Pet disease early warning method, device and equipment and storage medium

Abstract

The invention provides a pet disease early warning method, device and equipment and a storage medium, and the method strictly distinguishes the data processing of the obtained meteorological original data and biological original data from the data level, respectively processes the meteorological original data and biological original data, and provides a corresponding data basis for a subsequent double-tower structure. A meteorological sequence tensor obtained by preprocessing meteorological original data is processed by using an independent meteorological tower, and a coded biological vector obtained by preprocessing biological original data is processed by using an independent biological tower, so that both an input layer and a middle layer are not mixed, and corresponding weather characteristics and biological characteristics are separated at a structural level; and information from different sources is prevented from being mixed in the same network to influence a final prediction result, so that the prediction accuracy of the model is improved.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to a method, apparatus, device, and storage medium for early warning of pet diseases. Background Technology

[0002] With the increasing demand for pet ownership, common pets such as cats and dogs have become important members of families, and their health has become a major concern. Among them, upper respiratory tract infections in pets caused by pathogens such as feline herpesvirus (FHV), feline calicivirus (FCV), and influenza A virus (IAV) are common clinical diseases. They not only cause pets to cough and have runny noses, but in severe cases, they can also lead to complications, bringing economic burden and emotional distress to pet owners.

[0003] To mitigate the risk of infection in advance, there are already pet disease early warning methods based on environmental and individual characteristics. However, these methods generally suffer from insufficient practicality and low accuracy in disease prediction, making it difficult to support refined pet health management. Summary of the Invention

[0004] This disclosure provides a method, apparatus, device, and storage medium for early warning of pet diseases, in order to at least solve the above-mentioned technical problems existing in the prior art.

[0005] A first aspect of this disclosure provides a pet disease early warning method, the method being applied to a trained pet disease early warning model, the pet disease early warning model including a weather tower, a biological tower, an exposure layer, and a task head network, the method comprising: Acquire raw meteorological data and raw biological data, wherein the raw biological data is acquired by biological data acquisition terminals distributed at multiple clinics and uploaded to the server with timestamps and geographic location identifiers; The raw meteorological data and the raw biological data are preprocessed respectively to obtain meteorological sequence tensors and encoded biological vectors. The meteorological sequence tensor is input into the meteorological tower to obtain the corresponding meteorological feature vector, and the encoded biological vector is input into the biological tower to obtain the corresponding biological feature vector; The meteorological sequence tensor is input into the exposure layer to obtain the daily exposure scalar for each pathogen, and the exposure vector corresponding to the combination of all pathogens is determined based on the daily exposure scalar for each pathogen. Based on the meteorological feature vector and the biological feature vector, the corresponding intermediate gating feature is determined, and based on the intermediate gating feature, the corresponding gating fusion feature is determined; The gated fusion features are concatenated with the exposure vector and then input into the task head network to obtain the corresponding infection risk probability, so as to provide early warning of pet diseases.

[0006] In one possible implementation, the raw biological data includes age data, sex data, living environment data, seasonal data, and species data. Correspondingly, the preprocessing of the raw meteorological data and the raw biological data to obtain a meteorological sequence tensor and an encoded biological vector includes: Construct a weather sequence tensor of a preset length based on the raw meteorological data; The age data is normalized to obtain normalized age data; The gender data, living environment data, seasonal data, and variety data are mapped to integer indices to obtain indexed multi-source data; The normalized age data and the indexed multi-source data are encoded to obtain the encoded biological vector.

[0007] In one embodiment, the raw meteorological data includes daily maximum temperature data, daily minimum temperature data, relative humidity data, diurnal temperature range data, and air quality index data.

[0008] In one possible implementation, the step of inputting the meteorological sequence tensor into a meteorological tower to obtain the corresponding meteorological feature vector, and inputting the encoded biological vector into a biological tower to obtain the corresponding biological feature vector, includes: The meteorological sequence tensor of a preset length is input into the meteorological tower, and the time context information is extracted using a single-layer BiLSTM to obtain the daily meteorological feature vector. Based on the weights corresponding to the time attention layers, the daily meteorological feature vectors are used to obtain the corresponding meteorological feature vectors. The encoded biological vectors are input into the biological tower, and after processing by a multilayer perceptron and a normalization layer, the corresponding biological feature vectors are obtained.

[0009] In one possible implementation, the step of inputting the meteorological sequence tensor into the exposure layer to obtain the daily exposure scalar sequence corresponding to each pathogen, and determining the exposure vector corresponding to the combination of all pathogens based on the daily exposure scalar sequence of each pathogen, includes: The meteorological sequence tensor is input into the exposure layer to map the daily exposure scalar for each pathogen; For each pathogen, the daily exposure scalar of the pathogen is time-series modeled based on the corresponding LSTM unit, and the daily exposure scalar is weighted and accumulated based on a gating mechanism to obtain the accumulated daily exposure sequence for each pathogen. The cumulative daily exposure sequences of all pathogens are combined to obtain the corresponding exposure vector.

[0010] In one possible implementation, determining the corresponding intermediate gating feature based on the meteorological feature vector and the biological feature vector, and determining the corresponding gating fusion feature based on the intermediate gating feature, includes: The meteorological feature vector and the biological feature vector are concatenated and input into the shared gating backbone network to obtain the corresponding intermediate gating features; By linearly fusing the intermediate gating features with the gating coefficients of each pathogen, the gating fusion features of each pathogen are obtained.

[0011] In one possible implementation, the pet disease early warning model further includes an interpretation interface, and correspondingly, the method further includes: The relative contributions of the weather tower, the biological tower, and the exposure layer in the task head are evaluated using the interpretation interface for each pathogen-related task.

[0012] A second aspect of this disclosure provides a pet disease early warning device, the device being applied to a trained pet disease early warning model, the pet disease early warning model including a weather tower, a biological tower, an exposure layer, and a task head network, the device comprising: The data acquisition module is used to acquire raw meteorological data and raw biological data. The raw biological data is acquired by biological data collection terminals distributed at multiple clinics and uploaded to the server with timestamps and geographic location identifiers. The preprocessing module is used to preprocess the raw meteorological data and the raw biological data respectively, and obtain the meteorological sequence tensor and the encoded biological vector accordingly. The feature extraction module is used to input the meteorological sequence tensor into the meteorological tower to obtain the corresponding meteorological feature vector, and input the encoded biological vector into the biological tower to obtain the corresponding biological feature vector; The exposure determination module is used to input the meteorological sequence tensor into the exposure layer to obtain the daily exposure scalar sequence corresponding to each pathogen, and to determine the exposure vector corresponding to the combination of all pathogens based on the daily exposure scalar sequence of each pathogen. The feature fusion module is used to determine the corresponding intermediate gating features based on the meteorological feature vector and the biological feature vector, and to determine the corresponding gating fusion features based on the intermediate gating features; The result generation module is used to concatenate the gating fusion features with the exposure vector and input them into the task head network to obtain the corresponding infection risk probability, so as to provide early warning of pet diseases.

[0013] A third aspect of this disclosure provides an electronic device comprising: At least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the pet disease early warning method described herein.

[0014] A fourth aspect of this disclosure provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the pet disease early warning method described in this disclosure.

[0015] This disclosure discloses a pet disease early warning method, which includes acquiring raw meteorological data and raw biological data; preprocessing the raw meteorological and biological data separately to obtain meteorological sequence tensors and encoded biological vectors; inputting the meteorological sequence tensors into a meteorological tower to obtain corresponding meteorological feature vectors, and inputting the encoded biological vectors into a biological tower to obtain corresponding biological feature vectors; inputting the meteorological sequence tensors into an exposure layer to obtain daily exposure scalar sequences for each pathogen, and determining the exposure vector corresponding to the combination of all pathogens based on the daily exposure scalar sequences for each pathogen; determining corresponding intermediate gating features based on the meteorological and biological feature vectors, and determining corresponding gating fusion features based on the intermediate gating features; concatenating the gating fusion features with the exposure vectors and inputting them into a task head network to obtain the corresponding infection risk probability for early warning of pet diseases. The acquired raw meteorological and biological data are strictly processed separately at the data level, providing a corresponding data foundation for the subsequent dual-tower structure. A separate weather tower is used to process the meteorological sequence tensor obtained from the preprocessing of raw meteorological data, and a separate biological tower is used to process the encoded biological vector obtained from the preprocessing of raw biological data. Neither the input layer nor the intermediate layer is mixed, so that the corresponding weather features and biological features are separated at the structural level. This prevents information from different sources from being mixed in the same network and affecting the final prediction results, thereby improving the prediction accuracy of the model.

[0016] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0017] The above and other objects, features, and advantages of this disclosure will become readily apparent from the following detailed description of exemplary embodiments, taken in conjunction with the accompanying drawings. Several embodiments of this disclosure are illustrated in the drawings by way of example and not limitation, in which: In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts.

[0018] Figure 1 This illustration shows the implementation flow of a pet disease early warning method according to an embodiment of the present disclosure. Figure 1 ; Figure 2 This is a schematic diagram of the data uploading implementation device in a pet disease early warning method according to an embodiment of the present disclosure; Figure 3 A schematic diagram of a pet disease early warning device according to an embodiment of the present disclosure is shown; Figure 4 A schematic diagram of the composition structure of an electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0019] To make the objectives, features, and advantages of this disclosure more apparent and understandable, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0020] This disclosure provides a method for early warning of pet diseases, such as... Figure 1 As shown, this method is applied to a trained pet disease early warning model, which includes a weather tower, a biological tower, an exposure layer, and a task head network. The method includes: S101. Obtain raw meteorological data and raw biological data. The raw biological data is obtained by biological data collection terminals distributed at multiple clinics and uploaded to the server with timestamps and geographic location identifiers. In this step, the raw meteorological data includes multi-dimensional meteorological data from several days prior to the pet's veterinary visit, used to characterize the specific conditions of multi-dimensional factors such as temperature, humidity, and air quality during the period leading up to the visit. The raw biological data includes environmental and biological data. Environmental data characterizes the pet's living environment and the current season, while biological data characterizes fixed characteristics such as the pet's age, sex, and breed, serving as the data foundation for subsequent predictions of pet disease risk. During the model training phase, detection data is also acquired and used as prediction labels. This detection data includes positive and negative pathogen detection data; a negative result indicates no infection with the corresponding pathogen, and a positive result indicates infection with the corresponding pathogen, used to adjust model parameters so that the pet disease early warning model's predictive ability meets practical needs. It should be noted that, as... Figure 2As shown, the raw biological data is acquired by biological data collection terminals distributed across multiple treatment sites, including portable nucleic acid testing modules, which work in conjunction with the accompanying APP to achieve the purpose of data collection. The data is then uploaded to the server with timestamps and geolocation identifiers for centralized data storage.

[0021] S102. Preprocess the raw meteorological data and raw biological data respectively to obtain the meteorological sequence tensor and the encoded biological vector. In this step, the acquired raw meteorological and biological data are preprocessed separately. The preprocessed meteorological data yields a corresponding meteorological sequence tensor, and the preprocessed biological data yields a corresponding encoded biological vector. The preprocessing of the raw meteorological data includes missing value imputation and handling of extreme outliers. Specifically, when missing values ​​exist in the raw meteorological data, they are first imputed by returning to the meteorological data source based on date and location. If returning to the source is unavailable, the average observation values ​​of city-level data within a preset number of days before and after the visit are used for imputation. For extreme outliers, they are first verified and repaired by returning to the source. If repair is impossible, the average observation values ​​of nearby days from city-level data within the same API data source are used as a substitute. If still unavailable, the data is truncated according to a preset quantile threshold or physically reasonable upper and lower limits. The imputed and truncated meteorological data are then standardized to form a fixed-length meteorological sequence tensor for subsequent meteorological tower processing. In one scenario, raw meteorological data from 14 days prior to a medical visit is collected, and correspondingly, after preprocessing, a meteorological sequence tensor is obtained.

[0022] It should also be noted that after preprocessing the raw biological data, corresponding encoded biological vectors are obtained. This means that the acquired raw biological data is uniformly encoded according to preset rules for subsequent biological tower processing. In other words, the boundaries between the two types of features are distinguished at the data level, and the data is strictly differentiated during the input stage, providing the corresponding data foundation for the subsequent dual-tower structure and channel-level contribution decomposition.

[0023] S103. Input the meteorological sequence tensor into the meteorological tower to obtain the corresponding meteorological feature vector, and input the encoded biological vector into the biological tower to obtain the corresponding biological feature vector. In this step, based on the dual-tower structure consisting of a meteorological tower and a biological tower, the corresponding meteorological feature vectors and biological feature vectors are determined respectively. The meteorological feature vector is obtained by feeding the 5-dimensional meteorological sequence tensor from the 14 days prior to the visit into the meteorological tower, and the corresponding biological feature vector is obtained by feeding the encoded biological vector into the biological tower. It should be noted that the meteorological tower and biological tower in this step are not used interchangeably, either in the input layer or the intermediate layer. That is, the model structure explicitly distinguishes between the short-term meteorological channel and the static biological / environmental channel. This ensures that weather information and individual biological information, such as individual susceptibility, are separated at the structural level, avoiding the mixing of information from different sources in the same network, which could lead to interpretation difficulties and generalization risks.

[0024] S104. Input the meteorological sequence tensor into the exposure layer to obtain the daily exposure scalar for each pathogen, and determine the exposure vector corresponding to the combination of all pathogens based on the daily exposure scalar for each pathogen. In this step, an exposure layer is introduced after the weather tower. That is, the weather sequence tensor is processed by the exposure layer before being sent to the subsequent task header to obtain the daily exposure scalar for each pathogen. It should be noted that the pathogens in this step include FHV, FCV, and IAV. After obtaining the weather sequence tensor, the daily weather vectors corresponding to the weather sequence tensor are input into the exposure layer. That is, for each time step, the multidimensional weather vectors are mapped to the daily exposure scalars corresponding to each pathogen, forming an independent daily exposure sequence for each pathogen.

[0025] Furthermore, the daily exposure sequences for each pathogen are weighted and accumulated over time to obtain the cumulative exposure R for each pathogen. d R of all pathogens d After integration, the corresponding exposure vector R is obtained.

[0026] S105. Determine the corresponding intermediate gating features based on meteorological feature vectors and biological feature vectors, and determine the corresponding gating fusion features based on the intermediate gating features; In this step, meteorological feature vectors and biological feature vectors are concatenated and input into a shared gating backbone network to obtain corresponding intermediate gating features. An independent gating head is then configured for each pathogen to output corresponding gating coefficients. These gating coefficients, ranging from 0 to 1, characterize whether the pathogen is more dependent on the meteorological or biological tower. Based on the fusion of the gating coefficients and intermediate gating features, the corresponding gating fusion features are obtained.

[0027] S106. After concatenating the gating fusion features with the exposure vector, input the data into the task head network to obtain the corresponding infection risk probability, so as to provide early warning of pet diseases.

[0028] In this step, the obtained gated fusion features and exposure vectors are concatenated and input into the task head network corresponding to the pathogen, thereby outputting the corresponding infection risk probability. In addition, while obtaining the infection risk probability of each pathogen, interpretability analysis is performed using the model's internal time kernel and tower-level gating weights. Accordingly, the daily exposure sequence is multiplied by the corresponding time kernel and normalized to obtain the relative contribution of each day before medical visit to the current risk, thus identifying the target weather conditions from the day before medical visit as the key exposure window for this risk. On the other hand, the input dimensions are split using the first-layer weights of the task head, and the contribution ratios of the weather channel, individual channel, and cumulative exposure to the risk are calculated by combining the meteorological feature vector, biological feature vector, and exposure vector of the current sample. Accordingly, by integrating the above time dimension and module dimension information with the pet's biological characteristics, a structured risk profile is generated, and a pre-set template is automatically filled to form the corresponding warning text, thereby providing early warning for pet diseases.

[0029] In one possible implementation, the raw biological data includes age data, sex data, habitat data, seasonal data, and species data. Correspondingly, the raw meteorological data and raw biological data are preprocessed to obtain a meteorological sequence tensor and an encoded biological vector, including: Construct a weather sequence tensor of a preset length based on raw meteorological data; The age data is normalized to obtain normalized age data; Gender data, living environment data, seasonal data, and variety data are mapped to integer indices to obtain indexed multi-source data; The normalized age data and the indexed multi-source data are encoded to obtain the encoded biological vectors.

[0030] In this embodiment, for the pet currently seeking medical attention, the raw biological data includes the pet's age, sex, living environment, breed, and the corresponding current season. Specifically, age data represents the pet's current age; sex data includes whether it is male or female; the biological environment data includes information such as whether it is kept indoors, semi-indoors, or outdoors; breed data includes the corresponding pet breed such as American Shorthair, British Shorthair, or domestic shorthair; and the season data includes information such as spring, summer, autumn, and winter to represent the current season.

[0031] It should be noted that when preprocessing the raw biological data to determine the corresponding encoded biological vectors, age data is normalized by converting it to months. Missing ages are imputed based on the mean of the training set, and the converted months are linearly scaled to the [0,1] interval to obtain normalized age data. In addition, for discrete data, including gender, living environment, season, and breed data, indexing is performed. Gender data is mapped to 0 / 1, and biological environment data is mapped differently depending on the situation: indoor pets are mapped to 0, semi-indoor pets to 1, and outdoor pets to 2. Breed data is based on a predefined breed list, mapping the current pet's breed to an integer index. Seasonal data is mapped from 0 to 3 for spring, summer, autumn, and winter, with the specific mapping result determined based on the current season.

[0032] Furthermore, the aforementioned integer indices are converted into corresponding vectors in the model using one-hot encoding. These vectors are then input into the MLP encoding corresponding to the normalized age data to obtain encoded biological vectors. The resulting encoded biological vectors are used only for subsequent biological towers.

[0033] In one possible implementation, the raw meteorological data includes daily maximum temperature data, daily minimum temperature data, relative humidity data, diurnal temperature range data, and air quality index data.

[0034] In this embodiment, the raw meteorological data includes five-dimensional short-time series meteorological data for the 14 days prior to the visit, specifically including daily maximum temperature data, daily minimum temperature data, daily relative humidity data, daily diurnal temperature range data, and daily air quality index (AQI) for the 14 days prior to the visit, which are denoted as sequences accordingly:

[0035] in, For time indexing, This indicates the date of the medical visit. sky( The day before the visit, (14 days prior to the visit) For the first Weather feature vectors Its five components correspond to the daily maximum temperature, daily minimum temperature, daily relative humidity, daily diurnal temperature range, and daily air quality index (AQI) in that order. This represents a 14-day short-term meteorological sequence arranged in chronological order.

[0036] In this embodiment, the weather sequence tensor is input to the corresponding weather tower during subsequent processing, and is distinguished from the encoded biological vector at the data level to avoid interference between different data and affect the accuracy of the prediction results.

[0037] It should be noted that when training the pet disease early warning model, for the original training samples, each sample corresponds to one animal's medical visit record, including meteorological training data and biological training data. The meteorological training data consists of short-term meteorological data for the 14 days prior to the medical visit, including daily maximum and minimum temperatures, relative humidity, diurnal temperature range, and air quality index. The biological training data includes age training data, sex training data, living environment training data, seasonal training data, and breed training data. All of the above biological training data is structured and encoded in the same way as in the application phase, resulting in encoded training biological vectors. Furthermore, during training, positive / negative labels are constructed for various upper respiratory tract pathogens. The labels correspond to the test results at the time of symptom onset and visit to the hospital, not the day of infection. That is, 0 / 1 labels are retained for FHV, FCV, and IAV, respectively, serving as the outputs of the three tasks in multi-task learning. Subsequently, the training set (train), test set (test), and validation set (val) are hierarchically divided. Specifically, a comprehensive hierarchical label is first constructed, corresponding to:

[0038] During training, the test set is first divided into two groups according to a 70%:15%:15% ratio. Then, the remaining samples are used to divide the data into training and validation sets, following the stratification process. label Stratified sampling was used to ensure that the positive / negative combinations of the three pathogens were distributed similarly across the three sets of data.

[0039] During training, for each training sample, 5-dimensional weather features are extracted for 14 days in the order of the previous 1 to the previous 14 days, forming a meteorological sequence tensor with the shape [N, 14, 5]. It should also be noted that for the training data of each weather dimension, if missing values ​​exist, they are first filled in by retrieving data from the meteorological data source according to date and location; if retrieving data from the source is unavailable, the average observations from the same API data source within a preset number of days before and after the visit are used for filling; if these values ​​are still unavailable and result in virtually no effective observations for that dimension within the month of the visit, the corresponding visit record is marked as a sample with severely missing meteorological information and is not included in the modeling dataset. For extreme outliers, they are first verified and repaired by retrieving data from the source; if repair is not possible, the average observations from the same API data source on nearby days are used as a substitute; if these values ​​are still unavailable, they are truncated according to a preset quantile threshold or a physically reasonable upper or lower limit to reduce the interference of a single extreme event on the model. For example, a pet hospital located in Hangzhou requires meteorological data for the 14 days prior to a visit. If humidity data for the 5th day prior to the visit is missing, it will be retrieved from the meteorological data source by date and location. If retrieval fails, the average humidity value of several days before and after the same API data source will be used to complete the data. If it is further detected that there are no valid humidity observations for the entire month of July 2024 at this station, and the monthly average cannot be calculated, the relevant medical records in July 2024 will be marked as samples with severely missing meteorological information and will not be included in the modeling dataset of this embodiment, or the data for that month will be re-collected after changing the meteorological data source before being included in the modeling.

[0040] During training, z-score standardization is performed based on the training set statistics. The mean and standard deviation are calculated for each of the five weather dimensions on the training set. The training, value, and test subsets are standardized using the training set parameters. The corresponding formulas are as follows:

[0041] in, This represents the original meteorological characteristic values ​​to be standardized. Represents the standardized eigenvalues; and These represent the mean and standard deviation of this meteorological dimension calculated on the training set, respectively. It is worth noting that the training, validation, and test sets all use the same dataset (...). , Standardization is performed to prevent data leakage. Furthermore, after standardization, missing or outlier values ​​are replaced to ensure input stability. The obtained meteorological sequence training tensor and the encoded training biological vectors are then input into the meteorological and biological towers for training.

[0042] In one possible implementation, a meteorological sequence tensor is input into a meteorological tower to obtain the corresponding meteorological feature vector, and the encoded biological vector is input into a biological tower to obtain the corresponding biological feature vector, including: A meteorological sequence tensor of a preset length is input into the meteorological tower, and a single-layer BiLSTM is used to extract the time context information to obtain the daily meteorological feature vector. Based on the weights corresponding to the time attention layers, the daily meteorological feature vectors are obtained. The encoded biological vectors are input into the biological tower, and after processing by a multilayer perceptron and a normalization layer, the corresponding biological feature vectors are obtained.

[0043] In this embodiment, a dual-tower structure of a meteorological tower and a biological tower is used to separately encode short-term meteorological data from the 14 days prior to medical visit, along with static characteristics such as animal age, sex, rearing environment, season, and breed. Based on this, pathogen perception lag effect distribution LSTM units (EpiLSTM) are introduced to model the distribution of pathogen perception lag effects for three pathogens: FHV, FCV, and IAV, respectively, to explicitly characterize the delayed and cumulative effects of pre-medical visit meteorological exposure on the current risk of disease on illness.

[0044] Among them, "lag effect" refers to the fact that the impact of external exposures such as meteorology on the risk of disease is not immediate, but may gradually manifest itself in several lag days after the observation date; "lag effect window" refers to a preset time range that goes back from the observation date (e.g., 14 days before the visit); "lag effect distribution (lag weight / time kernel)" refers to the weight distribution of the risk contribution of each lag day within the window, which is used to weight and accumulate the daily exposure to obtain the pathogen-specific lag effect cumulative exposure amount, and to identify high contribution lag intervals for interpreting the output.

[0045] Specifically, the data input to the weather tower is a weather sequence of length T=14, corresponding to: The five dimensions are: maximum temperature, minimum temperature, relative humidity, diurnal temperature range, and AQI, all standardized using z-scores based on the training set mean and standard deviation. Temporal context information is extracted using a single-layer bidirectional LSTM from the weather tower; the corresponding formula is:

[0046] in, This indicates that the bidirectional LSTM is at time step The hidden state vector is used to represent the state before the visit. The temporal characteristics of meteorological sequences after considering contextual information. , To hide the state dimension. It should be noted that, to depict the fact that not every day within the 14-day period contributes equally to the risk of disease, a time attention layer is added to the BiLSTM output, assigning learnable weights to different time steps. The corresponding formula is:

[0047] This corresponds to the meteorological feature vector w:

[0048] in, For time step Attention score, used to measure the features of that time step. The importance of overall risk characterization; It is a linear transformation matrix. For learnable attention vectors, It is a non-linear activation function; The attention weights after softmax normalization satisfy... .based on The meteorological feature vector is obtained by weighted summation of the hidden states at each time step. ,in , These are the feature dimensions after aggregation.

[0049] It should be noted that the WeatherTower in this embodiment only processes meteorological time series and does not involve the processing of any biological or static environmental features, thus preventing interference from biological or static environmental features.

[0050] It should also be noted that the encoded biological vectors are input into the BioTower, where the corresponding biological feature vectors are extracted through a multi-layer perceptron mechanism and a normalization layer. The input data for the BioTower is normalized age data. The discrete features after one-hot encoding, namely, discrete features such as gender, living environment, season, and species, are respectively g, e, s, and r after one-hot encoding.

[0051] Furthermore, the normalized age data is input into a small MLP for processing, and the corresponding formula is:

[0052] in, This represents the normalized age scalar. This represents an age coding network used to map scalar age to a low-dimensional continuous representation; The age embedding vector; The dimension of the age embedding vector.

[0053] Furthermore, the age data processed by the MLP is concatenated with the encoded discrete features to obtain the concatenated feature m, and the corresponding formula is:

[0054] It should be noted that the obtained spliced ​​feature m is processed through multiple MLP layers and normalization layers to obtain the corresponding biometric feature vector b, and the corresponding formula is: .

[0055] in, This represents the bio-tower coding network.

[0056] The biotower in this embodiment only involves the processing of biological characteristics or static environmental characteristics, and does not involve any meteorological sequence. That is, the meteorological tower and the biotower are strictly separated. Structurally, the meteorological information channel and the individual susceptibility channel are distinguished to prevent the interpretation difficulties and generalization risks caused by the mixed processing of information from different sources in the same network in related technologies.

[0057] In one possible implementation, a meteorological sequence tensor is input into the exposure layer to obtain a daily exposure scalar sequence for each pathogen, and an exposure vector corresponding to the combination of all pathogens is determined based on the daily exposure scalar sequence for each pathogen, including: Meteorological sequence tensors are input into the exposure layer to map daily exposure scalars for each pathogen; For each pathogen, the daily exposure scalar of the pathogen is time-series modeled based on the corresponding LSTM unit, and the daily exposure scalar is weighted and accumulated based on a gating mechanism to obtain the accumulated daily exposure sequence for each pathogen. The cumulative daily exposure sequences of all pathogens are combined to obtain the corresponding exposure vector.

[0058] In this embodiment, the exposure layer is an EpiLSTM exposure layer. Based on the epidemiological consensus that cumulative exposure over several days prior to medical visit is more critical for disease development, the EpiLSTM layer models the exposure process within a lag window for each of the three pathogens. For each time step t and pathogen task d, the 5-dimensional meteorological vector is first linearly mapped to a daily exposure input scalar, where the pathogen task... The corresponding daily exposure input: in, For pathogen mission The corresponding meteorological feature linear mapping weight vector, This is the bias term for the task, used to characterize the baseline exposure level under baseline meteorological conditions, thereby obtaining the scalar sequence corresponding to each pathogen task.

[0059] Furthermore, for each pathogen task d, based on the constructed independent EpiLSTM units, the one-dimensional exposure sequence is processed. For one-dimensional time series modeling, the corresponding formula is:

[0060] in, Indicates the task Optimized epidemiological LSTM unit; and They are time steps The hidden state vector and memory unit state vector are used to capture the task. Temporal dependence and cumulative effects in exposure sequences; , This refers to the state at the previous time step.

[0061] It should be noted that the internal control parameters address the pathogen learning lag effect, i.e., which time points are more sensitive to the eventual onset of disease.

[0062] Furthermore, the EpiLSTM exposure layer performs weighted accumulation of the exposed inputs in the time dimension to obtain a single scalar of accumulated exposure for each task, with the corresponding formula as follows:

[0063] in, Implemented by a dedicated reading layer for a single task, it essentially corresponds to the weighted sum of the exposure contributions of each day within the hysteresis window, used to hide the time-series state sequence. Integrate into a single scalar indicator This refers to the cumulative exposure amount of a specific pathogen within the observation window.

[0064] Accordingly, the cumulative daily exposure sequences of all pathogens are combined, that is, the cumulative exposure amounts of the three tasks are combined to form the exposure vector R, and the corresponding formula is as follows:

[0065] In this embodiment, an EpiLSTM exposure layer is introduced after the weather tower. During training, the lag effect of pathogen perception is explicitly modeled as a cumulative exposure. After obtaining the daily weather vector, it is not directly fed into the task header. Instead, an EpiLSTM exposure layer is constructed first. For each time step t, the 5-dimensional weather vector is mapped to the daily exposure scalar corresponding to each pathogen, forming an independent daily exposure sequence for each pathogen. Then, an independent LSTM unit is configured for each pathogen to perform time-series modeling on the one-dimensional exposure sequence. The weights of different lag days are automatically learned through a gating mechanism, which is the lag effect time kernel of the pathogen. Finally, a weighted accumulation is performed in the time dimension to obtain the pathogen-specific lag effect cumulative exposure. and the pathogens The exposure variable R is composed of these variables.

[0066] It should be noted that, unlike related technologies that use GBDT and conventional LSTM to implicitly learn temporal relationships, this embodiment explicitly models daily exposure sequences and gated accumulation based on the EpiLSTM exposure layer. This allows the model to automatically learn the cumulative amplification effects of environmental factors such as sustained low temperature, high humidity, and high AQI on different pathogens within the hysteresis window, and through... This creates exposure characteristics with clear physical meaning. In this embodiment, the exposure layer explicitly separates the daily exposure input, the lag effect time kernel, and the cumulative exposure output. This not only characterizes the complex relationship between exposure and lag but also directly provides pathogen-level exposure quantities with clear epidemiological implications.

[0067] In one possible implementation, corresponding intermediate gating features are determined based on meteorological feature vectors and biological feature vectors, and corresponding gating fusion features are determined based on the intermediate gating features, including: Meteorological feature vectors and biological feature vectors are concatenated and input into a shared gating backbone network to obtain the corresponding intermediate gating features; By linearly fusing the intermediate gating features with the gating coefficients of each pathogen, the gating fusion features of each pathogen are obtained.

[0068] In this embodiment, after obtaining the meteorological feature vector w, the biological feature vector b, and the exposure vector R, the contributions of the meteorological channel and the biological channel are adaptively weighted according to the task through a tower-level gating mechanism.

[0069] Specifically, the meteorological feature vector w and the biological feature vector b are concatenated and input into a shared gating backbone network to obtain the corresponding intermediate gating features. The corresponding formula is:

[0070] Furthermore, for each pathogen task d, an independent gating head is set to output a scalar gating coefficient, and the corresponding formula is:

[0071] In the formula Mapping the score to the Sigmoid function Gating coefficients are obtained from the interval .

[0072] Based on the aforementioned gating coefficients, for each pathogen task d, the features corresponding to the meteorological tower and the features corresponding to the biological tower are linearly fused according to the tower-level gating coefficients to obtain the gated fused features. The corresponding formula is:

[0073] It should be noted that the above The closer to 1, the more the identification of pathogen d depends on the weather pattern; the closer to 0, the more the identification of pathogen d depends on individual susceptibility and the rearing environment.

[0074] Furthermore, the aforementioned gating fusion features After being concatenated with the exposure vector d, it serves as the input data for the task header to predict the corresponding infection risk probability. Specifically, for each pathogen task d... Extract the corresponding exposure scalar from the exposure vector R. The tower-level gating fusion features are concatenated with the exposure scalar to obtain the corresponding task input vector. The corresponding formula is as follows:

[0075] Furthermore, the task input vector is fed into a two-layer fully connected network, including ReLU activation, LayerNorm, and Dropout, to obtain the log-odd probability of task d. The corresponding formula is Furthermore, regarding Applying the Sigmoid function yields the infection risk probability of pathogen d. .

[0076] Therefore, this embodiment uses a tower-level gating mechanism to adaptively allocate the weights of the meteorological tower and biological tower according to the pathogen, and combines them with the exposure amount for discrimination. After obtaining the meteorological feature vector w, biological feature vector b, and exposure vector R, instead of concatenating the three together and sharing a single fully connected layer as used in related technologies, a tower-level gating module is first constructed. The meteorological feature vector w and biological feature vector b are concatenated and input into a shared gating backbone network to obtain intermediate gating features. Then, an independent gating head is configured for each pathogen, with a gating coefficient of 0-1 as input to indicate whether the pathogen is more dependent on the meteorological tower or the biological tower. The meteorological feature vector w and biological feature vector b are then linearly fused according to the gating coefficient to obtain a task-level fusion representation. This representation is then concatenated with the corresponding task scalar and input into the task head network of the pathogen to output the corresponding infection risk probability.

[0077] This embodiment's structure enables joint modeling of meteorological models, individual susceptibility, and lag effect cumulative exposure, while maintaining a controllable network size. This allows for effective prediction of multi-pathogen respiratory infection risks. Compared to related technologies that simply share a single fusion method for multiple pathogens, this embodiment uses a tower-level gating structure combined with exposure data splicing. This allows different pathogens to automatically learn differentiated discrimination strategies based on weather-driven or individual susceptibility, reducing inter-task interference and providing quantifiable evidence for subsequent task-based analysis of the relative contributions of weather and individual channels.

[0078] In one possible implementation, the pet disease early warning model also includes an interpretation interface; correspondingly, the method also includes: The relative contributions of the weather tower, biological tower, and exposure layer in the task head are evaluated using the interpretation interface for each pathogen-related task.

[0079] It should be noted that the dual-tower structure in this embodiment supports the `explain()` interface, which is used to form time-dimensional risk contributions and module-level contributions based on internal features, and then convert them into user-understandable natural language text to achieve early warning. For the time-dimensional risk contribution, it mainly characterizes which days of weather are most dangerous. In the explanation interface, for each task `d` and sample `i`, the daily exposure `expo[:, :, d]` within the exposure layer and the hysteresis time kernel are used. Multiplying these together yields the original risk contribution for each day, which is:

[0080] in, Indicates the sample index. Indicates the pathogen task index. Indicates the first day before the medical visit sky; Indicates sample In the Day, in response to the task The "daily exposure scalar"; Indicates task The lag time kernel in the first The weight of the heavens; Indicates the first Heaven's mission The original contribution value of the risk; This is used to truncate the contribution to non-negative values, making it easier to interpret as "contribution intensity".

[0081] Furthermore, within 14 days Normalization, the corresponding formula is as follows:

[0082] This yields the target task contribution daycontrib[task], corresponding to the shape [B, 14], where the above... Characterizes the relative contribution of day t before the visit to the risk of pathogen d on that day. To prevent extremely small constants with a denominator of 0, analysis is performed. The maximum number of days can be used to identify the corresponding results. For example, the continuous low temperature and high humidity 3-5 days before the visit is the key period for this high-risk situation.

[0083] Furthermore, module-level contributions are determined, the interface is interpreted for each task d, and the relative contributions of the weather tower, biotower, and exposure in the task header are roughly evaluated. Specifically, the first-layer fully connected weight matrix of the task header is extracted and divided into weight blocks for the weather vector according to the input dimension. Weight blocks for biological vectors and weighting of exposure For each sample, using the corresponding input vectors w, b, and R... d Performing a linear transformation and taking the average of the absolute values, the result is: , , Furthermore, after normalizing the proportions, the corresponding formulas are as follows: ， ,

[0084] in, These represent the original contribution intensities of the meteorological channel, biological channel, and exposure channel, respectively; normalization yields... Used to represent the three channels for the task The relative contribution percentage, satisfying .

[0085] Corresponding target module-level contribution:

[0086] This indicates that the main drivers of this high risk are overall weather patterns, individual physical condition / environment, or concentrated exposure within a lag window.

[0087] It should also be noted that this embodiment converts the interpretation results into a structured risk profile. For each high-risk disease task, such as... Then, based on the contribution of the target task Identify the days with the highest contribution and merge them into corresponding continuous time windows, such as 6-3 days before the visit. Review the original meteorological data corresponding to these time periods and summarize pattern labels such as "persistent low temperature," "high humidity," "drastic temperature fluctuations," and "high AQI." Based on the contribution of the target module level... To identify the primary source of risk, the highest exposure level indicates that the risk is mainly caused by high exposure within the lag window prior to the observation date. Based on the biotatter input characteristics, age is discretized into kitten / adult / senior cat, and the environment is mapped to indoor / semi-indoor / outdoor. Breed, season, etc., are directly converted into Chinese descriptions to generate profiles such as "4 months old, outdoor living, British Shorthair kitten, winter".

[0088] Finally, the structured explanatory information is generated into user-friendly, natural language text descriptions that pet owners or veterinarians can understand. This can be achieved using pre-set templates and rules. The templates include risk conclusions, explanations of key causes, vulnerable individual profiles, and actionable protective recommendations. For example: Risk conclusion: "The model assesses that the current risk of infection with [disease name] is relatively high (approximately [risk probability])". Key reason explanation: "If the exposure rate is the highest and the high contribution window is "day 6 to 3, low temperature and high humidity", the main reason is that there was a continuous low temperature and high humidity weather process 6 to 3 days before the visit, and the exposure layer judged that the environmental conditions during this period may promote the transmission conditions." Susceptible individual profile: "At the same time, the animal is a [breed] with [age tag] and [living environment tag], making it more susceptible to infection with the pathogen under the aforementioned weather conditions"; for example, an older dog with a risk score of 0.8 for outdoor activities is considered a high-risk group; Actionable protective recommendations: "It is recommended that during future periods of similar continuous cooling and high humidity, people of [age tag] reduce their outdoor activity time, pay attention to keeping warm and disinfecting the environment, and complete vaccinations and regular check-ups under the guidance of a veterinarian."

[0089] Therefore, this embodiment can output a readable and usable warning result to the user, rather than just giving a black box probability that is difficult to interpret.

[0090] In summary, a trained model is used for risk prediction, outputting corresponding explanations and warnings. For the individual seeking medical attention, a 5-dimensional meteorological sequence and corresponding static features for the 14 days prior to the visit are constructed according to the same rules as during the training phase. After standardization and encoding, these are input into the meteorological tower and biological tower respectively, yielding meteorological feature vector w and biological feature vector b. Simultaneously, the daily meteorological feature vectors are mapped to the daily exposure sequences of each pathogen, and weighted and accumulated in the time dimension through the corresponding EpiLSTM exposure layer to obtain the lag effect exposure accumulation vector R. Subsequently, the weights of the meteorological tower and biological tower are calculated for each pathogen using tower-level gating. The meteorological feature vector w and biological feature vector b are fused according to the task and then compared with the corresponding... The data is concatenated and forward-computed using the task head to obtain the infection risk probability or risk level of multiple pathogens such as FHV, FCV, and IAV, achieving simultaneous output of risk prediction results for multiple upper respiratory tract pathogens in a single forward inference. Furthermore, while obtaining the risk probability of each pathogen, interpretability analysis is performed using the EpiLSTM time kernel and tower-level gating weights within the model: On one hand, the daily exposure sequence is multiplied by the corresponding time kernel and normalized to obtain the relative contribution of each day before the visit to the current risk, thereby automatically identifying "persistent low temperature, high humidity, or elevated AQI from day [number] to day [number] before the visit" as the key exposure window for this risk; on the other hand, the input dimensions are split using the first-level weights of the task head, combined with the current sample's w, b, and The system calculates and normalizes the contribution ratios of weather channels, individual channels, and cumulative exposure to risk. It combines the aforementioned time and module dimensions with animal characteristics such as age, rearing environment, breed, and season to generate a structured risk profile. This profile is then automatically populated using a preset template to form a natural language warning text, including risk conclusions, explanations of key causes, descriptions of susceptible individuals, and actionable protective measures. This results in a readable and usable warning, rather than simply providing an inexplicable black-box probability.

[0091] Under the same data conditions, the method disclosed herein can achieve more accurate and stable prediction of multi-pathogen infection risk compared with the models of related technologies. It can finely characterize the differentiated impact of environmental conditions on different pathogens, quantify the contribution ratio of weather channels, individual channels, and cumulative exposure to a certain high risk, and convert it into structured early warning information and natural language prompts that can be directly used for clinical and prevention and control decisions, significantly improving the practicality and interpretability of the model.

[0092] To implement the above method, one example of this application also provides a pet disease early warning device 200, such as... Figure 3 As shown, the device is applied to a trained pet disease early warning model. The pet disease early warning model includes a weather tower, a biological tower, an exposure layer, and a task head network. The device includes: The data acquisition module 201 is used to acquire raw meteorological data and raw biological data. The raw biological data is acquired by biological data collection terminals distributed at multiple clinics and uploaded to the server with timestamps and geographic location identifiers. Preprocessing module 202 is used to preprocess the raw meteorological data and raw biological data respectively, and obtain meteorological sequence tensors and encoded biological vectors accordingly; The feature extraction module 203 is used to input the meteorological sequence tensor into the meteorological tower to obtain the corresponding meteorological feature vector, and input the encoded biological vector into the biological tower to obtain the corresponding biological feature vector; The exposure determination module 204 is used to input the meteorological sequence tensor into the exposure layer to obtain the daily exposure scalar corresponding to each pathogen, and to determine the exposure vector corresponding to the combination of all pathogens based on the daily exposure scalar of each pathogen. The feature fusion module 205 is used to determine the corresponding intermediate gating features based on meteorological feature vectors and biological feature vectors, and to determine the corresponding gating fusion features based on the intermediate gating features; The result generation module 206 is used to concatenate the gating fusion features and the exposure vector, and then input them into the task head network to obtain the corresponding infection risk probability, so as to provide early warning of pet diseases.

[0093] By way of example, this application also provides an electronic device, including: processor; Memory used to store processor-executable instructions; A processor is used to read executable instructions from memory and execute the instructions to implement the methods described above.

[0094] By way of example, this application also provides a computer-readable storage medium storing a computer program for performing the above-described pet disease early warning method.

[0095] Figure 4 A schematic block diagram of an example electronic device 300 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0096] like Figure 4 As shown, device 300 includes a computing unit 301, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 302 or a computer program loaded from storage unit 308 into random access memory (RAM) 303. The RAM 303 may also store various programs and data required for the operation of device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via bus 304. Input / output (I / O) interface 305 is also connected to bus 304.

[0097] Multiple components in device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of monitors, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0098] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as a pet disease early warning method. For example, in some embodiments, a pet disease early warning method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed on device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by the computing unit 301, one or more steps of a pet disease early warning method described above may be performed. Alternatively, in other embodiments, the computing unit 301 may be configured to perform a pet disease early warning method by any other suitable means (e.g., by means of firmware).

[0099] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0100] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0101] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0102] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0103] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0104] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0105] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0106] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this disclosure, "a plurality of" means two or more, unless otherwise explicitly specified.

[0107] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A method for early warning of pet diseases, characterized in that, The method is applied to a trained pet disease early warning model, which includes a weather tower, a biological tower, an exposure layer, and a task head network. The method includes: Acquire raw meteorological data and raw biological data, wherein the raw biological data is acquired by biological data acquisition terminals distributed at multiple clinics and uploaded to the server with timestamps and geographic location identifiers; The raw meteorological data and the raw biological data are preprocessed respectively to obtain meteorological sequence tensors and encoded biological vectors. The meteorological sequence tensor is input into the meteorological tower to obtain the corresponding meteorological feature vector, and the encoded biological vector is input into the biological tower to obtain the corresponding biological feature vector; The meteorological sequence tensor is input into the exposure layer to obtain the daily exposure scalar sequence for each pathogen, and the exposure vector corresponding to the combination of all pathogens is determined based on the daily exposure scalar sequence for each pathogen. Based on the meteorological feature vector and the biological feature vector, the corresponding intermediate gating feature is determined, and based on the intermediate gating feature, the corresponding gating fusion feature is determined; The gated fusion features are concatenated with the exposure vector and then input into the task head network to obtain the corresponding infection risk probability, so as to provide early warning of pet diseases.

2. The pet disease early warning method according to claim 1, characterized in that, The raw biological data includes age data, sex data, habitat data, seasonal data, and species data. Correspondingly, the preprocessing of the raw meteorological data and the raw biological data yields a meteorological sequence tensor and an encoded biological vector, including: Construct a weather sequence tensor of a preset length based on the raw meteorological data; The age data is normalized to obtain normalized age data; The gender data, living environment data, seasonal data, and variety data are mapped to integer indices to obtain indexed multi-source data; The normalized age data and the indexed multi-source data are encoded to obtain the encoded biological vector.

3. The pet disease early warning method according to claim 2, characterized in that, The raw meteorological data includes daily maximum temperature data, daily minimum temperature data, relative humidity data, diurnal temperature range data, and air quality index data.

4. The pet disease early warning method according to claim 1, characterized in that, The process of inputting the meteorological sequence tensor into the meteorological tower to obtain the corresponding meteorological feature vector, and inputting the encoded biological vector into the biological tower to obtain the corresponding biological feature vector, includes: The meteorological sequence tensor of a preset length is input into the meteorological tower, and the time context information is extracted using a single-layer BiLSTM to obtain the daily meteorological feature vector. Based on the weights corresponding to the time attention layers, the daily meteorological feature vectors are used to obtain the corresponding meteorological feature vectors. The encoded biological vectors are input into the biological tower, and after processing by a multilayer perceptron and a normalization layer, the corresponding biological feature vectors are obtained.

5. The pet disease early warning method according to claim 1, characterized in that, The step of inputting the meteorological sequence tensor into the exposure layer to obtain the daily exposure scalar sequence corresponding to each pathogen, and determining the exposure vector corresponding to the combination of all pathogens based on the daily exposure scalar sequence of each pathogen, includes: The meteorological sequence tensor is input into the exposure layer to map the daily exposure scalar for each pathogen; For each pathogen, the daily exposure scalar of the pathogen is time-series modeled based on the corresponding LSTM unit, and the daily exposure scalar is weighted and accumulated based on a gating mechanism to obtain the accumulated daily exposure sequence for each pathogen. The cumulative daily exposure sequences of all pathogens are combined to obtain the corresponding exposure vector.

6. The pet disease early warning method according to claim 1, characterized in that, The step of determining the corresponding intermediate gating feature based on the meteorological feature vector and the biological feature vector, and determining the corresponding gating fusion feature based on the intermediate gating feature, includes: The meteorological feature vector and the biological feature vector are concatenated and input into the shared gating backbone network to obtain the corresponding intermediate gating features; By linearly fusing the intermediate gating features with the gating coefficients of each pathogen, the gating fusion features of each pathogen are obtained.

7. The pet disease early warning method according to any one of claims 1-6, characterized in that, The pet disease early warning model also includes an interpretation interface; correspondingly, the method also includes: The relative contributions of the weather tower, the biological tower, and the exposure layer in the task head are evaluated using the interpretation interface for each pathogen-related task.

8. A pet disease early warning device, characterized in that, The device is applied to a trained pet disease early warning model, which includes a weather tower, a biological tower, an exposure layer, and a task head network. The device includes: The data acquisition module is used to acquire raw meteorological data and raw biological data. The raw biological data is acquired by biological data collection terminals distributed at multiple clinics and uploaded to the server with timestamps and geographic location identifiers. The preprocessing module is used to preprocess the raw meteorological data and the raw biological data respectively, and obtain the meteorological sequence tensor and the encoded biological vector accordingly. The feature extraction module is used to input the meteorological sequence tensor into the meteorological tower to obtain the corresponding meteorological feature vector, and input the encoded biological vector into the biological tower to obtain the corresponding biological feature vector; The exposure determination module is used to input the meteorological sequence tensor into the exposure layer to obtain the daily exposure scalar sequence corresponding to each pathogen, and to determine the exposure vector corresponding to the combination of all pathogens based on the daily exposure scalar sequence of each pathogen. The feature fusion module is used to determine the corresponding intermediate gating features based on the meteorological feature vector and the biological feature vector, and to determine the corresponding gating fusion features based on the intermediate gating features; The result generation module is used to concatenate the gating fusion features with the exposure vector and input them into the task head network to obtain the corresponding infection risk probability, so as to provide early warning of pet diseases.

9. An electronic device, characterized in that, include: At least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform claim 1. The pet disease early warning method described in any one of the 7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the pet disease early warning method according to any one of claims 1-7.

Images