Folding umbrella anti-mis-touch triggering control method based on umbrella handle holding pressure distribution sequence

Abstract

The application is based on the folding umbrella false touch trigger control method of umbrella handle holding pressure distribution sequence, relates to the technical field of pressure sensor data processing and false touch behavior discrimination, and aims to improve the false touch recognition ability of the umbrella handle holding interface. The core scheme is to synchronously collect multi-channel original pressure signals through the umbrella handle ring array pressure sensor, to extract rhythm coordination features by grouping and sliding window processing, to obtain smooth multi-channel instantaneous phase angle sequence by Hilbert transform and phase expansion algorithm, to realize high confidence discrimination of false touch behavior based on phase consistency and phase difference entropy indexes, to perform feature fusion combined with time domain pressure gradient sequence, and finally to output fine-grained false touch type and trigger false touch control by using graph convolution network classification reasoning. The method can accurately capture and identify different types of false touch behavior, and improve the reliability and adaptability of the system false touch intervention.

Description

Technical Field

[0001] This invention relates to the field of signal processing and human-computer interaction safety control technology for smart wearable devices, and in particular to a method for preventing accidental triggering of folding umbrellas based on the umbrella handle grip pressure distribution sequence. Background Technology

[0002] Current smart wearable devices, especially smart folding umbrellas, primarily focus on improving the recognition and protection against accidental touches during handholding. Mainstream technologies in the industry typically revolve around analyzing the raw signals from multi-point pressure or tactile sensors. This involves methods such as amplitude variation analysis, threshold judgment, sliding window statistics, time / frequency domain analysis, and feature extraction based on deep learning to distinguish various user operation states. Most existing products employ simple methods like raw pressure signal denoising, feature interpolation reconstruction, or end-to-end nonlinear modeling of the pressure distribution sequence using convolutional neural networks or Long Short-Term Memory (LSTM) networks. These approaches can, to some extent, distinguish between normal handholding and typical accidental touches (such as hand slippage or excessive squeezing), enabling interactive electronic components like the umbrella handle to meet basic requirements for accidental operation warnings and protection.

[0003] In recent years, with advancements in miniaturized pressure sensor technology, signal acquisition rates, and embedded computing capabilities, significant progress has been made in the spatial and temporal resolution of pressure distribution sequences. Some studies have attempted to introduce cutting-edge algorithms such as multi-channel signal feature fusion and intelligent adaptive calibration of sensor arrays to improve robustness in complex scenarios, including adaptability to high-temperature, high-humidity, and high-noise environments. However, despite the fact that existing technologies can achieve a certain degree of high-frequency anomalous signal capture and large-amplitude motion resolution, key shortcomings still exist.

[0004] First, because the pressure distribution during dynamic gripping fluctuates dramatically with changes in microdynamics, the acquired pressure amplitude is often affected by short-term physiological hand tremors, occasional changes in force, or localized aging of the sensor array. Conventional methods struggle to reliably extract the core temporal features that distinguish between accidental touches and normal operations. Existing threshold discrimination methods, primarily based on amplitude and amplitude change trends, have significant blind spots for slight sliding or slow squeezing-type accidental touches, resulting in high false alarm rates and poor robustness. Most deep learning end-to-end modeling methods heavily rely on the scale of pressure signal data and the consistency of the acquisition environment, exhibiting limited noise resistance and forward prediction capabilities. They struggle to capture early signals of accidental touches and are susceptible to interference from physiological variables such as new users and prolonged gripping.

[0005] Secondly, current technologies generally overlook the rhythmic coordination characteristics of multi-channel pressure distribution in the temporal dimension. The potential features such as temporal synchronization and phase correlation between multi-channel signals have long been neglected by traditional methods. Mainstream algorithms have failed to fully exploit the phase coupling information within the pressure signals, thus the stability of feature extraction in noisy environments urgently needs improvement. Especially in resource-constrained embedded platforms and real-time response scenarios, existing high-complexity convolutional neural networks or temporal modeling algorithms require significant computational power and have high response latency, making them difficult to implement in practical applications.

[0006] Furthermore, in order to adapt to the ever-changing individual physiological characteristics and usage habits, traditional anti-mistouch recognition methods based on static benchmarks or offline models are difficult to continuously and adaptively update, resulting in a significant drift in recognition accuracy when held continuously for a long time (such as in the case of umbrella handle use).

[0007] In summary, existing technologies in the field of pressure distribution sequence feature extraction and accidental touch prevention for smart grip devices such as folding umbrellas mainly face the following unresolved technical challenges: lack of in-depth exploration of the rhythmic coordination characteristics of multi-channel pressure time-series signals; the pressure feature extraction process is sensitive to noise and physiological drift, making it difficult to achieve robust and rapid identification and classification in the early stages of accidental touch actions; and related solutions generally consume a lot of computing power and are not conducive to the deployment of low-power embedded real-time control systems. Summary of the Invention

[0008] This application provides a method for preventing accidental triggering of folding umbrellas based on the umbrella handle grip pressure distribution sequence, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0009] The folding umbrella anti-accidental triggering control method based on the umbrella handle grip pressure distribution sequence provided in this application specifically includes: S1: Acquire the original pressure time-series signal and divide the original pressure time-series signal into one master control reference group and four subordinate subgroups based on spatial proximity; S2: Perform Hilbert transform processing on the original pressure timing signals in the master control reference group and the four subordinate subgroups respectively to generate a multi-channel instantaneous phase angle sequence; S3: Based on the multi-channel instantaneous phase angle sequence, calculate the mean and standard deviation of the sliding window phase difference between the average phase of each subordinate subgroup and the reference phase, and generate a phase consistency index sequence. S4: Using the monotonically increasing trend and increase data of the phase consistency index sequence, combined with the preset dynamic threshold, a preliminary judgment is made to generate the initial judgment result of the accidental touch action initiation stage. S5: Calculate the phase difference entropy value by statistically analyzing the dispersion of the phase difference distribution of each subordinate subgroup, and compare the phase difference entropy value with the 95th percentile value of the historical baseline to generate auxiliary criterion results; S6: Based on the preliminary judgment result and the auxiliary judgment result, when it is determined that the false touch behavior is established, freeze the current phase consistency model parameters and only update the drift compensation term of the reference phase trajectory to generate the reference phase trajectory. S7: Perform heterogeneous feature fusion processing on the phase consistency index sequence and the time-domain pressure gradient sequence to generate a fused feature vector; S8: Input the fused feature vector into the graph convolutional network for classification reasoning, output fine-grained classification results, and trigger the corresponding anti-accidental touch control command.

[0010] The folding umbrella anti-accidental triggering control method based on the umbrella handle grip pressure distribution sequence provided in this application has the following beneficial effects: (1) By constructing a ring array based on spatially distributed pressure sensing units and introducing a cross-channel instantaneous phase dynamic consistency analysis mechanism, the problems of high misjudgment rate and response lag caused by the reliance on amplitude threshold judgment and complex signal preprocessing in traditional touch recognition methods are effectively overcome. This scheme abandons the high-overhead processing procedures such as denoising, interpolation or deep feature extraction of the original pressure signal, and instead focuses on the temporal phase evolution relationship between each sensing subgroup and the main control reference group during the user's grip. The instantaneous phase angle sequence is extracted using Hilbert transform, and the phase consistency index sequence is constructed using the mean and standard deviation of the phase difference within the sliding window, realizing early sensitive detection of the erroneous touch action initiation stage. In particular, when the index shows a monotonically increasing trend and the increase exceeds the limit within 5 consecutive time windows, a preliminary warning is triggered, which significantly improves the forward prediction capability of the system for small unintentional operations, greatly reduces the probability of false response caused by sudden pressure fluctuations, and enhances the stability and reliability of human-computer interaction.

[0011] (2) Further integrate phase difference entropy value as an auxiliary criterion. By quantifying the dispersion of phase difference distribution among multiple subgroups, the system monitors the degree of disruption of the overall rhythmic coordination of the system while the phase consistency trend changes. When the entropy value exceeds the 95th percentile of the historical baseline, the confidence of the false touch warning is enhanced, thus forming a two-dimensional dynamic verification mechanism to effectively distinguish between real control intentions and non-functional hand micro-movements or environmental disturbances. On this basis, the system adopts an adaptive strategy that combines parameter freezing and drift compensation. After determining that the false touch development stage has been reached, only the slow drift term of the reference phase trajectory is updated without adjusting the core model parameters. It can maintain the stability of the learned pattern and adapt to the phase change phenomenon caused by physiological factors such as muscle relaxation and body position changes during long-term holding, avoiding the memory erosion problem caused by continuous learning in traditional models. At the same time, it integrates the phase consistency index with the temporal pressure gradient sequence through heterogeneous feature fusion and inputs it into a lightweight graph convolutional network to achieve fine-grained classification of three typical mis-touch behaviors (sliding mis-touch, squeezing mis-touch, and hovering jitter), which balances classification accuracy and computational efficiency. It is suitable for resource-constrained embedded control scenarios of folding umbrellas and has system-level advantages of low power consumption, high real-time performance and strong robustness.

[0012] The aforementioned technical approach fundamentally changes the existing design paradigm of "emphasizing amplitude and neglecting rhythm" in pressure sensing interaction. For the first time, it takes cross-channel temporal phase coordination as the core discrimination criterion, which not only significantly improves the sensitivity and specificity of accidental touch recognition, but also constructs a lightweight decision-making closed loop that does not require complex parameter tuning and has online adaptive capabilities. It provides a scalable and interpretable new generation of human-machine co-sensing architecture for miniaturized intelligent handheld devices. Attached Figure Description

[0013] Figure 1 This is the main flowchart of a folding umbrella anti-accidental trigger control method based on the umbrella handle grip pressure distribution sequence.

[0014] Figure 2 This is a sub-flowchart of a folding umbrella anti-accidental trigger control method based on the umbrella handle grip pressure distribution sequence.

[0015] Figure 3 This is another sub-flowchart of the folding umbrella anti-accidental trigger control method based on the umbrella handle grip pressure distribution sequence. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0018] like Figure 1 As shown, this application provides a method for preventing accidental triggering of a folding umbrella based on the umbrella handle grip pressure distribution sequence, specifically including: S1: Acquire the original pressure time-series signal and divide the original pressure time-series signal into one master control reference group and four subordinate subgroups based on spatial proximity; S2: Perform Hilbert transform processing on the original pressure timing signals in the master control reference group and the four subordinate subgroups respectively to generate a multi-channel instantaneous phase angle sequence; S3: Based on the multi-channel instantaneous phase angle sequence, calculate the mean and standard deviation of the sliding window phase difference between the average phase of each subordinate subgroup and the reference phase, and generate a phase consistency index sequence. S4: Using the monotonically increasing trend and increase data of the phase consistency index sequence, combined with the preset dynamic threshold, a preliminary judgment is made to generate the initial judgment result of the accidental touch action initiation stage. S5: Calculate the phase difference entropy value by statistically analyzing the dispersion of the phase difference distribution of each subordinate subgroup, and compare the phase difference entropy value with the 95th percentile value of the historical baseline to generate auxiliary criterion results; S6: Based on the preliminary judgment result and the auxiliary judgment result, when it is determined that the false touch behavior is established, freeze the current phase consistency model parameters and only update the drift compensation term of the reference phase trajectory to generate the reference phase trajectory. S7: Perform heterogeneous feature fusion processing on the phase consistency index sequence and the time-domain pressure gradient sequence to generate a fused feature vector; S8: Input the fused feature vector into the graph convolutional network for classification reasoning, output fine-grained classification results, and trigger the corresponding anti-accidental touch control command.

[0019] Step S1: Acquire the original pressure time-series signal and divide it into one master reference group and four subordinate subgroups based on spatial proximity. Specifically, this includes: S1.1: Perform synchronous sampling control on the ring array pressure sensor embedded in the umbrella handle grip interface to obtain a set of twelve raw pressure timing signals containing timestamp information.

[0020] Hardware-level synchronization triggering commands are executed on the ring array pressure sensors embedded in the umbrella handle grip interface to ensure that the analog-to-digital converters (ADCs) of the twelve independent channels start sampling on the same microsecond clock edge, eliminating time skew errors in multi-channel data acquisition. Based on a preset fixed sampling frequency, the signal acquisition module discretizes and quantizes the analog voltage signal of each pressure sensing unit at a frequency of no less than 100Hz, generating a raw digital pressure value sequence containing precise timestamps. The discrete pressure data from the twelve channels are aligned along the time axis to construct a raw pressure time series matrix of dimension N×12, where N represents the number of sampling points and 12 represents the number of spatially distributed sensing unit channels, forming a complete set of twelve-channel raw pressure time series signals. A global synchronization time index is added to the generated raw pressure time series signal set to ensure a strict correspondence between the data of each channel in the time domain during subsequent sliding window analysis, providing a high-fidelity spatiotemporal reference data source for phase consistency calculations.

[0021] Through the above-mentioned synchronous sampling and data structuring processing, the dispersed sensor analog signals are transformed into a set of twelve raw pressure time-series signals with strict spatiotemporal alignment characteristics, realizing precise synchronization of multi-channel signals on a microsecond-level time scale, laying a high-quality data foundation for the early identification of accidental touch behavior based on phase difference analysis.

[0022] S1.2: Perform sliding window variance calculation processing based on the set of the twelve-channel original pressure time-series signals to generate a dynamic stability index sequence that characterizes the response sensitivity and fluctuation amplitude of each sensing unit.

[0023] The twelve-channel raw pressure time-series signal set containing timestamp information, output from step S1.1, is received as the input data source for calculating the dynamic stability index. A sliding window truncation operation is performed on the raw pressure time-series signal of each channel, with a sliding window length of N sampling points and a step size of M sampling points to construct a local time-domain analysis unit. Within each sliding window, the arithmetic mean of all sampling points within that window is calculated as a central trend quantity characterizing the current pressure baseline level. Based on the arithmetic mean, the sum of squares of the differences between each sampling point within the window and the mean is calculated to quantify the dispersion of the pressure signal within the local time window. The sum of squares is divided by the degree of freedom coefficient of the window length minus one to obtain the sample variance value under that sliding window, used to eliminate statistical errors caused by sampling quantity deviations.

[0024] The calculated sample variance values ​​are square-rooted to generate a standard deviation sequence characterizing the pressure fluctuation amplitude, restoring it to the same dimensional units as the original pressure signal. An exponentially weighted moving average is applied to the standard deviation sequence, assigning higher weights to recent data to enhance the indicator's sensitivity to sudden changes in grip state while suppressing severe jitter caused by transient noise. The smoothed standard deviation sequence is defined as a dynamic stability index sequence, where each element corresponds to the pressure fluctuation intensity within a specific time window. Through sliding window variance calculation and exponentially weighted smoothing, the original pressure time-series signal from the previous step is transformed into a dynamic stability index sequence characterizing the response sensitivity and fluctuation amplitude of each sensing unit. This achieves robust extraction of the micro-fluctuation characteristics of grip pressure, providing a quantitative basis for subsequent selection of the master control reference group.

[0025] S1.3: Using the dynamic stability index sequence, three adjacent sensing units with moderate fluctuation amplitude and the most sensitive response are selected to construct a master control reference group signal set as a phase reference source.

[0026] S1.4: Execute a neighboring region division algorithm based on the spatial location coordinates of the master control reference group signal set to classify the remaining nine sensing units into four spatially continuous subordinate subgroup signal sets.

[0027] Obtain the spatial coordinate indices of the three adjacent sensing units contained in the master control reference group signal set determined in step S1.3, as well as the complete spatial topology distribution matrix of the remaining nine non-master control sensing units. The three sensing units of the master control reference group form a continuous arc-shaped region in the ring array, which is defined as the core radiation source of the phase reference.

[0028] Based on the cylindrical geometric unfolding model of the umbrella handle grip interface, a local polar coordinate system is established with the geometric center of the master control reference group as the origin. The angular and radial distances of the remaining nine sensing units relative to the center of the master control reference group are calculated, and a spatial proximity weight matrix is ​​constructed. This weight matrix quantifies the physical coupling strength between each subordinate unit and the master control reference source, providing a geometric constraint basis for subsequent subgroup division.

[0029] Based on the principle of spatial continuity, the remaining nine sensing units are linearly ordered along the circumference of the ring array. Using the two boundary units of the master control reference group as dividing points, the remaining circumference is divided into four independent arc segments. Each arc segment contains at least two sensing units, ensuring that each subordinate subgroup has sufficient spatial sampling density to characterize the changing trend of the local pressure field.

[0030] A connectivity check is performed on the sensing units within each arc segment to confirm that all units within the segment are physically adjacent and uninterrupted. If the number of units in a segment exceeds a preset limit, the farthest unit is redistributed to adjacent segments based on its angular distance from the boundary of the master control reference group, in order to balance the load of each subgroup and maintain the compactness of the spatial topology.

[0031] The four segmented intervals are mapped to the first, second, third, and fourth subordinate subgroups, respectively. The sensor unit indices within each subgroup are encapsulated into independent data structures and logically associated with the master control reference group signal set. This partitioning method ensures that each subordinate subgroup spatially corresponds to a different gripping area of ​​the user's palm (such as the base of the thumb, hypothenar eminence, thenar eminence, etc.).

[0032] Using the aforementioned neighborhood partitioning algorithm, the nine dispersed non-master sensing units are classified into four spatially continuous and topologically defined subordinate subgroup signal sets. This achieves signal grouping and structuring based on spatial proximity, laying a solid spatial topological foundation for the subsequent extraction of phase difference consistency indices of each subgroup relative to the master reference.

[0033] S1.5: Perform data structure encapsulation processing based on the master control reference group signal set and the four subordinate subgroup signal sets to generate a grouped pressure time-series data stream for subsequent Hilbert transform processing.

[0034] Step S2: Perform Hilbert transform processing on the original pressure time-series signals in the master control reference group and the four subordinate subgroups respectively to generate a multi-channel instantaneous phase angle sequence. Specifically, this includes: S2.1: Perform empirical mode decomposition preprocessing on the original pressure time series signals collected from the main control reference group and the four subordinate subgroups respectively to extract the intrinsic mode components that characterize the dynamic changes of the user's grip and remove high-frequency noise interference to generate a denoised pure pressure time series signal sequence.

[0035] S2.2: Based on the denoised pure pressure time-series signal sequence, construct an analytical signal model, and use the Hilbert transform algorithm to perform orthogonal mapping processing on the signal of each channel to generate a complex domain analytical signal vector containing the real part of the original signal and the imaginary part of the Hilbert transform result.

[0036] The system receives a denoised and clean pressure time-series signal sequence generated after empirical mode decomposition preprocessing in step S2.1. This sequence contains discrete sampling point data of each sensing unit in the time dimension within the master control reference group and four subordinate subgroups, which serves as the input basis for constructing the analytical signal model.

[0037] For each channel of the denoised and clean pressure time-series signal, the Hilbert transform algorithm is used to perform orthogonal mapping processing, aiming to construct orthogonal components with the same amplitude spectrum as the original signal but with a phase lag of 90 degrees, so as to completely characterize the instantaneous state of the signal in the complex domain.

[0038] The Hilbert transform and analytic signal construction process described above is executed in parallel for the three adjacent sensing units in the master reference group and the remaining nine sensing units in the four subordinate subgroups, ensuring that all twelve channels are strictly aligned on the timestamps, and generating an analytic signal vector set containing twelve complex sequences.

[0039] By constructing a complex domain analytic signal model, the one-dimensional time-domain pressure signal is extended into a rotating vector on a two-dimensional complex plane, realizing the mapping from a single amplitude domain to the amplitude-phase joint domain. This provides a mathematically complete signal representation for subsequent calculation of the instantaneous phase angle and analysis of the phase consistency between multiple channels, effectively avoiding the phase ambiguity problem caused by noise interference in traditional time-domain analysis.

[0040] S2.3: Calculate the arctangent function ratio based on the real and imaginary part values ​​of the complex domain analytic signal vector to solve for the instantaneous phase angle value of each pressure sensing unit at each sampling time, and generate an initial instantaneous phase angle time series.

[0041] Receive the complex domain analytic signal vector generated in step S2.2. This vector contains the real part of the original signal component and the imaginary part of the Hilbert transform result component corresponding to each pressure sensing unit at each sampling time.

[0042] For each channel data in the complex domain analytic signal vector, extract the real part numerical sequence and the imaginary part numerical sequence as input variables for instantaneous phase angle calculation.

[0043] For each sampling time t, read the real part signal value x of the i-th sensing unit at the current time. i (t) and the imaginary part of the signal value y i (t), construct two-dimensional orthogonal coordinate points.

[0044] The angle of the orthogonal coordinate point relative to the real axis is calculated using the arctangent function to obtain the initial instantaneous phase angle value before expansion.

[0045] The four-quadrant arctangent function atan2 is used for calculation to ensure that the phase angle range covers the complete [-π,π] interval and avoids the loss of phase information due to the limitation of a single quadrant.

[0046] Iterate through all sensing units in the master reference group and subordinate subgroups, and perform the above phase calculation operation on each unit at all sampling times.

[0047] The calculated discrete phase angle values ​​are arranged in chronological order to form the initial instantaneous phase angle time sequence corresponding to each sensing unit.

[0048] The generated initial instantaneous phase angle time series is encapsulated in a data structure, retaining the timestamp index and channel identifier, so that subsequent steps can be performed for phase expansion and consistency analysis.

[0049] By using the four-quadrant arctangent operation, the complex domain analytical signal from the previous step is transformed into initial instantaneous phase angle time series data that reflects the rhythmic characteristics of pressure fluctuations. This achieves the mapping from the amplitude-frequency domain to the phase domain, providing basic phase data for capturing the early phase discretization trend of accidental touch actions.

[0050] S2.4: Perform a phase unfolding operation on the initial instantaneous phase angle time series to eliminate discontinuities caused by periodic angle jumps, and perform channel alignment calibration on the phase data according to the spatial topology of the umbrella handle grip interface to generate a smooth and continuous multi-channel instantaneous phase angle sequence.

[0051] Receive the initial instantaneous phase angle time series generated by step S2.3. This series contains the original phase values ​​of each pressure sensing unit in the master control reference group and the four subordinate subgroups at discrete sampling times. The value range is limited to the main value interval from negative π to positive π, and there are phase jump points caused by the periodic truncation of the arctangent function.

[0052] The initial instantaneous phase angle time series is subjected to point-by-point differential detection processing to calculate the difference in phase values ​​between adjacent sampling times in order to identify abnormal jump points that exceed a preset phase continuity threshold. The preset phase continuity threshold is set to 0.8 times π to distinguish between normal continuous phase changes and periodic boundary jumps.

[0053] When the absolute value of the phase difference between adjacent time points is detected to be greater than the preset phase continuity threshold, it is determined that a phase winding phenomenon has occurred, and the winding direction is determined according to the positive or negative sign of the difference. If the difference is positive and exceeds the threshold, it is determined to be downward winding, and a compensation of -2π is required; if the difference is negative and exceeds the threshold, it is determined to be upward winding, and a compensation of +2π is required.

[0054] Based on the determined winding direction, the phase values ​​of the current moment and all subsequent sampling points are accumulated with a corresponding compensation amount that is an integer multiple of 2π. The phase unfolding algorithm is then executed to reconstruct the sawtooth phase sequence that was originally limited to the main value range into a continuous linear phase trajectory that monotonically increases or decreases with time, thus eliminating the discontinuity caused by the periodicity of the angle.

[0055] A spatial topology constraint mechanism is introduced into the unfolded multi-channel phase sequence. Based on the physical arrangement order of the ring array sensors at the umbrella handle grip interface, a phase gradient consistency verification model between adjacent sensing units is established, and the difference in the unfolded phase of two spatially adjacent sensing units at the same time is calculated.

[0056] If the phase difference between adjacent sensing units exceeds the upper limit of the spatial phase gradient set based on the biomechanical characteristics of gripping, it is determined that there is local phase distortion or cumulative error drift in the channel. The spatial smoothing correction program is then initiated, and the abnormal channel is interpolated and corrected using the phase mean of the adjacent normal channels to ensure that the phase continuity in the spatial distribution conforms to the physical law of the coordinated contraction of human hand muscles.

[0057] Perform time axis alignment calibration on the spatially corrected phase data. Using the phase sequence of the most sensitive central sensing unit in the master control reference group as the time reference, calculate the cross-correlation coefficient of the phase sequences of other subordinate subgroup sensing units relative to the reference sequence, and find the time delay offset corresponding to the peak value of the cross-correlation coefficient.

[0058] Based on the calculated time delay offset, fractional delay compensation is performed on the phase sequence of each subordinate subgroup to eliminate the phase asynchrony between channels caused by differences in signal transmission paths or micro-hysteresis in sensor response, thereby achieving strict alignment of multi-channel phase data in the time dimension.

[0059] The aligned multi-channel phase sequence is processed by low-pass smoothing filter, with the cutoff frequency set to 1.5 times the highest frequency component of the gripping action. This filters out high-frequency quantization noise and minor jitter that may remain during phase unfolding, generating a smooth multi-channel instantaneous phase angle sequence with high signal-to-noise ratio, temporal continuity, and spatial consistency.

[0060] Through a chain of processing steps including phase unrolling, spatial topology correction, and time delay alignment, the initial phase sequence obtained in the previous step, which contains jumps and asynchrony defects, is transformed into a smooth and continuous multi-channel instantaneous phase angle sequence. This enables accurate characterization of the dynamic rhythm of grip pressure changes and lays a solid data foundation for the subsequent extraction of highly robust phase consistency indicators.

[0061] S2.5: The smooth and continuous multi-channel instantaneous phase angle sequence is structurally reorganized according to the division logic of the master control reference group and four subordinate subgroups to form a standardized multi-channel instantaneous phase angle sequence output set for characterizing cross-channel rhythm coordination.

[0062] like Figure 2 As shown, step S3 involves calculating the mean and standard deviation of the sliding window phase difference between the average phase of each subordinate subgroup and the reference phase, based on the multi-channel instantaneous phase angle sequence, to generate a phase consistency index sequence. Specifically, this includes: S3.1: Perform time-dimensional mean aggregation processing on the master reference group data in the multi-channel instantaneous phase angle sequence to generate the master reference group average phase trajectory as a time-series alignment benchmark.

[0063] S3.2: Based on the average phase trajectory of the master control reference group, perform sliding window truncation operation on the multi-channel instantaneous phase angle sequences in the four subordinate subgroups respectively to generate subordinate subgroup sliding window phase data blocks containing local timing information.

[0064] Receive the average phase trajectory of the master control reference group generated in step S3.1 as the timing alignment reference, and the structured and recombined multi-channel instantaneous phase angle sequence of the four subordinate subgroups output in step S2.5.

[0065] Set the length and step size parameters of the sliding time window, where the window length corresponds to the minimum time span for the early characteristics of accidental touch to appear, and the step size determines the real-time resolution of the phase consistency index update.

[0066] For each subordinate subgroup, based on the preset sliding time window parameters, the multi-channel instantaneous phase angle sequence is continuously truncated on the time axis to obtain a local phase data segment composed of the current time and several historical sampling points.

[0067] The extracted local phase data segments are arranged in a matrix according to the sensor unit channel index to construct a sliding window phase data block of subordinate subgroups containing spatial and temporal dimensions, ensuring that each data block contains complete phase evolution information of all sensor units in the subgroup within a specified time window.

[0068] The phase value segments of the average phase trajectory of the master reference group within the same time interval as the sliding window of the subordinate subgroup are extracted synchronously to form a reference phase reference vector that is strictly aligned with the time of the data block of the subordinate subgroup.

[0069] Boundary integrity checks are performed on the generated subordinate subgroup sliding window phase data blocks to remove incomplete window data caused by insufficient data length at the signal start or end stages, ensuring the statistical validity of subsequent phase difference calculations.

[0070] By using a sliding window truncation processing method, the globally continuous multi-channel instantaneous phase angle sequence is transformed into a subordinate subgroup sliding window phase data block with local temporal correlation characteristics. This achieves a discretized local representation of the dynamic change process of the grip pressure phase, providing a standardized input data structure for subsequent calculation of local rhythm deviation and phase consistency index.

[0071] S3.3: Perform point-by-point phase subtraction operation using the corresponding window segment of the sliding window phase data block of the subordinate subgroup and the average phase trajectory of the master control reference group to generate a subordinate subgroup sliding window phase difference sequence characterizing local rhythm deviation.

[0072] The system receives a corresponding window segment of the subordinate subgroup sliding window phase data block and the master control reference group average phase trajectory as the input data source. The subordinate subgroup sliding window phase data block contains the instantaneous phase angle sequence of four spatially adjacent sensing units within a specific time window, and the master control reference group average phase trajectory is the arithmetic mean sequence of the phases of three reference sensing units within the same time window.

[0073] For each channel phase sequence in the sliding window phase data block of the subordinate subgroup, a strict time axis alignment operation with the average phase trajectory of the master reference group is performed. This ensures a one-to-one correspondence between the two sequences at the sampling time index, eliminates time offset errors caused by data transmission delays or asynchronous processing, and constructs a synchronous phase comparison dataset.

[0074] For each pair of phase data points after synchronization, a point-by-point phase subtraction operation is performed. Since the phase angle has a periodic characteristic, direct subtraction may cause the result to exceed the principal value range; therefore, a phase winding correction mechanism is required. The difference between the instantaneous phase angle of the i-th sensing unit in the subordinate subgroup at time t and the average phase angle of the master control reference group at time t is calculated.

[0075] The calculated original phase difference is normalized modulo 2π to map it to the interval [-π, π] to eliminate the numerical discontinuity caused by jumps to integer multiples of 2π. Specifically, if the original phase difference is greater than π, then 2π is subtracted; if the original phase difference is less than -π, then 2π is added.

[0076] By iterating through all time points within the sliding window and all sensing units within the subordinate subgroups, the above subtraction and correction steps are repeated to generate a two-dimensional matrix structure of the subordinate subgroup sliding window phase difference sequence. The row dimension of this sequence corresponds to the sampling points within the time window, and the column dimension corresponds to the sensing unit channels within the subordinate subgroups.

[0077] Outlier removal is performed on the generated phase difference sequence. Outlier phase jump points caused by momentary poor sensor contact or electromagnetic interference are identified and removed. Linear interpolation is used to fill in the missing data to ensure the smoothness of the phase difference sequence and the consistency of its physical meaning.

[0078] By performing point-by-point phase subtraction and periodic correction, the multi-channel phase data from the previous step is transformed into a sliding window phase difference sequence of subordinate subgroups that characterizes local rhythm deviations. This enables precise quantification of the phase lag or lead of each subordinate region relative to the master control reference, providing a standardized deviation data basis for subsequent statistical phase consistency indicators.

[0079] S3.4: Perform first-order moment statistical calculation on the sliding window phase difference sequence of the subordinate subgroup to generate the mean of the sliding window phase difference that characterizes the trend of cross-channel phase synchronization center.

[0080] Receive the phase difference sequence of the sliding window of the subordinate subgroup. This sequence contains the instantaneous deviation values ​​of each subordinate sensing unit relative to the reference phase of the master control reference group within the current sliding time window, and serves as the input data source for the first-order moment statistical calculation.

[0081] A channel-dimensional traversal operation is performed on the sliding window phase difference sequence of the subordinate subgroup. For all sensing units in each subordinate subgroup, the set of discrete phase difference sampling points within the current sliding window length N is extracted to construct a local data matrix for statistical moment calculation.

[0082] Based on the local data matrix, the arithmetic mean algorithm is applied to perform lateral aggregation of the phase difference values ​​of different sensing channels belonging to the same subordinate subgroup at the same time, thereby eliminating random fluctuation interference caused by small deformation of the contact surface or local noise of single-point sensors.

[0083] The calculated average phase difference values ​​of each subordinate subgroup are arranged in time stamp order to form a sliding window average phase difference value sequence that evolves over time. This sequence reflects the synchronous trend of the pressure distribution rhythm of each spatial region relative to the center of the main control region.

[0084] The generated sliding window phase difference mean sequence is smoothed and filtered, and a moving average filter is used to further suppress transient spikes caused by the Hilbert transform boundary effect, ensuring the continuity and stability of the mean trajectory.

[0085] Through the above-mentioned first-order moment statistical calculation method, the phase difference sequence of the subordinate subgroup sliding window is transformed into the average phase difference of the sliding window, which represents the trend of cross-channel phase synchronization center. This realizes the quantitative extraction of the center offset of rhythmic coordination of multi-channel pressure signals, and provides stable benchmark reference data for the subsequent generation of phase consistency index by combining the standard deviation of the second moment.

[0086] S3.5: Perform second-order moment statistical calculation processing based on the sliding window phase difference sequence of the subordinate subgroup and the generated average sliding window phase difference to generate a phase consistency index sequence that characterizes the degree of cross-channel rhythmic synergy fluctuation.

[0087] The system receives the phase difference sequence of the sliding window of the subordinate subgroups and the corresponding average phase difference of the sliding window, which serves as the basic data input for calculating the degree of fluctuation in cross-channel rhythm coordination. This step aims to quantify the phase dispersion of the pressure sensing units within each subordinate subgroup relative to the master control reference group, in order to capture the rhythm desynchronization phenomenon unique to the initial stage of accidental touch.

[0088] Iterate through all sampling times within the current sliding time window. For each subordinate subgroup, extract the instantaneous phase difference value of each sensing unit within that subgroup at a specific time. Using the average phase difference of the sliding window generated in the previous step as the central reference point, calculate the deviation between the instantaneous phase difference of each sensing unit and the average value at each sampling time.

[0089] The total phase fluctuation energy within the window period is obtained by summing the squared phase deviation values ​​at all sampling times within the sliding window.

[0090] Divide the sum of phase variances by the effective degrees of freedom within the window, i.e., the total number of sampling points minus one, to eliminate the influence of the number of samples on the assessment of the degree of fluctuation and generate an unbiased estimate of the phase variance value.

[0091] The standard deviations of the sliding window phase differences calculated for each of the four subordinate subgroups are weighted and averaged. The weighting coefficients are determined based on the number of sensing units in each subgroup and their spatial distribution relative to the master reference group, to generate a phase consistency index that comprehensively reflects the rhythmic coordination of the overall grip interface. The current phase consistency index (PCI) is generated using the following formula: Wherein, PCI is the phase consistency index, w k Let σ be the normalized weight coefficient of the k-th subordinate subgroup, satisfying that the sum of all weights is 1. k Let be the standard deviation of the sliding window phase difference for the k-th subordinate subgroup.

[0092] The calculated phase consistency index is stored in a time-series buffer to form a continuously updated phase consistency index sequence. This sequence can reflect the phase synchronization status of multi-channel pressure signals in real time under user grip conditions. When a mis-touch action occurs, the involuntary tremor or sliding of the finger muscles causes significant differences in the rhythm of pressure changes in different areas, resulting in a sharp increase in the phase consistency index.

[0093] Through the above-mentioned second-order moment statistical calculation and weighted fusion processing, the phase difference sequence of the subordinate subgroup sliding window is transformed into a phase consistency index sequence that characterizes the degree of cross-channel rhythmic synergy fluctuation. This achieves quantitative capture of the early phase discretization trend of accidental touch actions, effectively overcomes the shortcomings of traditional amplitude features that are easily affected by noise and lack foresight, and significantly improves the robustness and response speed of accidental touch recognition.

[0094] like Figure 3 As shown, step S4 involves using the monotonically increasing trend and increase data of the phase consistency index sequence, combined with a preset dynamic threshold, to perform a preliminary judgment and generate an initial judgment result for the accidental touch action initiation stage. Specifically, this includes: S4.1: Perform sliding time window truncation processing on the phase consistency index sequence to generate a local phase consistency index subsequence containing multiple consecutive sampling points, which serves as the basic data unit for trend analysis.

[0095] S4.2: Perform adjacent time point difference operation processing based on the local phase consistency index subsequence to generate an instantaneous rate of change sequence characterizing the direction of phase consistency change, which is used to quantify the slope of the index evolution over time.

[0096] S4.3: Perform positive and negative sign continuity determination processing on the instantaneous rate of change sequence to generate a monotonic flag sequence that indicates whether the phase consistency index shows a strictly monotonically increasing state, and filter out the rising segment data that meets the characteristics of accidental touch precursor.

[0097] Receive the instantaneous rate of change sequence generated in step S4.2, which contains differential data of the phase consistency index between adjacent sampling points within the sliding window, as the raw input data stream for determining monotonicity.

[0098] For each numerical element in the instantaneous rate of change sequence, a sign extraction operation is performed. By comparing the magnitude of each difference data with the zero value, the direction of change of the phase consistency index at each time step is determined.

[0099] Construct a Boolean flag array of the same length as the instantaneous rate of change sequence, initialize all flags to an invalid state, and use it to record whether each sampling point satisfies the positively increasing local condition.

[0100] Traverse the instantaneous rate of change sequence. When the difference data at a certain moment is detected to be strictly greater than zero, set the corresponding flag bit array element at that moment to a logical true value, indicating that there is an increasing trend of phase discretization at that moment.

[0101] When the differential data at a certain moment is detected to be less than or equal to zero, the corresponding flag array element at that moment is set to a logical false value, and the continuously incrementing counter is reset to indicate that the phase coherence at that moment has not deteriorated or shown signs of recovery.

[0102] Based on the generated Boolean flag sequence, a continuous truth segment detection algorithm is executed to identify continuous truth subsequences whose length reaches the preset minimum duration window.

[0103] The identified continuous truth subsequences are marked with boundaries, and their start time index and end time index are extracted to form several independent monotonically increasing time intervals.

[0104] Flags located within a monotonically increasing time interval are kept in an active state, while isolated truth points or discontinuous truth segments outside the interval are forcibly corrected to an inactive state in order to filter out brief false rising pulses caused by high-frequency noise from the sensor.

[0105] The final monotonicity flag sequence is generated. This sequence retains valid flags only when the phase consistency index shows a continuous and stable monotonically increasing trend, and is marked as invalid at other times.

[0106] By determining the continuity of positive and negative signs and processing noise filtering, the instantaneous rate of change sequence of the previous step is transformed into a monotonic flag sequence that characterizes the stability of the early warning features of false touch. This achieves early and robust screening of phase discretization trends and effectively improves the noise resistance of false touch detection.

[0107] For example, the sliding window length is set to 10 sampling points, and the sampling frequency is 50Hz, meaning each window covers a time span of 0.2 seconds. The instantaneous rate of change sequence contains 10 difference data points: [0.02, 0.03, -0.01, 0.04, 0.05, 0.06, 0.07, 0.08, -0.02, 0.01]. First, the sign of each data point is determined: 0.02 > 0 is recorded as 1, 0.03 > 0 as 1, -0.01 ≤ 0 as 0, 0.04 > 0 as 1, and so on, resulting in the initial Boolean sequence [1, 1, 0, 1, 1, 1, 1, 1, 0, 1]. The preset minimum duration window length is 4 consecutive points. Detecting continuous truth segments: Indexes 0-1, with a length of 2 (less than 4), are considered noise and their flags are corrected to 0; indices 3-7, with a length of 5 (greater than or equal to 4), are considered valid monotonically increasing segments and their flags are retained at 1; index 9, with a length of 1 (less than 4), is considered noise and corrected to 0. The final generated monotonicity flag sequence is [0,0,0,1,1,1,1,1,0,0]. This result accurately eliminates brief fluctuations at the beginning and end, retaining only the stable upward trend for 5 consecutive points (0.1 seconds) in the middle, providing a high-confidence data foundation for subsequent calculations of cumulative increases and significantly reducing the false alarm rate caused by hand shake.

[0108] S4.4: Perform cumulative increase calculation processing based on the continuous and valid rising segment data in the monotonicity flag sequence to generate a dynamic increase value that characterizes the degree of accumulation of erroneous momentum and reflects the degree of asynchrony of pressure distribution rhythm.

[0109] The monotonicity flag sequence generated in step S4.3 is received. This sequence identifies valid data segments where the phase consistency index exhibits a strictly monotonically increasing state within a continuous time window. The index intervals continuously marked as valid states in the monotonicity flag sequence are extracted to determine the time span between the start time and the current time of the false trigger precursor characteristics. Based on the determined time span, corresponding local subsequences are extracted from the phase consistency index sequence output in step S3 to construct a dynamic dataset characterizing the evolution of rhythm desynchronization. The first and last numerical differences are calculated on the extracted local subsequences to determine the absolute difference between the current phase consistency index value and the baseline value at the start time, thus obtaining the basic cumulative increase. A time decay weighting factor is introduced to nonlinearly correct the basic cumulative increase, suppressing increase jitter caused by sampling interval fluctuations or instantaneous noise and enhancing the stability of trend judgment. An exponentially weighted moving average algorithm is used to process the corrected increase data, smoothing high-frequency disturbance components and generating a dynamic increase value reflecting the degree of false trigger momentum accumulation. By using the above-mentioned cumulative increase calculation and weighted smoothing process, the monotonicity determination result of the previous step is transformed into a dynamic increase value that quantitatively represents the degree of asynchrony in pressure distribution rhythm. This achieves accurate capture of the energy accumulation trend during the initiation stage of accidental touch action, providing a high signal-to-noise ratio discrimination basis for subsequent threshold comparison.

[0110] For example, the sliding time window length is set to 20 sampling points, and the sampling frequency is 100Hz. Within a certain detection period, the monotonicity flag sequence is continuously marked as valid (i.e., a value of 1) in the interval between the 50th and 55th sampling points, indicating that the phase consistency index continues to rise within this 60ms period. The values ​​corresponding to the index intervals are extracted from the phase consistency index sequence. Assume the phase consistency index value at the 50th sampling point is 0.12 rad, and the value at the 55th sampling point is 0.18 rad. The first and last numerical differences are calculated, and the basic cumulative increase is 0.18 - 0.12 = 0.06 rad. A time decay weighting factor of 0.95 is set to correct for fluctuations at intermediate sampling points. If there is a slight drop at the intermediate point, the impact is suppressed after weighting. An exponentially weighted moving average algorithm is applied, with a smoothing coefficient α of 0.3, to smooth the recent increase trend. If the dynamic increase value at the previous moment was 0.05 rad, the updated dynamic increase value at the current moment is calculated as follows: Where A is the current dynamic increase value, ΔP is the basic cumulative increase value of 0.06 rad, and At-1 is the dynamic increase value of 0.05 rad at the previous moment. Substituting the values, we get A = 0.3 × 0.06 + 0.7 × 0.05 = 0.018 + 0.035 = 0.053 rad. This dynamic increase value of 0.053 rad accurately reflects the accumulation trend of phase discretization within this short time window. Compared with the original difference value, it effectively filters out single-point noise interference, significantly improves the sensitivity of identifying weak false trigger precursor signals, and ensures higher confidence when comparing with the dynamic threshold in subsequent steps.

[0111] S4.5: Perform a comparison logic judgment process between the dynamic increase value and the preset dynamic threshold to generate a preliminary judgment result of the accidental touch action start-up stage, indicating whether the accidental touch action has entered the start-up stage, and complete the preliminary capture of potential accidental touch behavior.

[0112] Receive the dynamic increment value representing the degree of accumulation of false touch momentum output from step S4.4. This value quantifies the degree of accumulation of the phase consistency index deviating from the normal grip reference within the current sliding time window.

[0113] The system calls a preset dynamic threshold generation module to read the current usage environment parameters of the folding umbrella and the statistical characteristics of the user's historical holding behavior, and calculates an adaptive judgment threshold suitable for the current scenario.

[0114] The calculation of dynamic thresholds is based on a basic static threshold, an ambient temperature compensation coefficient, and a user-specific grip strength correction factor, ensuring the robustness of the discrimination standard under different working conditions.

[0115] A dynamic threshold function is constructed using a weighted linear combination method. The basic static threshold is used as a benchmark, and a fluctuation margin caused by environmental noise is superimposed to form a threshold boundary that changes slowly over time.

[0116] The dynamic increase value generated by S4.4 is compared with the preset dynamic threshold T obtained in real time using a logical operation.

[0117] If the dynamic increase value is greater than or equal to the preset dynamic threshold T, it is determined that the discretization trend of the phase consistency index has exceeded the tolerance range of normal physiological grip fluctuation.

[0118] At this point, a high-level signal is generated to indicate the initial judgment result of the accidental touch action initiation stage, signifying that the system has entered an early warning state for accidental touch behavior.

[0119] If the dynamic increase value is less than the preset dynamic threshold T, it is determined that the current phase change is still within the range of normal grip adjustment or slight interference, and does not constitute a substantial threat of accidental touch.

[0120] At this point, a low-level signal is generated to maintain normal monitoring status, and the system continues to collect subsequent pressure timing data and update the sliding window statistics.

[0121] The generated preliminary judgment result signal is latched into a temporary register and stamped with the current timestamp, serving as the triggering or suppression condition for calculating the phase difference entropy value auxiliary criterion in step S5.

[0122] Through the dynamic threshold comparison logic processing described above, the dynamic increase value calculated in the previous step is transformed into a binary preliminary judgment result of the false touch action initiation stage. This achieves the initial capture and coarse-grained screening of potential false touch behaviors, effectively reducing the false alarm rate caused by instantaneous noise peaks, and providing a reliable candidate event window for subsequent fine-grained classification.

[0123] Step S5: Calculate the phase difference entropy value by statistically analyzing the dispersion of the phase difference distribution of each subordinate subgroup, and compare the phase difference entropy value with the 95th percentile value of the historical baseline to generate an auxiliary criterion result. Specifically, this includes: S5.1: Obtain the instantaneous phase angle data of four subordinate subgroups in the multi-channel instantaneous phase angle sequence, and perform phase difference mapping processing on the instantaneous phase angle data in each subordinate subgroup based on the average phase trajectory of the master control reference group as the reference phase to generate a set of subordinate subgroup phase difference sequences that characterize cross-channel rhythm synchronization deviation.

[0124] S5.1 acquires the instantaneous phase angle data of four subordinate subgroups in the multi-channel instantaneous phase angle sequence, and performs phase difference mapping processing on the instantaneous phase angle data in each subordinate subgroup based on the average phase trajectory of the master control reference group as the reference phase, so as to generate a set of subordinate subgroup phase difference sequences that characterize the cross-channel rhythm synchronization deviation.

[0125] This sub-step is located at the beginning of the confidence enhancement judgment of accidental touch behavior. It aims to transform the multi-channel instantaneous phase angle information extracted in the previous step into phase difference data that can quantify the degree of rhythm desynchronization between the local grip area and the overall grip reference, and provide a direct input object for subsequent calculation of phase difference entropy value.

[0126] Read the standardized multi-channel instantaneous phase angle sequence output set output by step S2, and separate the instantaneous phase angle time series of the three sensing units corresponding to the master control reference group, as well as the instantaneous phase angle time series of the sensing units contained in the four subordinate subgroups, to ensure that all data are strictly aligned on the time axis and have completed phase expansion processing.

[0127] A point-by-point arithmetic average is performed on the instantaneous phase angle time series of the three sensing units in the master control reference group to eliminate random phase jitter caused by small deformation of the contact surface of a single sensor, and generate a smooth average phase trajectory of the master control reference group. This trajectory represents the macroscopic rhythmic benchmark of the core gripping area of ​​the user's palm.

[0128] For each subordinate subgroup, iterate through all the sensing units contained within it, extract the instantaneous phase angle values ​​within the current sliding time window, and construct a local phase matrix for the subordinate subgroup, where the rows of the matrix correspond to the time sampling points and the columns correspond to the spatial sensing nodes within the subgroup.

[0129] Using the average phase trajectory of the master control reference group as the minuend and the instantaneous phase angle of each sensing unit in the subordinate subgroup as the subtrahend, phase subtraction is performed time-by-time to calculate the instantaneous phase offset of each sensing unit in each subordinate subgroup relative to the master control reference.

[0130] Considering the periodicity of the phase angle, the calculated instantaneous phase offset is normalized to modulo 2π to ensure that the phase difference value converges to the interval [-π,π], thus avoiding numerical jumps caused by phase wrapping that could interfere with the stability of subsequent statistical properties.

[0131] The instantaneous phase difference sequences calculated by all sensing units within the same subordinate subgroup are aggregated to form a phase difference sequence set exclusive to that subgroup. This set fully records the rhythmic lag or lead state of each node relative to the holding center within the spatial region.

[0132] The phase difference calculation and aggregation process is repeated for the remaining three subordinate subgroups in sequence, and finally a set of phase difference sequences of subordinate subgroups containing four independent elements is generated. Each element corresponds to the cross-channel rhythm synchronization deviation time series data of a spatial sub-region.

[0133] By using the phase difference mapping processing method described above, the multi-channel absolute phase data from the previous step is converted into relative phase difference data, eliminating the common-mode phase drift caused by changes in global gripping force, and achieving the technical effect of retaining only the discretized features of local rhythms that reflect the cause of accidental touch.

[0134] For example, the umbrella handle grip interface is configured with a circular array containing 12 pressure sensing units, a sampling frequency of 100Hz, and a sliding window length of 0.5 seconds (i.e., 50 sampling points). The master control reference group consists of spatially adjacent sensing units 4, 5, and 6. The remaining 9 units are divided into four subordinate subgroups: subgroup 1 contains units 1 and 2, subgroup 2 contains unit 3, subgroup 3 contains units 7 and 8, and subgroup 4 contains units 9, 10, 11, and 12. At t=1.2 seconds, the average phase angle of the three units in the master control reference group is calculated to be 0.85 radians. For sensing unit 1 in subordinate subgroup 1, its original instantaneous phase angle is 1.10 radians, and for sensing unit 2 it is -2.90 radians. According to the phase difference calculation formula, the phase difference of unit 1 is 0.25 radians; the phase difference of unit 2, after modulo 2π normalization, is 2.53 radians. Similarly, the phase difference of all units within subgroup 1 at 50 sampling points in the current window is calculated to form the phase difference sequence of subgroup 1. If the user experiences a slight premonition of a swipe-to-touch, the phase difference fluctuation of subgroup 1 (located on the thumb side) will be significantly greater than that of other subgroups. For example, the standard deviation of the phase difference in subgroup 1 increases from 0.15 radians under normal grip to 0.45 radians, while that in subgroup 3 (little finger side) remains at 0.18 radians. This differentiated set of phase difference sequences accurately captures the desynchronization of local grip rhythms, providing highly discriminative input data for subsequent entropy calculations and effectively improving the system's sensitivity to early spurious touches.

[0135] S5.2: Based on the set of phase difference sequences of subordinate subgroups, a local phase difference data segment at the current moment is extracted using a sliding time window mechanism, and a probability density function estimation process is performed on the local phase difference data segment to generate a local phase difference probability density distribution curve that characterizes the phase difference distribution pattern.

[0136] Step S5.2 aims to extract a local phase difference data segment at the current moment based on the set of phase difference sequences of subordinate subgroups using a sliding time window mechanism, and perform probability density function estimation processing on the local phase difference data segment to generate a local phase difference probability density distribution curve characterizing the phase difference distribution pattern. This step is located in the core link of generating the auxiliary criterion for false touch behavior, and follows the phase difference mapping result output by S5.1, providing an accurate probability distribution input for calculating Shannon entropy in S5.3.

[0137] The system receives the set of four subordinate subgroup phase difference sequences generated by S5.1 and defines them as input data streams. Each subgroup contains the real-time deviation value of each sensing unit in the corresponding spatial region relative to the reference phase of the master control reference group.

[0138] Set the length parameter W and the step size parameter S of the sliding time window. The window length W covers a number of sampling points sufficient to reflect the short-term dynamic characteristics of the gripping action, and the step size S determines the frequency of data updates. It is usually set to one-quarter of the window length to achieve overlapping sampling.

[0139] Based on the current system clock, a sliding window truncation operation is triggered to extract the data of the most recent W sampling points from the continuous phase difference sequence of each subordinate subgroup, and a local phase difference data matrix is ​​constructed to ensure the time alignment and integrity of the data.

[0140] Histogram statistical preprocessing is performed on the extracted local phase difference data segments to divide the entire phase difference value range into K equal-width intervals. The frequency of samples falling into each interval is counted to form an initial frequency distribution vector.

[0141] A kernel density estimation algorithm is introduced to smooth the initial frequency distribution. A Gaussian kernel function is selected as the kernel function, and the smoothing degree is controlled by the bandwidth parameter h to eliminate quantization noise caused by discrete sampling and restore continuous probability density features.

[0142] The kernel density estimation process is represented by the MathML formula, and the probability density estimate at any phase difference x is calculated. The bandwidth parameter h is automatically optimized according to the Silverman rule to balance bias and variance, ensuring that the probability density curve is neither overly smoothed and loses details, nor overfits noise.

[0143] Within a preset range of phase difference values, the probability density estimate f(x) of each point is calculated by traversing with a fixed step size, generating a set of discrete probability density point pairs.

[0144] The generated probability density point pairs are interpolated and fitted to construct a continuous and smooth local phase difference probability density distribution curve, ensuring that the total area under the curve is normalized to 1, which conforms to the basic properties of the probability density function.

[0145] The generated local phase difference probability density distribution curve is encapsulated into a standard data structure, including the curve coordinate point set, bandwidth parameter and statistical confidence index, as the direct input for S5.3 to calculate Shannon entropy.

[0146] By using kernel density estimation, the discrete phase difference data segments from the previous step are transformed into continuous and smooth local phase difference probability density distribution curves. This enables a refined characterization of the multi-channel phase synchronization deviation distribution under gripping conditions, effectively overcoming the feature instability problem caused by the sensitivity of bin partitioning in traditional histogram methods. It provides a robust probabilistic model foundation for subsequent false touch confidence assessment based on information theory.

[0147] S5.3: Based on the local phase difference probability density distribution curve, the probability density value is processed by nonlinear transformation using the Shannon entropy calculation formula to quantitatively extract the real-time phase difference entropy scalar value that characterizes the degree of disorder in the phase difference distribution.

[0148] The set of discretized probability density values ​​corresponding to the local phase difference probability density distribution curve is received as the input data source for Shannon entropy calculation. This data source characterizes the distribution of the phase deviation of each subordinate subgroup relative to the master reference group in numerical space within the current sliding time window.

[0149] A non-zero value screening process is performed on the discretized probability density value set to remove zero-probability terms caused by quantization errors or computational precision limitations, thus constructing an effective probability distribution vector. Zero-probability terms can lead to numerical overflow errors in logarithmic operations; this removal operation ensures the numerical stability of subsequent nonlinear transformations.

[0150] Based on the effective probability distribution vector, the natural logarithm function is applied to perform element-wise mapping on each probability density component, generating a probability logarithmic sequence. This step transforms the multiplicative relationship in the probability space into an additive relationship in the logarithmic space, laying the foundation for linear weighting of information.

[0151] The self-information contribution of each component is calculated by multiplying the probability density value with its corresponding probability logarithm. Specifically, each effective probability density component is multiplied by its corresponding negative natural logarithm to obtain the scalar contribution of information entropy carried by that component.

[0152] The self-information contribution values ​​of all components are summed using the Shannon entropy calculation formula to quantitatively extract the real-time phase difference entropy scalar value that characterizes the degree of disorder in the phase difference distribution.

[0153] The calculated real-time phase difference entropy value is subjected to a dimensional consistency check to confirm that its numerical range conforms to the positive semi-definite property defined in information theory. If the calculation result is negative, it is determined that the probability distribution estimation is abnormal, triggering the re-execution of the probability density function estimation process in step S5.2.

[0154] Through the above Shannon entropy nonlinear transformation, the local phase difference probability density distribution curve generated in the previous step is transformed into a single real-time phase difference entropy scalar, realizing a quantitative characterization of the discreteness of the rhythmic coordination of multi-channel pressure signals, and providing a core indicator for the subsequent false touch confidence enhancement criterion based on historical baseline.

[0155] For example, if the sliding time window length is set to 500ms and the sampling frequency is 100Hz, then the window contains 50 sampling points. The phase difference value range [-π,π] is uniformly divided into 32 discrete intervals, each interval having a width of approximately 0.196 radians. At a moment when a monotonically increasing phase consistency index is detected, the frequency of each subordinate subgroup's phase difference falling into each interval is counted, and the probability distribution vector is obtained after normalization. Assume that the probability density values ​​of the 5 main intervals are 0.35, 0.25, 0.20, 0.15, and 0.05, respectively, and the probability of the remaining intervals is close to 0 and is therefore eliminated. Substituting into the Shannon entropy formula, we calculate: H = -(0.35ln(0.35) + 0.25ln(0.25) + 0.20ln(0.20) + 0.15ln(0.15) + 0.05ln(0.05)). The contributions of each self-information element are calculated as follows: 0.35*(-1.0498)≈-0.3674, 0.25*(-1.3863)≈-0.3466, 0.20*(-1.6094)≈-0.3219, 0.15*(-1.8971)≈-0.2846, 0.05*(-2.9957)≈-0.1498. Summing and inverting these values ​​yields the real-time phase difference entropy value H≈1.4703 bits. This value is significantly higher than the baseline entropy value under normal stable grip conditions (typically below 1.2 bits), indicating that the rhythmic synchronicity of the pressure distribution is rapidly disintegrating, and the phase difference distribution is becoming more homogenized and disordered. Based on this, the system generates a high-confidence auxiliary criterion signal for the risk of accidental touch.

[0156] S5.4: Based on the real-time phase difference entropy scalar, call the pre-stored historical normal grip state phase difference entropy value database, perform percentile statistical retrieval processing, and generate the historical baseline 95th percentile threshold representing the upper limit of normal grip fluctuation.

[0157] Receive the real-time phase difference entropy scalar value generated by the preceding step S5.3. This scalar quantifies the degree of disorder in the phase distribution of the subordinate subgroup relative to the master reference group within the current sliding time window, and serves as the query index for dynamic threshold retrieval.

[0158] The system calls upon a pre-built historical normal grip state phase difference entropy value database in the embedded memory. This database stores historical sequence data of phase difference entropy values ​​collected and cleaned over a long period of time under non-accidental touch scenarios such as static gripping of a folding umbrella, slight adjustments, and walking vibrations.

[0159] The historical phase difference entropy value sequence in the database is sorted and preprocessed. The fast sorting algorithm is used to arrange the discrete entropy value samples in ascending order of value to construct an ordered set of statistics, which provides a data structure foundation for percentile retrieval.

[0160] Based on the set of ordered statistics, the total number of samples N is determined, and the index position of the target quantile is calculated according to the preset confidence level of 95%. The entropy threshold corresponding to the 95th percentile is accurately estimated between two adjacent ordered samples using linear interpolation.

[0161] Specifically, let E be an ordered sequence of historical phase difference entropy values. sorted ={e_1,e_2,...,e_N}, where e_i≤e_{i+1}, calculate the target index k=0.95×(N-1)+1. If k is an integer, then directly take e_i. k If k is not an integer, let the integer part of k be floor(k) and the fractional part be frac(k). Then, the historical baseline 95th quantile threshold T is calculated using the following formula. 95 : Among them, T 95 The historical baseline 95th percentile threshold, e floor (k) is the entropy value of the floor(k) position after sorting, e floor (k)+1 is the entropy value of the floor(k)+1 position after sorting, and k is the target quantile index.

[0162] The calculated historical baseline 95th quantile threshold T 95 The output is sent to the subsequent comparison module. This threshold characterizes the upper boundary of the dispersion of the phase difference distribution under 95% normal grip conditions.

[0163] Through the percentile statistical retrieval processing described above, static historical data is transformed into a dynamic discrimination benchmark that adapts to the current user's grip habits. This achieves adaptive matching of the false touch recognition threshold to individual user differences and environmental noise background, significantly reducing the false alarm rate caused by fixed thresholds.

[0164] S5.5: Based on the real-time phase difference entropy scalar value and the historical baseline 95th percentile threshold, perform numerical comparison processing. When the real-time phase difference entropy scalar value exceeds the historical baseline 95th percentile threshold, generate an auxiliary criterion result signal to mark the high-confidence false touch risk state.

[0165] Step S6: Based on the preliminary judgment result and the auxiliary judgment result, when it is determined that the false touch behavior has occurred, the current phase consistency model parameters are frozen and only the drift compensation term of the reference phase trajectory is updated to generate the reference phase trajectory. Specifically, this includes: S6.1: Based on the preliminary judgment result and the auxiliary judgment result, the false touch behavior determination signal is generated. The weight and standard deviation statistical coefficient of the sliding window phase difference mean in the current phase consistency model are calculated and the parameter freezing process is performed to generate a static phase consistency model parameter set in a locked state to prevent the feature extraction logic from changing abruptly at the moment of false touch triggering.

[0166] The system receives the initial judgment result signal of the accidental touch action initiation phase generated in step S4 and the auxiliary judgment result signal generated in step S5, which serve as the logical input conditions for triggering the freezing of phase consistency model parameters. The system generates a global accidental touch behavior establishment judgment flag bit only when the initial judgment result signal indicates that the accidental touch action has entered the initiation phase and the auxiliary judgment result signal confirms a high-confidence accidental touch risk state. This flag bit serves as the enable signal for subsequent parameter locking operations.

[0167] The system reads the sliding window phase difference mean value calculation weight coefficient matrix currently in use in the phase consistency calculation module. This matrix defines the contribution distribution ratio of different subordinate subgroups in the phase difference statistics. The system copies the matrix to the read-only storage area to form a static weight snapshot, blocking the adaptive update request of the weight coefficients in subsequent sampling periods.

[0168] The standard deviation statistical coefficient used to characterize the degree of cross-channel rhythm coordination fluctuation in the phase consistency calculation module at the current moment is read synchronously. This coefficient determines the sensitivity threshold of the phase consistency index sequence to instantaneous phase jitter. Its value is locked and written into the non-volatile register to ensure that the statistical benchmark on which the feature extraction logic depends remains constant at the moment of accidental triggering.

[0169] The aforementioned locked static weight snapshot is combined and encapsulated with fixed standard deviation statistical coefficients to construct a static phase consistency model parameter set in the locked state. This parameter set remains unchanged in value during subsequent accidental touch classification inference and control command generation, avoiding divergence or abrupt changes in model parameters caused by drastic pressure distribution changes due to accidental touch actions.

[0170] By performing parameter freezing, the dynamically changing phase consistency model is transformed into a static reference model, eliminating the feature extraction benchmark drift caused by drastic signal fluctuations at the moment of accidental touch triggering. This achieves the stability and robustness of the accidental touch recognition logic in complex dynamic grip scenarios and provides a stable constraint boundary for subsequent independent drift compensation of the benchmark phase trajectory.

[0171] S6.2: Using the static phase consistency model parameter set of the locked state as a constraint, perform low-frequency trend component separation processing on the average phase trajectory of the master control reference group to extract the low-frequency drift component sequence that characterizes the physiological baseline drift caused by the user's long-term natural grip.

[0172] The static phase consistency model parameter set and the average phase trajectory data of the master control reference group in the locked state are received as the input basis for low-frequency trend component separation processing. This step aims to extract the low-frequency components that characterize the physiological baseline drift caused by the user's prolonged natural grip from the original phase trajectory, which contains high-frequency grip jitter and accidental touch transient interference.

[0173] Multi-scale wavelet packet decomposition was performed on the average phase trajectory of the master control reference group. An orthogonal wavelet basis was constructed using Daubechies-4 wavelet basis functions to map the time-domain phase signal to the time-frequency joint domain, generating a decomposition coefficient sequence containing different frequency sub-bands. By setting the decomposition level to 5 levels, the signal frequency band was divided into 32 equal-width sub-bands. The lowest frequency sub-band corresponds to DC and extremely slowly changing physiological drift components, while the high-frequency sub-band corresponds to hand tremors and changes in force during accidental touches.

[0174] Based on the wavelet coefficients obtained from the decomposition, a low-frequency energy proportion evaluation index is constructed, and a set of sub-band coefficients with concentrated energy and frequencies below 0.5Hz is selected. A soft thresholding denoising algorithm is used to shrink the high-frequency coefficients of non-lowest frequency sub-bands. The threshold is selected based on the noise standard deviation estimate locked by the static phase consistency model parameter set, in order to suppress the aliasing effect of high-frequency random noise on low-frequency trend extraction.

[0175] The lowest frequency subband wavelet coefficients after screening are reconstructed using inverse wavelet transform to restore the time-domain signal shape. During the reconstruction process, a sliding window weighted smoothing mechanism is introduced, with a window length set to 2 seconds and a Hanning window used as the weighting function, to further eliminate the Gibbs phenomenon and boundary effects in the reconstructed signal and generate a preliminary low-frequency drift candidate sequence.

[0176] The initial low-frequency drift candidate sequence is subtracted point-by-point from the original average phase trajectory to obtain the residual high-frequency sequence. The instantaneous power spectral density of the residual high-frequency sequence is calculated to verify whether it meets the preset high-pass characteristic, i.e., the main energy distribution is above the 1Hz frequency band. If there is still significant low-frequency residual energy in the residual sequence, the threshold coefficients of wavelet decomposition are dynamically adjusted and reconstruction is performed again until the proportion of low-frequency energy in the residual sequence is lower than the preset tolerance threshold.

[0177] Through the above-mentioned multi-scale decomposition, threshold shrinkage and iterative reconstruction processing, the average phase trajectory of the master control reference group in the previous step is transformed into a low-frequency drift component sequence that characterizes the physiological baseline drift caused by the user's long-term natural grip. This achieves precise decoupling of slow-changing physiological characteristics and fast-changing accidental touch characteristics, providing a pure low-frequency reference for the subsequent adaptive compensation of the reference phase trajectory.

[0178] S6.3: Based on the low-frequency drift component sequence, perform a first-order inertial filtering algorithm to generate a smoothed reference phase trajectory drift compensation vector, filter out high-frequency noise interference, and retain slowly changing physiological adaptation characteristics.

[0179] Receive the low-frequency drift component sequence, which is the output of step S6.2, representing the physiological baseline drift caused by the user's long-term natural grip. This sequence serves as the raw input data stream for the first-order inertial filtering algorithm and contains the phase offset value that varies with time.

[0180] The time constant parameter of the first-order inertial filter is set. Based on the typical frequency range of human muscle micro-tremors and the slow-changing characteristics of physiological drift in the scenario of holding a folding umbrella, the cutoff frequency is configured between 0.5Hz and 1.0Hz to establish the filter's attenuation rate for high-frequency noise and its speed of following low-frequency trends.

[0181] Obtain the instantaneous value of the low-frequency drift component at the current sampling time t, and denote it as the input variable x(t). At the same time, read the output value after filtering at the previous sampling time t-1, and denote it as the historical state variable y(t-1). Construct the basic data pair for recursive filtering calculation.

[0182] Iterative smoothing is performed using a first-order linear difference equation to obtain the smoothed reference phase trajectory drift compensation vector at the current time.

[0183] Based on the clock frequency of the sampling system and the preset time constant τ, the value of the filtering coefficient α is dynamically calculated to ensure that while maintaining the real-time phase tracking, high-frequency random interference caused by sensor thermal noise or slight hand tremors is effectively suppressed.

[0184] The calculated smooth output value y(t) at the current time is stored in the system buffer, and the historical state variable is updated so that it becomes y(t-1) in the calculation of the next sampling period, thus maintaining the continuity of the filtering process and the memory effect.

[0185] Boundary condition checks are performed on the continuously generated smooth output value sequence to prevent the phase compensation amount from exceeding the physical reasonable range due to excessive transient response in the initial startup phase, and to ensure that the drift compensation vector is always within the controllable range.

[0186] The low-frequency drift component sequence extracted in the previous step is transformed into a smoothed reference phase trajectory drift compensation vector by processing with a first-order inertial filtering algorithm. This filters out high-frequency noise interference and retains the slowly changing physiological adaptation characteristics, thereby achieving stable adaptive updating of the reference phase trajectory during long-term holding.

[0187] S6.4: Perform dynamic superposition correction processing on the original reference phase trajectory according to the smoothed reference phase trajectory drift compensation vector to generate an adaptive reference phase trajectory that is updated in real time and has physiological adaptability, ensuring that subsequent phase difference calculations are always based on the latest normal grip state reference.

[0188] Obtain the smoothed reference phase trajectory drift compensation vector output from step S6.3. This vector represents the low-frequency physiological baseline offset caused by muscle fatigue or postural fine-tuning during prolonged user gripping.

[0189] Read the original reference phase trajectory data stored in the phase consistency model at the current moment. This data records the cumulative historical sequence of the average phase of the master control reference group from system initialization to the current moment.

[0190] Perform timestamp alignment verification on the original reference phase trajectory data to ensure that the sampling time point of the drift compensation vector corresponds strictly one-to-one with the discrete sampling point of the original reference phase trajectory on the time axis, thus eliminating data misalignment errors caused by asynchronous sampling.

[0191] Each component value in the smoothed reference phase trajectory drift compensation vector is then superimposed point by point onto the corresponding phase angle value of the original reference phase trajectory according to the corresponding time index.

[0192] Perform dynamic overlay correction processing to generate a real-time updated and physiologically adaptive reference phase trajectory: Phase normalization constraint processing is performed on the adaptive reference phase trajectory obtained after superposition calculation. The phase angle values ​​that exceed the range of [-π,π] are mapped back to the main value range by adding or subtracting integer multiples of 2π to prevent phase entanglement from disrupting the continuity of subsequent phase difference calculation.

[0193] The processed adaptive reference phase trajectory is written into the reference register of the phase consistency index sequence calculation module, replacing the original static reference data, thus completing the online update of the reference phase trajectory.

[0194] By using a dynamic overlay correction process, the drift compensation result obtained in the previous step is transformed into real-time updated adaptive reference phase trajectory data, which enables the reference phase to automatically follow and compensate for changes in the user's physiological grip. This ensures that subsequent phase difference calculations are always based on the latest normal grip reference, significantly improving the stability of accidental touch recognition in long-term grip scenarios.

[0195] For example, in a scenario where a user holds a folding umbrella handle for more than 10 minutes, the average phase of the master control reference group slowly drifts positively due to slight hand muscle fatigue. The system reads the original reference phase trajectory value at the current time t=600s as 1.57 radians (i.e., π / 2). Simultaneously, step S6.3 outputs a drift compensation vector of 0.05 radians after first-order inertial filtering. The system directly superimposes the 0.05 radians onto the 1.57 radians, calculating an adaptive reference phase trajectory of 1.62 radians. Since 1.62 radians is within the [-π, π] interval, no phase normalization mapping is required. If at another time t=605s, the original reference phase trajectory is 3.10 radians, the drift compensation vector is 0.10 radians, and the superposition result is 3.20 radians, still within the valid interval, and is directly updated. If the superposition result exceeds π (approximately 3.14159) at a certain moment, for example, if the original value is 3.10 radians and the compensation value is 0.20 radians, the superposition result is 3.30 radians. The system subtracts 2π from this, obtaining -2.98 radians as the final adaptive reference phase. This adaptive reference phase trajectory is immediately written to the reference register for use in calculating the phase consistency index in step S7. Compared to using a fixed initial reference, this method effectively eliminates reference deviation caused by prolonged holding, reducing the calculation error of the phase difference entropy value and significantly improving the robustness of accidental touch behavior classification in long-term interaction scenarios.

[0196] S6.5: Replace the original reference phase trajectory with the adaptive reference phase trajectory and input it into the phase consistency index sequence calculation module to complete the online calibration closed loop of the reference phase trajectory and ensure the accuracy of cross-channel rhythm coordination measurement in the continuous false touch discrimination cycle.

[0197] Step S7: The phase consistency index sequence and the time-domain pressure gradient sequence are subjected to heterogeneous feature fusion processing to generate a fused feature vector containing temporal rhythm synergy and amplitude change rate. Specifically, this includes: S7.1: Perform sliding window normalization on the phase consistency index sequence to eliminate the dimensional influence caused by differences in grip strength among different users and generate a standardized phase consistency index sequence.

[0198] S7.2: Based on the original pressure time-series signal, the first derivative is calculated using the central difference algorithm to quantify the pressure transient rate of each sensing unit and generate a time-domain pressure gradient sequence.

[0199] S7.3: Apply adaptive threshold truncation filtering to the time-domain pressure gradient sequence to remove abnormal gradient spikes caused by high-frequency noise from the sensor, and generate a denoised time-domain pressure gradient sequence.

[0200] Receive the time-domain pressure gradient sequence generated in step S7.2. This sequence contains the pressure change rate values ​​of each sensing unit at continuous sampling time, and serves as the input data object for adaptive threshold truncation filtering.

[0201] The sliding window statistical characteristics of the time-domain pressure gradient sequence are calculated. A sliding time window of length N is selected, and the local mean and local standard deviation of the pressure gradient values ​​within the window are calculated to characterize the baseline level and the degree of fluctuation dispersion of pressure changes near the current time.

[0202] A dynamic adaptive threshold boundary is constructed based on the local mean and local standard deviation, and upper and lower limits are set using the three-times-standard-deviation principle.

[0203] Each sampling point in the time-domain pressure gradient sequence is compared with the dynamic adaptive threshold boundary point by point to identify abnormal gradient spike data points that exceed the upper threshold or fall below the lower threshold, and these are marked as noise regions to be corrected.

[0204] Linear interpolation repair is performed on the abnormal gradient spike data points marked as noise regions. A linear fitting function is constructed using the effective gradient data points adjacent to the outlier points, and the replacement values ​​are calculated to replace the original outliers, thus maintaining the continuity and smoothness of the pressure gradient curve.

[0205] The original values ​​of normal gradient data points that are not marked as noise regions are retained unchanged. The repaired outlier values ​​are recombined with the retained normal values ​​in chronological order to generate a denoised temporal pressure gradient sequence.

[0206] By using the above adaptive threshold truncation filtering method, the noisy time-domain pressure gradient sequence generated in the previous step is transformed into a denoised time-domain pressure gradient sequence that eliminates high-frequency noise interference, thus achieving the expected technical effect of preserving the true characteristics of erroneous momentum change while suppressing random noise from the sensor.

[0207] S7.4: Perform time axis alignment interpolation based on the standardized phase consistency index sequence and the denoised temporal pressure gradient sequence to unify the temporal resolution of the two types of heterogeneous data and generate a time-synchronized dual-modal feature matrix.

[0208] S7.5: The time-synchronized dual-modal feature matrix adopts a channel splicing and linear mapping fusion strategy to integrate the temporal rhythm coordination information and amplitude change rate information, and generate a fused feature vector containing temporal rhythm coordination and amplitude change rate.

[0209] Obtain the time-synchronized dual-modal feature matrix after time-axis aligned interpolation. This matrix contains feature data in two dimensions: the standardized phase consistency index sequence and the denoised time-domain pressure gradient sequence.

[0210] A channel dimension concatenation operation is performed on the time-synchronized dual-modal feature matrix. The standardized phase consistency index sequence is used as the first feature channel, and the denoised temporal pressure gradient sequence is used as the second feature channel to construct the original fusion feature tensor with a dual-channel structure.

[0211] Based on the original fusion feature tensor, a learnable linear mapping weight matrix is ​​introduced. Through matrix multiplication, the high-dimensional or heterogeneous feature space is projected to a unified low-dimensional latent feature space to eliminate the semantic gap between different physical dimensions.

[0212] The intermediate feature vector after linear mapping is transformed nonlinearly by applying the ReLU activation function to enhance the model's ability to represent the complex nonlinear temporal rhythm coordination and amplitude change rate coupling relationship in false touch behavior.

[0213] The activated feature vectors are normalized using the L2 norm to constrain the feature vector magnitude, prevent feature value explosion caused by sudden pressure changes in individual sensing units, and improve the numerical stability of subsequent graph convolutional network classification inference.

[0214] By employing a channel concatenation and linear mapping fusion strategy, the time-synchronized dual-modal feature matrix from the previous step is transformed into a fused feature vector containing temporal rhythm synergy and amplitude change rate. This achieves efficient integration and semantic unification of heterogeneous features, providing high-quality input for fine-grained classification of subsequent lightweight graph convolutional networks.

[0215] For example, in a folding umbrella holding scenario, the time-synchronized bimodal feature matrix X_fusion has a dimension of 12×2, where 12 represents the 12 pressure sensing units of the umbrella handle ring array, and 2 represents the two feature channels: the standardized phase consistency index and the denoised temporal pressure gradient. The linear mapping weight matrix W is set to have a dimension of 2×8, and the bias vector b has a dimension of 8×1, aiming to map the bimodal features into an 8-dimensional fused feature vector. For a certain accidental touch event, the standardized phase consistency indices of the first to third sensing units are 0.85, 0.82, and 0.79, respectively, and the denoised temporal pressure gradients are -0.15, -0.12, and -0.10, respectively; the indices of the remaining sensing units are close to zero. Inputting X_fusion into the formula, and after weighted summation of the W matrix and superposition of the bias b, an intermediate feature vector is obtained. For example, the first fused feature component is calculated as (0.85×w_11+(-0.15)×w_21+...+b_1). Assuming w_11=0.5, w_21=-0.3, and b_1=0.1, the result is 0.85×0.5+(-0.15)×(-0.3)+0.1=0.425+0.045+0.1=0.57. After performing similar calculations on all 8 components, the ReLU activation function is applied, setting negative values ​​to zero and retaining positive values. Finally, L2 normalization is performed. Assuming the activated vector is [0.57,0.0,0.32,...,0.11], and its L2 norm is 0.75, the normalized fused feature vector F is [0.76,0.0,0.43,...,0.15]. This fused feature vector significantly enhances the coupling feature representation of sliding accidental touches in the phase discretization and pressure gradient directions, enabling subsequent graph convolutional networks to more accurately distinguish between sliding accidental touches and squeezing accidental touches, resulting in a significant improvement in classification accuracy.

[0216] Step S8: Input the fused feature vector into the graph convolutional network for classification inference, output fine-grained classification results, and trigger corresponding anti-accidental touch control commands. Specifically, this includes: S8.1: Based on the spatial topological relationship of each sensing unit in the fused feature vector, a dynamic adjacency matrix is ​​constructed, and the fused feature vector containing temporal rhythm coordination and amplitude change rate is mapped to graph structure data node attributes to generate a weighted graph signal representation with spatial correlation information.

[0217] S8.2: Perform multi-layer graph convolution aggregation operation on the weighted graph signal representation, and use learnable weight parameters to propagate and fuse node feature information in the neighborhood to generate a high-dimensional graph embedding feature representation containing the local pressure distribution evolution pattern.

[0218] The system receives a weighted graph signal representation with spatial association information generated in step S8.1. This representation includes graph nodes consisting of 12 pressure sensing units and their corresponding fused feature vectors, as well as a dynamic adjacency matrix constructed based on spatial topological relationships.

[0219] The weighted graph signal representation is subjected to a first-layer graph convolution aggregation operation, which uses learnable weight parameters to propagate feature information between directly adjacent sensing unit nodes to capture the microscopic spatial correlation of local grip pressure.

[0220] Based on the node connection relationship determined by the dynamic adjacency matrix, the fusion feature vector of each target node and the feature vector of its first-order neighbor nodes are weighted and summed to generate an intermediate state of node features after preliminary aggregation.

[0221] The intermediate state of the node features is input into the linear transformation layer, and the low-dimensional fused features are mapped to the high-dimensional latent space through matrix multiplication to enhance the nonlinear expressive power of the features and eliminate the interference caused by the difference in dimensions.

[0222] The feature vector after linear transformation is processed by nonlinear rectification using the ReLU activation function. Negative responses are eliminated while positive pressure change patterns with physical meaning are retained, generating the first layer of graph embedding feature representation.

[0223] The second-layer graph convolution aggregation operation is performed on the first-layer graph embedding feature representation to expand the receptive field range to capture the evolution trend of macroscopic pressure distribution across regions, thereby identifying global features of large-scale accidental touch actions such as sliding and squeezing.

[0224] Again, based on the dynamic adjacency matrix, the second-order neighbor node information of each node in the first-layer graph embedding feature representation is aggregated to simulate the diffusion effect of pressure waves on the umbrella handle grip interface.

[0225] The aggregated second-order neighbor features are coupled with the first-layer graph embedding features of the current node using residual connection processing to preserve the original local details and prevent the gradient vanishing problem in deep networks.

[0226] A second linear transformation and nonlinear activation process are performed on the feature vectors after residual connection to generate a high-dimensional graph embedding feature representation containing the local pressure distribution evolution pattern.

[0227] By using multi-layer graph convolution aggregation and residual connection processing, the weighted graph signal representation from the previous step is transformed into high-dimensional graph embedding feature representation data, achieving the expected technical effect of deep feature extraction of the spatial topology and temporal evolution of accidental touch behavior.

[0228] S8.3: Based on the high-dimensional graph embedding feature representation, a nonlinear transformation is performed through a fully connected layer to map the high-dimensional abstract features to a preset mis-touch behavior category space, so as to generate posterior probability distribution vectors corresponding to three types of targets: sliding mis-touch, squeezing mis-touch, and hovering jitter.

[0229] It receives a high-dimensional graph embedding feature representation generated by multi-layer graph convolution aggregation operation. This feature vector contains the local pressure distribution evolution pattern and cross-channel rhythmic coordination information of each sensing node on the umbrella handle grip interface under spatial topological association.

[0230] The high-dimensional graph embedding feature representation is flattened to transform the multidimensional tensor structure into a one-dimensional linear feature vector, in order to adapt to the input dimension requirements of the fully connected layer and ensure the compatibility of subsequent nonlinear mapping calculations.

[0231] A fully connected neural network architecture with two hidden layers is constructed. The first hidden layer is configured with 128 neurons. The ReLU activation function is used to introduce nonlinear transformation capability. Through the linear combination of the weight matrix and the bias vector and the activation function mapping, the high-order abstract features of accidental touch behavior are initially extracted.

[0232] The Dropout regularization mechanism is applied to the output of the first hidden layer to randomly discard some neuron connections in order to prevent the model from overfitting and enhance the generalization robustness of the classifier in complex grip noise environments.

[0233] The second hidden layer is configured with 64 neurons, which continues to perform deep nonlinear transformations on higher-order abstract features, further compressing the feature space and strengthening the separability boundaries between various accidental touch behaviors.

[0234] The output of the second hidden layer is mapped to the output layer, which is configured with three neurons, corresponding to three target categories: swiping accidental touch, squeezing accidental touch, and hovering jitter. The output value is normalized using the Softmax activation function.

[0235] Through the nonlinear transformation and Softmax normalization of the fully connected layer, the high-dimensional abstract graph embedding features are transformed into posterior probability distribution vectors representing the confidence of three types of targets: sliding accidental touch, squeezing accidental touch, and hovering jitter. This achieves a precise mapping from spatial topological features to semantic category probability space, providing a quantitative basis for subsequent maximum likelihood estimation decisions.

[0236] S8.4: Perform maximum likelihood estimation decision processing on the posterior probability distribution vector, select the category label with the highest probability value as the final judgment result, so as to generate a fine-grained accidental touch behavior classification result with clear semantic definition.

[0237] Receive the posterior probability distribution vector generated by the preceding step S8.3. This vector contains the confidence values ​​of three types of target behaviors—sliding accidental touch, squeezing accidental touch, and hovering jitter—in a preset category space.

[0238] Perform dimension traversal and numerical comparison operations on the posterior probability distribution vector, initialize the maximum probability value variable to zero, and initialize the corresponding category index label to an empty state to establish the initial benchmark for decision search.

[0239] Read the value of the first component corresponding to the swiping accidental touch category in the posterior probability distribution vector in sequence, and compare it with the current maximum probability value variable.

[0240] If the value of the first component is greater than the current maximum probability value variable, then update the maximum probability value variable to the value of that component and update the category index label to the swipe accidental touch indicator.

[0241] Continue reading the value of the second component in the posterior probability distribution vector corresponding to the squeeze accidental touch category, and compare it with the updated maximum probability value variable.

[0242] If the value of the second component is greater than the current maximum probability value variable, then the maximum probability value variable is updated again to the value of that component, and the category index label is updated to the squeeze accidental touch indicator.

[0243] Finally, the value of the third component corresponding to the hovering jitter category in the posterior probability distribution vector is read and compared with the current maximum probability value variable.

[0244] If the value of the third component is greater than the current maximum probability value variable, then update the category index label to the hover jitter indicator; otherwise, keep the original category index label unchanged.

[0245] Through the above successive comparison and conditional update mechanism, the behavior category corresponding to the component with the global maximum value in the posterior probability distribution vector is determined, realizing the mapping transformation from continuous probability space to discrete semantic label.

[0246] Based on the defined category index labels, a structured classification result object is constructed, which includes behavior type identifier, confidence score, and timestamp information.

[0247] Perform integrity checks on the generated structured classification results to ensure that the category identifier belongs to one of the three predefined sets of accidental touch behaviors and that the confidence value is within the legal range.

[0248] The structured classification result object that passes the verification is output as the final judgment result, which is used for subsequent matching and instruction generation in the control strategy library.

[0249] By using maximum likelihood estimation decision processing, the high-dimensional graph embedding feature mapping result from the previous step is transformed into a fine-grained classification result of accidental touch behavior with clear semantic definition, thereby achieving accurate parsing and ambiguity elimination of user intent and providing a highly reliable decision basis for accidental touch prevention control.

[0250] S8.5: Match the preset control strategy library according to the fine-grained accidental touch behavior classification results, call the corresponding execution logic to generate anti-accidental touch control instructions, so as to trigger the folding umbrella locking mechanism or motor drive unit to complete the anti-accidental touch action intervention.

[0251] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0252] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0253] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for preventing accidental triggering of a folding umbrella based on the umbrella handle grip pressure distribution sequence, characterized in that, Includes the following steps: S1: Acquire the original pressure time-series signal and divide the original pressure time-series signal into one master control reference group and four subordinate subgroups based on spatial proximity; S2: Perform Hilbert transform processing on the original pressure timing signals in the master control reference group and the four subordinate subgroups respectively to generate a multi-channel instantaneous phase angle sequence; S3: Based on the multi-channel instantaneous phase angle sequence, calculate the mean and standard deviation of the sliding window phase difference between the average phase of each subordinate subgroup and the reference phase, and generate a phase consistency index sequence. S4: Using the monotonically increasing trend and increase data of the phase consistency index sequence, combined with the preset dynamic threshold, a preliminary judgment is made to generate the initial judgment result of the accidental touch action initiation stage. S5: Calculate the phase difference entropy value by statistically analyzing the dispersion of the phase difference distribution of each subordinate subgroup, and compare the phase difference entropy value with the 95th percentile value of the historical baseline to generate auxiliary criterion results; S6: Based on the preliminary judgment result and the auxiliary judgment result, when it is determined that the false touch behavior is established, freeze the current phase consistency model parameters and only update the drift compensation term of the reference phase trajectory to generate the reference phase trajectory. S7: Perform heterogeneous feature fusion processing on the phase consistency index sequence and the time-domain pressure gradient sequence to generate a fused feature vector; S8: Input the fused feature vector into the graph convolutional network for classification reasoning, output fine-grained classification results, and trigger the corresponding anti-accidental touch control command.

2. The method for preventing accidental triggering of a folding umbrella based on the umbrella handle grip pressure distribution sequence according to claim 1, characterized in that, The S4 specifically includes: A sliding time window truncation process is performed on the phase consistency index sequence to generate a local phase consistency index subsequence containing multiple consecutive sampling points, which serves as the basic data unit for trend analysis. Based on the local phase consistency index subsequence, perform adjacent time point difference operation processing to generate an instantaneous change rate sequence characterizing the direction of phase consistency change, which is used to quantify the slope of the index's evolution over time. The instantaneous rate of change sequence is subjected to a positive and negative sign continuity determination process to generate a monotonic flag sequence that indicates whether the phase consistency index shows a strictly monotonically increasing state, and the rising segment data that meets the characteristics of accidental touch precursors are selected. Based on the continuous and valid rising segment data in the monotonic flag sequence, the cumulative increase calculation is performed to generate a dynamic increase value that characterizes the degree of accumulation of false trigger momentum, reflecting the degree of asynchrony of pressure distribution rhythm. The dynamic increase value is compared with a preset dynamic threshold to generate a preliminary judgment result of the accidental touch action start-up stage, indicating whether the accidental touch action has entered the start-up stage, thus completing the initial capture of potential accidental touch behavior.

3. The method for preventing accidental triggering of a folding umbrella based on the umbrella handle grip pressure distribution sequence according to claim 1, characterized in that, The S5 specifically includes: The instantaneous phase angle data of four subordinate subgroups in the multi-channel instantaneous phase angle sequence are acquired. Based on the average phase trajectory of the master control reference group as the reference phase, phase difference mapping processing is performed on the instantaneous phase angle data in each subordinate subgroup to generate a set of subordinate subgroup phase difference sequences that characterize cross-channel rhythm synchronization deviation. Based on the set of phase difference sequences of subordinate subgroups, a sliding time window mechanism is used to extract local phase difference data segments at the current moment, and probability density function estimation processing is performed on the local phase difference data segments to generate a local phase difference probability density distribution curve that characterizes the phase difference distribution pattern. Based on the local phase difference probability density distribution curve, the probability density value is processed by nonlinear transformation using the Shannon entropy calculation formula, and the real-time phase difference entropy scalar value characterizing the disorder of the phase difference distribution is extracted in a quantitative manner. Based on the real-time phase difference entropy scalar, the database of pre-stored historical normal grip state phase difference entropy values ​​is called, and percentile statistical retrieval processing is performed to generate the historical baseline 95th percentile value representing the upper limit of normal grip fluctuation. Based on the real-time phase difference entropy scalar value and the historical baseline 95th percentile value, a numerical comparison logic is performed. When the real-time phase difference entropy scalar value exceeds the historical baseline 95th percentile value, an auxiliary criterion result signal is generated to mark a high-confidence false touch risk state.

4. The method for preventing accidental triggering of a folding umbrella based on the umbrella handle grip pressure distribution sequence according to claim 1, characterized in that, S3 specifically includes: performing time-dimensional mean aggregation processing on the master control reference group data in the multi-channel instantaneous phase angle sequence to generate the master control reference group average phase trajectory as a time-series alignment benchmark.

5. The method for preventing accidental triggering of a folding umbrella based on the umbrella handle grip pressure distribution sequence according to claim 4, characterized in that, S3 further includes: based on the average phase trajectory of the master control reference group, performing a sliding window truncation operation on the multi-channel instantaneous phase angle sequences in the four subordinate subgroups respectively, so as to generate a subordinate subgroup sliding window phase data block containing local timing information.

6. The method for preventing accidental triggering of a folding umbrella based on the umbrella handle grip pressure distribution sequence according to claim 5, characterized in that, S3 further includes: performing point-by-point phase subtraction operations using the corresponding window segments of the subordinate subgroup sliding window phase data block and the average phase trajectory of the master control reference group to generate a subordinate subgroup sliding window phase difference sequence characterizing local rhythm deviation.

7. The method for preventing accidental triggering of a folding umbrella based on the umbrella handle grip pressure distribution sequence according to claim 6, characterized in that, S3 specifically includes: performing first-order moment statistical calculation on the sliding window phase difference sequence of the subordinate subgroup to generate the average sliding window phase difference that characterizes the trend of cross-channel phase synchronization center.

8. The method for preventing accidental triggering of a folding umbrella based on the umbrella handle grip pressure distribution sequence according to claim 7, characterized in that, S3 specifically includes: performing second-order moment statistical calculation processing based on the sliding window phase difference sequence of the subordinate subgroup and the generated average sliding window phase difference to generate a phase consistency index sequence characterizing the degree of cross-channel rhythmic synergy fluctuation.

9. The method for preventing accidental triggering of a folding umbrella based on the umbrella handle grip pressure distribution sequence according to claim 1, characterized in that, The original pressure timing signal in S1 is synchronously acquired by no fewer than 12 pressure sensing units distributed in a ring array on the umbrella handle grip interface.

10. The method for preventing accidental triggering of a folding umbrella based on the umbrella handle grip pressure distribution sequence according to claim 1, characterized in that, S6 specifically includes: generating a false touch behavior determination signal based on the initial judgment result and the auxiliary judgment result; performing parameter freezing processing on the weight and standard deviation statistical coefficient of the sliding window phase difference mean in the current phase consistency model; and generating a static phase consistency model parameter set in a locked state.

Images