Aquaculture environment control method and system based on machine learning

Abstract

The present application relates to the field of image recognition technology, and discloses a method and system for regulating and controlling aquaculture environment based on machine learning, the method comprising: collecting environmental parameters of a farm; using a camera system installed in the farm to capture images of animals in the farm; analyzing the images of the animals through a preset convolutional neural network to extract the behavioral feature vectors of the animals; inputting the environmental parameters of the farm and the behavioral feature vectors of the animals into a preset deep neural network, and outputting the comprehensive behavioral index of the animals; and regulating and controlling the environmental control equipment of the farm according to the comprehensive behavioral index of the animals. The present invention provides an intelligent method for regulating and controlling aquaculture environment by introducing advanced machine learning technology, especially convolutional neural networks (CNNs) and deep neural networks (DNNs). The method can realize accurate monitoring and analysis of environmental parameters of farms and animal behavioral characteristics, and dynamically adjust environmental control equipment according to the analysis results.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and more specifically, to a method and system for regulating aquaculture environment based on machine learning. Background Art

[0002] Environmental control in traditional farms mainly relies on manual experience and fixed parameter settings, such as temperature, humidity, light cycle, etc., which often cannot fully meet the specific environmental needs of animals of different species and growth stages, and may affect the growth efficiency, health status and product quality of animals.

[0003] In recent years, the development of machine learning technology and sensor technology has provided new ways to optimize the breeding environment. Various sensors installed in the farm can monitor environmental parameters in real time, such as temperature, humidity, ammonia concentration, carbon dioxide concentration, etc., and use these data for analysis and decision-making. At the same time, the advancement of image recognition technology makes it possible to use camera systems to capture and analyze animal behavior. Computer vision algorithms can extract animal behavior characteristics from images, such as eating, drinking, resting, exercise and other behavior patterns, which is of great significance for assessing the health status and welfare level of animals.

[0004] However, most of the image analysis in the existing technology remains at the surface level, lacking a deep understanding and quantification of animal behavior. In addition, the relationship between environmental parameters and animal behavior is complex, and it is difficult to accurately reflect the impact of environmental factors on animal behavior simply by relying on experience or simple data analysis. Therefore, there is an urgent need for a method that can comprehensively consider environmental parameters and animal behavior characteristics to guide the environmental regulation of farms in order to achieve more accurate and efficient management. Summary of the invention

[0005] In order to solve the above-mentioned technical problems, the present application is proposed to provide a method and system for aquaculture environment regulation based on machine learning, which aims to guide the environmental regulation of farms by comprehensively considering environmental parameters and animal behavior characteristics, so as to achieve more accurate and efficient management.

[0006] In a first aspect, the present invention provides a method for regulating aquaculture environment based on machine learning, comprising: using sensors installed in a farm to collect environmental parameters of the farm; using a camera system installed in the farm to capture images of animals in the farm; analyzing the images of the animals through a preset convolutional neural network to extract behavioral feature vectors of the animals; inputting the environmental parameters of the farm and the behavioral feature vectors of the animals into a preset deep neural network, and outputting a comprehensive behavioral index of the animals, indicating the level of influence of the farm environment on the behavior of the animals; and regulating the environmental control equipment of the farm according to the comprehensive behavioral index of the animals.

[0007] Optionally, in the aforementioned breeding environment control method based on machine learning, there are multiple animals, a marker is implanted in each animal, and images of the animals in the breeding farm are captured using a camera system installed in the breeding farm, including: detecting the positions of all markers implanted in the animals; calculating the center positions of all the markers; and adjusting the shooting direction and focal length of the camera system according to the center positions of all the markers.

[0008] Optionally, in the aforementioned breeding environment control method based on machine learning, the image of the animal is analyzed by a preset convolutional neural network to extract the behavioral feature vector of the animal, including: extracting multiple levels of feature maps of the image of the animal through multiple convolutional layers in the convolutional neural network; setting weight values ​​for the multiple levels of feature maps of the image of the animal according to the type of the animal; synthesizing the multiple levels of feature maps of the image, and the synthesized feature maps ,in, is the preset benchmark weight value, is the preset reference magnification value, is the type of animal, The reference size of the type animal. is the number of the multiple convolutional layers, It is a feature map generated by the i-th convolutional layer in the input to output direction of the convolutional neural network.

[0009] Optionally, in the aforementioned breeding environment control method based on machine learning, the image of the animal is analyzed by a preset convolutional neural network to extract the behavioral feature vector of the animal, and it also includes: performing preliminary dimensionality reduction on the synthesized feature map according to a preset linear dimensionality reduction method, and obtaining the values ​​of n features of the feature map after the preliminary dimensionality reduction; calculating the feature mean of multiple known feature maps; performing secondary dimensionality reduction on the feature map, and obtaining the value of the jth feature of the feature map after the secondary dimensionality reduction. ,in, is the feature mean of the multiple feature maps, is the number of all features contained in the feature graph before dimensionality reduction, is the value of the kth feature of the feature map before dimensionality reduction, is the variance of the kth feature, is the preset weight value of the j-th feature, used to indicate the importance of the j-th feature.

[0010] Optionally, in the aforementioned breeding environment control method based on machine learning, before inputting the environmental parameters of the breeding farm and the behavioral feature vectors of the animals into a preset deep neural network, the method further includes: detecting the absolute value of each weight value in the deep neural network; calculating the gradient information of each weight value in the deep neural network with respect to the loss function; counting the number of non-zero connections and the total number of connections at the level where each weight value in the deep neural network is located; calculating the contribution score of each weight value in the deep neural network to indicate the degree of influence of the weight value on the output result of the deep neural network, wherein the first Weight value Contribution score ,in, is the preset balance coefficient, For the Weight value The absolute value of and Respectively Weight value The number of non-zero connections and the total number of connections at the level, For the Weight value Gradient information about the loss function; selecting a group of weight values ​​from the deep neural network and pruning the connections of the group of weight values ​​according to the contribution score of each weight value in the deep neural network.

[0011] Optionally, in the aforementioned aquaculture environment control method based on machine learning, a group of weight values ​​are selected from the deep neural network and the connections of the group of weight values ​​are pruned according to the contribution score of each weight value in the deep neural network, including: calculating the maximum operation time of the deep neural network according to the preset maximum control interval of the environmental control equipment of the farm; calculating the maximum number of connections inside the deep neural network according to the maximum operation time of the deep neural network; and determining the number of connections to be pruned from the deep neural network according to the maximum number of connections inside the deep neural network.

[0012] Optionally, in the aforementioned breeding environment control method based on machine learning, before inputting the environmental parameters of the breeding farm and the behavioral feature vectors of the animals into a preset deep neural network, the method further comprises: when training the deep neural network, calculating the behavioral consistency loss of the deep neural network , where T is the preset time series length, and Respectively represent the deep neural network in Moment and The comprehensive behavior index output at all times, express and The Euclidean distance between them; According to the behavioral consistency loss of the deep neural network , calculate the value of the loss function of the deep neural network; and adjust the deep neural network according to the value of the loss function of the deep neural network.

[0013] Optionally, in the aforementioned aquaculture environment control method based on machine learning, before adjusting the deep neural network according to the value of the loss function of the deep neural network, the method further includes: calculating the sparsity loss of the deep neural network ,in, is the set of weight values ​​in the deep neural network, represents any weight value in the deep neural network; based on the sparsity loss of the deep neural network , the value of the loss function of the deep neural network is corrected.

[0014] In a second aspect, the present invention provides a breeding environment control system based on machine learning, comprising: a parameter acquisition module, which uses sensors installed in the breeding farm to collect environmental parameters of the breeding farm; an image capture module, which uses a camera system installed in the breeding farm to capture images of animals in the breeding farm; a feature extraction module, which analyzes the images of the animals through a preset convolutional neural network to extract behavioral feature vectors of the animals; an index calculation module, which inputs the environmental parameters of the breeding farm and the behavioral feature vectors of the animals into a preset deep neural network, and outputs a comprehensive behavioral index of the animals, indicating the level of influence of the environment of the breeding farm on the behavior of the animals; and an equipment control module, which controls the environmental control equipment of the breeding farm according to the comprehensive behavioral index of the animals.

[0015] The above one or more technical solutions of the present invention have at least one or more of the following beneficial effects:

[0016] According to the technical solution of the present invention, by introducing advanced machine learning technology, especially convolutional neural network (CNN) and deep neural network (DNN), an intelligent breeding environment control method is provided. The method can realize accurate monitoring and analysis of breeding farm environmental parameters and animal behavior characteristics, and dynamically adjust environmental control equipment according to the analysis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0018] Figure 1 It is a flow chart of a method for controlling aquaculture environment based on machine learning according to an embodiment of the present application;

[0019] Figure 2 It is a partial flow chart of the aquaculture environment control method based on machine learning according to an embodiment of the present application;

[0020] Figure 3 Another layout flow chart of the aquaculture environment control method based on machine learning according to an embodiment of the present application;

[0021] Figure 4 is another partial flow chart of the aquaculture environment control method based on machine learning according to an embodiment of the present application;

[0022] Figure 5 This is another partial flow chart of the aquaculture environment control method based on machine learning according to an embodiment of the present application;

[0023] Figure 6 It is a block diagram of a breeding environment control system based on machine learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] Some embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0025] It should be noted that “weight” and “weight value” have equivalent meanings in this application.

[0026] like Figure 1 As shown, in one embodiment of the present invention, a method for regulating and controlling aquaculture environment based on machine learning is provided, comprising:

[0027] Step S110, using sensors installed in the farm to collect environmental parameters of the farm;

[0028] Step S120, using a camera system installed in the farm to capture images of animals in the farm;

[0029] In this embodiment, by accurately capturing and analyzing the behavioral characteristics of animals, the needs and health status of animals can be more accurately understood, and environmental conditions can be adjusted in a timely manner, thereby improving the quality of life of animals, reducing stress reactions, and promoting their healthy growth;

[0030] Step S130, analyzing the image of the animal through a preset convolutional neural network to extract the animal's behavioral feature vector;

[0031] Step S140, inputting the environmental parameters of the farm and the behavioral feature vectors of the animals into a preset deep neural network, and outputting the comprehensive behavioral index of the animals, which indicates the level of influence of the farm environment on the behavior of the animals;

[0032] In this embodiment, by using deep learning algorithms, especially convolutional neural networks and deep neural networks, valuable behavioral feature vectors can be extracted from a large amount of image data, achieving a deep understanding and quantification of animal behavior;

[0033] Step S150, regulating the environmental control equipment of the breeding farm according to the comprehensive behavior index of the animals;

[0034] In this embodiment, the environmental control device performs intelligent regulation based on the comprehensive behavior index, thereby avoiding energy waste in the traditional fixed mode, improving the efficiency of energy and other resource utilization, and reducing operating costs.

[0035] According to the technical solution of this embodiment, by introducing advanced machine learning technology, especially convolutional neural network (CNN) and deep neural network (DNN), an intelligent breeding environment control method is provided. This method can realize accurate monitoring and analysis of farm environmental parameters and animal behavior characteristics, and dynamically adjust environmental control equipment according to the analysis results.

[0036] like Figure 2 As shown, another embodiment of the present invention further provides a method for regulating aquaculture environment based on machine learning. Compared with the aforementioned embodiment, the method for regulating aquaculture environment based on machine learning in this embodiment has a plurality of animals, and a marker is implanted in each animal. Step S120 includes:

[0037] Step S210, detecting the positions of all markers implanted in the animal.

[0038] In this embodiment, by implanting markers in each animal and accurately detecting the location of these markers, the behavior patterns of individual animals can be tracked and analyzed more accurately. Compared with relying solely on external image recognition, this method can provide more detailed and reliable behavior data, which helps to deeply understand the behavioral characteristics of animals.

[0039] Step S220, calculating the center positions of all markers.

[0040] Step S230: adjusting the shooting direction and shooting focal length of the camera system according to the center positions of all the markers.

[0041] According to the technical solution of this embodiment, the shooting direction and focal length of the camera system are automatically adjusted according to the center position of all markers, ensuring that the camera always focuses on the core area of ​​the animal group. This not only improves the quality of image capture, but also ensures that clear and stable images can be obtained even when the animals are active frequently or densely, thereby providing high-quality data support for subsequent behavioral analysis.

[0042] like Figure 3 As shown, another embodiment of the present invention further provides a method for regulating and controlling an aquaculture environment based on machine learning. Compared with the aforementioned embodiment, the method for regulating and controlling an aquaculture environment based on machine learning in this embodiment, step S130 includes:

[0043] Step S310, extracting feature maps of multiple levels of the animal image through multiple convolutional layers in the convolutional neural network.

[0044] In this embodiment, multiple layers of feature maps of animal images are extracted layer by layer through multiple convolutional layers in a convolutional neural network (CNN). This method can capture feature information of different scales and complexities in the image. Compared with single-level feature extraction, multi-level feature maps provide richer and more detailed descriptions of behavioral features, which helps to understand the behavioral patterns of animals more accurately.

[0045] Step S320, according to the type of the animal, weight values ​​are set for the feature maps of multiple levels of the image of the animal.

[0046] In this embodiment, weight values ​​are set for feature maps at each level according to the animal type. This process takes into account the differences in visual features of different types of animals. By tailoring feature map weight values ​​for each animal, more focus can be placed on the features that are most critical to the behavior analysis of a specific animal, thereby improving the relevance and accuracy of the analysis results.

[0047] Step S330, synthesizing the feature maps of multiple levels of the image, and the synthesized feature maps ,in, is the preset benchmark weight value, is the preset reference magnification value, For the type of animal, The reference size of the type animal. is the number of convolutional layers, It is the feature map generated by the i-th convolutional layer in the input to output direction of the convolutional neural network.

[0048] In this embodiment, the synthesis formula of the feature map not only combines the feature maps generated by all convolutional layers, but also adjusts them through factors such as the reference weight value, the magnification value, and the reference size. This intelligent synthesis method ensures that the final feature map retains the key information of the original image and can highlight the individual characteristics of the animal, laying a solid foundation for subsequent behavioral analysis.

[0049] Step S340, performing preliminary dimensionality reduction on the synthesized feature map according to a preset linear dimensionality reduction method, and obtaining the values ​​of n features of the feature map after the preliminary dimensionality reduction.

[0050] Step S350, calculating the feature means of the known multiple feature maps.

[0051] In this embodiment, the feature mean of the known feature graph is calculated and used in the quadratic dimensionality reduction formula. This step enhances the model's resistance to outliers or noise. By taking the feature mean into consideration, the model can better adapt to input data under different conditions, improving the stability and reliability of the overall system.

[0052] Step S360, perform secondary dimensionality reduction on the feature map, and the value of the jth feature of the feature map after secondary dimensionality reduction is ,in, is the feature mean of multiple feature maps, is the number of all features contained in the feature graph before dimensionality reduction, is the value of the kth feature of the feature map before dimensionality reduction, is the variance of the kth feature, The preset weight value of the j-th feature, used to indicate the importance of the j-th feature.

[0053] In this embodiment, the initial dimensionality reduction and secondary dimensionality reduction use a preset linear dimensionality reduction method and an adaptive dimensionality reduction formula to effectively reduce the dimension of the feature data while maximizing the representativeness of the data. The feature graph after dimensionality reduction not only reduces the computational complexity, but also improves the speed and efficiency of model training, allowing the system to respond quickly on larger data sets.

[0054] According to the technical solution of this embodiment, functions such as multi-layer feature map extraction, personalized weighting, intelligent synthesis and efficient dimensionality reduction are added on the original basis, which significantly improves the accuracy of behavioral feature extraction and the robustness of the model, and provides farms with more accurate and efficient management methods.

[0055] like Figure 4 As shown, another embodiment of the present invention further provides a method for regulating and controlling an aquaculture environment based on machine learning. Compared with the aforementioned embodiment, the method for regulating and controlling an aquaculture environment based on machine learning in this embodiment, before step S140, further includes:

[0056] Step S410, detecting the absolute value of each weight value in the deep neural network.

[0057] Step S420, calculating the gradient information of each weight value in the deep neural network with respect to the loss function.

[0058] Step S430, counting the number of non-zero connections and the total number of connections at the level where each weight value is located in the deep neural network.

[0059] In this embodiment, the absolute value, gradient information and number of connections of each weight value in the deep neural network are detected and counted, which provides a detailed data basis for the subsequent calculation of the contribution score of the weight value. In this way, the weight values ​​that have little impact on the model output results can be accurately located, thereby reducing unnecessary calculations without affecting the performance of the model.

[0060] Step S440, calculating the contribution score of each weight value in the deep neural network, which is used to indicate the influence of the weight value on the output result of the deep neural network, wherein Weight value Contribution score ,in, is the preset balance coefficient, For the Weight value The absolute value of and Respectively Weight value The number of non-zero connections and the total number of connections at the level, For the Weight value Gradient information about the loss function.

[0061] In this embodiment, the contribution scoring formula comprehensively considers the absolute value of the weight value, the proportion of non-zero connections and its gradient information about the loss function. This scoring mechanism can not only quantify the impact of each weight value on the model output, but also achieve fair comparison between different layers. This method helps to more finely identify which weight values ​​are really important, thereby providing a scientific basis for pruning operations.

[0062] Step S450, according to the contribution score of each weight value in the deep neural network, select a group of weight values ​​from the deep neural network and cut off the connection of the group of weight values.

[0063] In step S450, the following steps may be performed: (1) based on the preset maximum control interval of the environmental control equipment of the farm, the maximum operation time of the deep neural network is calculated; (2) based on the maximum operation time of the deep neural network, the maximum number of connections within the deep neural network is calculated; and (3) based on the maximum number of connections within the deep neural network, the number of connections to be pruned from the deep neural network is determined.

[0064] In this embodiment, the maximum control interval of the farm environment control equipment is specially taken into consideration, and the maximum operation time and the maximum number of internal connections of the deep neural network are calculated based on this. This design ensures that even after the network scale is reduced, the model can still complete the operation within the specified time, ensuring real-time performance and response speed, while maximizing the saving of computing resources.

[0065] According to the technical solution of this embodiment, by introducing a series of optimization measures for deep neural networks, including but not limited to weight value detection, gradient calculation, connection statistics, contribution score calculation and score-based weight value pruning, the model efficiency, generalization ability and computing resource utilization are significantly improved, while the maintenance cost is also reduced, injecting new vitality into the development and application of intelligent breeding management systems.

[0066] like Figure 5 As shown, another embodiment of the present invention further provides a method for regulating and controlling an aquaculture environment based on machine learning. Compared with the aforementioned embodiment, the method for regulating and controlling an aquaculture environment based on machine learning in this embodiment, before step S140, further includes:

[0067] Step S510, when training the deep neural network, calculate the behavior consistency loss of the deep neural network , where T is the preset time series length, and They represent the deep neural network in Moment and The comprehensive behavior index output at all times, express and The Euclidean distance between .

[0068] In this example, the behavioral consistency loss is designed to ensure that the comprehensive behavioral index output by the deep neural network at different time points remains consistent. By minimizing the Euclidean distance between the outputs at different time points, this method can effectively reduce the temporal fluctuations of the model prediction results and ensure the stability of animal behavior assessment, which is crucial for long-term monitoring and management of animal welfare.

[0069] Step S520, based on the behavioral consistency loss of the deep neural network , calculate the value of the loss function of the deep neural network.

[0070] In this embodiment, the introduction of behavioral consistency loss helps to improve the model's resistance to noisy data or abnormal situations. Even if inaccurate data input occurs at certain time points, the model can maintain a stable output and avoid affecting the overall evaluation results due to errors at a single time point.

[0071] Step S530, calculating the sparsity loss of the deep neural network ,in, is the set of weight values ​​in the deep neural network, Represents any weight value in the deep neural network.

[0072] In this embodiment, the sparsity loss encourages the model weight value to approach zero, thereby prompting the model to adopt a more concise and effective feature representation. This sparse representation not only reduces redundant information, but also improves the interpretability and comprehensibility of the model, allowing the model to focus more on key features, thereby enhancing its generalization ability.

[0073] Step S540, sparsity loss based on deep neural network , to correct the value of the loss function of the deep neural network.

[0074] In this embodiment, combining behavioral consistency loss and sparsity loss can significantly improve the overall performance of the model, including but not limited to prediction accuracy, computational efficiency, and resource utilization.

[0075] Step S550, adjusting the deep neural network according to the value of the loss function of the deep neural network.

[0076] According to the technical solution of this embodiment, behavioral consistency loss and sparsity loss are introduced on the original basis. By optimizing the training process of the deep neural network, the time series stability, robustness and long-term prediction ability of the model are significantly improved, while the risk of overfitting is also reduced, providing strong technical support for the development and application of intelligent breeding management systems.

[0077] like Figure 6 As shown, in one embodiment of the present invention, a farming environment control system based on machine learning is provided, comprising:

[0078] The parameter collection module 610 uses sensors installed in the farm to collect environmental parameters of the farm.

[0079] The image capture module 620 uses a camera system installed in the farm to capture images of animals in the farm.

[0080] In this embodiment, by accurately capturing and analyzing the behavioral characteristics of animals, the needs and health status of animals can be understood more accurately, and environmental conditions can be adjusted in a timely manner, thereby improving the quality of life of animals, reducing stress reactions, and promoting their healthy growth.

[0081] The feature extraction module 630 analyzes the image of the animal through a preset convolutional neural network to extract the animal's behavioral feature vector.

[0082] The index calculation module 640 inputs the environmental parameters of the farm and the behavioral feature vectors of the animals into a preset deep neural network, and outputs the comprehensive behavioral index of the animals, which represents the level of influence of the farm environment on the behavior of the animals.

[0083] In this embodiment, by using deep learning algorithms, especially convolutional neural networks and deep neural networks, valuable behavioral feature vectors can be extracted from a large amount of image data, achieving a deep understanding and quantification of animal behavior.

[0084] The equipment control module 650 controls the environmental control equipment of the farm according to the comprehensive behavior index of the animals.

[0085] In this embodiment, the environmental control device performs intelligent regulation based on the comprehensive behavior index, thereby avoiding energy waste in the traditional fixed mode, improving the efficiency of energy and other resource utilization, and reducing operating costs.

[0086] According to the technical solution of this embodiment, by introducing advanced machine learning technology, especially convolutional neural network (CNN) and deep neural network (DNN), an intelligent breeding environment control method is provided. This method can realize accurate monitoring and analysis of farm environmental parameters and animal behavior characteristics, and dynamically adjust environmental control equipment according to the analysis results.

[0087] The basic principles of the present application are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present application are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present application. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, not for limitation, and the above details do not limit the present application to being implemented by adopting the above specific details.

[0088] The block diagrams of the devices, apparatuses, equipment, and systems involved in this application are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagram. As will be appreciated by those skilled in the art, these devices, apparatuses, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including", "comprising", "having", etc. are open words, referring to "including but not limited to", and can be used interchangeably with them. The words "or" and "and" used here refer to the words "and / or" and can be used interchangeably with them, unless the context clearly indicates otherwise. The words "such as" used here refer to the phrase "such as but not limited to", and can be used interchangeably with them.

[0089] It should also be noted that in the apparatus, device and method of the present application, each component or each step can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present application.

[0090] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

[0091] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A method for controlling aquaculture environment based on machine learning, characterized in that: include: Using sensors installed in the farm to collect environmental parameters of the farm; Using a camera system installed in the farm to capture images of animals in the farm; Analyzing the image of the animal through a preset convolutional neural network to extract the behavioral feature vector of the animal; Inputting the environmental parameters of the farm and the behavioral feature vectors of the animals into a preset deep neural network, and outputting a comprehensive behavioral index of the animals, indicating the level of influence of the farm environment on the behavior of the animals; According to the comprehensive behavioral index of the animals, the environmental control equipment of the breeding farm is regulated. The method of analyzing the image of the animal by a preset convolutional neural network and extracting the behavioral feature vector of the animal includes: Extracting feature maps of multiple levels of the image of the animal through multiple convolutional layers in the convolutional neural network; According to the type of the animal, weight values ​​are set for feature maps of multiple levels of the image of the animal respectively; The feature maps of multiple levels of the image are synthesized, and the synthesized feature maps Wherein, w0 is the preset reference weight value, i0 is the preset reference magnification value, type is the type of the animal, s(type) is the reference size of the type animal, m is the number of the multiple convolutional layers, F i is a feature map generated by the i-th convolutional layer in the input-to-output direction of the convolutional neural network, The method further comprises: analyzing the image of the animal by using a preset convolutional neural network to extract the behavioral feature vector of the animal. Performing preliminary dimensionality reduction on the synthesized feature map according to a preset linear dimensionality reduction method, and obtaining values ​​of n features of the feature map after the preliminary dimensionality reduction; Calculate the feature mean of multiple known feature maps; Perform secondary dimensionality reduction on the feature map, and the value of the jth feature of the feature map after the secondary dimensionality reduction Among them, F avg is the feature mean of the multiple feature maps, q is the number of all features contained in the feature map before dimensionality reduction, F k is the value of the kth feature of the feature map before dimensionality reduction, V(F k ) is the variance of the k-th feature, w j is the preset weight value of the j-th feature, used to indicate the importance of the j-th feature.

2. The method for controlling the aquaculture environment based on machine learning according to claim 1, characterized in that: There are multiple animals, each of which has a marker implanted in its body, and a camera system installed in the farm is used to capture images of the animals in the farm, including: detecting the location of all markers implanted in the animal; Calculating the center positions of all the markers; According to the center positions of all the markers, the shooting direction and shooting focal length of the camera system are adjusted.

3. The method for controlling the aquaculture environment based on machine learning according to claim 1, characterized in that: Before inputting the environmental parameters of the farm and the behavioral feature vectors of the animals into a preset deep neural network, the method further includes: Detecting the absolute value of each weight value in the deep neural network; Calculating the gradient information of each weight value in the deep neural network with respect to the loss function; Counting the number of non-zero connections and the total number of connections at the level where each weight value in the deep neural network is located; Calculate the contribution score of each weight value in the deep neural network to indicate the influence of the weight value on the output result of the deep neural network, where the ath weight value w a Contribution score Among them, γ is the preset balance coefficient, |w a | is the ath weight value w a The absolute value of num(w a ) and num(t) are the ath weight value w a The number of non-zero connections and the total number of connections at the level, is the ath weight value w a Gradient information about the loss function; According to the contribution score of each weight value in the deep neural network, a group of weight values ​​is selected from the deep neural network and the connections of the group of weight values ​​are pruned.

4. The method for controlling the aquaculture environment based on machine learning according to claim 3, characterized in that: According to the contribution score of each weight value in the deep neural network, a group of weight values ​​are selected from the deep neural network and connections of the group of weight values ​​are cut off, including: Calculating the maximum operation time of the deep neural network according to the preset maximum control interval of the environmental control equipment of the farm; Calculating a maximum number of connections within the deep neural network according to a maximum operation time of the deep neural network; The number of connections to be pruned from the deep neural network is determined based on the maximum number of connections within the deep neural network.

5. The method for controlling the aquaculture environment based on machine learning according to claim 1, characterized in that: Before inputting the environmental parameters of the farm and the behavioral feature vectors of the animals into a preset deep neural network, the method further includes: When training the deep neural network, the behavioral consistency loss of the deep neural network is calculated Where T is the preset time series length, f(t) and f(t ′ ) represent the deep neural network at time t and t ′ The comprehensive behavior index output at each moment, d(f(t),f(t ′ )) represents f(t) and f(t ′ ) between ; Calculating a value of a loss function of the deep neural network according to a behavioral consistency loss L1 of the deep neural network; The deep neural network is adjusted according to the value of the loss function of the deep neural network.

6. The method for controlling the aquaculture environment based on machine learning according to claim 5, characterized in that: Before adjusting the deep neural network according to the value of the loss function of the deep neural network, the method further includes: Compute the sparsity loss of the deep neural network Among them, W tot is a set of weight values ​​in the deep neural network, and w represents any weight value in the deep neural network; Based on the sparsity loss L2 of the deep neural network, the value of the loss function of the deep neural network is corrected.

7. A breeding environment control system based on machine learning, characterized in that: include: A parameter collection module collects environmental parameters of the farm using sensors installed in the farm; An image capture module, using a camera system installed in the farm to capture images of animals in the farm; A feature extraction module, which analyzes the image of the animal through a preset convolutional neural network to extract the behavioral feature vector of the animal; An index calculation module, which inputs the environmental parameters of the farm and the behavioral feature vectors of the animals into a preset deep neural network, and outputs a comprehensive behavioral index of the animals, indicating the level of influence of the farm environment on the behavior of the animals; The equipment control module controls the environmental control equipment of the farm according to the comprehensive behavior index of the animals. The method of analyzing the image of the animal by a preset convolutional neural network and extracting the behavioral feature vector of the animal includes: Extracting feature maps of multiple levels of the image of the animal through multiple convolutional layers in the convolutional neural network; According to the type of the animal, weight values ​​are set for feature maps of multiple levels of the image of the animal respectively; The feature maps of multiple levels of the image are synthesized, and the synthesized feature maps Wherein, w0 is the preset reference weight value, i0 is the preset reference magnification value, type is the type of the animal, s(type) is the reference size of the type animal, m is the number of the multiple convolutional layers, F i is a feature map generated by the i-th convolutional layer in the input-to-output direction of the convolutional neural network, The method further comprises: analyzing the image of the animal by using a preset convolutional neural network to extract the behavioral feature vector of the animal. Performing preliminary dimensionality reduction on the synthesized feature map according to a preset linear dimensionality reduction method, and obtaining values ​​of n features of the feature map after the preliminary dimensionality reduction; Calculate the feature mean of multiple known feature maps; Perform secondary dimensionality reduction on the feature map, and the value of the jth feature of the feature map after the secondary dimensionality reduction Among them, F avg is the feature mean of the multiple feature maps, q is the number of all features contained in the feature map before dimensionality reduction, F k is the value of the kth feature of the feature map before dimensionality reduction, V(F k ) is the variance of the k-th feature, w j is the preset weight value of the j-th feature, used to indicate the importance of the j-th feature.

Images