Milk goat phenotype determination method and system based on adaptive information fusion

Abstract

The invention provides a milk goat phenotype determination method and system based on adaptive information fusion, and belongs to the technical field of intelligent breeding, and the method comprises the following steps: preprocessing multi-dimensional phenotype information of a milk goat; extracting high-order feature representation from different data sources of the preprocessed multi-dimensional phenotypic information by using an integrated learning model, and calculating weights of different data sources; performing weighted decision according to the high-order feature representation and the data source weight to obtain a fusion result; inputting the fusion result into a phenotype calculation module for training to obtain a trained phenotype calculation module; and performing performance evaluation on the trained phenotype calculation module, and using a self-learning optimization process when an error exceeds a set threshold value to realize self-adaptive updating of model parameters. According to the method, the problems of low efficiency, large error, insufficient multi-source fusion and poor model adaptability in traditional phenotype determination are effectively solved, and the automation, robustness and precision of milk goat phenotype determination in a complex environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent breeding technology, and in particular to a method and system for determining the phenotypic characteristics of dairy goats based on adaptive information fusion. Background Technology

[0002] With the rapid development of smart animal husbandry and precision breeding technologies, the automated collection and intelligent analysis of animal phenotypic data have become crucial components of modern breeding systems. Dairy goats, as a primary dairy-producing herbivorous animal, have phenotypic parameters such as body size, milk yield, and milk composition as key indicators for evaluating their genetic potential and production performance. Traditional manual measurements rely heavily on measurement and visual inspection, which suffers from high workload, low efficiency, and strong subjectivity. These methods are easily affected by operational experience, lighting conditions, and posture changes, making it difficult to meet the needs of large-scale population assessment and dynamic monitoring. In recent years, computer vision and intelligent sensing technologies have been introduced into animal husbandry breeding, enabling partially automated measurements by collecting animal body size and behavioral information through cameras, depth cameras, or sensors. However, existing studies often rely on single modalities, lacking multi-source data collaboration, and thus failing to comprehensively reflect the overall phenotypic characteristics of dairy goats. Differences in spatial scale, temporal resolution, and noise characteristics between different modalities make simple data stacking or static weight fusion insufficient to adapt to individual differences and environmental fluctuations, resulting in insufficient model generalization and stability. Furthermore, existing systems generally lack self-learning and dynamic optimization capabilities, failing to update parameters and adaptively adjust based on measurement errors.

[0003] Therefore, there is an urgent need to construct an intelligent phenotyping method that integrates multi-dimensional features and possesses adaptive weight adjustment and self-learning optimization mechanisms. By combining deep learning and fuzzy logic, dynamic coordination and conflict correction of multi-source information can be achieved, enabling stable output of high-precision body size parameters in complex environments, providing reliable data support for the selection of high-quality breeding sheep, precision nutrition management, and intelligent farming. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for determining the phenotype of dairy goats based on adaptive information fusion, so as to solve the problems of low efficiency, large error, insufficient fusion of multi-source data and poor model adaptability of traditional phenotype determination methods.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for determining the phenotype of dairy goats based on adaptive information fusion, comprising: Step 1: Obtain multidimensional phenotypic information of dairy goats; the multidimensional phenotypic information includes: body size and posture image data acquired based on computer vision, milk yield and milk composition data and basic attribute data acquired based on sensors; Step 2: Preprocess the multidimensional phenotypic information to obtain preprocessed multidimensional phenotypic information; Step 3: Use an ensemble learning model to extract high-order feature representations from different data sources of preprocessed multidimensional phenotypic information and calculate the weights of different data sources; Step 4: Perform a weighted decision based on the higher-order feature representation and the data source weights to obtain the fusion result; Step 5: Input the fusion result into the phenotypic solution module for training to obtain the trained phenotypic solution module; Step 6: Evaluate the performance of the trained phenotypic solution module, and use a self-learning optimization process when the error exceeds a set threshold to achieve adaptive updating of model parameters.

[0006] Preferably, in step 1, a time synchronization method is used to achieve time alignment between the video frames and the sensor data, and the time synchronization formula is:

[0007] in, For the target time point that needs to be synchronized, The sensor data value corresponding to the target time point. and For the time markers of the two points before and after, and These are the sensor data values ​​at the corresponding time points; When the light intensity is below 10 lux, an adaptive smoothing algorithm is used to process the body size and posture image data. The specific calculation formula is as follows:

[0008] in, These are the pixel values ​​of the original image. These are the weighting factors of the locally weighted filter function. The radius of the locally weighted range. It is the pixel factor after smoothing; When the illumination intensity exceeds 1000 lux, a brightness correction algorithm is introduced to process the body size and posture image data. The specific calculation formula is as follows:

[0009] in, The pixel values ​​of the repaired image. These are the pixel values ​​of the original image. This is the contrast adjustment factor. This is the brightness correction constant; When the occlusion rate exceeds 30%, a region restoration algorithm is used to process the body size and posture image data. The specific calculation formula is as follows:

[0010] in, For the missing frame time points that need interpolation, and These are the pixel values ​​of the previous and next frames, respectively. These are the interpolated image pixel values.

[0011] Preferably, in step 2, the outlier identification method based on data quantile boundaries calculates the interval difference:

[0012] in, For interval difference, This is the boundary of the first quantile segment. This marks the boundary of the third quantile segment; Set a threshold, and calculate the acceptable range of the data as follows: or

[0013] in, The lower realm The upper bound is below the lower bound. or higher than the upper limit The value is considered an outlier and is used to remove or replace it with the correct data value; Transform the data after removing outliers into a standard Gaussian distribution so that all dimensions of the data have the same scale. The formula is:

[0014] in, For the original data items, The center value of the data. This is a measure of the deviation of the data.

[0015] Preferably, in step 3, the base learners in the ensemble learning model are utilized. Process the multidimensional phenotypic information and take the input data Mapping to a higher-order feature space, the formula for generating the prediction result is:

[0016] in, For the parameters or hyperparameters of the base learners, the output of each base learner is... The final high-order feature representation is obtained by fusion through a weighted decision mechanism, and the calculation formula is as follows:

[0017] in, For the first The weights of each base learner The number of base learners.

[0018] Preferably, in step 3, the weight of each data source is adjusted through a dynamic fuzzy adjustment mechanism, and the calculation formula is as follows:

[0019] in, The initial weights of the data source, Indicates the first Data source weight at any given time It is an adjustment factor. This represents the weight increment adjusted according to the fuzzy rules.

[0020] Preferably, in step 4, adaptive fusion is performed on each data source based on its features and corresponding weights. The specific fusion formula is as follows:

[0021] in, For the final output, It is a data source Information representation, ; The weighted decision averaging method is used to optimize the final output, and the calculation formula is as follows:

[0022] in, For the final fusion result, The data source weights have been optimized and adjusted, and The adjustment is achieved through the correlation between data sources. For the first Information representation of each data source; The formula for quantifying the correlation between different data sources is:

[0023] in, A measure of the correlation between each data source, representing the data sources. and The correlation.

[0024] Preferably, in step 5, the phenotypic calculation module includes a feature reconstruction layer, a regression calculation layer, and a result output layer; the feature reconstruction layer processes the fusion result. Feature mapping is performed to obtain the reconstructed high-dimensional feature representation:

[0025] in, For the mapping weight matrix, For bias terms, It is a non-linear activation function; The regression calculation layer establishes a nonlinear regression relationship based on a multilayer perceptron network structure, which is used to output key phenotypic parameters of dairy goats:

[0026] in For the hidden layer activation function, , These are the weight matrices for the hidden layer and the output layer, respectively. , For bias terms; The output layer is used to output body height, body length, chest circumference, milk yield, and milk protein content.

[0027] Preferably, in step 6, the measurement accuracy is evaluated based on the mean absolute error and root mean square error; when the error exceeds a threshold... At that time, a small-sample incremental learning mechanism is triggered to fine-tune the weights and fuzzy logic parameters of the deep feature extraction model; wherein, the error evaluation function is:

[0028] in, The predicted phenotypic values ​​output by the model. These are manually calibrated values, and N is the sample size. when or At this point, the system enters the parameter optimization phase, adjusting the weights using an error gradient-based approach. The calculation formula is as follows:

[0029] in, For adaptive learning rate, The comprehensive loss function is defined as follows: , and These are weighting coefficients used to balance the effects of the two types of errors; New weights after adjustment With fuzzy logic rule parameters Synchronous updates, and through a dynamic learning rate decay function. Achieving model convergence completes the continuous optimization and self-learning loop of the model; among which, Indicates the first The model weight parameter matrix of the next iteration This represents the weight increment calculated based on the error gradient. The updated weight parameter matrix, Indicates the first The learning rate for the next iteration. This is the learning rate decay coefficient, used to control the learning rate as the number of iterations increases. The rate of decrease.

[0030] The present invention also provides a system for determining the phenotypic characteristics of dairy goats based on adaptive information fusion, comprising: A multi-source data acquisition module is used to acquire multidimensional phenotypic information of dairy goats; the multidimensional phenotypic information includes: body size and posture image data acquired based on computer vision, milk yield and milk composition data and basic attribute data acquired based on sensors; The data preprocessing module is used to preprocess multidimensional phenotypic information to obtain preprocessed multidimensional phenotypic information; The data source feature extraction module is used to extract high-order feature representations from different data sources of preprocessed multidimensional phenotypic information using an ensemble learning model, and to calculate the weights of different data sources. The adaptive information fusion module is used to perform weighted decision-making based on high-order feature representations and data source weights to obtain the fusion result; The phenotypic solution module is used to input the fusion result into the phenotypic solution module for training to obtain the trained phenotypic solution module; The model feedback and dynamic optimization module is used to evaluate the performance of the trained phenotypic solution module and to use a self-learning optimization process when the error exceeds a set threshold to achieve adaptive updates of the model parameters.

[0031] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) This invention collected body size and posture image data, milk yield and milk composition data, as well as basic attribute data such as breed and age of dairy goats, and constructed a multidimensional phenotypic information system for dairy goats. Through multi-source data acquisition and preprocessing, the limitations of traditional single-modal measurement were overcome, and comprehensive and automated measurement of key phenotypic parameters such as body height, body length, chest circumference, milk yield, and milk protein content of dairy goats was realized, providing comprehensive data support for precision breeding and efficient farming.

[0032] (2) This invention constructs an adaptive information fusion module based on multidimensional ensemble learning and adaptive fuzzy control, realizing high-order feature extraction and dynamic weighted fusion of multi-source heterogeneous data. This module mines deep correlation features of different data sources through an ensemble learning model, and uses an adaptive fuzzy algorithm to adjust the weights of each modality in real time. Combined with an information conflict optimization mechanism, it effectively corrects the inconsistencies between multiple sources, thereby significantly improving the stability, robustness and generalization performance of phenotypic measurement results in complex environments.

[0033] (3) This invention designs a phenotypic solution model that includes feature reconstruction and nonlinear regression calculation, and introduces a model feedback and dynamic optimization mechanism based on error evaluation. The system can automatically trigger small-sample incremental learning based on the error between the measurement results and the manually calibrated values, fine-tuning and optimizing the model parameters, thus realizing the model's self-learning and continuous performance improvement. Combined with a data storage module with hierarchical storage and version management, the traceability of measurement results and the reliability of model updates are ensured, forming a complete "measurement-evaluation-optimization" closed loop. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 A flowchart of the method for determining the phenotype of dairy goats based on adaptive information fusion provided in an embodiment of the present invention; Figure 2 A schematic diagram illustrating the principle of the adaptive information fusion-based method for determining the phenotype of dairy goats provided in this embodiment of the invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] like Figure 1 As shown, this invention provides a method for determining the phenotype of dairy goats based on adaptive information fusion, comprising: Step 1: Obtain multidimensional phenotypic information of dairy goats; the multidimensional phenotypic information includes: body size and posture image data acquired based on computer vision, milk yield and milk composition data and basic attribute data acquired based on sensors; In this invention, multidimensional phenotypic data includes, but is not limited to: body size and posture image data, milk yield and milk composition data, and basic attribute data; the body size and posture image data includes, but is not limited to: body height, body length, and chest circumference; the milk yield and milk composition data includes, but is not limited to: daily milk yield and milk protein content; the basic attribute data includes, but is not limited to: dairy goat breed and dairy goat age in months.

[0039] In this embodiment, the multidimensional phenotypic data structure of dairy goats in step 1 is divided into three tables: a body size and body shape data table, a milk yield and milk composition data table, and a basic attribute data table, as shown in Tables 1 to 3, respectively. The body size and body shape data table, the milk yield and milk composition data table, and the basic attribute data table are associated with the main database through Goat_ID as a foreign key.

[0040] Table 1 Body Size and Posture Data

[0041] Table 2 Milk Yield and Milk Composition Data

[0042] Table 3 Genomic Molecular Markers

[0043] In this embodiment, the multi-source data acquisition unit uses a time synchronization module to align video frames with sensor data, ensuring that time series from different data sources can be analyzed at the same time step. The time series are aligned using a time synchronization formula, specifically:

[0044] in, For the target time point that needs to be synchronized, The sensor data value corresponding to the target time point. and For the time markers of the two points before and after, and These represent the sensor data values ​​at corresponding time points. Using a time synchronization formula, missing data between time markers can be filled in, ensuring that all time series data are aligned.

[0045] The multi-source data acquisition unit introduces an automatic compensation acquisition mechanism to address the adverse effects of ambient lighting and occlusion, ensuring the quality and continuity of video data. Specifically, when the light intensity is below 10 lux or above 1000 lux, or when the occlusion rate of a video frame exceeds 30%, the system automatically triggers the compensation mechanism and activates corresponding image enhancement techniques to improve video quality.

[0046] Specifically, for low-light environments (below 10 lux), to improve the quality of video data in low-light conditions, the system employs an adaptive smoothing algorithm. This algorithm reduces noise and improves image clarity by performing local pixel-weighted smoothing on the image. Specifically, the algorithm performs weighted correction within the neighborhood of each pixel, using the weighted average of the surrounding pixel values ​​as the new value for the target pixel. The specific calculation formula is as follows:

[0047] in, These are the pixel values ​​of the original image. These are the weighting factors of the locally weighted filter function. The radius of the local weighted filter. It is the pixel factor after smoothing.

[0048] Furthermore, in high-light environments, when the light intensity exceeds 1000 lux, video frames may become overexposed, leading to loss of image details. To address this, the multi-source data acquisition unit introduces a brightness correction algorithm, which adjusts the brightness and contrast of the video image to repair overexposure caused by high light. The brightness correction algorithm restores the image to a normal brightness range by adjusting its brightness and contrast, preventing the loss of important information due to overexposure. The specific calculation formula is as follows:

[0049] in, The pixel values ​​of the repaired image. These are the pixel values ​​of the original image. This is the contrast adjustment factor. This is the brightness correction constant.

[0050] Furthermore, when the occlusion rate exceeds 30%, the automatic compensation acquisition mechanism utilizes a region restoration algorithm to repair the missing video data caused by occlusion. The region restoration algorithm compensates for the correlation between consecutive frames; the specific calculation formula is as follows:

[0051] in, For the missing frame time points that need interpolation, and These are the pixel values ​​of the previous and next frames, respectively. These are the interpolated image pixel values.

[0052] Step 2: Preprocess the multidimensional phenotypic information to obtain preprocessed multidimensional phenotypic information; In step 2, the collected multidimensional phenotypic data undergoes comprehensive processing, including outlier identification and removal, data time-series alignment, and eigenvalue normalization. To ensure data quality, outliers and inconsistencies in each data source must be addressed. Outlier removal is the first step in data cleaning. An outlier identification and removal method based on data quantile boundaries is used to determine the effective data range for each data source. The formula for calculating the interval difference is:

[0053] in, For interval difference, This is the boundary of the first quantile segment. This is the boundary of the third quantile segment. The data is calculated... and This allows us to obtain the normal range of the data.

[0054] Furthermore, an appropriate threshold is set, typically a constant multiple (such as 1.5) to determine the acceptable range of data. The calculation formula is: or

[0055] in, The lower realm The upper bound is below the lower bound. or higher than the upper limit Values ​​that are considered outliers are either removed or replaced with reasonable data values. This process effectively eliminates abnormal data caused by sensor malfunctions, environmental changes, etc., thereby ensuring the accuracy of subsequent data analysis.

[0056] Eigenvalue normalization employs a uniform variability method, which transforms the data into a standard Gaussian distribution, ensuring that all dimensions of the data have the same scale. The calculation formula is as follows:

[0057] in, For the original data items, The center value of the data. This serves as a measure of data bias. This unified variability method helps avoid the excessive impact of dimensional differences between data from different dimensions on model training, ensuring the robustness and stability of the model.

[0058] Step 3: Use an ensemble learning model to extract high-order feature representations from different data sources of preprocessed multidimensional phenotypic information and calculate the weights of different data sources; In this embodiment, step 3 utilizes a multi-source heterogeneous data processing system with advanced ensemble learning and fuzzy control techniques to extract deep-level features from multi-dimensional data from different sources and perform adaptive fusion. The system achieves multi-level and comprehensive processing of dairy goat phenotypic data and improves the accuracy and reliability of model predictions through weighted information fusion and optimization methods.

[0059] Specifically, multidimensional ensemble learning models are used to extract high-order features from multi-source heterogeneous data. To extract deep-level features from different sensors and data sources (such as body shape image data, milk yield and composition data, basic attribute data, etc.), ensemble learning methods are employed. Specifically, the base learners in the ensemble learning model... Each input data Mapped to a higher-order feature space. Each base learner The input data is processed to generate prediction results. The calculation formula is:

[0060] in, For input data, represent features obtained from a specific data source (such as body shape images, milk production data, etc.), specifically in the form of... , For the first Data points, For the first The base learner, as a base learner The parameters or hyperparameters. For the prediction result, the output value of each base learner represents the feature representation extracted from the data source.

[0061] Multidimensional ensemble learning models combine the predictions of multiple base learners, making weighted decisions based on the weights of different data sources and features, thereby extracting more accurate and robust features. The outputs of the base learners are fused through a weighted decision mechanism. In multidimensional ensemble learning models, the weight of each base learner is determined by the reliability or accuracy of the data source. The calculation formula is as follows:

[0062] in, This represents the final output of the multidimensional ensemble learning model, i.e., the high-order feature representation. For the first The output of the base learner represents the value from the first base learner. Features extracted from data sources. For the first The weights of each base learner are determined based on the performance of each learner or its contribution to the data source. The number of base learners represents the total number of base learners in the ensemble learning model.

[0063] Weights of base learners Learning can be automated through the training process, typically adjusted by evaluating the performance of each base learner on the validation set. The weights can reflect the contribution of each base learner. For weighted decision mechanisms, the weights can be optimized based on the accuracy or other metrics of each base learner. For example, the weight vector can be dynamically adjusted based on the performance of the base learners or the reliability of the data source. This leads to a more accurate and robust feature representation.

[0064] Specifically, the adaptive fuzzy control algorithm is applied to dynamically adjust the weights of data sources based on their importance. To maximize the fusion effect, the algorithm employs a dynamic fuzzy adjustment mechanism to adjust the weights of each data source, ensuring the accuracy of information fusion. The specific calculation formula is as follows:

[0065] in, The initial weights of the data source, Indicates the first Data source weight at any given time It is an adjustment factor. This represents the weight increment adjusted according to fuzzy rules. Through this dynamic adjustment mechanism, the model can optimize the weights of each data source in real time based on current fusion requirements. Specifically, if a feature from a certain data source contributes significantly to the final target output, its weight will be increased; conversely, it will be decreased. Fuzzy control algorithms can flexibly handle nonlinear relationships and uncertainties between different data sources. In some cases, there may be complex nonlinear correlations between goat body size data and milk composition data. Dynamic adjustment of fuzzy rules can effectively capture these nonlinear relationships and optimize the data fusion results.

[0066] Step 4: Perform a weighted decision based on the higher-order feature representation and the data source weights to obtain the fusion result; In step 4, the weighted information fusion module adaptively fuses the data sources based on their features and corresponding weights. The target output is set as the fused result, and the specific formula is as follows:

[0067] in, For the final output, It is a data source Information representation, By using a weighted information fusion method, the model can rationally integrate information from different data sources according to their importance, thereby obtaining more accurate and comprehensive prediction results.

[0068] The information conflict optimization mechanism is primarily used to further optimize potential conflicts or inconsistencies between different data sources. Considering that different data sources may provide conflicting information, this mechanism employs a weighted decision averaging method to optimize the final fusion result. The calculation formula is as follows:

[0069] in, For the final fusion output, The optimized and adjusted data source weights For the first Information representation from each data source. To effectively handle conflicts or inconsistencies between different data sources, the first step is to use correlation metrics. To quantify the correlation between different data sources, the similarity calculation formula is as follows:

[0070] in, A measure of the correlation between each data source, representing the data sources. and The correlation, This refers to specific methods for measuring relevance. To resolve conflicts between data sources, relevance measurement is first used. This method quantifies the correlation between different data sources. Based on the similarity between data sources, the weight of each data source is adjusted so that the final fusion result minimizes conflicts between data sources.

[0071] Step 5: Input the fusion result into the phenotypic solution module for training to obtain the trained phenotypic solution module; In step 5, the phenotypic solution module transforms the comprehensive feature vector output by the adaptive information fusion module into key phenotypic parameters of dairy goats, realizing a nonlinear mapping from the feature space to the physiological phenotypic space; this module includes a feature reconstruction layer, a regression calculation layer, and a result output layer.

[0072] The comprehensive feature vector output by the adaptive information fusion module The input is fed into the feature reconstruction layer. The feature reconstruction layer synthesizes the feature vector. After feature mapping, the reconstructed high-dimensional features are represented as follows: ,in, For the mapping weight matrix, For bias terms, This is a non-linear activation function. In this embodiment, the ReLU activation function is used to improve feature sparsity and non-linear expressive power. Testing showed that when n=128 and m=256, a better feature expressive power can be obtained. The role of this layer is to improve the separability and stability of the fused features in the phenotypic space. The regression calculation layer uses a three-layer perceptron network as input to the reconstructed features. The output is a five-dimensional phenotypic parameter vector.

[0073] in, These represent five key phenotypic parameters: body height, body length, chest circumference, milk yield, and milk protein content. Their nonlinear mapping relationship is defined as follows:

[0074] in For the hidden layer activation function, , These are the weight matrices for the hidden layer and the output layer, respectively. , This is the bias term. In this embodiment, the number of hidden layer nodes is set to 128, the number of output layer nodes is set to 5, and the Adam optimizer is used for parameter training with an initial learning rate of 0.001 and a batch size of 32. In actual implementation, the system collects 1200 sets of calibration sample data (including image features and milk production indicators), which are randomly divided into a training set of 900 sets and a validation set of 300 sets. Through stepwise training of regression relationships, the model can converge after about 80 iterations, with the average absolute error of prediction for each phenotypic parameter being less than 1.5%, and maintaining high stability under different lighting and pose conditions. The output layer outputs the regression results in vector form and transmits them to the model feedback module and the data storage module. By comparing the results of manual measurement with those of the system output, the phenotypic calculation module controls the average deviation of core parameters such as body height, chest circumference, and milk yield to within 2%. The overall measurement accuracy is improved by more than 10% compared with traditional manual methods, providing a reliable computational basis for subsequent dynamic optimization and intelligent measurement.

[0075] Step 6: Evaluate the performance of the trained phenotypic solution module, and use a self-learning optimization process when the error exceeds a set threshold to achieve adaptive updating of model parameters.

[0076] In step 6, the model performance is evaluated based on the error between the measurement results and the manually calibrated values, and a self-learning optimization process is triggered when the error exceeds a set threshold, so as to achieve adaptive updating of model parameters and stable performance improvement. The error evaluation function is defined as follows:

[0077] in, The predicted phenotypic values ​​output by the model. Here, N represents the manually calibrated values, and N is the sample size. or When the system enters the parameter optimization phase, the threshold in this embodiment is... =0.05; The self-learning parameter optimization process adjusts the weights based on the error gradient, as shown in the formula: ,in, For adaptive learning rate, The comprehensive loss function is defined as follows: , and These are weighting coefficients used to balance the effects of the two types of errors.

[0078] The adaptive update process involves adjusting the new weights. With fuzzy logic rule parameters Synchronous updates, and through a dynamic learning rate decay function. Achieve stable convergence of the model; among which, Indicates the first The model weight parameter matrix of the next iteration This represents the weight increment calculated based on the error gradient. The updated weight parameter matrix, Indicates the first The learning rate for the next iteration. This is the learning rate decay coefficient, used to control the learning rate as the number of iterations increases. The deceleration rate was gradually reduced. During implementation, approximately 50 recent samples were selected to form an incremental training set, with each optimization iteration lasting about 10-20 rounds. The model converged rapidly on the incremental data. After optimization, the system's R-value decreased by an average of about 9.8% and M-value decreased by about 12.3% on the new batch of data, verifying the adaptive capability of the dynamic optimization module under different environmental and individual differences.

[0079] In this invention, it also includes: a data storage module, which is used to perform hierarchical management of model parameters, measurement results and historical version information generated during system operation, including three parts: model parameter storage area, measurement result storage area and historical version management area; The model parameter storage area is used to store the latest weights of the adaptive information fusion model. With fuzzy logic parameters After each model update, based on the update identifier Create a version index table:

[0080] in Represents a timestamp, used to mark the time when the model is updated. This represents a versioned set of model parameters; The measurement results storage area is used to save the phenotypic measurement results of each individual dairy goat. And establish index mapping relationships ,in Assign an individual number, To determine time, This index indicates the model version number used, which allows for tracing phenotypic results for a specific individual, time period, and version. The historical version management area is used to store model update logs and performance evaluation metrics, and to define performance evolution functions. ,in, Indicates the first The root mean square error of the version model, This represents the initial model error. By comparing the performance of successive versions, the system can evaluate the extent of model optimization and performance stability. When the performance improvement ratio Below the performance threshold Time (in this example) If the performance improvement is less than 10%, the system determines that the model optimization effect is insufficient and automatically rolls back to the previous stable version. The rollback operation directly restores the model parameters through the index table. The system then reloads the data into the main computing module, enabling rapid switching and performance recovery. Simultaneously, the system records the rollback reason, trigger time, and performance difference ΔPerf in the version management area for subsequent manual review or automatic retraining. Historical version data supports retraining operations, allowing users to select a historical high-performance version as the base model and re-optimize it using new data, thereby improving the controllability of the model's long-term evolution.

[0081] In actual operation, the hierarchical storage module recorded approximately 200 model update logs, with an average version switching response time of less than 0.8 seconds. Through performance evolution function monitoring, the system can automatically roll back within 3 seconds when performance degradation exceeds 10%. Experimental results show that this mechanism effectively avoids overfitting and performance oscillations in the small-sample incremental learning phase, ensuring that the system error remains stable below 0.08 over long-term operation, thus guaranteeing the reliability and interpretability of model predictions.

[0082] Furthermore, experimental verification shows that the adaptive information fusion model proposed in this invention outperforms traditional methods in both accuracy and efficiency of body size measurement. Specific performance comparisons are shown in Table 4. Compared with methods such as monocular vision, binocular vision, and 3D point cloud analysis, the model of this invention exhibits the best performance across all error indices. The average relative error is only 1.62%. This invention improves the accuracy of core phenotypic parameter measurement by more than 10% and the measurement efficiency by more than 50% compared to traditional manual methods. It can also adapt to differences in breed, age, and breeding environment, providing reliable data support for the selection of high-quality dairy goats and precise breeding decisions, effectively promoting the intelligent upgrading of animal husbandry.

[0083] Table 4 Performance Comparison Table

[0084] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0085] This document illustrates the principle and implementation of a method for determining the phenotype of dairy goats based on adaptive information fusion through specific examples. The descriptions of these embodiments are merely illustrative of the method and its core ideas. Furthermore, those skilled in the art will recognize that modifications may be made to the specific implementation methods and the scope of the method for determining the phenotype of dairy goats based on adaptive information fusion, based on the principles of this invention. Therefore, the content of this specification should not be construed as limiting the scope of this invention.

Claims

1. A method for determining the phenotype of dairy goats based on adaptive information fusion, characterized by, The method comprises the following steps: Step 1: obtaining multi-dimensional phenotype information of dairy goats; The multi-dimensional phenotype information comprises body size and posture image data collected based on computer vision, milk yield and milk component data collected based on sensors, and basic attribute data; Step 2: preprocessing the multi-dimensional phenotype information to obtain preprocessed multi-dimensional phenotype information; Step 3: extracting high-order feature representation from different data sources of the preprocessed multi-dimensional phenotype information by using an ensemble learning model, and calculating weights of the different data sources; Step 4: performing weighted decision according to the high-order feature representation and the data source weights to obtain a fusion result; Step 5: inputting the fusion result into a phenotype solving module for training to obtain a trained phenotype solving module; Step 6: performing performance evaluation on the trained phenotype solving module, and using a self-learning optimization process to realize adaptive updating of model parameters when an error exceeds a set threshold.

2. The method according to claim 1, wherein, In step 1, a time synchronization method is used to realize time alignment of video frames and sensor data, and the time synchronization formula is as follows: ； wherein, is a target time point for which synchronization is required, is a sensor data value corresponding to the target time point, and are time stamps of the two points before and after, and are sensor data values corresponding to the time points, respectively. When the light intensity is less than 10 lux, an adaptive smoothing algorithm is used to process the body size and posture image data, and the specific calculation formula is as follows: ； wherein, is a pixel value of the original image, is a weight factor of the local weighted filter function, is a radius of the local weighted range, is a pixel factor after smoothing processing; When the light intensity is greater than 1000 lux, a brightness correction algorithm is introduced to process the body size and posture image data, and the specific calculation formula is as follows: ； wherein, is a pixel value of the repaired image, is a pixel value of the original image, is a contrast adjustment factor, is a luminance repair constant; When the occlusion rate is greater than 30%, a region repair algorithm is used to process the body size and posture image data, and the specific calculation formula is as follows: ； wherein, is the time point of the missing frame for which interpolation is required, and are the pixel values of the preceding and following frames, respectively, is the interpolated image pixel value.

3. The method according to claim 2, wherein the method is characterized by, In step 2, a data quantile interval limit outlier identification method is used to process the sensor data, first calculating the interval gap as follows: ； wherein is an interval gap, is a first quantile limit, is a third quantile limit; A threshold is set, and the acceptable range of data is calculated as follows: or ; wherein, is a lower bound, is an upper bound, values below the lower bound or above the upper bound are considered outliers and are either rejected or replaced by correct data values; The data after removing outliers is converted to a standard Gaussian distribution, so that all dimensions of data have the same dimension, and the formula is as follows: ； wherein, is the original data item, is the central value of the data, is the measure of deviation of the data.

4. The method according to claim 3, wherein the method is characterized by, In step 3, the base learners in the ensemble learning model are utilized The multi-dimensional phenotype information is processed, and input data is mapped to a high-order feature space to generate a prediction result formula as follows: ； wherein, are parameters or hyperparameters of the base learners, and the output of each base learner is are fused by a weighted decision mechanism to obtain the final high-order feature representation, and the calculation formula is: ； wherein, is the weight of the th base learner, is the number of base learners.

5. The method according to claim 4, wherein the method is characterized by, In step 3, the weight of each data source is adjusted through a dynamic fuzzy adjustment mechanism, and the calculation formula is as follows: ； wherein, is the initial weight of the data source, denotes the data source weight at the time instant, is the adjustment factor, is the weight increment adjusted according to the fuzzy rule.

6. The dairy goat phenotyping method based on adaptive information fusion according to claim 5, characterized in that, In step 4, each data source is adaptively fused according to the characteristics and corresponding weights of each data source, and the specific fusion formula is as follows: ； wherein, is the final output result, is the data source information representation, is the dynamic weight of the first data source; The final output result is optimized by using a weighted decision average method, and the calculation formula is as follows: ； wherein, is the final fusion result, is the data source weight after optimization adjustment, and is adjusted through the correlation between data sources, is the information representation of the first data source. The formula for quantifying the correlation between different data sources is as follows: ； wherein, is a measure of relevance between each data source, representing the relevance of data source and to data source 7. The method according to claim 6, wherein, In step 5, the phenotype solving module comprises a feature reconstruction layer, a regression calculation layer and a result output layer; The feature reconstruction layer is used for reconstructing the fusion result The feature mapping is performed to obtain a reconstructed high-dimensional feature representation ； wherein, is a mapping weight matrix, is a bias term, is a non-linear activation function; The regression calculation layer establishes a nonlinear regression relationship based on a multilayer perceptron network structure, and is used to output key phenotype parameters of dairy goats: ； wherein, is a hidden layer activation function, , are weight matrices for the hidden and output layers, respectively, , is a bias term; The result output layer is used to output body height, body length, chest circumference, milk yield and milk protein content.

8. The method according to claim 7, wherein the method is characterized by, In step 6, the determination accuracy is evaluated based on the mean absolute error and the root mean square error, and when the error exceeds a threshold value a small sample incremental learning mechanism is triggered to fine-tune the deep feature extraction model weights and the fuzzy logic parameters; wherein the error evaluation function is: ； wherein, is a predicted phenotype value output by the model, is a human-labeled value, and N is the number of samples. When or The system enters the parameter optimization phase, and the weight is adjusted based on the error gradient, and the calculation formula is: ； wherein, is the adaptive learning rate, is the comprehensive loss function, defined as , and is the weight coefficient, used to balance the influence of the two errors; adjusted new weights with fuzzy logic rule parameters is updated synchronously, and is decayed by a dynamic learning rate decay function achieves model convergence, thereby completing the continuous optimization and self-learning closed loop of the model; wherein, denotes the model weight parameter matrix of the iteration, denotes the weight increment calculated based on the error gradient, is the updated weight parameter matrix, denotes the model weight parameter matrix of the iteration, is the learning rate of the iteration, and is a learning rate decay coefficient used to control the decreasing rate of the learning rate with the number of iterations 9. A dairy goat phenotype measurement system based on adaptive information fusion, characterized by, The method comprises the following steps: A multi-source data acquisition module is used to obtain multi-dimensional phenotype information of dairy goats; The multi-dimensional phenotype information comprises body size and posture image data collected based on computer vision, milk yield and milk component data collected based on sensors, and basic attribute data; A data preprocessing module is used to preprocess the multi-dimensional phenotype information to obtain preprocessed multi-dimensional phenotype information; A data source feature extraction module is used to extract high-order feature representation from different data sources of the preprocessed multi-dimensional phenotype information by using an ensemble learning model, and calculate weights of the different data sources; A data source feature extraction module is used to extract high-order feature representation from different data sources of the preprocessed multi-dimensional phenotype information by using an ensemble learning model, and calculate weights of the different data sources; An adaptive information fusion module is configured to obtain a fusion result by weighting decision according to the high-order feature representation and the data source weight; A phenotype solving module is configured to input the fusion result into the phenotype solving module for training to obtain a trained phenotype solving module; A model feedback and dynamic optimization module is configured to evaluate the performance of the trained phenotype solving module, and use a self-learning optimization process to realize adaptive update of model parameters when the error exceeds a set threshold.

Images