Abnormal separation alarm system and method for mobile positioning terminal

Abstract

The invention relates to the technical field of safety monitoring, in particular to an abnormal separation alarm system and method for a mobile positioning terminal, and the method comprises the steps: synchronously collecting multi-modal physiological and position signals under a unified time reference, and packaging the multi-modal physiological and position signals into a data set; and reconstructing a relative coordinate of the terminal according to the position signal, mapping a role identifier and each modal deviation into node features, constructing a space-time hypergraph containing a constraint hyperedge, and inputting the feature pulse into a pulse graph neural network to obtain an interactive excitation value and a consistency value. The terminal generates a risk vector and then uploads the risk vector in a homomorphic encryption manner, and the gateway solves a minimum value of a risk function in a ciphertext domain according to weight and topology, obtains a global risk index and divides risk levels; and issuing alarms to drive acousto-optic and access control according to levels, writing the alarms into a distributed account book, and remotely issuing updated weights and thresholds to realize closed-loop self-adaption. The method has the advantages of high precision, low delay and privacy security.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of safety monitoring, in particular to a mobile positioning terminal abnormal separation alarm system and method. BACKGROUND

[0002] In the personal safety supervision of special scenarios, the management party usually implements the "three-number" rule: each group of three managed personnel must act simultaneously and be within a distance of no more than 5 meters from each other; the distance limit can be temporarily canceled when the supervisor is close; the supervisor needs to have close contact with the managed personnel once a day for conversation and transformation; in addition, the position of the driver of the vehicle entering the site and the contact distance with the managed personnel also need to be monitored in real time and an alarm is triggered when there is an abnormal approach. The mobile positioning terminal can present the personnel distribution and trajectory in real time, but the existing solutions rely on single-channel RSSI, Bluetooth heartbeat or infrared access control to determine the terminal state, which has three major defects: ① The collected signal is single, which cannot identify complex abnormalities such as terminal detention or impersonation, and is easy to break the three-number constraint; ② The determination rule is fixed, which is difficult to adapt to the electromagnetic interference on site, changes in personnel density and supervisor exemption situations; ③ The data is uploaded in plaintext, the exposure risk of physiological and location information is high, and the privacy and compliance pressure is great. The above shortcomings limit the reliability and timeliness of the supervision of high-risk groups, emergency disposal and responsibility tracing. SUMMARY

[0003] In view of the many problems existing in the prior art, the present application provides a mobile positioning terminal abnormal separation alarm system and method, which collects multi-modal signals on a unified time base, reconstructs the relative coordinates of the terminals and constructs a time-space hypergraph containing constraint hyperedges; the node features are output by a pulse graph neural network to obtain interaction excitation values and consistency values, and the risk vectors are homomorphically encrypted to the gateway after being generated, and the global risk index is obtained by minimizing in the ciphertext domain; sound and light and access control are triggered according to the level, and the results are written to the chain and the weight is written back. Realize privacy protection, real-time alarm and online self-adaptation.

[0004] A mobile positioning terminal abnormal separation alarm system, comprising: a time-based acquisition module for acquiring multi-modal physiological and location signals under a unified time reference, and packaging them into an integrated data set; a hypergraph reasoning module for reconstructing the relative coordinates between terminals according to the location signals, combining role identification and each modal deviation information into node features, generating a time-space hypergraph containing constraint hyperedges, and converting the node features into pulse sequences to input a neural network model to obtain interaction excitation values and consistency values; a homomorphic aggregation module for constructing a risk vector for each node, encrypting the risk vector and transmitting it to the gateway, calculating the minimum value of the risk function in the ciphertext domain according to the preset weight and hypergraph topology, obtaining the global risk index and determining the risk level; An adaptive response module is configured to issue alarm information to drive terminal sound and light prompts and access control linkage according to the risk level, write risk indicators and risk levels into a distributed ledger, and send updated weights and thresholds to terminals in a remote upgrade manner to form a feedback loop.

[0005] Preferably, the time base acquisition module is provided with a time synchronization unit, which adds a timestamp to each acquisition event using a unified hardware time base, and packs the multi-modal physiological and position signals into a single data tensor within a preset time window.

[0006] Preferably, the hypergraph reasoning module reconstructs the relative coordinates between terminals according to the relative distance signals by a weighted least squares method, and sets a three-connected constraint hyperedge and a living body exempt hyperedge in the reconstructed coordinate system to form a space-time hypergraph.

[0007] Preferably, the hypergraph reasoning module converts node features into pulse sequences through logarithmic time quantization, and performs message passing in an isometric pulse graph neural network to output interaction excitation values and consistency values.

[0008] Preferably, when the consistency value is lower than a preset consistency threshold and the interaction excitation value is higher than a preset interaction threshold, the hypergraph reasoning module adds a fake person hyperedge to the space-time hypergraph and recalculates the interaction excitation value.

[0009] Preferably, the homomorphic aggregation module encodes the risk vector using a homomorphic encryption scheme supporting floating point operations, and constructs a risk function in the ciphertext domain according to preset weight items and hypergraph topology items.

[0010] Preferably, the homomorphic aggregation module generates a global risk indicator after minimizing the risk function, and determines a risk level according to a risk indicator grading rule.

[0011] Preferably, the adaptive response module generates a weight threshold update package according to the dual gradient calculation result after writing the risk indicator and the risk level into the distributed ledger, and uses a digital signature method to ensure the integrity of the update package.

[0012] Preferably, after receiving and verifying the update package, the positioning terminal loads the updated weights and thresholds in the next data acquisition period, recalculates the interaction excitation value, and makes the neural network model cyclically adaptive based on the updated weights and thresholds.

[0013] A mobile positioning terminal abnormal separation alarm method applied to the mobile positioning terminal abnormal separation alarm system, comprising the following sequentially executed steps: Collecting multi-modal physiological signals and position signals under a unified time reference, and packaging the synchronized data into an integrated data set; Reconstructing inter-terminal relative coordinates according to the position signal, combining the role identifier and each modal deviation information into node features, generating a space-time hypergraph containing constraint hyperedges, converting the node features into pulse sequence input into a neural network model to obtain interaction excitation values and consistency values; Constructing a risk vector containing the interaction excitation values and the consistency values and a multi-modal anomaly indication for each node, performing encrypted transmission of the risk vector to a gateway, and calculating a risk function minimum value in a ciphertext domain according to a preset weight and a hypergraph topology to obtain a global risk indicator and determine a risk level; According to the risk level, issuing alarm information to drive terminal sound and light prompts and access control linkage, writing the global risk indicator and the risk level into a distributed ledger, and sending updated weights and thresholds to the terminal through remote upgrade to realize a feedback loop.

[0014] Compared with the prior art, the advantages and beneficial effects of the present application are that: Through the technical means of "multi-modal synchronous acquisition + pulse hypergraph reasoning", real-time anomaly detection of cross-signal fusion and personnel role constraint is realized; Through the technical means of "floating point homomorphic encryption aggregation", the security calculation effect of not leaking privacy but still being able to globally determine the level is realized; Through the technical means of "blockchain storage + dual gradient remote upgrade", the closed-loop optimization of data traceability and model self-learning is realized; Through the technical means of "double-link broadcast + local firmware fault tolerance", the millisecond-level rapid response of sound and light prompts and access control linkage is realized. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is an interaction schematic diagram of the system of the present application; Figure 2 is a flowchart of the method of the present application. DETAILED DESCRIPTION

[0016] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments can be practiced without these specific details.

[0017] As Figure 1 shown, a mobile positioning terminal anomaly separation alarm system includes: A time-based acquisition module for acquiring multi-modal physiological and position signals under a unified time reference and packaging into an integrated data set; The time base acquisition module is the signal entrance and global synchronization core of the mobile positioning terminal abnormal separation alarm system. Its task is to convert the multi-modal original electrical signals generated by each wearing terminal in a special scene into a discrete event stream arranged according to a unified time reference, and encapsulate it into a structured data set within a fixed time window for direct retrieval by the subsequent hypergraph reasoning module. The module is composed of a high-stability clock unit, a synchronization coordination unit, an event sampling unit, and a window buffering unit. The high-stability clock unit provides a unique hardware time base, the synchronization coordination unit is responsible for error suppression, the event sampling unit completes multi-channel quantization, and the window buffering unit is responsible for zero-copy packaging.

[0018] The high-stability clock unit uses a temperature-compensated crystal oscillator with a nominal frequency of 10 MHz. It generates a terminal time sequence by step alignment with an external satellite time scale through a local loop phase-locked circuit. To ensure consistent time across the network, the synchronization coordination unit uses a timestamp exchange and residual compensation mechanism. The synchronization frame contains a reference count value. After receiving it, the terminal calculates the residual error:

[0019] wherein represents the current count value of the terminal, represents the reference count value. The synchronization coordination unit iteratively converges within a single cycle , and the residual error is compressed to seconds, making the network-wide time error not exceed 1 millisecond. The result of the cooperation between the high-stability clock unit and the synchronization coordination unit is that a globally consistent timestamp can be obtained at any terminal sampling instant, which is the basis for subsequent multi-node time-space alignment.

[0020] The event sampling unit configures independent sampling periods for the five types of channels: human body capacitance-coupled signals at 1 kHz for monitoring wearing integrity; ultra-wideband ranging signals at 200 Hz for calculating relative distances between nodes; fiber Bragg grating strain signals at 250 Hz for detecting wristband material stretching deformation; millimeter wave vital sign micro-vibration signals at 20 Hz for extracting respiratory and heartbeat main peaks; and tunneling magnetometer magnetic flux density signals at 20 Hz for monitoring magnetic field disturbances. The sampling value and timestamp form an event triple <timestamp, channel index, quantized value>. To avoid traditional waveform redundancy storage, the module uses address event representation method to directly push the triple to the ring buffer, and the entire process does not go through unnecessary copying, reducing storage jitter.

[0021] The window buffering unit maintains a double-pointer structure at the hardware level: the write pointer increments in the order of event arrival, and when the pointer difference reaches the 100 millisecond window length, packaging is triggered. Packaging is divided into two steps: first, the event stream is stably sorted according to the timestamp; second, the five types of channel data are aggregated into row vectors and filled with valid bit masks using node number as the index, finally generating a structure like wherein is the number of nodes, is the number of channels multiplied by the number of sampling points. The tensor header records the window number and the start timestamp, which facilitates quick positioning by downstream modules. If a terminal is missing some channel data within a window, the mask can prompt the inference module to handle it with reduced weight, avoiding amplification of errors caused by local missing data.

[0022] Due to the complex electromagnetic environment of special population activity sites, the collection port must have an anti-jitter mechanism. The system integrates an adaptive filter at the event sampling layer to suppress first-order difference of high-frequency jitter introduced by temperature drift and mechanical resonance. After filtering, the capacitively coupled baseline shows a stable ladder, and if there is a line drop spike, the upper algorithm can quickly identify the error removal behavior. The tunneling magnetometer is also sensitive to random drift of the environmental magnetic field, and the module performs short-time averaging on the output incremental signal to remove negligible background perturbations to reduce false positives.

[0023] Example 1: After deploying 32 terminals in a 600 square meter special scene for 24 hours, the distribution of multi-channel timestamp difference is counted, with a maximum difference of 0.9 milliseconds and a median difference of 0.31 milliseconds. Compared with the software synchronization baseline scheme, the time-based collection module reduces the cross-node time drift by 92%, and reduces the variance of the interaction excitement value calculation by 45%, significantly improving the confidence of abnormal identification.

[0024] Example 2: The window length is shortened to 50 milliseconds to simulate a high-concurrency sampling scenario; the ring buffer does not overflow in 10 hours of stress testing, and the window packing completion rate is 100%. This shows that this module can support higher frequency sensing requirements under the premise that the subsequent algorithm remains unchanged, and has sufficient flexibility for extreme scene early warning.

[0025] The effect of this module is reflected in two aspects: first, it provides strictly time-sequenced node features for subsequent hypergraph reasoning, preventing incorrect edge weight evaluation due to time jitter; second, it reduces data arrangement delay through zero-copy packing, allowing the interaction excitement value and consistency value to be calculated within 10 milliseconds after the window ends, thereby ensuring that the end-to-end delay of the alarm level does not exceed 150 milliseconds. Overall, the time-based collection module lays the foundation for the system to perform unified reasoning on multi-modal high-order coupled information, directly determining the real-time and accuracy of the abnormal separation alarm system.

[0026] Preferably, the time-based collection module sets a time synchronization unit, which adds a timestamp to each collection event using a unified hardware time base, and packs the multi-modal physiological and position signals into a single data tensor within a preset time window.

[0027] The time-based acquisition module is the first signal gateway of the system, and its essential task is to accurately time, denoise, and queue the multi-modal physiological and position signals generated by each mobile terminal in a special scene under a unified time reference, and organize them into a single data tensor according to a fixed time window to provide a highly consistent input for subsequent hypergraph reasoning. To achieve this goal, the module is internally arranged with a high-stability clock unit, a synchronization coordination unit, an event sampling unit, and a window buffering unit, which are tightly coupled in a pipeline through hardware triggers and interrupt loops to minimize signal timing drift and data transfer delay.

[0028] The high-stability clock unit uses a temperature-compensated crystal oscillator supplemented by a local phase-locked loop circuit. The crystal oscillator has a nominal frequency of 10 MHz and can maintain a daily drift of less than 1 ppm within a temperature range of -20°C to 60°C. This unit first shapes the external timing pulse into an on-chip reference signal, then uses a phase comparator to generate a correction amount to drive a voltage-controlled crystal oscillator to fine-tune the output frequency, achieving microsecond-level alignment. The synchronization coordination unit is responsible for propagating the reference count value between terminals, calculating the residual through timestamp exchange, and closed-loop compensation. The residual is defined as: wherein represents the current hardware count value of the terminal, represents the reference count value. After one synchronization period iteration, the residual is compressed to seconds, which can ensure that the time error between any two terminals within a window length of 100 milliseconds is not more than 1 millisecond. This index is particularly critical in the present invention because the subsequent pulse coding model is extremely sensitive to time drift. If the timing error exceeds 2% of the window width, the upper graph reasoning algorithm will exhibit false alarm amplification.

[0029] The event sampling unit configures independent sampling timers for the five sensing channels: human body capacitance coupling 1 kHz, ultra-wideband ranging 200 Hz, fiber Bragg grating strain 250 Hz, millimeter wave vital sign micro-vibration 20 Hz, and tunneling magnetometer magnetic flux density 20 Hz. Whenever the timer overflows, the analog-to-digital converter captures the instantaneous voltage and generates a triple 〈timestamp, channel index, quantized value〉. To reduce CPU load, the module uses an address event representation mechanism to directly write the triple into a shared ring buffer, which avoids the cache jitter caused by the traditional packet-aggregation process.

[0030] The window buffering unit operates in a write-pointer-driven manner: when the difference between the write pointer and the read pointer reaches the preset window length, a hardware trigger triggers a packaging interrupt. The packaging process first globally sorts according to the timestamp, then aggregates all channel sampling points within the same window according to the terminal number to form a row vector. The final output tensor has dimensions wherein is the number of terminals, ​The number of channels is multiplied by the number of sampling points. The tensor header records the window number and the starting timestamp, and the tail records the valid bit mask. The mask is used to prompt the subsequent algorithm whether to weight the missing values. This design takes into account data continuity and distortion tolerance, ensuring that even if individual terminals are temporarily offline in extreme scenarios, the data of the remaining terminals can still be aligned within the same tensor.

[0031] In order to reduce the false alarm rate of abnormal triggering, a short-time smoothing filter is built into the time-based acquisition module. Taking the human body capacitance coupling signal as an example, the module calculates the sliding mean of the sampling sequence in real time and detects the mutation. If the mutation amplitude exceeds the set threshold and the duration is less than 5 milliseconds, it is determined as a disturbance spike and discarded automatically. Similar mechanisms are also applied to the magnetic flux density increment signal to filter the daily environmental magnetic field jitter. This avoids misjudging occasional electromagnetic interference as wristband removal or shielding operation, providing stable input for the subsequent risk vector construction link.

[0032] Example 3: 24 positioning terminals are arranged in a 500 square meter special scene, and continuously run for 48 hours. The maximum observed deviation of the five-channel cross-correlation time difference is 0.8 milliseconds, which is reduced by more than 90% compared with the software synchronization baseline. Through the analysis of the node feature reconstruction error in the same window, the error variance is reduced by 37%, directly improving the subsequent interaction excitement value threshold discrimination, and the false alarm rate is reduced from 3.2% to 1.8%.

[0033] Example 4: The window length is reduced by half to 50 milliseconds and the sampling rate of all channels is increased by 50%. The module has no overflow in the ring buffer during 12 hours of stress testing, and the window packing completion rate maintains 100%. This shows that the module is scalable and can support more intensive data flow without modifying the upper interface.

[0034] The core principle of the time-based acquisition module is "hardware time base unification - residual compensation - event encoding - window alignment". Through this chain design, "time" and "event" are directly bound, rather than storing waveforms first and then doing offline synchronization, so that interference can be shielded in time at the acquisition end and the convolution window can be ensured to be complete at the inference end. Effectively, the module significantly improves the node feature alignment accuracy, reduces the noise injection of time drift on the pulse neural network, and controls the window encapsulation delay within 5 milliseconds, thereby ensuring that the end-to-end delay of the system alarm does not exceed 150 milliseconds.

[0035] The hypergraph inference module is used to reconstruct the relative coordinates between terminals according to the position signal, combine the role identification and the modal deviation information into node features, generate a space-time hypergraph containing constraint hyperedges, and convert the node features into pulse sequences to input into a neural network model to obtain an interaction excitement value and a consistency value. The hypergraph reasoning module is located in the middle of the computing link of the mobile positioning terminal abnormal separation alarm system, and is specially responsible for converting the unified time window data set output by the time-based collection module into a calculable graph structure, and performing event-driven reasoning on the graph structure to give real-time interaction excitement value and consistency value, and providing numerical basis for the node risk vector generated by the homomorphic aggregation module. The module is composed of four stages of pipeline: coordinate reconstruction, node feature construction, hypergraph generation, and pulse reasoning.

[0036] Coordinate reconstruction. In special scenarios, multiple positioning terminals obtain a distance matrix through ultra-wideband transceiving ranging pulses. The hypergraph reasoning module uses a weighted least squares method to generate a residual term for each distance measurement, and maps the ranging signal-to-noise ratio as a weight to solve the relative coordinates. The core objective function is:

[0037] wherein represents the two-dimensional coordinate vector of terminal , and represents the ranging value of terminal to terminal , and is the confidence coefficient of distance measurement. The coordinate reconstruction does not depend on fixed anchors, and can maintain stable layout in signal shielding or structural shielding environment, while the translational and rotational invariance ensures that the downstream reasoning results do not fluctuate with the overall drift of the scene.

[0038] Node feature construction. Each terminal outputs five-dimensional features in a single time window: role identification, human body capacitance coupling deviation, magnetic flux density deviation, vital sign main peak deviation, and optical fiber strain variance. The role identification discretely encodes special groups and management personnel; the four deviations respectively reflect the wearing integrity, magnetic shielding attempt, living body, and wristband mechanical damage risk. The module performs window linear normalization on the five-dimensional vector, so that the feature distribution of each dimension falls within the interval, avoiding over-saturation or over-sparse problems in pulse coding stage.

[0039] Hypergraph generation. The module imports the reconstructed coordinates and node features into the hyperedge generator. According to the scene specification, the three-number constraint hyperedge forms a high-order relationship among the three special personnel nodes in the same group, which is used to monitor the distance within the group in real time; the living exempt hyperedge temporarily shields the three-number constraint when the management personnel is close; the dummy hyperedge is responsible for amplifying the possible disguise risk online. The hyperedge is represented as a multi-dimensional tensor in data structure, and the high-order connection form allows single message passing to span multiple nodes, reducing repeated traversal.

[0040] Pulse inference, node features are mapped to pulse sequences by logarithmic time quantization, and the pulse interval is inversely proportional to the absolute value of the feature, so it can maintain numerical resolution on low-power hardware. The pulse sequence is input into a multi-layer isometric pulse graph neural network to realize the synchronous propagation of messages in the structure and time domain. The network outputs the interaction excitation value and the consistency value: the former represents the degree of disturbance of the node to the topological stability, and the latter represents the consistency of the physiological signal and the wearing state of the node. When the consistency value is lower than the consistency threshold and the interaction excitation value is higher than the interaction threshold, the module immediately inserts a dummy super-edge and quickly performs two-layer re-inference to generate a new interaction excitation value. The dynamic edge insertion process is first proposed in this invention and can be regarded as an online amplifier, which can significantly increase the risk of potential camouflage events within a single window without waiting for the next learning period.

[0041] Example 5: 36 terminals are deployed in a special scene with an area of 500 square meters, and 100 wristband replacement behaviors are artificially manufactured for 24 hours of continuous operation. The hypergraph inference module inserts a dummy super-edge and outputs a high-risk label within 200 milliseconds in 98 times, and the false detection rate is reduced to 1.8%. Example 2: The time window is shortened to 50 milliseconds and the sampling rate of each channel is increased by 50%, and the window buffer still maintains a 100% packing success rate. The pulse inference delay is controlled within 12 milliseconds, proving that the module has high throughput expansion capability.

[0042] Actual deployment shows that when the three-in-a-row constraint is broken but the living body exempt super-edge exists, the interaction excitation value remains low, and the consistency value remains high, the system will not misreport; when the magnetic flux density deviation and the capacitance deviation are abnormal at the same time, even if the distance meets the constraint, the interaction excitation value still rises rapidly, triggering the subsequent aggregation module for high-weight processing.

[0043] Preferably, the hypergraph inference module reconstructs the relative coordinates between terminals according to the relative distance signal by the weighted least squares method, and sets the three-in-a-row constraint super-edge and the living body exempt super-edge in the reconstructed coordinate system to form a spatio-temporal hypergraph.

[0044] The hypergraph inference module is the "nerve center" of the mobile positioning terminal abnormal separation alarm system, responsible for converting the unified time window tensor output by the time-based acquisition module into a structured spatio-temporal hypergraph, and completing real-time semantic inference on the hypergraph. Its workflow can be divided into four levels: coordinate reconstruction, node feature construction, constraint super-edge injection, and pulse inference.

[0045] Coordinate reconstruction, the system deploys several terminals in a special scene, and each pair of terminals periodically exchanges ultra-wideband ranging pulses and obtains a distance matrix. In order to get rid of the dependence on fixed anchors, the module performs weighted least squares optimization on the distance matrix to solve the relative coordinates of the terminals in the current window. The optimization objective function is denoted as:

[0046] wherein is a two-dimensional coordinate vector of the terminal ; is a ranging value of the terminal to the terminal ; is a confidence coefficient calculated according to the ranging signal-to-noise ratio. The coordinate system obtained by solving has translational and rotational invariance, and can remain stable in a metal shielding or structural shielding environment. The traditional method often uses a multi-station method, which needs to deploy fixed reference points in advance. The present application reduces the deployment cost through pure mobile point composition.

[0047] Node feature construction, the module extracts a five-dimensional feature vector for each terminal in the same time window: role identification - distinguishes special groups and supervisors, encoded as 0 and 1; coupling capacitance deviation - reflects the integrity of the wristband fitting; magnetic flux density deviation - monitors magnetic shielding or disassembly attempts; vital sign main peak deviation - takes the relative change of the millimeter wave radar breathing peak and heartbeat peak; fiber strain variance - measures the stress state of the wristband material. All dimensions are linearly normalized to the interval in this window.

[0048] Constraint super-edge injection, in the reconstructed coordinate system, the system needs to pay attention to both spatial topology and supervision rules. The module injects three types of super-edges according to the rules: A, three consecutive number constraint super-edge, connects three special group nodes in the same group as a three-order relationship, used to detect group dissolution; B, living body exemption super-edge, created when the distance between a supervisor and any special group node is less than a set threshold, used to temporarily weaken the three consecutive number restriction; C, fake person super-edge, dynamically added when the consistency value is lower than a threshold and the interaction excitement value is higher than a threshold, used to amplify the possible disguise risk. High-order super-edges can affect multiple nodes through a single message passing, which can significantly reduce the computational overhead compared to judging the distance threshold one by one.

[0049] Pulse reasoning, after the node feature vector is quantized by logarithmic time, it is mapped to a pulse sequence, the greater the absolute value of the feature, the shorter the corresponding pulse interval. The module uses a multi-layer isometric pulse graph neural network to perform event-driven message passing on a spatio-temporal hypergraph, outputting an interaction excitement value and a consistency value. The interaction excitement value measures the degree of influence of the node on the topology stability; the consistency value measures the degree of consistency of the physiological signal and the wearing state. When the consistency value is lower than a preset consistency threshold and the interaction excitement value is higher than a preset interaction threshold, the module immediately inserts a fake person super-edge and performs two-layer re-reasoning to strengthen anomaly detection. The dynamic edge insertion mechanism is first used in the present application, which can significantly amplify the disguise risk in a single window.

[0050] In the special scene of 500 square meters, 36 terminals are deployed, and 100 wristband replacement behaviors are artificially manufactured for 24 hours of continuous collection. The hypergraph reasoning module inserts the false person hyperedge within 200 milliseconds in 98 events and outputs a high-risk label, with a false negative rate of 1.8%, which is reduced by 70% compared with the baseline without the false person hyperedge.

[0051] In example 7, the time window is shortened to 50 milliseconds, and the five-channel sampling rate is increased by 50%. The window packing completion rate still remains 100%, the pulse reasoning delay is controlled within 12 milliseconds, and the proof module has expansion capability for large data throughput.

[0052] Through the present application, pure mobile point coordinate reconstruction eliminates the deployment of anchor points and is naturally adapted to shielding environments. High-order hyperedges act on multiple nodes at a time, reducing logical judgment and traversal delay. Pulse event-driven reasoning has low power consumption and is suitable for long-term operation of power-limited terminals. Dynamic false person hyperedge provides self-enhancing feedback, amplifies the risk signal when an abnormal event occurs, and reduces false negatives. The isometric structure ensures that the reasoning result is robust to overall coordinate drift and rotation.

[0053] Preferably, the hypergraph reasoning module converts the node features into a pulse sequence through logarithmic time quantization and performs message passing in the isometric pulse graph neural network to output interaction excitation values and consistency values.

[0054] The pulse encoding-reasoning link of the hypergraph reasoning module converts multi-dimensional node features into sparse event streams and completes low-power and high-robustness semantic reasoning with the help of the isometric pulse graph neural network. This section focuses on the two core mechanisms of "logarithmic time quantization" and "pulse message passing", and explains their principles, implementation paths, and application effects in special scenarios.

[0055] The principle of logarithmic time quantization is that the node feature vector is composed of five-dimensional values of role identification, coupled capacitance deviation, magnetic flux density deviation, vital sign main peak deviation, and strain variance. In order to efficiently represent continuous amplitudes on neuromorphic hardware, the present application uses logarithmic time quantization to map amplitudes to pulse firing intervals. The quantization process consists of two steps: (1) Amplitude compression: first map each feature to (2) Logarithmic mapping: convert the normalized result to a positive pulse interval , and the mapping formula is:

[0056] wherein is the reference interval constant; is the current feature value. It can be seen that the larger the absolute value of the feature, the shorter the pulse interval, and the higher the event frequency. This formula has exponential compression characteristics and can cover a wide amplitude dynamic range within a limited time window. In the formula is the basic unit of time quantization, which does not change with dimensions.

[0057] Pulse sequence generation, after logarithmic mapping, the module creates five independent pulse sequences for each terminal's five-dimensional features. The pulses within the sequence are periodically emitted. All pulses are accompanied by the node to which they belong and the feature index, achieving a two-dimensional identification of spatial dimensions and feature dimensions. Compared with the amplitude direct writing scheme, the pulse sequence can be encoded with a Boolean event array, significantly reducing storage and bus conflicts.

[0058] Isometric pulse graph neural network, the network runs in a three-stage "structure-message-aggregation" mode: Structure preservation: Use the previous coordinate reconstruction and hyperedge injection results to form the adjacency tensor. Isometric design ensures that the message propagation order and weight reuse method remain consistent when the overall graph is translated or rotated.

[0059] Message generation: At each pulse arrival time step, the source node pulse sequence is decoded into a time increment and spliced with the edge weight to form a message vector.

[0060] Aggregation update: The target node accumulates messages from the neighborhood in a gated weighted sum manner, and outputs the next layer of pulses after triggering the threshold discharge.

[0061] Since all calculations are triggered only when events arrive, the silent period does not require computing resources, and the power consumption is linearly related to the pulse density. Finally, the network outputs two types of results for each node: interaction excitation value and consistency value. The interaction excitation value is obtained by normalizing the cumulative count of neighborhood pulses, and a large value indicates that the node has strong abnormal influence in the current topology. The consistency value is generated by comparing the phase difference between the coupled capacitor pulse and the vital sign pulse, indicating the matching degree of the wearing state and the living body.

[0062] Dynamic dummy hyperedge trigger logic, if the consistency value is lower than the consistency threshold and the interaction excitation value is higher than the interaction threshold, the module immediately inserts a dummy hyperedge for the node in the hypergraph, and the corresponding hyperedge weight is higher than the regular edge. The edge insertion operation will modify the adjacency tensor at the next time step, so that the node will get a larger negative contribution in subsequent message passing, thereby amplifying the risk signal and achieving "fast gain - fast decay" online regulation.

[0063] Embodiment 8, 40 terminals are arranged in a special scene with an area of 500 square meters. The time window is set to 100 milliseconds, The pulse graph neural network adopts a 4-layer structure with a setting of 1 millisecond. After continuous collection for 24 hours, statistics show that the average pulse density of a single node is about 80 Hz, and the power consumption is less than one third of that of a traditional floating-point graph network; 120 artificial wearing replacement events are manufactured, and the system outputs a red risk level within 180 milliseconds in 118 events, with a missed detection rate of 1.7%; under the same data, the pulse network is replaced by a dense floating-point network, and the end-to-end delay rises to 400 milliseconds, and the missed detection rate rises to 6.5%, verifying the dual advantages of the pulse-isovar design in real-time and accuracy.

[0064] According to the application, the logarithmic time quantization completes the lossless mapping of continuous amplitude to sparse pulses, and ensures that the characteristics of a large dynamic range can be represented on a digital neuromorphic hardware without scaling; the isovar pulse graph neural network keeps the output consistent under any rotation or static drift, and is suitable for a complex barrier wall scene; the dynamic dummy superedge mechanism strengthens the risk weight of a low-consistency-high-excitement node, and realizes millisecond-level abnormal amplification; event-driven calculation significantly reduces the power consumption, so that the terminal meets the long-time endurance requirement.

[0065] Preferably, when the consistency value is lower than a preset consistency threshold and the interaction excitement value is higher than a preset interaction threshold, the supergraph reasoning module adds a dummy superedge to the spatiotemporal supergraph and recalculates the interaction excitement value.

[0066] The supergraph reasoning module introduces a “dummy superedge” dynamic amplification mechanism in the real-time alarm link of the system, which is used for quickly highlighting abnormal situations such as wearing camouflage and badge replacement. The core idea is to temporarily add a high-weight superedge to the suspected camouflage node on the spatiotemporal supergraph, so as to change the adjacency tensor and immediately recalculate the interaction excitement value, thereby amplifying a low-probability event to a high-confidence risk signal within a single time window.

[0067] The determination principle is that the application defines two types of node state indicators: (1) a consistency value, which is used to measure the matching degree of the node vital sign pulse sequence and the wearing-related signal in phase or energy distribution; and (2) an interaction excitement value, which is used to measure the information propagation activity of the node in the spatiotemporal supergraph.

[0068] Under the condition of normal wearing and no camouflage, the consistency value is high and the interaction excitement value is at a medium or low level; when the badge is taken off, the wristband is removed, or the camouflage dummy is present, the node will show a low consistency value, and the interaction excitement value will rise due to the sudden change in position or constraint relationship. The module combines the two, and determines the abnormality by using double thresholds:

[0069] wherein represents the consistency value in the current window, represents the interaction excitement value, is a consistency threshold, are interaction thresholds. Both thresholds are data-driven at model initialization and can be updated in adaptive iteration steps.

[0070] The implementation process is as follows: after the feature transmission is completed, the module calculates locally at the node and If and are met at the same time, the "fake candidate" event is triggered; a fake superedge is inserted for the candidate node in the current hypergraph, and the weight of the superedge is set to a multiple factor of the regular edge weight, which is derived from the historical false positive rate calibration; only two rounds of fast forward reasoning are performed in the affected subgraph, and the new interaction excitement value is output; and is written into the node risk vector and synchronized to the homomorphic aggregation module.

[0071] The mathematical model is that the insertion of the superedge is equivalent to adding a connection in the corresponding dimension of the adjacency tensor , and the degree tensor is updated synchronously. The single-layer message passing of the pulse graph neural network is as follows:

[0072] is given, where is the pulse stack of the node in the layer, is the weight matrix of the same layer, is the pulse activation function. When the high-weight fake superedge is added, the matrix increases in the corresponding row and column, causing the information to be concentrated in the candidate node set at the next time step, resulting in a significant increase. Due to the model's equivariant property, the edge insertion operation does not destroy the invariance to global coordinate rotation and translation.

[0073] In embodiment 9, 32 terminals are deployed in a special scene with an area of 600 square meters, 24 hours of continuous data are recorded, and 120 fake events are manually inserted, which are divided into three categories: the badge is taken off and placed in the original position and then left; the wristband is cut and left in place; the management personnel's badge is taken away and taken away.

[0074] The system settings are , , and the weight factor of the fake superedge is 2. The test results show that: the trigger mechanism takes an average of 15 milliseconds to complete the superedge insertion; the time consumed by re-reasoning after edge insertion is 6 milliseconds; the comprehensive missed detection rate is 1.5%, which is reduced by 68% compared with the non-inserted version; the average false positive rate is increased from 2.8% to 3.1%, but the total risk list is more discriminative, facilitating the adjustment of the weight by the aggregation module. This embodiment proves that the dynamic fake superedge can significantly reduce the missed detection within an acceptable false positive increase, and the end-to-end alarm delay is controlled within 200 milliseconds.

[0075] Through the application, edge insertion and re-reasoning are completed in a local subgraph, and the time consumption is in a sub-linear relationship with the node scale. Re-reasoning is triggered only when a double threshold condition is met, and the typical window trigger rate is less than 3%, which has a negligible impact on the overall power consumption. Even if the disguised node attempts to destroy or reduce part of the indicators by avoiding three consecutive numbers or magnetic shielding, as long as the two types of features appear in reverse, the double threshold condition can still be met. The threshold parameters can be automatically adjusted according to the historical false alarm rate and the missed detection rate in the adaptive iteration stage, and no repeated manual correction is required for new scene deployment.

[0076] A homomorphic aggregation module is configured to construct a risk vector for each node, encrypt the risk vector, and transmit the encrypted risk vector to a gateway, and calculate a minimum value of a risk function in a ciphertext domain according to a preset weight and a hypergraph topology to obtain a global risk indicator and determine a risk level. The homomorphic aggregation module is responsible for safely collecting local risk information distributed on each mobile terminal to the central gateway, and completing a one-time global risk level determination in the ciphertext domain. The module working link includes four stages of risk vector generation, homomorphic encoding transmission, ciphertext risk function minimum value calculation, and risk level mapping, all of which are closed in a single time window to ensure data freshness and alarm real-time performance.

[0077] In the risk vector generation stage, the terminal assembles a five-dimensional vector for the node after the pulse graph neural network reasoning is completed: interaction excitation value, coupling consistency value, coupling breakage indication, strain overrun indication, and magnetic flux anomaly indication. The first three dimensions are floating-point quantities, and the last two dimensions are Boolean quantities. In order to be compatible with homomorphic computation, the Boolean quantity is first mapped to a floating-point number of 0 and 1. After the vector is constructed, it immediately enters the homomorphic encoding stage. The module selects the CKKS encoding scheme that supports floating-point operations to pack the real value sequence into polynomial coefficients. A uniform scale is set for each window during encoding to avoid accumulation of decoding errors at the central end. After encoding is completed, the ciphertext is uploaded to the gateway through a low-power wide-area network link; the encryption process is completed in the hardware acceleration unit on the terminal side, and the one-way delay is less than 5 milliseconds.

[0078] After the central end receives all the ciphertexts, a risk function is constructed without decryption:

[0079] In the formula, x represents the ciphertext risk vector of the node; x represents the ciphertext risk vector of the node; x represents the ciphertext risk vector of the node; is a weight column vector, which is calibrated by the historical false alarm rate and the missed detection rate; is a constraint penalty coefficient; is a constraint penalty coefficient; and is the relative coordinate vector obtained in the coordinate reconstruction stage, which has been encrypted on the terminal side and transmitted. The two terms measure the local risk of the node and the punishment for violating the topology constraints, respectively. Since CKKS supports homomorphic addition and multiplication, the gateway can directly perform linear operations on vectors and coefficients in the ciphertext domain without decryption and without seeing any individual physiological indicators. The risk function can be completed in one multiplication-addition iteration in the current window, with a calculation depth of 2, which meets the current prime chain length.

[0080] After obtaining the minimum value, the gateway decrypts to obtain the global risk indicator . To facilitate deployment, the system divides the risk level into four grades: green, yellow, orange, and red, and the thresholds are automatically updated through the historical window percentile. After the level is determined, an alarm frame is broadcast, and the terminal triggers an audible and visual prompt and triggers the access control subsystem. At the same time, the gateway writes the , risk level) hash into the distributed ledger, and any node cannot tamper with it afterwards; the blockchain record also provides an auditable basis for tracing the cause of a false positive.

[0081] The key innovation of the module design is the two-level coupling of “encrypted aggregation-ciphertext optimization”. On the one hand, the risk vector itself contains complex features amplified by pulse reasoning; on the other hand, the ciphertext constraint term integrates the high-order relationship of three consecutive nodes into the same minimization, eliminating the need for post-processing rule tables and reducing the logic complexity. Since all intermediate results are always in the ciphertext state, the cloud cannot see any node's original feature values, meeting the requirement of isolating personal information in sensitive scenarios.

[0082] In embodiment 10, 30 terminals are deployed in a 500 square meter special scenario, the window length is 100 milliseconds, and the risk vector upload occupies an average of 12 Kbps of link bandwidth; the gateway takes 18 milliseconds to find the minimum value in the ciphertext domain, and the end-to-end level determination delay (from terminal upload to terminal receiving broadcast) is 82 milliseconds. Compared with the traditional plaintext aggregation scheme, the delay only increases by 11 milliseconds, but the data leakage surface is reduced to 0. After reviewing 24 hours of data, the matching rate of red alerts and real abnormal events is 98%, indicating that the double-layer risk function achieves a balance between privacy and performance without sacrificing level determination accuracy.

[0083] The module also supports weight adaptation. The ledger record generates a weight increment package after averaging by the dual gradient, which is delivered to the terminal in encrypted differential form the next morning. After loading the new weights, the terminal will regenerate the risk vector and synchronize the scale in the next window, thereby closing the “upload-aggregation-writeback” cycle, so that the threshold converges gradually with the environment and crowd behavior.

[0084] By integrating homomorphic encoding, ciphertext risk optimization, and blockchain storage into a single pipeline, the present application achieves low-latency, high-privacy, and high-traceability global risk level determination in special scenarios, providing reliable input for subsequent adaptive response modules.

[0085] Preferably, the homomorphic aggregation module encodes the risk vector using a homomorphic encryption scheme supporting floating-point operations, and constructs a risk function in the ciphertext domain according to preset weight items and hypergraph topology items.

[0086] The design goal of the homomorphic aggregation module is to complete global fusion of local risk assessment results from multiple terminals in the ciphertext domain without revealing any personal physiological or location information. The homomorphic encryption scheme supporting floating-point operations (typically implemented as CKKS encoding) is selected because the risk vector contains both discrete indicators and continuous values between 0 and 1. If integer homomorphic encoding is used, it will be difficult to balance accuracy and dynamic range. The specific process is as follows: First, each terminal generates a risk vector within the current time window, where the elements are interaction excitement value, consistency value, coupling breakage marker, strain overrun marker, and magnetic flux anomaly marker in turn. The terminal uses the hardware multiplication array in the chip to perform polynomial encoding, maps the floating-point sequence to ciphertext polynomial coefficients, and uses a unified scale for all terminals in the same window to ensure that the decryption error on the center side is controllable. Second, the ciphertext vector is uploaded to the gateway and can participate in weighted operations without decryption. In order to include both node self-risk and topology constraints, the module constructs the following weighted risk function in the ciphertext domain:

[0087] wherein is the number of terminals in the window; represents the risk vector ciphertext of terminal ; is the weight column vector, which is offline calibrated by differential evolution algorithm according to historical false positive and false negative ratios; and are the relative coordinate ciphertext obtained in the coordinate reconstruction stage; is the topology penalty coefficient, used to amplify the distance square term that violates the three-connected number constraint. The gateway only needs to perform addition and multiplication in the ciphertext domain, and does not need to perform comparison operations to complete the objective function accumulation, so it will not trigger high-complexity operations such as homomorphic square root, ensuring that the delay and power consumption are controllable.

[0088] After the homomorphic operator outputs , the gateway decrypts to obtain the global risk indicator . Since decryption only acts on a single scalar, not the original vector of each node, the data exposure surface is compressed to a minimum. Subsequently, the module maps the indicator to four risk levels of green, yellow, orange, and red according to a preset threshold and broadcasts an alarm frame. The alarm frame does not carry any personal data, but only contains the window number, risk level, and affected node bitmap. After receiving the frame, the terminal drives the audible and visual prompts and the access control to act in coordination, realizing a low-latency closed loop.

[0089] To illustrate the effect, a special scene with an area of 500 square meters is selected for comparison experiment: under the condition of 32 terminals working simultaneously and window length of 100 milliseconds, the average ciphertext upload bandwidth of the homomorphic aggregation module of the application is 12 kilobits per second, the time consumption of homomorphic summation and multiplication is 18 milliseconds, and the end-to-end judgment delay is 82 milliseconds. The control group adopts the plaintext upload plus central floating-point calculation scheme, although the calculation time consumption is reduced by only 7 milliseconds, but the complete vector of each terminal needs to be transmitted in the channel, the bandwidth rises by 5 times and the personal breathing and heartbeat information is exposed. Further analysis of the 24-hour log shows that the matching rate of red alarm and artificially marked abnormal events reaches 98%, indicating that the introduction of the topology penalty term does not weaken the judgment accuracy.

[0090] The homomorphic aggregation module also has dynamic weight self-adaptive capability. The system counts in the distributed ledger every day, , risk level, generates weight increment by using dual gradient averaging. The increment is packaged and signed and issued by the remote upgrade mechanism, and the terminal loads the new weight and synchronizes the scale in the next collection period, so that the threshold converges automatically with the scene environment and crowd behavior. This "encrypted aggregation-blockchain storage-weight rewriting" closed loop improves the adaptability of the model to environmental drift, and at the same time, with the help of the tamper-proofing feature of the ledger, it provides an audit basis for misjudgment traceability.

[0091] Preferably, the homomorphic aggregation module generates a global risk indicator after completing the minimization of the risk function, and determines the risk level according to the risk indicator grading rules.

[0092] The homomorphic aggregation module is responsible for collecting the local risk vectors output by each mobile terminal to the central gateway, and completing a global risk judgment without exposing any personal physiological or location information. Its working link is divided into four steps: risk vector generation, homomorphic encoding transmission, ciphertext risk function minimization, and risk level mapping.

[0093] The module first generates a risk vector on the terminal side. Each terminal obtains interaction excitement value, consistency value, coupling break indication, strain overrun indication, and magnetic flux anomaly indication in a single time window, a total of five-dimensional values. The first two dimensions are floating-point quantities, and the last three dimensions map Boolean quantities to 0 or 1. The vector is denoted as , where is the terminal number. In order to maintain floating-point precision in the ciphertext, the application selects the homomorphic encryption scheme CKKS that supports floating-point operations. The terminal encodes into ciphertext , and uses a uniform scale to ensure that the center decoding error is controlled. The encoding work is accelerated by the local hardware multiplication array, with a single delay of about 5 milliseconds.

[0094] After the ciphertext is chained, the gateway constructs a global risk function in the pure ciphertext domain: , where is the total number of terminals in the window; is a weight vector, calibrated jointly with historical false negatives and false positives; is a topological penalty coefficient; and are relative coordinates ciphertexts obtained in the coordinate reconstruction phase. The first term measures local risk at the node, and the second amplifies the square of the distance that violates the triplets constraint. All symbols are unique in the full text, always refer to the ciphertext object after homomorphic encryption.

[0095] CKKS supports homomorphic addition and multiplication, and the gateway can directly perform linear operations on the ciphertext without decryption. Since the risk function only contains a single vector inner product and a single distance square summation, the homomorphic circuit depth is 2, which can be completed within a prime chain length of 60 bits, with an operation delay of about 18 milliseconds. After obtaining , the gateway decrypts to obtain the global risk indicator . Decryption only acts on a single scalar, avoiding the leakage of node-level details.

[0096] The risk level mapping adopts a four-grade system: green, yellow, orange, and red. The threshold vector is automatically updated through the percentile of the past 1000 window indicators, ensuring that the stable alarm proportion is maintained when the group behavior changes over a long period of time. Typical initialization is red , orange , yellow , and green . After determining the level, the gateway generates an alarm frame containing 〈window number, risk level, node bitmap〉 and broadcasts it. The terminal triggers an audio-visual prompt and access control action according to the bitmap.

[0097] To ensure the results are auditable, the gateway hashes 〈window number, , risk level〉 and writes it into the distributed ledger. Any subsequent node cannot fake or delete the record. The ledger entry also provides gradient information for model adaptation: daily statistics of the past 24 hours indicators and levels, calculation of the average dual gradient, and updating of the weights . The update package is signed and distributed in the form of remote upgrade, and the terminal loads the new weights in the next collection period, closing the "upload-aggregation-rewrite" cycle.

[0098] Example 11: In a special scenario with an area of 500 square meters, 30 terminals are deployed, the window length is 100 milliseconds, and it runs continuously for 24 hours. The average bandwidth of ciphertext upload is 12 kilobits per second, and the end-to-end level determination delay is 82 milliseconds; compared with the plaintext aggregation control group, the delay only increases by 11 milliseconds, but the leakage risk is reduced to 0. The matching rate of red alarms and manually labeled anomalies reaches 98%, indicating that the introduction of the topological penalty term does not affect the level determination accuracy. The weight adaptation mechanism reduces the first-day missed detection rate from 3.4% to 1.9% on the second day, indicating that the feedback closed loop is effective.

[0099] An adaptive response module is configured to issue alarm information to drive terminal sound and light prompts and access control linkage according to the risk level, write risk indicators and risk levels into a distributed ledger, and send updated weights and thresholds to the terminal in a remote upgrade manner to form a feedback loop.

[0100] The adaptive response module is located at the terminal control layer of the mobile positioning terminal abnormal separation alarm system, and is responsible for converting the risk level output by the homomorphic aggregation module into executable sound and light prompts, access control actions and model self-learning updates, thereby closing the full-link feedback. The module is divided into four sub-units in chronological order: an alarm broadcast unit, a field execution unit, a ledger storage unit and a remote upgrade unit.

[0101] The alarm broadcast unit generates an alarm frame immediately after receiving the global risk indicator. The frame field contains the window number, risk level, node bitmap and cyclic redundancy check code. The risk level is represented by 2 bits in binary, and the node bitmap is converted into a byte sequence according to the number of terminals. The frame is first sent via a low-power wide-area network, and if the 3-way handshake fails, the cellular backup link is automatically switched to ensure that the key alarm has a 99% success rate in one delivery.

[0102] The field execution unit analyzes the risk level and triggers actions according to the following rules: green level is silent; yellow level turns on yellow light and writes local log; orange level turns on orange light and beeps once; red level turns on red light and continuous beeping, and writes lock command to the nearest access controller. The access controller executes the lock for 30 seconds and maintains high priority until the system is revoked. The firmware has a repeated trigger suppression logic to avoid sound and light flickering caused by broadcast jitter.

[0103] The ledger storage unit writes <window number, risk indicator, risk level> after SHA-256 hashing into the consortium chain, adopts 3-gateway endorsement consensus, and the normal transaction delay is 120 milliseconds. Only the digest is saved on the chain, and the original physiological or location information is not included, meeting the privacy compliance and audit requirements. The ledger data also provides gradient basis for weight adaptation.

[0104] The remote upgrade unit takes the past 24 hours of ledger records as input and calculates the average gradient:

[0105] wherein is the number of windows, is the decrypted risk vector of the i-th window. The new weight vector is updated by the following formula:

[0106] ​The fixed learning rate is used. After the differential update package is signed by elliptic curve digital signature, it is sent through over-the-air upgrade at 02:30-03:00 every day. After the terminal verifies the signature, it is written into the standby partition, and is loaded in the next collection cycle; if the self-check fails, the previous version is automatically rolled back.

[0107] In example 12, 32 terminals are deployed in a 600 square meter special scene, and the window length is 100 milliseconds. The red level is triggered 200 times, the average sound and light delay of the terminal is 60 milliseconds, the access control locking delay is 110 milliseconds, and the end-to-end delay is 142 milliseconds. After the system is continuously operated for 72 hours, the average change of the weight vector is 6%, the false alarm rate is reduced from 3.2% to 1.9%, and the real-time of the verification fast path and the adaptation of the slow path both meet the expectations.

[0108] Preferably, after the adaptive response module writes the risk indicators and risk levels into the distributed ledger, it generates a weight threshold update package according to the results of the dual gradient calculation, and uses digital signature to ensure the integrity of the update package.

[0109] After the adaptive response module broadcasts the alarm and writes it into the distributed ledger, it will immediately start the parameter self-learning process. This process uses the confirmed window number, global risk indicator and risk level triplets in the ledger as input, generates a weight threshold update package through the dual gradient method, and ensures the integrity of the update package through digital signature, thereby forming a model closed loop.

[0110] First, the module caches the records in the ledger in the last 24 hours locally. For the first window, the decrypted risk vector is denoted as , which includes the interaction excitement value, the consistency value and three Boolean indicators. The module calculates the time-weighted average gradient:

[0111] wherein is the number of windows, is the weight based on the risk level mapping, and the red window has the maximum value and the green window has the minimum value. The vector reflects the cumulative trend of the contribution of each dimension of risk in the past period. Then, the current weight vector is iteratively updated with a fixed learning rate and the threshold set :

[0112] In the formula, the is the sub-vector of the Boolean dimension in , which is used to synchronize the convergence threshold. represents the risk vector weighting coefficient, represents the hierarchical threshold array, is the learning rate constant. To reduce the communication load, the module calculates the difference:

[0113] and writes the binary complement form into the update package. The update package structure is 〈version number, window range, difference length, difference data, check value〉. The difference data is hashed by SHA-256, and a signature field is generated by the elliptic curve digital signature algorithm. The integrity and source can be verified on the terminal side.

[0114] The update package is delivered by using a fragmented over-the-air upgrade strategy: the gateway splits the package into several frames during the idle period of the service, and each frame contains a sequence number and a segment check. After receiving all the fragments and passing the signature verification, the terminal writes them into the standby storage area. If the write or self-check after restart fails, it will trigger automatic rollback to the previous stable version, ensuring system continuous availability.

[0115] In embodiment 13, multiple rounds of behavior pattern changes occur within a month. The initial weights of the system are uniformly set to 1. Through the dual gradient self-learning, the interaction excitement weight is increased to 3.4, the consistency weight is decreased to 2.1, and the Boolean item weight is differentiated to 4.0, 2.8, and 2.5. The measured false negative rate is reduced from 3.2% to 1.7%, and the false positive rate remains at about 2.9%, proving that the adaptive response module can automatically optimize the judgment sensitivity in real time according to the scene. The digital signature verification failure rate is 0, indicating that the package integrity is fully protected. The average size of the entire update package is 620 bytes, and the transmission time is 240 milliseconds, which has negligible impact on the daily link bandwidth.

[0116] Preferably, after receiving and verifying the update package, the positioning terminal loads the updated weights and thresholds in the next data collection period, recalculates the interaction excitement value, and makes the neural network model cyclically adaptive based on the updated weights and thresholds.

[0117] The positioning terminal undertakes the terminal landing of model parameters and real-time reasoning tasks. To avoid inconsistency caused by over-the-air upgrades, the present application designs a “three-stage hot replacement” and “single-cycle recalculation” mechanism on the terminal side. The overall process is divided into three stages: verification writing, hot replacement loading, and adaptive reasoning.

[0118] The first stage is verification writing. After receiving the update package through the low-power wide-area network, the terminal first performs two-level integrity verification: the first level calls the elliptic curve public key to verify the package signature, ensuring that the source is trustworthy; the second level checks the cyclic redundancy check code of the package body to prevent radio frequency link bit flipping. After passing the double check, the difference segment is written into the standby flash memory area, and the old version remains in the main area. After writing is completed, the upgrade version number and window number are recorded in the on-chip security fuse area for subsequent audit backtracking. It is worth noting that the write operation is completed in the closed external interrupt state to prevent clock interruption from causing the write to jump in the middle.

[0119] The second stage is hot replacement loading. To ensure timing continuity, the terminal does not immediately switch to new parameters, but triggers a soft reset process at the end of the current collection window: turn off event sampling timer and pulse stack refresh logic; copy the new weight vector and threshold array to static random access memory; move old weights and thresholds to rollback partition; modify pulse graph neural network initialization pointer to call new weight address for subsequent windows; restore event sampling timer and clear the "upgrade ready" flag.

[0120] If a parameter verification failure or memory parity error is detected within a 15-second self-check period, the loader will restore the old parameters from the rollback partition and record the error code, ensuring that the terminal can still output interaction excitement values in extreme cases.

[0121] The third stage is adaptive inference. After the terminal enters the next collection cycle, the pulse graph neural network uses the updated weight vector to recalculate the interaction excitement value:

[0122] where represents the interaction excitement value of node , is the message statistics vector accumulated by node in the current window, is the new weight vector. Since the new weight is generated under the driving of dual gradient, its numerical distribution is more consistent with the abnormal pattern in the last 24 hours, which increases the weight of high-probability risk dimensions and reduces the weight of low-contribution dimensions, making more sensitive to real abnormalities. The threshold array is updated synchronously and used for subsequent consistent double-threshold judgment and risk level mapping.

[0123] In principle, the differential update package only contains the increment of weights and thresholds, with a package body length of about 620 bytes and a transmission time of 240 milliseconds; hot replacement does not restart the operating system, only performs pointer redirection at the microcontroller layer, with a switching delay of 50 milliseconds; and the recalculation of interaction excitement values completely follows the existing message queue, with unchanged calculation load.

[0124] Embodiment 14: 32 terminals are deployed in a special scene with an area of 600 square meters, and the window length is 100 milliseconds. After starting the adaptive function, the system receives the first update package at 02:30 on the first day at midnight; each terminal completes hot replacement at 02:30:50, and enters stable operation at 03:00. Compared with before upgrading, the average value of interaction excitement value decreases by 8%, the variance of high-risk nodes increases by 7%, the accuracy of red alarm increases from 96% to 98%, and the false alarm rate decreases from 3.2% to 1.9%. The whole month statistics show that the upgrade fails to trigger rollback event twice, with a failure rate of 0.02%, meeting the high availability requirement.

[0125] As shown in Figure 2 A mobile positioning terminal abnormal separation alarm method applied to the mobile positioning terminal abnormal separation alarm system, comprising the following sequentially executed steps: Collecting multi-modal physiological signals and position signals under a unified time reference, and packaging the synchronized data into an integrated data set; Reconstructing the relative coordinates between terminals according to the position signals, combining the role identification and multi-modal deviation information into node features, generating a spatio-temporal hypergraph containing constraint hyperedges, and converting the node features into pulse sequences to input a neural network model to obtain interaction excitement values and consistency values; Constructing a risk vector for each node containing the interaction excitement value, the consistency value, and multi-modal abnormality indication, encrypting the risk vector and transmitting it to the gateway, and calculating the minimum value of the risk function in the ciphertext domain according to the preset weight and hypergraph topology to obtain a global risk indicator and determine the risk level; According to the risk level, issuing alarm information to drive terminal sound and light prompts and access control linkage, writing the global risk indicator and the risk level into a distributed ledger, and sending updated weights and thresholds to the terminal through remote upgrading to realize feedback loop.

[0126] Under a unified time reference, each terminal synchronously collects multi-modal physiological signals and position signals, and packs them into a unified data set; then reconstructs the relative coordinates of the terminals according to the position signals, and combines the role identification and multi-modal deviation into node features to construct a spatio-temporal hypergraph containing constraint hyperedges; then pulse the node features to input a neural network, and calculate the interaction excitement value and consistency value of each node; then generate a risk vector for each node containing the interaction excitement value, the consistency value, and the multi-modal abnormality indication, upload it to the gateway after homomorphic encryption, calculate the minimum value of the risk function in the ciphertext domain according to the weight and topology, obtain the global risk indicator and divide the risk level; finally, issue an alarm according to the risk level, make the terminal execute sound and light prompts and access control linkage, write the risk indicator and level into a distributed ledger, and issue new weights and thresholds to the terminal through remote upgrading, thereby forming a closed-loop adaptation.

[0127] The above merely provides an example of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the scope of claims of the present application.

Claims

1. A mobile positioning terminal abnormal separation alarm system, characterized in that, include: The time-base acquisition module is used to acquire multimodal physiological and positional signals under a unified time base and encapsulate them into an integrated dataset; The hypergraph inference module is used to reconstruct the relative coordinates between terminals based on the location signal, combine the role identifier and the modal deviation information into node features, generate a spatiotemporal hypergraph with constrained hyperedges, and convert the node features into pulse sequences to input into the neural network model to obtain interaction excitation value and consistency value. The homomorphic aggregation module is used to construct risk vectors for each node, encrypt the risk vectors and transmit them to the gateway, calculate the minimum value of the risk function in the encrypted domain based on preset weights and hypergraph topology, obtain the global risk index and determine the risk level. The adaptive response module is used to issue alarm information based on the risk level, drive the terminal's audio-visual prompts and access control linkage, write the risk indicators and risk levels into the distributed ledger, and send the updated weights and thresholds to the terminal through remote upgrades to form a feedback loop.

2. The system according to claim 1, characterized in that, The time base acquisition module sets up a time synchronization unit, uses a unified hardware time base to add a timestamp to each acquisition event, and packages multimodal physiological and positional signals into a single data tensor for output within a preset time window.

3. The system according to claim 1, characterized in that, The hypergraph reasoning module reconstructs the relative coordinates between terminals based on the relative distance signal using the weighted least squares method, and sets up triple-constraint hyperedges and liveness exemption hyperedges in the reconstructed coordinate system to form a spatiotemporal hypergraph.

4. The system according to claim 3, characterized in that, The hypergraph inference module converts node features into pulse sequences through log-time quantization and performs message passing in the equivariant pulse graph neural network to output interactive excitation and consistency values.

5. The system according to claim 4, characterized in that, When the consistency value is lower than the preset consistency threshold and the interaction excitement value is higher than the preset interaction threshold, the hypergraph inference module adds a dummy hyperedge to the spatiotemporal hypergraph and recalculates the interaction excitement value.

6. The system according to claim 1, characterized in that, The homomorphic aggregation module uses a homomorphic encryption scheme that supports floating-point operations to encode the risk vector and constructs a risk function in the ciphertext domain based on preset weight terms and hypergraph topology terms.

7. The system according to claim 6, characterized in that, After minimizing the risk function, the homomorphic aggregation module generates a global risk index and determines the risk level according to the risk index classification rules.

8. The system according to claim 1, characterized in that, After writing the risk indicators and risk levels into the distributed ledger, the adaptive response module generates a weight threshold update package based on the dual gradient calculation results and uses digital signatures to ensure the integrity of the update package.

9. The system according to claim 8, characterized in that, After receiving and verifying the update packet, the positioning terminal loads the updated weights and thresholds in the next data acquisition cycle, recalculates the interaction excitation value, and enables the neural network model to cyclically adapt based on the updated weights and thresholds.

10. A method for alarming abnormal separation of a mobile positioning terminal, applied to the mobile positioning terminal abnormal separation alarm system according to any one of claims 1 to 9, characterized in that, The steps are executed in the following order: Multimodal physiological and positional signals are collected under a unified time reference, and the synchronized data is encapsulated into an integrated dataset. The relative coordinates between terminals are reconstructed based on the location signals. The role identifier and the modal deviation information are combined into node features to generate a spatiotemporal hypergraph containing constrained hyperedges. The node features are converted into pulse sequences and input into the neural network model to obtain interaction excitation value and consistency value. For each node, a risk vector containing the interaction excitement value, the consistency value, and the multimodal anomaly indication is constructed. The risk vector is encrypted and transmitted to the gateway. The minimum value of the risk function is calculated in the encrypted domain according to the preset weight and the hypergraph topology to obtain the global risk index and determine the risk level. According to the risk level, alarm information is issued to drive the terminal's audio and visual prompts and access control linkage. The global risk indicators and the risk level are written into the distributed ledger, and the updated weights and thresholds are sent to the terminal through remote upgrades to achieve a feedback loop.

Images