Method and system for predicting incidence of infectious disease based on three-dimensional parallel convolution

Abstract

This invention relates to a method and system for predicting infectious disease incidence rates based on three-layer parallel convolution, belonging to the fields of artificial intelligence and public health monitoring technology. Addressing the shortcomings of existing prediction methods, such as difficulty in collaboratively modeling multi-source data, balancing long-term trends and short-term fluctuations, and lack of adaptive feature fusion, this method first preprocesses the input time-series data and decomposes it into coarse, medium, and fine-grained features. Then, it extracts long-term trend, medium-term cycle, and short-term fluctuation features through three parallel branches: a kernel adaptive network, a multi-scale temporal convolutional network, and a temporal convolutional module. Next, it uses a gated fusion unit to adaptively weight and fuse the multi-scale features, and finally outputs the predicted value through a fully connected layer. This method is mainly used for the accurate prediction of infectious disease incidence rates, providing support for disease early warning and prevention resource allocation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and public health monitoring technology. More specifically, this invention relates to a method and system for predicting the incidence of infectious diseases based on three-dimensional parallel convolution. Background Technology

[0002] Accurate prediction of infectious disease incidence rates is a crucial task in public health management. With the development of artificial intelligence, machine learning and deep learning methods have gradually become the mainstream for infectious disease incidence prediction. Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), and Gated Recurrent Units (GRUs) are suitable for capturing long-term dependencies, while Convolutional Neural Networks (CNNs) and Temporal Convolutional Networks (TCNs) excel at extracting local features and multi-scale patterns. These methods outperform traditional statistical and epidemiological models in prediction accuracy, but they still have the following limitations:

[0003] First, many models rely primarily on single-dimensional time series analysis, failing to fully integrate data features from other sources such as population movement and meteorological conditions. Because data from different sources differ in both temporal and spatial scales, effectively co-modeling these heterogeneous data sets presents challenges, limiting the models' ability to capture the complex patterns of epidemic evolution.

[0004] Secondly, infectious disease data typically encompasses both long-term trends and short-term sudden fluctuations. Existing models often struggle to balance these two different time scales in their structural design. Some models perform well in modeling long-term dependencies but are inadequate in responding to short-term mutations; others do the opposite. This imbalance affects the predictive stability of models across different stages of an epidemic.

[0005] Furthermore, when fusing features from multiple sources or at multiple scales, existing methods typically employ simple concatenation or fixed-weighting. Because infectious disease data inherently possesses high noise and non-stationary characteristics, this simple fusion mechanism struggles to adaptively adjust the importance of different features, resulting in poor robustness of the model in complex real-world scenarios and significant fluctuations in prediction results.

[0006] Therefore, providing a prediction method that can effectively integrate multi-source data, collaboratively model features at different time scales, and has adaptive fusion capabilities remains a technical problem that needs to be solved in this field. Summary of the Invention

[0007] One objective of this invention is to provide a method and system for predicting the incidence of infectious diseases based on three-layer parallel convolution, so as to at least solve the above-mentioned problems.

[0008] To achieve the objectives and other advantages of this invention, a method for predicting the incidence rate of infectious diseases based on three-layer parallel convolution is provided, comprising:

[0009] Step 1: Obtain historical infectious disease incidence time series data and related exogenous characteristic data, perform standardization and alignment processing to form a uniform and regular input sequence;

[0010] Step 2: Input the input sequence simultaneously into three parallel convolutional branches, each using a large-scale convolutional kernel. k l Mesoscale convolution kernel k m and small-scale convolution kernels k s Perform convolution operations and then linear projection to output coarse-grained features with uniform dimensions. Medium-grained characteristics and fine-grained features ,in, k l > k m > k s ;

[0011] Step 3, The input is fed into the coarse-grained prediction branch, where it undergoes nonlinear mapping via a kernel adaptive network, outputting long-term trend features. O l ;Will The input is fed into the mid-granularity prediction branch, where features are extracted through a multi-scale temporal convolutional network containing dilated convolutions and residual connections, outputting mid-term periodic features. O m ;Will The input is fed into the fine-grained prediction branch, where local features are extracted using a temporal convolution module. Combined with residual connections and pooling operations, the output is short-term fluctuation features. O s ;

[0012] Step 4: O l , O m , O s The components are concatenated by using a gated weight generation network to calculate the corresponding adaptive weights, which are then weighted and summed to obtain the fused features. O fusion ;

[0013] Step 5: O fusion The input is fed into a fully connected layer, which maps the predicted incidence rates of infectious diseases over the next H time steps. .

[0014] Preferably, in the infectious disease incidence prediction method based on three-layer parallel convolution, the exogenous feature data in step one includes population flow data, meteorological environment data, and holiday marking data.

[0015] Preferably, in the infectious disease incidence prediction method based on three-layer parallel convolution, in step three, the kernel adaptive network performs nonlinear transformation on the input features through learnable basis functions; the learnable basis functions include B-spline basis functions.

[0016] Preferably, in the infectious disease incidence prediction method based on three-layer parallel convolution, in step three, the multi-scale temporal convolutional network is composed of multiple residual blocks stacked together, and each residual block contains a one-dimensional causal convolution with a different dilation rate.

[0017] Preferably, in the infectious disease incidence prediction method based on three-dimensional parallel convolution, in step three, the temporal convolution module includes two one-dimensional causal convolutions, and the output of the second convolution is residually connected to the input of the temporal convolution module through a 1×1 convolution; the pooling operation adopts adaptive average pooling to compress the time dimension to a fixed length.

[0018] Preferably, in the infectious disease incidence prediction method based on three-layer parallel convolution, in step four, the gated weight generation network consists of a fully connected layer and a Sigmoid activation function.

[0019] Preferably, in the infectious disease incidence prediction method based on three-layer parallel convolution, step five involves optimizing the model parameters using a mean squared error loss function, defined as follows: , where Y i This is the true incidence rate value. B is the model prediction value, and B is the batch size.

[0020] This invention also provides an infectious disease incidence prediction system based on three-layer parallel convolution, comprising:

[0021] The data processing unit is used to acquire historical infectious disease incidence time-series data and related exogenous characteristic data, perform standardization and alignment processing, and form a uniform and regular input sequence.

[0022] The three-layer pyramid convolutional feature decomposition unit is used to simultaneously input the input sequence into three parallel convolutional branches, each using a large-scale convolutional kernel. k l Mesoscale convolution kernel k m and small-scale convolution kernels k sPerform convolution operations and then linear projection to output coarse-grained features with uniform dimensions. Medium-grained characteristics and fine-grained features ,in, k l > k m > k s ;

[0023] Multi-scale feature extraction unit, which is used to extract features from multiple scales. The input is fed into the coarse-grained prediction branch, where it undergoes nonlinear mapping via a kernel adaptive network, outputting long-term trend features. O l ;Will The input is fed into the mid-granularity prediction branch, where features are extracted through a multi-scale temporal convolutional network containing dilated convolutions and residual connections, outputting mid-term periodic features. O m ;Will The input is fed into the fine-grained prediction branch, where local features are extracted using a temporal convolution module. Combined with residual connections and pooling operations, the output is short-term fluctuation features. O s ;

[0024] Multi-scale gating fusion unit, which is used to... O l , O m , O s The components are concatenated by using a gated weight generation network to calculate the corresponding adaptive weights, which are then weighted and summed to obtain the fused features. O fusion ;

[0025] A fully connected output unit, which is used to output... O fusion The input is fed into a fully connected layer, which maps the predicted incidence rates of infectious diseases over the next H time steps. .

[0026] The present invention also provides an electronic device, comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the above-described method for predicting the incidence of infectious diseases based on three-dimensional parallel convolution.

[0027] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for predicting the incidence of infectious diseases based on three-dimensional parallel convolution.

[0028] The present invention has at least the following beneficial effects:

[0029] First, by employing a three-layer pyramid convolution decomposition unit, parallel feature extraction at the coarse-grained, medium-grained, and fine-grained levels is achieved. This not only takes into account both long-term trends and short-term fluctuations but also fully utilizes multi-source data features such as population flow, meteorological environment, and historical cases. In multi-dimensional and multi-scale data scenarios, this invention can significantly improve prediction accuracy.

[0030] Second, the coarse-grained kernel adaptive network module, the medium-grained multi-scale temporal convolutional network module, and the fine-grained temporal convolutional module have clear divisions of labor: they respectively focus on long-term trends, cross-scale periodic patterns, and short-term sudden anomalies. Then, through the multi-scale gating fusion unit mechanism, the adaptive fusion of multi-level information is achieved, so that the model can still maintain good predictive stability when facing seasonal fluctuations and sudden epidemics.

[0031] Third, by introducing a multi-scale gating fusion mechanism, the features of different branches are dynamically weighted, which can automatically adjust the feature contribution ratio according to the distribution of input data, thereby effectively suppressing noise interference and improving the stability and robustness of prediction results.

[0032] Fourth, the overall structure of the method of this invention is an end-to-end parallel convolution and gating fusion architecture, which has good scalability and portability. It can be directly deployed in public health monitoring systems to realize real-time prediction and early warning of infectious disease incidence, providing technical support for resource allocation and prevention and control strategy evaluation by public health departments. Compared with traditional methods that rely on expert experience to set parameters, this method can reduce manual intervention, shorten the modeling cycle, and improve the efficiency of practical applications.

[0033] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating an embodiment of the infectious disease incidence prediction method based on three-layer parallel convolution according to the present invention. Detailed Implementation

[0035] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings, so that those skilled in the art can implement it based on the description.

[0036] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0037] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation plan are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified.

[0038] In a first aspect, the present invention provides a method for predicting the incidence rate of infectious diseases based on three-layer parallel convolution, comprising:

[0039] Step 1: Obtain historical infectious disease incidence time series data and related exogenous characteristic data, perform standardization and alignment processing to form a uniform and regular input sequence;

[0040] Step 2: Input the input sequence simultaneously into three parallel convolutional branches, each using a large-scale convolutional kernel. k l Mesoscale convolution kernel k m and small-scale convolution kernels k s Perform convolution operations and then linear projection to output coarse-grained features with uniform dimensions. Medium-grained characteristics and fine-grained features ,in, k l > k m > k s ;

[0041] Step 3, The input is fed into the coarse-grained prediction branch, where it undergoes nonlinear mapping via a kernel adaptive network, outputting long-term trend features. O l ;Will The input is fed into the mid-granularity prediction branch, where features are extracted through a multi-scale temporal convolutional network containing dilated convolutions and residual connections, outputting mid-term periodic features. O m ;Will The input is fed into the fine-grained prediction branch, where local features are extracted using a temporal convolution module. Combined with residual connections and pooling operations, the output is short-term fluctuation features. O s ;

[0042] Step 4: O l , O m , O s The components are concatenated by using a gated weight generation network to calculate the corresponding adaptive weights, which are then weighted and summed to obtain the fused features. O fusion Step 5: Ofusion The input is fed into a fully connected layer, which maps the predicted incidence rates of infectious diseases over the next H time steps. .

[0043] In the above implementation, during the data preprocessing stage, historical infectious disease incidence rate time-series data and related exogenous characteristic data are acquired, including population flow data, meteorological and environmental data, and holiday marker data. The batch size B can be set to 32, 64, or 128, the time step T can be set to 30, 60, or 90, and the feature dimension D can be set to 10, 20, or 30. Standardization can be performed using the Z-score method, and alignment can be performed using time series interpolation methods. Data acquisition can be based on a server and database system, such as the relational database MySQL or the non-relational database MongoDB. Data processing can be performed using the Python programming language and the Pandas library. The workflow includes reading raw data from the database, checking for missing values ​​and filling them with linear interpolation, then standardizing the numerical features using the Z-score method to ensure all features are on the same scale, and time alignment is achieved through a unified timestamp, forming an input sequence in tensor format.

[0044] In the three-layer parallel convolutional feature decomposition stage, the input sequence is simultaneously input into three convolutional branches. The kernel size can be set as follows: large-scale convolutional kernel kl is 7, 9, or 11; medium-scale convolutional kernel km is 5, 7, or 9; and small-scale convolutional kernel ks is 3, 5, or 7. The output dimension of the linear projection layer can be uniformly set to 64, 128, or 256. Convolutional operations can be implemented based on deep learning frameworks such as TensorFlow or PyTorch. The workflow includes inputting the input data into three convolutional branches in parallel, applying different scale convolutional kernels to extract features, and then adjusting the feature dimension through linear projection to output coarse-grained, medium-grained, and fine-grained features.

[0045] In the multi-scale feature extraction and fusion output stage, a coarse-grained feature input kernel adaptive network uses B-spline basis functions for nonlinear mapping; the number of basis functions can be set to 10, 20, or 30. A medium-grained feature input multi-scale temporal convolutional network uses a residual block count of 3, 5, or 7, and a dilation rate of 1, 2, or 4. A fine-grained feature input temporal convolutional module uses two layers of one-dimensional causal convolution; the kernel size can be set to 3 or 5, and the pooling operation output length can be set to 16, 32, or 64. Feature fusion is achieved through a gated weight generation network, which consists of fully connected layers and a sigmoid activation function; the hidden layer dimension can be set to 128, 256, or 512. The fully connected layer outputs predicted values ​​for the next H time steps, where H can be set to 7, 14, or 30. The loss function uses mean squared error, and the optimizer can be either Adam or SGD. The workflow involves each branch processing the input features: the kernel adaptive network maps long-term trends using learnable basis functions; the multi-scale temporal convolutional network extracts mid-term periodic features using dilated convolutions; the temporal convolutional module combines residual connections and pooling to extract short-term fluctuation features; then, the multi-scale features are concatenated and input into a gating network to generate adaptive weights, which are then weighted and fused. Finally, the results are mapped to predicted values ​​through a fully connected layer. Parameter settings can be optimized using backpropagation to improve the basis function parameters and minimized using gradient descent to minimize the loss function.

[0046] This implementation method achieves effective integration of multi-source data and extraction of multi-scale temporal features. By balancing the contributions of long-term trends, medium-term cycles, and short-term fluctuations through an adaptive fusion mechanism, it improves the accuracy and stability of infectious disease incidence prediction and provides reliable support for public health management.

[0047] According to another embodiment of the present invention, in the method for predicting the incidence of infectious diseases based on three-layer parallel convolution, in step one, the exogenous feature data includes population flow data, meteorological environment data, and holiday marker data.

[0048] In the above implementation, population flow data can originate from base station signaling data of mobile network operators or public transportation card swipe records. Data dimensions can include daily pedestrian traffic, cross-regional movement ratio, or dwell time. Numerical ranges, for example, pedestrian traffic can be in units of tens of thousands, and movement ratios can be expressed as percentages. Meteorological environmental data can come from meteorological monitoring stations, including temperature, humidity, precipitation, and wind speed. The temperature range can be set to -10℃ to 40℃, humidity to 0% to 100%, and precipitation in millimeters. Holiday marker data can use binary encoding, with 0 representing non-holidays and 1 representing statutory holidays or special event days. The workflow involves collecting exogenous feature data from multiple data sources in real-time or periodically. Population flow data is obtained from mobile operators via API interfaces, meteorological environmental data is downloaded from meteorological department databases, and holiday marker data is parsed from public calendar sources. Data preprocessing includes format conversion, missing value imputation, and time alignment. For example, linear interpolation is used to handle missing meteorological data, and moving averages are used to smooth fluctuations in population flow data. Parameter settings are determined by historical data analysis to determine thresholds, such as setting the temperature anomaly threshold to 35℃ and the peak pedestrian traffic threshold to 1.5 times the daily average.

[0049] This implementation method enhances the model's ability to perceive environmental impacts and social factors by integrating multi-source exogenous feature data, thereby improving the comprehensiveness and reliability of infectious disease incidence prediction.

[0050] According to another embodiment of the present invention, in the infectious disease incidence prediction method based on three-layer parallel convolution, in step three, the kernel adaptive network performs nonlinear transformation on the input features through learnable basis functions; the learnable basis functions include B-spline basis functions.

[0051] In the above implementation, the number of basis functions in the kernel adaptive network can be set to 10, 20, or 30, and the input feature dimension can be set to 64, 128, or 256. The degree of the B-spline basis function can be set to 2, 3, or 4, and the number of control points can be set to 10, 20, or 30. The working process includes inputting the input features into the kernel adaptive network, which performs nonlinear mapping through learnable basis functions. The B-spline basis functions, defined according to the control points and degree, smoothly transform the features and output the mapped features. Parameter settings are optimized using backpropagation and gradient descent to optimize the basis function parameters, for example, by using the Adam optimizer to adjust the control point positions.

[0052] This implementation method achieves flexible nonlinear transformations in the kernel adaptive network through learnable basis functions (such as B-spline basis functions), which can better capture the complex patterns of input features, improve the model's ability to model long-term trends, and thus enhance the accuracy and adaptability of predictions.

[0053] According to another embodiment of the present invention, in the infectious disease incidence prediction method based on three-layer parallel convolution, in step three, the multi-scale temporal convolutional network is composed of multiple residual blocks stacked together, and each residual block contains a one-dimensional causal convolution with a different dilation rate.

[0054] In the above implementation, the number of residual blocks can be set to 3, 5, or 7, and the input and output dimensions of each residual block can be set to 64, 128, or 256. The process involves inputting the input feature sequence into a multi-scale temporal convolutional network. The network processes the data sequentially through stacked residual blocks. Each residual block contains a convolution operation, an activation function, and residual connections, ensuring effective gradient propagation and preventing vanishing gradients. Each residual block contains a one-dimensional causal convolution with a different dilation rate, which can be set to 1, 2, 4, or 8. The kernel size of the one-dimensional causal convolution can be set to 3, 5, or 7. The process involves applying a one-dimensional causal convolution in each residual block, using different dilation rates to expand the receptive field to capture multi-scale temporal dependencies. The causal convolution ensures that the output depends only on the current and past inputs, preventing the leakage of future information.

[0055] This implementation enhances the model's ability to capture multi-scale temporal dependencies by stacking residual blocks and using one-dimensional causal convolutions with different dilation rates, effectively extracting mid-term periodic features and improving the stability and accuracy of predictions.

[0056] According to another embodiment of the present invention, in the infectious disease incidence prediction method based on three-dimensional parallel convolution, in step three, the temporal convolution module includes two one-dimensional causal convolutions, and the output of the second convolution is residually connected to the input of the temporal convolution module through a 1×1 convolution; the pooling operation adopts adaptive average pooling to compress the time dimension to a fixed length.

[0057] In the above implementation, the kernel size can be set to 3, 5, or 7, and the number of input and output channels can be set to 64, 128, or 256. The process involves the input feature sequence first undergoing local feature extraction through a first-layer one-dimensional causal convolution. Causal convolution ensures that the output depends only on the current and past inputs. Then, non-linearity is introduced through an activation function such as ReLU, followed by further refinement of features through a second-layer one-dimensional causal convolution. The output of the second-layer convolution and the input of the temporal convolution module are residually connected through a 1×1 convolution. The number of input and output channels of the 1×1 convolution can be set to 64, 128, or 256, and the scaling factor for the residual connection can be set to 1.0. The process involves adjusting the number of channels using a 1×1 convolution on the input of the temporal convolution module to match the output dimension of the second-layer convolution, and then performing element-wise addition to achieve residual connections, thereby promoting gradient flow and preventing network degradation.

[0058] The pooling operation employs adaptive average pooling, compressing the time dimension to a fixed length, which can be set to 16, 32, or 64. The process involves inputting the output of the convolutional module into the adaptive average pooling layer, automatically calculating the pooling kernel size to compress a time series of arbitrary length to a preset fixed length, and outputting a feature vector of uniform dimension.

[0059] This implementation effectively extracts local temporal features by combining two layers of one-dimensional causal convolution with residual connections, and achieves dimensionality standardization through adaptive average pooling, thereby enhancing the model's ability to capture short-term fluctuations and its robustness in handling variable-length inputs, thus improving the stability and adaptability of predictions.

[0060] According to another embodiment of the present invention, in the infectious disease incidence prediction method based on three-layer parallel convolution, in step four, the gated weight generation network consists of a fully connected layer and a Sigmoid activation function.

[0061] In the above implementation, the input dimension of the fully connected layer can be based on the total dimension of the concatenated features, such as 192, 384, or 768, and the output dimension can be set to 3, corresponding to the weights of the three features: long-term trend, medium-term cycle, and short-term fluctuation. The hidden layer dimension can be set to 64, 128, or 256. The Sigmoid activation function ensures that the output weight values ​​are between 0 and 1. The working process includes inputting the concatenated multi-scale features into a gated weight generation network, first performing a linear transformation through the fully connected layer to map the input features to the hidden layer representation, then converting the hidden layer output into normalized weights through the Sigmoid activation function. These weights correspond to the contribution ratios of features at different scales, and finally using weighted summation to obtain the fused features.

[0062] This implementation uses a gated weight generation network constructed with fully connected layers and a sigmoid activation function to achieve adaptive weighted fusion of multi-scale features. It can dynamically adjust the contribution ratio of each feature according to the input data, thereby improving the stability and accuracy of the prediction model.

[0063] According to another embodiment of the present invention, in step five of the infectious disease incidence prediction method based on three-layer parallel convolution, the mean squared error loss function is used to optimize the model parameters. The loss function is defined as follows: , where Y i This is the true incidence rate value. B is the model prediction value, and B is the batch size.

[0064] In the above implementation, the batch size B can be set to 32, 64, or 128, and the true incidence rate value Y... i and predicted value The format is floating-point. The workflow includes the following steps during the model training phase: inputting the predicted sequence and the actual incidence rate sequence from the model into the loss function module, calculating the squared error for each sample, then calculating the batch average to obtain the loss value, calculating the gradient through the backpropagation algorithm, and using the optimizer to update the model parameters to minimize the loss.

[0065] This implementation optimizes the model parameters using the mean squared error loss function, making the predicted values ​​closer to the true incidence rate and improving the stability of model training and the accuracy of prediction.

[0066] Secondly, this invention provides an infectious disease incidence prediction system based on three-layer parallel convolution, comprising:

[0067] The data processing unit is used to acquire historical infectious disease incidence time-series data and related exogenous characteristic data, perform standardization and alignment processing, and form a uniform and regular input sequence.

[0068] The three-layer pyramid convolutional feature decomposition unit is used to simultaneously input the input sequence into three parallel convolutional branches, each using a large-scale convolutional kernel. k l Mesoscale convolution kernel k m and small-scale convolution kernels k s Perform convolution operations and then linear projection to output coarse-grained features with uniform dimensions. Medium-grained characteristics and fine-grained features ,in, k l > k m > k s ;

[0069] Multi-scale feature extraction unit, which is used to extract features from multiple scales. The input is fed into the coarse-grained prediction branch, where it undergoes nonlinear mapping via a kernel adaptive network, outputting long-term trend features. O l ;Will The input is fed into the mid-granularity prediction branch, where features are extracted through a multi-scale temporal convolutional network containing dilated convolutions and residual connections, outputting mid-term periodic features. O m ;Will The input is fed into the fine-grained prediction branch, where local features are extracted using a temporal convolution module. Combined with residual connections and pooling operations, the output is short-term fluctuation features. O s ;

[0070] Multi-scale gating fusion unit, which is used to...O l , O m , O s The components are concatenated by using a gated weight generation network to calculate the corresponding adaptive weights, which are then weighted and summed to obtain the fused features. O fusion ;

[0071] A fully connected output unit, which is used to output... O fusion The input is fed into a fully connected layer, which maps the predicted incidence rates of infectious diseases over the next H time steps. .

[0072] The above-mentioned infectious disease incidence prediction system based on three-dimensional parallel convolution is based on the same inventive concept as the above-mentioned infectious disease incidence prediction method based on three-dimensional parallel convolution, and can achieve the same technical effect. To avoid duplication, it will not be described again here.

[0073] Thirdly, the present invention provides an electronic device.

[0074] The electronic device includes a processor, a memory, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements each step of the above-mentioned method for predicting the incidence of infectious diseases based on three-dimensional parallel convolution and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0075] Fourthly, the present invention provides a computer-readable storage medium.

[0076] The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the aforementioned infectious disease incidence prediction method based on three-layer parallel convolution, achieving the same technical effect. To avoid repetition, this will not be elaborated further. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0077] Below, as Figure 1 As shown, a specific embodiment will be used to further explain and illustrate the method of the present invention.

[0078] I. Data Preprocessing and Alignment.

[0079] The historical infectious disease incidence time series data and related exogenous characteristic data are obtained and standardized and aligned to form a uniform and regular input sequence.

[0080] Specifically, the original sequence (Including historical incidence rate and exogenous features), where B is the batch size, T is the time step, and D is the feature dimension.

[0081] Standardization can be achieved using the following formula: In the formula, b represents the batch index, t represents the time step index, d represents the feature dimension index, and X... b,t,d μ represents the original value of the b-th sample at time step t and feature dimension d. d σ represents the mean of the feature dimension d. d ε represents the standard deviation of the feature dimension d, and ε represents a small constant added to avoid division by zero error.

[0082] Missing values ​​can be filled using forward imputation and local linear interpolation; discrete features such as holidays / weather can be concatenated using embedding (dimension merged into D).

[0083] Second and third layer pyramid convolution feature decomposition.

[0084] The input sequence is simultaneously fed into three parallel convolutional branches, each using a large-scale convolutional kernel. k l Mesoscale convolution kernel k m and small-scale convolution kernels k s Perform convolution operations and then linear projection to output coarse-grained features with uniform dimensions. Medium-grained characteristics and fine-grained features ,in, k l > k m > k s .

[0085] Specifically, we first perform "equal-length causal convolution (default stride s=1)", using three sets of convolution kernels with different receptive fields for parallel operation:

[0086] coarse-grained convolution kernel (large kernel) k l ): Covers a long-term window to extract long-term smooth change features;

[0087] Medium kernel k m ): Covering a medium time span, extracting periodic and phased features;

[0088] Fine-grained convolution kernel (small kernel) ks ): Covers a short window to extract local high-frequency and burst patterns.

[0089] Formal representation:

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] in, F l , F m , F s These represent the three features before alignment; ϕ (·) represents the activation function, either ReLU or SiLU; Conv k δ represents the convolution operation, where the kernel size k determines the receptive field of the feature; δ represents the bias term in the convolution operation. This represents the input sequence after standardization and time alignment as described above, with dimensions B×T×D, used for convolutional feature extraction; d l , d m , d s These are the dimensions of the three features.

[0095] The three features obtained are as follows:

[0096] ;

[0097] Where T′ represents the aligned time step, which may be shortened after convolution, and the output length T′=T when the time step is the same.

[0098] To ensure alignment of the three outputs in the subsequent fusion stage, linear projection is used to map the three features to a unified hidden dimension d′, resulting in the final output:

[0099] .

[0100] 3. Re-extraction of features from three paths.

[0101] Will The input is fed into the coarse-grained prediction branch, where it undergoes nonlinear mapping via a kernel adaptive network, outputting long-term trend features. O l ;Will The input is fed into the mid-granularity prediction branch, where features are extracted through a multi-scale temporal convolutional network containing dilated convolutions and residual connections, outputting mid-term periodic features. O m ;Will The input is fed into the fine-grained prediction branch, where local features are extracted using a temporal convolution module. Combined with residual connections and pooling operations, the output is short-term fluctuation features. O s .

[0102] (1) For coarse-grained features The kernel adaptive network first uses a linear mapping layer to process the input features. Mapping to intermediate representation:

[0103] in, d h H represents the input feature dimension of the kernel adaptive network, and H represents the intermediate feature representation after linear mapping, which is used for subsequent kernel basis function mapping. W h This is the trainable weight matrix of the linear mapping layer; b h This is the bias vector.

[0104] Subsequently, a learnable kernel basis function (such as the B-spline function Spline(·)) is used to perform a nonlinear transformation on the features. For the input feature H b,t,j Its kernel mapping form is:

[0105] ;

[0106] Among them, Spline m (·) represents the m-th basis function, with parameter a j,m During training, it can be learned, where M is the number of basis functions and H is... b,t,j Ψ represents the input feature value of the b-th sample at time step t and feature channel j; b,t This represents the feature vector of the b-th sample after being mapped by M basis functions at time step t, which is used for subsequent linear combination and fusion.

[0107] The obtained kernel mapping features are then combined using matrices: ;in, These are learnable weights.

[0108] Finally, a coarse-grained output is obtained through linear transformation: ;in, All of these are trainable parameters.

[0109] Final output This represents the long-term trend characteristics of the coarse-grained level, which is used for subsequent fusion.

[0110] (2) For medium-grained characteristics Multi-scale temporal convolutional networks first apply one-dimensional convolution to the input medium-granular features. Perform linear projection: In the formula, U (0) This represents the initial representation of the medium-granularity feature input after a 1×1 convolutional linear projection and activation function, with dimension R. B×T×d It is equivalent to the input layer features of a multi-scale temporal convolutional network.

[0111] A multi-scale temporal convolutional network internally stacks R residual blocks, each containing a different dilation rate. d r The convolution for the r-th residual block:

[0112] The input layer features are processed by LayerNorm and causal convolution to obtain intermediate features: ;

[0113] The output increment is then obtained by performing a 1×1 convolution mapping (dimensionality reduction or channel adjustment):

[0114]

[0115] Combine it with input U (r−1) Add them together to form a residual connection: ;

[0116] in, V (r) This represents the result of layer normalization and causal convolution in the r-th residual block ( The intermediate feature maps are obtained by applying the activation function ϕ(·) after the initial feature map. U (r) This represents the output characteristic of the r-th residual block, derived from the output of the previous residual block. U (r−1) With the current convolution result V (r) The features are obtained by performing 1×1 convolutions and summing them (i.e., the features after residual connections). (·) denotes a one-dimensional causal convolution operation with a kernel size of k. n The expansion rate is δ r , is a causal convolution, ensuring that the output depends only on the input at the current and past time steps; LN(·) represents Layer Normalization, used to stabilize training and speed up convergence; δ r =2 r-1This ensures that the receptive field expands exponentially with the number of layers; all convolutions are causal convolutions with left padding. p L =δ r ( k m -1), right fill p R =0, to avoid future information leaks.

[0117] The output of each residual block is added to the input to ensure gradient stability. ;in, This represents the input features (i.e., U) of the r-th residual block in a multi-scale temporal convolutional network. (r−1) By stacking R such residual blocks, efficient modeling of multi-scale time dependencies can be achieved.

[0118] After stacking R residual blocks, the final medium-granularity feature is obtained:

[0119] Effective sensory field: .

[0120] (3) For fine-grained features The temporal convolution module uses a one-dimensional convolutional kernel that slides along the time dimension to extract local high-frequency patterns. In the formula, Q1 represents the result of the first convolution layer, and Q2 represents the result of the second convolution layer. The kernel size for both one-dimensional convolution layers is k. s All are causal convolutions, ensuring no leakage of future information. The first convolutional layer mainly extracts short-term high-frequency features from the original input, and the second convolutional layer further combines these patterns.

[0121] To enhance the model's multi-level representation of short-term patterns, a second convolutional layer is added and residual connections are introduced: Input is connected via residual connection By fusing with convolutional features Q2, we can ensure that features are stable while capturing fine-grained features, thus alleviating the gradient decay problem in deep networks.

[0122] Pooling is applied to the local convolution results to compress the time dimension and obtain a global representation of the short-term features: Pooling operations (average pooling or max pooling) are performed in the time dimension to obtain compressed representations of short-term features, which are used to enhance the model's robustness to local patterns. The output results... .

[0123] IV. Multi-scale gating fusion.

[0124] Will O l , O m ,O s The components are concatenated by using a gated weight generation network to calculate the corresponding adaptive weights, which are then weighted and summed to obtain the fused features. O fusion .

[0125] Specifically, the three types of output features O l , O m , O s Concatenate along the feature dimension: F cat =Concat ( O l 、O m 、O s ); and unify dimensions through linear mapping. Multi-scale gating weights are generated using a fully connected layer and a sigmoid activation function. G=σ ( W g ⋅F align +b g In the formula, σ(⋅) is the Sigmoid activation function; Wg∈R d’×3 b is a trainable weight matrix; g ∈R 3 The bias term; G is a gated weight tensor (matrix) generated by a fully connected layer mapping + Sigmoid activation, which generates a weight vector for each time step t and batch b. G= [ g l 、g m 、g s ], g l +g m +g s ≈1, g l 、g m 、 g s These represent the weights of coarse-grained, medium-grained, and fine-grained output features in the fusion process, respectively.

[0126] The final fusion output is: , .

[0127] V. Prediction of fully connected layers.

[0128] Will O fusion The input is fed into a fully connected layer, which maps the predicted incidence rates of infectious diseases over the next H time steps. .

[0129] Specifically, the weighted fusion features O fusion The input is fed into one or more fully connected layers, which process it through linear transformations and activation functions to obtain an intermediate representation Z: Where FC represents a fully connected layer, and d′′ is the output dimension. d′′ can be equal to or less than d′, used for further compression of feature representation.

[0130] Finally, a linear transformation is used to obtain the final predicted value. It is a linear regression result based on fusion features, representing the predicted value of incidence rate.

[0131] Among them, FC out This represents the final fully connected output layer, which transforms the intermediate representation Z into a scalar prediction value. .

[0132] To optimize the prediction results, mean squared error (MSE) is used as the loss function to measure the difference between the model's predicted values ​​and the actual incidence rates. The loss function is defined as follows:

[0133]

[0134] Where Y i This is the true incidence rate value. B is the model prediction value, and B is the batch size.

[0135] To verify the effectiveness of the proposed method, a comparative experiment was conducted between the proposed infectious disease incidence prediction model based on three-layer parallel convolution and three representative single models in the current temporal prediction field: TCN (Temporal Convolutional Network), LSTM (Long Short-Term Memory Network), and Informer (a long sequence prediction model based on self-attention mechanism). These three models represent convolutional, recurrent, and attention-based structures, respectively, and comprehensively reflect the improvements in accuracy, precision, and stability of the proposed method. The experiments were conducted under the same dataset and hardware conditions, and the evaluation metrics included mean absolute error (MAE), mean squared error (MSE), and training time. The experimental results are shown in the table below.

[0136] Test results of different methods

[0137]

[0138] As shown in the table above, the method of this invention significantly outperforms existing single models in both the MAE and MSE key metrics. Compared to the TCN model, MAE is reduced by approximately 17.5%, and MSE by approximately 13.1%; compared to the LSTM model, MAE is reduced by approximately 11.8%, and MSE by approximately 8.7%; compared to the Informer model, MAE is reduced by approximately 8.8%, and MSE by approximately 6.4%. These results demonstrate that the method of this invention exhibits higher prediction accuracy and generalization performance in multi-scale feature modeling and gated adaptive fusion.

[0139] In terms of computational efficiency and convergence performance, under the same hardware environment and training rounds, the average training time of the method of this invention is approximately 22 minutes, which is about 29% and 24% less than that of LSTM and Informer models, respectively. Thanks to the parallel convolutional structure and multi-scale gating fusion mechanism, this invention significantly improves the computational efficiency and convergence speed of the model while maintaining high accuracy.

[0140] In summary, the infectious disease incidence prediction method based on three-layer parallel convolution of this invention exhibits significant advantages in the following aspects: in terms of accuracy, the average MSE and MAE are reduced by approximately 10% to 40%; in terms of computational efficiency, the model has fewer parameters, a lightweight structure, and faster training and inference speeds. Compared with existing single models, the method of this invention achieves significant improvements in multiple dimensions, including accuracy, precision, stability, and computational efficiency.

[0141] The number of devices and processing scale described herein are for simplification of the invention. Applications, modifications, and variations of the present invention's method and system for predicting infectious disease incidence rates based on three-layer parallel convolution will be readily apparent to those skilled in the art.

[0142] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for predicting the incidence of infectious diseases based on three-dimensional parallel convolution, characterized in that, include: Step 1: Obtain historical infectious disease incidence time series data and related exogenous characteristic data. The exogenous characteristic data includes population flow data, meteorological environment data, and holiday marking data. Standardize and align the above data to form a uniform and regular input sequence. Step two, input the input sequence into three parallel convolution branches at the same time, respectively using large scale convolution kernel k l , medium scale convolution kernel k m and small scale convolution kernel k s Convolution operation is carried out, and linear projection is carried out, and the output dimension of the coarse-grained feature is unified , medium-grained feature and fine-grained feature , wherein k l > k m > k s ; Step 3, The input is fed into the coarse-grained prediction branch, where it undergoes nonlinear mapping via a kernel adaptive network, outputting long-term trend features. O l ;Will The input is fed into the mid-granularity prediction branch, where features are extracted through a multi-scale temporal convolutional network containing dilated convolutions and residual connections, outputting mid-term periodic features. O m ;Will The input is fed into the fine-grained prediction branch, where local features are extracted using a temporal convolution module. Combined with residual connections and pooling operations, the output is short-term fluctuation features. O s ; Step 4: O l , O m , O s Concatenated along the feature dimension to form concatenated feature F cat The spliced ​​features F are obtained through linear mapping. cat The unified dimension is the alignment feature F align , Where B is the batch size, T′ is the aligned time step, and d′ is the uniform hidden dimension; the corresponding adaptive weights are calculated through a gated weight generation network consisting of fully connected layers and a sigmoid activation function, specifically: gated weights G=σ ( W g ⋅F align +b g ), where σ(⋅) is the Sigmoid activation function, Wg∈R d’×3 b is a trainable weight matrix g ∈R 3 For bias terms, G= [ g l 、g m 、g s ], g l +g m +g s ≈1, g l 、g m 、g s These represent the weights of coarse-grained, medium-grained, and fine-grained output features in the fusion process, respectively, and are adjusted according to adaptive weights. O l , O m , O s We perform a weighted summation to obtain the fusion feature O. fusion , , ; Step five, the O fusion to a fully connected layer to map to the predicted values of the infectious disease incidence rate for the next H time steps .

2. The method for predicting the incidence of infectious diseases based on three-layer dimensional parallel convolution according to claim 1, wherein, In step three, the kernel adaptive network performs a nonlinear transformation on the input features using learnable basis functions; Learnable basis functions include B-spline basis functions. 3.The method of predicting the incidence of infectious diseases based on three-dimensional parallel convolution according to claim 1, wherein, In step three, the multi-scale temporal convolutional network is composed of multiple residual blocks stacked together, each residual block containing a one-dimensional causal convolution with a different dilation rate. 4.The method of predicting the incidence of infectious diseases based on three-dimensional parallel convolution according to claim 1, wherein, In step three, the temporal convolution module contains two layers of one-dimensional causal convolution, and the output of the second convolution layer is residually connected to the input of the temporal convolution module through a 1×1 convolution; the pooling operation adopts adaptive average pooling to compress the temporal dimension to a fixed length.

5. The infectious disease incidence prediction method based on three-layer parallel convolution as described in claim 1, characterized in that, In step five, the model parameters are optimized using the mean square error loss function, which is defined as: where Y i is the true incidence value, is the model prediction value, and B is the batch size.

6. A system for predicting the incidence of infectious diseases based on three-dimensional parallel convolution, characterized in that, include: The data processing unit is used to acquire historical infectious disease incidence time-series data and related exogenous feature data, including population flow data, meteorological environment data and holiday marking data. The data is standardized and aligned to form a uniform and regular input sequence. The three-layer pyramid convolutional feature decomposition unit is used to simultaneously input the input sequence into three parallel convolutional branches, each using a large-scale convolutional kernel. k l Mesoscale convolution kernel k m and small-scale convolution kernels k s Perform convolution operations and then linear projection to output coarse-grained features with uniform dimensions. Medium-grained characteristics and fine-grained features ,in, k l > k m > k s ; Multi-scale feature extraction unit, which is used to extract features from multiple scales. The input is fed into the coarse-grained prediction branch, where it undergoes nonlinear mapping via a kernel adaptive network, outputting long-term trend features. O l ;Will The input is fed into the mid-granularity prediction branch, where features are extracted through a multi-scale temporal convolutional network containing dilated convolutions and residual connections, outputting mid-term periodic features. O m ;Will The input is fed into the fine-grained prediction branch, where local features are extracted using a temporal convolution module. Combined with residual connections and pooling operations, the output is short-term fluctuation features. O s ; Multi-scale gating fusion unit, which is used to... O l , O m , O s Concatenated along the feature dimension to form concatenated feature F cat The spliced ​​features F are obtained through linear mapping. cat The unified dimension is the alignment feature F align , Where B is the batch size, T′ is the aligned time step, and d′ is the uniform hidden dimension; the corresponding adaptive weights are calculated through a gated weight generation network consisting of fully connected layers and a sigmoid activation function, specifically: gated weights G=σ ( W g ⋅F align +b g ), where σ(⋅) is the Sigmoid activation function, Wg∈R d’×3 b is a trainable weight matrix g ∈R 3 For bias terms, G= [ g l 、g m 、g s ], g l +g m +g s ≈1, g l 、g m 、g s These represent the weights of coarse-grained, medium-grained, and fine-grained output features in the fusion process, respectively, and are adjusted according to adaptive weights. O l , O m , O s We perform a weighted summation to obtain the fusion feature O. fusion , , ; A fully connected output unit, which is used to output... The input is fed into a fully connected layer, which maps the predicted incidence rates of infectious diseases over the next H time steps. .

7. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the infectious disease incidence prediction method based on three-dimensional parallel convolution as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the infectious disease incidence prediction method based on three-dimensional parallel convolution as described in any one of claims 1 to 5.

Images