode.

[0087] The combination of the two correction modes reduces the computing power requirement by more than 90% compared to traditional deep learning models, perfectly adapts to low-cost embedded hardware, requires no external computing power module, has extremely low integration difficulty, and is easy to scale up for mass production.

[0088] like Figure 3-4 The three-dimensional force of the unloaded force value before decoupling calibration , , The three-dimensional force of the unloaded force after decoupling calibration , , Both are close to 0, indicating a significant decoupling effect.

[0089] Preferably, the acquisition of the multiple raw signals further includes an intermittent adaptive sampling step: Periodically calculate the peak-to-peak value of the original signals of all sensing units within the sliding window; In response to the fact that the peak-to-peak value of all sensing units is less than the preset silence threshold and the duration is greater than the preset standby threshold, the signal is acquired at the first sampling frequency. When the signal is acquired at the first sampling frequency, in response to the mean feature of the multiple raw signals of any sensing unit deviating from the baseline by more than a preset wake-up threshold, the sampling frequency is switched to the second sampling frequency to acquire the signal. Specifically, a hysteresis protection mechanism is used to switch between the first sampling frequency and the second sampling frequency. The preset switching threshold for switching back from the second sampling frequency to the first sampling frequency is less than the preset wake-up threshold for switching from the first sampling frequency to the second sampling frequency.

[0090] In this embodiment, the sampling frequency of the sliding window can also be adaptively changed. Specifically, firstly, the peak-to-peak value of the multiple raw signals from all sensing units within the sliding window is periodically calculated, and each peak-to-peak value is compared with a preset silence threshold. The preset silence threshold is determined based on a multiple of the peak-to-peak value of the sensor's background noise. In this embodiment, the background noise is 5mV, and the preset silence threshold is preferably twice the peak-to-peak value of the background noise. When the peak-to-peak value is less than the preset silence threshold and the duration is greater than a preset standby threshold, it is determined that the sensor is in a state without external force, and the MCU controller enters a low-power standby mode. The sliding window then acquires signals at a first sampling frequency. The preset standby threshold is preferably 2 seconds, and the first sampling frequency is preferably 10Hz.

[0091] After the MCU controller enters low-power standby mode, it polls the mean characteristic M of the multiple raw signals from all sensing units at a first sampling frequency. When the mean characteristic of the multiple raw signals from any sensing unit deviates from the current baseline by more than a preset wake-up threshold, an interrupt is immediately triggered to wake up the MCU controller, which then switches to the second sampling frequency to acquire signals. The response time for the MCU to complete the transition from standby to full-speed acquisition is less than or equal to 5ms. The preset wake-up threshold is determined based on a multiple of the sensor's noise floor standard deviation, preferably 3 times, approximately 8MV. The second sampling frequency is preferably 1kHz, but it can also be set to the default frequency.

[0092] To avoid frequent switching in critical states, this embodiment introduces a hysteresis protection mechanism. Based on a preset wake-up threshold, a preset switching threshold is set; preferably, this preset switching threshold is half of the preset wake-up threshold. This ensures the stability of the MCU control system during multiple state transitions, such as the sliding window's sampling frequency switching from a first sampling frequency to a second sampling frequency and then back to the first sampling frequency. It should be noted that the preset switching threshold here is an updated preset silence threshold.

[0093] Preferably, the filtering of the multiple digital signals, the calculation of the three-dimensional force rough value, and the correction of the three-dimensional force rough value are all performed using a fixed-point number format with signed integer calculations.

[0094] This embodiment addresses the low-power, offline requirements of wearable devices and portable smart terminals. It employs a full-process optimization strategy for edge-side computation, simplifying the method at the instruction level and eliminating redundant calculation steps to adapt to the computing power architecture of low-cost MCUs such as STM32, ESP32, and 51 microcontrollers. Specifically, the method uses fixed-point arithmetic throughout. Particularly in the processes of filtering multiple digital signals, calculating the coarse three-dimensional force value, and correcting the coarse three-dimensional force value, the computational power loss of floating-point calculations is avoided, significantly improving the edge-side computation speed. The fixed-point number used in this embodiment is preferably in Q format, specifically Q8.8 format, with the quantization formula as follows:

[0095] The scaling factor Q = 2^8 = 256.

[0096] The inverse quantization formula is:

[0097] Specifically, in the forward propagation formula of the first fully connected neural network, each parameter is in Q8.8 format, the product of the parameters is in Q16.16 format, and the sum needs to be right-shifted by 8 bits to truncate it back to Q8.8 format before being added to the remaining parameters to ensure that the final output is in Q8.8 format. The entire process requires only integer multiplication and shift operations, without the need for a floating-point unit, and is well-compatible with low-cost MCUs such as Cortex-M0 / M3 that lack a hardware FPU. The fixed-point representation of the aforementioned activation function is as follows:

[0098] The above expression is directly equivalent in the fixed-point domain as: If If the result is zero, set it to 0; otherwise, leave it unchanged. This process requires no multiplication or division operations, only a single integer comparison and conditional assignment instruction. Actual testing shows that the output deviation between fixed-point and floating-point inference results is less than or equal to 0.5%, having no substantial impact on the final decoupling accuracy of the three-dimensional force. Through the above optimization, hardware degradation calculation adaptation is achieved at the algorithm level, enabling the algorithm to run entirely on low-cost microcontrollers without floating-point units, greatly satisfying the ultra-low-cost control requirements of flexible intelligent electronic skin.

[0099] The above methods have yielded the following beneficial technical effects: 1) By organically combining signal cleaning and preprocessing, coarse calculation and extraction with dynamic correction mode, the problem of unbalanced accuracy, stability and computing power consumption when array-type flexible sensors are decoupled at the end side three-dimensional force is solved.

[0100] 2) In the preprocessing stage, adaptive correction of temperature drift, suppression of inter-channel crosstalk, and filtering of power frequency noise ensure the signal-to-noise ratio of the physical signal from the source. In the decoupling stage, real-time adaptive switching between "high precision - heavy computing power" and "low computing power - high efficiency" correction modes achieves a synergistic effect of saving computing resources under high force values ​​and ensuring decoupling accuracy under minute torque changes. This ensures excellent real-time decoupling performance and anti-interference capabilities with extremely low computing power consumption, making it ideal for low-cost, resource-constrained embedded deployments.

[0101] Example 2

[0102] Based on the principles described in Example 1, a three-dimensional force decoupling system 100 for flexible sensors based on a lightweight hybrid algorithm is also proposed, such as... Figure 5 As shown, it includes: The first unit 110 is used to collect multiple raw signals output by each electrode in the target sensing unit of the flexible sensor and convert them into multiple digital signals. The second unit 120 is used to perform baseline correction based on adaptive baseline adjustment, neighborhood crosstalk suppression based on neighborhood signal correlation analysis model, and filtering processing based on adaptive filtering architecture on the multiple digital signals respectively, so as to filter the multiple digital signals and obtain cleaned multiple signals. The third unit 130 is used to extract the core features of the cleaned multi-channel signal and calculate the three-dimensional force rough value based on the core features, wherein the core features include mean features, difference features and gradient features. The fourth unit 140 is used to determine the current correction mode based on the real-time resultant force amplitude of the target sensing unit at the current moment. The fifth unit 150 is used to correct the coarse calculation value of the three-dimensional force based on a determined correction mode, and to perform sliding mean smoothing filtering on the corrected force value to output the decoupled final three-dimensional force.

[0103] Example 3

[0104] Furthermore, an electronic device is proposed, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement a three-dimensional force decoupling method based on a lightweight hybrid algorithm in Embodiment 1 when executing executable instructions.

[0105] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0106] Furthermore, it should be noted that the use of terms such as "first," "second," and "a" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified. The terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two elements or the interaction between two elements, unless otherwise explicitly specified. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0107] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0108] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A three-dimensional force decoupling method for flexible sensors based on a lightweight hybrid algorithm, characterized in that, include: The system collects multiple raw signals output from each electrode within the target sensing unit of the flexible sensor and converts them into multiple digital signals. The multiple digital signals are subjected to baseline correction based on adaptive baseline adjustment, neighborhood crosstalk suppression based on neighborhood signal correlation analysis model, and filtering processing based on adaptive filtering architecture to filter the multiple digital signals and obtain cleaned multiple signals. The core features of the cleaned multi-channel signals are extracted, and the three-dimensional force rough value is calculated based on the core features. The core features include mean features, difference features, and gradient features. The current correction mode is determined based on the real-time resultant force amplitude of the target sensing unit at the current moment; Based on the determined correction mode, the coarsely calculated three-dimensional force value is corrected, and the corrected force value is smoothed by sliding mean filtering to output the decoupled final three-dimensional force.

2. The method according to claim 1, characterized in that, Baseline correction based on adaptively adjusted baselines includes: Calculate the rate of change of variance of the multiple digital signals within the current sliding window; The window length of the sliding window is dynamically adjusted based on the comparison between the variance change rate and the preset change rate threshold. Based on the number of sampling points in the adjusted sliding window, calculate the current window mean and window standard deviation to determine the current dynamic threshold range; In response to the current sampling point falling within the dynamic threshold range, the baseline is updated using an exponentially weighted moving average. The drift error is corrected by subtracting the updated baseline from the current multi-channel digital signal.

3. The method according to claim 1, characterized in that, Neighborhood crosstalk suppression based on neighborhood signal correlation analysis models includes: Centered on the target sensing unit, acquire the signals of the target sensing unit and its eight neighboring sensing units within the sliding window period; Calculate the Pearson correlation coefficient between the target sensing unit and each adjacent sensing unit, and normalize the influence weight of the adjacent sensing units based on the Pearson correlation coefficient to construct the neighborhood influence weight matrix. The crosstalk component estimate is obtained by weighting and summing the weights of each neighborhood influence in the neighborhood influence weight matrix with the multi-channel digital signals of the corresponding adjacent sensing units. The crosstalk component is removed by subtracting the crosstalk component estimate from the signal of the target sensing unit.

4. The method according to claim 1, characterized in that, Filtering processes based on adaptive filtering architecture include: The multi-channel digital signals are input to the front-end power frequency notch filter to filter out power frequency interference. The dead zone threshold is set based on the noise standard deviation of the multiple digital signals; The error term of the subsequent minimum mean square adaptive filter is ternarily quantized using the dead zone threshold. In response to the absolute value of the error term before quantization being greater than the dead zone threshold, the weight coefficients of the adaptive filter are updated using the error term after ternary quantization. In response to the absolute value of the error term before quantization being less than or equal to the dead zone threshold, the current adaptive filter weight coefficient update is frozen.

5. The method according to claim 1, characterized in that, The current correction mode is determined based on the real-time resultant force amplitude of the target sensing unit at the current moment, including: The real-time resultant force amplitude is calculated using the aforementioned three-dimensional force coarse calculation value; In response to the real-time resultant force amplitude of the target sensing unit being less than a preset force amplitude threshold, the first correction mode is determined; In response to the real-time resultant force amplitude of the target sensing unit being greater than or equal to the preset force amplitude threshold, a second correction mode is determined; The decoupling accuracy of the first correction mode is greater than that of the second correction mode, and the single-frame decoupling time in the first correction mode is greater than that in the second correction mode.

6. The method according to claim 5, characterized in that, Based on the determined correction mode, the coarsely calculated three-dimensional force values ​​are corrected, including: In the first correction mode, the core features and the coarse calculation value of the three-dimensional force are input into the trained first fully connected neural network, and the corresponding compensation amount of the three-dimensional force is output. The coarse calculation value of the three-dimensional force is added to the compensation amount to obtain the decoupled three-dimensional force of the target sensing unit. In the second correction mode, the coarse three-dimensional force value is linearly corrected according to a preset linear correction coefficient to obtain the decoupled three-dimensional force of the target sensing unit.

7. The method according to any one of claims 1-6, characterized in that, The acquisition of the multiple raw signals also includes an intermittent adaptive sampling step: Periodically calculate the peak-to-peak value of the original signals of all sensing units within the sliding window; In response to the fact that the peak-to-peak value of all sensing units is less than the preset silence threshold and the duration is greater than the preset standby threshold, the signal is acquired at the first sampling frequency. When the signal is acquired at the first sampling frequency, in response to the mean feature of the multiple raw signals of any sensing unit deviating from the baseline by more than a preset wake-up threshold, the sampling frequency is switched to the second sampling frequency to acquire the signal. Specifically, a hysteresis protection mechanism is used to switch between the first sampling frequency and the second sampling frequency. The preset switching threshold for switching back from the second sampling frequency to the first sampling frequency is less than the preset wake-up threshold for switching from the first sampling frequency to the second sampling frequency.

8. The method according to claim 7, characterized in that, The filtering of the multiple digital signals, the calculation of the three-dimensional force rough value, and the correction of the three-dimensional force rough value are all performed using a fixed-point number format with signed integer calculations.

9. A three-dimensional force decoupling system for flexible sensors based on a lightweight hybrid algorithm, characterized in that, include: The first unit is used to collect multiple raw signals output by each electrode in the target sensing unit of the flexible sensor and convert them into multiple digital signals. The second unit is used to perform baseline correction based on adaptive baseline adjustment, neighborhood crosstalk suppression based on neighborhood signal correlation analysis model, and filtering processing based on adaptive filtering architecture on the multiple digital signals respectively, so as to filter the multiple digital signals and obtain cleaned multiple signals. The third unit is used to extract the core features of the cleaned multi-channel signals and calculate the three-dimensional force rough value based on the core features, wherein the core features include mean features, difference features and gradient features. The fourth unit is used to determine the current correction mode based on the real-time resultant force amplitude of the target sensing unit at the current moment. The fifth unit is used to correct the coarsely calculated three-dimensional force value based on a determined correction mode, and to perform sliding mean smoothing filtering on the corrected force value to output the decoupled final three-dimensional force.

10. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the three-dimensional force decoupling method for flexible sensors based on a lightweight hybrid algorithm as described in any one of claims 1-8 when executing executable instructions.

Images