Chicken flock average weight estimation method and system based on multi-feature deep fusion learning

Abstract

The application discloses a flock average weight estimation method and system based on multi-feature deep fusion learning, comprising the following steps: collecting flock average weight through an intelligent acquisition device, and storing flock average weight data and flock related information in the database of a server after processing by a data processing program; creating a flock average weight estimation model based on historical flock average weight data and flock related prior knowledge, and predicting flock weight; displaying prediction data through a data visualization module, and converting data information into easy-to-analyze graphics and chart forms, so that the trend and law of the data can be more clearly understood. The application effectively improves the accuracy of livestock and poultry weight prediction, and provides technical support for enterprise flock selling plan and breeding plan making.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent breeding, more particularly, to a chicken flock average weight estimation method and system based on prior knowledge and multi-feature deep fusion learning. BACKGROUND

[0002] In recent years, with the growth of global population and the improvement of living standards, livestock meat production continues to grow, and livestock breeding has become one of the core pillars of the agricultural economy. In the breeding process, livestock weight is one of the focuses, which is directly related to the production performance and economic benefits of livestock breeding. With the deep integration of new generation information technology and livestock breeding, various automatic livestock weight prediction algorithms have been proposed. According to the different technical means used, livestock weight prediction algorithms can be roughly divided into methods based on growth curves, methods based on vision, and methods based on time series.

[0003] The growth curve model has achieved good results in livestock weight prediction, but in actual application scenarios, due to the sparsity, high dynamics and diversity of commercial livestock data, it is difficult for this method to capture the internal correlation information in the data, so the prediction accuracy is limited. Livestock weight estimation based on vision avoids the stress response and injury to livestock and significantly improves the measurement efficiency. However, this method mostly focuses on the extraction of visual features of broilers, ignoring the influence of differences such as broiler species, age, gender, etc. on their weight. In addition, factors such as broiler posture changes and lighting environment can also affect the accuracy of weight estimation. Therefore, it is of great significance to develop a chicken flock average weight estimation method that combines growth curve model and deep network for intelligent management and intelligent decision-making of the breeding industry. SUMMARY

[0004] To solve the above technical problems, the present application provides a chicken flock average weight estimation method and system based on multi-feature deep fusion learning, which can automatically extract high-order features in the data and learn complex nonlinear relationships through a multi-layer neural network structure, fitting the growth curve function model parameters and solving the problem of insufficient parameter fitting degree. Combining the growth curve model and the deep learning network model is particularly suitable for improving the accuracy and timeliness of livestock weight estimation.

[0005] The first aspect of the present application provides a chicken flock average weight estimation method based on multi-feature deep fusion learning, comprising:

[0006] Collecting chicken flock average weight data and chicken flock related information, preprocessing the collected data, and storing the preprocessed data in a server database;

[0007] Based on the historical average weight data and chicken flock related information in the database, a chicken flock average weight estimation model is constructed, and low-order feature learning and high-order feature learning are performed in the chicken flock average weight estimation model.

[0008] The chicken group related information is taken as prior knowledge, parameters fitted by high-order feature learning are used to construct a growth curve function with the prior knowledge, and a chicken group average weight prediction value is output by the growth curve function;

[0009] The output chicken group average weight prediction value is returned to the server, and the chicken group average weight prediction value is visually displayed in the server.

[0010] In this scheme, the chicken group average weight data to be measured and the chicken group related information are collected, the collected data is preprocessed, and the preprocessed data is stored in the server database, specifically:

[0011] The chicken group to be measured is precisely weighed by using a full-automatic whole chicken weighing device, and the chicken group average weight is automatically calculated and obtained, at the same time, the chicken group related information is collected, including the original age;

[0012] The collected data is subjected to outlier, repeated value elimination and missing value completion operation, a preprocessed historical average weight data set and a chicken group related information data set are constructed, and the preprocessed historical average weight data set and the chicken group related information data set are stored in the server database.

[0013] In this scheme, the chicken group average weight estimation model includes a low-order feature learning module, a high-order feature learning module and a prediction module;

[0014] The historical average weight data in the database is imported into the chicken group average weight estimation model, in the low-order feature learning module, an embedding layer composed of a full connection layer is used for feature embedding, and the full connection layer is subjected to Dropout processing to prevent overfitting;

[0015] The historical average weight data is subjected to feature dimension reduction by using a full connection layer, the original high-dimensional sparse vector is mapped into a low-dimensional dense vector, and an embedding vector v x of chicken growth information features is obtained, v n = {x n v i}, wherein v i ∈R is an embedding vector of the i-th historical average weight data corresponding feature vector, x x ∈R represents the feature vector corresponding to the i-th historical average weight data, i = 1, 2, …, n, and n represents the total number of feature vectors;

[0016] The low-dimensional dense vector v BI obtained by the embedding layer is imported into the second-order cross layer in the low-order feature learning module as an input vector, a pairwise cross product operation is performed on the input vector, then the operation result is added, and a low-dimensional cross dense vector f x is obtained.

[0017] In the chicken group average weight estimation model, the low-dimensional cross-dense vector learned by the low-order feature learning module is input into the high-order feature learning module, and the high-order feature learning module includes an MLP neural network and a SoftPlus function.

[0018] The MLP neural network is used to learn the low-dimensional cross-dense vector, capture the high-order complex internal relationship between feature vectors, fit the growth curve prediction function parameters, obtain the key growth parameters of the output of the MLP neural network, and use the SoftPlus function for further parameter fitting.

[0019] In the present scheme, the dimensions of the MLP neural network are set to [512, 128, 64, 32, 8], the network parameters of the MLP neural network are updated and optimized using stochastic gradient descent, and the model formula is represented as:

[0020] z1=σ1(W1f BI (v x )+b1)

[0021] z2=σ2(W2z1+b2)

[0022] z3=σ3(W3z2+b3)

[0023] z4=σ4(W4z3+b4)

[0024] z5=σ5(W5z3+b5)

[0025] y=z5={slope,height,offset}

[0026] Where Z l represents the l-th hidden layer, l=1, 2, 3, 4, 5, W l ,b l and σ represent the weight, bias vector and activation function of the l-th layer, respectively, f BI (v x ) represents a low-dimensional cross-dense vector, slope represents the growth rate of the chicken group, y represents the key growth parameter, height represents the growth rate of the chicken group, and offset represents the offset.

[0027] In the present scheme, the prediction module of the chicken group average weight estimation model constructs a growth curve function based on the key growth parameters fitted by the high-order feature learning module and the original age characteristics, uses the growth curve function to predict the chicken group average weight, and outputs the chicken group average weight prediction value.

[0028] The growth curve function f(slope, height, offset, age) is represented as:

[0029]

[0030] Wherein, slope represents the growth rate of the chicken population, height represents the chicken population growth rate, offset represents the offset, and age represents the original age characteristics.

[0031] In the scheme, the average weight of the chicken population estimated by the chicken population average weight prediction value output by the model is visualized and displayed, and the chicken population average weight prediction value is converted into a graphical or tabular form.

[0032] The second aspect of the present application provides a chicken population average weight estimation system based on multi-feature deep fusion learning, which comprises a data acquisition module, a chicken population average weight estimation module and a data visualization module.

[0033] The data acquisition module is responsible for weighing the chicken population to be measured by a full-automatic whole chicken weighing device, collecting chicken population average weight data and chicken population related information, preprocessing the collected data, and storing the processed data in a database server.

[0034] The chicken population average weight prediction module is responsible for constructing a chicken population average weight estimation model based on historical average weight data and chicken population related information in the database, predicting the chicken population average weight for a specified length of time using the chicken population average weight estimation model, and storing the chicken population average weight prediction value in the database of the server.

[0035] The data visualization module is responsible for visualizing the chicken population average weight prediction value, retrieving the chicken population average weight prediction value from the database, and displaying the data in a graphical or tabular form.

[0036] Compared with the prior art, the present application has the following advantages:

[0037] (1) The present application designs a new growth curve prediction function, which has the characteristics of efficient and accurate prediction of livestock and poultry body weight compared with the previous growth curve function. The combination of the growth curve model and the deep learning network model is particularly suitable for improving the accuracy and timeliness of livestock and poultry weight estimation.

[0038] (2) The present application constructs a chicken population average weight estimation model based on prior knowledge and multi-feature deep fusion learning, which combines a low-order feature learning module, a high-order feature learning module and a prediction module, effectively improves the accuracy of livestock and poultry weight prediction, and through a multi-layer neural network structure, the chicken population average weight estimation model can automatically extract high-order features from historical chicken population average weight data and learn complex nonlinear relationships to fit the growth curve function model parameters, thereby solving the problem of insufficient parameter fitting degree. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or examples. Obviously, the drawings in the following description only some of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0040] Figure 1 A flowchart of a chicken flock average weight estimation method based on multi-feature deep fusion learning is shown.

[0041] Figure 2 A framework diagram of a chicken flock average weight estimation model in the embodiment is shown.

[0042] Figure 3 A framework diagram of a chicken flock average weight estimation system based on multi-feature deep fusion learning is shown. DETAILED DESCRIPTION

[0043] In order to more clearly illustrate the technical solutions in the embodiments or examples of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or examples. Obviously, the drawings in the following description only some of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0044] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0045] Figure 1 A flowchart of a chicken flock average weight estimation method based on multi-feature deep fusion learning is shown.

[0046] As Figure 1 shown, the present embodiment provides a chicken flock average weight estimation method based on multi-feature deep fusion learning, comprising:

[0047] S102, collecting chicken flock average weight data to be measured and chicken flock related information, preprocessing the collected data, and storing the preprocessed data in a server database;

[0048] S104, constructing a chicken flock average weight estimation model based on historical average weight data and chicken flock related information in the database, and performing low-order feature learning and high-order feature learning in the chicken flock average weight estimation model;

[0049] S106, taking the chicken flock related information as prior knowledge, constructing a growth curve function using the parameters fitted out by the high-order feature learning and the prior knowledge, and outputting a chicken flock average weight prediction value through the growth curve function;

[0050] S108, return the output flock average weight prediction value to the server, and visualize the flock average weight prediction value in the server.

[0051] It should be noted that the to-be-tested flock is sent into the full-automatic whole chicken weighing device, the full-automatic whole chicken weighing device includes an electronic scale and a sensor, the to-be-tested flock is accurately weighed by using the full-automatic whole chicken weighing device, and the flock average weight is automatically calculated and obtained, meanwhile, related information of the flock is collected, including original age; in order to ensure the accuracy and integrity of the data, the collected data is subjected to outlier, repeated value elimination and missing value completion operations, a preprocessed historical average weight dataset and a flock related information dataset are constructed, and the preprocessed historical average weight dataset and the flock related information dataset are stored in a server database, for subsequent data analysis and breeding management tasks.

[0052] Figure 2 A framework diagram of the flock average weight estimation model in the embodiment is shown.

[0053] According to the embodiment of the present application, the flock average weight estimation model includes a low-order feature learning module, a high-order feature learning module and a prediction module, specifically:

[0054] The low-order feature learning module includes embedding learning and second-order feature learning, the historical average weight data in the database is imported into the flock average weight estimation model, in the low-order feature learning module, an embedding layer composed of a full connection layer is used for feature embedding, the full connection layer is subjected to Dropout processing to prevent overfitting; a full connection layer is used for feature dimension reduction of the historical average weight data, an original high-dimensional sparse vector is mapped into a low-dimensional dense vector, a challenge brought by data sparsity is solved, and an embedding vector v x of the flock growth information feature is obtained. n , n} is obtained, where v i ∈R is an embedding vector of the feature vector corresponding to the i-th historical average weight data, x i ∈R represents the feature vector corresponding to the i-th historical average weight data, i=1, 2, …, n, and n represents the total number of feature vectors.

[0055] The low-dimensional dense vector v x obtained by the embedding layer is imported into a second-order cross layer in the low-order feature learning module as an input vector, two-by-two cross product operation is performed on the input vector, then the operation result is added, and the second-order feature interaction pooling layer outputs a low-dimensional cross dense vector f BI (V x ), the features are interacted two by two, so that the model learns more implicit relationships between the features, and the formula for calculating the vector of the second-order feature interaction pooling layer is:

[0056]

[0057] wherein x i , x j denote the feature vectors corresponding to the i-th and j-th historical average weight data respectively, v i , v j denote the embedding vectors of the feature vectors corresponding to the i-th and j-th historical average weight data respectively, and denotes the product between the feature vectors.

[0058] In the flock average weight estimation model, the low-dimensional cross-dense vector learned by the low-order feature learning module is input into the high-order feature learning module, and the high-order feature learning module includes an MLP neural network and a SoftPlus function; the MLP neural network is used to learn the low-dimensional cross-dense vector, capture the high-order complex internal relationship between the feature vectors, and further alleviate the problems caused by data sparsity, high dynamicity and diversity. The MLP neural network is used to fit the growth curve prediction function parameters, output the key growth parameters, and the SoftPlus function is used for further parameter fitting. The SoftPlus function is responsible for further fitting the key growth parameters learned by the MLP neural network to avoid negative numbers, outliers, etc., thereby improving the prediction accuracy and robustness of the model.

[0059] In the MLP neural network learning, the low-dimensional cross-dense vector learned in the low-order feature learning module is input into the MLP neural network to capture the complex internal relationship between the cross features. The dimension of the MLP neural network is set to [512, 128, 64, 32, 8], and the model formula is represented as:

[0060] z1 = σ1(W1f BI (v x )+b1)

[0061] z2 = σ2(W2z1+b2)

[0062] z3 = σ3(W3z2+b3)

[0063] z4 = σ4(W4z3+b4)

[0064] z5 = σ5(W5z3+b5)

[0065] y = z5 = {slope, height, offset}

[0066] wherein Z l denotes the l-th hidden layer, l = 1, 2, 3, 4, 5, W l , b land σ represent the weight, bias vector and activation function of the l-th layer, preferably, the activation function is a nonlinear activation function Relu, f BI (v x ) represents a low-dimensional cross-dense vector, slope represents the growth rate of the chicken population, y represents the key growth parameter, height represents the growth rate of the chicken population, and offset represents the offset.

[0067] The loss function of the MLP neural network is generated according to the mean square error, and the network parameters of the MLP neural network are updated and optimized using stochastic gradient descent, a training sample x is randomly selected, each network parameter is updated in the direction of its negative gradient, and the update process of the network parameter is represented as:

[0068]

[0069] where θ represents the trainable network parameter, η represents the learning rate that controls the gradient descent step size, y(x) represents the predicted value of the training sample x.

[0070] Before the chicken population average weight estimation model receives data, BN normalization is performed on each feature vector to reduce the internal covariate shift of each layer in the MLP neural network, thereby accelerating the convergence process of the model. Let the input vector x of a certain layer be i ∈R, and all input vectors of each batch are B=x i The formula of BN normalization is as follows where μ represents the small batch mean, σ represents the small batch variance, γ and β represent trainable parameters for scaling to improve the representation ability of the model, and ∈ represents a constant to prevent the variance from being 0.

[0071] The parameters slope and height fitted based on the MLP neural network model are further parameter fitted via SoftPlus fitting, and the SoftPlus function is defined as follows: x represents the function parameter. By fitting the key growth parameters slope and height in the chicken population growth curve function, the generalization and accuracy of the model are improved, thereby effectively alleviating the problems of sparse, highly dynamic and diverse chicken population growth information.

[0072] The prediction module of the chicken population average weight estimation model constructs a growth curve function based on the key growth parameters fitted by the high-order feature learning module combined with the original age characteristics, uses the growth curve function to predict the chicken population average weight, and outputs the chicken population average weight prediction value;

[0073] The growth curve function f(slope, height, offset, age) is represented as:

[0074]

[0075] Wherein, slope represents the growth rate of the chicken population, height represents the growth rate of the chicken population, offset represents the offset, and age represents the original age characteristics.

[0076] The chicken population average weight prediction value output by the chicken population average weight estimation model is visualized, and the chicken population average weight prediction value is converted into a graphical or tabular form, so that the trend and rule of the data can be more clearly understood, thereby providing strong support for decision-making.

[0077] Figure 3 A framework diagram of a chicken population average weight estimation system based on multi-feature deep fusion learning is shown.

[0078] The second embodiment of the present application provides a chicken population average weight estimation system based on multi-feature deep fusion learning, which comprises a data acquisition module, a chicken population average weight estimation module and a data visualization module.

[0079] The data acquisition module is responsible for weighing the to-be-tested chicken population by a full-automatic whole chicken weighing device, acquiring chicken population average weight data and chicken population related information, preprocessing the acquired data, and storing the processed data into a database server.

[0080] The data acquisition module acquires data in a day granularity, acquires chicken population data in batches through an intelligent data device, and stores the data into a server database. The data preprocessing program takes out the chicken population data from the server, performs preprocessing such as outlier and duplicate value elimination and missing value completion on the data, and stores the processed data set into the server database.

[0081] The chicken population average weight prediction module is responsible for constructing a chicken population average weight estimation model based on historical average weight data and chicken population related information in the database. In the chicken population average weight estimation model, the fitted chicken population growth rate, chicken population growth rate, offset and other parameters are put into the proposed growth curve function together with the prior knowledge such as age, and the chicken population average weight estimation model is used to predict the chicken population average weight in a specified time length. The chicken population average weight prediction value is stored into the server database.

[0082] The data visualization module is responsible for visualizing the chicken population average weight prediction value, taking out the chicken population average weight prediction value from the database, and displaying the data in a graphical or tabular form. By converting the data information into an easily analyzed graphical or tabular form, the trend and rule of the data can be more clearly understood, thereby providing strong support for decision-making.

[0083] The third embodiment of the present application provides a computer readable storage medium, wherein a chicken flock average weight estimation method based on multi-feature deep fusion learning is stored in the computer readable storage medium.

[0084] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other manners. The embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling, or direct coupling or communication connection between the components can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0085] In addition, each function unit in the embodiments of the present application can be integrated into one processing unit, each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in the form of hardware, or in the form of hardware plus software function unit.

[0086] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by a program instruction related hardware, and the foregoing program can be stored in a computer readable storage medium, and the program is executed to perform the steps of the above-mentioned method embodiments; and the foregoing storage medium includes: mobile storage equipment, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk and various storage program codes.

[0087] Alternatively, the integrated unit of the present application, if implemented in the form of a software function module and sold or used as an independent product, can also be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, and includes several instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the methods described in the embodiments of the present application. The foregoing storage medium includes: mobile storage equipment, ROM, RAM, magnetic disk or optical disk, and various storage program codes.

[0088] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A chicken flock average weight estimation method based on multi-feature deep fusion learning, characterized in that, The method comprises the following steps: Collecting the average weight data of the test chicken group and the related information of the chicken group, preprocessing the collected data, and storing the preprocessed data in the server database; Based on the historical average weight data and the related information of the chicken group in the database, a chicken group average weight estimation model is constructed, and low-order feature learning and high-order feature learning are performed in the chicken group average weight estimation model; The related information of the chicken group is used as prior knowledge, and the parameters fitted by high-order feature learning are combined with the prior knowledge to construct a growth curve function, and the chicken group average weight prediction value is output by the growth curve function; The output chicken group average weight prediction value is returned to the server, and the chicken group average weight prediction value is visualized in the server; The chicken group average weight estimation model comprises a low-order feature learning module, a high-order feature learning module, and a prediction module; The historical average weight data in the database is imported into the chicken group average weight estimation model, and the embedding layer composed of a fully connected layer is used for feature embedding in the low-order feature learning module, and the fully connected layer is subjected to Dropout processing to prevent overfitting; The historical average weight data is subjected to feature dimension reduction by using a full connection layer, and original high-dimensional sparse vectors are mapped into low-dimensional dense vectors to obtain embedding vectors of chicken growth information features wherein it is assumed that is an embedding vector of a feature vector corresponding to the i-th historical average weight data, is a feature vector corresponding to the i-th historical average weight data, , indicates the total number of feature vectors;​​ The embedding layer obtains a low-dimensional dense vector The second-order cross layer in the low-order feature learning module is imported as an input vector, a two-by-two cross product operation is performed on the input vector, then the operation result is added, and a low-dimensional cross dense vector is obtained ; In the chicken group average weight estimation model, the low-dimensional cross-dense vector learned by the low-order feature learning module is input into the high-order feature learning module, and the high-order feature learning module comprises an MLP neural network and a SoftPlus function; The MLP neural network is used to learn the low-dimensional cross-dense vector, capture the high-order complex internal relationship between the feature vectors, fit the growth curve prediction function parameters, obtain the key growth parameters of the MLP neural network output, and use the SoftPlus function for further parameter fitting; The dimension of the MLP neural network is set as The network parameters of the MLP neural network are updated and optimized using stochastic gradient descent, and the model formula is represented as: ， ， ， ， ， ， wherein, denote the first layer hidden layer, , and denote the first layer weight, bias vector and activation function, respectively, denote the low-dimensional cross-dense vector, denote the growth rate of the flock, denote the key growth parameters, denote the growth rate of the flock, denote the offset The prediction module of the chicken group average weight estimation model constructs a growth curve function based on the key growth parameters fitted by the high-order feature learning module and the original age characteristics, uses the growth curve function to predict the chicken group average weight, and outputs the chicken group average weight prediction value; The growth curve function is represented as: ， wherein, represents the growth rate of the flock, represents the growth rate of the flock, represents the offset, represents the original age characteristic.

2. The flock average weight estimation method based on multi-feature deep fusion learning according to claim 1, characterized in that, Collecting the average weight data of the test chicken group and the related information of the chicken group, preprocessing the collected data, and storing the preprocessed data in the server database, specifically: The test chicken group is accurately weighed by using a full-automatic whole chicken weighing device, and the chicken group average weight is automatically calculated and obtained, and at the same time, the related information of the chicken group is collected, including the original age; The collected data is subjected to outlier and repeated value elimination and missing value completion operations, and a preprocessed historical average weight data set and a chicken group related information data set are constructed, and the preprocessed historical average weight data set and the chicken group related information data set are stored in the server database.

3. The flock average weight estimation method based on multi-feature deep fusion learning according to claim 1, characterized in that, The chicken group average weight prediction value output by the chicken group average weight estimation model is visualized, and the chicken group average weight prediction value is converted into a graph or chart form.

4. A chicken flock average weight estimation system based on multi-feature deep fusion learning, characterized in that, The chicken group average weight estimation method based on multi-feature deep fusion learning is realized, and the system comprises a data collection module, a chicken group average weight estimation module, and a data visualization module; The data collection module is responsible for weighing the test chicken group by using a full-automatic whole chicken weighing device, collecting the chicken group average weight data and the related information of the chicken group, preprocessing the collected data, and storing the processed data in the database server; The flock average weight prediction module is responsible for constructing a flock average weight estimation model based on historical average weight data and flock related information in the database, predicting flock average weight in a specified time length using the flock average weight estimation model, and storing the flock average weight prediction value in the database of the server. The data visualization module is responsible for visualizing the flock average weight prediction value, taking the flock average weight prediction value from the database, and displaying the data in the form of graphs and charts.

Images