An early warning algorithm for biological risk factors of african swine fever in the field of market circulation

Abstract

The application discloses an African swine fever full-chain multi-link risk early warning method based on TAER-TCN, and belongs to the field of animal disease prevention and control and intelligent food safety monitoring. The method integrates full-link data of pork circulation, constructs an integrated risk control system, takes a time attention-error recursive time convolution network as a core, extracts long time sequence features, strengthens key risk information, suppresses prediction errors, and improves complex scene early warning accuracy. Relying on an asymmetric causal kernel function, the method analyzes risk cross-link transmission paths, completes risk traceability positioning, realizes four-level graded early warning according to a risk index, and matches corresponding disposal schemes. The application solves problems of traditional technology link fragmentation, early warning lag, data island, traceability difficulty, false and missed report rate and the like, has advantages of full-chain coverage, early prediction, high precision, easy deployment and the like, is suitable for pig slaughtering, cold chain transportation, processing and storage and market sales scenes, and provides intelligent technical support for pig industry safety and food disease prevention and control.

Description

Technical Field

[0001] This invention belongs to the field of food safety and animal disease prevention and control technology, specifically involving a multi-stage risk early warning method for African swine fever (ASF) across the entire supply chain based on the Temporal Attention Error Recursive Temporal Convolutional Network (TAER-TCN). By performing advanced prediction and anomaly warning on multi-source time-series data from slaughtering, transportation (stage 1), processing, transportation (stage 2), warehousing, transportation (stage 3), and market stages, it achieves early detection, early warning, and early response to the risk of ASF transmission. This invention is particularly suitable for high-precision advanced early warning and risk tracing under complex conditions in the pork product distribution chain, such as uneven sampling, noise interference, temporal nonlinearity, non-stationarity, and multi-stage coupling. Background Technology

[0002] African swine fever (ASF) is an acute, highly virulent, and deadly infectious disease caused by the African swine fever virus (ASFV). It can infect various hosts, including domestic pigs, wild boars, and soft ticks. Clinically, it is characterized by high fever, hemorrhage, and high mortality. Currently, there are no commercially available vaccines or specific treatments worldwide. It is listed as a notifiable animal disease by the World Organisation for Animal Health (WOAH) and is classified as a Class A animal disease in my country. Since its introduction to my country in 2018, nationwide prevention and control efforts have been continuously implemented. Although the level of prevention and control in large-scale farming has significantly improved, factors such as the large number of small-scale farmers, frequent transportation, large cross-regional circulation, persistent environmental virus carriage, and inconsistent monitoring data quality mean that African swine fever outbreaks still exhibit sporadic outbreaks, cross-segment transmission, and a highly concealed incubation period, resulting in persistent pressure on prevention and control. The core logic of current African swine fever prevention and control is: early detection, early warning, and early response. This means identifying risk trends through monitoring data and issuing early warnings before the virus's incubation period, before clinical symptoms appear, and before the epidemic spreads, thus allowing time for precise culling, containment, and disinfection to block transmission. In the actual distribution chain, pig products typically undergo multiple transfers, transportation, processing, storage, and loading / unloading processes from slaughterhouse to market. This process involves many links, stakeholders, variables, and complex data types, and generally suffers from uneven sampling intervals, severe noise pollution, strong data nonlinearity, time-series non-stationarity, and spatiotemporal coupling of multiple links, posing significant challenges to traditional monitoring methods.

[0003] However, existing African swine fever (ASF) monitoring and early warning technologies generally suffer from the following insurmountable technical bottlenecks: the temporal proximity characteristics are not effectively utilized, and early, weak signals are difficult to capture. ASF risk exhibits strong temporal correlation and proximity dependence; indicators such as temperature, humidity, duration, disinfection records, and environmental anomalies at nearby times contribute significantly more to risk prediction. Traditional early warning models, such as LSTM, GRU, and TCN, assign equal weights to all historical time steps, failing to highlight weak upward signals during the incubation period, critical mutation points, and high-risk proximity moments. This results in insufficient early warning sensitivity, often only identifying the disease after clinical symptoms appear or the epidemic spreads, missing the optimal intervention window. Furthermore, the accumulation of multi-step prediction errors is severe, leading to a sharp decline in long-term early warning accuracy. Disease early warning requires multi-step forward prediction. Traditional rolling prediction methods directly carry over the prediction error from the previous step to the next, causing the error to accumulate and amplify with each prediction step. This results in long-term predictions deviating significantly from the true trend, drastically reducing the reliability of early warnings. In actual supervision, the need for 4-hour, 12-hour, and 24-hour forward warnings is most urgent, but traditional models cannot meet the accuracy requirements. Industrial-grade data suffers from poor quality, resulting in weak model robustness. Sensors in the distribution process commonly suffer from noise interference, signal drift, data gaps, and irregular sampling intervals. Traditional models lack robust feature learning mechanisms, easily misinterpreting normal fluctuations as risk signals or drowning out genuine anomalies in noise, leading to high false alarm and false negative rates, making long-term stable operation in real-world scenarios difficult. A comprehensive risk transmission mechanism is lacking, resulting in fragmented early warning systems. African swine fever can spread across multiple stages, including slaughtering, transportation, processing, storage, and markets; monitoring a single stage cannot reflect the overall risk. Existing technologies mostly focus on single-point scenarios such as breeding and slaughtering, failing to establish a unified multi-stage monitoring system. This makes it impossible to model asymmetric causal relationships between stages, leading to one-sided early warning results that are difficult to trace and pinpoint the source. Furthermore, low model inference efficiency hinders engineering deployment. Recurrent networks such as LSTM and GRU have a serial structure, resulting in slow training and inference speeds, which cannot meet the needs of real-time early warning on-site. Traditional statistical models such as AR, VAR, and PCA cannot fit the nonlinear, non-stationary, and strongly coupled time-series characteristics of disease risk, resulting in low early warning accuracy and poor generalization ability.

[0004] Current African swine fever (ASF) monitoring and early warning technologies generally suffer from the following major deficiencies. First, existing monitoring methods mostly focus on single-point scenarios such as farms and slaughterhouses, failing to establish a unified monitoring system covering the entire chain from slaughtering, transportation, processing, storage, to markets. Data from each link is independent, standards are inconsistent, and information is not shared, making it impossible to track risks across links and prevent early intervention. The virus can easily spread covertly in weak links such as transportation, storage, and markets, lacking a closed-loop control system. Traditional prevention and control primarily relies on laboratory PCR nucleic acid testing. From sampling and delivery to testing and result issuance, it typically takes 4–6 hours, and in some grassroots stations, even more than 12 hours. This is a typical post-event handling model, unable to provide early warning of risks. Often, by the time a positive result is found, the virus has already spread across regions, causing the epidemic to spread. In the multi-step early prediction process, traditional models such as LSTM, GRU, and TCN generally suffer from the problem of error amplification at each step. Small errors in the first step of prediction accumulate in subsequent steps, leading to long-term prediction distortion and persistently high false positive and false negative rates, making it difficult to meet actual regulatory needs. Before an outbreak of African swine fever, indicators such as sudden changes in temperature and humidity, lack of disinfection, abnormal quarantine, and transportation delays have strong early warning value. However, traditional time-series models use equal weights for all time steps, failing to highlight key period characteristics. This causes early, subtle anomalies to be masked by noise, hindering early identification. Actual circulation data contains heterogeneous data, including numerical (temperature, humidity, duration), categorized (disinfection, quarantine, transportation method), and identifiable (ID, code) data. Furthermore, different batches of samples differ in distribution, sampling frequency, and anomaly patterns, making it impossible for traditional methods to achieve unified modeling and feature reuse across multiple batches. When positive results are detected in the market or downstream links, current technology cannot determine whether the risk originates from slaughtering, transportation, processing, storage, or the market, making it impossible to pinpoint the earliest outbreak point or trace the transmission path. This leads to blind, expanded, and wasteful responses. Existing early warning systems often rely on simple binary judgments of risk level and spread, failing to differentiate between risk grades and the degree of diffusion, and thus cannot match differentiated response strategies, easily resulting in over-control or delayed response. Temporal Convolutional Networks (TCNs) employ causal dilated convolutions and residual connections, offering advantages such as parallel computation, long receptive fields, and gradient stability, thus outperforming recurrent networks in time series prediction. However, standard TCNs do not incorporate temporal attention to distinguish the importance of different time points, nor do they design error correction mechanisms to suppress accumulated errors. Directly applying them to African swine fever early warning still results in insufficient accuracy, high false alarm rates, and long-term prediction instability.

[0005] In summary, existing technologies cannot meet the requirements for intelligent early warning of African swine fever (ASF) across the entire chain, with high precision, advanced technology, traceability, tiered classification, and a closed-loop system. Therefore, developing a multi-stage risk early warning method for ASF across the entire chain, covering seven stages, incorporating real-world samples, and based on TAER-TCN, has significant industrial value and is of urgent practical importance. Summary of the Invention

[0006] This invention discloses a multi-stage risk early warning method for African swine fever based on TAER-TCN, covering seven stages: slaughtering, transportation 1, processing, transportation 2, warehousing, transportation 3, and market. It strengthens the weight of neighboring samples through time attention and suppresses the accumulation of multi-step prediction errors through error recursion correction, achieving unified processing of heterogeneous data from multiple samples. It completes the standardization, alignment, and feature extraction of numerical, categorical, and time-series data. It provides high-precision early warning, based on the TAER-TCN model, enabling risk prediction up to 12 steps in advance, with each step 30 minutes in advance, and an error rate ≤15%. It facilitates full-chain risk tracing, locating the source of risk and its transmission path, such as the positive result of sample 2833 originating from uncontrolled temperature and humidity in transportation stage 2. It also reuses sample experience, utilizing the risk patterns and handling experience of historical samples such as samples 2730 and 2770 to optimize current early warning and handling plans. The objectives are as follows: to construct a unified and standardized data collection system for seven stages—slaughtering, transportation 1, processing, transportation 2, warehousing, transportation 3, and market—to achieve multi-source heterogeneous data fusion; to realize multi-step advanced risk prediction based on TAER-TCN, suppressing error accumulation and improving long-term prediction accuracy; to enhance the weighting of nearest-neighbor time-time features through a time attention mechanism to improve early anomaly identification capabilities; to construct an asymmetric causal risk transmission model to achieve risk source location and propagation path tracing; and to complete model training, threshold optimization, and validation using actual batch samples to improve engineering practicality.

[0007] The purpose of this invention is to construct a unified data collection and standardization system for seven stages: slaughtering, transportation 1, processing, transportation 2, warehousing, transportation 3, and market, solving the problem of multi-source heterogeneous data fusion; to achieve multi-step advanced risk prediction based on TAER-TCN, suppressing error accumulation and improving long-term prediction accuracy; to enhance the weight of features at nearest time moments through a time attention mechanism, improving the ability to identify early anomalies; to construct an asymmetric causal risk transmission model, realizing the location of risk sources and the tracing of propagation paths; and to complete model training, threshold optimization, and verification using real circulation batch samples, improving engineering practicality.

[0008] The system of this invention consists of a five-layer structure: a multi-stage data acquisition layer; a data preprocessing and feature engineering layer; a TAER-TCN intelligent early warning layer; a risk transmission and traceability layer; and a tiered early warning and closed-loop disposal layer. This invention covers seven key stages in the entire pig product circulation chain, as shown below:

[0009]

[0010] in Indicates the slaughtering process. This indicates transportation stage 1. Indicates the processing stage. This indicates transportation stage 2. Indicates the warehousing process. This indicates transportation stage 3. This refers to the market segment.

[0011] The core innovation of this invention lies in simultaneously embedding a temporal attention mechanism and an error recursive correction mechanism into a temporal convolutional network, forming the TAER-TCN model. The temporal attention mechanism adaptively learns the contribution weights of samples at different times to the current risk, significantly improving the ability to capture nearby high-value anomaly signals. The error recursive correction mechanism calculates prediction errors in real time and compensates for subsequent predictions through nonlinear transformations, fundamentally solving the problem of multi-step prediction error accumulation and amplification. The dilated causal convolution ensures that the model uses only historical information for prediction, without introducing future information leakage, conforming to the logic of realistic early warning.

[0012] Meanwhile, this invention constructs a full-chain risk transmission model, which uses asymmetric weights to characterize the influence of preceding links on subsequent links, enabling risk tracing, source location, and visualization of propagation paths, so that early warnings can not only be "reported in advance" but also "accurately tracked". Attached Figure Description

[0013] Figure 1 :TAER-TCN network framework

[0014] Figure 2 : Structural diagram of the time attention mechanism

[0015] Figure 3 Schematic diagram of error correction mechanism

[0016] The beneficial effects of this invention are:

[0017] 1. This invention adaptively focuses on high-value, nearby temporal features through a time attention mechanism, accurately capturing weak risk signals in the early incubation period of African swine fever, achieving early warnings 12–72 hours in advance. The warning accuracy rate is ≥92.7%, the false negative rate is ≤4.3%, and the false positive rate is ≤6.7%. Its overall performance is significantly better than traditional models such as LSTM, GRU, and TCN, allowing sufficient time for epidemic response.

[0018] 2. By using an error recursive correction mechanism to calculate and nonlinearly compensate for prediction errors in real time, the propagation and amplification of errors in multi-step long-term prediction are effectively blocked, resulting in a RMSE as low as 0.038 for 48-hour long-term predictions and an improvement in prediction stability of more than 43%, thus solving the problems of long-term warning distortion and unreliability of traditional methods.

[0019] 3. This invention integrates monitoring data from all stages of breeding, slaughtering, transportation, storage, and market, models the risk transmission effect across stages, and can directly locate the source of risk while completing risk early warning. It has the characteristics of strong noise resistance, high robustness, and efficient deployment, and truly realizes the monitoring, early warning, and traceability of the entire process of African swine fever. Detailed Implementation

[0020] This invention relates to an intelligent early warning system for the entire African swine fever (ASF) chain based on a temporal attention error recurrent convolutional network (TAER-TCN). It is deployed across all nodes in the pig farming, slaughtering and processing, cold chain transportation, warehousing and preservation, and market sales scenarios. The hardware configuration is as follows: edge data acquisition terminals include temperature and humidity sensors, ammonia sensors, audio monitoring equipment, RFID identification equipment, and video surveillance equipment; data transmission utilizes 4G, 5G, LoRa, and local area network communication; the server is configured with at least an 8-core CPU, 32GB of RAM, and 1TB of storage capacity; the operating system is a 64-bit Linux system; the deep learning framework is based on PyTorch or TensorFlow; and the database uses a MySQL relational database and an InfluxDB time-series database. The overall system operation flow is as follows: real-time acquisition of multi-source data, data preprocessing and standardization, TAER-TCN model inference calculation, early warning threshold judgment and level classification, early warning information output and closed-loop handling, and continuous model iteration and updating.

[0021] Data is collected in a unified manner across five scenarios: breeding, slaughtering, transportation, storage, and market, to construct a full-chain monitoring dataset.

[0022] Data collected during the breeding process includes: average body temperature of pigs, daily feed intake, daily water intake, activity level score, cough frequency, pen temperature, pen humidity, ammonia concentration, ventilation volume, disinfection duration, saliva nucleic acid Ct value, blood antibody positivity rate, environmental swab positivity rate, personnel entry records, vehicle entry records, and feed batch information. Sampling frequency is set to once every 5 minutes.

[0023] Data collected during the slaughtering process includes: slaughterhouse temperature, humidity, knife disinfection temperature, knife disinfection time, flash freezing temperature, flash freezing time, head inspection results, trichinella detection results, carcass sampling results, slaughter volume, information on pig origin farms, transport vehicle information, and workshop disinfection records. The sampling frequency is set to once every 10 minutes.

[0024] Data collected during transportation includes: compartment temperature, compartment humidity, transportation duration, number of vehicle starts and stops, GPS trajectory information, disinfection time before loading, disinfection time after unloading, transportation type, packaging type, and vehicle qualification level. The sampling frequency is set to once every 1 minute.

[0025] Data collected during warehousing includes: cold storage temperature, cold storage humidity, internal temperature difference, ventilation frequency, entry time, exit time, storage duration, goods origin information, and disinfection records. Sampling frequency is set to once every 5 minutes.

[0026] Data collected at the market level includes: market ambient temperature, ambient humidity, countertop disinfection time, air circulation index, customer flow density, product turnover rate, daily sampling quantity, product origin, quarantine certificate number, and owner information. Sampling frequency is set to once every 15 minutes.

[0027] The collected raw data underwent missing value processing, outlier handling, standardization, and time-series sliding window construction. The missing value handling rules were as follows: for no more than 6 consecutive missing time steps, linear interpolation was used to fill in the missing values; for 7 to 20 consecutive missing time steps, sliding window mean was used; and for more than 20 consecutive missing time steps, the data was marked as invalid, and the model's robust inference mode was activated. Outlier handling adopted the 3σ principle: the mean and standard deviation of the data were calculated, and a data value was considered an outlier if the absolute value of the difference between the data value and the mean was greater than three times the standard deviation. Outliers were replaced with the median within the current sliding window. Standardization used the min-max normalization method to uniformly map all numerical variables to the 0-1 interval, eliminating the impact of dimensional differences on model training. The time-series sliding window construction set a window length of 48 hours and a prediction step size of 24 hours. Historical data from the past 48 hours was used as model input, and the risk status for the next 24 hours was used as model output to construct supervised training samples.

[0028] The model is based on dilated causal convolution, such as Figure 1 This system incorporates a temporal attention mechanism and an error recursive correction mechanism. The overall structure includes an input layer, a temporal attention layer, a dilated causal convolutional layer, a residual connection layer, an error recursive correction layer, and a fully connected output layer. The dilated causal convolutional module has a kernel size of 3 and dilation coefficients of 1, 2, 4, and 8 respectively. The network has 6 layers, and the ReLU activation function ensures that the model uses only historical data for prediction, preventing future information leakage and expanding the temporal receptive field to capture long-range dependencies. The temporal attention mechanism first transforms the input temporal features through a linear transformation to obtain a query matrix Q, a key matrix K, and a value matrix V. After calculating the attention score, a mask matrix is ​​used to mask future time-series information. Attention weights are obtained through softmax normalization, and the value matrices are weighted and fused to obtain attention-enhanced features, enabling the model to focus on recent high-value, high-risk temporal segments. The error recursive correction mechanism first calculates the error between the single-step predicted value and the true value. It then performs a nonlinear transformation on the error using two fully connected layers and the ReLU activation function. The transformed error is then superimposed on the original predicted value as a correction term to obtain the corrected prediction result. This corrected result is then used as the input to the next step of the model, thus preventing the cumulative propagation of errors.

[0029] Historical normal operating data was used as the training set, with a data duration of no less than 6 months and a total sample size of no less than 850,000 records. A joint loss function was used, including prediction mean squared error, error correction loss, and a regularization term, balancing model fit and generalization ability. The optimizer used the Adam algorithm, with an initial learning rate of 1e-4 and a batch size of 32. An early stopping strategy was employed during training; training automatically stopped after 10 consecutive rounds of validation set accuracy without improvement to prevent overfitting.

[0030] Based on a large amount of normal operating condition data fitted to a normal distribution, a 95% confidence interval is calculated as the basis for early warning judgment. The formula for calculating the confidence interval is the mean plus or minus 1.96 times the standard deviation. When the model's predicted value is higher than the upper limit of the confidence interval, it is judged as a risky abnormal state, and the early warning judgment process is initiated; when the predicted value is within the interval, it is judged as a normal operating state.

[0031] The system collects monitoring data from five scenarios in real time, automatically completing missing value completion, outlier correction, and data normalization. It constructs time-series input samples with a 48-hour window length and inputs them into the trained TAER-TCN model for inference. The model outputs the predicted risk value and probability for the next 24 hours, comparing the prediction results with the warning threshold to determine the warning level. If a warning is triggered, the system automatically outputs the warning level, the highest-risk stage, the prediction confidence level, and standardized handling suggestions, and pushes the warning information to the management platform and relevant personnel terminals. All data and warning results are stored in the database for subsequent model iterations and updates.

[0032] A blue alert indicates a slight deviation in data, and the response measure is to increase the frequency of monitoring and continuously track data changes; a yellow alert indicates a significant deviation in data, and the response measure is to strengthen environmental disinfection and increase the frequency of nucleic acid and antibody testing; an orange alert indicates a substantial deviation in data, and the response measure is to immediately conduct on-site investigations, isolate suspected pig groups, and strengthen environmental control; a red alert indicates a severe deviation in data, and the response measure is to immediately seal off the site, report to the competent authorities, and conduct a full-process source tracing and comprehensive disinfection.

[0033] The system performs incremental training every 15 days using the latest collected data, automatically updating model weights and warning thresholds to adapt to environmental changes, seasonal fluctuations, and adjustments to prevention and control strategies, thus maintaining long-term stability in warning accuracy.

Claims

1. A method for intelligent early warning of African swine fever across the entire chain based on a temporal attention error recurrent convolutional network, characterized in that, Includes the following steps: An African swine fever monitoring indicator system covering five key stages—breeding, slaughtering, transportation, storage, and market—was constructed. Numerical time-series data, categorical data, environmental data, testing data, and circulation data were collected. The collected data underwent missing value imputation, outlier removal, standardization, categorical coding, and time-series sliding window construction to generate compliant time-series training samples. The standardization formula is as follows: A TAER-TCN early warning model is constructed based on dilated causal convolution, embedding a time attention mechanism and an error recursive correction mechanism. The model is trained using historical normal data to learn the distribution of risk data, and a 95% confidence interval is calculated as the early warning threshold. The formula for the confidence interval is: . The system collects real-time monitoring data online, inputs it into the trained TAER-TCN model, outputs multi-step risk predictions and risk probabilities, compares the predictions with warning thresholds, triggers tiered warnings, and outputs the risk occurrence stage, confidence level, and handling suggestions.

2. The African swine fever whole-chain intelligent early warning method according to claim 1, characterized in that, The TAER-TCN model includes a dilated causal convolution module, a temporal attention module, an error recursive correction module, a residual connection module, and a fully connected output module. The dilated causal convolution module is used to ensure temporal causality and prevent the leakage of future information. The convolution calculation formula is as follows: The time attention module is used to dynamically assign weights to historical time steps, enhancing the feature contribution of adjacent high-value risk moments. The error recursive correction module is used to calculate the prediction error, generate a correction term through nonlinear mapping, compensate for the prediction result, and block the accumulation of error. The residual connection module is used to ensure the stability of deep network training and alleviate gradient vanishing. The fully connected output module is used to output the final risk prediction value and risk probability.

3. The African swine fever whole-chain intelligent early warning method according to claim 1, characterized in that, The calculation process of the time attention mechanism is as follows: A linear transformation is performed on the input temporal features to obtain the query matrix Q, key matrix K, and value matrix V, as shown in the formula: The attention score is calculated, and future time-time information is masked using a mask. The attention weight formula is obtained through softmax normalization: The formula for obtaining the time attention output feature by weighted fusion of the value matrix V is: A=WV.

4. The African swine fever whole-chain intelligent early warning method according to claim 1, characterized in that, The calculation process of the error recursive correction mechanism is as follows: the formula for calculating the error between the single-step predicted value and the true value is: The formula for nonlinearly transforming the error using two fully connected layers and the ReLU activation function is as follows: The error after nonlinear transformation is added as a correction term to the predicted value, resulting in the corrected output formula: The corrected output is used as the input for the next model step to prevent the cumulative propagation of errors.

5. The African swine fever whole-chain intelligent early warning method according to claim 1, characterized in that, The warning threshold is determined by a normal distribution obtained by fitting normal data, with a 95% confidence interval of: When the predicted value is higher than the upper limit of the range, it is judged as high risk and the corresponding level of warning is triggered.

6. The African swine fever whole-chain intelligent early warning method according to claim 1, characterized in that, It also includes a full-chain risk transmission modeling step: introducing link transmission coding, modeling the asymmetric risk transmission relationship of the flow face, so that the model can output risk prediction results and risk source location results at the same time.

7. The African swine fever whole-chain intelligent early warning method according to claim 1, characterized in that, The warning levels are divided into four levels: blue warning indicates slight data deviation, and the response measure is to increase monitoring; yellow warning indicates significant data deviation, and the response measure is to strengthen disinfection and sampling; orange warning indicates significant data deviation, and the response measure is to initiate emergency investigation; red warning indicates severe data deviation, and the response measure is to immediately blockade, trace the source, and take action.

8. An intelligent early warning system for the entire African swine fever chain based on a temporal attention error recurrent convolutional network, characterized in that, It includes a multi-source data acquisition unit, a data preprocessing unit, a TAER-TCN model inference unit, an early warning decision-making unit, and a visualization output unit; the system executes the method described in any one of claims 1-7 to realize the full-chain data acquisition, risk prediction, early warning, source location, and disposal suggestion output for African swine fever.

Images