A method for predicting sow farrowing using vulvar physiological temporal characterization and prototype correction
By extracting the physiological temporal features of the sow vulva using instance segmentation and deep learning techniques, and combining long short-term memory networks and prototype correction mechanisms, the inefficiency and high cost of sow farrowing monitoring are solved, achieving accurate farrowing prediction without contact.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING AGRICULTURAL UNIVERSITY
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-26
AI Technical Summary
Current technologies for monitoring sow farrowing mainly rely on manual observation, which is inefficient and prone to errors. Furthermore, the use of wearable sensors on pigs can cause stress and is costly, making it difficult to achieve low-cost, contactless, and accurate farrowing prediction.
The vulva region of sows is segmented using instance segmentation technology, and the physiological temporal features of the vulva are extracted through a deep learning model. The sow farrowing is predicted by combining long short-term memory network and prototype correction mechanism, and sow farrowing early warning is provided by computer vision technology without contact.
It enables rapid and accurate identification of the vulva area of sows and prediction of farrowing in a pigsty environment, improving the accuracy and efficiency of farrowing prediction and reducing stress on sows and economic costs.
Smart Images
Figure CN121661056B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent pig farming technology, specifically a method for predicting sow farrowing using deep learning technology and instance segmentation algorithms. Background Technology
[0002] In large-scale commercial pig farms, to reduce the probability of newborn piglets being crushed and killed by sows after farrowing, gestation crates are usually used to confine sows before their expected farrowing date. However, this can affect sow welfare. By using cameras to detect the physiological characteristics of the sow's vulva within the farrowing crate, a farrowing warning signal is generated. The gestation crate is then closed, and farm management personnel are notified. This approach ensures the welfare of sows about to farrow and allows farm staff to prepare for farrowing.
[0003] Currently, monitoring of sow farrowing is mainly based on experience-based manual observation, which is not only time-consuming and labor-intensive, but also prone to observational bias. With the development of sensor technology, wearable sensors are used to predict sow farrowing time. However, using wearable sensors on pigs is very challenging because sensors often trigger stress responses, and each sow needs a sensor, resulting in high economic costs. With the development of computer vision technology, non-contact behavioral monitoring based on computer vision has gained attention. Driven by the simultaneous needs of precision animal husbandry and intelligent pig farming, utilizing computer vision technology to quickly and accurately identify the physiological characteristics of the sow's vulva and predict farrowing in a low-cost, non-contact manner has significant practical implications. Under natural conditions, the sow's vulva is used to determine the sow's estrous cycle, reproductive performance, and to prevent farrowing diseases. Veterinary medicine shows that from 7 days before farrowing until the sow is about to give birth, the vulva gradually swells, darkens in color, the labia become red and swollen, loose, and wrinkled with reduced folds, accompanied by a small amount of mucus discharge. Therefore, it is important to propose a method for segmenting the vulva region of sows and extracting physiological features using instance segmentation technology, which is the basis for parturition prediction based on the physiological temporal representation of the vulva. Summary of the Invention
[0004] (a) Technical problems to be solved
[0005] Accurate segmentation of the sow's vulva is a prerequisite for extracting its physiological characteristics. A method is needed to detect and segment the sow's vulva in a farrowing environment to extract physiological changes in the vulva, determine the sow's impending parturition, and obtain parturition prediction results. This invention aims to solve the problem of segmenting the sow's vulva and extracting its physiological temporal characteristics using instance segmentation technology for parturition prediction.
[0006] (II) Technical Solution of the Invention
[0007] To address the aforementioned problems, this invention proposes a complete method for predicting sow farrowing using vulvar physiological temporal representation and prototype correction. It includes the following steps (S1-S6):
[0008] S1. Collect standing image data of sows at the pig farm 7 days before the sow's expected farrowing date and preprocess the data;
[0009] S1-1. Use a camera to capture images of the vulva of a sow in a standing position from a rear-view perspective;
[0010] S1-2. Perform operations such as cleaning, filtering, noise reduction, and image enhancement on the collected sow images to construct a sow vulva image dataset D1;
[0011] S2. Train the instance segmentation deep learning YOLOv8 model to obtain the sow vulva segmentation mask;
[0012] S2-1. Divide the collected sow vulva image dataset D1 into a training set D1_train, a validation set D1_valid, and a test set D1_test in a ratio of 6:2:2.
[0013] S2-2. Perform vulvar mask annotation on the training set D1_train, validation set D1_valid, and test set D1_test respectively. Divide the vulvar region into three sub-regions in the image: left labia, right labia, and clitoris, with categories left_lip, right_lip, and clitoris respectively, and mark the positions of the three sub-regions.
[0014] S2-3. Generate a corresponding txt file with the image name for each labeled image file, containing the target category and the coordinates of all points of the labeled mask. ,in , All values are normalized.
[0015] S2-4. Use the YOLOv8 model to train the labeled training set D1_train, and use the validation set D1_valid to validate it, thus obtaining the vulvar instance segmentation model.
[0016] S2-5. Input the test set D1_test divided in S2-2 into the trained vulvar instance segmentation model and evaluate the performance metrics of the model.
[0017] S2-6. Use the trained vulvar instance segmentation model to identify the vulva image of a sow and obtain the coordinates of all points of the instance segmentation mask for the three sub-regions in the vulva image. And the coordinates of the top-left corner of the bounding rectangle of the three sub-regions. Width of the detection frame and height Combining depth information and depth camera intrinsics Calculate the true coordinates of the sub-region mask after segmentation. ;
[0018] S3. Design a feature extraction network to extract the original physiological features of the sow's vulva area;
[0019] S3-1. Using the instance segmentation mask coordinates obtained in S2-6, calculate the width of the three sub-regions. ,high and area To obtain the morphological characteristics of the vulvar area;
[0020] S3-2. Using OpenCV color space conversion and statistical methods, calculate the mean and standard deviation of each channel (H, S, V) in each sub-region within the three sub-region segmentation mask coordinates, and normalize the generated mean and standard deviation to obtain the color features of the vulva region.
[0021] S3-3. Use the Gray-Level Co-occurrence Matrix (GLCM) method to calculate the texture features of the vulva region;
[0022] S3-4. The extracted morphological, color, and texture features of the vulva are spliced together to construct the original physiological features of the vulva. ;
[0023] S4. Perform temporal modeling of the original physiological characteristics of the vulva to obtain the temporal representation of the physiological characteristics of the vulva of sows;
[0024] S4-1. Input the image of the sow's vulva into the trained YOLOv8 instance segmentation model to obtain the instance segmentation mask coordinates;
[0025] S4-2. Input the segmentation mask coordinates into the feature extraction network model to extract the original physiological features of the sow's vulva region. ;
[0026] S4-3. Repeat steps S4-1 and S4-2 for the vulva images of sows for 7 consecutive days, and output the original physiological features of the vulva for 7 consecutive days to a CSV file, including date, morphological features, color features and texture features;
[0027] S4-4. Using the Long Short-Term Memory (LSTM) network method, time-series modeling is performed on the original physiological features in the CSV file to obtain dynamic features with long- and short-term dependencies, and time-series features are extracted. ;
[0028] S4-5. Use statistical methods to calculate the mean, variance, and skewness of the primitive physiological characteristics of the sow's vulva. ;
[0029] S5. The extracted vulvar features are fused to construct a physiological time-series representation variable for the sow vulva;
[0030] S5-1, Extracted original physiological characteristics Time series characteristics and Perform dimensional transformation to unify them to the same dimension;
[0031] S5-2. Assign learnable parameter weights to the three features, use feature variance as a quality indicator to initialize the weights, and adjust the learnable parameters as the classification model is trained.
[0032] S5-3, Weighted original physiological characteristics Time series characteristics and By splicing the data, we can construct a series of physiological characteristics of the vulva. ;
[0033] S6. Input the vulvar physiological time sequence characterization variables into the prototype calibration classification model to obtain the sow's farrowing status and predict the sow's farrowing.
[0034] S6-1. Perform steps S4 and S5 on all sows to obtain the vulvar physiological time-series characterization variables of all sows and construct dataset D2.
[0035] S6-2. Divide the dataset D2 containing all vulvar physiological time-series representation variables obtained in S5-1 into a training set D2_train, a validation set D2_valid, and a test set D2_test in a ratio of 8:1:1.
[0036] S6-3. Label the farrowing status categories of the training set D2_train, validation set D2_valid, and test set D2_test respectively. Label the characteristic variables of sows that farrow within 12 hours as category 1, and label the characteristic variables of sows that have not farrowed within 12 hours as category 0.
[0037] S6-4. Train the improved prototype network using the labeled training set D2_train. Calculate the similarity between query samples and support samples using an instance-level attention mechanism. Correct category bias through data expansion and prototype correction mechanisms to obtain the sow farrowing prediction results. Validate the model using the validation set D2_valid to obtain a classification model of vulvar physiological time-series representation variables and calculate the probability of belonging to each category.
[0038] S6-5. Input the test set D2_test divided in S6-2 into the trained vulvar physiological time-series representation variable classification model and evaluate the performance index of the model.
[0039] S6-6. Using a classification model of vulvar physiological time sequence characteristics, the sow's parturition status is obtained. When the model classification result indicates that the sow will parturient within 12 hours, a parturition warning is issued.
[0040] The beneficial effects achieved by this invention are as follows:
[0041] This invention applies an instance segmentation algorithm to the segmentation of the sow vulva, enabling the identification and segmentation of the sow vulva in a farrowing environment. It also extracts physiological features of the sow vulva region and uses temporal changes in the physiological characteristics of the sow vulva to predict farrowing. The technology disclosed in this invention provides technical support for farrowing prediction of sows in commercial farrowing environments based on computer vision. It can help pig farmers quickly and accurately grasp the sow's farrowing status and prepare in advance, showing broad application prospects in precision farming and smart animal husbandry. Attached Figure Description
[0042] Figure 1 This is an overall flowchart of the sow farrowing prediction method based on vulvar physiological time-series characterization and prototype correction of the present invention.
[0043] Figure 2 This is a flowchart illustrating the construction of the vulvar physiological time-series characterization variables of the present invention.
[0044] Figure 3 This is a schematic diagram of the vulvar physiological time-series characterization variables of the present invention.
[0045] Figure 4 This is a visualization diagram of the instance segmentation model results of the present invention. Detailed Implementation
[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments. In this embodiment, a method for predicting sow farrowing is provided, which utilizes instance segmentation technology to segment the sow vulva and employs vulvar physiological temporal representation and prototype correction. This method involves collecting and preprocessing image data in the field, using a deep learning algorithm to obtain a sow vulva segmentation mask, extracting the physiological temporal representation of the sow vulva through a feature extraction network and constructing variables, and classifying these vulvar physiological temporal representation variables to predict sow farrowing. The overall flowchart of this method is shown below. Figure 1 As shown, it mainly includes the following steps (S1 to S6):
[0047] S1. Collect standing image data of sows at the pig farm 7 days before the sow's expected farrowing date and preprocess the data;
[0048] S1-1. Use a camera to capture images of the vulva of a sow in a standing position from a rear-view perspective;
[0049] S1-2. Perform operations such as cleaning, filtering, noise reduction, and image enhancement on the collected sow images to construct a sow vulva image dataset D1;
[0050] S2. Train the instance segmentation deep learning YOLOv8 model to obtain the sow vulva segmentation mask;
[0051] S2-1. Divide the collected sow vulva image dataset D1 into a training set D1_train, a validation set D1_valid, and a test set D1_test in a ratio of 6:2:2.
[0052] S2-2. Perform vulvar mask annotation on the training set D1_train, validation set D1_valid, and test set D1_test respectively. Divide the vulvar region into three sub-regions in the image: left labia, right labia, and clitoris, with categories left_lip, right_lip, and clitoris respectively, and mark the positions of the three sub-regions.
[0053] S2-3. Generate a corresponding txt file with the image name for each labeled image file, containing the target category and the coordinates of all points of the labeled mask. ,in , All values are normalized.
[0054] S2-4. Use the YOLOv8 model to train the labeled training set D1_train, and use the validation set D1_valid to validate it, thus obtaining the vulvar instance segmentation model.
[0055] S2-5. Input the test set D1_test divided in S2-2 into the trained vulvar instance segmentation model, evaluate the performance metrics of the model, and visualize the segmentation model results as follows: Figure 4 As shown;
[0056] S2-6. Use the trained vulvar instance segmentation model to identify the vulva image of a sow and obtain the coordinates of all points of the instance segmentation mask for the three sub-regions in the vulva image. And the coordinates of the top-left corner of the bounding rectangle of the three sub-regions. Width of the detection frame and height Combining depth information and depth camera intrinsics Calculate the true coordinates of the sub-region mask after segmentation. The calculation formula is as follows:
[0057]
[0058] S3. Design a feature extraction network to extract the original physiological features of the sow's vulva area;
[0059] S3-1. Using the instance segmentation mask coordinates obtained in S2-6, calculate the width of the three sub-regions. ,high and area To obtain the morphological characteristics of the vulvar area;
[0060] S3-2. Using OpenCV color space conversion and statistical methods, calculate the mean and standard deviation of each channel (H, S, V) in each sub-region within the three sub-region segmentation mask coordinates, and normalize the generated mean and standard deviation to obtain the color features of the vulva region.
[0061] S3-3. Use the Gray-Level Co-occurrence Matrix (GLCM) method to calculate the texture features of the vulva region;
[0062] S3-4. The extracted morphological, color, and texture features of the vulva are spliced together to construct the original physiological features of the vulva. ;
[0063] S4, such as Figure 2 As shown in the flowchart of constructing variables for the temporal representation of vulvar physiology, temporal modeling of the original physiological characteristics of the vulva is performed to obtain the temporal representation of the physiological characteristics of the vulva of sows.
[0064] S4-1. Input the image of the sow's vulva into the trained YOLOv8 instance segmentation model to obtain the instance segmentation mask coordinates;
[0065] S4-2. Input the segmentation mask coordinates into the feature extraction network model to extract the original physiological features of the sow's vulva region. ;
[0066] S4-3. Repeat steps S4-1 and S4-2 for the vulva images of sows for 7 consecutive days, and output the original physiological features of the vulva for 7 consecutive days to a CSV file, including date, morphological features, color features and texture features;
[0067] S4-4. Using the Long Short-Term Memory (LSTM) network method, time-series modeling is performed on the original physiological features in the CSV file to obtain dynamic features with long- and short-term dependencies, and time-series features are extracted. ;
[0068] S4-5. Use statistical methods to calculate the mean, variance, and skewness of the primitive physiological characteristics of the sow's vulva. ;
[0069] S5. The extracted vulvar features are fused to construct a physiological time-series representation variable for the sow vulva;
[0070] S5-1, Extracted original physiological characteristics Time series characteristics and Perform dimensional transformation to unify them to the same dimension;
[0071] S5-2. Assign learnable parameter weights to the three features, use feature variance as a quality indicator to initialize the weights, and adjust the learnable parameters as the classification model is trained.
[0072] S5-3, Weighted original physiological characteristics Time series characteristics and By splicing the data, we can construct a series of physiological characteristics of the vulva. like Figure 3 As shown, the construction formula is as follows:
[0073]
[0074] S6. Input the vulvar physiological time sequence characterization variables into the prototype calibration classification model to obtain the sow's farrowing status and predict the sow's farrowing.
[0075] S6-1. Perform steps S4 and S5 on all sows to obtain the vulvar physiological time-series characterization variables of all sows and construct dataset D2.
[0076] S6-2. Divide the dataset D2 containing all vulvar physiological time-series representation variables obtained in S5-1 into a training set D2_train, a validation set D2_valid, and a test set D2_test in a ratio of 8:1:1.
[0077] S6-3. Label the farrowing status categories of the training set D2_train, validation set D2_valid, and test set D2_test respectively. Label the characteristic variables of sows that farrow within 12 hours as category 1, and label the characteristic variables of sows that have not farrowed within 12 hours as category 0.
[0078] S6-4. Train the improved prototype network using the labeled training set D2_train, and calculate the similarity between the query sample and the support samples using an instance-level attention mechanism. The calculation formula is as follows:
[0079]
[0080]
[0081]
[0082] in, It is an instance in the support set of category c. This is a query example. It is a linear layer, and ⊙ represents element-wise multiplication. It is an activation function, and sum {} is the sum of all elements in the vector. It is the importance score of the query instance relative to the support set instances. This is the calculated weighted prototype.
[0083] The category bias is corrected through data expansion and prototype correction mechanisms, calculated as follows:
[0084]
[0085]
[0086]
[0087] Where Bias is the class bias, and E is the prototype calculated in an ideal way. It is an instance that belongs to this category. is the distribution of all instances belonging to this category, where x is an instance of the available support set for this category. It is the distribution of available tag instances for this category. The pseudo-prototype is obtained by expanding the support set by selecting the top t query instances with the highest prediction confidence. It is the prototype using weighted average. The t query instances with the highest confidence scores among the predicted query instances. It is an instance from the original support set and the expanded set of t query instances, and uses a balancing parameter λ to balance the importance of the pseudo-prototype, thus calculating the final prototype for each category. ;
[0088] S6-5. Validate using the validation set D2_valid to obtain the classification model of the vulvar physiological time-series representation variables, and calculate the probability of belonging to each category. The calculation formula is as follows:
[0089]
[0090] in, Representative to A query instance after encoding. This is a query example. The probability of belonging to category c. Represents the distance calculation function. Represents one category among all prototype categories. This represents the summation over all categories. The test set D2_test, partitioned by S6-2, is input into the trained classification model of vulvar physiological time-series representation variables. The performance metrics of this model are evaluated, and the model accuracy is shown in Table 1.
[0091] Table 1: Model Performance Indicators
[0092]
[0093] S6-6. Using a classification model of vulvar physiological time sequence characteristics, the sow's parturition status is obtained. When the model classification result indicates that the sow will parturient within 12 hours, a parturition warning is issued.
[0094] Preferably, this embodiment uses YOLOv8 as the instance segmentation algorithm, but it can also be replaced with other instance segmentation algorithms, including but not limited to YOLOv5, YOLOv10, YOLOv11, and Mask R-CNN. This embodiment compares and selects the optimal solution based on existing data.
[0095] Preferably, this embodiment uses primitive vulvar features, statistical features, and temporal features to construct the physiological temporal representation variables of the sow's vulva. Other features and construction methods can also be used. This embodiment compares existing data and selects the optimal solution.
[0096] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.
Claims
1. A method for predicting sow farrowing using vulvar physiological time-series characterization and prototype correction, characterized in that... It includes: S1. Data Acquisition and Preprocessing: Images of the vulva of sows in a standing position are acquired before the sow's expected farrowing date and then cleaned, screened, enhanced, and normalized. S2. Instance segmentation: The vulva region is segmented using a deep learning instance segmentation model to obtain a sub-region mask containing the left labia, right labia and clitoris; S3. Feature Extraction: Based on sub-regions, the morphological, color, and texture features of the vulva are extracted using a feature extraction network model to form the original physiological features. S4. Temporal Representation: Temporal modeling of raw physiological characteristics over multiple consecutive days is performed to obtain dynamic features with long-term and short-term dependencies, and statistical methods are combined to extract statistical features such as mean, variance, and skewness. S5. Feature Fusion: The original features, temporal features and statistical features are fused to construct vulvar physiological temporal representation variables; S6. Prototype Correction Prediction: Input the characterization variables into the improved prototype correction classification model, calculate the similarity between the query sample and the support samples using an instance-level attention mechanism, and correct the category bias through data expansion and prototype correction mechanisms to obtain the sow farrowing prediction result; the construction steps of the improved prototype correction classification model are as follows: S6-1. Perform steps S4 and S5 on sows to obtain the vulvar physiological time-series characterization variables of all sows and construct dataset D2. S6-2. Divide the dataset D2 containing all vulvar physiological time-series representation variables obtained in S5-1 into a training set D2_train, a validation set D2_valid, and a test set D2_test according to the proportions. S6-3. Label the farrowing status categories of the training set D2_train, validation set D2_valid, and test set D2_test respectively. Label the characteristic variables of sows that farrow within 12 hours as category 1, and label the characteristic variables of sows that have not farrowed within 12 hours as category 0. S6-4. Train the improved prototype network using the labeled training set D2_train, calculate the similarity between the query sample and the support sample using the instance-level attention mechanism, and correct the category bias through data expansion and prototype correction mechanisms to obtain the sow farrowing prediction results; and use the validation set D2_valid to validate the model, obtain the classification model of vulvar physiological time-series representation variables, and calculate the probability of belonging to each category. S6-5. Input the test set D2_test divided in S6-2 into the trained vulvar physiological time-series representation variable classification model and evaluate the performance index of the model. S6-6. Using a classification model of vulvar physiological time sequence characteristics, the sow's parturition status is obtained. When the model classification result indicates that the sow will parturient within 12 hours, a parturition warning is issued.
2. The method according to claim 1, characterized in that, In step S2, the deep learning instance segmentation model uses the YOLOv8 model. Images of the vulva of sows in a standing position are collected before the sow's expected farrowing date, and then cleaned, filtered, enhanced, and normalized to obtain the dataset. The training steps of the model are as follows: S2-1. Divide the collected sow vulva image dataset D1 into training set D1_train, validation set D1_valid, and test set D1_test according to the proportion. S2-2. Perform vulvar mask annotation on the training set D1_train, validation set D1_valid, and test set D1_test respectively. Divide the vulvar region into three sub-regions in the image: left labia, right labia, and clitoris, with categories left_lip, right_lip, and clitoris respectively, and mark the positions of the three sub-regions. S2-3. Generate a corresponding txt file with the image name for each labeled image file, containing the target category and the coordinates of all points of the labeled mask. ,in , All values are normalized. S2-4. Use the YOLOv8 model to train the labeled training set D1_train, and use the validation set D1_valid to validate it, thus obtaining the vulvar instance segmentation model. S2-5. Input the test set D1_test divided in S2-2 into the trained vulvar instance segmentation model and evaluate the performance metrics of the model. S2-6. Use the trained vulvar instance segmentation model to identify the vulva image of a sow and obtain the coordinates of all points of the instance segmentation mask for the three sub-regions in the vulva image. And the coordinates of the top-left corner of the bounding rectangle of the three sub-regions. Width of the detection frame and height Combining depth information and depth camera intrinsics Calculate the true coordinates of the sub-region mask after segmentation. .
3. The method according to claim 2, characterized in that, In S3, a feature extraction network model is designed to extract the original physiological features of the sow's vulva region; specifically, the following steps are included: S3-1. Using the instance segmentation mask coordinates obtained in S2-6, calculate the width of the three sub-regions. ,high and area To obtain the morphological characteristics of the vulvar area; S3-2. Using OpenCV color space conversion and statistical methods, calculate the mean and standard deviation of each channel (H, S, V) in each sub-region within the three sub-region segmentation mask coordinates, and normalize the generated mean and standard deviation to obtain the color features of the vulva region. S3-3. Use the Gray-Level Co-occurrence Matrix (GLCM) method to calculate the texture features of the vulva region; S3-4. The extracted morphological, color, and texture features of the vulva are spliced together to construct the original physiological features of the vulva. .
4. The method according to claim 1, characterized in that, In S4, temporal modeling of the original physiological characteristics of the vulva is performed to obtain the temporal representation of the physiological characteristics of the vulva of sows; Specifically, the following steps are included: S4-1. Input the image of the sow's vulva into the trained deep learning instance segmentation model to obtain the instance segmentation mask coordinates; S4-2. Input the segmentation mask coordinates into the feature extraction network model to extract the original physiological features of the sow's vulva region. ; S4-3. Repeat steps S4-1 and S4-2 for sow vulva images from multiple consecutive days, and output the original physiological features of the vulva from multiple consecutive days to a CSV file, including date, morphological features, color features, and texture features; S4-4. Using the Long Short-Term Memory (LSTM) network method, time-series modeling is performed on the original physiological features in the CSV file to obtain dynamic features with long- and short-term dependencies, and time-series features are extracted. ; S4-5. Calculate the statistical characteristics of the primitive physiological features of the sow's vulva using statistical methods. This includes the mean, variance, and skewness.
5. The method according to claim 4, characterized in that... In S5, the extracted vulvar features are fused to construct a physiological time-series representation variable for the sow vulva; S5-1, Extracted original physiological characteristics Time series characteristics and Perform dimensional transformation to unify them to the same dimension; S5-2. Assign learnable parameter weights to the three features, use feature variance as a quality indicator to initialize the weights, and adjust the learnable parameters as the classification model is trained. S5-3, Weighted original physiological characteristics Time series characteristics and By splicing the data, we can construct a series of physiological characteristics of the vulva. .