Fry culture dynamic monitoring and feeding optimization method based on deep learning

Abstract

The invention discloses a fry culture dynamic monitoring and feeding optimization method based on deep learning. The method comprises the following steps: S1, installing a water quality sensor and a camera to collect water quality data and fry image data; s2, performing feature extraction on the fry image data by using an improved convolutional neural network; s3, analyzing a growth trend and an activity mode of the fry by using a long-short-term memory network to obtain a comprehensive score; s4, obtaining a feeding strategy based on a multi-target adaptive PPO algorithm, and forming a personalized feeding plan; s5, performing detection and early warning on abnormal data through an anomaly detection algorithm, and adjusting a breeding environment; and S6, continuously optimizing the fry culture dynamic monitoring model. According to the method, an efficient and scientific optimization scheme can be provided in dynamic monitoring and feeding of fry culture, and remarkable technical value and economic benefits are brought to practical application.

Description

Technical Field

[0001] The present invention relates to the technical field of fry breeding, and particularly to a dynamic monitoring and feeding optimization method for fry breeding based on deep learning. Background Art

[0002] With the rapid development of the aquaculture industry, traditional breeding methods have gradually become difficult to meet the requirements of high efficiency, intelligence, and precision. Especially in the process of fry breeding, how to accurately monitor the health status of fry and how to optimize the feeding strategy have become key factors affecting breeding efficiency and water quality management. Most traditional fry breeding management methods rely on manual observation and empirical judgment. Although they can play a certain role in some cases, there are many drawbacks. For example, the efficiency of manual monitoring is low, the feeding amount is inaccurate, and the breeding environment cannot be adjusted in real time, making it difficult to meet the requirements of refined management in modern breeding.

[0003] In the traditional process of fry breeding, water quality monitoring and feeding control are usually manually completed by breeders. The fluctuations of water quality parameters such as water temperature, dissolved oxygen, and pH value, as well as the growth status of fry, are often not effectively monitored and adjusted in a timely manner. For water quality monitoring, traditional methods rely on manual sampling and testing. However, this method is not only time-consuming and laborious, but also unable to obtain water quality changes in real time, easily overlooking potential water quality problems. The health status and growth progress of fry also usually rely on manual observation and recording, which makes it difficult for managers to comprehensively and real-time grasp the growth of fry, thereby affecting the accuracy and timeliness of feeding decisions. In addition, traditional feeding methods cannot be finely adjusted according to the actual needs of fry, resulting in overfeeding or underfeeding, which not only wastes feed resources, but also may cause water pollution, thus affecting the breeding environment and the growth of fry.

[0004] In recent years, with the development of deep learning and Internet of Things technologies, more and more intelligent management means have been introduced into the aquaculture field, but the existing intelligent management systems still face some technical bottlenecks.

[0005] First, most existing systems focus on water quality monitoring and fry behavior detection, but lack an effective comprehensive analysis and decision-making mechanism. Second, although there are some deep learning-based systems that attempt to analyze fry behavior and predict health status, most systems rely on simple image recognition techniques and ignore the complex dynamic interactions between fry and the environment. This method often fails to accurately capture complex factors in the aquaculture environment, such as water temperature changes and fry density fluctuations, resulting in poor performance of the model in practical applications. Finally, existing feeding optimization systems are mostly static or rule-based systems and cannot be flexibly adjusted according to the real-time growth status of fry and environmental changes, leading to low feeding efficiency and serious resource waste, which affects the aquaculture benefits. Therefore, there is an urgent need for a new type of fry aquaculture dynamic monitoring and feeding optimization method based on deep learning and the Internet of Things, which can obtain fry health status and water quality data in real time, comprehensively analyze this information and make intelligent decisions, so as to achieve precise feeding and efficient aquaculture management. This method should have an adaptive ability and be able to dynamically adjust according to changes in the aquaculture environment and fry to achieve the best aquaculture effect. Summary of the Invention

[0006] An object of the present invention is to propose a fry aquaculture dynamic monitoring and feeding optimization method based on deep learning. The present invention can provide an efficient and scientific optimization scheme in fry aquaculture dynamic monitoring and feeding, bringing significant technical value and economic benefits to practical applications.

[0007] A fry aquaculture dynamic monitoring and feeding optimization method based on deep learning according to an embodiment of the present invention includes the following steps:

[0008] S1. Install water quality sensors and cameras to collect water quality data and fry image data in the pool in real time, and transmit the collected water quality data and fry image data to a data processing center for storage;

[0009] S2. Use an improved convolutional neural network to extract features from fry image data. The improved convolutional neural network adopts transfer learning technology and is fine-tuned based on the pre-trained neural network ResNet. Extract fry image data features through the trained convolutional neural network and use a classifier to classify and identify the body shape, body posture changes and group behavior of fry;

[0010] S3. Use a long short-term memory network to analyze the growth trend and activity pattern of fry, combine water quality data and environmental changes, establish a time series model to predict the growth status of fry, and obtain a comprehensive score;

[0011] S4. Construct a multi-objective adaptive PPO algorithm, and train and optimize the multi-objective adaptive PPO algorithm through an experience replay mechanism. Automatically optimize the feeding strategy through the trained multi-objective adaptive PPO network to form a personalized feeding plan;

[0012] S5. Use a monitoring system to real-time track the fry density, activity range, and water quality changes in the pool, and combine an anomaly detection algorithm to detect and warn of abnormal data, and adjust the aquaculture environment;

[0013] S6. Continuously optimize the fry aquaculture dynamic monitoring model to improve the feeding efficiency and fry health management level under different aquaculture environments and water quality conditions.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Install multiple water quality sensors in the pool. The water quality sensors include a temperature sensor, a dissolved oxygen sensor, a pH value sensor, and a turbidity sensor. The temperature sensor is used to collect water temperature data, the dissolved oxygen sensor is used to collect the dissolved oxygen concentration data in the water, the pH value sensor is used to collect the acidity and alkalinity data of the water body, and the turbidity sensor is used to collect the turbidity data of the water body;

[0016] S12. Arrange a camera in the pool. The camera rotates to capture image data of the fry in the pool. The image data includes the body shape, activity state, and behavior characteristics of the fry;

[0017] S13. Through the Internet of Things technology, transmit the water quality data collected by the water quality sensors and the image data collected by the camera to the data processing center in real time through a wireless network.

[0018] Optionally, the S2 includes the following steps:

[0019] S21. Use an improved convolutional neural network to extract features from the fry image data. The filters in the convolutional layer are used to extract local features in the fry image. The filter size is 5×5. The convolutional operation identifies features by performing a sliding window process on the fry image. The convolutional result is the feature map F;

[0020] S22. In the feature map F, the pooling layer performs a downsampling operation on the feature map. The maximum pooling method is used to reduce the spatial dimension of the image. The pooling window size is 2×2. The pooling result is the downsampled feature map F p ;

[0021] S23. To further improve the accuracy of fry image feature extraction, the convolutional neural network is combined with the attention mechanism for improvement. The attention mechanism optimizes the model's attention to key features by weighting important regions in the feature map, enhancing the model's ability to identify fry in complex backgrounds. The attention mechanism is a spatial attention mechanism, and its calculation formula is:

[0022]

[0023] where A is the attention weight matrix, W 1 and W 2 are parameter matrices, F p is the pooled feature map, and softmax is the normalization function;

[0024] S24. Through the combination of multiple convolutional layers and pooling layers, the local features of fry in the image are extracted layer by layer. Finally, the local features of fry are mapped to the class labels of fry through the fully connected layer. To further improve the learning ability of the visual feature extraction model, depthwise separable convolution is used instead of traditional convolution. The formula for depthwise separable convolution is;

[0025] y = W 1 *x + W 2 *x;

[0026] where W 1 is the depthwise convolution kernel, W 2 is the pointwise convolution kernel, x is the input feature map, and * represents the convolution operation;

[0027] S25. The convolutional neural network model is initialized through the pre-trained neural network ResNet. The transfer learning technique is adopted, and the weights of ResNet are fine-tuned to adapt to the characteristics of fry image data. The process of transfer learning includes fine-tuning the ResNet model on the target dataset and fixing the parameters of some layers, only updating the parameters of the high-level network. The loss function of the fine-tuning process is:

[0028]

[0029] where y i is the target output label, is the predicted value, θ j is the network parameter, λ is the regularization coefficient, Loss(.) is the cross-entropy loss function, ∥.∥ 2 is the Frobenius norm, and n is the total number of samples in the dataset;

[0030] S26. Feature fusion, noise reduction, and data standardization are performed on the extracted fry image features. The formula for the fusion operation is:

[0031]

[0032] Among them, T f is the fused feature set, F i is the feature map of the i-th layer, w i is the importance weight of each layer of features, m is the number of feature layers, and principal component analysis is used to reduce the dimension of the fused features;

[0033] S27. Input the fused feature set T f into the classifier. An ensemble learning method is adopted to improve the accuracy by training multiple classifiers and fusing the output results of multiple classifiers:

[0034]

[0035] Among them, C final is the final classification result, w i is the weight of the i-th classifier, p i,c is the probability that the i-th classifier predicts the category c, and k is the number of classifiers.

[0036] Optionally, S3 includes the following steps:

[0037] S31. Combine the water quality data D w , including temperature, dissolved oxygen, pH value, turbidity, with the classification result C final of the fry image data to construct a multi-modal input data set X, X = {D w , C final}, which is used to analyze the growth trend and activity pattern of fry;

[0038] S32. Use a long short-term memory network to perform time series analysis on the multi-modal input data set X, and predict the health status and activity trend of fry in the future time period by inputting the historical data sequence;

[0039] S33. Based on the prediction results of the long short-term memory network, combined with the time series characteristics of the water quality data, use a support vector machine to further predict the growth trend of fry and obtain the predicted value Y pred , Y pred represents the predicted growth state of fry;

[0040] S34. Use the weighted average method to comprehensively obtain the comprehensive score S of the prediction results of different features:

[0041] S = w 1 ·Y pred + w 2 ·D w ;

[0042] Among them, w 1 and w2 is the weighting coefficient, Y pred is the predicted growth state of fry, D w is the water quality data.

[0043] Optionally, the S4 includes the following steps:

[0044] S41. Combine the water quality data D w and the classification result C final , and use the multi-objective adaptive PPO algorithm to obtain the feeding strategy, automatically adjust the feeding strategy to meet the needs of fry in different breeding environments, and update the strategy according to the fry health status and environmental feedback after each feeding;

[0045] S42. Define the state space S:

[0046] S = {S w , S f};

[0047] Among them, S w represents the water quality state, including temperature, dissolved oxygen, pH value, turbidity, and S f represents the health status and behavioral characteristics of fry, including body size change, activity level, feeding behavior;

[0048] S43. Define the action space A, A = {a 1 , a 2 ,..., a n}, and a 1 to a n are different combinations of feeding amounts, feeding frequencies, and feeding modes respectively;

[0049] S44. Optimize the strategy through the objective function of the multi-objective adaptive PPO algorithm, and the objective function is:

[0050]

[0051] Among them, r t (θ) is the probability ratio of the current strategy to the old strategy, is the advantage function of the fry health status, is the advantage function of resource utilization efficiency, is the advantage function of environmental protection, clip(.) represents the clipping operation, λ 1 , λ 2 are adjustment coefficients, ∈ is the threshold of clipping, and E t represents the expected value at time step t;

[0052] S45. Optimize the feeding strategy through the policy gradient algorithm based on the real-time monitored water quality data, fry health status, behavior characteristics, and historical feeding data, and dynamically adjust the weight of the reward in combination with the adaptive adjustment of the reward function. The reward function is as follows:

[0053] R t = w 1 ·R health + w 2 ·R efficiency + w 3 ·R environment ;

[0054] Among them, R health is the reward based on the fry health status, R efficiency is the reward based on the resource utilization efficiency, R environment is the reward based on environmental protection, and w 1 , w 2 , w 3 are the weight coefficients;

[0055] S46. Through the adaptive adjustment mechanism of the reward function, dynamically update the weight coefficients w 1 , w 2 , w 3 during the training process;

[0056] S47. Adopt the experience replay mechanism to train and optimize the multi-objective adaptive PPO algorithm:

[0057] Q′ t = τQ t + (1 - τ)Q′ t-1 ;

[0058] Among them, τ is the soft update factor, Q t is the Q value output by the current network, and Q′ t-1 is the Q value output by the target network;

[0059] S48. Through training, obtain the multi-objective adaptive PPO network, map the learned feeding strategy to the actual feeding control instruction, and the calculation formula for the feeding amount is:

[0060] a t = argmax a∈A Q(S t , a);

[0061] Among them, a t is the optimal feeding amount selected in the current state S t , A is the action space, and Q(S t , a) is the Q value of each action;

[0062] S49. The multi-objective adaptive PPO network adjusts the feeding mode according to the dynamic changes of the fry population. If the fry density is too high, the model reduces the feeding amount or changes the feeding frequency by analyzing the water quality status and the fry health status:

[0063]

[0064] where Density t is the fry density at the current time point, w 4 is the weight coefficient for adjusting the density, and exp(.) is the natural exponential function.

[0065] Optionally, S5 includes the following steps:

[0066] S51. Use an anomaly detection algorithm to analyze the water quality and the fry health status. The anomaly detection algorithm calculates the anomaly metric value E by comparing the current water quality data and fry behavior with the historical data model. The anomaly detection algorithm combines a deep autoencoder and an isolation forest algorithm. The deep autoencoder is used to learn the latent representation of the data and generate a reconstruction error to detect anomalies, and the isolation forest is used to further identify and filter extreme anomaly data;

[0067] The training process of the deep autoencoder includes inputting the water quality data and fry health status features at the current moment, mapping the water quality data and fry health status features to the latent space through the encoder, and then restoring the data through the decoder, and calculating the reconstruction error as the anomaly metric. The calculation formula for the reconstruction error E is:

[0068]

[0069] where X i is the original data of the i-th health status feature, is the reconstructed value of the i-th feature, n is the total number of feature data, and E AE is the reconstruction error, representing the difference between the current data and the reconstructed data. A larger error value indicates data anomaly;

[0070] S53. The isolation forest algorithm determines the anomaly points by constructing multiple decision trees and judging by the separation degree of the sample points. The points with low separation degree are anomaly points. The anomaly metric formula of the isolation forest is:

[0071]

[0072] where h(x) is the isolation depth of the sample point x, c(n) is the adjustment factor for the dataset size, n is the total number of samples, and E IF is the anomaly degree calculated by the isolation forest;

[0073] S54. Combine the anomaly metric values of the deep autoencoder and the Isolation Forest to generate a weighted anomaly metric value E combined :

[0074] E combined = w 1 ·E AE + w 2 ·E IF ;

[0075] where w 1 and w 2 are weighting coefficients;

[0076] S55. If the comprehensive anomaly metric value E combined is greater than the set threshold E threshold , then trigger an anomaly alarm and adjust the aquaculture environment according to the anomaly type. The alarm determination formula is:

[0077]

[0078] where Alert is the alarm signal, 1 indicates triggering the alarm, and 0 indicates not triggering the alarm.

[0079] The beneficial effects of the present invention are:

[0080] (1) The present invention can monitor and analyze the health status of fry, water quality parameters, and dynamic changes in the aquaculture environment in real time, and thus automatically adjust the feeding strategy according to these real-time data. This comprehensive solution based on deep learning and Internet of Things technology, through in-depth analysis of water quality data and the health status of fry, can not only accurately predict the health status of fry, but also optimize the feeding amount and feeding frequency according to environmental changes, avoiding the problems of overfeeding or underfeeding existing in traditional aquaculture processes, thereby improving the feeding efficiency, reducing feed waste, and lowering the aquaculture cost.

[0081] (2) The deep learning model introduced in the present invention, especially using convolutional neural network, long short-term memory network, and deep reinforcement learning technologies, can dynamically adjust the feeding strategy according to changes in the aquaculture environment, and has stronger adaptability and prediction ability. This model not only considers a single water quality factor, but also comprehensively considers the health status, behavior patterns of fry, and water quality data, providing a more comprehensive and accurate decision-making basis for feeding optimization. By integrating multiple algorithms, the present invention can provide personalized feeding schemes under different environmental conditions, ensuring that fry receive appropriate nutrient supply in the best growth environment, thereby promoting the healthy growth of fry and enhancing the aquaculture benefits.

[0082] (3) Through model fine-tuning, the present invention ensures the continuous optimization ability of the system. During the actual aquaculture process, the deep learning model can adaptively adjust according to the continuously collected new data, enabling the feeding strategy to be continuously optimized as the fry grow and the environment changes. This adaptive learning mechanism significantly improves the long-term stability and reliability of the system, and can effectively cope with various complex aquaculture environments and unforeseen changes. Compared with the prior art, the present invention not only improves the accuracy of the system, but also enhances the adaptability of the system in a dynamic environment, enabling it to continuously play a role in a changing aquaculture environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0084] Figure 1 is a flowchart of a method for dynamic monitoring and feeding optimization of fry aquaculture based on deep learning proposed by the present invention;

[0085] Figure 2 is a flowchart of a multi-objective adaptive PPO algorithm in a method for dynamic monitoring and feeding optimization of fry aquaculture based on deep learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0086] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0087] Refer to Figure 1 - Figure 2 , a method for dynamic monitoring and feeding optimization of fry aquaculture based on deep learning, includes the following steps:

[0088] S1. Install water quality sensors and cameras to collect water quality data and fry image data in the pool in real time, and transmit the collected water quality data and fry image data to the data processing center for storage;

[0089] S2. Use an improved convolutional neural network to extract features from the fry image data. The improved convolutional neural network adopts transfer learning technology and is fine-tuned based on the pre-trained neural network ResNet. Extract fry image data features through the trained convolutional neural network and use a classifier to classify and identify the body shape, body posture changes and group behaviors of the fry;

[0090] S3. Use a long short-term memory network to analyze the growth trend and activity pattern of the fry, combine the water quality data and environmental changes, establish a time series model to predict the growth state of the fry, and obtain a comprehensive score;

[0091] S4. Construct a multi-objective adaptive PPO algorithm, and train and optimize the multi-objective adaptive PPO algorithm through an experience replay mechanism. Automatically optimize the feeding strategy through the trained multi-objective adaptive PPO network to form a personalized feeding plan;

[0092] S5. Use a monitoring system to track the fry density, activity range, and water quality changes in the pool in real time, and combine an anomaly detection algorithm to detect and warn of abnormal data and adjust the breeding environment;

[0093] S6. Continuously optimize the fry breeding dynamic monitoring model to improve the feeding efficiency and fry health management level under different breeding environments and water quality conditions.

[0094] In this embodiment, S1 includes the following steps:

[0095] S11. Install multiple water quality sensors in the pool. The water quality sensors include a temperature sensor, a dissolved oxygen sensor, a pH value sensor, and a turbidity sensor. The temperature sensor is used to collect water temperature data, the dissolved oxygen sensor is used to collect the dissolved oxygen concentration data in the water, the pH value sensor is used to collect the acidity and alkalinity data of the water body, and the turbidity sensor is used to collect the turbidity data of the water body;

[0096] S12. Arrange a camera in the pool. The camera rotates to capture image data of the fry in the pool. The image data includes the body shape, activity state, and behavioral characteristics of the fry;

[0097] S13. Through the Internet of Things technology, transmit the water quality data collected by the water quality sensors and the image data collected by the camera to the data processing center in real time through a wireless network.

[0098] In this embodiment, S2 includes the following steps:

[0099] S21. Use an improved convolutional neural network to extract features from the fry image data. The filters in the convolutional layer are used to extract local features in the fry image. The filter size is 5×5. The convolutional operation identifies features by performing a sliding window process on the fry image. The convolutional result is the feature map F;

[0100] S22. In the feature map F, the pooling layer performs a downsampling operation on the feature map. The maximum pooling method is used to reduce the spatial dimension of the image. The pooling window size is 2×2. The pooling result is the downsampled feature map F p ;

[0101] S23. To further improve the accuracy of fry image feature extraction, a combination of a convolutional neural network and an attention mechanism is used for improvement. The attention mechanism optimizes the model's attention to key features by weighting important regions in the feature map, enhancing the model's ability to identify fry in complex backgrounds. The attention mechanism is a spatial attention mechanism, and its calculation formula is:

[0102]

[0103] where A is the attention weight matrix, W 1 and W 2 are parameter matrices, F p is the pooled feature map, and softmax is the normalization function;

[0104] S24. Through the combination of multiple convolutional layers and pooling layers, the local features of fry in the image are extracted layer by layer. Finally, the local features of fry are mapped to the class labels of fry through a fully connected layer. To further improve the learning ability of the visual feature extraction model, depthwise separable convolution is used instead of traditional convolution. The formula for depthwise separable convolution is;

[0105] y = W 1 *x + W 2 *x;

[0106] where W 1 is the depthwise convolution kernel, W 2 is the pointwise convolution kernel, x is the input feature map, and * represents the convolution operation;

[0107] S25. The convolutional neural network model is initialized through the pre-trained neural network ResNet. The transfer learning technique is adopted. By fine-tuning the weights of ResNet, it adapts to the characteristics of fry image data. The process of transfer learning includes fine-tuning the ResNet model on the target dataset and fixing the parameters of some layers, only updating the parameters of the high-level network. The loss function in the fine-tuning process is:

[0108]

[0109] where y i is the target output label, is the predicted value, θ j is the network parameter, λ is the regularization coefficient, Loss(.) is the cross-entropy loss function, ∥.∥ 2 is the Frobenius norm, and n is the total number of samples in the dataset;

[0110] S26. Feature fusion, noise reduction, and data standardization are performed on the extracted fry image features. The formula for the fusion operation is:

[0111]

[0112] Among them, T f is the fused feature set, F i is the feature map of the i-th layer, w i is the importance weight of each layer of features, m is the number of feature layers, and principal component analysis is used to reduce the dimension of the fused features;

[0113] S27. Input the fused feature set T f into the classifier. An ensemble learning method is adopted to improve the accuracy by training multiple classifiers and fusing the output results of multiple classifiers:

[0114]

[0115] Among them, C final is the final classification result, w i is the weight of the i-th classifier, p i,c is the probability that the i-th classifier predicts the category c, and k is the number of classifiers.

[0116] In this embodiment, S3 includes the following steps:

[0117] S31. Combine the water quality data D w , including temperature, dissolved oxygen, pH value, turbidity, with the classification result C final of the fry image data to construct a multi-modal input data set X, X = {D w , C final}, for analyzing the growth trend and activity pattern of fry;

[0118] S32. Perform time series analysis on the multi-modal input data set X using a long short-term memory network. By inputting the historical data sequence, predict the health status and activity trend of fry in the future time period;

[0119] S33. Based on the prediction results of the long short-term memory network, combined with the time series characteristics of the water quality data, use a support vector machine to further predict the growth trend of fry and obtain the predicted value Y pred , Y pred represents the predicted growth state of fry;

[0120] S34. Use the weighted average method to comprehensively obtain the comprehensive score S of the prediction results of different features:

[0121] S = w 1 ·Y pred + w 2 ·D w ;

[0122] Among them, w 1 and w2 is the weighting coefficient, Y pred is the predicted growth state of fry, D w is the water quality data.

[0123] In this embodiment, S4 includes the following steps:

[0124] S41. Combine the water quality data D w and the classification result C final , and use the multi-objective adaptive PPO algorithm to obtain the feeding strategy, automatically adjust the feeding strategy to meet the needs of fry in different breeding environments, and update the strategy according to the fry health status and environmental feedback after each feeding;

[0125] S42. Define the state space S:

[0126] S = {S w , S f};

[0127] Among them, S w represents the water quality state, including temperature, dissolved oxygen, pH value, turbidity, and S f represents the health status and behavioral characteristics of fry, including body size change, activity level, feeding behavior;

[0128] S43. Define the action space A, A = {a 1 , a 2 ,..., a n}, where a 1 to a n are different combinations of feeding amounts, feeding frequencies, and feeding modes respectively;

[0129] S44. Optimize the strategy through the objective function of the multi-objective adaptive PPO algorithm. The objective function is:

[0130]

[0131] Among them, r t (θ) is the probability ratio of the current strategy to the old strategy, is the advantage function of the fry health status, is the advantage function of resource use efficiency, is the advantage function of environmental protection, clip(.) represents the clipping operation, λ 1 , λ 2 are adjustment coefficients, ∈ is the threshold of clipping, and E t represents the expected value at time step t;

[0132] S45. Based on the real-time monitored water quality data, fry health status, behavioral characteristics, and historical feeding data, optimize the feeding strategy through the policy gradient algorithm, and dynamically adjust the reward weights in combination with the adaptive adjustment of the reward function. The reward function is as follows:

[0133] R t = w 1 ·R health + w 2 ·R efficiency + w 3 ·R environment ;

[0134] where R health is the reward based on the fry health status, R efficiency is the reward based on the resource utilization efficiency, R environment is the reward based on environmental protection, and w 1 , w 2 , w 3 are the weight coefficients;

[0135] S46. Through the adaptive adjustment mechanism of the reward function, dynamically update the weight coefficients w 1 , w 2 , w 3 during the training process;

[0136] S47. Adopt the experience replay mechanism for the training and optimization of the multi-objective adaptive PPO algorithm:

[0137] Q′ t = τQ t + (1 - τ)Q′ t-1 ;

[0138] where τ is the soft update factor, Q t is the Q value output by the current network, and Q′ t-1 is the Q value output by the target network;

[0139] S48. Through training, obtain the multi-objective adaptive PPO network, map the learned feeding strategy to the actual feeding control instruction, and the calculation formula for the feeding amount is:

[0140] a t = argmax a∈A Q(S t , a);

[0141] where a t is the optimal feeding amount selected in the current state S t , A is the action space, and Q(S t , a) is the Q value of each action;

[0142] S49. The multi-objective adaptive PPO network adjusts the feeding mode according to the dynamic changes of the fry population. If the fry density is too high, the model reduces the feeding amount or changes the feeding frequency by analyzing the water quality status and the health status of the fry:

[0143]

[0144] where Density t is the fry density at the current time point, w 4 is the weight coefficient for adjusting the density, and exp(.) is the natural exponential function.

[0145] In this embodiment, S5 includes the following steps:

[0146] S51. An anomaly detection algorithm is used to analyze the water quality and the health status of the fry. The anomaly detection algorithm calculates the anomaly metric value E by comparing the current water quality data and fry behavior with the historical data model. The anomaly detection algorithm combines a deep autoencoder and an isolation forest algorithm. The deep autoencoder is used to learn the latent representation of the data and generate a reconstruction error to detect anomalies, and the isolation forest is used to further identify and filter extreme anomaly data;

[0147] S52. The training process of the deep autoencoder includes inputting the water quality data and the fry health status features at the current moment, mapping the water quality data and the fry health status features to the latent space through the encoder, and then restoring the data through the decoder, and calculating the reconstruction error as the anomaly metric. The calculation formula of the reconstruction error E is:

[0148]

[0149] where X i is the original data of the i-th health status feature, is the reconstructed value of the i-th feature, n is the total number of feature data, and E AE is the reconstruction error, representing the difference between the current data and the reconstructed data. A larger error value indicates data anomaly;

[0150] S53. The isolation forest algorithm determines the anomaly points by constructing multiple decision trees and judging by the separation degree of the sample points. The points with low separation degree are anomaly points. The anomaly metric formula of the isolation forest is:

[0151]

[0152] where h(x) is the isolation depth of the sample point x, c(n) is the adjustment factor of the dataset size, n is the total number of samples, and E IF is the anomaly degree calculated by the isolation forest;

[0153] S54. Combine the anomaly measurement values of the deep autoencoder and the Isolation Forest to generate a weighted anomaly measurement value E combined :

[0154] E combined = w 1 ·E AE + w 2 ·E IF ;

[0155] Wherein, w 1 and w 2 are weighting coefficients;

[0156] S55. If the comprehensive anomaly measurement value E combined is greater than the set threshold E threshold , then trigger an anomaly alarm and adjust the aquaculture environment according to the anomaly type. The alarm determination formula is:

[0157]

[0158] Wherein, Alert is the alarm signal, 1 indicates triggering the alarm, and 0 indicates not triggering the alarm.

[0159] Example:

[0160] In a certain aquaculture farm, which is located in a coastal area of China and mainly farms seawater fry, including sea bass and grouper. The traditional aquaculture method faces problems such as difficult control of feeding amount, insufficient water quality monitoring, and lagging assessment of the health status of fry, resulting in resource waste and unbalanced growth of fry during the aquaculture process.

[0161] This aquaculture farm originally adopted traditional manual feeding methods and water quality monitoring methods. Aquaculture workers needed to check water quality parameters every few hours and manually adjust the feeding amount. However, due to the large fluctuations in water quality and the health status of fry, the traditional methods were often difficult to respond to environmental changes in a timely manner, resulting in too much or too little feeding amount, serious water pollution, and even the situation that some fry grew slowly or got sick. Especially when the aquaculture density was relatively high, the water quality changed more violently, leading to the deterioration of the aquaculture environment and the stagnation of fry growth.

[0162] To solve these problems, the breeding farm introduced the dynamic monitoring and feeding optimization method based on deep learning of the present invention. First, multiple water quality sensors and cameras were installed in the site to collect water quality data and fry image data in real time. The data of all sensors and cameras were transmitted to the data processing center through Internet of Things technology for real-time analysis and decision-making by the deep learning model. According to the real-time data, the deep learning model used convolutional neural network to analyze the image data, identify the health status, behavior patterns and density changes of the fry. At the same time, combined with the water quality data, the model predicted the growth trend of the fry through long short-term memory network and optimized the feeding strategy by using deep reinforcement learning to accurately adjust the feeding amount and frequency.

[0163] During the implementation process, the system first monitored the water quality data and the health status of the fry. For example, at the beginning of the experiment, the water temperature data collected by the sensor in real time was 24.5°C, the dissolved oxygen concentration was 6.5mg / L, the pH value was 7.8, and the turbidity was 15NTU. Based on these data, the deep learning model would predict the health status of the fry and calculate the optimal feeding amount according to the prediction result. For example, the feeding amount calculated by the model in the initial stage was 10kg of feed per hour, which was gradually adjusted based on the density of the fry and the water quality situation.

[0164] With the continuous accumulation of data during the breeding process, the system made the feeding strategy gradually more accurate through incremental learning and model fine-tuning. After several weeks of training and optimization, the system could dynamically adjust the feeding amount according to the growth and environmental changes of the fry. For example, in the test after one month, the system predicted that the improvement of the health status of the fry required an increase in the feeding amount, so it automatically adjusted the feeding amount per hour to 12kg. At the same time, it was monitored that the water temperature rose to 26.2°C, the dissolved oxygen concentration dropped to 5.8mg / L, and the pH value remained around 7.9. The system adjusted the feeding strategy in real time through the deep reinforcement learning algorithm, adjusted the feeding amount to 12kg, and increased the oxygen supply in the pool.

[0165] In addition, the system was also able to automatically adjust the feeding strategy according to the density of the fry. When the fry density was high, the system reduced the feeding amount through comprehensive evaluation of the water quality and health status, avoiding water pollution caused by excessive feeding. The system automatically adjusted the feeding amount and started to increase the workload of the oxygenation equipment in the pool to maintain the stability of the water quality.

[0166] By comparing with the traditional method, the specific data is as shown in Table 1 below:

[0167] Table 1 Comparison of water quality and fry growth data between the present invention and the traditional method during the breeding process

[0168] Project Traditional method Method of the present invention Average water temperature (°C) 25.2 25.6 Dissolved oxygen concentration (mg / L) 5.2 6.2 pH value 7.6 7.8 Turbidity (NTU) 18.5 12.1 Feeding rate (kg / h) 15 11 Average growth rate of fry (cm / day) 0.45 0.54

[0169] Throughout the embodiments, the implementer uses the method of the present invention to not only solve the problems of untimely water quality monitoring and difficult control of feeding amount in the traditional fry breeding process, but also improve the breeding efficiency and achieve precise optimization of fry growth and resource utilization.

[0170] By combining deep learning and Internet of Things technologies, the present invention realizes real-time monitoring and analysis of water quality data and fry health status. Through convolutional neural networks and long short-term memory networks, the model can accurately identify the health status of fry and predict their growth trends based on real-time data. Compared with the traditional method that relies on manual observation and fixed feeding amounts, the present invention optimizes the feeding strategy through deep reinforcement learning, automatically adjusting the feeding amount and frequency, thereby effectively reducing overfeeding and feed waste while ensuring the healthy growth of fry.

[0171] In the process of optimizing the feeding amount, the present invention introduces adaptive learning and model fine-tuning technologies, enabling the system to continuously adjust the feeding strategy according to the breeding environment and the growth dynamics of fry. Compared with the traditional static feeding method, the system of the present invention fully considers various factors such as water quality changes and fry density during feeding, greatly improving the resource utilization efficiency and avoiding common problems such as water pollution and inaccurate feeding in the breeding process.

[0172] Through intelligent and automated breeding management means, the present invention solves multiple pain points in traditional breeding methods, significantly improves breeding efficiency, reduces labor input and resource waste, and realizes the intelligence and refinement of breeding management through precise monitoring and adjustment of the breeding environment and fry health status, effectively promoting the sustainable development of the modern aquaculture industry.

[0173] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. A method for dynamic monitoring and feeding optimization of fry farming based on deep learning, characterized in that: The steps include: S1, install water quality sensor and camera, collect water quality data and image data of fry in the pool in real time, and transmit the collected water quality data and image data of fry to data processing center for storage; S2, using an improved convolutional neural network to extract features from the fry image data, wherein the improved convolutional neural network adopts transfer learning technology and is fine-tuned based on the pre-trained neural network ResNet, extracts features from the fry image data through the trained convolutional neural network, and uses a classifier to classify and identify the body shape, body shape changes, and group behavior of the fry; S3. Use long short-term memory networks to analyze the growth trends and activity patterns of fry, combine water quality data and environmental changes, establish a time series model to predict the growth status of fry, and obtain a comprehensive score; S4. Construct a multi-objective adaptive PPO algorithm, and train and optimize the multi-objective adaptive PPO algorithm through the experience replay mechanism. The multi-objective adaptive PPO network after training automatically optimizes the feeding strategy to form a personalized feeding plan. S5. Use the monitoring system to track the density, activity range and water quality changes of fry in the pond in real time, and use the abnormal detection algorithm to detect and warn abnormal data and adjust the breeding environment; S6. Continuously optimize the dynamic monitoring model of fry farming to improve feeding efficiency and fry health management level under different farming environments and water quality conditions.

2. According to a deep learning-based method for dynamic monitoring and feeding optimization of fry culture according to claim 1, it is characterized in that: The S1 comprises the following steps: S11, installing multiple water quality sensors in the pool, the water quality sensors including a temperature sensor, a dissolved oxygen sensor, a pH sensor, and a turbidity sensor, the temperature sensor is used to collect water temperature data, the dissolved oxygen sensor is used to collect dissolved oxygen concentration data in water, the pH sensor is used to collect pH data of the water body, and the turbidity sensor is used to collect turbidity data of the water body; S12, arranging a camera in the pool, wherein the camera takes image data of the fry in the pool in a rotating manner, wherein the image data includes the body shape, activity state, and behavioral characteristics of the fry; S13. Using Internet of Things technology, the water quality data collected by the water quality sensor and the image data collected by the camera are transmitted in real time to a data processing center via a wireless network.

3. The method for dynamic monitoring and feeding optimization of fry culture based on deep learning according to claim 1, characterized in that: The S2 comprises the following steps: S21, using an improved convolutional neural network to perform feature extraction on the fry image data, the filter of the convolution layer is used to extract local features in the fry image, the filter size is 5×5, the convolution operation identifies features by performing sliding window processing on the fry image, and the convolution result is a feature map F; S22. In the feature map F, the pooling layer performs a downsampling operation on the feature map, using the maximum pooling method to reduce the spatial dimension of the image. The pooling window size is 2×2, and the pooling result is the feature map F after dimensionality reduction. p ; S23. To further improve the accuracy of fry image feature extraction, a convolutional neural network and an attention mechanism are combined for improvement. The attention mechanism optimizes the model's attention to key features by weighting important areas in the feature map, thereby enhancing the model's ability to recognize fry in a complex background. The attention mechanism is a spatial attention mechanism, and the calculation formula is: Among them, A is the attention weight matrix, W1 and W2 are parameter matrices, and F p is the feature map after pooling, and softmax is the normalization function; S24, through the combination of multiple convolutional layers and pooling layers, the local features of the fry in the image are extracted layer by layer, and finally the local features of the fry are mapped to the category label of the fry through the fully connected layer. In order to further improve the learning ability of the visual feature extraction model, the deep separable convolution is used instead of the traditional convolution. The formula of the deep separable convolution is: y=W1*x+W2*x; Among them, W1 is the depth convolution kernel, W2 is the point convolution kernel, x is the input feature map, and * represents the convolution operation; S25, the convolutional neural network model is initialized by the pre-trained neural network ResNet, and the transfer learning technology is used to fine-tune the weights of ResNet. The transfer learning process includes fine-tuning the ResNet model on the target data set, fixing the parameters of some layers, and only updating the parameters of the high-level network. The loss function of the fine-tuning process is: Among them, y i Output label for the target, is the predicted value, θ j is the network parameter, λ is the regularization coefficient, Loss(. is the cross entropy loss function, ∥.∥ 2 is the Frobenius norm, n is the total number of samples in the data set; S26, feature fusion, noise reduction and data standardization are performed on the extracted fry image features, and the fusion operation formula is: Among them, T f is the fused feature set, F i is the feature map of the i-th layer, w i is the importance weight of each layer of features, m is the number of feature layers, and principal component analysis is used to reduce the dimension of the fused features; S27, the fused feature set T f Input into the classifier, using ensemble learning method to improve accuracy by training multiple classifiers and fusing the output results of multiple classifiers: Among them, C final is the final classification result, w i is the weight of the i-th classifier, p i,c is the probability of the th classifier predicting category c, and k is the number of classifiers.

4. The method for dynamic monitoring and feeding optimization of fry culture based on deep learning according to claim 1, characterized in that: The S3 comprises the following steps: S31, water quality data D w , including temperature, dissolved oxygen, pH value, turbidity, and classification results of fish fry image data C final Combined, construct a multimodal input dataset X, X = {D w ,C final }, used to analyze the growth trends and activity patterns of fry; S32, using a long short-term memory network to perform time series analysis on the multimodal input data set X, and predicting the health status and activity trend of the fry in the future time period by inputting the historical data sequence; S33. Based on the prediction results of the long short-term memory network, combined with the time series characteristics of the water quality data, the support vector machine is used to further predict the growth trend of the fry and obtain the predicted value Y pred , Y pred Represents the predicted growth status of fry; S34. Use the weighted average method to combine the prediction results of different features to obtain a comprehensive score S: S=w1·Y pred +w2·D w ; Among them, w1 and w2 are weighted coefficients, Y pred is the predicted growth status of fry, D w For water quality data.

5. The method for dynamic monitoring and feeding optimization of fry culture based on deep learning according to claim 1, characterized in that: The S4 comprises the following steps: S41. Combined with water quality data D w And the classification result C final , a multi-objective adaptive PPO algorithm is used to obtain the feeding strategy, automatically adjust the feeding strategy to adapt to the needs of fry in different breeding environments, and update the strategy according to the fry health status and environmental feedback after each feeding; S42. Define the state space S: S={S w ,S f }; Among them, S w Represents water quality, including temperature, dissolved oxygen, pH value, turbidity, S f Represent the health status and behavioral characteristics of fry, including changes in body size, activity level, and feeding behavior; S43, define the action space A, A = {a1, a2, ..., a n }, a1 to a n They are different combinations of feeding amount, feeding frequency and feeding mode; S44, the strategy is optimized through the objective function of the multi-objective adaptive PPO algorithm, and the objective function is: Among them, r t (θ) is the probability ratio of the current strategy to the old strategy, is the advantage function of the fry health status, is the advantage function of resource utilization efficiency, is the advantage function of environmental protection, clip(.) represents the clipping operation, λ1 and λ2 are adjustment coefficients, ∈ is the clipping threshold, and E t represents the expected value at time step t; S45. Based on the real-time monitored water quality data, fry health status, behavioral characteristics and historical feeding data, the feeding strategy is optimized through the policy gradient algorithm, and the weight of the reward is dynamically adjusted in combination with the adaptive adjustment of the reward function. The reward function is: R t =w1·R health +w2·R efficiency +w3·R environment ; Among them, R health is a reward based on the health status of the fry, R efficiency is a reward based on resource efficiency, R environment is a reward based on environmental protection, w1, w2, and w3 are weight coefficients; S46, dynamically updating weight coefficients w1, w2, w3 during training through an adaptive adjustment mechanism of the reward function; S47, using the experience replay mechanism to train and optimize the multi-objective adaptive PPO algorithm: Q′ t =τQ t +(1-τ)Q′ t-1 ; Among them, τ is the soft update factor, Q t is the Q value of the current network output, Q′ t-1 Q value output by the target network; S48. A multi-objective adaptive PPO network is obtained through training, and the learned feeding strategy is mapped into actual feeding control instructions. The calculation formula of the feeding amount is: a t =argmax a∈A Q(S t ,a); Among them, a t In the current state S t The optimal feeding amount selected under the condition, A is the action space, Q(S t ,a) is the Q value of each action; S49, the multi-objective adaptive PPO network adjusts the feeding mode according to the dynamic changes of the fry population. If the fry density is too high, the model reduces the feeding amount or changes the feeding frequency by analyzing the water quality and the health status of the fry: Among them, Density t is the fry density at the current time point, w4 is the weight coefficient for adjusting the density, and exp(.) is the natural exponential function.

6. The method for dynamic monitoring and feeding optimization of fry culture based on deep learning according to claim 1, characterized in that: The S5 comprises the following steps: S51, using an anomaly detection algorithm to analyze the water quality and the health status of the fry, the anomaly detection algorithm calculates the anomaly metric value E by comparing the current water quality data and the fry behavior with the historical data model, the anomaly detection algorithm combines the deep autoencoder and the isolation forest algorithm, the deep autoencoder is used to learn the potential representation of the data and generate reconstruction errors to detect anomalies, and the isolation forest is used to further identify and filter extreme abnormal data; S52, the training process of the deep autoencoder includes inputting the water quality data and the fry health status characteristics at the current moment, mapping the water quality data and the fry health status characteristics to the latent space through the encoder, then restoring the data through the decoder, and calculating the reconstruction error as the abnormality metric. The calculation formula of the reconstruction error E is: Among them, X i is the original data of the i-th health status feature, is the reconstructed value of the i-th feature, n is the total number of feature data, E AE is the reconstruction error, which indicates the difference between the current data and the reconstructed data. A larger error value indicates data anomaly; S53, the isolation forest algorithm constructs multiple decision trees and determines abnormal points by the separation degree of sample points. Points with low separation degree are abnormal points. The abnormal measurement formula of the isolation forest is: Among them, h(x) is the isolation depth of sample point x, c(n) is the adjustment factor of the data set size, n is the total number of samples, E IF Anomaly calculated for Isolation Forest; S54. Combine the anomaly metrics of the deep autoencoder and the isolation forest to generate a weighted anomaly metric E combined : AND combined =w1·E AE +w2·E IF ; Among them, w1 and w2 are weighting coefficients; S55, if the comprehensive abnormality metric value E combined Greater than the set threshold E threshold , an abnormal alarm is triggered, and the breeding environment is adjusted according to the abnormal type. The alarm judgment formula is: Among them, Alert is the alarm signal, 1 means triggering the alarm, and 0 means not triggering the alarm.

Images