Infectious disease early warning and monitoring method and system based on big data and deep learning

Abstract

The invention provides an infectious disease early warning and monitoring method and system based on big data and deep learning, and relates to the technical field of infectious disease early warning, and the method comprises the steps: obtaining infectious disease historical data, crowd movement track data, medical treatment data and environment monitoring data as a training data set; establishing a space-time propagation link diagram and carrying out dynamic segmentation coding to generate a propagation characteristic sequence; training a deep learning model based on the propagation feature sequence to predict an infectious disease propagation inflection point and a high-risk area propagation probability; and when the propagation probability exceeds a preset threshold value, generating early warning information and pushing the early warning information to a monitoring terminal. According to the invention, accurate prediction and timely early warning of the transmission trend of infectious diseases can be realized, and the epidemic prevention and control efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of infectious disease early warning, and particularly to an infectious disease early warning and monitoring method and system based on big data and deep learning. Background Art

[0002] Infectious disease early warning and monitoring is an important part of public health prevention and control. With the development of big data technology, by integrating multi-source data such as historical infectious disease data, population movement trajectories, medical treatment records, and environmental monitoring, a spatio-temporal evolution model of infectious disease transmission can be constructed. Currently, infectious disease early warning mainly analyzes historical data based on statistical methods and issues early warnings by setting fixed thresholds. However, the transmission of infectious diseases has significant spatio-temporal heterogeneity, and its transmission pattern will change dynamically with time, geographical location, and environmental factors.

[0003] However, there are still several deficiencies in existing infectious disease early warning methods. There is a lack of in-depth exploration of the spatio-temporal dynamic characteristics of infectious disease transmission, and it is impossible to accurately depict the transmission links between different regions; the early warning model relies too much on fixed statistical thresholds and fails to fully consider the influence of dynamic factors such as population movement and environmental changes; the early warning results lack pertinence and cannot effectively identify high-risk regions and key transmission nodes; the self-adaptability of the model is poor and it is difficult to dynamically adjust the early warning strategy according to the development trend of the epidemic.

[0004] In summary, by constructing a deep learning model based on spatio-temporal transmission links, achieving accurate modeling of the laws of infectious disease transmission, effectively integrating multi-source heterogeneous data, constructing an accurate spatio-temporal network of infectious disease transmission; identifying and quantifying the key factors affecting infectious disease transmission; accurately predicting the inflection points and high-risk regions of infectious disease transmission based on deep learning methods; establishing a dynamic and adaptive early warning mechanism to improve the accuracy and timeliness of early warning. The present invention can solve the problems in the prior art. Summary of the Invention

[0005] An embodiment of the present invention provides an infectious disease early warning and monitoring method and system based on big data and deep learning, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiment of the present invention, There is provided an infectious disease early warning and monitoring method based on big data and deep learning, including: Obtaining historical infectious disease data, population movement trajectory data, medical treatment data, and environmental monitoring data as a training data set; Based on the training data set, mapping the historical infectious disease data to a geographical space grid, calculating the transmission intensity between grids according to the population movement trajectory data, marking the susceptible areas of infectious diseases based on the medical treatment data, and identifying the transmission acceleration factor in combination with the environmental monitoring data to establish a spatio-temporal transmission link diagram; Perform dynamic segmentation encoding on the spatio-temporal propagation link diagram, divide the continuous time series into multiple propagation cycles, calculate the propagation intensity between grids and the temporal variation of the susceptible areas of infectious diseases for each propagation cycle, and generate a propagation feature sequence; Train a deep learning model based on the propagation feature sequence, predict the inflection point moment of infectious disease propagation according to the change trend of the propagation acceleration factor and the propagation intensity between grids, and calculate the propagation probability of high-risk areas; When the propagation probability exceeds the preset threshold, mark the corresponding grid as a warning area, generate warning information including the warning area and the propagation inflection point, and push the warning information to the monitoring terminal through a preset communication interface.

[0007] In an alternative embodiment, Based on the training dataset, map the historical data of infectious diseases to the geographical space grid, calculate the propagation intensity between grids according to the population movement trajectory data, mark the susceptible areas of infectious diseases based on the medical visit data, and identify the propagation acceleration factor by combining the environmental monitoring data. The spatio-temporal propagation link diagram established includes: Based on the population density distribution data, use the quadtree structure to divide the geographical space into grids to obtain geographical space grids; Map the historical data of infectious diseases to the geographical space grid, construct the spatio-temporal data structure of infectious diseases, and calculate the disease density distribution of the geographical space grid based on the spatial kernel function and the temporal kernel function; Perform abnormal trajectory filtering and sampling frequency normalization on the population movement trajectory data to obtain standard population movement trajectory data; based on the standard population movement trajectory data, calculate the population flow between the geographical space grids, and combine the disease density distribution to calculate the propagation intensity between the geographical space grids; Extract the characteristics of the visiting population from the medical visit data, construct a susceptibility index, and combine the disease density distribution to calculate the susceptibility degree of infectious diseases in the geographical space grid; Extract environmental characteristic parameters from the environmental monitoring data, calculate the deviation between the environmental characteristic parameters and the corresponding optimal propagation values, and identify the propagation acceleration factor; Construct a spatio-temporal propagation link diagram of infectious diseases according to the propagation intensity, susceptibility degree of infectious diseases, and propagation acceleration factor between the geographical space grids.

[0008] In an alternative embodiment, Extract environmental characteristic parameters from the environmental monitoring data, calculate the deviation between the environmental characteristic parameters and the corresponding optimal propagation values, and identifying the propagation acceleration factor includes: Obtain environmental monitoring data of multiple observation points, and extract environmental characteristic parameters from the environmental monitoring data; perform statistics on the environmental characteristic parameters, use the kernel density estimation method to identify the optimal distribution interval, and introduce a time decay weight to perform weighted processing on historical data to obtain an initial propagation optimum value; Analyze the change trend and fluctuation range of the environmental characteristic parameters in different time periods to determine the historical value retention coefficient; determine the observation point weight coefficient according to the influence degree of each observation point during the propagation process; Use the historical value retention coefficient and the observation point weight coefficient to perform weighted combination on the environmental characteristic parameters and the initial propagation optimum value of each observation point to obtain a dynamic propagation optimum value; Calculate the deviation between the observed value of the environmental characteristic parameter and the dynamic propagation optimum value. When the observed value is greater than the dynamic propagation optimum value, use the first deviation weight coefficient, and when the observed value is less than the dynamic propagation optimum value, use the second deviation weight coefficient to obtain the characteristic parameter deviation value; Combine the characteristic parameter deviation values in pairs and multiply them by the interaction coefficient to generate an interaction influence matrix; combine the single characteristic parameter deviation value with the interaction influence matrix, and obtain the comprehensive environmental influence intensity through successive multiplication operations; Calculate the mean and standard deviation of the comprehensive environmental influence intensity, and determine the propagation acceleration threshold in combination with a preset adjustment coefficient; identify the propagation acceleration factor based on the comprehensive environmental influence intensity and the propagation acceleration threshold.

[0009] In an alternative embodiment, Perform dynamic segmentation coding on the spatio-temporal propagation link diagram, divide the continuous time series into multiple propagation cycles, calculate the propagation intensity between grids and the temporal variation of the infectious disease susceptible area for each propagation cycle, and generate a propagation feature sequence including: Obtain the propagation data of the spatio-temporal propagation link diagram, and calculate the propagation trend function based on the propagation data. The propagation trend function is the cumulative value of the propagation intensity between grids in the spatio-temporal propagation link diagram; Construct a trend mutation detection function, and obtain the mutation characteristics of the propagation trend according to the change rate and change direction of the propagation trend function within adjacent time windows; Calculate the dynamic threshold according to the local variance, mark the time points where the absolute value of the trend mutation detection function exceeds the dynamic threshold and the mutation direction changes as demarcation points, and divide the continuous time series of the spatio-temporal propagation link diagram into multiple propagation cycles; For each of the propagation cycles, combine the propagation intensity between grids and the distribution state of the infectious disease susceptible area to form a cycle feature matrix; Perform a temporal difference operation on the periodic feature matrix to obtain difference features, extract temporal evolution features through non-linear dynamic analysis and statistical entropy analysis, and perform adaptive weighted fusion based on the feature discrimination ability to form a temporal feature vector; Construct an encoding mapping function based on the temporal feature vectors of adjacent propagation cycles, perform weighted combination on the temporal feature vectors and perform temporal smoothing to construct an associated code; Perform feature optimization and compression transformation on the associated code to generate the propagation feature sequence of the spatio-temporal propagation link graph.

[0010] In an alternative embodiment, Performing a temporal difference operation on the periodic feature matrix to obtain difference features, extracting temporal evolution features through non-linear dynamic analysis and statistical entropy analysis, and performing adaptive weighted fusion based on the feature discrimination ability to form a temporal feature vector includes: Perform a temporal difference operation on the periodic feature matrix to obtain first-order difference features and second-order difference features; Perform wavelet decomposition on the periodic feature matrix to obtain time-frequency features, extract the energy distribution features, kurtosis features and skewness features of the time-frequency features, and construct a multi-scale feature set; Calculate the sample entropy of the periodic feature matrix to obtain complexity features, perform phase space reconstruction on the periodic feature matrix to obtain a reconstructed phase space, extract dynamic invariants from the reconstructed phase space, calculate the Lyapunov exponent based on the reconstructed phase space to obtain chaotic characteristics, and combine the complexity features, the dynamic invariants and the chaotic characteristics to form a non-linear feature set; Calculate the sliding entropy of the periodic feature matrix to obtain uncertainty features, construct a dynamic mutual information matrix to calculate information flow features, and combine the uncertainty features and the information flow features to generate an information theory feature set; Construct a feature importance evaluation function to calculate the feature discrimination ability, and determine the adaptive weight coefficient according to the feature discrimination ability; Perform weighted combination on the first-order difference features, the second-order difference features, the multi-scale feature set, the non-linear feature set and the information theory feature set according to the adaptive weight coefficient, and associate with the duration of the propagation cycle to generate an enhanced temporal feature vector; Perform a robustness evaluation on the enhanced temporal feature vector, and screen key features to form a temporal feature vector.

[0011] In an alternative embodiment, Train a deep learning model based on the propagation feature sequence, and predict the inflection point moment of infectious disease transmission according to the change trends of the propagation acceleration factor and the propagation intensity between grids. Calculating the propagation probability of high-risk areas includes: Map the position information of each time step in the propagation feature sequence to sine function values and cosine function values respectively, and combine them with the features of the corresponding time step to obtain propagation position encoding features; Map the propagation position encoding features to a propagation query matrix, a propagation key matrix, and a propagation value matrix, calculate the similarity between the propagation query matrix and the propagation key matrix to obtain propagation attention weights, multiply the propagation attention weights by the propagation value matrix to obtain multiple propagation attention head features, and splice and linearly transform the multiple propagation attention head features to obtain propagation global dependence features; By constructing multiple parallel dilated causal convolution branches and dynamically adjusting the dilation rate based on the propagation acceleration factor, obtain multiple groups of propagation local temporal features, adaptively fuse the multiple groups of propagation local temporal features, and perform residual connection and layer normalization processing with the propagation feature sequence to obtain propagation normalized features; Splice the propagation global dependence features and the propagation normalized features and perform a non-linear transformation to obtain propagation fusion features. After combining the propagation fusion features with the propagation acceleration factor, input them into a feed-forward neural network to obtain infectious disease propagation inflection point prediction features; Calculate the infectious disease propagation inflection point probability based on the infectious disease propagation inflection point prediction features, and perform a weighted combination of the infectious disease propagation inflection point probability and the inter-grid propagation intensity to obtain the propagation probability of high-risk areas.

[0012] In an alternative embodiment, By constructing multiple parallel dilated causal convolution branches and dynamically adjusting the dilation rate based on the propagation acceleration factor, obtain multiple groups of propagation local temporal features, adaptively fuse the multiple groups of propagation local temporal features, and perform residual connection and layer normalization processing with the propagation feature sequence to obtain propagation normalized features, including: Input the propagation feature sequence into multiple parallel convolution branches, each of the parallel convolution branches is set with a different initial dilation rate, the initial dilation rates are distributed in an exponentially increasing manner, and each of the parallel convolution branches includes multiple layers of dilated causal convolution layers; Obtain the change rate of the propagation acceleration factor at the current moment, calculate the dilation rate adjustment coefficient according to the magnitude of the change rate, and multiply the dilation rate adjustment coefficient by the initial dilation rate of each of the parallel convolution branches to obtain the dynamic dilation rate; Perform dilated causal convolution operations on the propagation feature sequence in each of the parallel convolution branches using the dynamic dilation rate to obtain multiple groups of propagation local temporal features; Calculate the correlation coefficient between each group of the propagation local temporal features and the propagation acceleration factor, and input the correlation coefficient into a normalization function to obtain the feature importance weight; Weight the multiple groups of propagated local temporal features based on the feature importance weights, and obtain the fused propagated local temporal features by passing the weighted features through an adaptive fusion network; Perform residual connection on the fused propagated local temporal features and the propagated feature sequence, and perform layer normalization on the result of the residual connection to obtain the propagated normalized features.

[0013] In the second aspect of the embodiments of the present invention, Provide an infectious disease early warning and monitoring system based on big data and deep learning, including: A first unit for obtaining infectious disease historical data, population movement trajectory data, medical visit data, and environmental monitoring data as a training data set; A second unit for mapping the infectious disease historical data to a geographical space grid based on the training data set, calculating the propagation intensity between grids according to the population movement trajectory data, marking the infectious disease susceptible areas based on the medical visit data, and identifying the propagation acceleration factors in combination with the environmental monitoring data to establish a spatio-temporal propagation link graph; A third unit for dynamically segmenting and encoding the spatio-temporal propagation link graph, dividing the continuous time series into multiple propagation cycles, calculating the temporal changes of the propagation intensity between grids and the infectious disease susceptible areas for each propagation cycle, and generating a propagated feature sequence; A fourth unit for training a deep learning model based on the propagated feature sequence, predicting the inflection point moment of infectious disease propagation according to the change trends of the propagation acceleration factors and the propagation intensity between grids, and calculating the propagation probability of high-risk areas; A fifth unit for, when the propagation probability exceeds a preset threshold, marking the corresponding grid as a warning area, generating a warning message including the warning area and the propagation inflection point, and pushing the warning message to a monitoring terminal through a preset communication interface.

[0014] In the third aspect of the embodiments of the present invention, Provide an electronic device, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0015] In the fourth aspect of the embodiments of the present invention, Provide a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0016] In the embodiments of the present invention, by obtaining multi-source data and establishing a spatio-temporal propagation link graph, accurate modeling of the infectious disease transmission process is achieved, the accuracy of infectious disease early warning is improved, and the key nodes and high-risk areas of infectious disease transmission can be discovered in a timely manner; by adopting a dynamic segmentation coding method, the infectious disease transmission process is divided into multiple transmission cycles for analysis, which can effectively capture the dynamic characteristics and evolution laws of infectious disease transmission, and improve the timeliness and adaptability of the infectious disease early warning model; based on a deep learning model, the characteristics of infectious disease transmission are analyzed and predicted. By automatically identifying the transmission acceleration factor and the change trend of the transmission intensity between grids, the inflection point moment of infectious disease transmission can be accurately predicted, providing a scientific decision-making basis for disease prevention and control, and having important practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a schematic flowchart of the method for infectious disease early warning and monitoring based on big data and deep learning according to the embodiments of the present invention; Figure 2 is a radar chart of the optimal temperature distribution of dynamic transmission under different weight methods for the observation point; Figure 3 is a dynamic bubble chart for environmental monitoring data and identification of transmission acceleration factors; Figure 4 is a radial chart for principal component contribution and robustness evaluation of time series feature vectors. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0020] Figure 1 is a schematic flowchart of the method for infectious disease early warning and monitoring based on big data and deep learning according to the embodiments of the present invention, as Figure 1 shown, the method includes: Obtain historical infectious disease data, population movement trajectory data, medical treatment data, and environmental monitoring data as the training data set; Based on the training dataset, map the historical data of infectious diseases to the geospatial grid, calculate the transmission intensity between grids according to the population movement trajectory data, mark the susceptible areas of infectious diseases based on the medical visit data, combine the environmental monitoring data to identify the transmission acceleration factors, and establish a spatio-temporal transmission link diagram; Perform dynamic segmentation encoding on the spatio-temporal transmission link diagram, divide the continuous time series into multiple transmission cycles, calculate the temporal changes of the transmission intensity between grids and the susceptible areas of infectious diseases for each transmission cycle, and generate a transmission feature sequence; Train a deep learning model based on the transmission feature sequence, predict the inflection point moment of infectious disease transmission according to the change trends of the transmission acceleration factors and the transmission intensity between grids, and calculate the transmission probability of high-risk areas; When the transmission probability exceeds the preset threshold, mark the corresponding grid as a warning area, generate warning information including the warning area and the transmission inflection point, and push the warning information to the monitoring terminal through a preset communication interface.

[0021] In an optional embodiment, based on the training dataset, mapping the historical data of infectious diseases to the geospatial grid, calculating the transmission intensity between grids according to the population movement trajectory data, marking the susceptible areas of infectious diseases based on the medical visit data, combining the environmental monitoring data to identify the transmission acceleration factors, and establishing a spatio-temporal transmission link diagram includes: Based on the population density distribution data, use a quadtree structure to divide the geospatial into grids to obtain geospatial grids; Map the historical data of infectious diseases to the geospatial grid, construct an infectious disease spatio-temporal data structure, and calculate the disease density distribution of the geospatial grid based on a spatial kernel function and a temporal kernel function; Perform abnormal trajectory filtering and sampling frequency normalization processing on the population movement trajectory data to obtain standard population movement trajectory data; based on the standard population movement trajectory data, calculate the population flow between the geospatial grids, and combine the disease density distribution to calculate the transmission intensity between the geospatial grids; Extract the characteristics of the visiting population from the medical visit data, construct a susceptibility index, and combine the disease density distribution to calculate the susceptibility degree of the geospatial grid to infectious diseases; Extract environmental characteristic parameters from the environmental monitoring data, calculate the deviation between the environmental characteristic parameters and the corresponding optimal transmission values, and identify the transmission acceleration factors; Construct an infectious disease spatio-temporal transmission link diagram according to the transmission intensity, susceptibility degree to infectious diseases, and transmission acceleration factors between the geospatial grids.

[0022] In a specific implementation, after obtaining the population density distribution data of the area to be analyzed, an adaptive quadtree structure is used for grid division. First, the entire area is regarded as a root grid, and density evaluation and division are carried out level by level. When the population density difference within a certain grid exceeds the preset threshold, the grid is divided into four equal sub-grids. This division process is looped until the population density distribution within all grids meets the uniformity requirement. A unique identification code is assigned to each completed grid, and its boundary coordinate information, hierarchical information and other spatial attributes are recorded to form a multi-level spatial grid system.

[0023] The location information in the historical data of infectious diseases is converted into geographical coordinates and mapped to the corresponding spatial grid. A spatio-temporal data structure containing grid identification, time information and the number of cases is established to characterize the spatio-temporal characteristics of disease distribution. The Gaussian kernel function is applied in the spatial dimension to calculate the influence degree of cases on the surrounding grids, and the influence intensity decays with the increase of distance. In the time dimension, an exponential decay function is adopted to impose a time weight on historical cases, and the weight gradually decreases over time. The spatial influence and time weight are combined to obtain the disease density distribution of each grid.

[0024] The population movement trajectory data is preprocessed by setting speed threshold and residence time threshold to identify and eliminate trajectory points that do not conform to normal movement characteristics. The sampling frequencies of trajectory data from different sources are unified, and missing trajectory points are supplemented by interpolation method to ensure the continuity and consistency of trajectory data. The standardized trajectory points are mapped into the grid space, and the population flow volume between grids is counted within the set time window. According to the disease density distribution of the grid, the risk coefficient of the infectious disease carried by the mobile population is calculated, and then the transmission intensity representing the possibility of transmission between grids is obtained.

[0025] Population characteristic information is extracted from medical visit data, including multiple dimensions such as age characteristics, past medical history, immune status, etc., to construct a susceptibility scoring system. This scoring system considers the influence weights of each characteristic on susceptibility and calculates the distribution characteristics of susceptible populations within the grid. The susceptibility score is combined with the disease density distribution of the grid to obtain a susceptibility degree index reflecting the transmission risk of infectious diseases, which is used to identify high-risk areas.

[0026] When processing environmental monitoring data, environmental characteristic parameters including temperature, humidity, air quality, etc. are extracted. According to the transmission characteristics of infectious diseases, the optimal interval values of each environmental parameter are determined. The deviation between the actual environmental parameter and the optimal value is calculated to evaluate the influence degree of environmental factors on the transmission rate. The environmental factors that have a significant impact on the transmission rate are identified as transmission acceleration factors, which are used to adjust the calculation of transmission risk.

[0027] Finally, a spatio-temporal propagation link graph is constructed based on the propagation intensity between grids, the susceptibility of grids, and the propagation acceleration factor. The graph structure uses grids as nodes and propagation intensity as edges. The nodes contain susceptibility attributes, and the weights of the edges are adjusted by the propagation acceleration factor. This structure can dynamically display the propagation path and risk distribution of infectious diseases, providing a basis for subsequent propagation analysis and prediction.

[0028] Through the above implementation steps, the refined analysis and dynamic monitoring of the propagation characteristics of infectious diseases are realized, providing data support for the prevention and control of infectious diseases. This method can adapt to different scales and types of infectious disease propagation scenarios, with strong generality and scalability.

[0029] Exemplarily, in the initial grid division stage, after obtaining the population density data of the study area, the density difference threshold is set at 500 people per square kilometer. Starting from a 100km×100km root grid, when the density difference in densely populated areas such as commercial areas is detected to exceed the threshold, it is divided into four 50km×50km sub-grids. For sub-grids with still large population density differences, further division is continued. Finally, fine grids with a side length of 500m are formed in densely populated areas, and larger grid sizes are maintained in sparsely populated areas. Each grid is assigned a unique code, such as "A1-2-3" representing the third grid in the second layer of the first quadrant.

[0030] In the disease data mapping stage, the infectious disease case data for the most recent 90 days is collected. For example, in the grid numbered "A1-2-3", 5 cases were recorded on the 30th day and 3 cases on the 45th day. The Gaussian kernel function is used to calculate the spatial influence of cases, and the decay radius is set at 2km, which means that one case will have an impact on the grids within a range of 2km, but the impact intensity decreases with the increase in distance. At the same time, a time decay function is applied, and the decay period is set at 14 days, that is, the weight of historical cases more than 14 days old is reduced to half of the original.

[0031] In the processing of population movement trajectories, the abnormal speed threshold is set at 120km / h and the residence time threshold is set at 30 seconds. For example, in a certain trajectory, a recording point with an instantaneous speed reaching 150km / h is judged as abnormal and excluded. The sampling interval of all trajectory data is uniformly adjusted to 5 minutes, and linear interpolation is used to supplement the trajectory points with an interval greater than 5 minutes. Statistics show that within the time window from 8:00 to 9:00, the population flow volume from grid "A1-2-3" to the adjacent grid "A1-2-4" is 200 person-times.

[0032] In the susceptibility assessment stage, a scoring system is constructed, which includes age (weight 0.4), underlying diseases (weight 0.3), and immune status (weight 0.3). For example, in grid "A1-2-3", the proportion of people over 60 years old is 30%, the proportion of patients with chronic diseases is 20%, and the vaccination rate is 70%. Based on these data, the susceptibility score of this grid is calculated as 0.65 (with a full score of 1). Combining with the disease density index of 0.8 for this grid, the susceptibility level of infectious diseases in this grid is finally determined as the "high-risk" level.

[0033] In the analysis of environmental factors, the optimal temperature range for the spread of infectious diseases is 20 - 25°C, and the relative humidity is 60 - 70%. Monitoring data shows that the actual temperature in grid "A1-2-3" is 23°C, and the relative humidity is 75%. The temperature suitability is calculated as 0.9, and the humidity suitability is 0.8. Since the actual parameters are close to the optimal values, these environmental factors are identified as potential factors accelerating the spread.

[0034] Finally, in the transmission link diagram, grid "A1-2-3" is used as a node, and its susceptibility attribute is marked as "high-risk". The edge weight (transmission intensity) between it and the adjacent grid "A1-2-4" is 0.75. Considering the influence of environmental acceleration factors such as temperature and humidity, the final risk coefficient of this transmission path is adjusted to 0.85, indicating that this is a transmission path that requires key attention.

[0035] This example demonstrates the operation process of the technical solution in the actual scenario, reflecting the operability of the method. The specific parameters and thresholds in the example can be adjusted according to actual needs to ensure the flexibility and adaptability of the method.

[0036] In this embodiment, the dynamic granularity adjustment of the spatial analysis unit is realized. More refined grid division is adopted in densely populated areas, and larger grid scales are maintained in sparsely populated areas, which not only ensures the analysis accuracy but also improves the calculation efficiency; the spatio-temporal kernel function combination method is used to calculate the disease density distribution, which not only considers the spatial aggregation characteristics of cases but also introduces time decay weights, and can more accurately describe the spatio-temporal transmission law of infectious diseases, avoiding the limitations of traditional single-dimensional analysis methods; through the standardized processing of population movement trajectory data and the joint analysis of disease density, a calculation model for transmission intensity based on actual population movement is established. Compared with the simple geographical distance assessment method, it is more in line with the actual characteristics of the spread of infectious diseases through human contact; the susceptibility assessment and environmental factor analysis are integrated into the transmission link diagram to construct a multi-dimensional infectious disease transmission risk assessment system, which can dynamically reflect the changes in infectious risks in different regions and provide data support for precise prevention and control.

[0037] In an alternative embodiment, environmental characteristic parameters are extracted from the environmental monitoring data, the deviation between the environmental characteristic parameters and the corresponding optimal propagation values is calculated, and the identification of the propagation acceleration factor includes: Obtain the environmental monitoring data of multiple observation points, and extract environmental characteristic parameters from the environmental monitoring data; perform statistics on the environmental characteristic parameters, use the kernel density estimation method to identify the optimal distribution interval, and introduce a time decay weight to weight the historical data to obtain the initial optimal propagation value; Analyze the change trend and fluctuation range of the environmental characteristic parameters in different time periods to determine the historical value retention coefficient; determine the observation point weight coefficient according to the influence degree of each observation point in the propagation process; Use the historical value retention coefficient and the observation point weight coefficient to perform weighted combination on the environmental characteristic parameters and the initial optimal propagation value of each observation point to obtain the dynamic optimal propagation value; Calculate the deviation between the observed value of the environmental characteristic parameter and the dynamic optimal propagation value, use the first deviation weight coefficient when the observed value is greater than the dynamic optimal propagation value, and use the second deviation weight coefficient when the observed value is less than the dynamic optimal propagation value to obtain the characteristic parameter deviation value; Combine the characteristic parameter deviation values in pairs and multiply by the interaction coefficient to generate an interaction influence matrix; combine the single characteristic parameter deviation value with the interaction influence matrix, and obtain the comprehensive environmental influence intensity through consecutive multiplication operations; Calculate the mean and standard deviation of the comprehensive environmental influence intensity, and determine the propagation acceleration threshold in combination with a preset adjustment coefficient; identify the propagation acceleration factor based on the comprehensive environmental influence intensity and the propagation acceleration threshold.

[0038] In a specific implementation manner, obtain the environmental monitoring data of multiple observation points from the monitoring system, mainly including environmental characteristic parameters such as temperature, humidity, air pressure, wind speed, and air quality. Conduct statistical analysis on the historical data of each characteristic parameter, and use the kernel density estimation method to identify the dense interval of the data distribution. On this basis, introduce a time decay weight, assign a higher weight to the data in the relatively recent period, and assign a lower weight to the data in the relatively distant period, and obtain the initial optimal propagation value of each environmental characteristic parameter through weighted calculation.

[0039] Analyze the time series data of the environmental characteristic parameters, and calculate the change trend and fluctuation degree at different time scales. Determine the historical value retention coefficient according to the stability of the parameter, and the smaller the fluctuation of the parameter, the larger the historical value retention coefficient. At the same time, determine the weight coefficient of the observation point according to the correlation degree of each observation point in the historical propagation event, and the higher the correlation degree of the observation point, the greater the weight obtained.

[0040] Combine the historical value retention coefficient with the observation point weight coefficient to weight the environmental characteristic parameters of each observation point. Through this weighted combination method, adjust the initial propagation optimum value to a dynamic propagation optimum value considering spatio-temporal variations, enabling it to adapt to the dynamic changes in environmental conditions.

[0041] Calculate the difference between the observed value of the environmental characteristic parameter and the dynamic propagation optimum value in real time. When the observed value exceeds the optimum value, calculate using the first deviation weight coefficient to reflect the degree of influence beyond the optimum value; when the observed value is lower than the optimum value, calculate using the second deviation weight coefficient to reflect the degree of influence insufficient to reach the optimum value, thereby obtaining the deviation values of each characteristic parameter.

[0042] Pairwise combine the deviation values of different characteristic parameters and multiply them by a pre-set interaction coefficient to form a matrix representing the interaction effects of environmental factors. Combine the deviation value of a single characteristic parameter with the interaction influence matrix and obtain the comprehensive influence intensity reflecting the overall environmental influence through successive multiplication operations.

[0043] Calculate the mean and standard deviation of the comprehensive environmental influence intensity over a period of time, and combine with a pre-set adjustment coefficient to determine the threshold for identifying the propagation acceleration factor. When the comprehensive influence intensity caused by a certain combination of environmental characteristic parameters exceeds this threshold, identify this group of environmental factors as the propagation acceleration factor.

[0044] Exemplarily, in the analysis of the spread of a certain infectious disease, environmental monitoring data of 5 key observation points were selected. Through kernel density estimation analysis, it was found that the temperature parameter was most densely distributed in the range of 20 - 25°C. Using a 14-day time decay period, the weight of the data in the most recent 24 hours was 1, and the weight of the data 14 days ago decreased to 0.5. Thus, the initial propagation optimum value of the temperature parameter was calculated to be 23°C.

[0045] Analyze the 30-day change data of the temperature parameter and find that its daily fluctuation range is stably within 3°C, and determine its historical value retention coefficient to be 0.8. Observation point 1 had the most associated cases in historical transmission events, so its weight coefficient was determined to be 0.3, while observation point 5 had the fewest associated cases, and the weight coefficient was set to 0.1.

[0046] Through the weighted combination of the historical value retention coefficient and the observation point weight, adjust the initial propagation optimum value of the temperature, 23°C, to the dynamic propagation optimum value of 22.5°C. When the real-time temperature of 26°C is recorded at observation point 1, calculate the temperature deviation value using the first deviation weight coefficient of 0.6; when the temperature at observation point 2 is 19°C, calculate the deviation value using the second deviation weight coefficient of 0.4.

[0047] The interaction coefficient between temperature and relative humidity is set to 0.7. When the deviation values of the two are 0.3 and 0.4 respectively, the interaction effect value is 0.084. Combining the individual deviation values and interaction effects of all environmental factors, the comprehensive environmental impact intensity at the current moment is calculated to be 0.15.

[0048] After 30 consecutive days of monitoring, the mean value of the comprehensive environmental impact intensity is 0.1, and the standard deviation is 0.03. Using the adjustment coefficient 1.5, the propagation acceleration threshold is calculated to be 0.145. Since the current comprehensive impact intensity of 0.15 exceeds the threshold, this combination of environmental parameters is identified as a propagation acceleration factor, indicating that the current environmental conditions may accelerate the spread of infectious diseases.

[0049] The Figure 2 shows the influence of different deviation calculation methods on the deviation values of characteristic parameters. This technical solution uses a differential deviation weight coefficient. When the observed temperature is lower than the optimal temperature for transmission (left side in the figure), the second deviation weight coefficient of 0.4 is used. When the observed temperature is higher than the optimal temperature for transmission (right side in the figure), the first deviation weight coefficient of 0.6 is used. This asymmetric treatment reflects the different degrees of influence of temperature above and below the optimal value on transmission. When the temperature is 2°C lower, the calculated deviation value of the characteristic parameter is 0.20, while when the temperature exceeds the optimal value by 5°C, the deviation value reaches 0.80, showing that the accelerating effect of too high temperature on transmission is more significant. In contrast, the linear deviation method shows a symmetric linear relationship and cannot reflect the difference in the influence of high and low temperatures. Although the square deviation method has non-linear characteristics, the gradient is close to zero at the zero point, and the sensitivity to small deviations is insufficient. Experimental data show that the deviation change of this technical solution is smoother within the range of ±3°C, and the deviation value increases rapidly after exceeding this range, which is more in line with the actual transmission law. This differential treatment provides more accurate basic data for the subsequent analysis of the interaction effects of characteristic parameters.

[0050] In Figure 3The relationship between two key environmental parameters, temperature and humidity, and the identification effect of the transmission acceleration factor is intuitively shown in the dynamic bubble chart. The horizontal axis represents relative humidity (40%-90%), and the vertical axis represents temperature (15°C-30°C). The dashed rectangular area (21-24°C, 60%-70%) in the figure is the optimal transmission range. The size of the bubbles formed by 10 observation points represents the comprehensive environmental impact intensity, and the depth of gray represents the recognition accuracy rate. As the environmental parameters deviate from the optimal range, especially towards high temperature and high humidity, the bubble size and gray level gradually increase, indicating that the comprehensive environmental impact intensity and the recognition accuracy rate increase synchronously. The maximum impact intensity of 0.32 and the highest recognition accuracy rate of 87% are reached at point 6 (humidity 80%, temperature 29°C). The thick dashed line represents the distribution trend of historical observation data, showing a high correlation between the high temperature and high humidity combination and the transmission acceleration factor. The system can complete the identification and issue an early warning 112 hours before the environmental parameters enter the acceleration region. The entire evolution process clearly demonstrates how this technical solution realizes the early and accurate identification of the transmission acceleration factor by comprehensively analyzing the change trend and deviation value of environmental characteristic parameters, providing sufficient time for the deployment of prevention and control measures.

[0051] In this embodiment, by introducing a method combining time decay weight and kernel density estimation to calculate the optimal transmission value, not only the distribution characteristics of environmental parameters are considered, but also the importance of recent data is reflected, making the determination of the optimal transmission value more in line with the actual transmission law and improving the accuracy of environmental impact assessment; the historical value retention coefficient and the observation point weight coefficient are dynamically adjusted, enabling the optimal transmission value to be adaptively updated with changes in environmental conditions, overcoming the limitation of the traditional fixed threshold method that cannot adapt to dynamic environmental changes; a two-way evaluation mechanism based on different deviation weight coefficients is designed, considering the differential effects of environmental parameters above and below the optimal value respectively, and capturing the synergistic effects between multiple environmental factors through an interaction impact matrix, realizing the refined quantification of environmental impact; the transmission acceleration threshold is determined based on the statistical characteristics of the comprehensive environmental impact intensity, and a data-driven method for identifying the transmission acceleration factor is established, which can timely detect and warn of adverse environmental condition combinations, providing decision-making support for infectious disease prevention and control.

[0052] In an alternative embodiment, the spatio-temporal transmission link diagram is dynamically segmented and encoded, the continuous time series is divided into multiple transmission cycles, and the inter-grid transmission intensity and the temporal variation of the infectious disease susceptible area are calculated for each transmission cycle, generating a transmission feature sequence including: Obtain the transmission data of the spatio-temporal transmission link diagram, and calculate a transmission trend function based on the transmission data, where the transmission trend function is the cumulative value of the inter-grid transmission intensity in the spatio-temporal transmission link diagram; Construct a trend mutation detection function, and obtain the mutation characteristics of the transmission trend according to the change rate and change direction of the transmission trend function within adjacent time windows; Calculate the dynamic threshold according to the local variance, mark the time points at which the absolute value of the trend mutation detection function exceeds the dynamic threshold and the mutation direction changes as demarcation points, and divide the continuous time series of the spatio-temporal propagation link graph into multiple propagation cycles; For each of the propagation cycles, combine the inter-grid propagation intensity and the distribution state of the infectious disease susceptible areas to form a periodic feature matrix; Perform a time series difference operation on the periodic feature matrix to obtain difference features, extract time series evolution features through non-linear dynamic analysis and statistical entropy analysis, and perform adaptive weighted fusion based on the feature discrimination ability to form a time series feature vector; Construct an encoding mapping function based on the time series feature vectors of adjacent propagation cycles, perform weighted combination on the time series feature vectors and perform time series smoothing to construct an associated encoding; Perform feature optimization and compression transformation on the associated encoding to generate the propagation feature sequence of the spatio-temporal propagation link graph.

[0053] In a specific implementation, extract the time series data of the inter-grid propagation intensity from the spatio-temporal propagation link graph. For each time point, accumulate all the inter-grid propagation intensity values to form a trend function reflecting the overall propagation trend. This accumulation process takes into account the contributions of all active propagation paths.

[0054] Set a sliding time window, calculate the change rate and change direction of the trend function within adjacent windows. By comparing the change characteristics of different windows, identify the mutation points of the propagation trend. These mutation characteristics include rapid growth, rapid decline of the propagation intensity, or significant changes in the growth rate.

[0055] Calculate the dynamic threshold based on local data, which is adaptively adjusted according to the volatility of the time series data. When the absolute value of the trend mutation detection function exceeds the dynamic threshold and the change direction is reversed, mark this time point as the demarcation point of the propagation cycle. Thus, the continuous time series is segmented into multiple propagation cycles with significant feature differences.

[0056] Within each propagation cycle, combine the inter-grid propagation intensity data with the distribution state of the infectious disease susceptible areas to construct a matrix reflecting the propagation characteristics of this cycle. This matrix contains both the spatial distribution information of the propagation intensity and the distribution pattern of the susceptible areas.

[0057] Perform time series difference on the periodic feature matrix to obtain the change information of the propagation features. At the same time, perform non-linear dynamic analysis to extract the evolution law of the propagation process; through statistical entropy analysis, quantify the uncertainty of the propagation situation. Determine the feature weights according to the performance of different features in distinguishing the propagation states, and fuse multiple features to form a time series feature vector.

[0058] Analyze the relationship between the eigenvectors of adjacent propagation cycles and establish the feature mapping rules. Weightedly combine the eigenvectors according to the mapping rules, and eliminate the influence of noise through temporal smoothing to generate the correlation coding reflecting the propagation evolution law.

[0059] Optimize and compress the correlation coding, remove redundant information, extract the most representative feature combinations, and finally generate the feature sequence describing the propagation process.

[0060] Exemplarily, in the analysis of the spread of a certain infectious disease, 30 days of transmission data were recorded. Through cumulative calculation, it was found that the average daily transmission intensity from the 1st to the 10th day was 0.3, rose to 0.7 from the 11th to the 20th day, and dropped to 0.4 from the 21st to the 30th day.

[0061] Adopt a 5-day sliding window to analyze the change of the propagation trend, and it is found that the propagation intensity rises rapidly between the 10th and 11th days, with a change rate reaching 133%; it drops rapidly between the 20th and 21st days, with a change rate of -43%.

[0062] According to the standard deviation of the local 10-day data, the dynamic threshold is calculated to be 40%. Since the absolute values of the change rates on the 10th - 11th day and the 20th - 21st day exceed the threshold and are in opposite directions, the 11th day and the 21st day are marked as the cycle demarcation points, forming three propagation cycles.

[0063] Within the first cycle (1 - 10 days), construct a feature matrix containing the propagation intensity of 300 grids and the distribution status of 50 susceptible areas. Calculate the temporal difference of this matrix, and it is found that the average increase in propagation intensity is 0.02 per day, and the expansion rate of the susceptible area is 2 per day.

[0064] Through non-linear analysis, it is found that the propagation within the cycle shows a diffusion feature, and entropy analysis shows an increase in uncertainty. According to the discriminant ability of the features, weights of 0.4, 0.3, and 0.3 are assigned respectively, and they are fused to form the eigenvector of this cycle.

[0065] Analyze the relationship between the eigenvectors of adjacent cycles, and establish a mapping function reflecting the evolution of the propagation situation. Weightedly combine and smooth the eigenvectors to generate the correlation coding with temporal continuity.

[0066] Finally, compress and optimize the correlation coding, extract the most representative feature combinations, and generate the feature sequence reflecting the evolution law of the entire propagation process.

[0067] In this embodiment, the cycle division is performed by combining the propagation trend function and the dynamic threshold, realizing the adaptive segmentation of the propagation process, avoiding the feature loss that may be caused by the fixed cycle division, and improving the accuracy of subsequent analysis; the grid propagation intensity and the susceptible area distribution are characterized in a matrix form, and through multi-dimensional feature extraction and adaptive weight fusion, a comprehensive characterization of the propagation features is realized, overcoming the limitation that a single feature cannot fully express the propagation law; a coding mapping method based on adjacent cycle association is adopted to ensure the continuity of the feature sequence during cycle conversion, effectively avoiding the information break problem at the cycle junction of the traditional segmentation method; through feature optimization and compression transformation, the data redundancy is significantly reduced, the calculation efficiency is improved, and the key features of the propagation process are retained, providing high-quality input data for subsequent propagation prediction.

[0068] In an alternative embodiment, a time series difference operation is performed on the cycle feature matrix to obtain difference features, and time series evolution features are extracted through nonlinear dynamic analysis and statistical entropy analysis, and adaptive weighted fusion is performed based on the feature discrimination ability to form a time series feature vector, including: Perform a time series difference operation on the cycle feature matrix to obtain first-order difference features and second-order difference features; Perform wavelet decomposition on the cycle feature matrix to obtain time-frequency features, extract the energy distribution feature, kurtosis feature and skewness feature of the time-frequency features, and construct a multi-scale feature set; Calculate the sample entropy of the cycle feature matrix to obtain complexity features, perform phase space reconstruction on the cycle feature matrix to obtain a reconstructed phase space, extract dynamic invariants from the reconstructed phase space, calculate the Lyapunov exponent based on the reconstructed phase space to obtain chaotic characteristics, and combine the complexity features, the dynamic invariants and the chaotic characteristics to form a non-linear feature set; Calculate the sliding entropy of the cycle feature matrix to obtain uncertainty features, construct a dynamic mutual information matrix to calculate information flow features, and combine the uncertainty features and the information flow features to generate an information theory feature set; Construct a feature importance evaluation function to calculate the feature discrimination ability, and determine the adaptive weight coefficient according to the feature discrimination ability; Perform weighted combination of the first-order difference features, the second-order difference features, the multi-scale feature set, the non-linear feature set and the information theory feature set according to the adaptive weight coefficient, and associate with the duration of the propagation cycle to generate an enhanced time series feature vector; Perform a robustness evaluation on the enhanced time series feature vector, and screen key features to form a time series feature vector.

[0069] In a specific embodiment, the periodic feature matrix is first subjected to temporal difference processing. The differences between adjacent time points are calculated to obtain the first-order difference features, which reflect the change rate of the propagation features; then the first-order difference results are subjected to difference operations to obtain the second-order difference features, which reflect the acceleration characteristics of the propagation features. The difference operation adopts the central difference method, and unilateral difference processing is used for the sequence boundaries. By setting a sliding window of 10 time steps, the difference statistical features within the window are calculated, including the mean, standard deviation, maximum value, and minimum value.

[0070] When performing wavelet decomposition, the discrete wavelet transform under the multi-resolution analysis framework is selected, and the Daubechies 5 wavelet basis function is used to decompose the periodic feature matrix into 4 layers. Time-frequency features are extracted at each decomposition level: the energy distribution of the wavelet coefficients is calculated to obtain the energy proportion of different frequency bands; the kurtosis of the wavelet coefficients is calculated to reflect the sharpness of the data distribution; the skewness of the wavelet coefficients is calculated to characterize the asymmetry of the data distribution. The reconstruction error is calculated for the decomposition results of each layer to evaluate the decomposition quality. The features of all levels are combined to form a multi-scale feature set.

[0071] In the non-linear feature extraction step, the sample entropy of the periodic feature matrix is first calculated. The embedding dimension is set to 3, and the similarity tolerance is 0.2 times the standard deviation. The sample entropy values at different scales are calculated. Then phase space reconstruction is performed. The optimal delay time is determined by the time delay method, and the minimum mutual information criterion is used. The typical delay time is set to 4 - 6 time steps. The Cao method is used to determine the embedding dimension, which is usually between 5 - 8 dimensions. Dynamic invariants such as the correlation dimension and the largest Lyapunov exponent are extracted from the reconstructed phase space to characterize the chaotic characteristics of the system. These non-linear indicators are combined into a non-linear feature set.

[0072] For the extraction of information theory features, a sliding window mechanism is adopted. The window size is set to 1 / 4 of the propagation period length, and the step size is 1. The Shannon entropy is calculated within each window to obtain the sequence of entropy values changing with time, which reflects the uncertainty evolution of the propagation process. When constructing the dynamic mutual information matrix, the mutual information between grids at different time delays is calculated, and the delay range is from 1 to 1 / 3 of the propagation period length. By performing eigenvalue decomposition on the mutual information matrix, the main information flow patterns are extracted. The entropy change features and the information flow features are combined into an information theory feature set.

[0073] For feature importance evaluation, a feature selection method based on random forest is adopted. A random forest model containing 100 decision trees is constructed, and the Gini index is used to evaluate the classification contribution of features. The importance score is calculated for each feature, and the score is converted into a weight coefficient through the softmax function. To improve the stability of the evaluation, 10-fold cross-validation is adopted, and the average importance score is taken as the final result.

[0074] In the feature fusion stage, first, various types of features are standardized so that their mean is 0 and the standard deviation is 1. Different feature sets are weighted and combined according to the feature importance weights. Considering the duration information of the propagation cycle, the cycle length is converted into one-hot encoding and concatenated with the weighted features. Principal component analysis is used for dimensionality reduction, retaining the principal components that explain 95% of the variance, and an enhanced time-series feature vector is generated.

[0075] Finally, robustness evaluation is carried out using multiple methods: adding Gaussian noise to test the noise sensitivity of the features; randomly masking some features to test the redundancy of the features; using different data subsets to evaluate the stability of the features. According to the evaluation results, the feature dimensions with robustness scores in the top 70% are selected to form the final time-series feature vector. During the feature screening process, ensure that the retained feature set can balance the expression of various aspects of differential features, multi-scale features, non-linear features, and information-theoretic features.

[0076] In the entire feature processing process, the selection of key parameters uses the grid search method to determine the optimal values, and the evaluation metrics include the discriminability, stability, and computational efficiency of the features. At the same time, a feature quality monitoring mechanism is established. When the feature quality deteriorates, the relevant parameters are automatically adjusted to maintain the expression ability of the features. This multi-level and multi-angle feature extraction and optimization scheme ensures a comprehensive and stable feature expression of the propagation process.

[0077] Exemplarily, analyze the feature matrix within the propagation cycle of a certain infectious disease, including 10-day propagation data of 100 grid nodes. Through time-series difference calculation, it is found that the mean of the first-order difference of the propagation intensity is 0.05, indicating a 5% daily increase; the mean of the second-order difference is 0.01, indicating that the growth rate is gradually accelerating.

[0078] Perform 4-layer wavelet decomposition on the feature matrix. It is found that the energy proportion of the low-frequency part is 70%, reflecting the overall trend of the propagation; the kurtosis of the high-frequency part is 3.5 and the skewness is 0.8, reflecting significant local fluctuations.

[0079] The calculated sample entropy value is 2.3, indicating that the propagation process has a high complexity. In the reconstructed three-dimensional phase space, a dynamic invariant with a correlation dimension of 2.1 is extracted. The Lyapunov exponent is 0.15, indicating that the system has weak chaotic characteristics.

[0080] Use a 5-day sliding window to calculate the change in entropy value. It is found that the entropy value reaches a peak of 1.8 on the 7th day. Through dynamic mutual information analysis, it is found that the information transfer delay between adjacent grids is about 1 - 2 days.

[0081] Feature importance evaluation shows that the discriminant abilities of the first-order difference feature, energy distribution feature, and entropy value change feature are the strongest, and weights of 0.3, 0.25, and 0.25 are respectively assigned.

[0082] Combine all features according to their weights, and considering the duration feature that this propagation cycle lasts for 10 days, generate an enhanced feature vector containing 23 dimensions.

[0083] After robustness testing, 15 feature dimensions with high stability are selected to form the final time series feature vector.

[0084] In Figure 4 it shows the contribution degree of the principal components of the time series feature vector and the results of robustness evaluation. The 8 axes in the figure represent different types of features: first-order difference features, second-order difference features, energy distribution features, kurtosis and skewness features, sample entropy features, Lyapunov exponents, sliding entropy features, and information flow features. The numerical values represent the feature importance scores (0 - 1). This technical solution (solid line square) performs excellently in each dimension. The importance of the first-order difference feature is the highest (0.82), followed by the second-order difference feature (0.78), the sample entropy feature (0.76), and the energy distribution feature (0.75). The other features also maintain high values between 0.68 - 0.74, forming a balanced and extended polygon, and the overall robustness score reaches 0.87. In contrast, the traditional statistical features (dashed line circle) have certain performances (0.4 - 0.6) in three dimensions of the first-order difference, sample entropy, and energy distribution, but the performances in other dimensions are insufficient, and the overall robustness score is only 0.53. The simple spectral features (dash-dotted line triangle) generally have low feature importance in each dimension (0.2 - 0.4), and the overall robustness score is the lowest, which is 0.41. Through principal component analysis and robustness evaluation, this technical solution optimizes the original 23-dimensional feature vector into 15 dimensions, retains a 95% variance explanation rate, and ensures the comprehensiveness and stability of feature expression while significantly reducing the feature dimension.

[0085] In this embodiment, through a multi-level feature extraction strategy, including the comprehensive analysis of difference features, multi-scale features, non-linear features, and information theory features, a full-range feature description of the propagation process is realized, significantly improving the feature expression ability; introducing a feature importance evaluation and adaptive weight mechanism to achieve the optimal combination of features and avoid the feature redundancy problem that may be brought by the traditional equal-weight method; integrating the duration information of the propagation cycle into the feature vector, enriching the time dimension expression of the time series features, and improving the description accuracy of the feature sequence for the propagation process; through robustness evaluation and key feature screening, the stability and reliability of the feature vector are ensured, providing high-quality feature input for subsequent propagation prediction.

[0086] In an alternative embodiment, based on the propagation feature sequence, train a deep learning model, and according to the change trends of the propagation acceleration factor and the propagation intensity between grids, predict the inflection point moment of the infectious disease propagation. Calculating the propagation probability of high-risk areas includes: Map the position information at each time step in the propagation feature sequence to sine function values and cosine function values respectively, and combine them with the features at the corresponding time steps to obtain propagation position encoding features; Map the propagation position encoding features to a propagation query matrix, a propagation key matrix, and a propagation value matrix, calculate the similarity between the propagation query matrix and the propagation key matrix to obtain propagation attention weights, multiply the propagation attention weights by the propagation value matrix to obtain multiple propagation attention head features, and splice and linearly transform the multiple propagation attention head features to obtain propagation global dependence features; By constructing multiple parallel dilated causal convolution branches and dynamically adjusting the dilation rate based on the propagation acceleration factor, obtain multiple groups of propagation local temporal features, adaptively fuse the multiple groups of propagation local temporal features, and perform residual connection and layer normalization processing with the propagation feature sequence to obtain propagation normalized features; Splice the propagation global dependence features and the propagation normalized features and perform a non-linear transformation to obtain propagation fusion features. After combining the propagation fusion features with the propagation acceleration factor, input them into a feed-forward neural network to obtain infectious disease transmission inflection point prediction features; Calculate the infectious disease transmission inflection point probability based on the infectious disease transmission inflection point prediction features, and perform a weighted combination of the infectious disease transmission inflection point probability and the inter-grid propagation intensity to obtain the propagation probability of high-risk areas.

[0087] In a specific implementation, perform encoding conversion on the position information at each time step in the propagation feature sequence, and map the position information to periodic values through sine and cosine functions respectively. Combine the converted position encoding values with the original features at this time step to generate encoding features containing position information.

[0088] Generate a query matrix, a key matrix, and a value matrix respectively by performing three independent linear transformations on the propagation position encoding features. Calculate the dot product of the query matrix and the key matrix to obtain attention scores, and obtain attention weights after normalization processing. Multiply the attention weights by the value matrix to obtain multiple attention head features. Splice these features and perform a transformation through a linear layer to obtain features that capture global dependence relationships.

[0089] Construct multiple parallel dilated causal convolution branches, each branch using a different initial dilation rate. Dynamically adjust the dilation rate of each branch according to the intensity of the propagation acceleration factor so that the receptive field can adaptively cover different time ranges. Perform weighted fusion on the local temporal features extracted from multiple branches, and perform residual connection and layer normalization processing with the original feature sequence.

[0090] Concatenate the features representing global dependencies with the normalized local temporal features, and perform feature transformation through a non-linear activation function. Combine the transformed fused features with the propagation acceleration factor, and input them into a feed-forward neural network composed of multiple layers of perceptrons to obtain the feature representation for predicting the propagation inflection point.

[0091] Calculate the probability of a propagation inflection point occurring at each time step based on the inflection point prediction features. Combine the inflection point probability with the propagation intensity between grids through weighted combination to obtain the propagation risk probability for different regions.

[0092] Exemplarily, process the propagation feature sequence of a certain infectious disease for 10 days, and the position encoding at each time step is transformed through a sine-cosine function with a period of 10. For example, the position information on the 5th day is transformed into (0.59, 0.81), which is combined with the propagation features of that day to form an enhanced feature vector.

[0093] Construct 8 attention heads, and each head generates query, key, and value matrices with a dimension of 64 through independent linear transformations. The calculated attention scores show that the propagation features on the 5th day have strong correlations with those on the 3rd and 4th days, and the corresponding attention weights are 0.3 and 0.4 respectively.

[0094] Set 4 parallel convolutional branches with initial dilation rates of 1, 2, 4, and 8 respectively. When the acceleration factor of temperature increase is detected, adjust the dilation rates to 2, 4, 8, and 16 respectively to expand the receptive field range. The fused features show a rapid upward trend in propagation intensity.

[0095] After concatenating the global dependency features and the local temporal features, perform a non-linear transformation with a two-layer activation function of ReLU. Combining the acceleration factor information of the increase in relative humidity, the feed-forward network predicts that a propagation inflection point may occur on the 7th day.

[0096] Calculate that the probability of an inflection point occurring on the 7th day is 0.85. Combining the propagation intensity of 0.6 between grids at this time point, determine that the propagation probability of the high-risk area is 0.75.

[0097] In this embodiment, through the combination of position encoding and the multi-head attention mechanism, the long-range dependencies in the propagation feature sequence are effectively captured, overcoming the problem of insufficient long-distance information modeling ability of traditional temporal models; design an atrous convolution structure based on dynamic adjustment of the propagation acceleration factor to achieve adaptive receptive field adjustment for the propagation process and improve the model's response ability to sudden propagation events; adopt a fusion strategy of global dependency features and local temporal features, taking into account the overall propagation trend while retaining the fine-grained temporal patterns, improving the accuracy of inflection point prediction; jointly model the propagation inflection point probability and the grid propagation intensity to achieve precise quantification of the propagation risk, providing a reliable basis for the precise deployment of prevention and control measures.

[0098] In an alternative embodiment, by constructing multiple parallel dilated causal convolution branches and dynamically adjusting the dilation rate based on the propagation acceleration factor, multiple sets of propagated local temporal features are obtained. Adaptive fusion is performed on the multiple sets of propagated local temporal features, and residual connection and layer normalization processing are performed with the propagated feature sequence to obtain the propagated normalized features, including: Input the propagated feature sequence into multiple parallel convolution branches, each of the parallel convolution branches is set with a different initial dilation rate, the initial dilation rates are assigned in an exponentially increasing manner, and each of the parallel convolution branches includes multiple layers of dilated causal convolution layers; Obtain the change rate of the propagation acceleration factor at the current moment, calculate the dilation rate adjustment coefficient according to the magnitude of the change rate, and multiply the dilation rate adjustment coefficient by the initial dilation rate of each of the parallel convolution branches to obtain the dynamic dilation rate; Perform dilated causal convolution operations on the propagated feature sequence in each of the parallel convolution branches using the dynamic dilation rate to obtain multiple sets of propagated local temporal features; Calculate the correlation coefficient between each set of the propagated local temporal features and the propagation acceleration factor, and input the correlation coefficient into a normalization function to obtain the feature importance weight; Weight the multiple sets of propagated local temporal features based on the feature importance weight, and obtain the fused propagated local temporal features by passing the weighted features through an adaptive fusion network; Perform residual connection on the fused propagated local temporal features and the propagated feature sequence, and perform layer normalization processing on the result of the residual connection to obtain the propagated normalized features.

[0099] In a specific implementation manner, in the first stage of processing the propagated feature sequence, a parallel convolution branch network structure is constructed. Specifically, four exactly the same branches are constructed, and each branch internally includes three layers of dilated causal convolution layers. The initial dilation rates of these branches are set in an exponentially increasing manner. The initial dilation rate of the first branch is 1, the second branch is 2, the third branch is 4, and the fourth branch is 8. Inside each branch, the number of input channels of the convolution layer is consistent with the dimension of the propagated feature sequence, the number of output channels is uniformly set to 128, and the convolution kernel size is selected as 3. After each layer of convolution operation, a rectified linear unit activation function processing and a dropout operation are sequentially performed, and the dropout probability is set to 0.1. When performing causal convolution, specific padding operations are used to ensure that the calculation at the current moment can only utilize historical information, and at the same time, one-dimensional convolution operations are performed in the time dimension while keeping the sequence length unchanged.

[0100] Next, implement the dynamic adjustment mechanism for the dilation rate. First, use a sliding window of size 3 to calculate the short-term change trend of the propagation acceleration factor, and take the difference ratio between the value at the current moment and the average value within the window as the change rate. Set different adjustment strategies according to the magnitude of the change rate: when the change rate is less than 10%, the adjustment coefficient is 1.0; when the change rate is between 10% and 30%, the adjustment coefficient is 1.5; when the change rate exceeds 30%, the adjustment coefficient is 2.0. Multiply these adjustment coefficients by the initial dilation rates of each branch respectively to obtain the updated dynamic dilation rates. This update process is executed once at each time step, and at the same time, ensure that the updated dilation rate does not exceed one-third of the sequence length to maintain the effectiveness of feature extraction.

[0101] Before the convolution operation, first perform min-max normalization on the input propagation feature sequence and add positional encoding to enhance the temporal information. During the processing of the three-layer convolution, the first layer is mainly responsible for capturing local feature patterns, the second layer obtains information in a larger range by expanding the receptive field, and the third layer integrates feature information at multiple scales. After each layer of convolution processing, batch normalization, residual connection, and dropout regularization operations are performed in sequence to improve the stability of feature extraction and the generalization ability of the model.

[0102] Then, perform correlation analysis. Use the Pearson correlation coefficient to calculate the correlation between each group of propagation local temporal features and the propagation acceleration factor. Considering the possible lag effect during the propagation process, calculate the correlation coefficients at different time delays and select the maximum correlation coefficient as the index of feature importance. Then use the softmax function to normalize these correlation coefficients, set the temperature parameter to 0.5 to enhance the weight difference between different features, and ensure that the sum of the final obtained weights is 1.

[0103] When implementing feature fusion, first use the broadcasting mechanism to apply the feature importance weights to each group of features, keeping the dimension of the features unchanged. Then process the weighted features through an adaptive fusion network, which consists of two fully connected layers. The dimension of the first layer is twice the input dimension, and the dimension of the second layer is the same as the input dimension. Use the Gaussian error linear unit activation function and add a layer normalization layer. At the same time, introduce the multi-head attention mechanism, set 8 attention heads, with the dimension of each head being 16, and use scaled dot-product attention to calculate the attention weights.

[0104] In the last stage, residual connection and normalization are performed. The fused features are directly added to the original propagation feature sequence to ensure that their dimensions match. Then layer normalization is carried out, calculating the mean and variance of the features, initializing the normalization parameter γ to 1, β to 0, and adding a very small value of 1e-5 to prevent division by zero. Finally, the output is processed to ensure stable feature distribution and control the range within [-1, 1]. Throughout the process, an adaptive momentum optimizer is adopted, with the learning rate set to 0.001, the batch size to 32, the weight decay parameter to 0.01, and the gradient clipping threshold set to 1.0. The settings of these parameters jointly ensure the stability and effectiveness of model training.

[0105] In this embodiment, through the multi-branch parallel processing and dynamic dilation rate adjustment mechanism, multi-scale adaptive analysis of the propagation feature sequence is achieved, improving the model's ability to capture propagation patterns at different time scales; introducing a dynamic dilation rate adjustment strategy driven by a propagation acceleration factor enables the model to adaptively adjust the receptive field range according to environmental changes, enhancing the response ability to sudden propagation events; adopting a feature importance evaluation and adaptive fusion mechanism based on correlation analysis realizes the optimal combination of features at different time scales, improving the accuracy of feature expression; by combining residual connection and layer normalization, the retention of original feature information and the stability of feature distribution are effectively balanced, enhancing the training effect and generalization ability of the model.

[0106] The infectious disease early warning and monitoring system based on big data and deep learning includes: The first unit is used to obtain historical infectious disease data, population movement trajectory data, medical visit data, and environmental monitoring data as the training dataset; The second unit is used to map the historical infectious disease data to a geographical space grid based on the training dataset, calculate the propagation intensity between grids according to the population movement trajectory data, mark the infectious disease susceptible areas based on the medical visit data, identify the propagation acceleration factor in combination with the environmental monitoring data, and establish a spatio-temporal propagation link graph; The third unit is used to perform dynamic segment encoding on the spatio-temporal propagation link graph, divide the continuous time series into multiple propagation cycles, calculate the temporal changes of the propagation intensity between grids and the infectious disease susceptible areas for each propagation cycle, and generate a propagation feature sequence; The fourth unit is used to train a deep learning model based on the propagation feature sequence, predict the inflection point moment of infectious disease propagation according to the change trends of the propagation acceleration factor and the propagation intensity between grids, and calculate the propagation probability of high-risk areas; The fifth unit is used to mark the corresponding grid as a warning area when the propagation probability exceeds a preset threshold, generate warning information including the warning area and the propagation inflection point, and push the warning information to the monitoring terminal through a preset communication interface.

[0107] Provide an electronic device, comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0108] Provide a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0109] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing various aspects of the present invention.

[0110] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An infectious disease early warning and monitoring method based on big data and deep learning, characterized in that: include: Obtain historical data on infectious diseases, population movement trajectory data, medical treatment data, and environmental monitoring data as training data sets; Based on the training data set, historical data on infectious diseases are mapped to geographic space grids, the transmission intensity between grids is calculated based on the movement trajectory data of the crowd, and infectious disease-susceptible areas are marked based on medical treatment data. The transmission acceleration factor is identified in combination with environmental monitoring data to establish a spatiotemporal transmission link map. Dynamically segment and encode the spatiotemporal propagation link diagram, divide the continuous time series into multiple propagation cycles, calculate the temporal changes of the inter-grid propagation intensity and the infectious disease susceptible area for each propagation cycle, and generate a propagation feature sequence; Based on the transmission feature sequence, the deep learning model is trained to predict the turning point of infectious disease transmission and calculate the transmission probability in high-risk areas according to the changing trend of the transmission acceleration factor and the transmission intensity between grids; When the propagation probability exceeds the preset threshold, the corresponding grid is marked as a warning area, and warning information including the warning area and the propagation inflection point is generated, and the warning information is pushed to the monitoring terminal through the preset communication interface.

2. The method according to claim 1, characterized in that Based on the training data set, the historical data of infectious diseases are mapped to geographic space grids, the transmission intensity between grids is calculated based on the population movement trajectory data, the susceptible areas of infectious diseases are marked based on medical treatment data, and the transmission acceleration factors are identified in combination with environmental monitoring data. The spatiotemporal transmission link diagram is established, including: Based on the population density distribution data, the geographic space is gridded using a quadtree structure to obtain a geographic space grid; Mapping the historical data of infectious diseases to a geographic space grid, constructing a spatiotemporal data structure of infectious diseases, and calculating the disease density distribution of the geographic space grid based on a spatial kernel function and a temporal kernel function; Perform abnormal trajectory filtering and sampling frequency uniformity processing on the crowd movement trajectory data to obtain standard crowd movement trajectory data; based on the standard crowd movement trajectory data, calculate the crowd flow between the geographic space grids, and calculate the transmission intensity between the geographic space grids in combination with the disease density distribution; Extracting the characteristics of the medical consultation population from the medical consultation data, constructing a susceptibility index, and calculating the infectious disease susceptibility of the geographic space grid in combination with the disease density distribution; Extracting environmental characteristic parameters from the environmental monitoring data, calculating the deviation between the environmental characteristic parameters and the corresponding propagation optimum value, and identifying the propagation acceleration factor; According to the transmission intensity, infectious disease susceptibility and transmission acceleration factor between the geographic space grids, a spatiotemporal transmission link map of infectious diseases is constructed.

3. The method according to claim 2, characterized in that Extracting environmental characteristic parameters from the environmental monitoring data, calculating the deviation between the environmental characteristic parameters and the corresponding propagation optimum value, and identifying the propagation acceleration factor comprises: Acquire environmental monitoring data of multiple observation points, extract environmental characteristic parameters from the environmental monitoring data; perform statistics on the environmental characteristic parameters, use a kernel density estimation method to identify the optimal distribution interval, and introduce a time decay weight to perform weighted processing on historical data to obtain an initial propagation optimum value; Analyze the change trend and fluctuation range of the environmental characteristic parameters in different time periods to determine the historical value retention coefficient; determine the observation point weight coefficient according to the influence of each observation point in the propagation process; The historical value retention coefficient and the observation point weight coefficient are used to perform a weighted combination of the environmental characteristic parameters of each observation point and the initial propagation optimum value to obtain a dynamic propagation optimum value; Calculate the deviation between the observed value of the environmental characteristic parameter and the dynamic propagation optimum value, use a first deviation weight coefficient when the observed value is greater than the dynamic propagation optimum value, and use a second deviation weight coefficient when the observed value is less than the dynamic propagation optimum value, to obtain a characteristic parameter deviation value; The characteristic parameter deviation values ​​are combined in pairs and multiplied with the interaction coefficient to generate an interaction influence matrix; a single characteristic parameter deviation value is combined with the interaction influence matrix to obtain the comprehensive environmental impact intensity through a continuous multiplication operation; The mean and standard deviation of the comprehensive environmental impact intensity are calculated, and a propagation acceleration threshold is determined in combination with a preset adjustment coefficient; based on the comprehensive environmental impact intensity and the propagation acceleration threshold, a propagation acceleration factor is identified.

4. The method according to claim 1, characterized in that: The spatiotemporal propagation link diagram is dynamically segmented and coded, and the continuous time series is divided into multiple propagation cycles. For each propagation cycle, the temporal changes of the inter-grid propagation intensity and the infectious disease susceptible areas are calculated, and the propagation feature sequence is generated, including: Acquire propagation data of the space-time propagation link diagram, and calculate a propagation trend function based on the propagation data, wherein the propagation trend function is a cumulative value of propagation intensity between grids in the space-time propagation link diagram; Constructing a trend mutation detection function, and obtaining the mutation characteristics of the propagation trend according to the change rate and change direction of the propagation trend function in adjacent time windows; Calculate a dynamic threshold value according to the local variance, mark the time point when the absolute value of the trend mutation detection function exceeds the dynamic threshold value and the mutation direction changes as a dividing point, and divide the continuous time series of the space-time propagation link diagram into multiple propagation cycles; For each of the transmission cycles, the inter-grid transmission intensity and the distribution state of the infectious disease susceptible areas are combined to form a periodic characteristic matrix; Performing time series difference operation on the periodic feature matrix to obtain differential features, extracting time series evolution features through nonlinear dynamic analysis and statistical entropy analysis, and performing adaptive weighted fusion based on feature discrimination capability to form a time series feature vector; Constructing a coding mapping function based on the time series feature vectors of adjacent propagation cycles, performing weighted combination and time series smoothing on the time series feature vectors, and constructing an associated code; The associated codes are subjected to feature optimization and compression transformation to generate a propagation feature sequence of the space-time propagation link graph.

5. The method according to claim 4, characterized in that The periodic feature matrix is ​​subjected to time series difference operation to obtain differential features, and the time series evolution features are extracted through nonlinear dynamic analysis and statistical entropy analysis. Based on the feature discrimination capability, adaptive weighted fusion is performed to form a time series feature vector including: Perform time series difference operation on the periodic feature matrix to obtain first-order difference features and second-order difference features; Performing wavelet decomposition on the periodic feature matrix to obtain time-frequency features, extracting energy distribution features, kurtosis features and skewness features of the time-frequency features, and constructing a multi-scale feature set; Calculating the sample entropy of the periodic characteristic matrix to obtain complexity characteristics, performing phase space reconstruction on the periodic characteristic matrix to obtain a reconstructed phase space, extracting dynamic invariants from the reconstructed phase space, calculating the Lyapunov exponent based on the reconstructed phase space to obtain chaotic characteristics, and combining the complexity characteristics, the dynamic invariants and the chaotic characteristics to form a nonlinear feature set; Calculating the sliding entropy of the periodic characteristic matrix to obtain uncertainty characteristics, constructing a dynamic mutual information matrix to calculate information flow characteristics, and combining the uncertainty characteristics with the information flow characteristics to generate an information theory feature set; Constructing a feature importance evaluation function to calculate feature discrimination capability, and determining an adaptive weight coefficient according to the feature discrimination capability; The first-order difference feature, the second-order difference feature, the multi-scale feature set, the nonlinear feature set and the information theory feature set are weighted and combined according to the adaptive weight coefficient, and are associated with the duration of the propagation cycle to generate an enhanced time series feature vector; The enhanced time series feature vector is evaluated for robustness, and key features are screened to form a time series feature vector.

6. The method according to claim 1, characterized in that Based on the transmission feature sequence, the deep learning model is trained to predict the turning point of infectious disease transmission according to the changing trend of the transmission acceleration factor and the transmission intensity between grids, and the transmission probability of high-risk areas is calculated, including: Mapping the position information of each time step in the propagation feature sequence into sine function values ​​and cosine function values ​​respectively, and combining them with the features of the corresponding time step to obtain a propagation position coding feature; Mapping the propagation position encoding features into a propagation query matrix, a propagation key matrix and a propagation value matrix, calculating the similarity between the propagation query matrix and the propagation key matrix to obtain a propagation attention weight, multiplying the propagation attention weight by the propagation value matrix to obtain a plurality of propagation attention head features, and concatenating and linearly transforming the plurality of propagation attention head features to obtain a propagation global dependency feature; By constructing multiple parallel dilated causal convolution branches and dynamically adjusting the dilation rate based on the propagation acceleration factor, multiple groups of propagation local time series features are obtained, and the multiple groups of propagation local time series features are adaptively fused and residually connected and layer-normalized with the propagation feature sequence to obtain propagation normalized features; The global transmission dependency feature and the normalized transmission feature are concatenated and a transmission fusion feature is obtained through nonlinear transformation. The transmission fusion feature is combined with the transmission acceleration factor and then input into a feedforward neural network to obtain a prediction feature of the inflection point of infectious disease transmission. The probability of the infectious disease transmission inflection point is calculated based on the infectious disease transmission inflection point prediction characteristics, and the probability of the infectious disease transmission inflection point is weightedly combined with the transmission intensity between grids to obtain the transmission probability of the high-risk area.

7. The method according to claim 6, characterized in that By constructing multiple parallel dilated causal convolution branches and dynamically adjusting the dilation rate based on the propagation acceleration factor, multiple sets of propagation local time series features are obtained, the multiple sets of propagation local time series features are adaptively fused and residually connected and layer normalized with the propagation feature sequence, and the propagation normalized features obtained include: Inputting the propagation feature sequence into a plurality of parallel convolution branches, each of the parallel convolution branches is set with a different initial dilation rate, the initial dilation rate is allocated in an exponentially increasing manner, and each of the parallel convolution branches includes a plurality of dilated causal convolution layers; Obtaining the change rate of the propagation acceleration factor at the current moment, calculating the expansion rate adjustment coefficient according to the magnitude of the change rate, and multiplying the expansion rate adjustment coefficient by the initial expansion rate of each of the parallel convolution branches to obtain a dynamic expansion rate; Using the dynamic expansion rate to perform expansion causal convolution operations on the propagation feature sequence in each of the parallel convolution branches, to obtain multiple groups of propagation local time series features; Calculating the correlation coefficient between each group of the local time series features of the propagation and the propagation acceleration factor, and inputting the correlation coefficient into a normalization function to obtain a feature importance weight; Weighting the multiple groups of propagation local time series features based on the feature importance weights, and obtaining fused propagation local time series features through an adaptive fusion network. The fused propagation local time series features are residually connected with the propagation feature sequence, and the result of the residual connection is layer-normalized to obtain the propagation normalized features.

8. An infectious disease early warning and monitoring system based on big data and deep learning, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain historical data on infectious diseases, population movement trajectory data, medical treatment data, and environmental monitoring data as training data sets; The second unit is used to map the historical data of infectious diseases to geographic space grids based on the training data set, calculate the transmission intensity between grids based on the movement trajectory data of the crowd, mark the susceptible areas of infectious diseases based on the medical treatment data, identify the transmission acceleration factor combined with the environmental monitoring data, and establish a spatiotemporal transmission link map; The third unit is used to dynamically segment and encode the spatiotemporal propagation link diagram, divide the continuous time series into multiple propagation cycles, calculate the temporal changes of the inter-grid propagation intensity and the infectious disease susceptible area for each propagation cycle, and generate a propagation feature sequence; The fourth unit is used to train a deep learning model based on the transmission feature sequence, predict the turning point of infectious disease transmission according to the changing trend of the transmission acceleration factor and the transmission intensity between grids, and calculate the transmission probability in high-risk areas; The fifth unit is used to mark the corresponding grid as a warning area when the propagation probability exceeds a preset threshold, generate warning information including the warning area and the propagation inflection point, and push the warning information to the monitoring terminal through a preset communication interface.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images