A multi-modal handwriting dynamic feature extraction and authentication system

Abstract

The application relates to the field of biometric identification technology and discloses a multi-modal handwriting dynamic feature extraction and identification system. The system synchronously acquires macro kinematic sequences and micro acoustic emission sequences, performs time axis regularization based on a bottom-layer hardware interrupt signal, extracts instantaneous geometric singular points of the macro kinematic sequences, combines stress checking to intercept effective micro windows, performs phase space line integral calculation in the effective micro windows to extract macro dissipation work parameters, applies phase compensation to the micro acoustic emission sequences based on corresponding physical coordinates to realize beam forming, acquires micro friction frequency response chirp rates, fuses the dissipation work parameters and the chirp rates to generate high-dimensional identification feature vectors, and inputs the high-dimensional identification feature vectors into a classifier to output a judgment result. The application quantizes a real dynamic physical work process of writing, effectively resists high-precision mechanical arm impersonation attacks, overcomes the defects that a single-channel acoustic acquisition is susceptible to bottom noise and dispersion interference, and improves identification stability in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of biometric recognition technology, specifically a multimodal handwriting dynamic feature extraction and authentication system. Background Technology

[0002] Handwriting, as an important biometric feature, has extremely high practical application value in fields such as identity authentication, financial payment, and forensic identification. With the popularization of digital devices, traditional static image visual comparison is gradually evolving into dynamic handwriting anti-counterfeiting technology. Its core lies in capturing the writing action data of the signer on the electronic screen to confirm the true identity. This verification method based on dynamic time-series information has become an important barrier to protect the security of underlying information.

[0003] Existing dynamic handwriting anti-counterfeiting systems typically rely on touch panels or dedicated digital writing tablets to continuously extract the two-dimensional planar coordinates, movement speed, and discrete vertical downward stress of the pen tip during the user's writing process. To expand the feature dimensions, some traditional solutions add an additional single-channel microphone at the hardware level to synchronously record the microscopic acoustic waveforms generated by the friction between the pen tip and the screen. In this architecture, the system algorithm often performs isolated feature extraction on the macroscopic kinematic trajectory and the microscopic acoustic signal, and finally directly submits these conventional time-frequency domain features to a classification model for authenticity determination.

[0004] However, existing anti-counterfeiting verification technologies only focus on surface kinematic features, lacking continuous quantification of the unique phase-space physical work process during the exertion of human hand muscles and bones. This makes the system vulnerable to machine-generated forgery attacks. Existing solutions lack spatial multi-point collaboration and dispersion compensation mechanisms, resulting in severe phase distortion of extracted audio features, significantly weakening recognition stability in complex environments. The inherent sampling rate difference between macroscopic trajectories and microscopic audio makes traditional software-based polling acquisition mechanisms prone to random time delays, hindering accurate alignment of underlying data. Furthermore, the direct linear splicing of heterogeneous and dimensionless physical features fails to construct inherent nonlinear mapping constraints across modalities, leading to disordered topological distribution in the feature space. This makes it difficult for classification models to fit clear decision boundaries, ultimately resulting in a high misjudgment rate when collecting genuine handwriting. Therefore, this invention provides a multimodal handwriting dynamic feature extraction and authentication system to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multimodal handwriting dynamic feature extraction and authentication system. It solves the problems of traditional handwriting anti-counterfeiting technologies based on macroscopic two-dimensional trajectories and discrete pressure being susceptible to robotic arm trajectory imitation attacks and having difficulty quantifying the dynamic physical work process during actual writing; and the problem that anti-counterfeiting schemes that introduce single-channel audio are susceptible to interference from environmental background noise and mechanical operating noise due to the lack of spatial multi-channel collaboration and dispersion compensation mechanisms, leading to an increased false rejection rate during normal equipment acquisition.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a multimodal handwriting dynamic feature extraction and authentication system, based on a hardware carrier including a microcontroller unit, a touch panel, and a piezoelectric ceramic sensor array. The system includes: The signal acquisition and alignment module is used to synchronously acquire the macroscopic kinematic sequence and the microscopic acoustic emission sequence during the writing process through the touch panel and the piezoelectric ceramic sensor array, and to perform time axis normalization and alignment of the macroscopic kinematic sequence and the microscopic acoustic emission sequence based on the underlying hardware interrupt signal captured by the microcontroller unit. The micro-window anchoring module is used to calculate the instantaneous curvature of the macroscopic kinematic sequence to extract the instantaneous geometric singularity. Combined with the stress continuity verification condition, it extracts the effective micro-window on the time axis with the instantaneous geometric singularity as the center and discards the invalid window below the stress threshold. The dissipated work analysis module is used to perform interpolation sampling and time differentiation on the macroscopic kinematic sequence within the effective micro-window, complete the phase space line integral calculation along the open interval trajectory of the time window, and extract the macroscopic dissipated work parameters. The beamforming analysis module is used to extract the macroscopic physical coordinates corresponding to the effective micro-window to calculate the absolute physical distance difference from the sound source to all piezoelectric ceramic sensors. It calls the dispersion wavenumber curve of the touch panel medium and applies a phase rotation compensation factor based on the absolute physical distance difference in the frequency domain to achieve spatial beamforming. It extracts the peak value of the first time derivative of the main frequency energy ridge of the superposition matrix to obtain the micro-friction frequency response chirp rate. The feature mapping and counterfeit detection module is used to calculate the relationship between macroscopic dissipation power parameters and microscopic friction frequency response chirp rate to generate a cross-domain modulation index. It summarizes the extracted feature parameters and fuses them with global pen movement parameters to generate a high-dimensional counterfeit detection feature vector. The high-dimensional counterfeit detection feature vector is input into the support vector machine classifier to output the authenticity judgment result.

[0007] Preferably, the signal acquisition and alignment module is further divided into a macroscopic signal acquisition unit, a microscopic signal acquisition unit, and a clock alignment unit; the macroscopic signal acquisition unit continuously scans the surface of the touch panel screen at a first sampling rate to generate a macroscopic kinematic sequence containing two-dimensional plane absolute physical coordinates and vertical compressive stress; the microscopic signal acquisition unit concurrently acquires acoustic amplitude data of all channels of the array at a second sampling rate to generate a microscopic acoustic emission sequence.

[0008] Furthermore, the clock alignment unit connects the interrupt pin of the touch panel output pressure change threshold to the external hardware interrupt pin of the microcontroller unit; when the pressure value exceeds the preset non-zero trigger threshold in the initial stage of the touch, an edge transition level signal is generated, the system suspends the background task and enters the interrupt service subroutine to read the system clock count value, marks it as the global absolute reference zero point, and resets the timestamps of the macroscopic kinematic sequence and the microscopic acoustic emission sequence with the global absolute reference zero point as the starting point.

[0009] Preferably, the micro-window anchoring module includes a geometric singularity extraction unit, an initial interval interception unit, and a stress conservation verification unit; the geometric singularity extraction unit performs Gaussian smoothing filtering on the original two-dimensional plane discrete coordinate points, and calculates the instantaneous geometric curvature at each sampling time based on the smoothed spatial difference.

[0010] Furthermore, the geometric singularity extraction unit sets the instantaneous curvature to zero when the instantaneous scalar velocity is lower than the preset pause noise threshold, and uses non-maximum suppression logic to retain only the extreme point with the largest curvature value as an independent instantaneous geometric singularity within a time interval less than the preset physiological limit cycle.

[0011] Preferably, the stress conservation verification unit extracts the minimum compressive stress of all sampling points within the initial interval. If the minimum compressive stress is less than the pre-calibrated effective physical friction threshold, the interval is marked as invalid and discarded; otherwise, the interval is retained as the effective micro-window.

[0012] Furthermore, the dissipated work analysis module includes a data interpolation and differentiation unit and a dissipated work quantization unit; the data interpolation and differentiation unit uses a cubic spline interpolation algorithm to perform time-domain densification processing on the data within the effective micro-window and calculates the absolute amplitude of the tangential acceleration; the dissipated work quantization unit uses a numerical integration algorithm to perform phase space line integration, and its continuous calculation formula for the macroscopic dissipated work parameter is expressed as: ; In the formula, For the first The macroscopic dissipation power parameter calculated within an effective micro-window; This is the timestamp of the center moment of the micro-window; The one-sided offset time set when capturing the micro-window; This is the continuous function of the compressive stress generated through interpolation; It is a continuous function of the absolute magnitude of tangential acceleration; Representing time variables The differential.

[0013] Preferably, the beamforming analysis module includes a spatial distance calculation unit, a beamforming compensation unit, and a chirp rate extraction unit; the spatial distance calculation unit selects the sensor with the smallest absolute spatial coordinate difference from the sound source as the reference channel, and calculates the absolute physical distance difference of all other channels relative to the reference channel.

[0014] Furthermore, the beamforming compensation unit uses a Hamming window as the analysis window for short-time Fourier transform to convert the microscopic acoustic emission sequence into a two-dimensional time-frequency domain matrix, and independently applies a phase rotation compensation factor based on the absolute physical distance difference to different frequency sub-bands to complete coherent superposition.

[0015] Preferably, the chirp rate extraction unit extracts the frequency coordinates of the energy amplitude reaching the global maximum value in the superposition matrix and connects them to form the main frequency energy ridge. After smoothing the main frequency energy ridge using a median filtering algorithm, the absolute peak value of its first-order time derivative within the current micro-window is retrieved as the micro-friction frequency response chirp rate.

[0016] Furthermore, the feature mapping and counterfeit detection module includes a nonlinear mapping unit, and the equivalent formula for calculating the cross-domain modulation index by the nonlinear mapping unit is expressed as: ; In the formula, Indicates that for the first The cross-domain modulation index is calculated using an effective micro-window; This refers to the macroscopic dissipation power parameter within this window; The chirp rate of the micro-friction frequency response within this window; To match the constant coefficients of the units calibrated based on the positive sample set; This is a preset zero-constant parameter; It is the natural logarithm function.

[0017] Preferably, the feature mapping and anti-counterfeiting module includes a feature vector generation unit and a classification decision unit; the feature vector generation unit summarizes the cross-domain modulation index of all micro-windows, extracts the mean and variance, and horizontally concatenates them with the global average writing pressure and total pen stroke time, and performs standardized preprocessing by preset statistical mean and standard deviation to generate standard feature vectors; the classification decision unit constructs a single-class support vector machine with radial basis kernel function, calculates the topological distance in the mapping space and compares it with the judgment threshold to output the identification status.

[0018] A second aspect of the present invention provides a method for multimodal handwriting dynamic feature extraction and authentication, the method comprising: The macroscopic kinematic sequence and microscopic acoustic emission sequence during the writing process are synchronously acquired through the touch panel and piezoelectric ceramic sensor array, and the time axis is normalized and aligned based on the underlying hardware interrupt signal captured by the microcontroller unit. The instantaneous curvature of the macroscopic kinematic sequence is calculated to extract the instantaneous geometric singularity. The effective micro-window is then truncated with the instantaneous geometric singularity as the center, based on the stress continuity verification condition. Within the effective micro-window, interpolation sampling and time differentiation are performed on the macroscopic kinematic sequence, and phase space linear integration is performed along the open interval trajectory of the time window to extract the macroscopic dissipated work parameter. Extract the macroscopic physical coordinates corresponding to the effective micro-window and calculate the absolute physical distance difference from the sound source to all piezoelectric ceramic sensors. Call the dispersion wavenumber curve of the medium and apply a phase rotation compensation factor in the frequency domain to achieve spatial beamforming. Extract the peak value of the first time derivative of the main frequency energy ridge of the superposition matrix to obtain the micro-friction frequency response chirp rate. Based on the nonlinear mapping formula, the macroscopic dissipation power parameter and the microscopic friction frequency response chirp rate are calculated to generate the cross-domain modulation index within the micro-window. The feature parameters are summarized and fused with the global pen movement parameters to generate a high-dimensional anti-spoofing feature vector. The high-dimensional anti-spoofing feature vector is input into the classifier to output the judgment result.

[0019] Preferably, in the step of acquiring the macroscopic kinematic sequence and the microscopic acoustic emission sequence and performing time axis normalization and alignment, the macroscopic kinematic sequence with a first sampling rate and the microscopic acoustic emission sequence with a second sampling rate are acquired concurrently by the touch panel and the piezoelectric ceramic sensor array, wherein the first sampling rate and the second sampling rate are not equal; the edge jump level signal generated by the pressure stress breaking through the minimum non-zero threshold during the touch stage is monitored by the microcontroller unit to lock the global absolute reference zero point and complete the alignment.

[0020] Furthermore, in the step of calculating the instantaneous curvature of the macroscopic kinematic sequence to extract instantaneous geometric singularities and extracting effective micro-windows in combination with stress continuity verification conditions, the sequence containing two-dimensional plane absolute physical coordinates is subjected to Gaussian smoothing filtering and spatial difference operation to obtain instantaneous geometric curvature. Non-maximum suppression logic is used to discard redundant extreme points within a single force application cycle to extract independent instantaneous geometric singularities, and the stress continuity verification is performed by comparing the minimum downward compressive stress within the interval with the physical friction threshold.

[0021] Preferably, in the step of extracting the macroscopic dissipative work parameter, cubic spline interpolation is used within an effective micro-window to improve the time resolution of discrete data, the absolute amplitude of tangential acceleration is calculated based on the first and second time derivatives of the interpolated reconstructed coordinates, and the macroscopic dissipative work parameter is obtained by performing time step accumulation on the absolute amplitude of tangential acceleration and the continuous sequence of compressive stress using a numerical integration algorithm.

[0022] Furthermore, in the step of obtaining the micro-friction frequency response chirp rate, the waveform is converted into a time-frequency matrix using short-time Fourier transform, and the compensation angle is calculated in the frequency domain for the dispersion wavenumber curves of the independently combined medium at different frequency sub-bands. The phase of the remaining channels is aligned using the closest channel as the time reference. The energy maximum coordinates are connected in the superimposed spectrum, and after smoothing filtering, the differential solution is performed to obtain the chirp rate.

[0023] Preferably, in the step of generating a high-dimensional anti-counterfeiting feature vector and outputting the judgment result, the cross-domain modulation index within the micro-window is calculated using a logarithmic function carrying a zero-prevention constant parameter. The average value and variance parameters of all effective micro-window modulation indices are extracted, and these parameters are concatenated with the global average writing pressure and total pen stroke time of the total sequence. After standardization using the standard deviation zero-prevention division method based on the positive sample set parameters, the results are fed into a single-class support vector machine to output the true / false results.

[0024] This invention provides a multimodal handwriting dynamic feature extraction and authentication system. It has the following beneficial effects: 1. This invention uses low-level hardware interrupt signals to perform time-axis normalization on macroscopic kinematic sequences and microscopic acoustic emission sequences, and uses instantaneous geometric singularities and stress conservation verification conditions to extract effective micro-windows. This mechanism eliminates interference from background noise and invalid friction intervals during the writing pause phase, ensuring the data validity of subsequent multimodal feature extraction and avoiding the occupation of system computing resources by non-critical data.

[0025] 2. This invention quantifies the actual physical work process by calculating dissipative work parameters through phase space line integrals at the macroscopic level. Simultaneously, at the microscopic level, it introduces spatial beamforming technology based on the absolute physical distance difference and dispersion wavenumber curves to extract the frictional frequency response chirp rate. This two-dimensional physical parameter extraction method can identify the dynamic differences when a high-precision robotic arm imitates pen strokes and overcomes the shortcomings of traditional single-channel acoustic acquisition, which is susceptible to environmental noise interference, thus improving the noise resistance of the feature in complex environments.

[0026] 3. This invention utilizes a nonlinear logarithmic function carrying an anti-zero constant parameter to fuse macroscopic dissipation work parameters with microscopic friction frequency response chirp rate to generate a cross-domain modulation index. This index is then combined with global pen stroke parameters and anti-zero division standardization preprocessing to construct a high-dimensional anti-counterfeiting feature vector. This feature mapping logic eliminates dimensional differences between different physical modes and optimizes the topological distribution of feature data in the mapping space, thereby improving the boundary partitioning ability of a single-class support vector machine and reducing the false rejection rate of normal writing samples while maintaining anti-counterfeiting security. Attached Figure Description

[0027] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a timing diagram of a signal acquisition and alignment module according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the phase change microwindow anchoring logic according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the phase space line integration operation logic according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the prior guidance dispersion compensation beamforming logic according to an embodiment of the present invention; Figure 7 This is a schematic diagram of cross-modal feature fusion and classification decision logic according to an embodiment of the present invention; Figure 8 The diagram shows the multimodal anti-counterfeiting performance and feature distribution of the present invention, where (a) is a comparison diagram of ROC curves of the test set, and (b) is a cross-modal two-dimensional feature distribution and boundary profile diagram.

[0028] Among them, 10 is the signal acquisition and alignment module; 20 is the micro-window anchoring module; 30 is the power dissipation analysis module; 40 is the beamforming analysis module; and 50 is the feature mapping and counterfeit detection module. Detailed Implementation

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] See attached document Figure 1 , Figure 1 This is a system architecture diagram according to an embodiment of the present invention. The present invention provides a multimodal handwriting dynamic feature extraction and authentication system, which relies on a hardware carrier including a microcontroller unit, a touch panel, and a piezoelectric ceramic sensor array. The multimodal handwriting dynamic feature extraction and authentication system of the present invention may include: The signal acquisition and alignment module 10 is used to synchronously acquire the macroscopic kinematic sequence and microscopic acoustic emission sequence during the writing process through the touch panel and the piezoelectric ceramic sensor array, and to perform time axis normalization and alignment of the macroscopic kinematic sequence and microscopic acoustic emission sequence based on the underlying hardware interrupt signal captured by the microcontroller unit.

[0031] The micro-window anchoring module 20 is used to calculate the instantaneous curvature of the macroscopic kinematic sequence to extract the instantaneous geometric singularity. Combined with the stress continuity verification condition, it extracts the effective micro-window on the time axis with the instantaneous geometric singularity as the center and discards the invalid windows below the stress threshold.

[0032] The dissipated work analysis module 30 is used to perform interpolation sampling and time differentiation on the macroscopic kinematic sequence within the effective micro-window, complete the phase space line integral calculation along the open interval trajectory of the time window, and extract the macroscopic dissipated work parameters.

[0033] The beamforming analysis module 40 is used to extract the macroscopic physical coordinates corresponding to the effective micro-window to calculate the absolute physical distance difference from the sound source to each piezoelectric ceramic sensor. It calls the dispersion wavenumber curve of the touch panel medium to apply a phase rotation compensation factor based on the physical distance difference in the frequency domain to achieve spatial beamforming. It extracts the peak value of the first-order time derivative of the main frequency energy ridge of the superposition matrix to obtain the micro-friction frequency response chirp rate.

[0034] The feature mapping and anti-counterfeiting module 50 is used to calculate the relationship between macroscopic dissipation power parameters and microscopic friction frequency response chirp rate to generate a cross-domain modulation index, summarize and extract feature parameters and fuse them with global pen movement parameters to generate a high-dimensional anti-counterfeiting feature vector, and input the high-dimensional anti-counterfeiting feature vector into the support vector machine classifier to output the authenticity judgment result.

[0035] See attached document Figure 2 , Figure 2 This is a flowchart of a method according to an embodiment of the present invention. The present invention provides a multimodal handwriting dynamic feature extraction and authentication method, comprising the following steps: S10 synchronously acquires the macroscopic kinematic sequence and microscopic acoustic emission sequence during the writing process through the touch panel and piezoelectric ceramic sensor array, and performs time axis regularization and alignment based on the underlying hardware interrupt signal captured by the microcontroller unit.

[0036] S20 calculates the instantaneous curvature of the macroscopic kinematic sequence to extract the instantaneous geometric singularity, and combines the stress continuity verification condition to extract an effective micro-window centered on the instantaneous geometric singularity.

[0037] S30 performs interpolation sampling and time differentiation on the macroscopic kinematic sequence within the effective micro-window, completes phase space linear integration calculation along the open interval trajectory of the time window, and extracts macroscopic dissipative work parameters.

[0038] S40 extracts the macroscopic physical coordinates corresponding to the effective micro-window and calculates the absolute physical distance difference between the sound source and each piezoelectric ceramic sensor. It calls the dispersion wavenumber curve of the medium and applies a phase rotation compensation factor in the frequency domain to achieve spatial beamforming. It extracts the peak value of the first-order time derivative of the main frequency energy ridge of the superposition matrix to obtain the micro-friction frequency response chirp rate.

[0039] S50 calculates the macroscopic dissipation power parameter and the microscopic friction frequency response chirp rate based on the nonlinear mapping formula to generate the cross-domain modulation index within the micro-window. It summarizes the feature parameters and fuses them with the global pen movement parameters to generate a high-dimensional anti-spoofing feature vector. The high-dimensional anti-spoofing feature vector is then input into the classifier to output the judgment result.

[0040] See attached document Figure 3 , Figure 3 This is a timing diagram of a signal acquisition and alignment module 10 according to an embodiment of the present invention. In this embodiment, the signal acquisition and alignment module 10 is used to perform the operation of step S10 described above, and its internal components can be further divided into a macroscopic signal acquisition unit, a microscopic signal acquisition unit, and a clock alignment unit.

[0041] To construct a foundational dataset characterizing the dynamic features of handwriting, the system needs to concurrently acquire macroscopic sequences containing spatial information and microscopic sequences containing high-frequency physical feedback. Specifically, the macroscopic signal acquisition unit relies on the touch panel to obtain handwriting motion data. As a preferred approach, the touch panel embeds a touch driver chip to continuously scan the screen surface at a set first sampling rate.

[0042] When the writing terminal contacts the panel, the macroscopic signal acquisition unit acquires the absolute spatial coordinates and the downward pressure perpendicular to the panel at each sampling moment, thereby generating a macroscopic kinematic sequence. The first sampling rate is set to... To fully reproduce the physical movements of brushstrokes, the actual values ​​must meet the following requirements. Technical conditions. Based on general physical principles, a higher spatial sampling rate can more accurately capture the minute changes in coordinates and pressure caused by the slight tremors of the wrist muscles during pen strokes, providing continuous and smooth data support for subsequent calculations of dissipated work. The macroscopic kinematic sequence is represented as: ; In the formula, and Indicates the touch panel in the first Each sampling time Output two-dimensional plane absolute physical coordinates; Indicates the first The vertical compressive stress value output at each sampling time; This represents the total number of macroscopic sampling points during a continuous writing process. This represents a macroscopic kinematic sequence.

[0043] During the same time period of acquiring the macroscopic kinematic sequence, the microscopic signal acquisition unit obtains the high-frequency microscopic acoustic emission sequence through a piezoelectric ceramic sensor array attached to the edge of the panel. From a physical acoustics perspective, the friction of the pen tip against the panel medium excites elastic waves inside the solid; the piezoelectric ceramic sensor array receives these elastic waves and generates a weak electrical signal. The microscopic signal acquisition unit is equipped with a multi-channel audio analog-to-digital converter to concurrently acquire acoustic amplitude data from the array channels at a second sampling rate. The second sampling rate is set to... Its numerical requirements satisfy For the array contained The set of microscopic acoustic emission sequences for an independent piezoelectric ceramic sensor is represented as follows: ; In the formula, The physical channel numbering of the piezoelectric ceramic sensor in the array; Representing the The sensor at the first Each sampling time The output acoustic amplitude quantization value; Represents a set of microscopic acoustic emission sequences; This represents the total number of microscopic sampling points during the continuous writing process. Since the first sampling rate and the second sampling rate are inconsistent, and The values ​​differ by orders of magnitude.

[0044] For the driving logic of the bottom-layer scanning coordinates and analysis of the touch panel, those skilled in the art can use a conventional capacitive array scanning protocol to implement it. The specific implementation principle is well-known in the art and will not be elaborated here. For the analog-to-digital conversion process of the piezoelectric ceramic sensor converting mechanical vibration into electrical signals, those skilled in the art can use an audio analog-to-digital conversion circuit to build it. The circuit structure design is well-known in the art and will not be elaborated here.

[0045] After acquiring the aforementioned heterogeneous dual-modal signals, eliminating the hardware scheduling delay difference caused by the transmission of different data buses is a prerequisite for achieving accurate cross-modal feature mapping. The macroscopic kinematic sequence and the microscopic acoustic emission sequence are transmitted to the microcontroller unit via their respective independent hardware bus protocols. Conventional software systems receive data based on a task polling mechanism, which generates uncontrollable operating system scheduling delay differences, causing misalignment between coordinates and sound on the time axis. To avoid algorithm mapping dead zones caused by timing misalignment, the clock alignment unit relies on the hardware-level interrupt response mechanism at the microcontroller unit's underlying layer to solve the data asynchrony problem between different buses. Specifically, in the hardware interconnection, the clock alignment unit directly connects the dedicated interrupt pin for the touch panel's output pressure change threshold to the microcontroller unit's external hardware interrupt pin.

[0046] During the initial touch phase of writing, the moment the pen tip contacts the touch panel, the downward stress in the local area suddenly increases from zero. When the clock alignment unit detects that this downward stress value exceeds a preset minimum non-zero threshold, the hardware sensing circuit immediately converts this physical contact action into an edge-transition level signal. In this embodiment, the specific value of this minimum non-zero threshold can be dynamically determined by acquiring the root mean square value of the stress noise floor of the touch panel in a contactless, static state, and multiplying it by a safety margin coefficient, for example, between 1.5 and 2.0, thereby effectively preventing false triggering of the hardware circuit caused by environmental vibration. The external hardware interrupt controller of the microcontroller captures this transition level signal, immediately suspends other low-priority background tasks within the system, and enters the interrupt service subroutine without delay.

[0047] The interrupt service routine reads the current count value of the internal system clock of the microcontroller and marks the extremely narrow time window in which the touch occurred as the global absolute reference zero point. The clock alignment unit extracts the cached data frames from the memory buffer and the subsequent new input data stream, in order to Reset the timestamp of the underlying data frame at midnight.

[0048] Considering the slight thermal drift of the crystal oscillator inside the microcontroller during a relatively long writing session, as a further preferred approach, the system repeatedly triggers the aforementioned hardware interrupt process each time the user initiates and ends a stroke. The system then uses each new physical touch point to update the local absolute time reference for the corresponding stroke sequence. This segmented hardware clock warping mechanism completely eliminates the cumulative error of crystal oscillator drift over long periods.

[0049] After processing by the clock alignment unit, the relative time variables in the macroscopic kinematic sequence Relative time variables in microscopic acoustic emission sequences All were forcibly mapped to Starting from the same absolute physical timeline, this timing warping operation ensures that subsequent processing modules can establish a precise mapping between the macroscopic kinematic sequence and the microscopic acoustic emission sequence using a unified time index.

[0050] See attached document Figure 4 , Figure 4 This is a schematic diagram of phase transition microwindow anchoring logic according to an embodiment of the present invention. In this embodiment, the microwindow anchoring module 20 is used to perform the operation of step S20 described above. The microwindow anchoring module 20 includes a geometric singularity extraction unit, an initial interval truncation unit, and a stress conservation verification unit.

[0051] The distinguishability of features during the normal, smooth pen stroke phase is limited, while regions where the trajectory direction abruptly changes better reflect the writer's unique muscle control and damping feedback. To accurately pinpoint the timing of these physical phase transitions, the geometric singularity extraction unit receives the aligned macroscopic kinematic sequence from the upstream module and performs smoothing filtering on the trajectory coordinates to calculate the instantaneous geometric curvature. Considering the high-frequency quantization noise commonly present during analog-to-digital converter sampling, as a preferred approach, the system employs a Gaussian smoothing filter algorithm to smooth the original two-dimensional discrete coordinate points. For the selection of the window function and convolution operation in the Gaussian smoothing filter, those skilled in the art can utilize conventional filtering techniques from the field of digital signal processing; the process is well-known in the field and will not be elaborated upon here.

[0052] After smoothing and denoising, the geometric singularity extraction unit calculates the instantaneous geometric curvature at each sampling moment based on the spatial difference between discrete coordinate points. The discrete calculation formula for instantaneous geometric curvature is expressed as: ; In the formula, Representative moment The corresponding trajectory curvature value; and The first-order difference value of the coordinate with respect to time represents the component of the instantaneous pen movement speed in the corresponding axis; and The second-order difference value of the coordinate with respect to time represents the component of the instantaneous pen stroke acceleration in the corresponding axis.

[0053] Considering the unavoidable physical pauses during actual writing, when the first-order difference value approaches zero during a pause, it can cause a floating-point aberration in the curvature calculation formula where the denominator is zero. To eliminate this algorithmic dead zone, the system incorporates pause elimination logic before performing curvature calculation. Specifically, the system presets a pause speed threshold that approximates the system noise floor; when the sum of squares of the first-order differences at a given moment... When the velocity is less than the threshold for stopping, the system directly sets the instantaneous curvature at that moment to zero, thereby filtering out non-phase-change stationary stagnation points.

[0054] After calculating the instantaneous curvature of the entire sequence, the system iterates through the sequence to extract local maxima points that exceed a set sharp bend threshold. In this embodiment, the specific value of the sharp bend threshold is determined by pre-statistically analyzing the lower limit of the curvature distribution of the pen stroke angle region in a large amount of valid sample data.

[0055] To avoid extracting multiple redundant maxima at the same pen stroke angle due to touch panel sampling jitter, the system introduces non-maximum suppression logic. Specifically, the system sets a time interval threshold, which typically corresponds to the minimum physiological limit cycle of a single muscle exertion. When the time interval between two detected local maxima is less than this threshold, the system retains only the maxima with the largest curvature value, discarding the others. The local maxima retained after non-maximum suppression are defined as independent instantaneous geometric singularities, and the system records the timestamp set corresponding to each singularity.

[0056] After determining the singularity where the trajectory undergoes a physical phase transition, the initial interval extraction unit extracts the required data segments on the time axis, centered on the timestamps of each instantaneous geometric singularity. For each center moment in the timestamp set, the initial interval extraction unit extends a fixed offset time towards the positive and negative sides of the time domain to define the initial interval. The value of this offset time is set based on the physical response period of elastic oscillation generated by human muscles overcoming frictional resistance, typically selecting a fixed value within the range of 20ms to 50ms to cover a complete damping decay physical process. Let the first... The timestamps corresponding to the instantaneous geometric singularities are: The bias time is The corresponding initial interval is represented as: ; In the formula, Indicated by time The system performs the above truncation operation on all recorded instantaneous geometric singularities to generate a continuous time window containing the period before and after the phase transition.

[0057] In actual feature calculations, specific handwriting habits can easily cause data gaps in the algorithm. For example, when a writer is writing continuously, the pen tip briefly leaves the screen surface due to the pen-lifting motion. However, due to the data fitting characteristics of capacitive touch panels, continuous coordinate lines are still generated at the electrical level. Within this virtual connection range where no actual solid mechanical friction occurs, the piezoelectric ceramic sensor struggles to capture effective acoustic waveforms. If data containing such ranges is directly fed downstream for cross-modal feature calculations, the lack of microscopic acoustic feature variables will cause the subsequent calculation denominator to approach zero abnormally.

[0058] To avoid the aforementioned computational dead zone, the stress conservation verification unit is configured to perform window validity verification based on stress conservation. The stress conservation verification unit traverses the downward stress sequence within each initial interval, determining whether the writing terminal maintains sufficient mechanical coupling with the touch panel during that time period. As a preferred approach, the system pre-defines an effective physical friction threshold. This threshold is set through hardware experimental calibration based on the lowest touch pressure corresponding to the elastic wave with a acceptable signal-to-noise ratio that the piezoelectric ceramic sensor can receive. The stress conservation verification unit extracts the current initial interval... The system extracts the minimum value from all downward stress sampling points and compares it with the effective physical friction threshold. If the extracted minimum value is less than the effective physical friction threshold, the system determines that the pen tip is off the ground within that time interval, marks that interval as invalid, and removes it from the processing queue.

[0059] The system determines that the writing instrument and the medium surface maintain continuous physical contact with acoustic excitation capability only when the compressive stress at all sampling times within the initial interval is greater than or equal to the effective physical friction threshold. The time period retained by the stress conservation verification logic is defined as a strongly coupled effective micro-window. The stress conservation verification unit ultimately outputs a set of filtered effective micro-windows. This filtering mechanism removes invalid data segments that are prone to causing singular values ​​in calculations, providing a physically coherent temporal data foundation for subsequent analysis modules.

[0060] See attached document Figure 5 , Figure 5 This is a schematic diagram of the phase space line integral operation logic according to an embodiment of the present invention. In this embodiment, the dissipated work analysis module 30 is used to perform the operation of step S30 described above. The dissipated work analysis module 30 includes a data interpolation and differentiation unit and a dissipated work quantization unit.

[0061] After acquiring the filtered effective micro-windows, the system further calculates the mechanical work parameters that quantify the deep muscle exertion habits. Limited by the hardware scanning frequency limit of the touch panel, the effective micro-windows with a short time span typically contain only a limited number of discrete macroscopic sampling points. To avoid truncation errors caused by directly performing high-order kinematic differentiation and integration on sparse discrete data, the data interpolation and differentiation unit receives the discrete data within the effective micro-windows and performs interpolation upsampling operations.

[0062] As a preferred approach, the system employs a cubic spline interpolation algorithm to densify the macroscopic absolute spatial coordinates and compressive stress data within the micro-window in the temporal domain. In practice, the system sets the interpolated target time resolution to a specific multiple of the original sampling period. To achieve a balance between smoothing the fitted curve and consuming the computational power of the microcontroller unit, the upsampling factor is typically set between 4 and 10 times. Cubic spline interpolation can reconstruct the spatial physical trajectory of the pen stroke relatively closely while ensuring the continuity of the first and second derivatives of the data points. For the polynomial piecewise fitting and boundary condition solution of the cubic spline interpolation algorithm, those skilled in the art can use conventional interpolation techniques in the field of numerical analysis; the specific fitting process is well-known in this field and will not be elaborated upon here.

[0063] After obtaining the high-resolution continuous sequence after interpolation and upsampling, the data interpolation differentiation unit calculates the instantaneous tangential acceleration amplitude within the micro-window through difference operations. The data interpolation differentiation unit calculates the instantaneous scalar velocity based on the first-order time derivative of the two-dimensional plane coordinates, and the formula is expressed as: ; In the formula, For a moment The instantaneous scalar velocity; and These are the first derivatives of the interpolated and reconstructed two-dimensional plane absolute physical coordinates with respect to the time variable. Based on this, the data interpolation differentiation unit further performs time differentiation on the instantaneous scalar velocity to obtain the absolute amplitude of the tangential acceleration: ; In the formula, Indicates time The absolute amplitude of the tangential acceleration, which is used to characterize the velocity change of the pen tip in the tangential direction of the pen stroke trajectory; Represents instantaneous scalar velocity The differential; Representing time variables The differential.

[0064] After acquiring high-precision kinematic parameters within the micro-window, the dissipative work quantization unit performs transient phase space line integrals to calculate the total mechanical work. Traditional kinematic feature extraction algorithms often rely on topological space rules with closed start and end points when calculating the envelope area. However, in actual writing, due to the short time span of the micro-window, the captured pen stroke trajectory usually presents an open interval shape with unconnected beginnings and ends. If the algorithm forcibly connects the beginning and end of the trajectory to perform topological integration of the closed area, it is easy to introduce spatial deformation errors, which in turn affect the subsequent counterfeit detection results.

[0065] To meet the quantization requirements of open-interval trajectories, the dissipative work quantization unit, based on dynamic principles, performs phase-space linear integration of the compressive stress and tangential acceleration along the actual open-interval trajectory within the time window. At the physical mechanism level, the compressive stress is positively correlated with the instantaneous dynamic frictional resistance of the pen tip and the medium contact surface; while the tangential acceleration, to a certain extent, reflects the magnitude of the transient driving force output by the muscle group to overcome this frictional damping. The dissipative work quantization unit performs linear integration of the product of these two factors in the time domain to extract the macroscopic dissipative work parameter. This parameter characterizes the nonlinear mechanical work consumed by a specific muscle group at the moment of phase transition. The continuous calculation formula for the macroscopic dissipative work parameter is expressed as: ; In the formula, For the first The macroscopic dissipation power parameter calculated within an effective micro-window; This is the timestamp of the center moment of the micro-window; The one-sided offset time set when capturing the micro-window; This is the continuous function of compressive stress generated through interpolation.

[0066] In specific digital signal processing steps, since continuous integration is difficult to perform directly in the microcontroller unit, the dissipated power quantization unit uses a numerical integration algorithm to equivalently calculate the aforementioned continuous integration. As one implementation method, the system employs the trapezoidal rule or Simpson's rule, using the small time interval after interpolation upsampling as the discrete integration step size, and accumulating the product of physical quantities at each discrete time point to obtain an approximate integration result. Since the pre-processed stress conservation verification unit has already filtered out the invalid intervals of the pen-lifting loop, the compressive stress function participating in the integration... The results are presented as valid values ​​greater than zero. This reduces the risk of numerical integration results abnormally approaching zero and ensures the stable operation of subsequent cross-modal feature calculation logic.

[0067] See attached document Figure 6 , Figure 6This is a schematic diagram of the prior guidance dispersion compensation beamforming logic according to an embodiment of the present invention. In this embodiment, the beamforming analysis module 40 is used to perform the operation of step S40 described above. The internal structural logic of the beamforming analysis module 40 can be divided into a spatial distance calculation unit, a beamforming compensation unit, and a chirp rate extraction unit.

[0068] After acquiring the aforementioned effective micro-window and macroscopic dynamic parameters, the system proceeds to purify and extract features from the synchronously acquired microscopic acoustic emission sequence. Typically, the acoustic signal energy generated by the friction of the pen tip is relatively weak and easily interfered with by internal hardware vibrations of the touch panel and ambient noise. To improve the signal-to-noise ratio of the microscopic signal, the spatial distance calculation unit extracts the macroscopic absolute spatial coordinates corresponding to the center of the current effective micro-window, using this as the physical location reference for the sound source. The piezoelectric ceramic sensor array's physical coordinates are fixed and calibrated at the factory, and the spatial distance calculation unit calculates the absolute physical distance from the sound source to each piezoelectric ceramic sensor based on spatial geometry.

[0069] As a preferred approach, the system selects the sensor closest to the sound source as the reference channel and calculates the physical propagation distance difference of each of the other array channels relative to the reference channel. The aforementioned distance difference parameter establishes the temporal relationship of signal arrival between different channels, providing macroscopic prior guidance for subsequent spatial filtering.

[0070] In conventional airborne acoustic wave processing, beamforming often employs a simple time-delay superposition algorithm. However, considering the propagation characteristics of sound waves in solid media such as touch panels, friction-excited elastic waves exhibit significant dispersion characteristics, meaning that the propagation phase velocities of sound waves with different frequency components differ within the solid. Directly performing translational superposition in the time domain makes it difficult to align the phases of different frequency components, easily leading to distortion of the broadband waveform. To adapt to the acoustic characteristics of solid media, the beamforming compensation unit utilizes the factory-preset dispersion wavenumber curve of the touch panel. This curve characterizes the nonlinear mapping relationship between the spatial wavenumber and frequency for a specific panel material.

[0071] To apply compensation in the frequency domain, the beamforming compensation unit performs a short-time Fourier transform on the microacoustic emission sequence of each channel, converting it into a two-dimensional time-frequency domain matrix. In practice, the system typically uses a Hamming window as the analysis window for the short-time Fourier transform, setting the window length between 128 and 512 sampling points. To ensure sufficient time resolution for subsequent ridge extraction, the overlap rate of adjacent analysis windows is usually set between 50% and 75%.

[0072] In the constructed time-frequency domain matrix, the beamforming compensation unit independently applies a phase rotation compensation factor based on the aforementioned physical distance difference for different frequency sub-bands. This calculation operation compensates for the phase shift caused by dispersion effects, aligning the acoustic emission signals of each sensor to the same phase reference across multiple frequency bands, thereby achieving spatial coherent superposition. The compensation superposition formula is expressed as: ; In the formula, This is the integrated time-frequency matrix after beamforming compensation; This represents the total number of physical channels in the piezoelectric ceramic sensor array; For the first Complex time-frequency values ​​obtained by short-time Fourier transform of the original discrete signals of each channel; The imaginary unit; To retrieve the corresponding frequency from the medium dispersion wavenumber curve Wavenumber parameters under; For the first The physical propagation distance difference between each channel and the reference channel. After frequency domain phase rotation alignment, the broadband signal energy of the target friction sound is accumulated, while the environmental random noise is weakened to a certain extent. The system thus obtains a purified time-frequency matrix with a high signal-to-noise ratio.

[0073] After obtaining the integrated time-frequency matrix, the system further quantifies the dynamic characteristics of sound frequency evolution over time. Changes in pen stroke direction or fluctuations in downward pressure alter the microscopic contact state between the pen tip and the panel, causing instantaneous drift in the dominant frequency of microscopic frictional sound generation. The chirp extraction unit traverses the integrated time-frequency matrix. The system extracts the frequency coordinates where the energy amplitude reaches the global maximum within each time slice, and connects these coordinates in chronological order to form the dominant frequency energy ridge. To mitigate the discrete jumps in the ridge caused by sudden transient noise, the system employs a median filtering algorithm to smooth the dominant frequency energy ridge. The sliding window length of the median filter can be set to 3 to 5 time steps, which helps prevent outliers from causing singular dead zones in the differentiation operation.

[0074] After smoothing, the chirp extraction unit calculates the first-order time derivative of the dominant frequency energy ridge and retrieves its absolute peak value within the current micro-window time period. This peak value is defined as the micro-friction frequency response chirp, used to quantify the severity of the frequency abrupt change. Its feature extraction formula is as follows: ; In the formula, The calculated micro-friction frequency response chirp rate; The main frequency energy ridge line is a continuous function after median filtering and smoothing. The time interval that belongs to the currently set valid micro-window; Represents the differential operator; Representing time variables The feature parameter captures minute changes by utilizing the dynamic changes of microscopic high-frequency elastic waves. It captures the microsecond-level mechanical friction state that is difficult to reflect by low-frequency macroscopic coordinates, providing in-depth physical feature support for subsequent multimodal authentication.

[0075] See attached document Figure 7 , Figure 7 This is a schematic diagram of cross-modal feature fusion and classification decision logic according to an embodiment of the present invention. In this embodiment, the feature fusion and counterfeit detection decision module 50 is used to perform the operation of step S50 above. The feature fusion and counterfeit detection decision module 50 can be internally divided into a nonlinear mapping unit, a feature vector generation unit, and a classification decision unit.

[0076] Establishing a physical correlation between macroscopic dynamics and microscopic acoustic response is fundamental to feature fusion. The nonlinear mapping unit receives the macroscopic dissipation work and microscopic frictional frequency response chirp rate corresponding to each micro-window output from the upstream module. Physically, the macroscopic dissipation work characterizes the mechanical energy output by the writer during the pen stroke phase transition, while the microscopic chirp rate reflects the acoustic frequency drift caused by the abrupt change in the stress state of the medium. Given that these two parameters belong to different physical dimensions and have a large numerical range, direct linear calculations can easily lead to an imbalance in the feature space weights. Therefore, the nonlinear mapping unit constructs a logarithmic mapping equation with a singularity-preventing constant to calculate the cross-domain modulation index.

[0077] As a preferred method, the formula for calculating the cross-domain modulation index is expressed as: ; In the formula, Indicates that for the first The cross-domain modulation index is calculated using an effective micro-window; This refers to the macroscopic dissipation work within this window; The chirp rate of the micro-friction frequency response within this window; The dimensional matching coefficient; To prevent singular value constants, a value of 10 can be taken. -6 Positive real numbers of the order of magnitude; It is the natural logarithm function.

[0078] In this embodiment, the dimension matching coefficient The specific value is determined by the ratio of the mean macroscopic dissipation work to the mean microscopic chirp rate in the statistically valid sample set, and its value is usually within 10. 2 Up to 10 4 Between. Introduce a singularity prevention constant. It effectively avoids the computational dead zone caused by the division denominator being zero when the micro-chirp rate approaches zero, while the addition operation within the parentheses ensures that the parameter of the logarithmic function is always greater than or equal to 1, making the output stable within the non-negative real number range.

[0079] After calculating the cross-domain modulation index of each local micro-window, the system needs to convert it into a fixed-length global feature suitable for machine learning models. The feature vector generation unit traverses the entire handwriting sequence, summarizes all local cross-domain modulation indices, and extracts their statistical mean. With variance The mean is used to characterize the average conversion level of energy and frequency response throughout the entire stroke process, while the variance quantifies the fluctuation of this conversion level.

[0080] The system extracts the global average writing pressure of the entire handwriting sequence. and total time spent writing The feature vector generation unit concatenates the above four parameters horizontally to generate an initial multidimensional feature vector, which is represented as: ; In the formula, This represents the generated initial multidimensional feature vector.

[0081] To avoid weight bias in subsequent model kernel function calculations due to differences in numerical magnitudes across different physical dimensions, the feature vector generation unit performs Z-score standardization preprocessing after vector concatenation. Specifically, the system pre-calculates the mean and standard deviation of these four dimensions using a valid sample set, and transforms the initial vector into a standard feature vector with an approximate mean of 0 and a variance of 1 through standardization mapping. To prevent the calculated standard deviation from approaching zero due to extremely stable user-written features or insufficient early sample size, the system introduces a small constant to prevent division by zero in the denominator of the standardized division. The specific implementation equation for single-dimensional feature preprocessing is expressed as follows: ; In the formula, For the standardized first 1D feature components; The first initial vector to be measured Original numerical value; and These are the statistical mean and standard deviation of the valid sample set along this dimension, respectively. It is usually set to 10. -8 Scale. After the above robustness processing, the output standard feature vector is formally used as the input data for the classification model. Each dimension of the standardized multidimensional feature vector corresponds to the specific physical state of local mechanical coupling, local force stability, global absolute force, and global pen stroke rhythm in the writing behavior.

[0082] After the feature vectors are generated, the classification decision unit inputs them into the classification model to output the final authenticity judgment. Considering that handwriting authentication scenarios typically only allow the user's genuine signature as a positive sample, making it difficult to obtain forged negative samples, the classification decision unit in this embodiment uses a single-class support vector machine algorithm to construct a classification decision network. The internal hierarchical structure of this network includes an input layer, a kernel mapping hidden layer, and an output decision layer. In the data flow direction, the input layer receives the preprocessed four-dimensional feature vectors, the kernel mapping hidden layer uses the radial basis function to nonlinearly map the data to a high-dimensional feature space, and the output decision layer calculates the topological distance in the high-dimensional space based on the support vectors and outputs the discrimination state.

[0083] ; In the formula, Let be the normal vector of the hyperplane; Represents the set of sample slack variables; Indicates the sequence number of the training sample; The distance offset from the hyperplane to the origin; For the first Slack variables for each sample; The total number of training samples; The key hyperparameter controls the ratio of the lower bound of support vectors to the upper bound of outlier samples, and is typically set between 0.01 and 0.1. For solving the quadratic programming problem of this objective function, those skilled in the art can use a conventional sequential minimum optimization algorithm, the solution process of which is well-known in the field and will not be elaborated here.

[0084] In the actual reasoning and authentication phase, the classification decision unit inputs the feature vector of the handwriting to be tested, after undergoing the same standardized preprocessing, into the already trained decision network. The model calculates the signed topological distance of this input feature vector to the decision hyperplane in the mapping space. The system compares this distance with a preset judgment threshold, which typically corresponds to the hyperplane boundary, i.e., a value of 0. When the calculated topological distance is greater than or equal to 0, the classification decision unit determines that the test vector is located within a high-density legitimate sample region, and the system outputs a business status indicating that the authentication result is true; conversely, if the topological distance is less than 0, it indicates that the dynamic characteristics of the handwriting have deviated from the inherent habits of legitimate users, and the system outputs an interception signal indicating that the authentication result is false.

[0085] To better understand the technical solution of this invention, the following description is based on a specific application scenario.

[0086] To eliminate physiological fluctuations caused by uncontrollable factors such as fatigue and emotions in human test samples, this application example uses a high-precision six-axis programmable robotic arm to hold a standard capacitive stylus and execute a standardized preset writing trajectory on a touch panel equipped with four piezoelectric ceramic sensors.

[0087] The robotic arm controls the stylus to contact the screen with an initial downward pressure of 1.5N. The sudden change in stress triggers an external hardware interrupt in the microcontroller unit. The clock alignment unit immediately records the current system time as the absolute zero point. Subsequently, the macroscopic acquisition unit records the two-dimensional coordinates and pressure at 240Hz, and the microscopic acquisition unit records the acoustic emission audio of four channels at 44.1kHz. Both are timestamped with the absolute zero point as the reference.

[0088] When the robotic arm reaches a 90-degree acute angle turn in the pattern, the system calculates the instantaneous curvature of the two-dimensional coordinates at that moment. Due to the sharp turn, the curvature exceeds the sharp bend threshold, and the system anchors this moment as an instantaneous geometric singularity, extending it 30ms before and after, thus capturing an initial interval of 60ms in total. The stress conservation verification unit checks the minimum stress within this 60ms interval and finds that the lowest stress is 1.1N, determining that no false connection has occurred and establishing it as a valid micro-window.

[0089] Within the aforementioned 60ms effective micro-window, the system upscales the few discrete coordinate points to a dense sequence of 1200Hz using cubic spline interpolation. The instantaneous tangential acceleration at the bend is calculated and multiplied by the continuous compressive stress sequence. Numerical integration is then performed along the actual open-interval trajectory to finally extract the macroscopic dissipative work parameter within this micro-window. This macroscopic dissipative work parameter represents the nonlinear work done by the robotic arm's servo motor to overcome damping at this point.

[0090] The system extracts the macroscopic coordinates at the bend and calculates the physical distance difference between the sound source and the four piezoelectric ceramic sensors. It then uses the dispersion wavenumber curve of the touch glass medium and applies a phase rotation factor to the four acoustic channels in the frequency domain for coherent superposition. From the purified time-frequency matrix, the system extracts the dominant frequency energy ridge and calculates its first-order time derivative to obtain the micro-friction frequency response chirp rate. .

[0091] The system calculates the cross-domain modulation index. The mean and variance of the modulation index of the entire trajectory, along with the global average pressure and total time, are concatenated into a four-dimensional vector. After Z-score standardization, this vector is input into a pre-trained single-class support vector machine using baseline driving parameters. The network calculates the topological distance from this vector to the hyperplane as +1.42, which is greater than the decision threshold of 0. The final output is: True, Allow.

[0092] Experimental dataset construction: Positive sample, real-world verification set: The robotic arm is loaded with a standard dynamic drive configuration with fixed servo stiffness and fixed damping coefficient, and continuously writes the target pattern 1000 times to simulate the user's inherent and stable physical force application habits.

[0093] Negative samples, high-fidelity forgery set: The robotic arm loads a forged drive configuration and uses machine vision feedback to perfectly replicate the two-dimensional macroscopic trajectory of the target pattern. However, it deliberately changes the internal stiffness and acceleration curve of the servo motor to simulate the scene where the forger can only imitate the shape but cannot replicate the deep muscle tremors and dynamic pressing work when copying, and writes continuously 1000 times.

[0094] Control group setup: Control Group A: Only macroscopic sampling coordinates and compressive stress are extracted, and the sequence similarity is calculated using the dynamic time warping algorithm for counterfeit detection.

[0095] Control group B: Only single-channel microacoustic waveforms are extracted, without spatial beamforming and dispersion compensation. Mel frequency cepstral coefficients are directly extracted and sent to a standard SVM for authentication.

[0096] Experimental group: The system adopts a complete architecture that integrates macroscopic dissipation power, dispersion compensation beamforming acoustic chirp rate, and cross-modal modulation.

[0097] The system conducted a blind test on a total of 2000 samples. The evaluation metrics included the false acceptance rate (FAR) of misclassifying a forgery as genuine, the false rejection rate (FRR) of misclassifying a genuine work as a forgery, and the equal error rate (EER). The test results are shown in the table below: Blind Test Comparison Results of Handwriting Authentication Performance

[0098] See attached document Figure 8 Analysis of the test data revealed that control group A experienced a false acceptance rate as high as 14.2% when faced with the trajectory replication of a high-precision robotic arm. The main reason for this phenomenon is that traditional macroscopic anti-counterfeiting methods are easily deceived by highly similar trajectories, and they only extract simple discrete pressure values, making it difficult to effectively quantify the dynamic work changes generated by the writing tool during movement.

[0099] Control group B introduced microscopic acoustic features, which improved the system's sensitivity to concealed force differences to some extent. However, due to the lack of multi-channel spatial coordination and waveform compensation mechanisms, the acoustic signals acquired by a single channel are easily interfered with by the noise from the robotic arm's operation. This environmental interference directly led to an increase in the system's normal sample false rejection rate to 11.4%, making it difficult to meet the stability requirements in practical applications.

[0100] The experimental group of this invention achieved strict time alignment of data using underlying hardware mechanisms and successfully eliminated interference from external noise, accurately extracting the dissipated power and frequency chirp rate parameters that reflect the essence of writing force. Test results show that this invention significantly reduces the overall error rate to 0.85%. Although multi-dimensional feature calculations increase the time for a single authentication to 34 milliseconds, this time overhead is well within the allowable range of real-time response settings for smart terminals. The above results fully demonstrate that the solution of this invention has excellent reliability and anti-counterfeiting capabilities when resisting high-precision machine spoofing attacks.

[0101] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multimodal handwriting dynamic feature extraction and authentication system, based on a hardware carrier including a touch panel and a piezoelectric ceramic sensor array, characterized in that, The system includes: The signal acquisition and alignment module is used to simultaneously acquire macroscopic kinematic sequences and microscopic acoustic emission sequences, and to perform time axis regularization based on the underlying hardware interrupt signal; The micro-window anchoring module is used to extract the instantaneous geometric singularities of the macroscopic kinematic sequence and combine stress verification to extract the effective micro-window; The dissipated work analysis module is used to perform integral calculations on the macroscopic kinematic sequence within the effective micro-window to extract macroscopic dissipated work parameters. The beamforming analysis module is used to apply phase compensation to the microscopic acoustic emission sequence based on the macroscopic physical coordinates corresponding to the effective micro-window to achieve beamforming, so as to obtain the microscopic friction frequency response chirp rate. The feature mapping and counterfeit detection module is used to fuse the macroscopic dissipation power parameter and the microscopic friction frequency response chirp rate to generate a high-dimensional counterfeit detection feature vector to output the authenticity determination result.

2. The multimodal handwriting dynamic feature extraction and authentication system according to claim 1, characterized in that, The signal acquisition and alignment module includes a macroscopic signal acquisition unit, a microscopic signal acquisition unit, and a clock alignment unit; The macroscopic signal acquisition unit continuously scans the surface of the touch panel screen to generate a macroscopic kinematic sequence containing two-dimensional plane absolute physical coordinates and vertical compressive stress. At the same time, the microscopic signal acquisition unit concurrently acquires acoustic amplitude data from all channels of the array to generate a microscopic acoustic emission sequence. The clock alignment unit monitors the edge transition level signal generated when the compressive stress value exceeds the preset non-zero trigger threshold in the initial stage of the touch event, reads the system clock count value and marks it as the global absolute reference zero point, and drives the macroscopic kinematic sequence and the microscopic acoustic emission sequence to perform timestamp mark reset based on the global absolute reference zero point to complete the alignment association of the underlying time.

3. The multimodal handwriting dynamic feature extraction and authentication system according to claim 1, characterized in that, The micro-window anchoring module includes a geometric singularity extraction unit, an initial interval interception unit, and a stress conservation verification unit. The geometric singularity extraction unit performs Gaussian smoothing filtering and spatial difference on the two-dimensional plane absolute physical coordinates in the macroscopic kinematic sequence, calculates the instantaneous geometric curvature at each sampling moment, and combines a preset pause noise threshold and non-maximum suppression logic to filter redundant extreme points and extract independent instantaneous geometric singularities. The initial interval extraction unit extends the offset time to both sides of the time axis with the instantaneous geometric singularity as the center to form an initial interval, drives the stress conservation verification unit to extract the minimum value of the compressive stress in the initial interval, and converts the initial interval into an effective micro-window when it is determined that the minimum value of the compressive stress reaches the pre-calibrated effective physical friction threshold.

4. The multimodal handwriting dynamic feature extraction and authentication system according to claim 1, characterized in that, The dissipation work analysis module includes a data interpolation and differentiation unit and a dissipation work quantization unit. The data interpolation and differentiation unit uses a cubic spline interpolation algorithm to perform time-domain densification processing on the discrete macroscopic kinematic sequence based on the effective micro-window interval of the micro-window anchoring module to reconstruct a continuous sequence of compressive stress, and calculates a continuous sequence of absolute amplitude of tangential acceleration based on the time derivative of the coordinates. The dissipation work quantization unit uses a numerical integration algorithm to perform phase space line integration, and accumulates and integrates the continuous sequence of downward compressive stress and the continuous sequence of absolute amplitude of tangential acceleration over time steps to construct a macroscopic dissipation work parameter that quantifies the actual work done by the pen stroke.

5. The multimodal handwriting dynamic feature extraction and authentication system according to claim 1, characterized in that, The beamforming analysis module includes a spatial distance calculation unit, a beamforming compensation unit, and a chirp rate extraction unit. The spatial distance calculation unit extracts the sensor with the smallest absolute spatial coordinate difference from the sound source as a reference channel based on the macroscopic physical coordinates corresponding to the effective micro-window, and calculates and outputs the absolute physical distance difference matrix of all channels other than the reference channel relative to the reference channel. The beamforming compensation unit uses short-time Fourier transform to convert the microscopic acoustic emission sequence into a two-dimensional time-frequency domain matrix, and calls the dispersion wavenumber curve of the touch panel medium. It then calculates and applies a phase rotation compensation factor in conjunction with the absolute physical distance difference matrix to complete the coherent superposition of each frequency sub-band to construct a superimposed spectrum matrix.

6. The multimodal handwriting dynamic feature extraction and authentication system according to claim 5, characterized in that, The chirp rate extraction unit is configured as follows: In the superimposed spectrum matrix constructed by the beamforming compensation unit, the frequency coordinates where the energy amplitude reaches the global maximum value are extracted and connected to form the main frequency energy ridge. The main frequency energy ridge is smoothed using a median filtering algorithm; For the time range of the current effective micro-window, calculate the first time derivative of the main frequency energy ridge and retrieve the absolute peak value, and establish the absolute peak value as the micro-friction frequency response chirp rate.

7. The multimodal handwriting dynamic feature extraction and authentication system according to claim 1, characterized in that, The feature mapping and anti-spoofing module includes a nonlinear mapping unit, a feature vector generation unit, and a classification decision unit. The nonlinear mapping unit synchronously acquires the macroscopic dissipation power parameter and the microscopic friction frequency response chirp rate, and uses the natural logarithm function logic carrying a preset zero-constant parameter and a dimension-matching constant coefficient to extract the nonlinear correlation attribute between the macroscopic dissipation power parameter and the microscopic friction frequency response chirp rate, and generates the cross-domain modulation index corresponding to the single effective micro-window. The feature vector generation unit summarizes the mean and variance of all extracted cross-domain modulation indices from the entire sequence, drives the splicing logic, and horizontally connects them with the global average writing pressure and total pen stroke time. After standardization preprocessing by anti-zero division, a high-dimensional anti-spoofing feature vector is constructed.

8. The multimodal handwriting dynamic feature extraction and authentication system according to claim 7, characterized in that, The classification decision unit performs the following steps: Construct a single-class support vector machine with radial basis kernel function; The high-dimensional anti-spoofing feature vector constructed by the feature vector generation unit is analyzed, and the topological distance parameter in the nonlinear mapping space is calculated. The calculated topological distance parameter is compared with a preset judgment threshold, and a Boolean true or false judgment result is directly output according to the comparison logic.

9. The multimodal handwriting dynamic feature extraction and authentication system according to claim 3, characterized in that, The geometric singularity extraction unit is configured to execute the following processing logic when extracting the instantaneous geometric singularity: The system drives the calculation of the instantaneous scalar velocity of the current coordinate data, and when the instantaneous scalar velocity is lower than the preset pause noise threshold, it forces the instantaneous geometric curvature of the corresponding sampling time to be set to zero. By introducing a time interval parameter based on the physiological limits of physical pen movement, and comparing all non-zero geometric curvature values ​​within the time interval, the instantaneous geometric singularity that excludes local jitter interference is constructed.

10. The multimodal handwriting dynamic feature extraction and authentication system according to claim 7, characterized in that, The feature vector generation unit includes the following steps when performing the anti-zero division normalization preprocessing: Extract the global statistical mean and global standard deviation parameters calculated based on the pre-calibrated positive sample set of the system; Based on the global statistical mean parameter, a centering translation operation is performed on the currently spliced ​​feature data stream, and the global standard deviation parameter is driven to perform division and scaling calculations on the translated data in combination with the minimum anti-zero constant, thereby constructing the high-dimensional anti-spoofing feature vector.

Images