Pig health state dynamic evaluation and early disease warning system based on intelligent algorithm

Abstract

The application discloses a pig health state dynamic evaluation and early disease warning system based on an intelligent algorithm, and belongs to the technical field of intelligent livestock breeding. The system comprises a multi-modal data acquisition unit S100 for collecting pig time series data; a first evaluation unit S200 for outputting a health risk probability; a second evaluation unit S300 for outputing a disease, stress, environment and management four-type risk source probability; a third evaluation unit S400 for outputting a plurality of epidemic probabilities when the disease risk exceeds a threshold value by taking the discrete data matrix in the early stage of the disease as input; and a warning pushing unit S500 for grading alarm according to the maximum epidemic probability. The application realizes closed-loop monitoring from risk identification, cause discrimination to epidemic confirmation through three-level progressive evaluation, solves the problems of single dimension, lack of dynamic evaluation and source differentiation, and enhances the generalization ability by using a training method based on a similar sample average matrix, and improves the robustness under complex environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent livestock farming technology, and more specifically, to a dynamic assessment system for pig health status and early disease warning based on intelligent algorithms. Background Technology

[0002] Pig farming is rapidly developing towards large-scale and intensive operations, and intelligent monitoring of pig health status and early disease warning have become research hotspots in the field of smart farming. Traditional manual inspections rely on experience-based judgment, which suffers from problems such as high labor intensity, strong subjectivity, poor real-time performance, and high missed detection rate, making it difficult to meet the needs of modern farming. In recent years, intelligent monitoring technologies based on machine vision, sound recognition, wearable sensors, and thermal infrared imaging have made significant progress. However, existing technologies still have the following shortcomings.

[0003] (i) The health status assessment has a single dimension and a high false alarm and false negative rate: Most existing methods are based on single modal data for identification, such as only detecting cough sounds, only identifying abnormal activities, only measuring ear temperature, or only analyzing image posture. The data dimensions of a single perception modality are limited, making it difficult to comprehensively represent the true health status of pigs under complex group rearing conditions, which easily leads to false alarms or false negatives. For example, a decrease in activity may be due to disease, or it may be caused by environmental stress or feeding cycle, which is difficult to distinguish with a single feature.

[0004] (ii) Lack of a progressive assessment mechanism makes it difficult to distinguish the source of risk: Existing research focuses on "abnormal event identification" or "classification at a certain moment", lacking a progressive assessment framework from "probability of health risk" to "identification of risk source" and then to "specific types of epidemic". When the system detects an anomaly, farmers cannot determine whether the anomaly is caused by disease, stress and environmental fluctuations or management factors, resulting in poor alarm interpretation and insufficient targeted intervention measures, which affects farmers' trust in the intelligent system.

[0005] (iii) Lack of dynamic and continuous assessment and early warning capabilities for individuals: The occurrence of diseases in pigs is usually a gradual process, and data at a single moment is easily affected by factors such as occasional stress, environmental disturbances and feeding cycles. Existing methods are mostly biased towards the identification of anomalies at a single moment, and lack the ability to continuously and dynamically model by combining historical behavior, physiological signals and environmental data, making it difficult to achieve early warning of health change trends. Especially in large-scale group housing environments, the similarity and frequent occlusion of individual pigs make it difficult to track individual identities, further limiting the feasibility of dynamic assessment.

[0006] (iv) Epidemic early warning models rely on manual annotation and have insufficient generalization ability: For early warning of swine epidemics, existing studies mostly use supervised learning methods, which require a large amount of sample data labeled with specific epidemic types. However, it is difficult to obtain multimodal data before and in the early stages of an epidemic, the annotation cost is high, and the data distribution varies greatly under different pig farms, breeds, ages and environmental conditions, resulting in weak cross-scenario generalization ability of the model. In addition, existing models are mostly single-disease identification and lack the ability to simultaneously evaluate the probability of multiple epidemics.

[0007] (v) Complex pig farm environment affects algorithm stability: In actual pig houses, there are factors such as changes in light, pen blockage, manure pollution, pig gathering and high background noise, which significantly reduce the accuracy of visual recognition, sound detection and thermal infrared temperature measurement; existing algorithms are mostly verified in the laboratory or controlled environment, and there is still a significant gap in stability and robustness under real production conditions.

[0008] In summary, while existing pig health monitoring technologies have made progress in individual identification, they have not yet formed an integrated dynamic monitoring system capable of integrating multimodal data, achieving progressive health assessment, distinguishing risk sources, and providing early probabilistic warnings for various epidemics. Therefore, there is an urgent need to develop an intelligent assessment and early warning system that can overcome the above shortcomings and is suitable for complex pig farm environments. Summary of the Invention

[0009] To overcome the problems existing in the above technologies, the present invention provides a dynamic assessment system for pig health status and early disease warning based on intelligent algorithms, the system comprising: Multimodal data acquisition unit; used to collect the first multimodal data of pigs at time steps, forming the input sequence X = [x1, x2, ... x t , ...x T ], where x t ∈R D Let R represent the D-dimensional feature vector at time step t, where T indicates that there are a total of T time steps, and R is a real number.

[0010] The first evaluation unit; with a built-in first intelligent algorithm model f1(·), is used to map the input sequence X into a health risk probability. ; and output as the first health assessment result; specifically, the first intelligent algorithm model f1(·) is a temporal neural network based on a hybrid of long short-term memory network and gated recurrent unit, with an input dimension of D, a hidden layer dimension of H, and an output layer of single neuron using the Sigmoid activation function to output p. risk .

[0011] The second evaluation unit; includes a built-in second intelligent algorithm model f2(·), used to compare the input sequence X with the first health evaluation result p. riskAfter concatenation, it is mapped to a probability vector of risk sources:

[0012] Where ⊕ represents vector concatenation, P cause =[p disease p stress p env p manage The probabilities of the four causes—disease, stress, environment, and management—are respectively defined, and the following conditions are met: [0, 1] 4 This represents a four-dimensional unit interval vector, where 4 represents the dimension of the vector.

[0013] The second intelligent algorithm model f2(·) is a multilayer perceptron classifier, whose input is the global average pooling vector of X and p. risk The concatenated result has an output layer of 4 neurons that uses the Softmax activation function to output P. cause .

[0014] The third evaluation unit; used when p disease Triggered when ≥θ, where θ is a preset threshold; the third evaluation unit has a built-in third intelligent algorithm model f3(·), which extracts a matrix M∈R from discrete second multimodal data during a period before and in the early stage of the target epidemic through the multimodal data acquisition unit. T1×D1 Given the input, output the probability vector of each pig belonging to one of the K types of swine epidemics. Where T1 is the number of discrete sampling times, D1 is the dimension of the multimodal data, and R is a real number; based on the probability vector P epidemic The value of the element in the middle is used to generate an early warning message.

[0015] Each row of the input matrix M in the third evaluation unit consists of a second multimodal data vector. The second multimodal data includes at least: posture features, activity level features, and feeding behavior features extracted by machine vision; cough frequency and acoustic features extracted by a sound sensor; ear temperature features extracted by infrared thermal imaging; and temperature, humidity, and ammonia concentration features extracted by an environmental sensor. The second multimodal data is a subset and aggregated representation of the first multimodal data within the corresponding time window. That is, for each discrete sampling moment within the time window, the second multimodal data vector is generated by the first multimodal data vector at the corresponding moment through feature filtering, dimensionality reduction, and statistical aggregation techniques.

[0016] The preset threshold θ is set to a fixed value between 0.6 and 0.8, and can be dynamically updated based on the recent warning accuracy through a dynamic adjustment strategy.

[0017] The training process of the third intelligent algorithm model f3(·) of the third evaluation unit is as follows: First, obtain the sample set from historical data. M i ∈R T1×D1 Let y be the multimodal data matrix of the i-th sample before the onset of the epidemic and before its invention. i ∈{0,1} K True epidemic labels with unique heat codes, , k represents the kth type of epidemic; second, for each sample (M) in the sample set j y j M j Calculate its comparison with M in other samples in the sample set. i similarity s ij =sim(M i M j M j Let Γ(M) represent the multimodal data matrix of the j-th sample before the onset of an epidemic and before its invention, where i ≠ j. The L samples with the highest similarity are selected, and their index set is denoted as Γ(M). j ); 3. Calculate the average matrix of the L samples. and average label vector Fourth, The third intelligent algorithm model f3(·) is trained using the training data, and the optimization objective is to minimize the prediction probability. and Cross-entropy loss between; and The calculation formulas are as follows:

[0018]

[0019] Similarity s ij Using cosine similarity:

[0020] Where vec(·) is the matrix vectorization operation. It is an L2 norm.

[0021] The number of the top L samples with the highest similarity is an integer between 5 and 20, and L does not exceed 10% of the total number of historical samples N.

[0022] The third intelligent algorithm model f3(·) is a deep network combining convolutional neural networks and attention mechanisms, with input M. i ∈R T1×D1 The output is a K-dimensional probability vector, and the dimension of the output probability vector is equal to the number of swine epidemics.

[0023] The multimodal data acquisition unit also includes an identity association module, which uses a combination of RFID ear tags and visual re-identification algorithms to bind the multimodal data at each time step to a specific individual pig, forming an individual-level input sequence X and matrix M.

[0024] In addition, the present invention also provides an early warning push unit, when p disease ≥θ and max(P) epidemic When )≥η, max(P) epidemic The corresponding epidemic is taken as the first epidemic, and a first-level alarm message containing the epidemic type, probability value, and details of the source of risk of the first epidemic is generated, where η∈[0.5, 0.9] is the alarm threshold; when p disease ≥θ, max(P) epidemic When η < η, a secondary alarm message containing details of each epidemic type, probability value, and source of epidemic risk is generated.

[0025] The beneficial effects of this invention are as follows: This invention provides a dynamic assessment system for pig health status and early disease warning based on intelligent algorithms. This system has the following advantages: 1. By fusing the temporal feature vector X and discrete matrix M collected by the multimodal data acquisition unit, it achieves a comprehensive representation of pig health status, overcoming the problems of limited dimensionality and high false alarm / false negative rates of single-modal data; 2. By setting the first assessment unit to output the health risk probability p... risk The second assessment unit outputs a risk source probability vector P. cause The third assessment unit outputs the epidemic probability vector P. epidemic A three-tiered progressive assessment architecture of "risk identification—source determination—epidemic confirmation" was constructed, enabling effective differentiation of the causes of health abnormalities and solving the problem of poor alarm interpretability in existing technologies; 3. By adopting a temporal neural network based on a hybrid of long short-term memory networks and gated recurrent units as the first assessment unit, continuous time modeling of the input sequence X was performed, realizing dynamic assessment and early trend perception of pig health status, overcoming the deficiency of existing technologies in lacking continuous dynamic assessment capabilities; 4. The average matrix of similar samples was calculated when training the third intelligent algorithm model. and average label vector Furthermore, the third intelligent algorithm model f3(·) was trained, reducing the model's dependence on precisely labeled samples and enhancing its generalization ability across pig farms and breeds. This solved the problems of high labeling costs and weak generalization ability in epidemic early warning models; 5. By setting up an early warning push unit, when p disease ≥θ and max(P) epidemic When p ≥ η, a level 1 alarm message is generated; when p disease ≥θ, max(P) epidemicWhen η < η, a secondary alarm message is generated; this realizes graded early warning for high-confidence epidemics and low-confidence disease risks, which not only ensures accurate intervention when the disease is identified, but also avoids missed reports when the model is uncertain, thereby improving the system's adaptability in different early warning scenarios and the user experience of farmers. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the intelligent algorithm-based dynamic assessment and early disease warning system for pigs according to the present invention. Detailed Implementation

[0027] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments given herein are for illustration and explanation only and are not intended to limit the present invention.

[0028] like Figure 1 The diagram shows a schematic of the intelligent algorithm-based dynamic assessment and early disease warning system for pigs according to the present invention. The schematic diagram includes a multimodal data acquisition unit S100, a first assessment unit S200, a second assessment unit S300, a third assessment unit S400, and an early warning push unit S500.

[0029] The core innovations of this invention are found in the first assessment unit S200, the second assessment unit S300, and the third assessment unit S400. This invention employs a three-tiered, progressive assessment framework—"risk identification—source determination—epidemic confirmation"—through these three units, effectively differentiating the causes of health abnormalities and addressing the problem of poor alarm interpretability in existing technologies. Simultaneously, the early warning push unit S500 provides tiered early warnings for both high-confidence epidemics and low-confidence disease risks, ensuring accurate intervention when identifying the disease while avoiding missed reports when the model is uncertain. The following specific embodiments further illustrate the innovations of this invention: The multimodal data acquisition unit S100 is deployed above and around each pen in a large-scale pig house to synchronously collect the first multimodal data of each pig at fixed time intervals (e.g., every minute or every 5 minutes) and store it in time step order to form an individual-level input sequence X = [x1, x2, ... x...]. t , ...x T ]; where the feature vector x corresponds to each time step t. t ∈R D It consists of the following first multimodal data: Visual features: Images of pigs are captured by RGB depth cameras deployed above the pen, and object detection and pose estimation algorithms are used to extract the pigs' body features, gait features, ear and eye status, feeding posture and lying posture.

[0030] Activity characteristics: The activity intensity, distance traveled, standing time, and lying time of pigs within a given time step are obtained through visual tracking algorithms or wearable accelerometers.

[0031] Sound characteristics: Pig calls and ambient sounds were collected by a microphone array deployed around the pen, and acoustic features such as cough frequency, call energy and duration of abnormal calls were extracted using a sound event detection algorithm.

[0032] Body temperature characteristics: By pointing an infrared thermal imaging camera at the ear area of ​​the pig, the highest temperature and average temperature of the ear area are extracted through target detection and temperature correction algorithms.

[0033] Environmental characteristics: The ambient temperature, relative humidity, ammonia concentration, and carbon dioxide concentration at the current time step are collected by temperature and humidity sensors, ammonia sensors, and carbon dioxide sensors deployed in the pigsty.

[0034] The visual features, activity level features, sound features, body temperature features, and environmental features mentioned above are time-aligned and concatenated into a D-dimensional feature vector x. t The system continuously records data in chronological order, forming an input sequence X of length T, which is used for health risk analysis in the subsequent first and second assessment units.

[0035] In addition, the multimodal data acquisition unit also includes an identity association module, which combines RFID ear tag readers and visual re-identification algorithms to bind the collected multimodal data with specific individual pigs at each time step, ensuring the identity consistency of individual-level sequence X and avoiding individual confusion in a group-housed environment.

[0036] The first evaluation unit S200 receives the individual-level input sequence X = [x1, x2, ... x] output from the multimodal data acquisition unit. t , ...x T It incorporates a first intelligent algorithm model f1 (·); this model adopts a temporal neural network architecture based on a hybrid of long short-term memory network (LSTM) and gated recurrent unit (GRU).

[0037] Model Architecture: The hybrid temporal neural network consists of stacked recurrent layers, such as an LSTM layer as the first layer and a GRU layer as the second layer, with each layer having the same hidden state dimension. The input sequence X passes through the LSTM layer and the GRU layer in sequence, and the GRU hidden state at the last time step is taken. Then it passes through a fully connected layer, and finally is mapped to a scalar value between 0 and 1, i.e., the health risk probability p, through the Sigmoid activation function. risk =f1(X)∈[0,1]; This probability value represents the confidence level that the current pig has a health abnormality or disease risk.

[0038] Training process of the hybrid temporal neural network: In the offline training phase, a large number of historical pig individual-level input sequences X and their corresponding health status labels (e.g., 0 for healthy, 1 for disease) are collected to construct a training dataset; supervised learning is adopted, and the model parameters are optimized using a binary classification cross-entropy loss function; during training, each sequence X is input into the hybrid temporal network, and the predicted probability p is output. risk The loss is calculated with the real labels, and the weight parameters of LSTM, GRU and fully connected layers are updated through backpropagation algorithm and adaptive moment estimation optimizer. After multiple rounds of iterative training until the loss converges, the first trained intelligent algorithm model is obtained. In the online deployment stage, the model receives continuously collected sequences X in real time and dynamically outputs the health risk probability of each time window.

[0039] Output result p risk On the one hand, it serves as the first health assessment result, directly presented to the farmers; on the other hand, it serves as an important input feature (concatenated with X) for the subsequent second assessment unit, used to further identify the source of risk.

[0040] The second evaluation unit S300 receives the input sequence X = [x1, x2, ... x] from the multimodal data acquisition unit. t , ...x T ] and the health risk probability p from the output of the first assessment unit risk It also incorporates a second intelligent algorithm model f2(·), the model being... The model employs a multilayer perceptron classifier architecture; the specific process is as follows: Input processing: First, global average pooling is performed on the input sequence X along the time dimension, that is, the feature vectors at T time steps are averaged element by element to obtain a D-dimensional global average pooling vector. Then, the vector and scalar p risk The features are concatenated to form a (D+1) dimensional joint feature vector, which serves as the input to the multilayer perceptron classifier.

[0041] Model Structure: The multilayer perceptron classifier consists of an input layer, several hidden layers (e.g., two fully connected layers, each followed by a ReLU activation function and a Dropout layer), and an output layer. The output layer contains four neurons, uses the Softmax activation function, and outputs a 4-dimensional probability vector P. cause =[p disease p stress p env p manage ]; These correspond to the probabilities of four types of causes: disease, stress, environment, and management, and satisfy the following conditions: .

[0042] Model training process: During the offline training phase, a large number of historical samples are collected. Each sample contains an input sequence X and the corresponding first evaluation unit output p. risk (Can be generated from a pre-trained first model), and manually labeled risk source tags (one-hot encoding of four types of causes); for example, [1, 0, 0, 0] indicates that the risk source is disease, and [0, 0, 1, 0] indicates that the risk source is stress; after global average pooling of X for each sample, it is compared with p risk The concatenation is input into a multilayer perceptron classifier, and the output is a predicted probability vector P. cause The classification cross-entropy loss function is adopted, and the network weights are updated through backpropagation algorithm and optimizer (such as Adam). After multiple rounds of iterative training until the loss converges, the trained second intelligent algorithm model is obtained.

[0043] Output: P output by the second evaluation unit cause Used to determine the primary sources of health risk; where p disease As a condition for triggering the third evaluation unit (when p) disease (When the probability value is ≥θ, subsequent epidemiological analysis is triggered). Other probability values ​​can help farmers take targeted intervention measures (such as improving the environment, reducing stress, or adjusting management).

[0044] The third assessment unit S400 continuously monitors the disease risk probability p output by the second assessment unit. disease When p disease When p ≥ θ, the third evaluation unit is triggered, where θ is a preset threshold (e.g., a fixed value between 0.6 and 0.8, or dynamically adjusted based on recent warning accuracy); when p disease When the value is less than θ, the system determines that the current pig has not reached the disease risk level required to initiate an epidemiological analysis. At this time, the system does not trigger the third assessment unit, nor does it generate any epidemic warning information. However, the system will continuously monitor and record: the system maintains continuous updates to the individual-level time-series database through continuous monitoring and data recording; it outputs the normal health status: the system will output the health risk probability p from the first assessment unit. risk The risk source probability vector P output by the second evaluation unit cause The results of routine health monitoring are displayed periodically on the aquaculture management interface, but no alarms are triggered; auxiliary management suggestions: If P cause The probability of environmental or managerial causes exceeding a certain auxiliary threshold (e.g., p) env >0.6), the system can provide non-urgent prompts, such as "It is recommended to check the temperature and humidity of the pen" or "Adjust the feeding density," to help farmers with preventative management; dynamically adjust the monitoring frequency: when p disease When it approaches but does not reach θ (e.g., 0.8θ < p) disease<θ), the system can automatically shorten the data sampling interval or increase the feature extraction frequency to improve sensitivity to subsequent risk increases.

[0045] Upon triggering, the system automatically backtracks the individual pig's historical multimodal data within a time window before and during the pre-infection period of the target epidemic (i.e., the incubation period to the precursor period) as secondary multimodal data. Data is then extracted from this data at discrete time points to construct a two-dimensional matrix M∈R. T1×D1 T1 represents the number of discrete sampling moments, and D1 represents the dimension of the second multimodal data. Each row corresponds to a discrete sampling moment (e.g., every half hour or every hour), and each column corresponds to a second multimodal data feature. The second multimodal data in matrix M includes at least: posture features, activity level features, and feeding behavior features extracted by machine vision; cough frequency and acoustic features extracted by sound sensors; ear temperature features extracted by infrared thermal imaging; and temperature, humidity, and ammonia concentration features extracted by environmental sensors. The second multimodal data acquisition time interval is extended (to half-hour or hour-level aggregated values) to ensure that rapid physiological and behavioral changes in the prodromal period of an epidemic can be captured, while avoiding noise and computational burden caused by the original second-level or minute-level data. Therefore, the second multimodal data is a subset or aggregated representation of the first multimodal data within the time window. That is, for each discrete sampling moment within the time window, its second multimodal data vector is generated by the first multimodal data vector at the corresponding moment through feature filtering, dimensionality reduction, or statistical aggregation techniques.

[0046] The third intelligent algorithm model structure: The third evaluation unit has a built-in third intelligent algorithm model f3(·). In this embodiment, f3(·) adopts a deep network combining convolutional neural networks and attention mechanisms; its input training phase uses the average matrix. The output is a K-dimensional probability vector. These correspond to the incidence probabilities of K types of swine epidemics; specifically, this deep network first uses several two-dimensional convolutional layers to extract matrices. The correlation patterns in the local time-feature space are analyzed, and then attention mechanisms (such as channel attention or spatiotemporal attention) are introduced to adaptively weight different time steps and different feature dimensions. Finally, K-dimensional probabilities are output through global average pooling and fully connected layers.

[0047] The unique training process of the training unit: N samples are collected from historical data to form a sample set. , of which M i ∈R T1×D1 Let y be the multimodal data matrix of the i-th sample before the onset of the epidemic and before its invention. i ∈{0,1} K True epidemic labels with unique heat codes, That is, only one component is 1, and the rest are 0.

[0048] Similarity calculation: For each sample in the sample set (M) j y j M in ) j Calculate its comparison with M in other samples in the sample set. i similarity s ij =sim(M i M j This embodiment uses cosine similarity:

[0050] Where vec(·) is the matrix vectorization operation. Let L be the L2 norm; sort the samples from highest to lowest similarity, select the top L most similar samples (L is usually an integer between 5 and 20, and does not exceed 10% of the total number of samples N), and denote their index set as Γ(M). j ).

[0051] Construct a smooth training objective: Calculate the average matrix and average label vector of these L samples:

[0052] Average label vector Instead of one-hot encoding, a soft label (e.g., [0.6, 0.1, 0.2, ...]) is used, reflecting the epidemic distribution within the sample domain.

[0053] Model training: the obtained The model f3(·) is trained using the training data; the target bit is optimized to minimize the prediction probability. and Cross-entropy loss between layers; updating the parameters of convolutional layers, attention layers, and fully connected layers through backpropagation algorithm and optimizer (such as Adam); this training method enables the model to learn common patterns in local neighborhoods rather than overfitting to a single accurate label, thereby significantly improving the generalization ability across pig farms and breeds.

[0054] Early warning push unit S500: After the model training of the third evaluation unit S400 is completed, the system inputs the matrix M of the current pigs (constructed by sampling every half hour or one hour) into the trained f3(·) to obtain the epidemic probability vector P. epidemic Subsequently, according to max(P) epidemic The alarm is categorized by size.

[0055] Level 1 alarm: When max(P) epidemicWhen ≥η, where η∈[0.5, 0.9] is the alarm threshold (e.g., η=0.6); then max(P) epidemic The corresponding epidemic is designated as the first epidemic, and a data source containing the epidemic type, probability value, and details of the risk source of the first epidemic is generated (e.g., combined with the p output of the second assessment unit). cause The system will issue a Level 1 alarm for environmental and management factors and push it to the aquaculture management personnel's terminal.

[0056] Level 2 alarm: When max(P) epidemic When η < η, a secondary alarm message is generated containing the types of each epidemic and their probability values, as well as details of the sources of risk for each epidemic. This alarm does not give a single conclusion, but provides a complete probability distribution for farmers to make a comprehensive judgment based on clinical symptoms and environmental data.

[0057] Through the aforementioned hierarchical mechanism, the system can provide precise intervention suggestions when the model is highly certain, and avoid missed reports when the model is uncertain, thus achieving flexibility and reliability in early warning.

[0058] In this embodiment, the matrix M used by the third assessment unit can be extracted from the historical data stored in the multimodal data acquisition unit (i.e., the first multimodal data sequence X is resampled and aggregated at the half-hour or hour level within a specific time window), or it can be obtained by an independent dedicated acquisition channel; the identity association module (RFID + visual re-identification) ensures the correctness of the individual attribution of the backtracking data; the entire system forms a complete closed-loop dynamic assessment and early warning process from risk identification (first assessment unit) - source identification (second assessment unit) - epidemic confirmation (third assessment unit) → graded early warning.

Claims

1. A dynamic assessment system for pig health status and early disease warning based on intelligent algorithms, characterized in that, The system includes: Multimodal data acquisition unit; used to collect the first multimodal data of pigs at time steps, forming the input sequence X = [x1, x2, ... x t , ...x T ], where x t ∈R D Let R represent the D-dimensional feature vector at time step t, where T represents the total number of time steps and R is a real number. The first evaluation unit; with a built-in first intelligent algorithm model f1(·), is used to map the input sequence X into a health risk probability. And output as the first health assessment result; The second evaluation unit; includes a built-in second intelligent algorithm model f2(·), used to compare the input sequence X with the first health evaluation result p. risk After concatenation, it is mapped to a probability vector of risk sources: Where ⊕ represents vector concatenation, P cause =[p disease p stress p env p manage The probabilities of the four causes—disease, stress, environment, and management—are respectively defined, and the following conditions are met: [0, 1] 4 This represents a four-dimensional unit interval vector, where 4 represents the dimension of the vector; The third evaluation unit; used when p disease Triggered when ≥θ, where θ is a preset threshold; the third evaluation unit has a built-in third intelligent algorithm model f3(·), which extracts a matrix M∈R from discrete second multimodal data during a period before and in the early stage of the target epidemic through the multimodal data acquisition unit. T1×D1 Given the input, output the probability vector of each pig belonging to one of the K types of swine epidemics. Where T1 is the number of discrete sampling times, D1 is the dimension of the multimodal data, and R is a real number; based on the probability vector P epidemic The value of the element in the middle is used to generate an early warning message.

2. The intelligent algorithm-based dynamic assessment and early disease warning system for pigs according to claim 1, characterized in that, The first intelligent algorithm model f1(·) is a temporal neural network based on a hybrid of long short-term memory network and gated recurrent unit. Its input dimension is D, the hidden layer dimension is H, and the output layer is a single neuron that uses the sigmoid activation function to output p. risk .

3. The intelligent algorithm-based dynamic assessment and early disease warning system for pigs according to claim 1, characterized in that, The second intelligent algorithm model f2(·) is a multilayer perceptron classifier, whose input is the global average pooling vector of X and p. risk The concatenation result shows that the output layer has 4 neurons and uses the Softmax activation function to output P. cause .

4. The intelligent algorithm-based dynamic assessment and early disease warning system for pigs according to claim 1, characterized in that, Each row of the input matrix M in the third evaluation unit consists of a second multimodal data vector. The second multimodal data includes at least: posture features, activity level features, and feeding behavior features extracted by machine vision; cough frequency and acoustic features extracted by a sound sensor; ear temperature features extracted by infrared thermal imaging; and temperature, humidity, and ammonia concentration features extracted by an environmental sensor. The second multimodal data is a subset and aggregated representation of the first multimodal data within the corresponding time window. That is, for each discrete sampling moment within the time window, the second multimodal data vector is generated by the first multimodal data vector at the corresponding moment through feature filtering, dimensionality reduction, and statistical aggregation techniques.

5. The intelligent algorithm-based dynamic assessment and early disease warning system for pigs according to claim 1, characterized in that, The preset threshold θ is set to a fixed value between 0.6 and 0.8, and can be dynamically updated based on the recent warning accuracy through a dynamic adjustment strategy.

6. The intelligent algorithm-based dynamic assessment and early disease warning system for pigs according to claim 1, characterized in that, The training process of the third intelligent algorithm model f3(·) of the third evaluation unit is as follows: First, obtain the sample set from historical data. , of which M i ∈R T1×D1 Let y be the multimodal data matrix of the i-th sample before the onset of the epidemic and before its invention. i ∈{0,1} K True epidemic labels with unique heat codes, , k represents the kth type of epidemic; second, for each sample (M) in the sample set j y j M j Calculate its comparison with M in other samples in the sample set. i similarity s ij =sim(M i M j M j Let Γ(M) represent the multimodal data matrix of the j-th sample before the onset of an epidemic and before its invention, where i ≠ j. The L samples with the highest similarity are selected, and their index set is denoted as Γ(M). j ); 3. Calculate the average matrix of the L samples. and average label vector Fourth, The third intelligent algorithm model f3(·) is trained using the training data, and the optimization objective is to minimize the prediction probability. and Cross-entropy loss between; and The calculation formulas are as follows: ， 。 7. The intelligent algorithm-based dynamic assessment and early disease warning system for pigs according to claim 6, characterized in that, The similarity s ij Using cosine similarity: Where vec(·) is the matrix vectorization operation. It is an L2 norm; The number of the top L samples with the highest similarity is an integer between 5 and 20, and L does not exceed 10% of the total number of historical samples N.

8. The intelligent algorithm-based dynamic assessment and early disease warning system for pigs according to claim 6, characterized in that, The third intelligent algorithm model f3(·) is a deep network combining convolutional neural networks and attention mechanisms, with input M. i ∈R T1×D1 The output is a K-dimensional probability vector, and the dimension of the output probability vector is equal to the number of swine epidemics.

9. The intelligent algorithm-based dynamic assessment and early disease warning system for pigs according to claim 1, characterized in that, The multimodal data acquisition unit also includes an identity association module, which uses a combination of RFID ear tags and visual re-identification algorithms to bind the multimodal data at each time step with a specific individual pig, forming an individual-level input sequence X and matrix M.

10. The intelligent algorithm-based dynamic assessment and early disease warning system for pigs according to claim 1, characterized in that, It also includes an early warning push unit, when p disease ≥θ and max(P) epidemic When )≥η, max(P) epidemic The corresponding epidemic is taken as the first epidemic, and a first-level alarm message containing the epidemic type, probability value, and details of the source of risk of the first epidemic is generated, where η∈[0.5, 0.9] is the alarm threshold; when p disease ≥θ, max(P) epidemic When η < η, a secondary alarm message containing details of each epidemic type, probability value, and source of epidemic risk is generated.

Images