A sheep nutrition dynamic monitoring method and system based on multi-source data fusion

Abstract

The application provides a kind of sheep nutrition dynamic monitoring method based on multi-source data fusion, it is related to livestock breeding intelligent monitoring technical field, the method comprises: at multiple moments of current feeding cycle, the environmental index of pasture is collected, at the end of current feeding cycle, the physiological monitoring index of multiple sheep is collected, the dosage of feed in current feeding cycle is obtained, and the proportioning of multiple nutrients in feed, through the nutrition monitoring model after training, environmental index, physiological monitoring index, the dosage of feed and the proportioning of multiple nutrients are handled, the addition suggestion information of various nutrients is obtained, in the next feeding cycle, according to the addition suggestion information, feed is made. Can make nutrition monitoring model consider the multi-source state information of current feeding cycle and historical feeding state change simultaneously, to improve the accuracy and dynamic adaptability of nutrient addition suggestion.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for livestock farming, and in particular to a method and system for dynamic monitoring of sheep nutrition based on multi-source data fusion. Background Technology

[0002] In sheep farming, feed formulation and nutritional adjustments typically rely on human experience, fixed formulas, or periodic test results. While this method can meet general feeding and management needs, it struggles to reflect the dynamic relationships between changes in the pasture environment, the sheep's physical condition, and the state of feed nutrition. Especially when environmental factors such as temperature, humidity, and rainfall change, sheep's feed intake, metabolic state, and nutritional requirements may alter. Feeding solely based on fixed formulas can easily lead to insufficient nutrient supply or wasted nutrients.

[0003] The information disclosed in the background section of this application is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0004] This invention provides a method for dynamic monitoring of sheep nutrition based on multi-source data fusion, which can solve the problem that existing sheep nutrition management mainly relies on human experience, resulting in inaccurate judgment of sheep's nutritional needs by the model.

[0005] According to a first aspect of the present invention, a method for monitoring the dynamic nutrition of sheep based on multi-source data fusion is provided, comprising:

[0006] Environmental indicators of the pasture were collected at multiple points during the current feeding cycle;

[0007] At the end of the current feeding cycle, physiological monitoring indicators were collected from multiple sheep.

[0008] Obtain the amount of feed used in the current feeding cycle, as well as the ratio of various nutrients in the feed;

[0009] The trained nutrition monitoring model processes environmental indicators, physiological monitoring indicators, feed dosage, and the ratio of various nutrients to obtain recommendations for the addition of various nutrients.

[0010] In the next feeding cycle, prepare feed according to the added recommendations.

[0011] According to a second aspect of the present invention, a sheep nutrition dynamic monitoring system based on multi-source data fusion is provided, comprising:

[0012] The first data collection module collects environmental indicators of the pasture at multiple points in the current feeding cycle.

[0013] The second data collection module collects physiological monitoring indicators from multiple sheep at the end of the current feeding cycle.

[0014] The third data acquisition module obtains the amount of feed used in the current feeding cycle, as well as the ratio of various nutrients in the feed.

[0015] Add a suggestion information acquisition module, which processes environmental indicators, physiological monitoring indicators, feed dosage and ratio of various nutrients through the trained nutrition monitoring model to obtain suggestion information for the addition of various nutrients.

[0016] The feed preparation module prepares feed according to the added suggestion information in the next feeding cycle.

[0017] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0018] According to this invention, environmental description sequences at multiple time points can be processed through 1D convolutional layers to extract continuous variation features of environmental factors within the current feeding cycle. Furthermore, through first and second multi-layer perception layers, physical characteristics reflecting the sheep's nutritional metabolism and growth status, and nutritional characteristics reflecting the nutritional supply situation in the current feeding cycle, can be obtained. These feature vectors are then temporally processed with the sheep's latent state vector from the previous feeding cycle, enabling the model to simultaneously consider multi-source state information from the current feeding cycle and the continuous influence of historical feeding cycles. This avoids relying solely on single environmental data, single weight data, or a single feed ratio for nutritional judgments, improving the accuracy, continuity, and dynamic adaptability of nutrient supplementation recommendations. Clustering, sorting, screening, and averaging are performed separately for growing sheep and mature sheep. Representative clusters with similar physical states and a high proportion within the same growth stage are first screened, and finally, a weighted fusion is performed based on the number of samples in the selected clusters. On the one hand, this approach reduces the interference of abnormal individuals, outliers, or a few sheep in special states on the overall sheep population's physical condition assessment. On the other hand, it takes into account the differences in nutritional requirements, growth rate, and weight changes between growing and mature sheep, making the final sheep population physical condition description vector more representative of the current overall state of the flock. This provides reliable and representative data support for the subsequent training of the nutrition monitoring model, helping to improve the stability and reliability of the model's assessment of the flock's nutritional needs. In addition to constraining the difference between the model's output nutrient addition suggestions and labeled information through a suggestion loss function, a state consistency loss function is further introduced, obtained by comparing the state description vectors of non-true-matched negative sample sheep with those of true-matched sample sheep. This allows the model to enhance its ability to distinguish between true and mismatched feeding states during training, reducing misjudgments caused by insufficient correlation between environmental, physiological, and feed data in the training samples. Therefore, by further combining the above two loss functions, the model can improve both the predictive accuracy of nutrient addition suggestions and the rationality and robustness of multi-source data fusion results. In the calculation of the state consistency loss function, the consistency probability of the negative sample sheep state description vector and the sample sheep state description vector is judged by the fourth multi-layer perception layer. The state consistency loss function is constructed based on the cross-entropy loss, so that the model can learn whether the combination of multi-source data comes from the real matching feeding state. This makes the state consistency loss function not only have a clear training objective, but also facilitate the optimization of model parameters through backpropagation, so as to enhance the model's ability to distinguish between real matching states and non-matching states, and enable the model to learn the consistency relationship between multi-source data more accurately during the training process.

[0019] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Other features and aspects of the invention will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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 embodiments can be obtained based on these drawings without creative effort.

[0021] Figure 1 An exemplary flowchart of a sheep nutrition dynamic monitoring method based on multi-source data fusion according to an embodiment of the present invention is shown.

[0022] Figure 2 A schematic diagram of the structure of a sheep nutrition dynamic monitoring system based on multi-source data fusion according to an embodiment of the present invention is shown as an example. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0024] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0025] Figure 1 An exemplary flowchart illustrates a method for monitoring sheep nutrition dynamics based on multi-source data fusion according to an embodiment of the present invention, the method comprising:

[0026] Step S1: Collect environmental indicators of the pasture at multiple points in the current feeding cycle;

[0027] Step S2: At the end of the current feeding cycle, collect physiological monitoring indicators from multiple sheep.

[0028] Step S3: Obtain the amount of feed used in the current feeding cycle, as well as the ratio of various nutrients in the feed.

[0029] Step S4: Using the trained nutrition monitoring model, process environmental indicators, physiological monitoring indicators, feed dosage and the ratio of various nutrients to obtain information on the addition of various nutrients.

[0030] Step S5: In the next feeding cycle, prepare feed according to the added recommendation information.

[0031] The sheep nutrition dynamic monitoring method based on multi-source data fusion according to the present invention enables the nutrition monitoring model to simultaneously consider multi-source state information of the current feeding cycle and historical feeding state changes, thereby improving the accuracy and dynamic adaptability of nutrient supplementation recommendations.

[0032] According to an embodiment of the present invention, in step S1, environmental indicators of the pasture, such as temperature, humidity, rainfall, and wind speed, are collected at multiple times during the current feeding cycle.

[0033] According to an embodiment of the present invention, in step S2, at the end of the current feeding cycle, physiological monitoring indicators of multiple sheep are collected, including blood indicators and weight indicators, to illustrate the health status, body condition and weight changes of the sheep.

[0034] According to an embodiment of the present invention, in step S3, the amount of feed used in the current feeding cycle and the ratio of various nutrients in the feed are obtained. The amount of feed used may include one or more of the total feeding amount, average feeding amount and actual feed intake. The ratio of various nutrients in the feed can be expressed as the ratio of nutrients such as protein, fat and fiber.

[0035] According to an embodiment of the present invention, in step S4, the trained nutrition monitoring model processes environmental indicators, physiological monitoring indicators, feed dosage, and the ratio of various nutrients to obtain supplementary information for various nutrients, including: obtaining temperature, humidity, and rainfall indicators at multiple times based on the environmental indicators; obtaining sheep blood and weight indicators based on the physiological monitoring indicators; and processing the temperature, humidity, rainfall, blood, weight, feed dosage, and the ratio of various nutrients using the trained nutrition monitoring model to obtain supplementary information for various nutrients.

[0036] According to an embodiment of the present invention, before obtaining supplementary nutrient recommendations through a trained nutrient monitoring model, temperature, humidity, and rainfall indicators at multiple times within the current feeding cycle are first obtained based on environmental indicators. The temperature indicator can be the pasture ambient temperature, the humidity indicator can be the pasture humidity or relative humidity, and the rainfall indicator can be the daily rainfall. Further, based on physiological monitoring indicators, the sheep's blood and body weight indicators are obtained. Then, the temperature, humidity, rainfall, blood, and body weight indicators, feed dosage, and the proportions of various nutrients in the feed are input into the trained nutrient monitoring model for processing, thereby obtaining supplementary nutrient recommendations. The trained nutrient monitoring model can be a combination of multi-source data fusion based on neural networks and a time-series model.

[0037] According to an embodiment of the present invention, a trained nutrition monitoring model processes temperature, humidity, rainfall, blood, body weight, feed dosage, and the ratio of various nutrients to obtain supplementation recommendations for various nutrients. This includes: obtaining an environmental description vector for a single moment based on temperature, humidity, and rainfall; combining environmental description vectors from multiple moments to obtain an environmental description sequence; processing the environmental description sequence through a 1D convolutional layer of the trained nutrition monitoring model to obtain an environmental feature vector; combining the blood and body weight of a single sheep to obtain a sheep physical description vector; obtaining a sheep group physical description vector based on the sheep physical description vectors of multiple sheep; and processing the sheep group physical description vector through a first multi-layer perceptual layer of the trained nutrition monitoring model to obtain a sheep physical feature vector. The second multi-layer perception layer of the trained nutrition monitoring model processes the nutrition description vector, composed of feed dosage and the ratio of various nutrients, to obtain a nutrition feature vector. The environmental feature vector, sheep physical characteristic vector, and nutrition feature vector are concatenated to obtain a sheep state vector. The third multi-layer perception layer of the trained nutrition monitoring model processes the sheep state vector to obtain a sheep state description vector. The sheep state description vector for the current feeding cycle and the sheep state latent vector for the previous feeding cycle are input into the temporal processing layer of the trained nutrition monitoring model for further processing to obtain the sheep state latent vector for the current feeding cycle. If the current feeding cycle is the first feeding cycle, the sheep state latent vector for the previous feeding cycle is set to a zero vector. The sheep state latent vector for the current feeding cycle is then input into the fully connected layer and the activation layer to obtain suggestions for adding various nutrients.

[0038] According to an embodiment of the present invention, the temperature index, humidity index, and rainfall index at a single moment are first combined into a vector, namely, the environmental description vector at a single moment, which can be represented as follows: ,in, This represents the temperature index at time t. This represents the humidity index at time t. Let t represent the rainfall index at time t. This vector is the environmental description vector for that time. Then, by combining the environmental description vectors from multiple times, we obtain the environmental description sequence, such as... Furthermore, the environmental description sequence is processed through the 1D convolutional layers of the trained nutrition monitoring model to extract temporal variation features from the environmental sequence, such as the impact of high temperature, high humidity, and rainfall on sheep feed intake and metabolic state, thereby obtaining an environmental feature vector. Then, the blood and weight indicators of individual sheep are combined to obtain a sheep physical description vector, such as... As shown, where, It can represent the first A collection of blood parameters from a single sheep. This represents a weight index. Multiple sheep body composition description vectors are then statistically analyzed, weighted, or clustered and fused to obtain a group body composition description vector representing the overall flock status. Next, the first multi-layer perception level of the trained nutrition monitoring model performs non-linear feature extraction on the sheep group body composition description vector to obtain a sheep body composition feature vector. Furthermore, the second multi-layer perception level of the trained nutrition monitoring model processes the nutrition description vector, composed of feed quantity and the ratio of various nutrients, to characterize the nutritional supply received by the flock during the current feeding cycle, obtaining a nutrition feature vector. This nutrition description vector can be represented as... This indicates the amount of feed used. to This represents the proportions of nutrients from the 1st to the mth, including but not limited to protein, fat, fiber, vitamins, and trace elements. Furthermore, by fusing multi-source data such as environmental feature vectors, sheep physical feature vectors, and nutritional feature vectors, and concatenating these three, a sheep state vector can be obtained, such as... ,in, Environmental characteristics, representing the impact of the external environment. These are physical characteristics, representing the condition of the sheep flock itself. The sheep state vector represents nutritional characteristics and feed supply. However, the spliced ​​sheep state vector information is quite complex. This invention further fuses and compresses the sheep state vector through the third multi-layer perception layer of the trained nutrition monitoring model to obtain the sheep state description vector. Furthermore, the sheep state description vector of the current feeding cycle and the sheep state latent vector of the previous feeding cycle are input into the temporal processing layer to obtain the sheep state latent vector of the current feeding cycle. This allows the model to consider not only the data of the current cycle but also the state latent vectors of historical feeding cycles to preserve the historical trend of the sheep's nutritional status. If the current feeding cycle is the first cycle, the state latent vector of the previous feeding cycle is set to a vector of all zeros. The temporal processing layer can be a network structure such as RNN, GRU, or LSTM. Finally, the sheep state latent vector of the current feeding cycle is input into the fully connected layer and the activation layer to obtain the addition suggestions for various nutrients. The fully connected layer is used to map the sheep state latent vector to obtain the output results corresponding to various nutrients. The activation layer is used to normalize, range-constrain, or probabilize the output results to generate the addition suggestions for various nutrients. The activation layer can use one or more of the following: Sigmoid function, Softmax function, ReLU function, or linear activation function.

[0039] In this way, the environmental description sequence at multiple time points can be processed through 1D convolutional layers, which can extract the continuous change features of environmental factors within the current feeding cycle. Furthermore, through the first and second multi-layer perception layers, physical characteristics reflecting the nutritional metabolism and growth status of the sheep flock and nutritional characteristics reflecting the nutritional supply situation in the current feeding cycle can be obtained. Then, the above feature vectors are processed temporally with the sheep state latent vectors of the previous feeding cycle, enabling the model to simultaneously consider the multi-source state information of the current feeding cycle and the continuous influence of historical feeding cycles. This avoids nutritional judgments based solely on single environmental data, single weight data, or a single feed ratio, and improves the accuracy, continuity, and dynamic adaptability of nutrient addition recommendations.

[0040] According to an embodiment of the present invention, obtaining a sheep population physical description vector based on multiple sheep physical description vectors includes: clustering the sheep physical description vectors of growing sheep to obtain multiple first clusters; sorting the first clusters according to the number of sheep physical description vectors in the first clusters to obtain a first cluster sequence; filtering according to the sorting in the first cluster sequence until a first ratio between the number of sheep physical description vectors in the selected first clusters and the number of growing sheep reaches a first preset ratio threshold; averaging the sheep physical description vectors in the selected first clusters to obtain a first population physical description vector of growing sheep; and clustering the sheep physical description vectors of mature sheep to obtain... Multiple second clusters are formed. The second clusters are sorted according to the number of sheep physical description vectors in each cluster to obtain a second cluster sequence. The second cluster sequence is then filtered according to its sorting until a second ratio between the number of sheep physical description vectors in the selected second clusters and the number of sheep in the growth stage reaches a second preset ratio threshold. The sheep physical description vectors in the selected second clusters are averaged to obtain the second group physical description vectors of mature sheep. The second group physical description vectors and the first group physical description vectors are weighted and summed based on the number of sheep physical description vectors in the selected second clusters and the number of sheep physical description vectors in the selected first clusters to obtain the sheep group physical description vectors.

[0041] According to an embodiment of the present invention, before inputting sheep physical condition data into the trained nutrition monitoring model, it is necessary to extract the representativeness of the sheep physical condition descriptor vectors. In this part, firstly, clustering is performed based on the physical condition descriptor vectors of each growing sheep, that is, growing sheep with similar physical conditions are grouped into the same category to obtain multiple first clusters. Then, the first clusters are sorted according to the number of sheep physical condition descriptor vectors in the first clusters. The more numerous the clusters, the more representative they are, and therefore they are prioritized. This yields a sequence of first clusters, which is then filtered until the first ratio between the number of sheep physical condition descriptor vectors in the selected first clusters and the number of growing sheep reaches a first preset ratio threshold. For example, the first preset ratio threshold is set to 60%. Assuming there are 500 growing sheep, the selection starts from the most representative cluster and continues until the number of growing sheep selected reaches 300, in order to reduce the impact of a few abnormal individuals or outliers on the results. Finally, the sheep physical description vectors in the selected first cluster are averaged to obtain a comprehensive vector, namely the first group physical description vector of growing sheep, representing the overall physical condition of this type of sheep flock. Correspondingly, to ensure that the sheep group physical description vectors can simultaneously reflect the physical condition of sheep at different growth stages, after obtaining the first group physical description vectors of growing sheep, the second group physical description vectors of mature sheep are further obtained. Similar to the above method, the sheep physical description vectors of mature sheep are clustered to obtain multiple second clusters. These second clusters are then sorted to obtain a second cluster sequence, and a second preset ratio threshold is set. The selection is performed according to the sorting in the second cluster sequence until the second ratio between the number of sheep physical description vectors in the selected second clusters and the number of mature sheep reaches the second preset ratio threshold. Finally, the sheep physical description vectors in the selected second clusters are averaged to obtain the second group physical description vector of mature sheep. Then, the physical description vectors of the second group and the first group are weighted and summed to obtain the physical description vector of the sheep population. This process can be represented as follows: , This represents the physical description vector for the first group. This represents the physical description vector for the second group. This indicates the weight of sheep in their growth stage. This represents the weight corresponding to mature sheep. If there are more growing sheep, the weight of the first group's physical description vector is greater; if there are more mature sheep, the weight of the second group's physical description vector is greater. Therefore, this weight can be the ratio of the number of growing sheep to the total number of sheep, and the ratio of mature sheep to the total number of sheep, respectively.

[0042] This method involves clustering, sorting, screening, and averaging sheep during their growth and maturity stages, respectively. It first identifies representative clusters with similar physical conditions and a high proportion within the same growth stage, then weights and merges them based on the number of samples in each selected cluster. This approach reduces interference from abnormal individuals, outliers, or a few sheep in special conditions on the overall sheep population's physical condition assessment. Furthermore, it takes into account the differences in nutritional requirements, growth rate, and weight changes between growing and mature sheep, resulting in a more accurate representation of the flock's overall physical condition. This provides reliable and representative data support for training subsequent nutrition monitoring models, improving the stability and reliability of the model's assessment of sheep nutritional needs.

[0043] According to an embodiment of the present invention, the training steps of the nutrition monitoring model include: during the training period, acquiring the sample environment description sequence and its corresponding sample environment feature vector; acquiring the sample sheep physical description vector and its corresponding sample sheep physical feature vector; acquiring the sample feed dosage and sample ratio, and its corresponding sample nutrition feature vector; obtaining the sample sheep state description vector based on the sample environment feature vector, the sample sheep physical feature vector, and the sample nutrition feature vector; obtaining the loss function of the nutrition monitoring model based on the sample environment feature vector, the sample sheep physical feature vector, the sample nutrition feature vector, and the sample sheep state description vector; and training the nutrition monitoring model based on the loss function of the nutrition monitoring model to obtain the trained nutrition monitoring model.

[0044] According to an embodiment of the present invention, during the training process of the nutrition monitoring model, training sample data is first acquired. Specifically, temperature, humidity, and rainfall indicators at multiple times within the training period are combined into a sample environment description sequence, and this sequence is input into the 1D convolutional layer of the nutrition monitoring model to be trained to obtain a sample environment feature vector. Simultaneously, blood and weight indicators of the sample sheep are acquired and combined into a sample sheep physical description vector. This vector is then processed through a first multi-layer perceptron layer to obtain a corresponding sample sheep physical feature vector. Further, the sample feed dosage and ratio within the training period are acquired, and the feed dosage and the ratio of various nutrients are combined into a sample nutrition description vector. This vector is then processed through a second multi-layer perceptron layer to obtain a corresponding sample nutrition feature vector. These multi-features are then fused to obtain a sample sheep state description vector, which serves as the loss function for the nutrition monitoring model. The model is then trained to obtain the trained nutrition monitoring model.

[0045] According to an embodiment of the present invention, the loss function of the nutrition monitoring model is obtained based on the sample environmental feature vector, the sample sheep physical feature vector, the sample nutritional feature vector, and the sample sheep state description vector. This includes: obtaining a negative sample sheep state description vector based on the sample sheep physical feature vectors of sheep fed different seasons and different feeds, and the sample environmental feature vectors and sample nutritional feature vectors from the training period; obtaining a state consistency loss function based on the negative sample sheep state description vector and the sample sheep state description vector; obtaining sample addition suggestions for various nutrients based on the sample sheep state description vector; obtaining a suggestion loss function based on the sample addition suggestion information and annotation information; and obtaining the loss function of the nutrition monitoring model based on the state consistency loss function and the suggestion loss function.

[0046] According to an embodiment of the present invention, in order to improve the model's ability to identify matching relationships among multi-source data, physical feature vectors are extracted from sheep samples from different seasons and feeding conditions. These vectors are then combined with environmental and nutritional feature vectors from the current training cycle to create mismatched sample combinations, i.e., negative sample sheep state description vectors. A state consistency loss function is obtained by comparing the negative sample sheep state description vectors with the sample sheep state description vectors, thereby improving the model's learning ability regarding multi-source data matching relationships. Further, the sample sheep state description vectors are input into a suggestion output layer (such as a sigmoid function) composed of fully connected layers and activation layers to obtain sample addition suggestion information. This suggestion information can be represented as the addition ratio or probability of various nutrients predicted by the model. The sample addition suggestion information is then compared with annotation information to calculate the prediction error. The annotation information can be the nutrient addition results or adjustment suggestions predetermined for the sample feeding cycle, which can be derived from livestock experts or historical feeding records. Finally, the state consistency loss function and the suggestion loss function are weighted and summed to obtain the loss function of the nutrition monitoring model. This process can be expressed as follows: ,in, Represents the state consistency loss function. This indicates the proposed loss function. and This represents the loss weighting coefficient.

[0047] In this way, not only is the difference between the nutrient addition suggestions output by the model constrained by the suggestion loss function and the labeled information, but a state consistency loss function is further introduced, obtained by comparing the state description vectors of sheep in non-true matching negative samples with those of truly matching sample sheep. This allows the model to enhance its ability to distinguish between real and mismatched feeding states during training, reducing misjudgments caused by insufficient correlation between environmental, physiological, and feed data in the training samples. Therefore, by combining the above two loss functions, the model can improve both the predictive accuracy of nutrient addition suggestions and the rationality and robustness of multi-source data fusion results.

[0048] According to an embodiment of the present invention, obtaining a state consistency loss function based on a negative sample sheep state description vector and a sample sheep state description vector includes: inputting the negative sample sheep state description vector into a fourth multilayer perceptron layer to obtain first consistency probability information; inputting the sample sheep state description vector into the fourth multilayer perceptron layer to obtain second consistency probability information; constructing a first cross-entropy loss function corresponding to the negative sample sheep state description vector and a second cross-entropy loss function corresponding to the sample sheep state description vector based on the first consistency probability information and the second consistency probability information; and weighted summing the first cross-entropy loss function and the second cross-entropy loss function to obtain the state consistency loss function.

[0049] According to an embodiment of the present invention, the negative sample sheep state description vector is input into the fourth multilayer perceptron to obtain first consistency probability information, which is used to determine the consistency between the current negative sample sheep state description vector and the corresponding environmental and nutritional features. Since it is a negative sample, ideally, the first consistency probability information should be 0. Then, the sample sheep state description vector is input into the fourth multilayer perceptron to obtain second consistency probability information. This process is the same as above. Ideally, the second consistency probability information should be 1. Then, based on the first and second consistency probability information, the first cross-entropy loss function corresponding to the negative sample sheep state description vector and the second cross-entropy loss function corresponding to the sample sheep state description vector are constructed respectively. This process can be expressed as follows: and ,in, This represents the first cross-entropy loss function. This represents the second cross-entropy loss function. This is the first consistency probability information. The second consistency probability information refers to the higher the probability that a negative sample is classified as a consistent state by the model, the larger the first cross-entropy loss function becomes; conversely, the higher the probability that a true sample is classified as a consistent state by the model, the smaller the second cross-entropy loss function becomes. Finally, the first and second cross-entropy loss functions are weighted and summed to obtain the state consistency loss function. This process can be expressed as follows: ,in, and This represents the weighting coefficient for state consistency loss.

[0050] In this way, during the calculation of the state consistency loss function, the consistency probability of the negative sample sheep state description vector and the sample sheep state description vector is judged by the fourth multi-layer perception layer, and the state consistency loss function is constructed based on cross-entropy loss. This allows the model to learn whether the combination of multi-source data comes from the true matching feeding state. This makes the state consistency loss function not only have a clear training objective, but also facilitates the optimization of model parameters through backpropagation, thereby enhancing the model's ability to distinguish between true matching states and non-matching states. This allows the model to learn the consistency relationship between multi-source data more accurately during the training process.

[0051] According to an embodiment of the present invention, in step S5, feed is prepared according to the supplementary information in the next feeding cycle. For example, if the supplementary information output by the nutrient monitoring model is to increase crude protein by 2% and calcium by 0.2%, then when preparing feed in the next feeding cycle, the corresponding proportions of crude protein and calcium are added to the original basic formula so that the adjusted feed meets the nutrient supplementary information output by the model.

[0052] According to an embodiment of the present invention, a method for dynamic monitoring of sheep nutrition based on multi-source data fusion can process environmental description sequences at multiple time points through 1D convolutional layers. This allows for the extraction of continuous change features of environmental factors within the current feeding cycle. Furthermore, through a first and second multi-layer perception layer, it can obtain physical characteristics reflecting the sheep's metabolic and growth states, as well as nutritional characteristics reflecting the nutritional supply situation in the current feeding cycle. These feature vectors are then temporally processed with the sheep's latent state vector from the previous feeding cycle. This enables the model to simultaneously consider multi-source state information from the current feeding cycle and the continuous influence of historical feeding cycles, avoiding reliance on single environmental data, single weight data, or a single feed ratio for nutritional judgments. This improves the accuracy, continuity, and dynamic adaptability of nutrient supplementation recommendations. Clustering, sorting, screening, and averaging are performed separately for growing sheep and mature sheep. Representative clusters with similar physical states and a high proportion within the same growth stage are first selected, and then weighted fusion is performed based on the number of samples in the selected clusters. On the one hand, this approach reduces the interference of abnormal individuals, outliers, or a few sheep in special states on the overall sheep population's physical condition assessment. On the other hand, it takes into account the differences in nutritional requirements, growth rate, and weight changes between growing and mature sheep, making the final sheep population physical condition description vector more representative of the current overall state of the flock. This provides reliable and representative data support for the subsequent training of the nutrition monitoring model, helping to improve the stability and reliability of the model's assessment of the flock's nutritional needs. In addition to constraining the difference between the model's output nutrient addition suggestions and labeled information through a suggestion loss function, a state consistency loss function is further introduced, obtained by comparing the state description vectors of non-true-matched negative sample sheep with those of true-matched sample sheep. This allows the model to enhance its ability to distinguish between true and mismatched feeding states during training, reducing misjudgments caused by insufficient correlation between environmental, physiological, and feed data in the training samples. Therefore, by further combining the above two loss functions, the model can improve both the predictive accuracy of nutrient addition suggestions and the rationality and robustness of multi-source data fusion results. In the calculation of the state consistency loss function, the consistency probability of the negative sample sheep state description vector and the sample sheep state description vector is judged by the fourth multi-layer perception layer. The state consistency loss function is constructed based on the cross-entropy loss, so that the model can learn whether the combination of multi-source data comes from the real matching feeding state. This makes the state consistency loss function not only have a clear training objective, but also facilitate the optimization of model parameters through backpropagation, so as to enhance the model's ability to distinguish between real matching states and non-matching states, and enable the model to learn the consistency relationship between multi-source data more accurately during the training process.

[0053] Figure 2An exemplary schematic diagram of a sheep nutrition dynamic monitoring system based on multi-source data fusion according to an embodiment of the present invention is shown, the system comprising:

[0054] The first data collection module collects environmental indicators of the pasture at multiple points in the current feeding cycle.

[0055] The second data collection module collects physiological monitoring indicators from multiple sheep at the end of the current feeding cycle.

[0056] The third data acquisition module obtains the amount of feed used in the current feeding cycle, as well as the ratio of various nutrients in the feed.

[0057] Add a suggestion information acquisition module, which processes environmental indicators, physiological monitoring indicators, feed dosage and ratio of various nutrients through the trained nutrition monitoring model to obtain suggestion information for the addition of various nutrients.

[0058] The feed preparation module prepares feed according to the added suggestion information in the next feeding cycle.

[0059] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0060] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and any variations or modifications may be made to the implementation of the present invention without departing from the stated principles.

Claims

1. A method for dynamic monitoring of sheep nutrition based on multi-source data fusion, characterized in that, include: Environmental indicators of the pasture were collected at multiple points during the current feeding cycle; At the end of the current feeding cycle, physiological monitoring indicators were collected from multiple sheep. Obtain the amount of feed used in the current feeding cycle, as well as the ratio of various nutrients in the feed; The trained nutrition monitoring model processes environmental indicators, physiological monitoring indicators, feed dosage, and the ratio of various nutrients to obtain recommendations for the addition of various nutrients. In the next feeding cycle, prepare feed according to the added recommendations.

2. The method for dynamic monitoring of sheep nutrition based on multi-source data fusion according to claim 1, characterized in that, The trained nutrition monitoring model processes environmental indicators, physiological monitoring indicators, feed dosage, and the ratio of various nutrients to obtain supplementation recommendations for various nutrients, including: Based on the environmental indicators, temperature, humidity, and rainfall indicators at multiple times are obtained. Based on the physiological monitoring indicators, obtain the sheep's blood and body weight indicators; The trained nutrition monitoring model processes temperature, humidity, rainfall, blood parameters, body weight, feed dosage, and the ratio of various nutrients to obtain recommendations for nutrient addition.

3. The method for dynamic monitoring of sheep nutrition based on multi-source data fusion according to claim 2, characterized in that, The trained nutrition monitoring model processes temperature, humidity, rainfall, blood parameters, body weight, feed dosage, and the ratio of various nutrients to obtain recommended information for nutrient supplementation, including: Based on the temperature, humidity, and rainfall indicators at a single moment, an environmental description vector for that single moment is obtained. The environmental description vectors at multiple time points are combined to obtain the environmental description sequence; The environmental description sequence is processed by the 1D convolutional layer of the trained nutrition monitoring model to obtain the environmental feature vector. The blood and weight indicators of a single sheep are combined to obtain a sheep physical description vector. Based on the physical description vectors of multiple sheep, obtain the physical description vector of the sheep population; The sheep's physical condition description vector is processed by the first multi-layer perception layer of the trained nutrition monitoring model to obtain the sheep's physical condition feature vector. The nutritional feature vector is obtained by processing the nutritional description vector composed of feed dosage and the ratio of various nutrients through the second multi-layer perception layer of the trained nutritional monitoring model. The environmental feature vector, sheep physical feature vector, and nutritional feature vector are concatenated to obtain the sheep state vector. The sheep state vector is processed by the third multi-layer perception layer of the trained nutrition monitoring model to obtain the sheep state description vector. The sheep state description vector of the current feeding cycle and the sheep state latent vector of the previous feeding cycle are input into the time-series processing layer of the trained nutrition monitoring model for processing to obtain the sheep state latent vector of the current feeding cycle. If the current feeding cycle is the first feeding cycle, the sheep state latent vector of the previous feeding cycle is set to an all-zero vector. The hidden vector of the sheep's current feeding cycle is input into the fully connected layer and the activation layer to obtain suggestions for adding various nutrients.

4. The method for dynamic monitoring of sheep nutrition based on multi-source data fusion according to claim 3, characterized in that, Based on the physical description vectors of multiple sheep, a physical description vector for the entire sheep population is obtained, including: The physical description vectors of sheep during the growth period are clustered to obtain multiple first clusters; Based on the number of sheep physical description vectors in the first cluster, the first cluster is sorted to obtain the first cluster sequence; The selection is performed according to the order in the first cluster sequence until the first ratio between the number of sheep physical description vectors in the selected first cluster and the number of sheep in the growth period reaches the first preset ratio threshold. The average of the sheep physical description vectors in the selected first cluster is used to obtain the first group physical description vector of the growing sheep. The physical description vectors of mature sheep are clustered to obtain multiple secondary clusters; The second clusters are sorted according to the number of sheep physical description vectors in the second clusters to obtain the second cluster sequence; The selection is performed according to the order in the second cluster sequence until the second ratio between the number of sheep physical description vectors in the selected second cluster and the number of sheep in the growth stage reaches the second preset ratio threshold. The average of the sheep physical description vectors in the selected second cluster is used to obtain the second group physical description vector of mature sheep. Based on the number of sheep physical description vectors in the selected second cluster and the number of sheep physical description vectors in the selected first cluster, the physical description vectors of the second group and the physical description vectors of the first group are weighted and summed to obtain the physical description vector of the sheep group.

5. The method for dynamic monitoring of sheep nutrition based on multi-source data fusion according to claim 3, characterized in that, The training steps for the nutrition monitoring model include: During the training period, obtain the sample environment description sequence and its corresponding sample environment feature vector; Obtain the sample sheep's physical description vector and its corresponding sample sheep's physical feature vector. Obtain the sample amount and ratio of feed, and the corresponding sample nutrient feature vector; Based on the sample environmental feature vector, the sample sheep physical feature vector, and the sample nutritional feature vector, the sample sheep state description vector is obtained. The loss function of the nutrition monitoring model is obtained based on the sample environmental feature vector, the sample sheep physical feature vector, the sample nutrition feature vector, and the sample sheep state description vector. The nutrition monitoring model is trained based on its loss function to obtain the trained nutrition monitoring model.

6. The method for dynamic monitoring of sheep nutrition based on multi-source data fusion according to claim 5, characterized in that, Based on the sample environmental feature vector, sample sheep physical feature vector, sample nutritional feature vector, and sample sheep state description vector, the loss function of the nutrition monitoring model is obtained, including: Based on the physical feature vectors of sample sheep fed with different feeds in different seasons, and the environmental feature vectors and nutritional feature vectors of the sample sheep in the training period, the negative sample sheep state description vector is obtained. Based on the negative sample sheep state description vector and the sample sheep state description vector, the state consistency loss function is obtained; Based on the sample sheep state description vector, sample supplementation suggestions for various nutrients are obtained; Based on the samples, add suggestion and annotation information to obtain the suggestion loss function; Based on the state consistency loss function and the proposal loss function, the loss function of the nutrition monitoring model is obtained.

7. The method for dynamic monitoring of sheep nutrition based on multi-source data fusion according to claim 6, characterized in that, Based on the negative sample sheep state description vector and the sample sheep state description vector, the state consistency loss function is obtained, including: Input the negative sample sheep state description vector into the fourth multi-layer perception layer to obtain the first consistency probability information; Input the sample sheep state description vector into the fourth multi-layer perception layer to obtain the second consistency probability information; Based on the first consistency probability information and the second consistency probability information, construct the first cross-entropy loss function corresponding to the negative sample sheep state description vector and the second cross-entropy loss function corresponding to the sample sheep state description vector, respectively. The state consistency loss function is obtained by weighted summation of the first cross-entropy loss function and the second cross-entropy loss function.

8. A sheep nutrition dynamic monitoring system based on multi-source data fusion, characterized in that, include: The first data collection module collects environmental indicators of the pasture at multiple points in the current feeding cycle. The second data collection module collects physiological monitoring indicators from multiple sheep at the end of the current feeding cycle. The third data acquisition module obtains the amount of feed used in the current feeding cycle, as well as the ratio of various nutrients in the feed. Add a suggestion information acquisition module, which processes environmental indicators, physiological monitoring indicators, feed dosage and ratio of various nutrients through the trained nutrition monitoring model to obtain suggestion information for the addition of various nutrients. The feed preparation module prepares feed according to the added suggestion information in the next feeding cycle.

Images