Dangerous behavior identification method and device based on multi-source physiological signal fusion

Abstract

The invention discloses a dangerous behavior recognition method and device based on multi-source physiological signal fusion, and aims to generate a physiological synchronization value reflecting a heart-brain coupling state by monitoring electroencephalogram and electrocardiosignals and calculating a phase locking value. Strong coupling signal pairs are screened based on physiological synchronization values, abnormal delay propagation paths and cascade mutation points are extracted, and dangerous trigger nodes and high-risk accumulation areas are identified through energy density analysis; the abnormal gradient field is subjected to rotation modulation innovatively through electroencephalogram signals, and a dynamic gradient field is generated to determine a dangerous behavior boundary; establishing an electrocardio resonance matching mechanism to generate a self-adaptive monitoring sequence and a grading early warning strategy; dynamically optimizing resource configuration and constructing a danger convergence field by adopting an adaptive acquisition scheme triggered by rhythm variability; and finally, early warning information is converted into a specific dangerous behavior category based on an accurate identification window, and an omnibearing intelligent monitoring solution is provided for high-risk operation safety management.

Description

Technical Field

[0001] The present invention relates to the technical field of physiological signal monitoring and behavior recognition, and in particular to a dangerous behavior recognition method and device based on multi-source physiological signal fusion. Background Art

[0002] In high-risk work environments, workers' physiological state directly impacts operational safety. Traditional safety monitoring methods, which primarily rely on external observation or monitoring of a single physiological parameter, struggle to accurately capture real-time changes in workers' physiological state, particularly early signs of dangerous behaviors such as fatigue, distraction, and stress reactions. While existing physiological monitoring systems can collect multiple physiological signals, such as electrocardiogram (ECG) and electroencephalogram (EEG), they lack in-depth analysis of the inherent correlations between these multiple signals and are unable to effectively identify abnormal changes in the heart-brain coupling state, resulting in inaccurate and in-time warnings of dangerous behaviors.

[0003] Furthermore, traditional fixed sampling strategies cannot adapt to the dynamic changes in physiological states and may miss important information at critical moments. Continuous high-density sampling also wastes computing resources and causes data redundancy. Therefore, a method is urgently needed to address at least one of these issues. Summary of the Invention

[0004] The present invention discloses a dangerous behavior identification method and device based on the fusion of multi-source physiological signals. The method aims to deeply analyze the phase-locked relationship between EEG and ECG signals, integrate the brain-heart coupling mechanism, and systematically extract key features such as physiological synchronization values, strongly coupled signal pairs, and abnormal propagation paths. Then, through adaptive acquisition triggered by gradient field rotation modulation driven by cognitive intensity and rhythm variation, the temporal and spatial evolution laws of dangerous behaviors are revealed, and ultimately a multi-level identification system with dangerous convergence fields, precise identification windows, and graded warning strategies is formed, providing real-time and accurate behavioral risk assessment for application scenarios such as high-altitude work safety monitoring, confined space work protection, heavy machinery operation warning, and extreme environment construction management.

[0005] A first aspect of the present invention provides a dangerous behavior identification method based on multi-source physiological signal fusion, comprising the following steps: Monitoring multi-source physiological signals of an operator, the multi-source physiological signals including an electroencephalogram (EEG) signal and an electrocardiogram (ECG) signal, and performing phase locking on the multi-source physiological signals to generate a physiological synchronization value; screening strongly coupled signal pairs based on the physiological synchronization value, extracting abnormal delay propagation paths from the strongly coupled signal pairs, detecting cascade mutation points along the abnormal delay propagation paths to form dangerous trigger nodes, and performing energy density analysis on the dangerous trigger nodes to identify high-risk cluster areas; extracting an abnormal gradient field from the high-risk cluster area, performing rotational modulation on the abnormal gradient field using the EEG signal to generate a dynamic gradient field, and determining a dangerous behavior boundary based on the dynamic gradient field; determining a dynamic detection accuracy based on the dangerous behavior boundary and the physiological synchronization value, performing resonance matching between the electrocardiogram signal and the dynamic detection accuracy to generate resonance detection parameters, generating an adaptive monitoring sequence based on the resonance detection parameters, performing sparse processing on the adaptive monitoring sequence to extract key detection windows, and constructing a graded early warning strategy based on the key detection windows; Performing rhythm analysis on the electrocardiogram signal to detect rhythm variability, dynamically adjusting acquisition density according to the rhythm variability to generate an adaptive acquisition plan, determining a detection period based on the adaptive acquisition plan and the graded warning strategy, and performing intermittent feature extraction according to the detection period to construct a danger convergence field; An accurate identification window is determined based on the dangerous convergence field, and the graded warning strategy is converted into a final dangerous behavior category according to the accurate identification window.

[0006] A second aspect of the present invention provides a dangerous behavior identification device based on multi-source physiological signal fusion, comprising: A signal monitoring module is used to monitor the multi-source physiological signals of the operator, the multi-source physiological signals including EEG signals and ECG signals, and perform phase locking on the multi-source physiological signals to generate physiological synchronization values; an anomaly tracking module, configured to screen strongly coupled signal pairs based on the physiological synchronization value, extract abnormal delay propagation paths from the strongly coupled signal pairs, detect cascade mutation points along the abnormal delay propagation paths to form dangerous trigger nodes, and perform energy density analysis on the dangerous trigger nodes to identify high-risk cluster areas; a gradient modulation module, configured to extract an abnormal gradient field from the high-risk cluster area, rotationally modulate the abnormal gradient field using the EEG signal to generate a dynamic gradient field, and determine a dangerous behavior boundary based on the dynamic gradient field; a resonance detection module, configured to determine a dynamic detection accuracy based on the dangerous behavior boundary and the physiological synchronization value, resonantly match the electrocardiogram signal with the dynamic detection accuracy to generate resonance detection parameters, generate an adaptive monitoring sequence based on the resonance detection parameters, perform a sparse process on the adaptive monitoring sequence to extract a key detection window, and construct a graded warning strategy based on the key detection window; a danger convergence module, configured to perform rhythm analysis on the electrocardiogram signal to detect rhythm variability, dynamically adjust acquisition density according to the rhythm variability to generate an adaptive acquisition scheme, determine a detection period based on the adaptive acquisition scheme and the graded warning strategy, and perform intermittent feature extraction according to the detection period to construct a danger convergence field; A behavior recognition module is used to determine an accurate recognition window based on the dangerous convergence field, and convert the graded warning strategy into a final dangerous behavior category according to the accurate recognition window.

[0007] The beneficial effects of the present invention are reflected in the following points: First, through the phase-locked analysis of multi-source physiological signals and the screening of strongly coupled signal pairs, deep-level correlation mining of the heart-brain coupling state is achieved, which can identify abnormal delayed propagation paths and cascade mutation points, accurately locate dangerous trigger nodes, and significantly improve the accuracy of early warning of dangerous behaviors compared with traditional single parameter monitoring. Secondly, through the detection of ECG rhythm variability and adaptive acquisition density adjustment, combined with resonance matching technology to improve signal detection quality, a physiological state-driven intelligent monitoring mechanism is established, which automatically increases the acquisition density when the rhythm variability increases and reduces the acquisition frequency during the stable period, achieving the optimal balance between monitoring accuracy and system efficiency. Finally, by adopting the dangerous convergence field construction and precise identification window positioning technology, combined with the graded warning strategy, the accurate conversion from physiological signals to specific dangerous behavior categories is achieved, which can distinguish between various dangerous states such as fatigue work, distraction, and impending disability and provide differentiated intervention suggestions, providing scientific and refined decision support for work safety management.

[0008] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The accompanying drawings herein illustrate specific examples of the technical solutions described in the present invention, and together with the specific implementation methods constitute a part of the specification, and are used to explain the technical solutions, principles and effects of the present invention.

[0010] Unless otherwise specified or defined, the same reference numerals in different drawings represent the same or similar technical features, and the same or similar technical features may also be represented by different reference numerals.

[0011] Figure 1 It is a flow chart of a dangerous behavior identification method based on multi-source physiological signal fusion according to the present invention.

[0012] Figure 2 This is a structural block diagram of a dangerous behavior identification device based on multi-source physiological signal fusion according to the present invention. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0014] It should be noted that all directional indications in the embodiments of the present application (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0015] It should also be noted that when an element is referred to as being "fixed on" or "disposed on" another element, it may be directly on the other element or there may be an intermediate element. When an element is referred to as being "connected to" another element, it may be directly connected to the other element or there may be an intermediate element.

[0016] In addition, the descriptions of "first", "second", etc. in this application are for descriptive purposes only and should not be understood as indicating or implying their relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that they can be implemented by ordinary technicians in this field. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0017] like Figure 1 As shown, an embodiment of the present invention provides a dangerous behavior identification method based on multi-source physiological signal fusion, including the following steps S110 to S160: Step S110 , monitoring the multi-source physiological signals of the operator, where the multi-source physiological signals include EEG signals and ECG signals, and performing phase locking on the multi-source physiological signals to generate physiological synchronization values.

[0018] Specifically, EEG signals are acquired through a multi-channel EEG acquisition device integrated into the worker's industrial hard hat. While maintaining its original protective function, this smart hard hat optimizes the acquisition device's layout using the international 10-20 system electrode layout. 32 electrodes are embedded in specific locations where the hard hat lining contacts the scalp, covering key brain regions such as the frontal, parietal, temporal, and occipital lobes. The EEG sampling rate is set to 1000Hz to ensure capture of high-frequency gamma-band neural oscillations. Each electrode contacts the scalp via a flexible conductive paste, with electrode impedance kept below 5kΩ to ensure signal acquisition quality. The electrodes are made of silver / silver chloride, which offers excellent biocompatibility and signal conductivity. Simultaneously, ECG signals are acquired using a three-lead ECG monitoring system. Electrodes are integrated at the junction of the hard hat's chin strap and shoulder straps, positioned at the right subclavian, left subcostal arch, and left subclavian locations, forming a standard lead II configuration. The ECG sampling rate is 500Hz, with a dynamic range of ±5mV and a resolution of 16 bits, ensuring accurate capture of subtle variations in the QRS complex. The data includes precise timestamps and millisecond-level synchronization, ensuring strict time alignment between different signal sources. The original physiological signal is filtered through a 50Hz notch filter to remove power frequency interference and a 0.5-100Hz bandpass filter to retain the effective physiological components. The filter uses a zero-phase design to avoid signal distortion.

[0019] In some embodiments, the phase locking of the multi-source physiological signals to generate a physiological synchronization value includes: generating a cognitive rhythm band through the EEG signal; forming a signal-rhythm correlation characteristic based on the ECG signal and the cognitive rhythm band; extracting a synchronization control interval within the signal-rhythm correlation characteristic; and forming a physiological synchronization value based on the locking strength of the synchronization control interval.

[0020] Generate cognitive rhythm bands from EEG signals. Perform spectral analysis on the collected multi-channel EEG signals. Use Fast Fourier Transform to convert the time domain signals to the frequency domain, with a transformation window length of 4 seconds and an overlap rate of 50%. Identify the characteristic frequency components in the EEG signals, including the five classic frequency bands of delta waves (0.5-4Hz), theta waves (4-8Hz), alpha waves (8-13Hz), beta waves (13-30Hz), and gamma waves (30-100Hz). Calculate the power spectral density (PSD(f)) = |X(f)| for each frequency band. ² / T, where X(f) is the frequency domain signal amplitude and T is the time window length. The power spectrum uses the Welch method to improve the estimation stability. Identify the peak frequency and bandwidth of each frequency band and determine the dominant and secondary components of the cognitive rhythm. Focus on theta waves related to cognitive load and beta waves related to attention. These two frequency bands can best reflect the real-time cognitive status of the operator. Use Morlet wavelet transform to extract instantaneous phase information. The wavelet basis function ψ(t)=exp(-t ² / 2σ ² )exp(i2πf0t), where σ controls the time-frequency resolution tradeoff and f0 is the center frequency. Complex wavelet transforms are used to simultaneously obtain the instantaneous phase φ(t) and instantaneous amplitude A(t) of each cognitive rhythm band, with a phase accuracy of 0.1 radian. Rhythmic information from different brain regions is spatially integrated to construct a brain-wide cognitive rhythm band atlas. The rhythm band data structure contains four-dimensional information: frequency, phase, amplitude, and spatial distribution, stored in tensor form, comprehensively describing the brain's rhythmic activity patterns and spatial synchronization characteristics.

[0021] Signal-rhythm correlation features were generated based on ECG signals and cognitive rhythm strips. The R-wave peak moment was extracted from the preprocessed ECG signal, and an adaptive threshold algorithm was used to ensure R-wave detection accuracy exceeding 99%. A series of consecutive heartbeat intervals (RR intervals) was calculated, and outliers caused by abnormal heartbeats and premature beats were eliminated. Heart rate variability analysis was performed on the RR interval series, extracting time-domain features including mean heart rate (HR), heart rate standard deviation (SDNN), and root mean square (RMSSD) of the difference between adjacent RR intervals. Frequency-domain features were calculated by resampling the RR interval series to a 4Hz uniform sequence using cubic spline interpolation. Spectral analysis was then performed to identify the low-frequency component (LF) (0.04-0.15Hz) and the high-frequency component (HF) (0.15-0.4Hz). The LF / HF ratio reflects the balance of the autonomic nervous system. A precise temporal correspondence between the ECG signal features and the cognitive rhythm strips was established, and dynamic features of both signals were synchronously extracted using a 5-second sliding time window. The mutual information (MI) between heart rate variability and EEG rhythm phase was calculated as ΣΣp(x,y)log(p(x,y) / (p(x)p(y))), where p(x,y) is the joint probability distribution estimated using the histogram method, and p(x) and p(y) are marginal probability distributions. The mutual information quantifies the nonlinear statistical dependency between the heart and brain signals. A signal-rhythm correlation matrix was constructed, with the dimensions being the number of ECG features × the number of cognitive rhythm bands, where the element values ​​represent the correlation strength between corresponding feature pairs. Singular value decomposition (SVD) was used to extract the principal components of the correlation matrix, retaining those with a cumulative contribution of 90%, and identifying the dominant pattern of heart-brain coupling. The signal-rhythm correlation characteristic data includes multi-dimensional information such as coupling strength, coupling time delay, coupling directionality, and coupling stability.

[0022] Extract synchronization control intervals within the signal-rhythm correlation characteristics. Analyze the temporal evolution of the signal-rhythm correlation characteristics and calculate the dynamic changes in correlation strength using a sliding window method with a window length of 10 seconds and a sliding step of 1 second. Identify time periods where the correlation strength continuously exceeds the set threshold as potential synchronization events. Use the phase synchronization indicator Calculate the phase locking degree of the heart and brain signals, where φ1 and φ2 are the instantaneous phases of the ECG-derived signal and the EEG rhythm, respectively. The PLV represents the average operation within the time window. The PLV value ranges from [0 to 1]. When the PLV value is close to 1, the two signals are highly phase-locked, while when it is close to 0, the phase relationship is randomly distributed and irregular. A synchronization threshold, PLV_th, is set to 0.7, and consecutive time periods with PLV greater than PLV_th are marked as candidate synchronization intervals. Candidate intervals are screened for duration, eliminating transient pseudo-synchronization events with durations less than 2 seconds and retaining stable and reliable synchronization control intervals. Detailed characteristic parameters for each synchronization control interval are calculated, including interval length, average synchronization strength, synchronization establishment time, synchronization maintenance stability, and synchronization release rate. The temporal distribution patterns of synchronization control intervals are analyzed, including interval frequency, interval interval time, interval duty cycle, and periodicity. Regular synchronization patterns are identified, reflecting the intrinsic physiological rhythms and regulatory mechanisms of heart-brain coupling. A complete mathematical description of the synchronization control interval is established, recording the start time t_start, end time t_end, synchronization strength-time curve PLV(t), and interval quality score.

[0023] The physiological synchronization value is formed based on the locking strength of the synchronization control interval. The comprehensive locking strength within each synchronization control interval is calculated. The locking strength is defined as LS=PLV×(1-σ_PLV / μ_PLV)×exp(-CV), where σ_PLV is the standard deviation of PLV within the interval, μ_PLV is the mean of PLV, and CV is the coefficient of variation. This formula comprehensively considers the three dimensions of synchronization strength, stability, and consistency. The smaller the standard deviation and the lower the coefficient of variation, the more stable and reliable the locking state. The time-weighted average of the locking strength of all identified synchronization control intervals is performed, and the weight factor w_i=T_i / T_total×Q_i, where T_i is the duration of the i-th interval, T_total is the total observation time, and Q_i is the interval quality score. Calculate the continuous instantaneous physiological synchronization value PSV(t)=ΣiLSi×K(tt i ), where LSi is the locking strength of the ith interval, K(tt i ) is the time kernel function. The time kernel function uses the Gaussian kernel with adaptive bandwidth K(τ)=exp(-τ ² / 2σ ²(t)), and the bandwidth σ(t) is dynamically adjusted based on the local synchronization density to ensure the temporal continuity and smoothness of the synchronization value. The physiological synchronization value is normalized and mapped to the standard interval [0, 1] using the sigmoid function, where 0 indicates complete loss of synchronization and 1 indicates perfect synchronization. A time series model of the physiological synchronization value is established to analyze its changing trends, periodic components, and random fluctuation characteristics. The final output physiological synchronization value sequence is provided at a 100Hz sampling rate, reflecting the dynamic changes in the operator's heart-brain coupling in real time.

[0024] Step S120 , screening strongly coupled signal pairs based on the physiological synchronization value, extracting abnormal delay propagation paths from the strongly coupled signal pairs, detecting cascade mutation points along the abnormal delay propagation paths to form dangerous trigger nodes, and performing energy density analysis on the dangerous trigger nodes to identify high-risk cluster areas.

[0025] Specifically, strongly coupled signal pairs were screened based on physiological synchronization values. From the physiological synchronization value time series, segments with sustained high synchronization values ​​were identified. Periods with PSVs > 0.8 and lasting for more than 5 seconds were labeled as strongly coupled periods. During these strongly coupled periods, the original EEG and ECG signals were backtracked to extract the corresponding signal segments. The coupling strength between different EEG channels and ECG signals was calculated, and the linear correlation between the signals was quantified using the cross-correlation function R(τ)=∫x(t)y(t+τ)dt, where x(t) is the EEG signal, y(t) is the ECG signal, and τ is the time delay. For each channel pair, the maximum cross-correlation coefficient and the corresponding delay time were calculated to construct a coupling strength matrix. A strong coupling threshold of 0.75 was set, and signal pairs with cross-correlation coefficients exceeding the threshold were selected as strongly coupled signal pairs. Statistical analysis showed that the coupling between the frontal EEG and ECG was the most significant, reflecting the regulatory effect of cognitive control on heart rate. The screened strongly coupled signal pairs were paired, with each pair consisting of an EEG channel signal and the corresponding ECG signal segment. Record the coupling parameters of each strongly coupled signal pair, including coupling strength, dominant delay time, coupling frequency band, and spatial location. Establish a database of strongly coupled signal pairs, including signal pair numbers, time tags, coupling characteristics, and raw signal data.

[0026] In some embodiments, extracting the abnormal delay propagation path from the strongly coupled signal pair includes: dividing the strongly coupled signal pair into a leading signal and a following signal; transmitting a timing scan path from the leading signal to the following signal; collecting the positions of abnormal response points on the timing scan path to form a response point set; and selecting the trajectory with the largest delay in the response point set as the abnormal delay propagation path.

[0027] Strongly coupled signal pairs are segmented into leading and following signals. The causal relationship of each strongly coupled signal pair is analyzed, and the Granger causality test is used to determine the driving direction between the signals. The direction and strength of the causal influence are determined by comparing the predictive power of one signal with and without the inclusion of historical information about the other. In most cases, EEG signals from specific brain regions serve as leading signals, while ECG signals serve as following signals, reflecting the central nervous system's regulation of cardiac activity. Feature points are labeled for the leading signal, identifying key features such as peaks, valleys, and zero crossings. Corresponding response features are identified in the following signal, and a mapping relationship between the feature points is established. The time delay between each feature point pair is calculated to form a delay sequence. The statistical characteristics of the delay sequence are analyzed. Under normal circumstances, delays should remain relatively stable, while abnormal delays manifest as sudden increases or dramatic fluctuations. The signal pairs are segmented into 5-second windows, and the leader-follower relationship is independently determined within each segment to capture dynamically changing causal patterns. A complete description of the leading and following signals is established, including signal type, sampling rate, time range, and feature point sequence.

[0028] A temporal scan path is launched from the leader signal to the follower signal. A scan starting point is defined at each sampling point of the leader signal, and a scan ray is projected onto the time axis of the follower signal. The direction of the scan ray is determined by the local signal gradient, which reflects the signal's changing trend. The temporal scan path equation P(t)=S0+∫v(τ)dτ is constructed, where P(t) is the path position at time t, S0 is the starting point position, and v(τ) is the scan velocity function. The integral represents the cumulative effect of velocity. The scan velocity is adaptively adjusted based on local signal characteristics, speeding up the scan during stable signal segments and slowing down during feature-rich segments. A scan window width of 100ms is set to ensure capture of the complete neural conduction process. Signal amplitude changes, phase evolution, and frequency drift along the scan path are recorded during the scan. A dynamic time warping algorithm is used to optimize the scan path to minimize the distance between the leader signal pattern and the follower signal response. Multiple parallel scan paths are established to cover different time scales and propagation modes. A path integral is calculated for each scan path to quantify the cumulative change in the signal along the path. The geometric characteristics of the scan path, including path length, curvature, and torsion, reflect the complexity of signal propagation. Visualization techniques are used to project the sequential scan path onto a two-dimensional time-amplitude plane, providing an intuitive representation of the signal propagation process.

[0029] Exemplarily, the collecting the positions of abnormal response points on the timing scanning path to form a response point set includes: analyzing the response conversion key points in the timing scanning path; determining the available natural abnormal direction at the response conversion key points; evaluating the detection reliability of the natural abnormal direction and generating a reliability evaluation result; generating a response point set based on the natural abnormal direction and the reliability evaluation result.

[0030] Analyze key response transition points along the temporal scan path. Locations along the temporal scan path where qualitative changes occur in the signal response pattern are identified; these locations correspond to transitions in physiological states. Curvature changes along the path are calculated; points with maximum curvature indicate sharp shifts in the response. Wavelet transforms are used to detect signal singularities; at these points, the wavelet coefficients exhibit a characteristic conical distribution. Criteria for identifying key response transition points include: an amplitude jump exceeding 2 standard deviations, a phase abrupt change exceeding π / 4, and a sudden change in frequency components. Local feature vectors are extracted at each key point, encompassing amplitude, phase, frequency, and path curvature. Statistical analysis shows that an average of 3-5 key response transition points occur every 10 seconds. Excessive frequency suggests a disturbance in physiological regulation. A transition probability matrix is ​​constructed between key points to analyze the switching patterns of response patterns. The time intervals between key points follow a specific probability distribution, which normally approximates a Poisson distribution, but exhibits clustered bursts when abnormal.

[0031] Exploitable natural anomaly directions are identified at key response transition points. Within the neighborhood of each key response transition point, the possible directions of signal evolution are analyzed. A vector field in the signal state space is calculated to describe the dynamic evolution of the signal. Natural anomaly directions are defined as characteristic directions that deviate from the normal evolution trajectory and are obtained by calculating the unstable eigenvectors of the locally linearized system. Eigenvectors with positive eigenvalues ​​point to unstable directions, along which the signal is prone to abnormal evolution. A local linearization model is constructed at the key points to predict the short-term evolution trend of the signal. Directions that cause rapid divergence are identified as the main anomaly directions; these directions are associated with the instability patterns of the physiological system. The angle between the anomaly direction and the normal evolution direction is calculated; a larger angle indicates a more severe anomaly. Taking into account physiological constraints, physiologically impossible anomaly directions are eliminated. A parameterized representation of the anomaly direction is established, including the starting direction, a unit anomaly vector, and an anomaly intensity parameter.

[0032] Evaluate the reliability of detecting natural abnormal directions and generate reliability assessment results. Calculate a directional stability index and evaluate the consistency of directional angles across multiple tests. Use the Monte Carlo method to add varying levels of noise to the signal to test the robustness of abnormal direction detection. Calculate reliability scores based on three dimensions: stability, consistency, and physiological plausibility. Consistency is assessed by comparing the similarity of abnormal directions at adjacent key points; a high degree of similarity indicates a reliable detection result. Physiological plausibility is determined based on a medical knowledge base, eliminating test results that violate physiological laws. Establish a reliability grading standard: high reliability, medium reliability, and low reliability. Generate assessment results with reliability labels to provide quality control information for subsequent processing. Perform secondary confirmation or elimination of low-reliability test results to ensure the accuracy of the analysis results.

[0033] Generate a set of response points based on the natural anomaly direction and reliability assessment results. Filter anomaly directions with high and medium reliability levels and include the corresponding key points in the response point set. Give each response point a complete attribute description, including timestamp, path location, anomaly direction, reliability, and feature vector. Use the convex hull algorithm to determine the boundaries of the response point set and identify the impact range of the anomaly response. Calculate the spatial distribution characteristics of the response points and analyze the uniformity of the anomaly distribution. Establish an association network between the response points, and the edge weights represent the strength of the mutual influence between point pairs. Identify key nodes and community structures in the network through graph theory analysis, and reveal the organizational pattern of anomaly propagation. Temporally sort and spatially index the response point set to support efficient query and analysis operations.

[0034] The trajectory with the largest delay in the response point set is selected as the abnormal delay propagation path. In the response point set, the propagation delay of each point relative to the leading signal is calculated. Delay time T i = t response,i -t trigger,i , where t response,i is the response time of the i-th point, t trigger,i The corresponding trigger moment is denoted by the delay time. All response points are sorted by delay time, and a subset of points with significantly increased delay is identified. A dynamic programming algorithm is used to find the optimal path between the response points, with the objective function being to maximize the cumulative delay along the path. Path constraints include temporal monotonicity and spatial continuity. A delay propagation graph is constructed, with vertices representing response points and edges representing connections that satisfy the constraints. Edge weights comprehensively consider both delay increment and spatial distance. A shortest path algorithm is used to search the graph for the optimal path from the earliest response point to the point with the largest delay. The identified abnormal delay propagation paths contain an average of 20-30 response points, with total path delays reaching 2-5 times the normal value. The geometric characteristics of the paths reveal a non-linear propagation pattern, with significant detours and stagnation. The physiological plausibility of the paths is verified to ensure that the paths conform to the basic laws of neural conduction. Complete information on the abnormal delay propagation paths is recorded, including the path point sequence, delay time for each segment, cumulative delay curves, and path visualization data.

[0035] Cascading mutation points along the abnormal delay propagation path are detected to form dangerous trigger nodes. A mutation intensity metric, M(t) = |dF / dt| / |F|, is defined, where F is a comprehensive feature vector containing amplitude, frequency, and phase information, dF / dt is the time derivative of the feature vector, and |·| represents the modulus of the vector. When the mutation intensity exceeds a threshold of 3, it is marked as a potential mutation point. A Bayesian change point detection algorithm is used to precisely locate the moment of mutation occurrence. The algorithm identifies jumps in model parameters by calculating posterior probabilities. Cascading mutations are defined as multiple mutation points occurring consecutively within a short period of time, indicating linked instability in a physiological system. Statistics show that cascading mutations typically contain 3-7 mutation points and exhibit an avalanche-like spread pattern. The causal relationships between mutation points are analyzed, and the direction and intensity of information flow are quantified using transfer entropy. Clusters of mutation points with strong causal connections are identified as dangerous trigger nodes, which may trigger systemic physiological disorders. Characteristics of dangerous trigger nodes include high mutation intensity, strong cascading effects, and low resilience. Detailed information for each dangerous trigger node is recorded, including location, time, mutation type, and impact range.

[0036] Perform energy density analysis on the danger trigger nodes to identify high-risk clusters. Calculate the signal energy Ei = ∫ |S(t)| for each danger trigger node. ² dt, where S(t) is the signal amplitude at the node and the integration interval is the node influence period. Energy density is defined as the amount of energy per unit space-time volume. The spatial distribution of energy density is calculated using the kernel density estimation method ρ(x)=ΣK_h(x-xi)E i , where K_h is the kernel function with bandwidth h, x is the spatial position, xi is the position of the i-th node, and E i is the energy of the node. A Gaussian kernel function is selected, and the bandwidth is adaptively adjusted according to the density of the node distribution. Areas with energy density exceeding 2 times the average density are identified as high-risk areas. High-risk clusters are usually elliptical or irregular in shape, reflecting the non-uniform distribution of abnormal energy. Calculate the geometric parameters of the cluster, including area, perimeter, ratio of major to minor axis, and compactness. Analyze the temporal evolution of the cluster, and track the formation, development, and dissipation of high-risk areas. Establish a risk level classification: three levels: extremely high risk, high risk, and medium risk. Generate a risk report for each high-risk cluster, including location, range, peak energy density, duration, and potential hazard assessment.

[0037] Step S130 , extracting the abnormal gradient field of the high-risk cluster area, performing rotational modulation on the abnormal gradient field through the EEG signal to generate a dynamic gradient field, and determining the dangerous behavior boundary based on the dynamic gradient field.

[0038] Specifically, the abnormal gradient field of the high-risk cluster area is extracted. From the high-risk cluster area, the spatial gradient of energy density is calculated. , where ρ is the energy density function, and The gradient vectors are the partial derivatives of energy density in the x and y directions, respectively. The gradient vector points in the direction of the fastest increase in energy density, and the magnitude of the gradient reflects the magnitude of the energy change. The gradient value at each point is calculated using the finite difference method, with the difference step size adaptively adjusted based on the spatial resolution of the cluster, typically between 0.5 and 2 mm. A complete gradient field distribution map is constructed for each high-risk cluster. Each location in the map has a corresponding gradient vector, forming a vector distribution similar to a flow field. Anomalous patterns in the gradient field are identified, including gradient reversals, vortex structures, and singularities. Gradient reversals manifest as opposite gradient directions in adjacent regions, indicating instability in energy flow and potential antagonistic patterns. Vortex structures are identified by calculating the curl of the gradient field. Regions with non-zero curl exhibit circulation patterns, potentially leading to localized energy accumulation. Singularities are defined as locations where the gradient is zero but the surrounding gradients vary dramatically. These points are centers of energy accumulation or sources of divergence. Statistical analysis shows that each high-risk cluster contains an average of 2-5 singularities and 1-3 vortex structures, with the density of singularities positively correlated with the risk level. A complex representation of the abnormal gradient field is established, encoding both the magnitude and direction information of the gradient.

[0039] In some embodiments, the rotational modulation of the abnormal gradient field by the EEG signal to generate a dynamic gradient field includes: performing cognitive intensity recognition on the EEG signal to generate an intensity distribution area; performing rotation potential evaluation based on the intensity distribution area to form a rotation factor; using the rotation factor to perform angular interpolation rotation on the abnormal gradient field to generate a continuous rotation distribution; and performing gradient transformation based on the continuous rotation distribution to generate a dynamic gradient field.

[0040] Cognitive intensity recognition is performed on electroencephalogram (EEG) signals to generate an intensity distribution area. Extract the feature frequency band components related to cognitive load from the multi-channel EEG signals collected in step S110. Focus on analyzing two frequency bands, the theta wave (4 - 8 Hz) and the gamma wave (30 - 100 Hz), which respectively reflect working memory load and cognitive processing intensity. Use the short-time Fourier transform to obtain the time-frequency representation, with a window length of 2 seconds and an overlap of 75% to ensure capturing the dynamic changes of the cognitive state. Define the cognitive intensity index CI = P_gamma / (P_theta + ε), where P_gamma is the power of the gamma band, P_theta is the power of the theta band, and ε is a small positive number (take 0.01) to prevent division by zero error. This ratio reflects the activation degree of higher cognitive processing relative to the basic cognitive load. Calculate the cognitive intensity for each of the 32 electrode channels to form discrete intensity values covering the entire scalp surface. Use the spherical spline interpolation method to expand the discrete electrode data into a continuous intensity field, and select the thin plate spline as the interpolation basis function to ensure smoothness. Identify the high-value areas in the intensity distribution, which correspond to the active cognitive processing centers of the brain and reflect the distribution of cognitive demands for the current task. The high-intensity areas in the prefrontal lobe are usually related to executive control, planning, and decision-making, the high-intensity areas in the parietal lobe are related to spatial attention and sensorimotor integration, and the high-intensity areas in the temporal lobe are related to memory retrieval. Establish a hierarchical system for the intensity distribution area: high-intensity area (CI > 2.0), medium-intensity area (1.0 < CI ≤ 2.0), low-intensity area (CI ≤ 1.0). Each intensity distribution area details its spatial range, geometric center position, average intensity value, peak intensity, and time stability index.

[0041] Rotational potential is assessed based on the intensity distribution to generate rotation factors. The spatial correspondence and interaction between the cognitive intensity distribution and the anomalous gradient fields are analyzed. The spatial gradient of the intensity distribution is calculated; the gradient direction indicates the spatial trend of cognitive intensity, while the gradient magnitude reflects the severity of the change. The directional difference between the two gradient fields is assessed to quantify the potential regulatory effect of cognitive activity on energy distribution. When the cognitive gradient and the energy gradient are oriented in different directions, this indicates the presence of cognitive regulation, generating the potential for rotational modulation. The local rotation angle is calculated for each spatial location, representing the deflection angle of the cognitive gradient relative to the energy gradient. The spatial distribution characteristics of the rotation angles are statistically analyzed; a normal distribution indicates uniform rotational modulation, while a multimodal distribution indicates the presence of multiple rotation centers. Based on the rotational potential assessment results, a complex rotation factor is calculated, encoding both the magnitude and direction of the rotation. The rotational strength is determined by the degree of non-collinearity between the two gradient fields, with the maximum rotational strength occurring when the two gradient fields are perfectly perpendicular. The rotation factor is spatially smoothed, and an adaptive Gaussian filter is used to remove local noise. The filter bandwidth is dynamically adjusted based on the characteristic scale of the intensity distribution. A temporal evolution model of the rotation factor is established to analyze the dynamic characteristics and periodicity of the rotation pattern. The attractor structure of the rotation factor is identified by phase space reconstruction technology. The stable limit cycle represents periodic rotation, and the chaotic attractor represents complex rotation dynamics.

[0042] A rotation factor is used to perform angular interpolation rotation on the anomalous gradient field, generating a continuous rotation distribution. The rotation operation is implemented through complex multiplication, maintaining the gradient field's magnitude while changing its direction. To ensure the spatial continuity of the rotation field, bilinear interpolation is used to smoothly transition angles between discrete grid points. The interpolation process considers the local topological structure of the gradient field, employing a more refined interpolation strategy near singular points. An adaptive interpolation algorithm is designed to dynamically adjust the interpolation density based on the rate of change of the local rotation angle, adding interpolation points in areas of drastic change. The fundamental topological invariance of the gradient field is strictly maintained during the rotation process, with the location of singular points remaining fixed while the surrounding field lines rotate in coordination. Quaternion representation is used to handle large rotations, avoiding singularities in Euler angle representation and ensuring numerical stability. A boundary handling strategy for the rotation field is established, employing natural boundary conditions to ensure a smooth transition of gradient directions tangentially at the boundaries. The divergence and curl of the rotated gradient field are calculated to verify the physical plausibility and energy conservation properties of the field. The researchers generated high-resolution, continuous rotation distribution maps with a spatial resolution of 1mm and a temporal resolution of 100ms, fully capturing the spatiotemporal details of rotational modulation. Statistical analysis of the rotation distribution revealed the intensity and range of cognitive modulation, with average rotation angles ranging from 30-60 degrees and reaching a maximum of 90 degrees, demonstrating the significant impact of cognition on energy flow.

[0043] Based on the continuous rotation distribution, the gradient transformation is implemented to generate a dynamic gradient field. The static rotation gradient field is expanded into a dynamic field that changes continuously over time, reflecting the real-time interaction between cognitive state and energy distribution. The time-varying gradient field G_d(x,y,t)=G ' (x,y)·M(t), where G ' (x,y) represents the rotated spatial gradient field, and M(t) represents the temporal modulation function, ranging from [0,1]. The temporal modulation function is obtained by analyzing the dynamic characteristics of EEG signals and reflects transient changes and rhythmic fluctuations in cognitive state. The envelope of the EEG signal is extracted using the Hilbert transform, and the normalized value of the envelope is used as the modulation intensity. The gradient field is subjected to spatiotemporal filtering, using edge-preserving anisotropic diffusion filtering in the spatial dimension and adaptive Kalman filtering in the temporal dimension. The filtering parameters are dynamically adjusted based on the local characteristics of the signal, reducing the filtering intensity in feature-rich regions to preserve details. Key characteristic quantities of the dynamic gradient field are calculated, including instantaneous intensity, rate of change, stability index, and dominant frequency component. Stability analysis uses a local linearization method, and the stability of dynamic behavior is determined by eigenvalues. Typical patterns in the dynamic gradient field are identified: stable equilibrium, periodic oscillation, quasi-periodic motion, and chaotic wandering. The generated dynamic gradient field integrates multidimensional information on spatial distribution, temporal evolution, cognitive regulation, and energy flow, forming a complete dynamic risk assessment framework.

[0044] Determine the boundary of dangerous behavior based on the dynamic gradient field. Analyze the spatial structure of the dynamic gradient field and determine the danger threshold through the statistical distribution of the gradient intensity. Define the gradient threshold G_th=μ_G+2σ_G, where μ_G is the spatial average value of the gradient field and σ_G is the standard deviation. This threshold corresponds to the upper bound of the 95% confidence interval. Areas where the gradient intensity exceeds the threshold are marked as potential danger zones. These areas have drastic energy changes and are prone to triggering abnormal behavior. Use the level set method to accurately track the boundary evolution of the dangerous area. The boundary is defined as a specific isosurface of the gradient field. The level set evolution equation is: , where φ is the level set function (positive value indicates safe area, negative value indicates dangerous area), F is the boundary normal movement speed, is the modulus of the gradient of the level set function. The boundary movement speed F is jointly determined by the local gradient field strength and the cognitive regulation strength, reflecting the dynamic diffusion process of risk. Calculate the geometric characteristic parameters of the boundary, including instantaneous perimeter, enclosed area, average curvature and shape complexity. High curvature segments represent unstable parts of the boundary, which are prone to topological changes such as splitting or merging. Monitor topological events of the boundary, including the generation of new danger zones, the expansion or contraction of existing danger zones, and the fusion of multiple danger zones. Establish a quantitative criterion for boundary stability, comprehensively considering the boundary movement speed, curvature change rate and area change rate. When any indicator exceeds the critical value, it is determined that the boundary is in an unstable state and requires key monitoring.

[0045] Step S140: Determine the dynamic detection accuracy based on the dangerous behavior boundary and the physiological synchronization value, resonate the ECG signal with the dynamic detection accuracy to generate resonance detection parameters, generate an adaptive monitoring sequence based on the resonance detection parameters, perform sparse processing on the adaptive monitoring sequence to extract key detection windows, and construct a graded warning strategy based on the key detection windows.

[0046] Specifically, dynamic detection accuracy is determined based on the dangerous behavior boundary and the physiological synchronization value. Dynamic features of the boundary are extracted from the dangerous behavior boundary, including boundary movement speed, deformation rate, and stability index. The distance d_min from the operator's current physiological state point to the nearest dangerous boundary is calculated, reflecting the size of the safety margin. Combined with the physiological synchronization value (PSV), a risk sensitivity index RS = PSV / d_min is defined. Risk sensitivity increases sharply when the physiological synchronization value is high and the distance to the boundary is close. A dynamic detection accuracy function P(t) = P_base × (1 + α·RS) is established, where P_base is the base detection accuracy, α is the adjustment coefficient, and RS is the risk sensitivity. The base detection accuracy is set at a 100Hz sampling rate and 16-bit quantization accuracy to meet normal monitoring requirements. As risk sensitivity increases, detection accuracy is dynamically improved, reaching a maximum of 1000Hz sampling rate and 24-bit quantization accuracy. The spatial distribution of detection accuracy is adjusted based on the geometry of the dangerous boundary, with higher detection resources allocated to areas with greater boundary curvature. In the temporal dimension, detection accuracy is dynamically adjusted with fluctuations in the physiological synchronization value, achieving optimal resource allocation. Dynamic detection accuracy encompasses not only sampling parameters but also the complexity of feature extraction, the sophistication of the analysis algorithm, and the sensitivity of the warning threshold. By adaptively adjusting detection accuracy, we optimize system efficiency while maintaining effective monitoring results.

[0047] In some embodiments, resonant matching the ECG signal with the dynamic detection accuracy to generate resonance detection parameters includes: performing rhythm decomposition evaluation on the ECG signal to generate a rhythm distribution map; establishing a resonance frequency scheme based on the rhythm distribution map to form a frequency control table; frequency aligning the frequency control table with the dynamic detection accuracy to obtain an adjustment parameter group; and resonant matching based on the adjustment parameter group to generate resonance detection parameters.

[0048] Exemplarily, the rhythm decomposition evaluation of the ECG signal to generate a rhythm distribution map includes: tracking the rhythm of the ECG signal across cycles to obtain rhythm evolution data; analyzing the contribution effect of each rhythm element based on the rhythm evolution data to construct a rhythm weight matrix, wherein each rhythm element includes basal heart rate, respiratory variability, and stress variability; and performing spatiotemporal distribution mapping on the rhythm weight matrix to generate a rhythm distribution map.

[0049] Cross-cycle rhythm tracking is performed on ECG signals to obtain rhythm evolution data. A sliding analysis window is established with a window length of 5 minutes, encompassing approximately 300-400 heartbeat cycles, to ensure statistical reliability. The window slides forward in 30-second steps to continuously track rhythm changes. Within each analysis window, the average heart rate, heart rate variability indicators, and frequency domain feature parameters are calculated. An autoregressive model is used to predict the rhythm characteristics of the next window, with the prediction error reflecting the degree of rhythm abruptness. The cross-cycle tracking algorithm uses dynamic time warping to align rhythm patterns across cycles and identify similarities and differences. A state-space model of rhythm evolution is established, with the state vector containing the current rhythm parameters and the state transition matrix describing the dynamic evolution of the rhythm. Key rhythm events are marked during the tracking process, such as rhythm pattern switching, frequency jumps, and amplitude abrupt changes. Rhythm similarity between adjacent cycles is calculated and quantified using mutual correlations. A decrease in similarity indicates a change in physiological state. A rhythm evolution dataset is generated, containing time series rhythm parameters, evolution trajectories, and event markers. Rhythm evolution data fully records the dynamic changes of ECG rhythm.

[0050] Based on rhythm evolution data, the contribution of each rhythmic component was analyzed and a rhythmic weight matrix was constructed. Rhythm evolution data were decomposed into three main components: basal heart rate (HR), which reflects overall metabolic status; respiratory variability (RV), which reflects respiratory-circulatory coupling; and stress variability (SRV), which reflects acute stress responses. Multiple regression analysis was used to assess the contribution of each component to overall HRV. The contribution of basal heart rate was calculated as the proportion of variance in the detrended residuals, with a typical value of 30-40%. The contribution of respiratory variability was assessed by integrating the high-frequency power spectrum, which accounts for 40-50% under normal conditions. Stress variability, manifested as sudden accelerations or decelerations in heart rate, was weighted by identifying abnormal events. A rhythmic weight matrix, W, was constructed, where the matrix element w_ij represents the weight of the i-th component in the j-th time window. Weight calculation took into account the stability, variability, and correlation of the component with the overall rhythm. Principal component analysis was used to verify the rationality of the weight assignments; the first three principal components should explain at least 85% of the total variance. The weight matrix was dynamically updated to reflect the changing importance of each component under different physiological states.

[0051] The rhythm weight matrix is ​​mapped to the time-space distribution to generate a rhythm distribution map. The one-dimensional time series weight matrix is ​​mapped to the two-dimensional time-frequency space to form an intuitive visual representation. The mapping process uses the kernel density estimation method to calculate the continuous distribution of rhythm intensity on the time-frequency plane. The Gaussian kernel function K(t,f)=exp(-(t-t_i) ² / 2σ_t ² -(f-f_i) ² / 2σ_f ²), where (t_i, f_i) are the time-frequency coordinates of the rhythmic event, and σ_t and σ_f are the time and frequency bandwidths. The bandwidth parameters are adaptively adjusted according to the time-frequency resolution requirements of the rhythm, with narrow bandwidth used for fast-changing rhythms and wide bandwidth used for slow-changing rhythms. The contributions of all rhythmic elements are superimposed to form a comprehensive rhythm distribution density field. Rhythm intensity is represented using a logarithmic scale to enhance the visibility of weak rhythmic components. A color mapping scheme is applied, with warm colors representing high-intensity rhythm areas and cool colors representing low-intensity areas. Key frequency lines are marked on the rhythm distribution map, such as respiratory rate, heart rate fundamental frequency, and its harmonics. Contour lines are added to show the gradient changes in rhythm intensity, making it easier to identify rhythm centers and boundaries.

[0052] Based on the rhythm distribution map, a resonant frequency scheme is established to form a frequency control table. Energy concentration areas within the rhythm distribution map are analyzed to identify stable dominant frequency components. Dominant frequencies typically include respiratory frequency (0.2-0.3Hz), baroreflex frequency (0.1Hz), and ultra-low frequency modulation (0.04Hz). For each dominant frequency, its time occupancy and average power are calculated to serve as the basis for importance scoring. A tiered resonant frequency scheme is designed, prioritizing high-importance rhythm frequencies. The first tier selects the strongest dominant frequency, typically respiratory-related frequencies, to achieve basic physiological synchronization. The second tier adds harmonic resonances of secondary frequencies to expand the frequency coverage of detection. The third tier considers cross-frequency terms to capture interactions between different rhythms. A frequency control table is constructed, with each row corresponding to a resonant mode, listing the center frequency, bandwidth, quality factor, and applicable conditions. The frequency control table supports dynamic switching, selecting the optimal resonant mode based on the characteristics of the current rhythm distribution. A frequency lock range is set, allowing for ±10% frequency deviation to accommodate natural variations in physiological rhythms.

[0053] Align the frequency control table with the dynamic detection accuracy to obtain a tuning parameter set. Read the current dynamic detection accuracy parameters, including the sampling rate, analysis window length, and update frequency. Calculate the detection system's natural frequency characteristics to determine its frequency response range and optimal operating point. Match the resonant frequency in the frequency control table with the detection system's operating frequency. Use a frequency comparison algorithm to find a parameter combination that makes the two frequencies integer multiples. When the detection frequency is an integer multiple of the resonant frequency, synchronous sampling is achieved, maximizing information acquisition efficiency. Calculate the frequency alignment error ε = |f_detect / f_resonance - round(f_detect / f_resonance)|. An error less than 0.05 indicates good alignment. For each feasible alignment solution, evaluate its implementation complexity and resource consumption. Generate a tuning parameter set, including the optimal sampling rate, buffer size, FFT point count, and filter parameters. This parameter set takes into account hardware limitations and real-time requirements to ensure feasibility on existing systems.

[0054] Resonance matching is performed based on the adjustment parameter set to generate resonance detection parameters. The optimal adjustment parameter set is selected and the resonance detection system is initialized. Precise clock synchronization is implemented to ensure that the sampling time is aligned with the characteristic points of the ECG signal. An adaptive buffering strategy is designed to dynamically adjust the buffer size to match heart rate changes. When the heart rate increases, the buffer size is reduced to improve response speed; when the heart rate decreases, the buffer size is increased to improve frequency resolution. The resonance efficiency metric η is calculated as P_signal / P_noise, where P_signal is the useful signal power and P_noise is the noise power. Resonance matching improves the signal-to-noise ratio by 3-5dB, significantly improving the detection capability of weak physiological signals. A complete resonance detection parameter configuration is generated, including sampling parameters (sampling rate, quantization bit number, trigger mode), analysis parameters (window function type, overlap ratio, frequency resolution), and synchronization parameters (phase calibration, delay compensation, and clock source selection). A dynamic parameter adjustment mechanism is established to optimize parameter settings based on real-time resonance quality feedback. The resonance detection parameters not only optimize the acquisition of a single signal but also coordinate the synchronization relationship between multiple physiological signals. Through resonance matching, intelligent monitoring driven by physiological rhythms is achieved, and the sensitivity and specificity of abnormality detection are improved.

[0055] Adaptive monitoring sequences are generated based on resonance detection parameters. The data acquisition system is configured using the resonance detection parameters to initiate synchronized multi-channel physiological signal monitoring. Monitoring sequences are generated according to the resonance period, with each sequence segment corresponding to an integer number of circadian rhythm cycles. Sequence length is dynamically adjusted based on the current risk level: shortening sequences for high-risk conditions improves temporal resolution, and lengthening sequences for low-risk conditions improves statistical reliability. A synchronization marker is embedded in each monitoring sequence, aligned with the peak of the ECG R wave, serving as a temporal reference. The sequence's information entropy is calculated as H = -Σp_ilog(p_i), where p_i represents the probability distribution of the signal and reflects the complexity of the physiological state. The sequence sampling strategy is adaptively adjusted, increasing the sampling density during periods of high entropy to capture rapidly changing physiological information. A sequence quality assessment mechanism is established to remove interfering segments and ensure data reliability. A sliding window technique is used to generate overlapping monitoring sequences with a 50% overlap rate to ensure that critical events are not missed. Real-time feature extraction is performed on the generated monitoring sequences, including time-domain statistics, frequency-domain power spectra, and nonlinear dynamics metrics. The adaptive monitoring sequence achieves close coupling of data acquisition and physiological rhythms, improving the pertinence and efficiency of monitoring.

[0056] The adaptive monitoring sequence is sparsified to extract key detection windows. The information distribution of the monitoring sequence is analyzed to identify time periods containing key physiological events. A change point detection algorithm is used to identify mutation points in the sequence, which mark the transition of physiological states. The local information content of the sequence is calculated as I(t) = -log(p(x_t|x_{t-1})), where p(x_t|x_{t-1}) is the conditional probability. The peak information content corresponds to the time when a key event occurs, such as arrhythmia, stress response, etc. An information content threshold I_th is set, and time periods exceeding the threshold are marked as candidate key windows. Cluster analysis is performed on the candidate windows, and windows with similar time are merged to avoid over-segmentation. Using sparse representation theory, the monitoring sequence is represented as a linear combination of a few key windows. Solve the sparse optimization problem: min||x-Φα|| ² +λ||α||1, where x is the original sequence, Φ is the dictionary matrix, α is the sparsity coefficient, and λ is the regularization parameter. The non-zero elements of the sparsity coefficient correspond to the positions and weights of key detection windows. The number of key detection windows extracted is reduced to 10-20% of the original sequence, while retaining over 95% of the diagnostic information.

[0057] Build a tiered early warning strategy based on key detection windows. Analyze risk indicators within key detection windows, including the degree, duration, and trend of abnormal physiological parameters. Establish a three-tiered early warning system: green (normal), yellow (caution), and red (danger). Green corresponds to physiological parameters within the normal range and risk indicators less than 1.5 times the baseline value. Yellow triggers: moderate abnormalities in any key window, such as a 30% decrease in heart rate variability, a physiological synchronization value below 0.5, or approaching the dangerous behavior boundary. Red triggers: persistent abnormalities in multiple key windows, extreme physiological parameters, or crossing the dangerous boundary. Design early warning timing logic to avoid false alarms caused by transient fluctuations. Yellow alerts require abnormalities to persist for more than 30 seconds, while red alerts require abnormalities to persist for more than 10 seconds or acute deterioration. Establish spatial early warning logic to account for coordinated abnormalities across multiple physiological systems. Simultaneous abnormalities in both ECG and EEG automatically raise the alert level. Develop differentiated response strategies: Green maintains routine monitoring; Yellow increases monitoring frequency and reminds workers to rest; Red immediately halts operations and initiates emergency procedures. Early warning information includes the current risk level, key abnormality indicators, recommended measures, and estimated recovery time. The graded early warning strategy enables accurate identification and graded management of risks, ensuring safety while avoiding excessive interference with normal operations.

[0058] Step S150, perform rhythm analysis on the ECG signal to detect rhythm variability, dynamically adjust the acquisition density according to the rhythm variability to generate an adaptive acquisition plan, determine the detection period based on the adaptive acquisition plan and the graded warning strategy, and perform intermittent feature extraction according to the detection period to construct a dangerous convergence field.

[0059] Specifically, rhythm analysis is performed on the ECG signal to detect rhythm variability. A continuous sequence of RR intervals is extracted from the ECG signal collected under the guidance of the resonance detection parameters in step S140. A sequence of differences between adjacent RR intervals, ΔRR_i = RR_{i+1} - RR_i, is calculated. This difference reflects the instantaneous changes in heart rate. Statistical analysis is performed on this difference sequence, and statistical quantities such as mean, standard deviation, skewness, and kurtosis are calculated. A rhythm variability index, RV, is defined as σ_ΔRR / μ_RR × 100%, where σ_ΔRR is the standard deviation of the RR interval differences and μ_RR is the mean RR interval. This index quantifies the degree of irregularity in the heart rate rhythm. Local rhythm variability is calculated using a sliding window technique with a window length of 60 seconds and a sliding step size of 10 seconds to capture dynamic changes in variability. Spectral analysis is performed on the rhythm variability to identify periodic and random components of the variability. Periodic variability is often associated with physiological processes such as respiration and blood pressure regulation, while random variability may reflect pathological conditions or stress responses. A multiscale analysis framework for rhythm variability was established, calculating variability indicators at different time scales, such as 5 seconds, 30 seconds, and 5 minutes. Short-term variability reflects rapid autonomic nervous system regulation, while long-term variability reflects slow humoral regulation and circadian rhythms. Two risk patterns, abnormally high variability and abnormally low variability, were identified. High variability indicates arrhythmia risk, while low variability indicates impaired autonomic nervous system function.

[0060] In some embodiments, dynamically adjusting the acquisition density according to the rhythm variability to generate an adaptive acquisition plan includes: converting the rhythm variability into a variation vector distribution; finding a stable center point from the variation vector distribution; performing density diffusion using the stable center point as a seed to form a preliminary acquisition area; and optimizing the boundaries of the preliminary acquisition area to form an adaptive acquisition plan.

[0061] Rhythm variability is converted into a variation vector distribution. A multidimensional feature space is constructed, with dimensions including instantaneous variability, short-term variation trend, long-term variation mean, and variation acceleration. Feature vectors are calculated for rhythm variability at each time point, forming a four-dimensional variation vector. Dimensionality reduction is performed using principal component analysis, retaining principal components with a cumulative contribution of 95%, typically the first two to three. In the reduced feature space, each time point corresponds to a location, forming the spatial distribution of variation vectors. The probability density distribution of the variation vectors is calculated, and a continuous density field is obtained using kernel density estimation. A Gaussian kernel function is selected, and the bandwidth is optimized through cross-validation to ensure the accuracy of the density estimation. Multiple modes are identified in the density distribution, each corresponding to a typical variation pattern. Normal variation modes typically exhibit a unimodal distribution, while abnormalities exhibit multimodal or long-tailed distributions. A temporal evolution model of the variation vector distribution is constructed to track the dynamic changes in the distribution morphology. A spreading distribution indicates increasing variability, while a contracting distribution indicates stabilization. The variation vector distribution provides a comprehensive description of rhythm variability and a geometric decision space for optimizing acquisition strategies.

[0062] Find stable centers from the distribution of mutation vectors. In the probability density distribution of mutation vectors, search for local density maxima as candidate centers. Use the mean shift algorithm to iteratively find density peaks. The algorithm converges to the point where it is a potential stable center. Calculate the stability index of each candidate center, including local density value, density gradient, and attraction domain size. Stability score , where ρ_c is the center density, is the gradient modulus, and V_basin is the volume of the attraction domain. The point with the highest stability score is selected as the primary stable center, representing the most common normal variation pattern. Secondary stable centers are identified. These points, although less dense, persist and may represent specific physiological states. The temporal persistence of stable centers is calculated; centers with longer durations are more reliable. The transition probabilities between stable centers are analyzed, and a state transition diagram is constructed to understand the switching patterns of variation patterns. The physiological significance of stable centers is verified to correspond to known physiological states (such as sleep, exercise, and stress). The coordinates and features of the identified stable centers are recorded as the starting point for subsequent density diffusion. Accurate positioning of stable centers ensures that acquisition resources are optimally allocated around key states.

[0063] Using the stable center point as the seed, density diffusion is performed to form a preliminary collection area. Starting from the identified stable center point, the diffusion equation is used to simulate the spatial propagation of the collection density. Diffusion equation , where ρ is the acquisition density, D is the diffusion coefficient, and f(ρ) is the source term. The diffusion coefficient is adaptively adjusted according to the local characteristics of the distribution of the mutation vector. High-variance areas diffuse slowly and require intensive acquisition. The source term is positive at the stable center, providing a driving force for density growth, and decays away from the center. The diffusion equation is numerically solved using the finite difference method, and the time step and spatial grid are determined according to the stability condition. The stopping condition for diffusion is set, and diffusion stops when the density falls below the threshold ρ_min or reaches the predetermined boundary. The interaction of multiple stable centers is considered during the diffusion process, and the density field is the superposition of the contributions of each center. The resulting density field defines the preliminary acquisition area. High-density areas correspond to high acquisition rates, while low-density areas reduce acquisition frequency. The coverage and resource consumption of the preliminary acquisition area are calculated to ensure maximum information acquisition under limited resources. The preliminary acquisition area provides a basic framework for refined boundary optimization.

[0064] The boundary of the preliminary acquisition area is optimized to form an adaptive acquisition scheme. The boundary characteristics of the preliminary acquisition area are analyzed to identify jagged, discontinuous, or overextended areas. Morphological operations are used to smooth the boundary, first eroding and then dilating to eliminate small protrusions and depressions. The active contour model is used to accurately adjust the boundary position, and the energy function E=∫(α|v ' | ² +β|v '' | ² )ds+∫f(v)ds, where v(s) is the parameterized boundary curve, s is the arc length parameter, and v ' The first derivative of the boundary curve represents the tangent vector, v '' The second-order derivative represents the curvature vector, and ds is the arc length element. The first term ∫(α|v ' | ² +β|v '' | ² )ds is the internal energy term, controlling the length and smoothness of the boundary. α is the weight coefficient controlling the boundary length, and β is the weight coefficient controlling the boundary smoothness. The second term, ∫f(v)ds, is the external energy term, or data term, where f(v) is the image gradient or edge strength function. This ensures that the boundary aligns with the actual image characteristics and acquisition requirements. The optimization process considers acquisition efficiency constraints, ensuring that boundary adjustments do not result in under-acquisition of critical areas. Risk weights are introduced, and the boundaries of high-risk areas are expanded outward to provide a safety margin. The area and shape complexity of the optimized acquisition region are calculated, simplifying overly complex boundaries to reduce implementation complexity. The continuous acquisition density field is discretized into a practical and feasible acquisition point configuration. A final adaptive acquisition plan is generated, specifying the acquisition parameter settings for each spatial location. The plan includes an acquisition frequency map, a channel activation table, accuracy configuration, and switching trigger conditions. This adaptive acquisition plan intelligently allocates acquisition resources, improving system efficiency while ensuring monitoring quality.

[0065] The detection period is determined based on an adaptive acquisition scheme and a hierarchical warning strategy. First, the spatial density distribution is determined based on the acquisition frequency map in the adaptive acquisition scheme. High-density, medium-density, and low-density areas in the scheme are mapped to different temporal modulation coefficients. A base detection period is set based on the current warning level: 5 minutes for green, 1 minute for yellow, and 10 seconds for red. The density modulation of the adaptive acquisition scheme is superimposed on the base period. The number of active channels is determined using the channel activation table in the acquisition scheme. Regions with more active channels have shorter detection periods, while those with fewer channels have longer periods. The comprehensive detection period, T_d, is calculated as T_base × (2-D_norm) × (1 + 0.5 × cos(2πf_physio × t)), where T_base is the base period, f_physio is the circadian rhythm frequency, t is the current time, and D_norm is the normalized acquisition density in the adaptive acquisition scheme. The cosine term introduces synchronization with the circadian rhythm. Based on the switching trigger conditions in the adaptive acquisition scheme, the detection period is automatically adjusted when parameter changes in boundary regions are detected. The lower limit of the detection cycle is set to 1 second to ensure a rapid response to acute events; the upper limit is 10 minutes to avoid monitoring blind spots.

[0066] In some embodiments, the intermittent feature extraction is performed according to the detection cycle to construct a hazard convergence field, including: locating and identifying hazard sources in the detection cycle to generate primary hazard sources and secondary hazard sources; using the primary hazard sources to evaluate the impact of the secondary hazard sources to form a hazard matrix; generating a dominant hazard vector through eigenvalue decomposition of the hazard matrix; and implementing field construction based on the dominant hazard vector to generate a hazard convergence field.

[0067] Hazard sources are located and identified during the detection cycle to generate primary and secondary hazard sources. Within each detection cycle, key physiological features are extracted and their hazard levels assessed. Hazard sources are defined as spatiotemporal locations where physiological parameters significantly deviate from the normal range. Anomaly detection algorithms are used to identify hazard signals such as sudden heart rate changes, rhythm disturbances, and abnormal waveforms. The hazard intensity of each anomaly point is calculated as R_i = |x_i - μ| / σ × duration × scope. Here, x_i is the outlier value, μ and σ are the mean and standard deviation of the normal parameters, duration is the duration, and scope is the scope of influence. All hazard sources are ranked by intensity. The highest intensity is defined as the primary hazard source, typically numbering 1-2. Secondary hazard sources are defined as outliers with lower intensity but still exceeding the threshold, numbering 3-10. The spatiotemporal distribution of hazard sources is analyzed; primary hazard sources are often located at key nodes in the physiological system. Detailed characteristics of each hazard source are recorded: location coordinates, intensity value, type label, timestamp, and associated parameters. Topological relationships between hazard sources are established to identify causal links and collaborative patterns.

[0068] The impact of primary hazard sources on secondary hazard sources is assessed to form a hazard matrix. The mechanism of influence propagation from primary hazard sources to secondary hazard sources is analyzed, taking into account the network connectivity characteristics of physiological systems. The impact function I(i,j) = R_i × exp(-d_ij / λ) × cos(θ_ij) is defined, where R_i represents the intensity of primary hazard source i, d_ij represents the distance to secondary hazard source j, λ represents the impact attenuation length, and θ_ij represents the angle between the propagation direction and the dominant path. The impact attenuation length is set based on the characteristics of the physiological system: λ≈10cm for neural conduction and λ≈50cm for blood circulation. The impact of all primary-secondary hazard source pairs is calculated, and the impact matrix I is constructed. The matrix element I_ij represents the impact of the i-th primary hazard source on the j-th secondary hazard source. The row sum of the matrix reflects the total influence of the primary hazard source, and the column sum reflects the degree of influence of the secondary hazard source. Strong impact pathways (I_ij > 0.7) are identified, as these pathways may form a cascading transmission of the hazard. The time delay of the impact is analyzed, with delays in neural pathways measured in milliseconds and in bodily fluid pathways measured in seconds to minutes. Construct an augmented hazard matrix, add the mutual influence between sub-hazard sources, and form a complete hazard network.

[0069] The dominant hazard vector is generated by eigenvalue decomposition of the hazard matrix. The eigenvalue decomposition of the hazard matrix I is performed as I=VΛV^(-1), where Λ is the eigenvalue diagonal matrix and V is the eigenvector matrix. The eigenvalue λ_k reflects the hazard intensity of the corresponding mode, and the largest eigenvalue corresponds to the most dangerous propagation mode. The distribution spectrum of the eigenvalue is calculated. A steep spectrum indicates a high concentration of hazard, while a flat spectrum indicates a dispersed hazard. The eigenvectors corresponding to the first k largest eigenvalues ​​are selected as the dominant hazard vectors, where k is determined by the cumulative contribution rate. The first dominant vector v_1 describes the most important hazard propagation mode, and the absolute value of the vector elements indicates the degree of participation of each node. The positive and negative distribution of the dominant vectors is analyzed. Nodes with the same sign change synchronously, while nodes with different signs change in opposition. The orthogonality between the dominant vectors is calculated to verify that they represent independent hazard modes. The hazard matrix is ​​reconstructed using the dominant vectors, retaining the main hazard propagation structure and filtering out random noise. The dominant hazard vector provides core directional information for the construction of the hazard convergence field.

[0070] Based on the dominant hazard vector, the field construction is implemented to generate the hazard convergence field. The hazard convergence field is used to describe the distribution law of the physiological risks of workers in time and space. The dominant hazard vector is used as the basis function to construct a continuous hazard field distribution. The hazard convergence field is defined as H(x,y,t)=Σ_kα_k(t)φ_k(x,y)v_k, where H(x,y,t) represents the hazard intensity value at the physiological state space coordinate (x,y) at time t, α_k(t) is the activation coefficient of the kth hazard mode at time t, φ_k(x,y) is the spatial influence function of the kth mode at the state position (x,y), and v_k is the characteristic weight of the kth dominant hazard vector. The activation coefficient is obtained by projecting the current hazard state onto the dominant vector, reflecting the activity level of each hazard mode. The spatial influence function uses a radial basis function, which is attenuated with the hazard source as the center to simulate the propagation characteristics of the hazard impact. By calculating the gradient of the field Identify the direction of fastest-growing danger, with contour lines depicting the spatial distribution of similar danger levels. Key features for identifying dangerous convergence fields include the convergence center, or the point of maximum field strength, which indicates the physiological state of workers most prone to dangerous behaviors; the watershed, or the boundary with zero gradient, which separates different areas of danger transmission; and the dangerous channel, or the high-gradient corridor, which indicates the path of rapid deterioration of dangerous conditions. Establish a field strength grading standard, classifying the physiological state of workers into three levels: safe, caution, and dangerous. Areas exceeding the threshold require immediate safety measures.

[0071] Step S160 : determining an accurate identification window based on the dangerous convergence field, and converting the graded warning strategy into a final dangerous behavior category according to the accurate identification window.

[0072] Specifically, a precise identification window is determined based on the hazard convergence field. Local maximum points of the hazard convergence field are calculated, corresponding to the centers of dangerous energy. A gradient ascent algorithm is used, starting from multiple initial points, to trace the paths to the maximum points and determine the attraction domain of each maximum point. For each maximum point, the hazard intensity integral I_max = ∫∫_DH(x,y,t)dxdy is calculated, where D is the attraction domain of the point. The integral value reflects the overall hazard level of the area. A hazard intensity threshold I_th is set, and regions with integral values ​​exceeding the threshold are selected as candidate identification windows. The temporal evolution characteristics of the candidate windows are analyzed, and the moving velocity of the window center v_c = |dr_c / dt| is calculated, where r_c is the center position vector. The moving velocity reflects the dynamic characteristics of hazard propagation: a rapidly moving window indicates the spread of hazard, while a slowly moving window indicates a stable accumulation of hazard. The spatiotemporal boundaries of each candidate window are determined. The spatial boundary is defined by the field strength contour H = 0.5H_max, and the temporal boundary is defined by the hazard duration. The window geometry is optimized, using a minimum bounding rectangle or ellipse to fit irregular boundaries, facilitating subsequent feature extraction. Adjacent windows with spatial overlap exceeding 50% are merged to avoid duplicate identification of the same hazardous area. The final set of precise identification windows, W = {W_i(x_i, y_i, t_i, Δx_i, Δy_i, Δt_i)}, is determined. Each window contains information about its center position, spatial dimensions, and time span. The precise identification window discretizes the continuous hazardous field into actionable analysis units, providing spatiotemporal localization for the precise identification of hazardous behaviors.

[0073] The hierarchical warning strategy is converted into the final dangerous behavior category based on the precise identification window. In each precise identification window, a multidimensional feature vector is extracted, including the statistical characteristics, dynamic characteristics and topological characteristics of the dangerous field. The statistical characteristics include the average field strength, maximum field strength, field strength variance and skewness within the window, reflecting the intensity and distribution characteristics of the danger. The dynamic characteristics are obtained by calculating the time derivative of the field strength. Obtain information including growth rate, oscillation frequency, and phase. Topological features are analyzed for the distribution of critical points within the window, counting the number and configuration of maxima, minima, and saddle points. The extracted feature vectors are matched with the hierarchical warning strategy. A mapping function f:R^n→{green, yellow, red} is established from feature space to warning level, and classification is performed using a support vector machine or decision tree. The corresponding warning level is calculated for each window to form a preliminary risk assessment. Warning information from multiple windows is integrated; the risk level increases significantly when multiple red warning windows exist. Specific risk behavior categories are defined: fatigue (continuous low field intensity yellow warning), distraction (rapid field intensity fluctuation yellow warning), impending disability (high field intensity red warning), sudden syncope risk (sharp field intensity increase red warning), and arrhythmia tendency (periodic oscillation red warning). Fuzzy logic is used to handle the uncertainty between warning levels and risk characteristics, allowing a window to belong to multiple categories simultaneously. The membership degree μ_i∈[0,1] is calculated for each risk behavior category, with the highest membership degree corresponding to the most likely risk type. A final risky behavior identification report is generated, including primary hazard categories, secondary risks, probability of occurrence, and recommended interventions. Accurate identification of risky behavior categories translates from abstract field distribution to specific behavioral risks, providing actionable decision-making information for safety management.

[0074] In order to implement a dangerous behavior identification method based on multi-source physiological signal fusion corresponding to the above method embodiment, and to achieve the corresponding functions and technical effects. Figure 2 , Figure 2 The following is a block diagram of a dangerous behavior identification device 200 based on multi-source physiological signal fusion according to an embodiment of the present application. For ease of explanation, only the parts related to this embodiment are shown. The dangerous behavior identification device 200 based on multi-source physiological signal fusion according to an embodiment of the present application includes: The signal monitoring module 201 is used to monitor the multi-source physiological signals of the operator, wherein the multi-source physiological signals include EEG signals and ECG signals, and perform phase locking on the multi-source physiological signals to generate physiological synchronization values; An anomaly tracking module 202 is configured to screen strongly coupled signal pairs based on the physiological synchronization value, extract abnormal delay propagation paths from the strongly coupled signal pairs, detect cascade mutation points along the abnormal delay propagation paths to form dangerous trigger nodes, and perform energy density analysis on the dangerous trigger nodes to identify high-risk cluster areas; a gradient modulation module 203 for extracting an abnormal gradient field from the high-risk cluster area, performing rotational modulation on the abnormal gradient field using the EEG signal to generate a dynamic gradient field, and determining a dangerous behavior boundary based on the dynamic gradient field; a resonance detection module 204 configured to determine a dynamic detection accuracy based on the dangerous behavior boundary and the physiological synchronization value, resonately match the electrocardiogram signal with the dynamic detection accuracy to generate resonance detection parameters, generate an adaptive monitoring sequence based on the resonance detection parameters, perform a sparse process on the adaptive monitoring sequence to extract key detection windows, and construct a graded warning strategy based on the key detection windows; a danger convergence module 205 for performing rhythm analysis on the ECG signal to detect rhythm variability, dynamically adjusting acquisition density according to the rhythm variability to generate an adaptive acquisition scheme, determining a detection period based on the adaptive acquisition scheme and the graded warning strategy, and performing intermittent feature extraction according to the detection period to construct a danger convergence field; The behavior recognition module 206 is configured to determine an accurate recognition window based on the dangerous convergence field, and convert the graded warning strategy into a final dangerous behavior category according to the accurate recognition window.

[0075] The aforementioned dangerous behavior identification device 200 using multi-source physiological signal fusion can implement the dangerous behavior identification method using multi-source physiological signal fusion described in the aforementioned method embodiment. The optional options in the aforementioned method embodiment also apply to this embodiment and will not be described in detail here. The remaining contents of the embodiments of this application can be referred to the contents of the aforementioned method embodiment and will not be further described in this embodiment.

[0076] The above description is only a partial or preferred embodiment of the present application. Neither the text nor the drawings can limit the scope of protection of the present application. Any equivalent structural transformation made by using the contents of the present application specification and drawings under the overall concept of the present application, or direct / indirect application in other related technical fields, is included in the scope of protection of the present application.

Claims

1. A dangerous behavior identification method based on multi-source physiological signal fusion, characterized in that: include: Monitoring multi-source physiological signals of an operator, the multi-source physiological signals including an electroencephalogram (EEG) signal and an electrocardiogram (ECG) signal, and performing phase locking on the multi-source physiological signals to generate a physiological synchronization value; screening strongly coupled signal pairs based on the physiological synchronization value, extracting abnormal delay propagation paths from the strongly coupled signal pairs, detecting cascade mutation points along the abnormal delay propagation paths to form dangerous trigger nodes, and performing energy density analysis on the dangerous trigger nodes to identify high-risk cluster areas; extracting an abnormal gradient field from the high-risk cluster area, performing rotational modulation on the abnormal gradient field using the EEG signal to generate a dynamic gradient field, and determining a dangerous behavior boundary based on the dynamic gradient field; determining a dynamic detection accuracy based on the dangerous behavior boundary and the physiological synchronization value, performing resonance matching between the electrocardiogram signal and the dynamic detection accuracy to generate resonance detection parameters, generating an adaptive monitoring sequence based on the resonance detection parameters, performing sparse processing on the adaptive monitoring sequence to extract key detection windows, and constructing a graded early warning strategy based on the key detection windows; Performing rhythm analysis on the electrocardiogram signal to detect rhythm variability, dynamically adjusting acquisition density according to the rhythm variability to generate an adaptive acquisition plan, determining a detection period based on the adaptive acquisition plan and the graded warning strategy, and performing intermittent feature extraction according to the detection period to construct a danger convergence field; An accurate identification window is determined based on the dangerous convergence field, and the graded warning strategy is converted into a final dangerous behavior category according to the accurate identification window.

2. The method according to claim 1, characterized in that The phase-locking the multi-source physiological signals to generate a physiological synchronization value includes: generating a cognitive rhythm band through the EEG signal; forming a signal-rhythm correlation characteristic based on the electrocardiogram signal and the cognitive rhythm band; extracting a synchronization control interval within the signal-rhythm correlation characteristic; A physiological synchronization value is formed based on the locking strength of the synchronization control interval.

3. The method according to claim 1, characterized in that The extracting the abnormally delayed propagation path from the strongly coupled signal pair comprises: Splitting the strongly coupled signal pair into a leading signal and a following signal; Transmitting a timing scan path from the leading signal to the following signal; Collecting the positions of abnormal response points on the sequential scanning path to form a response point set; The trajectory with the largest delay in the response point set is selected as the abnormal delay propagation path.

4. The method according to claim 1, wherein The step of rotationally modulating the abnormal gradient field by using the EEG signal to generate a dynamic gradient field includes: Performing cognitive intensity recognition on the EEG signal to generate an intensity distribution area; A rotation potential is evaluated based on the intensity distribution area to form a rotation factor; Performing angle interpolation rotation on the abnormal gradient field using the rotation factor to generate a continuous rotation distribution; A gradient transformation is performed based on the continuous rotation distribution to generate a dynamic gradient field.

5. The method according to claim 1, wherein The resonant matching of the electrocardiogram signal with the dynamic detection accuracy to generate a resonance detection parameter includes: Performing rhythm decomposition and evaluation on the electrocardiogram signal to generate a rhythm distribution map; Establishing a resonant frequency scheme based on the rhythm distribution diagram to form a frequency control table; Performing frequency alignment on the frequency control table and the dynamic detection accuracy to obtain an adjustment parameter group; Resonance matching is performed according to the adjustment parameter group to generate resonance detection parameters.

6. The method according to claim 1, characterized in that The dynamically adjusting the acquisition density according to the rhythm variability to generate an adaptive acquisition plan includes: Converting the rhythm variability into a variability vector distribution; Finding a stable center point from the distribution of the mutation vector; Using the stable center point as a seed, density diffusion is performed to form a preliminary collection area; Boundary optimization is performed on the preliminary acquisition area to form an adaptive acquisition solution.

7. The method according to claim 1, characterized in that The intermittent feature extraction is performed according to the detection period to construct a dangerous convergence field, including: Positioning and identifying the hazard sources during the detection period to generate primary hazard sources and secondary hazard sources; Using the primary hazard source to evaluate the impact of the secondary hazard source to form a hazard matrix; Generate a dominant hazard vector by performing eigenvalue decomposition on the hazard matrix; A field construction is implemented based on the dominant hazard vector to generate a hazard convergence field.

8. The method according to claim 3, characterized in that The collecting the positions of abnormal response points on the sequential scanning path to form a response point set includes: Analyzing response transition key points in the timing scan path; determining an exploitable natural anomaly direction at the response transition key point; evaluating the detection reliability of the natural anomaly direction and generating a reliability evaluation result; A response point set is generated according to the natural anomaly direction and the reliability evaluation result.

9. The method according to claim 5, characterized in that The step of performing rhythm decomposition and evaluation on the electrocardiogram signal to generate a rhythm distribution map includes: Performing cross-cycle rhythm tracking on the electrocardiogram signal to obtain rhythm evolution data; Analyzing the contribution effect of each rhythm element based on the rhythm evolution data to construct a rhythm weight matrix, wherein each rhythm element includes basal heart rate, respiratory variability, and stress variability; Performing spatiotemporal distribution mapping on the rhythm weight matrix to generate a rhythm distribution map.

10. A dangerous behavior recognition device based on multi-source physiological signal fusion, characterized in that: include: A signal monitoring module is used to monitor the multi-source physiological signals of the operator, the multi-source physiological signals including EEG signals and ECG signals, and perform phase locking on the multi-source physiological signals to generate physiological synchronization values; an anomaly tracking module, configured to screen strongly coupled signal pairs based on the physiological synchronization value, extract abnormal delay propagation paths from the strongly coupled signal pairs, detect cascade mutation points along the abnormal delay propagation paths to form dangerous trigger nodes, and perform energy density analysis on the dangerous trigger nodes to identify high-risk cluster areas; a gradient modulation module, configured to extract an abnormal gradient field from the high-risk cluster area, rotationally modulate the abnormal gradient field using the EEG signal to generate a dynamic gradient field, and determine a dangerous behavior boundary based on the dynamic gradient field; a resonance detection module, configured to determine a dynamic detection accuracy based on the dangerous behavior boundary and the physiological synchronization value, resonantly match the electrocardiogram signal with the dynamic detection accuracy to generate resonance detection parameters, generate an adaptive monitoring sequence based on the resonance detection parameters, perform a sparse process on the adaptive monitoring sequence to extract a key detection window, and construct a graded warning strategy based on the key detection window; a danger convergence module, configured to perform rhythm analysis on the electrocardiogram signal to detect rhythm variability, dynamically adjust acquisition density according to the rhythm variability to generate an adaptive acquisition scheme, determine a detection period based on the adaptive acquisition scheme and the graded warning strategy, and perform intermittent feature extraction according to the detection period to construct a danger convergence field; A behavior recognition module is used to determine an accurate recognition window based on the dangerous convergence field, and convert the graded warning strategy into a final dangerous behavior category according to the accurate recognition window.

Images