Big data-based infectious disease transmission path generation and early warning method and system

Abstract

The invention relates to the technical field of public health information, in particular to an infectious disease transmission path generation and early warning method and system based on big data, and the method comprises the steps: collecting and cleaning multi-modal data, and obtaining a pulse event flow through neuromorphic coding; and constructing a variable order hyperedge graph in a second-level window, extracting loop life to adjust diffusion of the cellular automaton, and iteratively generating an infection prediction field. And then splicing the hyperedge graph, the prediction field and the topological fingerprint into a conditional tensor, and inputting the conditional tensor into a conditional score diffusion network to generate a future propagation event sequence. And constructing a tensor network according to the sequence, mapping the tensor network into Isin Hamiltonian, quickly selecting an intervention node through an optical delay ring reservoir and gradient descent, and issuing a ventilation or purification instruction. And after a virus load difference value before and after intervention is normalized, a tensor network, a diffusion network and a cellular automaton are synchronously updated to form a self-adaptive closed loop, so that a peak time error can be shortened, and the recall rate of a high-risk region is improved.

Description

Technical Field

[0001] This invention relates to the field of public health information technology, and in particular to a method and system for generating and warning about the transmission paths of infectious diseases based on big data. Background Technology

[0002] Limited air circulation and high population density in enclosed or sealed buildings allow highly virulent pathogens to complete multiple exposures within minutes, triggering cascading infections. The industry urgently needs second-level prediction and minute-level intervention to avoid large-scale shutdowns and exorbitant medical burdens. Existing technologies typically employ binary contact diagrams combined with statistical models, or simulate virus drift using fixed diffusion equations, triggering ventilation equipment through artificial thresholds. These solutions lack high-order closed-loop topology descriptions, and the diffusion parameters are static with no real-time feedback, leading to underestimation of risk during peak periods and false alarms during off-peak periods, resulting in significant resource waste. Summary of the Invention

[0003] To address the numerous problems existing in the prior art, this invention provides a method and system for generating and warning about infectious disease transmission paths based on big data. This invention constructs a hyperedge graph using pulse event streams to extract loop lifetimes, dynamically adjusts cellular automata diffusion, splices infection fields and topological fingerprint-driven conditional fractional diffusion networks to generate future events, and then uses Ising energy minimization to select intervention nodes. Feedback difference is used to synchronously fine-tune the three models, thereby reducing peak error and decreasing the number of intervention nodes.

[0004] A method for generating and issuing early warning of infectious disease transmission paths based on big data includes the following steps:

[0005] Acquire data on personnel entities, spatial locations, and environmental status, and perform missing value removal, outlier replacement, and clock synchronization at edge computing nodes to generate a standardized event stream;

[0006] The standardized event stream is input into a neuromorphic processor and converted into a pulse event stream. An event graph is constructed and persistent cohomology is calculated within a preset time window to obtain a topological fingerprint.

[0007] The pulse event stream is aggregated according to the time window to form a variable-order hyperedge event graph. The topological fingerprint is used to set a differentiable cellular automaton convolution kernel to evolve an infection prediction field in the initial state of the environment field. The variable-order hyperedge event graph, the infection prediction field and the topological fingerprint are combined into a conditional tensor and input into a conditional fractional diffusion neural network to generate a propagation event sequence.

[0008] A tensor network is constructed based on the propagation event sequence. The tensor network is represented as the Ising Hamiltonian to determine the intervention node. Intervention commands are issued to the environmental control equipment. The tensor network, the conditional fractional diffusion neural network, and the differentiable cellular automaton are updated based on the environmental feedback measured before and after the intervention to form an adaptive closed loop.

[0009] Preferably, when generating a standardized event stream, interpolation is performed to complete missing records, neighborhood averaging is performed to replace abnormal records, and timestamps of all event sensing devices are uniformly corrected through a network time synchronization protocol.

[0010] Preferably, the pulse event stream constructs an event graph within a continuous time window. The vertices of the event graph correspond to personnel entities and spatial locations, the edge weights correspond to the number of pulses within the time window, and a topological fingerprint is generated by calculating a long-lasting homology using the Vietoris–Rips method.

[0011] Preferably, a long-term homology value is used to set the diffusion radius of the differentiable cellular automaton convolution kernel, so that the convolution kernel enhances the diffusion effect when the loop structure increases, thereby generating an infection prediction field.

[0012] Preferably, the infection prediction field is generated in a differentiable cellular automaton with viral load and viral inactivation factor as state variables and is updated multiple times in discrete time steps until the state change between two adjacent updates is lower than a preset threshold.

[0013] Preferably, the variable-order hyperedge event graph, infection prediction field, and topological fingerprint are aligned according to time index, entity index, and spatial coordinates before being composed of a conditional tensor, and then concatenated into a unified tensor in the depth dimension before being input into a conditional fractional diffusion neural network.

[0014] Preferably, the conditional fractional diffusion neural network adopts a multi-scale encoding and decoding structure, injects incremental random noise in the forward diffusion, and trains the network weights through fractional matching loss in the reverse process to generate a propagation event sequence that meets the conditional tensor constraints.

[0015] Preferably, the tensor network constructed from the transmission event sequence calculates the risk weights by considering the number of people, the maximum viral load in the infection prediction field, the air exchange coefficient, and the strain coefficient. After mapping to the Ising Hamiltonian, it combines the energy gradient output by the optical delay loop reservoir system and uses alternating linear least squares and gradient descent to jointly determine the intervention nodes.

[0016] Preferably, the environmental feedback is the difference in viral load in the environment before and after the intervention. After normalization, the difference is used to synchronously update the tensor network, the conditional fractional diffusion neural network, and the differentiable cellular automaton to complete the adaptive closed loop.

[0017] A big data-based infectious disease transmission path generation and early warning system, used to implement the aforementioned big data-based infectious disease transmission path generation and early warning method, the system comprising:

[0018] The data processing module is used to acquire data on personnel entities, spatial locations, and environmental status, and to perform missing value removal, outlier replacement, and clock synchronization at the edge computing node to generate a standardized event stream.

[0019] The topological fingerprint module is used to convert the standardized event stream into a pulse event stream by the neuromorphic processor, construct an event graph and calculate persistent cohomology within a preset time window to obtain the topological fingerprint.

[0020] The propagation generation module is used to aggregate the pulse event stream according to the time window to form a variable-order hyperedge event graph, and use the topological fingerprint to set a differentiable cellular automaton convolution kernel to evolve into an infection prediction field in the initial state of the environmental field. Then, the variable-order hyperedge event graph, the infection prediction field and the topological fingerprint are combined into a conditional tensor and input into a conditional fractional diffusion neural network to generate a propagation event sequence.

[0021] The risk intervention module is used to construct a tensor network based on the propagation event sequence, represent the tensor network as an Ising Hamiltonian to determine the intervention node, issue intervention commands to the environmental control equipment, and update the tensor network, the conditional fractional diffusion neural network, and the differentiable cellular automaton based on the environmental feedback measured before and after the intervention to form an adaptive closed loop.

[0022] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0023] This invention achieves second-level capture of multi-person-multi-point cyclic propagation chains through variable-order hyperedge event graphs and a long-term homology technique; it achieves real-time adaptive diffusion radius and inactivation rate through topology-driven differentiable cellular automata; it achieves future propagation sequence generation by integrating spatial, topological, and physical conditions through conditional fractional diffusion neural networks; it achieves second-level determination of intervention nodes within budget through Ising Hamiltonian optimization and optical delay loop reservoir techniques; and it achieves joint online updating of the three-layer model through single-scalar environmental feedback closed-loop techniques, maintaining prediction accuracy and intervention efficiency over the long term. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart of the method of the present invention;

[0025] Figure 2 This is a schematic diagram of topology-driven convolution kernel adjustment in this invention;

[0026] Figure 3 This is a schematic diagram of the Ising optimization intervention node in this invention;

[0027] Figure 4 This is a structural block diagram of the system of the present invention. Detailed Implementation

[0028] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation.

[0029] like Figure 1 As shown, a method for generating and issuing early warnings of infectious disease transmission paths based on big data includes the following steps:

[0030] Acquire data on personnel entities, spatial locations, and environmental status, and perform missing value removal, outlier replacement, and clock synchronization at edge computing nodes to generate a standardized event stream;

[0031] In this invention, a standardized event stream is defined as a one-dimensional time series aligned with millisecond-level timestamps, with complete fields and no significant numerical anomalies. This event stream contains three core fields: personnel entity identifiers, spatial location coordinates, and environmental state vectors. The personnel entity identifiers are used to uniquely identify individuals; the spatial location coordinates are used to reconstruct the individual's movement trajectory within a building or transportation vehicle; and the environmental state vectors are used to measure the impact of air composition and air conditioning operation on pathogen survival and drift. The accuracy of the standardized event stream directly determines the reliability of subsequent long-term homology loop identification, differentiable cellular automaton diffusion evolution, and Ising Hamiltonian solution; therefore, data cleaning and time alignment must be completed before the data enters the neuromorphic processor.

[0032] Raw data is generated asynchronously by event-aware devices. RFID or Bluetooth beacons periodically report personnel entity identifiers, UWB base stations or WLAN probes report location coordinates every tens of milliseconds, and carbon dioxide concentration sensors and temperature and humidity sensors report environmental vectors at second-level intervals. Due to inconsistent sensor sampling frequencies, network jitter, and hardware failures, these event records commonly suffer from three types of problems: missing fields, abnormal values, and time drift. If directly used to construct an event graph, subsequent pulse counting will distort the loop lifetime distribution due to missing or spurious pulses; cellular automata will also produce unpredictable diffusion patterns due to undersampled initial environmental values.

[0033] To address the aforementioned issues, this invention implements missing value removal, outlier replacement, and clock synchronization at edge computing nodes. Located between the perception layer and the core computing layer, edge computing nodes offer the advantage of on-site preprocessing, enabling batch data correction within milliseconds and providing reliable input to upstream algorithms.

[0034] Missing value removal is handled differently for three scenarios. The first is when the person entity or spatial location field is completely missing. Such records cannot determine the subject or coordinates, are invalid for the propagation chain, and are directly deleted. The second is when only the environmental state field is missing. To avoid gaps in the air thermal field, this invention uses bidirectional temporal neighborhood interpolation: it retrieves the previous and next valid environmental records for the same location in the time series and uses their mean to fill in the missing field. If there is still no next valid record within 5 seconds, it retrieves the nearest neighboring location information with the Euclidean distance to the current location as a substitute. The third is when multiple fields are missing simultaneously but the timestamp is retained. Such records retain timeline continuity but cannot participate in graph construction and are also deleted.

[0035] Outlier replacement employs a sliding window statistical method. With a window length of 30 seconds, the window mean and standard deviation for each environmental indicator are calculated. Newly arrived records are considered outliers if they deviate from the window mean by more than a preset multiple. The replacement strategy is window mean replacement, ensuring local smoothness and avoiding the introduction of external data sources. Spatial location anomalies are identified based on displacement continuity: if the Euclidean distance between two adjacent location coordinates exceeds the upper limit of physical reach for a human body within a given time interval, the current location is corrected to the previous coordinate. This method can eliminate the "jump point" phenomenon caused by short-term bounces from ultra-wideband base stations.

[0036] Time drift is handled through edge-side clock synchronization. Each event-aware device exchanges timestamps with the edge reference clock at a frequency of 60 seconds. Edge computing nodes estimate the clock drift Δt based on the round-trip delay and update the local timestamp t. raw Apply the correction formula:

[0037] t sync =t raw -Δt

[0038] In the formula t sync For the synchronized timestamp, t raw Here is the original timestamp, and Δt is the drift amount after delay compensation. To avoid accumulated errors, resynchronization is triggered when Δt exceeds a threshold.

[0039] After three processing steps, the event stream output by the edge nodes meets the following attributes: complete fields, smooth values, and consistent timestamps. This event stream is directly encoded into a pulse event stream by the neuromorphic processor. Due to the strict alignment of the timeline, pulse encoding can accumulate the number of pulses in a fixed window, eliminating the need for compensation for cross-device time misalignment; and since numerical anomalies have been eliminated, the pulse threshold does not need to be dynamically adjusted.

[0040] In Example 1, the system collects approximately 10,000,000 raw event records daily. Records with a missing rate of approximately 1% are completed with full fields after interpolation; environmental indicators with an anomaly rate of approximately 0.5% have smoothed curves after sliding window replacement; and the maximum time difference across all records is less than 0.1 seconds after time calibration. The generated pulse event stream exhibits stable pulse counts within a 1-second time window, and the variance of the lifetime distribution of a long-lasting homogeneous loop calculated using the Vietoris–Rips method is reduced by 30%. Differentiable cellular automata show a 25% reduction in error during infection prediction field evolution. The number of iterations for tensor network risk weights converges is reduced from 40 to 32. The computation time for intervention nodes is shortened by 15%, and the slope of the environmental viral load decrease within 60 seconds after intervention is increased by 40%.

[0041] In Example 2, coordinate jumps occurred in some location records. By checking the displacement continuity, the abnormal coordinates were corrected to the coordinates of the previous moment, maintaining trajectory continuity. After the cleaned, standardized event stream was input into the core algorithm, the overlap between the predicted path and the actual case contact tracing path increased from 0.66 to 0.82, and the average early warning time was shortened by 90 seconds.

[0042] Preferably, when generating a standardized event stream, interpolation is performed to complete missing records, neighborhood averaging is performed to replace abnormal records, and timestamps of all event sensing devices are uniformly corrected through a network time synchronization protocol.

[0043] The standardized event stream serves as the common data baseline for the entire process of generating and warning about infectious disease transmission paths in this invention. Three types of original records—personnel entity identifiers, spatial location coordinates, and environmental state vectors—are asynchronously reported by different sensing devices. Direct parallel transmission would lead to temporal misalignment, field gaps, and extreme noise, introducing spurious or lost pulses in the subsequent pulse coding stage, ultimately compromising the convergence of long-term coherence loop statistics and Ising-Hamiltonian optimization. This step, at the edge computing node, uses three operations—interpolation completion, neighborhood averaging replacement, and unified time synchronization—to fuse discrete asynchronous events into a continuous sequence with complete fields and time alignment, termed the standardized event stream.

[0044] Interpolation is used to complete all missing records. Missing records can be categorized into three types: identity missing, location missing, and environment missing. Records with missing identity or location cannot participate in the propagation chain inference and are directly deleted. Environment missing records are completed using temporal neighborhood interpolation: the two most recent valid environment records at the same location node are searched in time, and their values ​​are averaged as the estimated value. If no valid value is found within a set time window, the valid environment record at the spatial nearest location is used instead. This strategy utilizes both temporal smoothing and spatial correlation to match the completed value with the local scene; in an experiment in an airport waiting area, the original missing rate was about 1%, and the environment field completeness rate reached 100% after interpolation.

[0045] Neighborhood averaging is used to replace abnormal records. Environmental anomalies are often caused by sensor drift or transient interference, and location anomalies typically manifest as jump points. Edge computing nodes maintain the window mean and standard deviation of each metric using a fixed-length sliding window. If a new value deviates from the window mean by more than a dynamic threshold, it is judged as an environmental anomaly and replaced with the window mean; if the Euclidean distance between two consecutive location coordinates exceeds the physical reach threshold, the current location is corrected to the previous coordinate. This correction weakens the impact of extreme noise on viral load estimation and maintains trajectory continuity. A shopping mall deployment case shows that the anomaly rate decreased from 0.5% to 0%, and successfully eliminated jump coordinates caused by ultra-wideband base station reflections.

[0046] Unified time synchronization is achieved through a network time synchronization protocol. Each sensing device periodically sends its local timestamp to the edge nodes, which then estimate the clock drift variable Δt based on the round-trip delay. For the original timestamp t... raw implement:

[0047] t sync =t raw -Δt

[0048] This allows records from different devices to be mapped to a unified millisecond time axis. Where t... sync This represents the timestamp after calibration. To avoid accumulated errors, resynchronization is performed whenever Δt exceeds a threshold. In the airport implementation, the maximum time difference between multi-source records after calibration is less than 0.1 seconds.

[0049] The standardized event stream obtained through three steps possesses three key characteristics. First, the identities, location coordinates, and environmental vectors of individuals within the same time window correspond one-to-one, ensuring no temporal misalignment in the event graph construction. Second, the numerical distribution exhibits no sharp anomalies, providing low-noise input for maintaining long-term homology. Third, the recording intervals are uniform, facilitating encoding with a fixed impulse threshold by a neuromorphic processor.

[0050] The effects are demonstrated through two comparative experiments. In Experiment 1, in a relatively closed environment, using the unwashed raw event stream, the proportion of spurious pulses in the pulse event stream was approximately 8%, and the lifetime curve of a long-lasting homogeneous loop exhibited a long-tailed distribution. After adopting the standardized event stream, the proportion of spurious pulses decreased to 1%, the loop lifetime variance decreased by 30%, and the mean absolute error between the infection prediction field generated by the differentiable cellular automata and the real air sampling curve decreased by 25%. In Experiment 2, in a closed environment, the comparison before and after time calibration showed that misaligned events within the sliding window decreased by two orders of magnitude, the distribution of intervention nodes in the tensor network solution became more concentrated, the number of Ising Hamiltonian iterations decreased from 40 to 32, and the slope of the environmental viral load decrease increased by 40% within 60 seconds after the intervention command was issued.

[0051] The interpolation completion and neighborhood averaging replacement rely on local time window statistics and spatial nearest neighbor relationships to maintain reasonable completion values ​​while avoiding computational delays caused by introducing external global models. Unified time synchronization ensures millisecond-level synchronization among different sensing devices, providing precise boundaries for pulse window settings. By performing preprocessing at the edge, this invention transforms raw asynchronous and noisy data into a unified, smooth, and reliable standardized event stream, providing a solid data foundation for subsequent pulse coding, topology analysis, diffusion generation, and intervention decisions, thereby supporting real-time prediction and responsive early warning of infectious disease transmission paths.

[0052] like Figure 2 As shown, the standardized event stream is input into a neuromorphic processor and converted into a pulse event stream. An event graph is constructed within a preset time window, and persistent cohomology is calculated to obtain a topological fingerprint.

[0053] The purpose of feeding the standardized event stream to the neuromorphic processor is to transform discontinuous discrete events into a pulse event stream with uniform temporal resolution, providing low-latency input for subsequent topology analysis using hardware-level pulse coding. The neuromorphic processor employs a spiking neural network architecture, triggering a pulse on-chip for each received person entity-spatial location pairing event. Pulse generation is implemented based on a leakage integral model, which can be abstracted as follows:

[0054] s(t) = exp[-β(t-t0)]

[0055] Where s(t) represents the pulse intensity, t is the current time, t0 is the time of the last triggering of the same entity at the same location, and β is the attenuation coefficient. This formula means that if the same combination occurs repeatedly within a short period, the pulse intensity accumulates rapidly, highlighting high-frequency interaction. The pulse stream does not require numerical storage within the device; it is only output externally in the form of events, thus resulting in low latency and low power consumption. The attenuation coefficient β is chosen to ensure that the attenuation time constant is slightly smaller than the preset sliding window to prevent cross-window pulse superposition.

[0056] The system uses a fixed window to perform sliding accumulation on the pulse event stream. The window length is set according to the activity rhythm of the scene; 1 second was measured in an airport waiting area, and 2 seconds in a nursing home corridor scene. An event graph is constructed for each window: the vertex set contains both personnel entity identifiers and spatial location identifiers, and the edge weights are the pulse counts within the window. After obtaining the event graph, the system executes a Vietoris-Rips filtering sequence, connecting vertices step by step and filling higher-order simplexes with distance thresholds corresponding to the pulse counts. The output of the filtering sequence is a series of simplex complexes, and loops that appear or disappear during scale growth constitute a homology class. To measure loop stability, the concept of persistent homology is introduced. For each loop, the birth scale b and death scale d are recorded, and the loop lifetime is defined as:

[0057] L=db

[0058] A larger L value indicates that the loop remains connected across multiple scales, representing genuine cyclic interaction rather than noise. This invention extracts only one long-lasting homology, i.e., loop information, as the topological fingerprint. The topological fingerprint includes three indicators: the lifetime of the longest-lived loop, the mean lifetime, and the variance of lifetime, used to describe the intensity of cyclic propagation within a window.

[0059] This invention chooses persistent cohomology instead of traditional clustering metrics because common connected components can only identify group size and cannot distinguish closed propagation chains. Persistent cohomology captures loops that directly reflect closed-loop contact structures formed by multiple entities at multiple locations. For example, if person A and person B enter and exit two rooms successively within a short period and then meet again, this pattern easily leads to repeated exposure. By measuring loop lifetime, the potential "repeated contact rate" can be estimated macroscopically, providing a basis for subsequent adjustment of the diffusion radius using differentiable cellular automata.

[0060] In the relatively enclosed environment of Example 1, the system uses Bluetooth beacons to establish identification relationships between personnel and fixed facilities, acquiring approximately 50,000 pulses per second. With a window length of 1 second, an event graph is constructed for 120 static nodes. Calculations show that the average lifetime of the longest-lived loop during peak hours is 40% higher than during off-peak hours, consistent with on-site operation frequency records. After inputting this loop lifetime into a differentiable cellular automaton, the system automatically amplifies the diffusion radius, increasing the overlap between high-risk areas in the generated infection prediction field and subsequently tracked actual infection trajectories by 0.16.

[0061] Example 2 was conducted in a separate, enclosed environment. Offline video playback confirmed that a cyclical queuing trajectory with a cycle of approximately 4 minutes was captured by a long-lasting, homogeneous, stable algorithm. Traditional density-threshold-based clustering algorithms failed to identify this loop due to fluctuations in the rate of personnel entering and exiting. Topological fingerprinting, based on differences in loop lifetime, categorized this loop as one of the top three high-risk patterns, and the system accordingly deployed mobile air purification devices at the bends in the queue. Subsequent monitoring showed a 35% reduction in the peak concentration of viral aerosols.

[0062] Hardware pulse coding in neuromorphic processors offers two advantages over software sampling. First, it eliminates the need for continuous event streaming; pulses are triggered as soon as an event arrives, reducing CPU load. Second, the leak integral model and persistent coherence analysis are a natural match, as the pulse intensity decay is already reflected in graph distance weighting, reducing the need for additional smoothing operators. In real-world testing on the same edge hardware platform, hardware pulse coding combined with persistent coherence reduces computational latency by 60% and power consumption by 45% compared to traditional event clustering with loop closure detection.

[0063] Preferably, the pulse event stream constructs an event graph within a continuous time window. The vertices of the event graph correspond to personnel entities and spatial locations, the edge weights correspond to the number of pulses within the time window, and a topological fingerprint is generated by calculating a long-lasting homology using the Vietoris–Rips method.

[0064] The pulse event stream is a discrete pulse sequence triggered by a neuromorphic processor at the millisecond level. Each pulse is bound to a key-value pair of "person entity - spatial location" and a precise timestamp. To measure the closed-loop contact structure formed by multiple entities across multiple locations within a window, this invention accumulates the pulse event stream using a fixed-duration sliding window and constructs an event graph within each window. The event graph consists of a vertex set composed of person entity identifiers and spatial location identifiers: if person entity A triggers a pulse n times at location X within the current window, a weighted edge is added between vertex A and vertex X, with the edge weight denoted as the pulse count n. Since multiple entities may trigger mixed pulses at the same location within the same window, the event graph naturally encompasses three interaction modes: person-location, person-human, and location-location. This graph has a much smaller number of vertices and edges than the original number of pulse entries, thus providing a significant dimensionality compression advantage for subsequent topological operations.

[0065] To extract cyclic propagation patterns from event graphs, this invention introduces a Vietoris-Rips filtering sequence from a topology data analysis framework. Specifically, it sets a set of incremental distance thresholds {∈0,∈1,…,∈…}. m For any two vertices v i With v j Define “graph distance”:

[0066]

[0067] Where w ij Indicates v i With v j The edge weights between the points are defined by δ, a minimal positive number used to avoid zero denominators. Smaller distances indicate more frequent interactions. When the distance is no greater than the current threshold ∈ […]. k In a complex, connect the edges; if the pairwise distance between any three vertices is not greater than ∈ k If , then a triangle is filled; and so on, higher-order simplexes can be generated. As ∈ increases, the complex gradually evolves from a sparse graph to a complete graph. A long-term homology is calculated for this filtering sequence, i.e., tracking the generation and disappearance of all loops at different thresholds. The birth threshold is denoted as b, the death threshold as d, and the lifetime is defined as db. Loops with long lifetimes indicate that they remain closed and connected over a wider threshold range, and are therefore more likely to correspond to genuine collective cyclic contact rather than sampling noise.

[0068] This invention constructs a topological fingerprint vector by combining the most significant loop lifetime of each sliding window, the mean loop lifetime within the window, and the variance of the lifetime distribution, to characterize the closed-loop propagation strength within the window. Compared to traditional methods that use the number or density threshold of connected components, a long-term homology approach can simultaneously capture the higher-order structure of "loops" and exhibit a smooth response to scale parameters. The topological fingerprint is then input into a differentiable cellular automaton to dynamically adjust the diffusion radius of the convolution kernel: the larger the mean lifetime, the larger the convolution kernel diffusion radius is adjusted to enhance the spatial diffusion simulation of potential loop propagation chains.

[0069] For example, in a relatively enclosed environment with dense queues, the window length is set to 1 second. The event graph includes person entities and location points in three functional areas. After Vietoris-Rips filtering, the system found that the longest-lived loop remained closed across three consecutive threshold levels, and its vertex sequence corresponded to the cyclical movement trajectory at the bends in the queue. By comparing the fluctuations in the average loop lifetime over different time periods, the system automatically determined that the cyclical interactions were most frequent during peak hours and preemptively increased the airflow of air purification equipment in that area. Subsequent validation showed that the peak aerosol concentration decreased by approximately 35% compared to the uninterrupted state.

[0070] In another closed environment with periodic patrols, personnel patrolled approximately every 10 minutes, and the loop lifetime within the same window exhibited significant periodic fluctuations. The system captured the high-frequency patrol periods corresponding to the lifetime peaks, dynamically amplified the convolution kernel diffusion radius, and increased the overlap between the generated infection prediction field viral load curve and the air sampling curve by 0.15. Because loop statistics do not rely on preset trajectory templates, the system can maintain equal recognition and adaptability in different building layouts.

[0071] From a hardware efficiency perspective, the spiking neuromorphic processor operates in an event-driven manner, without high-frequency polling. Spiking intensity decays automatically based on a leakage integral model, naturally encoding interaction frequencies and suppressing spatiotemporal noise. Persistent coherence computation is performed within a window graph, typically with 10 to 10^5 vertices. 3 By combining sparse complex compression, the memory footprint is significantly lower than that of conventional high-order graph mining algorithms. Edge embedded platform tests show that the topological fingerprint calculation for each frame takes approximately 5 milliseconds, fully meeting real-time prediction requirements.

[0072] Preferably, a long-term homology value is used to set the diffusion radius of the differentiable cellular automaton convolution kernel, so that the convolution kernel enhances the diffusion effect when the loop structure increases, thereby generating an infection prediction field.

[0073] To ensure that the infection prediction field promptly reflects the propagation amplification effect caused by closed-loop interactions, this invention proposes dynamically controlling the diffusion radius of the convolution kernel of a differentiable cellular automaton using a long-lasting homogeneous lifetime statistic. When an increase in loop structures with longer lifetimes is detected within the window, the automaton expands the virus cross-grid propagation step size on the spatial grid; when closed loops are sparse, it automatically contracts the diffusion. This coupling strategy establishes real-time feedback between the topological layer and the physical diffusion layer, which can significantly improve the positioning accuracy of high-risk areas while avoiding excessive diffusion in low-risk stages.

[0074] The calculation of long-term homogeneous lifetime is based on an impulse event graph. The impulse event graph uses vertices composed of personnel entities and spatial locations, with edge weights taken as the reciprocal of the number of impulses within a sliding window as the "distance". A Vietoris-Rips filtered complex is generated according to an increasing distance threshold sequence, recording the existence intervals of loops at different scales. A birth threshold *b* and a death threshold *d* are defined for any loop, and lifetime is measured by L = *db*. The window topological fingerprint is derived from the longest lifetime *L*. max and mean lifespan L μ Composition. L max Reflecting extreme closed-loop events, such as dense queuing; L μ Summarize the overall circulation density. The longer the lifespan, the more stable the closed loop, meaning that people may repeatedly come into contact with the virus at multiple locations, which amplifies the spread of the virus.

[0075] Differentiable cellular automata discretize the building space into a uniform three-dimensional grid, with each grid storing the viral load ρ and the viral inactivation factor. State vector writing The state of the automaton at time step t is updated as follows:

[0076]

[0077] Where σ is the element-wise compression function, W R A spherical convolution kernel with radius R, where * denotes convolution, c nbr Let represent the neighborhood tensor, and b be the bias. The core coupling formula sets the diffusion radius:

[0078] R = αL max +βL μ

[0079] α and β are constant mapping coefficients. As the closed-loop lifetime within the window increases, R increases, allowing the automaton to spread the virus in a larger neighborhood; as the closed-loop lifetime decreases, R decreases, reducing unnecessary spread. The convolution kernel weights are generated using a spherical mask: all grid points within a lattice point at Euclidean distance R from the center grid are assigned the same spread weight, while other locations have zero weight. This kernel is updated once per second, ensuring the propagation model's sub-second response to topological changes.

[0080] During the project implementation, the edge server immediately calculates (L) after the sliding window ends.max ,L μ Substituting into the formula, we get R. Kernel updates are equivalent to mask switching and do not change the depth of the tensor computation graph; the measured computation time only increases by about 8% compared to the fixed kernel mode. The reasons for choosing a linear mapping instead of a complex nonlinear mapping are: first, the linear relationship is sufficient to distinguish between high-risk and low-risk windows; second, the linear coefficients can be obtained through offline grid search and can also be fine-tuned using gradients in the online phase, facilitating deployment and maintenance.

[0081] In terms of technical effectiveness, the dynamic radius brings three improvements. First, improved time alignment: during the peak of the closed-loop process, the time difference between the predicted viral peak and the airborne monitoring peak is reduced from 90 seconds to 60 seconds. Second, reduced spatial error: the ratio of predicted high-risk volumes to measured positive volumes is reduced from 1.8 to 1.2. Third, enhanced resource efficiency: the tensor network risk weights are more concentrated, reducing intervention nodes by 20% while increasing the viral load decline slope by 40%, indicating that resources are focused on truly high-risk locations.

[0082] Example 1 (relatively enclosed environment). 28,800 window fingerprints were collected over eight hours, with α=1 and β=0.5. R adaptively varied within a 2–4 ​​grid range, reducing the predicted peak position error from 4.6 meters to 3.1 meters and the peak time error from 85 seconds to 60 seconds. The air purification valve triggered 45 seconds in advance, resulting in a 38% decrease in peak virus concentration.

[0083] Example 2 (Closed Environment). The loop lifetime fluctuates periodically over 600 seconds, with a peak of 1.2 and a trough of 0.3. The system increases the R grid to 3.5 during the peak lifetime period and decreases it back to 2 grids during the trough period. The difference between the predicted high-risk volume and the measured volume converges to within 0.2 times, and the tensor network iterations are reduced from 40 to 32. The number of intervention nodes is reduced by 22%, and the virus descent slope is increased by 42%.

[0084] Compared to the fixed-radius model, the dynamic-radius strategy maintains controllable computational load while achieving a double improvement in spatiotemporal prediction accuracy. Furthermore, it does not rely on scene-specific path templates and supports rapid migration across building layouts. Since updating convolutional kernel weights does not modify model parameters, the overall inference pipeline can still be trained or fine-tuned end-to-end, preserving differentiability.

[0085] like Figure 3 As shown, the pulse event stream is aggregated according to the time window to form a variable-order hyperedge event graph. The topological fingerprint is used to set the differentiable cellular automaton convolution kernel to evolve into an infection prediction field in the initial state of the environmental field. The variable-order hyperedge event graph, the infection prediction field and the topological fingerprint are combined into a conditional tensor and input into a conditional fractional diffusion neural network to generate a propagation event sequence.

[0086] After standardization and neuromorphic encoding, the pulse event stream is aligned on a millisecond-level timeline, retaining the ternary information of "person entity—spatial location—time". To simultaneously capture propagation potential from three dimensions—spatial, topological, and dynamic—this invention sequentially performs variable-order hyperedge event graph construction, differentiable cellular automaton evolution, and conditional fractional diffusion generation. These three processes are concatenated at the tensor level as a unified input, driving the conditional fractional diffusion neural network to infer future propagation event sequences, achieving minute-level early warning.

[0087] Variable-order hyperedge event graphs aggregate impulse events into hyperedges within a sliding window, chronologically. A hyperedge contains several vertices participating in the same "contact encounter," which can be either individuals or spatial locations. The order of a hyperedge is the number of vertices; the higher the entity aggregation density within the window, the more high-order hyperedges are generated. Compared to traditional binary relation graphs, hyperedge event graphs can directly represent shared exposure of multiple subjects in the same area, such as multiple people simultaneously staying in a closed room. The weight of a hyperedge is set as the sum of all impulse counts within the window; a larger weight indicates a higher contact frequency. Because the window length is fixed, the variable order allows the graph model to maintain both temporal consistency and order flexibility: second- and third-order hyperedges dominate during low-density periods, while fourth- and even higher-order hyperedges dynamically appear during high-density periods, while maintaining a constant dimension of the single-window feature vector, facilitating tensor splicing.

[0088] Differentiable cellular automata and topological adaptation, topological fingerprinting by L max With L μ The descriptions represent the longest lifetime loop and the average lifetime of the window, respectively. These are mapped to the convolution kernel radius:

[0089] R = αL max +βL μ

[0090] Where α and β are constants. A larger convolution kernel radius allows the virus to jump between more neighboring cells. The convolution kernel generation logic is as follows: if the Euclidean distance from the grid center to the neighboring cell does not exceed R, the corresponding weight is 1; otherwise, it is 0. Global normalization is then applied to ensure diffusion conservation. State vector. According to the formula:

[0091]

[0092] Update, where σ is the compression function and b is the bias. After 20 consecutive iterations, the infection prediction field is obtained. A three-dimensional tensor. The dynamic radius ensures that as closed-loop interactions increase, the diffusion radius expands synchronously, quickly reflecting the risk of cyclic propagation.

[0093] Conditional tensor and fractional diffusion: To allow the generative model to simultaneously consider crowd interaction structure, environmental load distribution, and topological cycle strength, the three sources of information are concatenated in the tensor depth dimension as follows:

[0094]

[0095] in Let be the adjacency tensor of a variable-order hyperedge event graph. For the infection prediction field, Φ is the topological fingerprint broadcast to and The conditional fractional diffusion neural network uses a three-dimensional convolution-deconvolution symmetric structure. In the forward pass, Gaussian noise is added progressively according to a noise schedule. In the reverse pass, the noise gradient is learned through fractional matching, ultimately outputting a sequence of propagating events within the next τ seconds. Compared to the unconditional model, the conditional tensor explicitly injects spatial, topological, and dynamic information, significantly improving the interpretability and physical consistency of the generated sequence.

[0096] This invention achieves several improvements: **Peak prediction time alignment:** During closed-loop peak periods, the time difference between the predicted peak infection rate and the air sampling peak rate is reduced from 90 seconds to 60 seconds. **High-risk area localization:** The ratio of predicted high-risk volume to measured positive volume decreases from 1.8 to 1.2, reducing the false alarm area by approximately 33%. **Intervention resource conservation:** While the concentration of risk weights in the tensor network increases by 19% and the number of intervention nodes decreases by 22%, the viral load reduction slope increases by 40%, indicating greater control benefits with the same resources.

[0097] Example 1: The system ran continuously for 8 hours in a relatively closed environment, collecting data from 28,800 windows. The average topological fingerprint L... max =0.9,L μ =0.5, α=1, β=0.5, R adaptively varies within a 2–4 ​​grid range. The distance between the predicted peak location and the sampled peak decreased from 4.6 meters to 3.1 meters, and the peak time difference decreased from 85 seconds to 60 seconds. Opening the air purification valve 45 seconds earlier reduced the peak virus concentration by 38%.

[0098] Example 2: A closed-loop fluctuation with a 600-second cycle occurred in a closed environment, with a peak loop lifetime of 1.2 and a valley lifetime of 0.3. The system dynamically increased the RRR to 3.5 grids in the peak region and decreased it back to 2 grids in the valley region. The predicted volume error converged to 0.2 times, and the number of Ising Hamiltonian iterations decreased from 40 to 32. The number of intervention nodes decreased by 22%, and the virus descent slope increased by 42%.

[0099] Preferably, the infection prediction field is generated in a differentiable cellular automaton with viral load and viral inactivation factor as state variables and is updated multiple times in discrete time steps until the state change between two adjacent updates is lower than a preset threshold.

[0100] This invention uses differentiable cellular automata to generate an infection prediction field at a spatial grid scale. The core idea is to use airborne viral load ρ and viral inactivation factor. As state variables, these two elements adaptively diffuse, decay, and couple in the discrete time domain, thus characterizing the spatial distribution of infectious particles in the air over several seconds. Unlike common solutions to continuous partial differential equations, differentiable cellular automata can be trained end-to-end within a tensor framework and can be linked in real time with the convolution kernel radius and topological fingerprint.

[0101] Building or transportation spaces are discretized into uniform three-dimensional grids, with grid indices denoted as (x, y, z). Each grid stores a state vector.

[0102]

[0103] Where ρ represents the viral load per unit volume of air; This indicates the virus inactivation factor; the higher the value, the easier it is for the virus to become inactivated.

[0104] State evolution follows:

[0105]

[0106] t is the discrete time step; W R The symbol represents a spherical convolution kernel with radius R; "*" indicates a 3D convolution. Let be the neighborhood state tensor; b be the bias vector; σ be the element-wise compression function, ensuring that ρ is non-negative after the update. It is between 0 and 1. The kernel radius R is determined by the topological fingerprint, as shown in the formula:

[0107] R = αL max +βL μ

[0108] Where L max L represents the longest loop lifetime in the current window. μ The lifetime mean is given; α and β are mapping coefficients. When closed-loop interactions are frequent, both increase, allowing R to spread further; conversely, they shrink, achieving topological-physical coupling.

[0109] During the spread, the virus becomes inactive through an exponential decay function:

[0110]

[0111] Updated separately, where γ is the deactivation rate and Δt is the time step. This decay term is written into the bias b, which can be completed together after the convolution operation. Iteration continues until the grid state changes between two adjacent times satisfy:

[0112]

[0113] θ is the convergence threshold. The predicted field is considered to have stabilized when this condition is met.

[0114] The algorithm process includes four steps: 1. Mesh initialization: Based on the real-time sensor, assign spatial interpolation values ​​to the initial p. 1. Obtain the device calibration constant. 2. Convolution kernel generation: Read the latest αL... max +βL μ 3. State Iteration: A convolution-compression-summing loop is executed on the GPU, calculating the maximum norm difference at the end of each round; the process stops if the difference is less than θ or the maximum number of steps is reached. 4. Result Encapsulation: A three-dimensional tensor is output. It serves as an infection prediction field for subsequent tensor splicing and diffusion models.

[0115] Since the convolution kernel weights are sparse masks and the radius is usually no more than 4 grids, the time and space complexity of a single round of convolution is equivalent to that of a conventional 3×3×3 kernel. Multiple iterations are updated in-place in the GPU memory, without the need for frequent memory relocation. In actual tests, it takes about 12ms under the conditions of 64×64×8 grid size and 32-step update.

[0116] This invention introduces an adaptive mechanism through dynamic diffusion radius, significantly improving the sensitivity of the prediction field to the cyclic propagation chain. During the peak of the closed-loop lifetime, the model automatically increases the diffusion range, shortening the time difference between the peak virus concentration and the peak air sampling concentration by approximately 30 seconds. During the sparse closed-loop period, the convolution kernel radius automatically converges, avoiding misjudging large areas of high risk. Compared to a fixed-radius baseline, the mean absolute error decreases by 25%, the recall rate for high-risk grids improves by 0.14, while the inference latency increases by only 8%.

[0117] In an example, after continuous operation for 8 hours in a relatively enclosed environment, the average fingerprint density of the sliding window (L) was... max ,L μ ) = (0.9, 0.5), α = 1, β = 0.5. The upper limit of iteration is set to 32 steps, and the convergence threshold θ = 10. -4 The peak segment converged in 26 iterations, while the low-peak segment converged in 18 iterations. The distance between the predicted peak position and the sampled peak position decreased from 4.6m to 3.1m, and the peak time difference decreased from 85s to 60s. The air purification valve was triggered 45s earlier than the baseline, and the peak virus concentration decreased by 38%.

[0118] In a closed environment, there is a closed-loop fluctuation with a period of 600s. With a peak R=3.5 grid and a trough R=2 grid, the ratio of predicted high-risk volume to sampled volume converges to 1.2. The risk weights of the tensor network are more concentrated, the number of intervention nodes is reduced by 22%, and the slope of virus descent is increased by 42%.

[0119] Preferably, the variable-order hyperedge event graph, infection prediction field, and topological fingerprint are aligned according to time index, entity index, and spatial coordinates before being composed of a conditional tensor, and then concatenated into a unified tensor in the depth dimension before being input into a conditional fractional diffusion neural network.

[0120] Variable-order hyperedge event graphs, infection prediction fields, and topological fingerprints characterize propagation potential from three information chains: behavioral structure, physical field evolution, and topological cycle, respectively. If these three are fed into the model independently, different coordinate systems and resolutions will lead to spatiotemporal misalignment, making it difficult to maintain physical consistency in the generated results. This invention first aligns the three-source tensors along axes according to the time index t, entity index e, and three-dimensional spatial coordinates (x, y, z), and then concatenates them in the depth dimension into a unified conditional tensor. Ultimately, the conditional fractional diffusion neural network is driven to invert the propagation event sequence in a noisy space. This step achieves a cross-layer fusion of propagation prediction based on big data, moving from "static splicing" to "spatiotemporal coupling".

[0121] Three-source tensor alignment logic, variable-order hyperedge event graph adjacency tensor Indexed by (t, e1, e2), with the range being the hyperedge weights; Infection prediction field tensor Indexed by (t,x,y,z), with a value range of viral load; the topological fingerprint tensor Φ is extended to the same level as... Synchronized (t,e1,e2) scale. To eliminate resolution differences, firstly... Spatial index mapping to entity index: Let entity e be in the grid (x e ,y e ,z e The cell containing Φ is assigned the corresponding ρ value, which is then broadcast to the (e1,e2) axis. Next, Φ is copied to the (x,y,z) axis to align it with... Same dimension. Ultimately, all three have a unified primary index (t, e, x, y, z), thus ensuring that data from the same time slice and the same entity in the same spatial unit can be mapped bit-by-bit. Conditional tensors are concatenated and sequentially stacked along the depth dimension d to obtain:

[0122]

[0123] The symbol “⊕” indicates channel splicing. After splicing... The shape is (T,E,X,Y,Z,D), where D=3. To be compatible with 3D convolutional networks, (E,X,Y,Z) is first unfolded into four-dimensional voxels, and then T is used as the time channel input to the model. This preserves the discrete structure of the hyperedge weights, embeds the continuous gradient of the viral field, and carries the loop lifetime prior, avoiding information gaps.

[0124] Conditional Score Diffusion Neural Network, the model's forward process The noise intensity at time step s is written as follows:

[0125] ε s =ε0·exp(ks)

[0126] ε0 represents the initial noise amplitude, and k represents the exponential gain. The reverse process learns the noise gradient using fractional matching loss. The propagation event sequence with future window length τ is generated through iterative inversion. Because the conditional tensor contains entity pairs, spatial coordinates, and topological priors, the network can trace the propagation trajectory that best fits the three-source constraints in a high-dimensional noisy space.

[0127] This invention ensures that the prediction at any time is based on the same source data, eliminating the "scene floating" phenomenon; the hyperedge structure provides multi-subject contact weights, the topological fingerprint provides the closed-loop strength, and the infection field provides the physical gradient. After the three are fused, the time error of the generated sequence is reduced by 30% compared with the unconditional model; the spatial recall rate is improved, and the ratio of predicted high-risk volume to measured volume is reduced from 1.8 to 1.2; the number of network convergence steps is reduced by about 15%, and the computational energy consumption is reduced by 12%.

[0128] In this example, 28,800 sliding windows were collected in a relatively enclosed environment. After alignment of the three-source tensors, they were stitched together into a tensor of (28800, 64, 32, 32, 8, 3). The positive noise schedule step size was 100, and k = 0.04. After 15 rounds of model training, the overlap between the generated sequence and the case retrospective path reached 0.82, and the average early warning time was shortened by 90 seconds; the viral load decline slope within 60 seconds after intervention was triggered increased by 40%. These results verify the high efficiency and accuracy of the conditional tensor alignment and stitching strategy in the fusion of complex population and multi-source environmental data, providing key support for the second-level transmission early warning of this invention.

[0129] Preferably, the conditional fractional diffusion neural network adopts a multi-scale encoding and decoding structure, injects incremental random noise in the forward diffusion, and trains the network weights through fractional matching loss in the reverse process to generate a propagation event sequence that meets the conditional tensor constraints.

[0130] Conditional fractional diffusion neural networks are responsible for processing multi-source conditional tensors. The task is to parse the data into a sequence of future propagation events. Unlike traditional sequence-to-sequence recursive models, the diffusion model first constructs an easily sampled stochastic process in the noisy space, and then gradually denoises the data in the reverse process by learning a fractional function, so that the final sample is consistent with the true conditional distribution. This mechanism naturally possesses flexible multimodal generation capabilities and a stable gradient backpropagation path, making it particularly suitable for highly uncertain and multi-constrained infectious disease propagation scenarios.

[0131] Forward diffusion transforms the conditional tensor By gradually injecting random noise, a noise sequence is obtained. The stochastic process at step s is defined as:

[0132]

[0133] Where βs Let ε be the noise intensity at step s in the scheduling table. s This is a Gaussian noise tensor of the same shape. Set β... s The noise increases exponentially, significantly increasing the proportion of noise in the later stages, until the final state... It is approximately an isotropic Gaussian distribution. The inverse process requires learning a fractional function:

[0134]

[0135] Given the conditional context (i.e., the variable-order hyperedge graph, infection prediction field, and topological fingerprint in the same time slice), this invention approximates the gradient using a score-matching loss during the training phase. During inference, it sequentially proceeds from... Sampling direction Inversion is the process of obtaining the sequence of propagation events within the next τ seconds.

[0136] A symmetrical multi-scale encoder-decoder backbone is employed, consisting of a stacked interleaved structure of 3D convolutional and deconvolutional blocks. The encoding branches compress spatial resolution and expand channel depth layer by layer, capturing spatial gradients within the local receptive field. Skip connections stitch early details to the corresponding decoding branches, restoring the high-resolution structure. Each convolutional layer is followed by a conditional residual block, with input consisting of the current feature and the time-step encoding γ(s), where γ maps integer time steps to a high-dimensional vector using a positional encoding function. This allows the network to simultaneously perceive the conditional tensor, the diffusion phase, and the local context across all scales.

[0137] The training strategy involves partitioning the dataset by sliding windows. Each sample contains a variable-order hyperedge graph of historical sliding windows, an infection prediction field, a topological fingerprint, and a tensor of real future propagation events. During training, time steps s and noise ∈ [the range] are randomly sampled. s Generate noisy samples Network output score estimation Loss function:

[0138]

[0139] Differentiable across the entire link. Optimization is achieved using adaptive moment estimation, with a linear learning rate followed by 5 warm-up rounds and cosine annealing. β is maintained during training. s The fixed schedule facilitates comparison of different experiments.

[0140] The reasoning process begins with Gaussian sampling to obtain... Iterate backwards from s = S to 1, following:

[0141]

[0142] in For the correction term, ζ s This is for the re-injection of small noises. Obtained after completing the S-step. The future entity-location-time triple sequence can be recovered by resolving the tensor channels. Since the network has learned fractional functions under various scale conditions, the generated sequence naturally satisfies the three constraints of hyperedge weights, viral gradients, and topological lifetimes, ensuring physical consistency.

[0143] In an example, in a relatively closed environment, the scheduling steps S = 100 and the maximum noise β is set. S =0.02, trained for 15 epochs. Compared with the unconditional diffusion model: the prediction peak time error decreased from 90 seconds to 60 seconds; the overlap between the propagation path and the backtracking path increased from 0.66 to 0.82; the number of intervention nodes in the tensor network decreased by 22%, but the viral load reduction slope increased by 40%.

[0144] In a closed environment, with the same configuration, the early warning time is reduced by an average of 90 seconds, and the recall rate for locating high-risk areas is improved by 0.14. The number of training convergence epochs is about 30% less than that of the standard U-Net because the conditional tensor has integrated multimodal priors in the input layer, resulting in more stable gradients.

[0145] A tensor network is constructed based on the propagation event sequence. The tensor network is represented as the Ising Hamiltonian to determine the intervention node. Intervention commands are issued to the environmental control equipment. The tensor network, the conditional fractional diffusion neural network, and the differentiable cellular automaton are updated based on the environmental feedback measured before and after the intervention to form an adaptive closed loop.

[0146] In this invention, tensor networks serve as a bridge between "propagation sequence" and "intervention decision". First, the future propagation event sequence given by the conditional fractional diffusion neural network is parsed into a set of triples <personnel entity e>. i Spatial location p i Time of occurrence t i For predictions spanning T seconds, all events can be mapped to a six-dimensional tensor:

[0147]

[0148] The first to third dimensions correspond to the entity index, spatial index, and temporal index, respectively; the fourth dimension represents the binary contact state, the fifth dimension the binary infection state, and the sixth dimension contains K numerical features (pulse intensity, dwell time, local virus concentration, etc.). To reduce computational complexity, column tensor decomposition is used to decompose T into tensor columns of length D. Ensure that high-level interactive information is compressed without being lost.

[0149] Modeling intervention nodes using the Ising Hamiltonian can be approached by viewing the tensor column as a weighted graph, where each potential intervention unit (optional location of a vent, purifier, or disinfection robot) is denoted as vertex v. i Define a binary spin variable s. i∈{-1,1}——s i =1 indicates that node v is selected. i Implement intervention, s i =-1 indicates not selected. The propagation probability π is reconstructed from the tensor sequence. ij and node risk ρ i Construct the Ising-Hamiltonian:

[0150]

[0151] Among them, the coupling term J ij =λ1π ij ρ i ρ j Measure the virus reduction gain resulting from the linkage between two nodes; field item h i =λ2ρ i The reduction value of a single node is measured, with λ1 and λ2 being the adjustment coefficients. Lower energy indicates a smaller expected viral load. Minimizing H is equivalent to selecting the optimal set of intervention nodes under budget constraints.

[0152] To solve for the minimum energy state in real time, this invention utilizes an optical delay loop reservoir to obtain an approximate gradient: {J ij ,h i The light pulse intensity is encoded and injected into a ring fiber. The physical evolution of simulated annealing is completed by interference attenuation. After receiving the light output on the digital side, it can converge in no more than 32 iterations using alternating linear least squares combined gradient descent, satisfying the second-level response.

[0153] Closed-loop intervention and online updates to determine Subsequently, commands are sent to the corresponding environmental control devices via the edge controller, such as activating the HEPA filter, increasing the local airflow, or turning on the UV lamps. The system continuously collects the differences in airborne viral load (Δρ), carbon dioxide (ΔCO2), and temperature and humidity (ΔH) before and after the intervention, normalizing them into a feedback vector.

[0154] f = tanh([Δρ, ΔCO2, ΔH])

[0155] The feedback vector is used for three online incremental learning processes:

[0156] Tensor Networks – Fine-tuning the tensor column weights using ridge regression with target f, so that the propagation probability π ijIt instantly reflects actual changes after prevention and control; the conditional fractional diffusion network adds f to the fractional loss in the next batch of training, prompting the model to correct over- or under-propagation estimates; the differentiable cellular automaton adjusts the inactivation rate γ and diffusion mapping coefficients (α,β) according to Δρ, allowing the predicted field to quickly fit the measured gradient. This forms a closed loop of "prediction → intervention → feedback → re-prediction", with each cycle taking no more than 60 seconds, and the system adapts to the environment.

[0157] In one example, the system operated continuously for 8 hours in a relatively closed environment, selecting an average of 353,535 intervention nodes per cycle. Within 60 seconds of the initial intervention, the peak viral load decreased from 350 copies / m³ to 210 copies / m³. After two rounds of closed-loop updates, only 29 nodes were needed to maintain the same rate of decline for the same level of control, resulting in a 29% reduction in the energy function. In another closed environment, 48 nodes were initially selected, which were reduced to 36 after two rounds of feedback optimization, increasing the viral load decline rate by 42%.

[0158] Preferably, the tensor network constructed from the transmission event sequence calculates the risk weights by considering the number of people, the maximum viral load in the infection prediction field, the air exchange coefficient, and the strain coefficient. After mapping to the Ising Hamiltonian, it combines the energy gradient output by the optical delay loop reservoir system and uses alternating linear least squares and gradient descent to jointly determine the intervention nodes.

[0159] Tensor network risk modeling aims to transform the future propagation event sequence generated by conditional fractional diffusion neural networks into a computable objective of "where priority intervention is necessary." The core idea is to first assign spatiotemporal weights to each entity-location-time triple in the sequence, then calculate the node risk by comprehensively considering factors such as population size, virus intensity, ventilation capacity, and strain characteristics. Subsequently, this risk system is mapped to the Ising Hamiltonian to solve for the minimum energy structure. The energy gradient is output using an optical delay loop reservoir, and intervention nodes are locked within seconds using alternating linear least squares combined with gradient descent.

[0160] The risk weighting formula, the propagation event sequence is first arranged into a three-dimensional sparse matrix ε. e,p,t Non-zero elements record the frequency of interaction between entity e and location p within time t. The node arrival rate N is obtained by summing the values ​​along both the column direction (entity dimension) and the row direction (location dimension) of the matrix. e With stay rate S p Then, extract the maximum viral load ρ at position p from the infection prediction field tensor. max,p Ventilation capacity is expressed in terms of the air exchange coefficient (ACH). p Quantification; the strain factor is represented by the strain coefficient σ. The four factors are combined to form the risk weight:

[0161]

[0162] The parameters α, β, γ, and δ were calibrated offline using a grid search: α controls the influence of pedestrian density; β weighs virus intensity; γ suppresses the weight of high-ventilation areas; and δ filters the high infectivity of strains. p After normalization, the tensor diagonal blocks are filled to form a node diagonal matrix W. Simultaneously, the propagation probability π is used... pq Construct an adjacency matrix A using off-diagonal elements. Risk Tensor Network The edge weights are determined by Hadamard multiplication. The approach takes into account both the risks of the nodes themselves and the synergistic effects of the connected edges.

[0163] In the Ising Hamiltonian construction, to select the most valuable intervention node within the resource budget B, each location node p is assigned a spin variable s. p ∈{-1,1}, sp=1 indicates the execution of intervention. Symmetric coupling term:

[0164] J pq =λ1π pq w p w q

[0165] Single-point outdoor items:

[0166] h p =λ2w p

[0167] Overall Desin Energy:

[0168]

[0169] The first term minimizes the collaborative energy among high-risk nodes; the second term encourages the selection of single high-risk nodes; and the third term represents a soft budget constraint. The coefficients λ1, λ2, and λ3 are determined by sensitivity analysis, with λ3 typically being much larger than λ1 to strictly limit the number of nodes.

[0170] Energy gradient and numerical optimization: directly finding the minimum energy for discrete spins is an NP-complete problem. This invention utilizes optical delay loop reservoir hardware to quickly approximate the gradient: 1. [The following text appears to be a separate, unrelated section:] ... matrix {J... pq The encoding is a sequence of relative light intensities for pulse pairs, with a single-node potential h. p The encoding is the pulse bias intensity. 2. After the pulse passes through a single-mode fiber loop delay line, the output optical power decreases approximately over time after multiple interferences. 3. The optical power is converted into a digital gradient vector g via photoelectric conversion.

[0171] With the gradient in place, a two-stage solution is used: first, alternating linear least squares is used to find the continuous slack variables. Then use gradient descent combined with sign function projection to... The energy decrease convergence rate exceeds 90% within 32 iterations, satisfying real-time requirements.

[0172] Closed-loop intervention and updates, selecting intervention set V * ={p|s p After =1}, the edge controller sends the "start / stop signal - target location" message to the ventilation, filtration, UV, or disinfection equipment. Within 60 seconds before and after the intervention, the viral load difference Δρ, carbon dioxide difference ΔCO2, and humidity difference ΔH are continuously collected and normalized into a feedback vector f. The feedback is embedded in a three-layer model for fine-tuning: a tensor network, using f to update {w}. p} and {π pq The diffusion network adds f to the fractional loss to correct overfitting; the cellular automaton adjusts the inactivation rate and diffusion mapping coefficients based on Δρ. This closed-loop refresh cycle is 60 seconds, consistent with the computation rhythm of the sliding window topology.

[0173] Example: Operating in a relatively closed environment for 24 hours: Intervention is triggered 60 seconds before the predicted peak; in the first round, 35 nodes are selected, and the virus peak decreases from 350 replications / m³ to 210; after two rounds of feedback, only 29 nodes need to maintain the same rate of decrease, resulting in a 29% reduction in energy. 12-hour test in a closed environment: Initially 48 nodes are selected, reduced to 36 after feedback; the rate of decrease increases from 0.6 to 0.85 replications / m³ / second; total energy consumption is reduced by 18%, and strategy convergence is completed within 5 rounds.

[0174] Preferably, the environmental feedback is the difference in viral load in the environment before and after the intervention. After normalization, the difference is used to synchronously update the tensor network, the conditional fractional diffusion neural network, and the differentiable cellular automaton to complete the adaptive closed loop.

[0175] The environmental feedback loop connects the "feeling" and "action" of the prediction-intervention chain, ensuring that the system does not become inaccurate due to dynamic changes in the scenario. The core approach is to continuously monitor the viral load in the air, calculate the difference Δρ before and after intervention, and after normalization and compression, form a dimensionless feedback scalar f. Then, f is simultaneously injected into a tensor network, a conditional fractional diffusion neural network, and a differentiable cellular automaton. Each of the three adjusts its parameters according to a preset update law, achieving second-level adaptiveness.

[0176] Take n load observations before and after the intervention, and calculate the moving average for each:

[0177]

[0178] Difference:

[0179] Δρ=ρ after -ρ before

[0180] Directly reflects the effectiveness of the intervention: negative values ​​indicate successful reduction, while positive values ​​indicate ineffectiveness or rebound. To eliminate differences in room volume or detection accuracy, a reference load ρ is introduced. ref (The median viral load of the most recent day can be used) for proportional normalization, and then compressed to the (-1,1) interval using hyperbolic tangent:

[0181] f = tanh(Δρ / ρ) ref )

[0182] Thus, the absolute value of f represents the magnitude of the effect, and the sign indicates the direction.

[0183] Tensor networks enable rapid reweighting; tensor networks have passed node risk assessment. p With the right of collaboration J pq Approximate propagation path. After receiving f, perform an incremental ridge regression:

[0184]

[0185] Where H is the Ising energy and η1 is the learning rate. If f < 0, the risk of the intervened nodes is reduced, and the risk of the remaining high-residual nodes is appropriately increased; if f > 0, the adjustment is reversed to guide the selection of new nodes in the next round.

[0186] Fine-tuning of the conditional fractional diffusion network; the original formula for the fractional loss of the diffusion model is: Add a feedback regular expression:

[0187]

[0188] η² controls the amplitude. When f < 0, the gradient is regularly reduced to prevent overestimation; when f > 0, the gradient is reduced. Only the weights of the last two convolutional layers are fine-tuned, resulting in an inference latency increment of less than 2ms.

[0189] Differential cellular automata parameter correction: The automata diffusion-degradation equation is controlled by the inactivation rate γ and diffusion mapping coefficients α and β. Feedback update formula:

[0190] γ←γ-η3f,α←α+η4f,β←β+η4f

[0191] When f < 0, it indicates that the diffusion radius is sufficient, so the deactivation rate should be appropriately increased and the radius should be reduced; when f > 0, the deactivation rate should be reduced while the radius is enlarged to make the prediction field more sensitive.

[0192] In an example, during a one-day continuous test (in a relatively closed environment), the initial intervention value f = -0.24. After adjusting the three models, the number of intervention nodes required for the next round decreased from 35 to 29, while the peak viral load further decreased by 12%. After eight cycles, the absolute value of f stabilized at around 0.05, and the system entered maintenance mode. Compared to the baseline without feedback: the peak occurrence time error decreased from 60s to 45s; the recall rate of high-risk grids improved by 0.14; the average number of intervention nodes decreased by 18%; the relative decrease in viral peak value increased by 13%; and the additional computation time accounted for less than 10% of the total inference time. Closed-environment experiments also showed a similar trend, with a 42% increase in the viral load decline slope and a 15% reduction in energy consumption.

[0193] like Figure 4 As shown, a big data-based infectious disease transmission path generation and early warning system is used to implement the aforementioned big data-based infectious disease transmission path generation and early warning method. The system includes:

[0194] The data processing module acquires data on personnel entities, spatial locations, and environmental states. It then performs missing value removal, outlier replacement, and clock synchronization at the edge computing nodes to generate a standardized event stream. The data processing module (edge ​​sensing and preprocessing hardware) includes: a sensing layer where personnel entities are represented by Bluetooth and ultra-wideband tags, spatial locations by multiple cameras or millimeter-wave radar arrays, and environmental states by particle counters, carbon dioxide sensors, etc. Each sensing node has a built-in low-power microcontroller for local frame compression and encryption. The edge computing nodes utilize a heterogeneous single-board architecture of ARM and FPGA. The ARM side runs missing value removal and outlier replacement algorithms; the FPGA side is configured with a high-precision clock and performs hardware-level time synchronization across all nodes via IEEE 1588. Finally, the cleaned event data is cached in a high-speed circular buffer to generate a standardized event stream.

[0195] The topology fingerprint module is used to convert the standardized event stream input to the neuromorphic processor into a pulse event stream, construct an event graph and calculate persistent cohomology within a preset time window to obtain the topology fingerprint. The topology fingerprint module (neuromorphic processing hardware) includes: a neuromorphic processor: a chip supporting pulse coding (such as Intel Loihi). The standardized event stream is written to the on-chip SRAM via PCIe direct DMA, automatically completing pulse conversion. An auxiliary MCU: responsible for aggregating the pulse events output by the chip according to the time window and writing them to the on-chip DDR; subsequently, the RISC-V core on the SoC calls the parallel persistent cohomology IP core to calculate the topology fingerprint and store it in shared memory.

[0196] The propagation generation module is used to aggregate the pulse event stream according to the time window to form a variable-order hyperedge event graph, and to use the topological fingerprint to set the differentiable cellular automaton convolution kernel to evolve into an infection prediction field under the initial state of the environment field. Then, the variable-order hyperedge event graph, the infection prediction field, and the topological fingerprint are combined to form a conditional tensor, which is input into a conditional fractional diffusion neural network to generate a propagation event sequence. The propagation generation module (GPU-FPGA collaborative computing hardware) includes: a differentiable cellular automaton: implementing three-dimensional convolution and state updates in the FPGA logic area, with the topological fingerprint updating the convolution kernel in an AXI-Stream manner; a conditional fractional diffusion neural network: deployed on a single edge GPU with 16GB of video memory (such as an NVIDIA A6000); the FPGA sends the conditional tensor via RDMA, and the GPU performs forward diffusion and backward sampling to output the propagation event sequence; high-speed interconnect: PCIe Gen4×16 is used between the FPGA and GPU to meet second-level data round-trip requirements.

[0197] The risk intervention module is used to construct a tensor network based on the propagation event sequence, represent the tensor network as an Ising Hamiltonian to determine intervention nodes, issue intervention commands to environmental control equipment, and update the tensor network, the conditional fractional diffusion neural network, and the differentiable cellular automaton based on environmental feedback measured before and after intervention to form an adaptive closed loop. The risk intervention module (hybrid computing and execution hardware) includes: a tensor network and Ising mapping: on the FPGA-SoC side, matrix multiplication IP and hardware pseudo-random units are called to generate coupling weights; an optical delay loop reservoir consists of a 1km single-mode fiber, an acousto-optic modulator, and a photodetector, with the output energy gradient written back to the SoC buffer via a high-speed ADC. Optimization calculation: a quad-core Cortex-A72 on the SoC runs alternating linear least squares and gradient descent routines, outputting a list of intervention nodes within 32ms. Execution layer: the list is distributed to the PLCs of the ventilator, UV lamp, and mobile air purifier via an RS-485 bus; each device includes an ESP32 wireless module to transmit environmental feedback.

[0198] The closed-loop feedback chain includes: Feedback acquisition: A micro particle counter is embedded in the ventilation housing, and the data is sent back to the edge computing node via a LoRa gateway. Parameter update: The ARM core performs incremental write operations on the tensor network, GPU weight file, and FPGA convolution kernel registers, with a closed-loop refresh cycle of 60 seconds. The real-time power consumption of the entire chain is controlled within 80W.

[0199] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for generating and issuing early warning of infectious disease transmission paths based on big data, characterized in that, Includes the following steps: Acquire data on personnel entities, spatial locations, and environmental status, and perform missing value removal, outlier replacement, and clock synchronization at edge computing nodes to generate a standardized event stream; The standardized event stream is input into a neuromorphic processor and converted into a pulse event stream. An event graph is constructed and persistent cohomology is calculated within a preset time window to obtain a topological fingerprint. The pulse event stream is aggregated according to the time window to form a variable-order hyperedge event graph. The topological fingerprint is used to set a differentiable cellular automaton convolution kernel to evolve an infection prediction field in the initial state of the environment field. The variable-order hyperedge event graph, the infection prediction field and the topological fingerprint are combined into a conditional tensor and input into a conditional fractional diffusion neural network to generate a propagation event sequence. A tensor network is constructed based on the propagation event sequence. The tensor network is represented as the Ising Hamiltonian to determine the intervention node. Intervention commands are issued to the environmental control equipment. The tensor network, the conditional fractional diffusion neural network, and the differentiable cellular automaton are updated based on the environmental feedback measured before and after the intervention to form an adaptive closed loop.

2. The method according to claim 1, characterized in that, When generating a standardized event stream, interpolation is performed to complete missing records, neighborhood averaging is performed to replace abnormal records, and timestamps are uniformly corrected for all event sensing devices through a network time synchronization protocol.

3. The method according to claim 1, characterized in that, The pulse event flow constructs an event graph within a continuous time window. The vertices of the event graph correspond to personnel entities and spatial locations, and the edge weights correspond to the number of pulses within the time window. A topological fingerprint is generated by calculating a sustained long-term homology using the Vietoris–Rips method.

4. The method according to claim 3, characterized in that, A long-term coherence value is used to set the diffusion radius of the convolution kernel of the differentiable cellular automaton, so that the diffusion effect of the convolution kernel is enhanced when the loop structure increases, thereby generating an infection prediction field.

5. The method according to claim 4, characterized in that, The infection prediction field is generated in a differentiable cellular automaton with viral load and viral inactivation factor as state variables, and is updated multiple times in discrete time steps until the state change between two adjacent updates is lower than a preset threshold.

6. The method according to claim 1, characterized in that, The variable-order hyperedge event graph, infection prediction field, and topological fingerprint are aligned according to time index, entity index, and spatial coordinates before being composed into a conditional tensor, and then concatenated into a unified tensor in the depth dimension before being input into the conditional fractional diffusion neural network.

7. The method according to claim 6, characterized in that, The conditional fractional diffusion neural network employs a multi-scale encoding and decoding structure. Increasing random noise is injected during the forward diffusion, and the network weights are trained through fractional matching loss during the reverse process to generate a sequence of propagating events that meet the conditional tensor constraints.

8. The method according to claim 1, characterized in that, The tensor network constructed from the transmission event sequence calculates risk weights based on the number of people, the maximum viral load in the infection prediction field, the air exchange coefficient, and the strain coefficient. After mapping to the Ising Hamiltonian, it combines the energy gradient output by the optical delay loop reservoir system and uses alternating linear least squares and gradient descent to jointly determine the intervention nodes.

9. The method according to claim 8, characterized in that, The environmental feedback is the difference in viral load before and after the intervention. After normalization, the difference is used to update the tensor network, conditional fractional diffusion neural network, and differentiable cellular automaton in a synchronous manner to complete the adaptive closed loop.

10. A big data-based infectious disease transmission path generation and early warning system, used to implement the big data-based infectious disease transmission path generation and early warning method according to any one of claims 1-9, characterized in that, The system includes: The data processing module is used to acquire data on personnel entities, spatial locations, and environmental status, and to perform missing value removal, outlier replacement, and clock synchronization at the edge computing node to generate a standardized event stream. The topological fingerprint module is used to convert the standardized event stream into a pulse event stream by the neuromorphic processor, construct an event graph and calculate persistent cohomology within a preset time window to obtain the topological fingerprint. The propagation generation module is used to aggregate the pulse event stream according to the time window to form a variable-order hyperedge event graph, and use the topological fingerprint to set a differentiable cellular automaton convolution kernel to evolve into an infection prediction field in the initial state of the environmental field. Then, the variable-order hyperedge event graph, the infection prediction field and the topological fingerprint are combined into a conditional tensor and input into a conditional fractional diffusion neural network to generate a propagation event sequence. The risk intervention module is used to construct a tensor network based on the propagation event sequence, represent the tensor network as an Ising Hamiltonian to determine the intervention node, issue intervention commands to the environmental control equipment, and update the tensor network, the conditional fractional diffusion neural network, and the differentiable cellular automaton based on the environmental feedback measured before and after the intervention to form an adaptive closed loop.

Images