Schooling fish feeding desire evaluation method and device coupling adaptive forgetting learning and deep learning

Abstract

The application discloses a fish school feeding desire evaluation method and device coupling adaptive forgetting learning and deep learning. The method comprises the following steps: collecting videos during fish school breeding and feeding for training a recurrent neural network (RNN) fish school feeding desire evaluation model, and optimizing the model by using an adaptive forgetting learning technology. The adaptive forgetting learning technology mainly comprises three innovative technologies, i.e., an adaptive label flipping strategy, a double gradient correction mechanism and multi-round forgetting cumulative error analysis and control. Without retraining the model from the beginning, the model can forget the early collected but outdated training data, effectively improving the accuracy and self-updating adaptive ability of the fish school feeding desire evaluation model. The method is accurate, simple and efficient, the device structure is simple, the problem that the accuracy of the feeding desire evaluation model in aquaculture is reduced due to outdated data can be effectively solved, the model can be self-adapted and self-optimized in long-term operation, and the intelligentization of aquaculture management can be realized.

Description

Technical Field

[0001] This invention belongs to the field of industrialized intelligent aquaculture, and relates to a method and device for assessing the feeding desire of fish, particularly a method and device for assessing the feeding desire of fish that couples adaptive forgetting learning and deep learning. Background Technology

[0002] Currently, numerous studies have focused on the application of deep learning in aquaculture, with an increasing number of researches on assessing and predicting feeding desire intensity. These studies often propose or improve various deep learning models, which can provide relatively accurate results in experimental environments. However, real aquaculture environments are far more complex and diverse, making it difficult to find two completely identical environments. Therefore, most models developed in laboratories do not perform ideally in actual production and are difficult to realize their practical application value. Even under the same aquaculture environment, due to changes in various dynamic factors, early-collected training data may become outdated and lose its reference value, or even become a detrimental factor. Changes in farmed species, batch sizes, water quality, ambient light, and animal growth all cause changes in the timeliness of data. The traditional solution is to frequently retrain the model using new data, but this method is computationally very expensive on large-scale datasets and complex models, making it difficult to promote and use extensively in the aquaculture industry.

[0003] Based on the above background, this invention proposes a method and apparatus for assessing fish feeding desire by coupling adaptive forgetting learning and deep learning. It utilizes adaptive forgetting learning technology to optimize the fish feeding desire assessment model, effectively forgetting outdated data to mitigate its adverse effects without requiring retraining the model from scratch, thereby improving the model's accuracy and self-updating adaptive capability. This invention effectively solves the problem of decreased accuracy in current aquaculture feeding desire assessment models due to outdated data, ensuring the model's self-adaptive optimization during long-term operation, which is beneficial for achieving intelligent aquaculture management. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of decreased accuracy in fish feeding desire assessment models in aquaculture due to outdated data, and to provide a method and apparatus for assessing fish feeding desire by coupling adaptive forgetting learning and deep learning. This method can achieve: 1. Real-time acquisition and efficient processing of video data on fish farming and feeding; 2. Adaptive forgetting learning for outdated data, enabling the model to self-optimize and update; 3. Effectively improving the long-term accuracy and adaptability of the fish feeding desire assessment model.

[0005] The technical solution adopted in this invention is as follows:

[0006] A method for assessing the feeding desire of fish swarms by coupling adaptive forgetting learning and deep learning includes:

[0007] Video data from fish rearing and feeding periods is collected and labeled with fish feeding desire tags to form a training set. This training set is then used to train a recurrent neural network model for assessing fish feeding desire. During training, an adaptive forgetting learning technique is employed to optimize the model, mitigating the adverse effects of outdated data and improving its accuracy and self-updating adaptability. This adaptive forgetting learning technique involves creating a forgotten dataset and a retained dataset, and then using an adaptive label flipping strategy and a dual-gradient correction mechanism to obtain training gradients for optimizing the training of the fish feeding desire assessment model through forgetting learning.

[0008] Furthermore, in the above technical solution, the adaptive label flipping strategy includes:

[0009] The training set is divided proportionally as needed to create a forgotten dataset. For each sample in the forgotten dataset, the maximum value of the Hadamard product of its corresponding model predicted probability vector and the mask vector of the true label is calculated as an overfitting index. Based on this overfitting index, the label flipping intensity is dynamically adjusted according to the principle that the higher the overfitting degree, the more thorough the flipping needs to be. ; intensify the label flipping When applied to the real label mask, only the probability values ​​of the real category positions are retained. The other positions are completely preserved, generating flipped labels; finally, based on the flipped labels and combined with the degree of overfitting, the forgetting loss function is derived.

[0010] Furthermore, the flipped label The target label vector after adaptive label flipping is a probability distribution vector, with the sum of all elements being 1. Its generation method is as follows:

[0011]

[0012] in, For the Hadamard product operator, For label flip strength, Predict probability vectors for the model. The mask vector for the actual label. This represents the probability that the model predicts a sample belongs to the j-th food desire category. This is the value of the j-th food desire category in the real label mask vector.

[0013] Furthermore, the formula for calculating the forgetting loss function is as follows:

[0014]

[0015] in As an indicator of overfitting, The overfit detection threshold, These are the true class labels of the samples. To flip the label, Here, represents the model's predicted probability vector, and CrossEntropy is the multi-class cross-entropy loss function.

[0016] Furthermore, the dual gradient correction mechanism includes:

[0017] Calculate the forget gradient on the forgotten dataset and the retain gradient on the retained dataset respectively; calculate the angle between the forget gradient and the retain gradient, and determine whether the two gradients need to be orthogonalized based on the angle. If so, calculate the conflict coefficient and orthogonally correct the forget gradient and the retain gradient according to the conflict coefficient; weight the corrected forget gradient and the retain gradient to obtain the final gradient.

[0018] Furthermore, the forgotten gradient is a gradient vector obtained by taking the partial derivative of the forgotten loss function on the forgotten dataset, and the retained gradient is a gradient vector obtained by taking the partial derivative of the standard loss function during the model training process on the retained dataset.

[0019] Furthermore, the angle between the forgotten gradient and the retained gradient is calculated. , Represents the forgetting gradient vector norm, Indicates preserving the gradient vector The norm, when When, no correction is needed; when When, the two gradients are orthogonal; when When calculating the conflict coefficient: Orthogonalization correction is applied to the two gradients: , The final gradient is obtained by weighted synthesis.

[0020] Furthermore, the adaptive forgetting learning technique also includes multi-round forgetting accumulation error analysis and control, specifically including:

[0021] A cumulative error model is established, and the retention loss after multiple rounds of forgetting is predicted based on the cumulative error model; a dynamically adjusted learning rate is introduced to control the growth of the cumulative error; a cumulative error bound is defined, and the cumulative error is monitored in real time during the forgetting process. When the cumulative error exceeds the maximum allowable value, the forgetting process is stopped.

[0022] Furthermore, a cumulative error model is established, in which the increase in loss on retained data after the i-th round of forgetting training is proportional to the norm of the forgetting gradient of the i-th round of forgetting, thus enabling the prediction of retention loss after the n-th round of forgetting, in order to understand the degree of impact of multiple rounds of forgetting on retained data in advance; a dynamically adjusted learning rate is introduced in each round of forgetting training, and the learning rate gradually decreases as the number of forgetting rounds increases, thereby controlling the cumulative error.

[0023] The present invention also provides a fish feeding desire assessment device coupled with adaptive forgetting learning and deep learning, for implementing the method described in any of the preceding claims, the device comprising:

[0024] Video acquisition module: used to collect video data during fish farming and feeding; Deep learning module: used to train a recurrent neural network model for evaluating the feeding desire of fish using the video data; Adaptive forgetting learning module: used to optimize the model using adaptive forgetting learning technology. The adaptive forgetting learning module includes an adaptive label flipping strategy submodule, a dual gradient correction mechanism submodule, or, based on these, a multi-round forgetting cumulative error analysis and control submodule.

[0025] The beneficial effects of this invention are:

[0026] This invention effectively analyzes data from the aquaculture process by coupling adaptive forgetting learning and deep learning technologies. The adaptive forgetting learning technique mitigates the adverse effects of outdated data on the model, significantly improving the accuracy and self-updating adaptability of the fish feeding desire assessment model. The adaptive forgetting learning technique primarily includes three innovative technologies: an adaptive label flipping strategy, a dual-gradient correction mechanism, and further multi-round forgetting cumulative error analysis and control. This allows the model to forget previously collected but outdated training data without needing to retrain from scratch. This invention is precise, simple, and efficient, with a simple device structure. It effectively solves the problem of decreased accuracy in current aquaculture feeding desire assessment models due to outdated data, ensuring the model's self-adaptive optimization during long-term operation and facilitating intelligent aquaculture management. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of a fish feeding desire assessment device that couples adaptive forgetting learning and deep learning in this invention. Detailed Implementation

[0028] The invention will now be further described with reference to the accompanying drawings.

[0029] This invention provides a method and apparatus for assessing the feeding desire of fish swarms by coupling adaptive forgetting learning and deep learning. First, real-time video data of the fish rearing and feeding process is collected using a camera. A feeding desire assessment model is trained using a recurrent neural network (RNN) and deep learning algorithms. Then, the model is optimized using adaptive forgetting learning technology. The steps include: analyzing and processing outdated data using an adaptive label-flipping strategy; precisely controlling gradient and model changes using a dual-gradient correction mechanism during the forgetting learning process; and verifying and ensuring the effectiveness of the forgetting learning using multi-round forgetting cumulative error analysis and control techniques. The method of this invention mainly involves the following aspects:

[0030] 1. A feeding desire assessment device for fish swarms that couples adaptive forgetting learning and deep learning.

[0031] The device includes hardware such as a breeding pond, camera, display, and computer, as well as software functional modules including a video acquisition module, an RNN model training and inference module, and an adaptive forgetting learning module, wherein:

[0032] Video Acquisition System: The camera is positioned directly above the aquaculture pond, with a top-down view covering the entire pond. The frame rate is no less than 30fps, and the resolution is no less than 1920×1080. The camera is connected to the computer's input terminal for real-time video data transmission. The acquisition device uses a waterproof metal casing, and the image acquisition surface of the camera is made of transparent acrylic glass. Internal lighting uses artificial light sources with uniform LED illumination, ensuring no projection interference in the acquired images.

[0033] Deep learning system: The computer is equipped with an NVIDIA RTX 4060 Ti GPU and the PyTorch software framework for training and inference of RNN models. It uses recurrent neural networks (RNNs) to learn time-series features of labeled fish farming and feeding processes.

[0034] The adaptive forgetting learning system includes an adaptive label flipping strategy, a dual gradient correction mechanism, and a multi-round forgetting cumulative error analysis and control submodule. It is used to optimize the RNN model, enabling it to forget outdated data and reduce the adverse effects of outdated data on the model.

[0035] 2. Use an adaptive label flipping strategy for the analysis and processing of outdated data.

[0036] By introducing an overfitting index to dynamically adjust the label flipping intensity, refined processing of outdated data is achieved. Specific steps include:

[0037] Overfit analysis: For each sample (x, y) in the forgotten dataset, calculate the model's predictions. Then, the maximum value of the elements in the Hadamard product (i.e., element-wise dot product) between the model's predicted probability vector and the mask vector of the true label is calculated as an indicator of overfitting. This metric reflects the degree to which the model has learned and memorized the sample, among which This is the predicted probability vector for the model, used in the fish feeding desire assessment model of this method. Each element This represents the probability that the model predicts the sample belongs to the j-th class. All elements are non-negative and their sum is 1. The five categories represent five different intensities of appetite: extremely strong, strong, moderate, weak, and extremely weak. The mask vector is a one-hot encoded vector that represents the true label. This metric directly reflects the model's confidence in predicting the true label. Only the position corresponding to the true category is 1, and all other positions are 0.

[0038] Adaptive flip strength calculation: based on overfitting index Based on the principle that the higher the degree of overfitting, the more thorough the flipping needs to be, the label flipping intensity is dynamically adjusted: The formula was derived through comparison and optimization experiments. When the overfit is 0.5, α = 0, and no label flipping is performed; when the overfit is 0.75, α = 0.5, and 50% of the true class probability is masked; when the overfit is 0.95, α = 0.9, and 90% of the true class probability is masked.

[0039] Flip Label Generation: Generate flip labels based on the calculated flip strength. , It is the target label vector after adaptive label flipping. It is a probability distribution vector with the sum of all elements being 1, and the value range is between 0 and 1. This vector is used to calculate the forgetting loss function to guide the model to learn the features of outdated data.

[0040] in For a weighted mask vector, this operation will flip the intensity. When applied to the real label mask, the value at the real category position is... The value is 0 in other positions; The retained coefficient vector is the difference between the all-ones vector and the weighted mask vector. This vector represents the probability proportion of each position retained, with a value of 1- at the true class position. (Retain 1-) The probability of the true class is 1 in other positions (fully preserved). The purpose of this design is to mask part of the probability of the true class and let the model learn the distribution of other classes. It is the flipped prediction vector. This operation preserves the coefficients applied to the model's predicted probabilities, resulting in an unnormalized vector (the sum of its elements is not necessarily 1). This is the normalization factor, which is the sum of the probabilities of all categories after flipping. This value is used to normalize the flipped vector to a probability distribution, ensuring the final normalization. It is an effective probability distribution.

[0041] Forgetting Loss Calculation: The forgetting loss function guides how the model handles outdated data and is a key indicator in the forgetting learning process. The magnitude of the forgetting loss directly affects the direction and magnitude of model parameter updates. When the forgetting loss is large, the model will adjust its parameters more aggressively to forget outdated data; when the forgetting loss is small, the model's adjustments will be more gradual. This method calculates the forgetting loss... Its simplified form is ,in The overfitting detection threshold is used to determine whether the data has been sufficiently learned; the value used in this method is [value missing]. ; K is the true class label of the sample; K is the total number of classes; N is the total number of samples in the forgetting set; CrossEntropy is the cross-entropy loss function for multi-class classification, a commonly used classification loss function used to measure the difference between the model's prediction and the target label. In this formula, for samples with high overfitting, label flipping is used to calculate the forgetting loss; for samples with low overfitting, negative loss is used for gradient ascent. The advantage of this adaptive strategy is that it correlates the forgetting intensity with the learning state of the data, making the forgetting process more refined and controllable. A stronger flipping intensity is used for fully learned data to ensure effective forgetting; a weaker flipping intensity is used for insufficiently learned data to avoid excessive forgetting; through continuous nonlinear mapping, a smooth transition from "no flipping" to "complete flipping" is achieved.

[0042] 3. Employ a dual-gradient correction mechanism to precisely control the forgetting learning process.

[0043] A precise balance between the forgetting gradient and the retention gradient is achieved by calculating and analyzing the conflict between them. The specific steps include:

[0044] Gradient calculation: Calculate the forgetting gradient on the forgotten dataset and the retention gradient on the retained dataset respectively. , ,in Forgotten datasets are collections of all outdated data that need to be forgotten. In the assessment of feeding desire in aquaculture, forgotten datasets contain training samples that have lost their reference value due to the passage of time or changes in the environment. To preserve the dataset; Forgetting loss function; This is a predetermined standard loss function during model training; These are the initial model parameters, i.e., the model parameters before calculating the gradient; this is the initial point for gradient calculation. The symbol is for partial derivatives, indicating that taking the partial derivatives with respect to all model parameters yields a gradient vector. Therefore, we can conclude that... The gradient vector is obtained by taking the partial derivative of the forgetting loss function. It represents the rate of change of the forgetting loss function with respect to the model parameters, and guides how the model parameters are adjusted to achieve the forgetting of outdated data. The magnitude of the forgetting gradient reflects the magnitude of the parameter change required, and the direction of the gradient indicates the direction of the parameter change. Preserving gradients is about preserving the dataset. The partial derivative of the standard loss function is obtained, which guides the model on how to improve its performance on the retained data.

[0045] Gradient Angle Analysis: Calculate the angle between the forgotten gradient and the retained gradient. ,

[0046] in Represents the forgetting gradient vector The norm of the gradient vector, which is the size or length of the gradient vector, is a scalar value used to measure the overall size of the gradient vector. Similarly, this angle reflects the degree of similarity between the two gradient directions. When When the gradients point in similar directions, the two gradients point in similar directions; when... When, the two gradients point in opposite directions; when At that time, the two gradients are orthogonal.

[0047] Conflict coefficient calculation: Calculate the conflict coefficient based on the gradient angle. ,

[0048] when hour, This indicates that the two gradients do not conflict and do not require correction; when hour, This indicates that there is a conflict between the two gradients, which needs to be corrected; the conflict coefficient ranges from 0 to 1.

[0049] Bi-gradient correction: Orthogonalizing and correcting two gradients. , The geometric meaning of the correction is to remove the component parallel to the retained gradient from the forgotten gradient, and at the same time remove the component parallel to the forgotten gradient from the retained gradient, so that the two corrected gradients are orthogonal.

[0050] Corrected gradient synthesis: The two corrected gradients are then weighted and synthesized. ,in The gradient mixing ratio represents the weight of the forgotten target. To preserve the weights of the objectives, this weight ratio is derived from experimental optimization and achieves an optimal balance between the two objectives.

[0051] Forgetting learning: Use the corrected gradient from the previous step to perform forgetting learning.

[0052] Compared to unidirectional correction, the dual gradient correction mechanism can simultaneously consider the impact of the forgetting gradient on the retained data and the impact of the retained gradient on the forgetting effect; by orthogonalizing the gradients, it can achieve a better balance between the two objectives in parameter updates; it can effectively handle various situations of gradient conflict and improve the stability of forgetting learning; it has a clear geometric meaning and is easy to understand and verify.

[0053] 4. Employ multi-round forgetting cumulative error analysis and control techniques to ensure effective forgetting learning.

[0054] By establishing a cumulative error model and a dynamic learning rate adjustment mechanism, the stability of model performance during multiple rounds of forgetting is ensured. Specific steps include:

[0055] Cumulative error model establishment: Let the model parameters after the i-th round of forgetting be... The increase in loss of retained data is: (in (Using the standard loss function), assuming this increase is proportional to the norm of the forgetting gradient: , where c is a constant that can be estimated through the previous rounds of forgetting. The rationale for this assumption is that the magnitude of the forgetting gradient directly reflects the magnitude of parameter change, and the greater the magnitude of parameter change, the greater the impact on the retained data.

[0056] Retention Loss Prediction: Based on the cumulative error model, the retention loss after the nth round of forgetting can be predicted. This prediction formula can help us understand in advance the extent to which multiple rounds of forgetting affect the retained data.

[0057] Dynamic learning rate adjustment: To control cumulative error, a dynamically adjusted learning rate is introduced in each round of forgetting. This learning rate gradually decreases as the number of forgetting rounds increases, thereby controlling for cumulative error. In this method, the learning rate is... The initial learning rate is the value that achieves the optimal balance between forgetting effects and computational stability; The decay coefficient is a value that ensures the learning rate decays at a moderate rate with each round, neither too fast leading to insufficient forgetting in subsequent rounds, nor too slow leading to excessive cumulative error.

[0058] Cumulative error bound definition: Define the cumulative error bound to quantify the overall effect of multiple rounds of forgetting: In this method This represents the maximum allowable cumulative error. This limit is derived from experimental optimization and achieves an optimal balance between ensuring the forgetting effect and retaining data performance.

[0059] This method monitors and controls the forgetting process, tracking the accumulated error in real time. After each round of forgetting, the current accumulated error value is calculated. If the accumulated error does not exceed the maximum allowable value, the next round of forgetting continues. If the accumulated error exceeds the maximum allowable value, the system automatically stops the forgetting process to protect the performance of the retained data. The advantages of this method are: it can predict the impact of multiple rounds of forgetting on the retained data and take control measures in advance; it effectively controls the growth of accumulated error through dynamic learning rate adjustment; the accumulated error bound provides a clear stopping condition, ensuring the stability of model performance; and it maintains the accuracy of the model and avoids severe performance degradation during long-term operation.

[0060] Reference Figure 1 This is a simple example of a fish feeding desire assessment device that couples adaptive forgetting learning and deep learning according to the present invention. It includes a rearing pond 1, a camera 2, a computer 3, and a display screen 4. The camera 2 is installed directly above the rearing pond 1 and connected to the computer 3. It transmits real-time video data of the fish rearing and feeding process to the computer 3. The computer 3 inputs this data into its built-in deep learning system and adaptive forgetting learning system, and finally outputs the real-time feeding desire intensity result on the display screen 4. The rearing pond is connected to an external recirculating water treatment system to improve the water resource utilization rate of the entire recirculating aquaculture system, which aligns with the concept of sustainable development.

[0061] The method for assessing the feeding desire of fish using the above-mentioned device, which couples adaptive forgetting learning and deep learning, includes the following steps:

[0062] (1) Use a camera to capture real-time footage of the fish rearing and feeding process in the aquaculture pond, with a frame rate of no less than 30fps and a resolution of no less than 1920×1080. The camera is connected to the input terminal of the computer to transmit the video footage to the computer in real time;

[0063] (2) The feeding desire assessment model of fish is trained by using deep learning algorithms and recurrent neural networks. After a certain breeding stage and a certain amount of data is collected, the forgotten dataset and the retained dataset are divided and input into the adaptive forgetting learning system together with the feeding desire assessment model.

[0064] (3) The adaptive forgetting learning system executes an adaptive label flipping strategy and a dual gradient correction mechanism to form the final correction gradient. Learn by forgetting.

[0065] (4) The adaptive forgetting learning process uses multi-round forgetting cumulative error analysis and control throughout the entire process.

[0066] It should be noted that the recurrent neural network fish feeding desire assessment model involved in this invention is not the only model that can be used with the method of this invention. Any model based on recurrent neural networks is applicable to the method of this invention, regardless of the model input, structure, parameters, etc. In addition, this method can also be used in other fields.

[0067] The following is an application example of the above-described embodiments to demonstrate the actual effect of the method of the present invention. The research subject in this example is the largemouth bass, provided by Suzhou Taicang Aikeman Technology Co., Ltd. The initial average weight of the experimental fish was (100.0±3.0) g, and they were randomly divided into three groups: A, B, and C. An experiment was conducted in a recirculating aquaculture system for three months, with a stocking density of 60 kg / m³ (referring to actual production standards). The daily light cycle was set at 12 hours of light: 12 hours of darkness (12L:12D). Tongwei brand floating pellet feed (crude protein ≥48%, crude fat ≥6%, particle size 3.0 mm) was used in the experiment. Before the experiment, the fish were acclimatized in the recirculating aquaculture system for three weeks, during which the water temperature was maintained at (25±0.5) ℃ and the dissolved oxygen concentration at ≥5.5 mg / L. During the experiment, group A was placed in a factory-style aquaculture environment to simulate an environment that would generate outdated data; groups B and C were placed in a different factory-style aquaculture environment than group A, representing the existing aquaculture environment. The fish rearing and feeding process in the aquaculture pond was recorded in real time using cameras at a frame rate of at least 30fps and a resolution of at least 1920×1080. Each group of 1000 valid videos was collected and labeled, resulting in a total of 3000 valid data points. The first 2000 data points were assigned to the training set, and the last 1000 to the test set. A fish feeding desire assessment model was trained using deep learning algorithms and recurrent neural networks. The training set was further divided into a forgetting dataset (first 25%) and a retention dataset (last 75%), which, along with the feeding desire assessment model, were input into an adaptive forgetting learning system for experiments.

[0068] To verify the effectiveness of the method of this invention, the same data was used to perform forgetting learning through the full retraining method, the Fisher information matrix method, and the gradient ascent method for comparison. After the experiment, we compared the performance of the four methods, as shown in Table 1 below. In terms of forgetting effect, the loss increase of the full retraining method reached the highest at 0.85, which is the ideal situation. The loss increase of the Fisher information matrix method reached 0.52, the gradient ascent method reached 0.32, and the method of this invention reached 0.48. In terms of retention utility, the loss change of the full retraining method remained the best at 0.0, the loss change of the Fisher information matrix method was 0.0089, the gradient ascent method was 0.0156, and the method of this invention was 0.0042. In terms of test set prediction accuracy, the accuracy of the model after full retraining on the test set reached 89.64%, the accuracy of the Fisher information matrix method was 84.15%, the accuracy of the gradient ascent method was only 80.39%, and the accuracy of the method of this invention reached 87.10%. In terms of computation time, full retraining took 35,781 seconds, the Fisher information matrix method took 11,322 seconds, the gradient ascent method took 29 seconds, and the method of this invention took 47 seconds. Therefore, compared to the other three methods, this method balances forgetting learning time with model prediction accuracy, achieving good forgetting learning results with less forgetting learning time, and significantly increasing the model's prediction accuracy in the new breeding environment.

[0069] Table 1 Comparison of the learning effects of different methods for forgetting

[0070] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0071] It should be understood that the methods of the present invention can be implemented by computer program instructions, which can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor or other programmable data processing device to produce a machine, such that the instructions, which are executed by the processor of the computer or other programmable data processing device, produce methods for implementing the present invention.

[0072] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement one or more steps of the method of the present invention.

[0073] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide functionality for implementing one or more steps in the method of the present invention.

[0074] The above-disclosed embodiments are merely specific examples of this solution, but this solution is not limited thereto. For those skilled in the art, any modifications made without departing from this invention should be considered to fall within the scope of protection of this invention.

Claims

1. A method for assessing the feeding desire of fish swarms by coupling adaptive forgetting learning and deep learning, characterized in that, include: Video data during fish farming and feeding was collected, labeled with fish feeding desire tags, and a training set was formed. The training set was then used to train a recurrent neural network model for evaluating fish feeding desire. During training, the adaptive forgetting learning technique is used to optimize the fish feeding desire assessment model, so as to reduce the adverse effects of outdated data on the model during the breeding process and improve the model's accuracy and self-updating adaptive ability. The adaptive forgetting learning technique includes forming a forgotten dataset and a retained dataset, and on this basis, using an adaptive label flipping strategy and a dual gradient correction mechanism to obtain training gradients for optimizing the training of the fish feeding desire assessment model to perform forgetting learning. The adaptive label flipping strategy includes: dividing the training set into proportions to extract samples as a forgetting dataset; for each sample in the forgetting dataset, calculating the maximum value of the Hadamard product of its corresponding model prediction probability vector and the mask vector of the true label as an overfitting index; and dynamically adjusting the label flipping intensity based on the overfitting index, adhering to the principle that the higher the overfitting degree, the more thorough the flipping needs to be. ; intensify the label flipping When applied to the real label mask, only the probability values ​​of the real category positions are retained. The other positions are completely preserved, generating flipped labels; finally, based on the flipped labels and the degree of overfitting, the forgetting loss function is derived. The flip label The target label vector after adaptive label flipping is a probability distribution vector, with the sum of all elements being 1. Its generation method is as follows: ， in, For the Hadamard product operator, For label flip strength, Predict probability vectors for the model. The mask vector for the actual label. This represents the probability that the model predicts a sample belongs to the j-th food desire category. This is the value of the j-th food desire category in the real label mask vector.

2. The method for assessing fish feeding desire by coupling adaptive forgetting learning and deep learning according to claim 1, characterized in that, The formula for calculating the forgetting loss function is as follows: ， in As an indicator of overfitting, The overfit detection threshold, These are the true class labels of the samples. To flip the label, Here, represents the model's predicted probability vector, and CrossEntropy is the multi-class cross-entropy loss function.

3. The method for assessing fish feeding desire by coupling adaptive forgetting learning and deep learning according to claim 1, characterized in that, The dual gradient correction mechanism includes: Calculate the forget gradient on the forgotten dataset and the retain gradient on the retained dataset respectively; calculate the angle between the forget gradient and the retain gradient, and determine whether the two gradients need to be orthogonalized based on the angle. If so, calculate the conflict coefficient and orthogonally correct the forget gradient and the retain gradient according to the conflict coefficient; weight the corrected forget gradient and the retain gradient to obtain the final gradient.

4. The method for assessing fish feeding desire by coupling adaptive forgetting learning and deep learning according to claim 3, characterized in that, The forgetting gradient is a gradient vector obtained by taking the partial derivative of the forgetting loss function on the forgotten dataset, and the retaining gradient is a gradient vector obtained by taking the partial derivative of the standard loss function during the model training process on the retained dataset.

5. The method for assessing fish feeding desire by coupling adaptive forgetting learning and deep learning according to claim 3, characterized in that, Calculate the angle between the forgotten gradient and the retained gradient. , Represents the forgetting gradient vector norm, Indicates preserving the gradient vector The norm, when When, no correction is needed; when When, the two gradients are orthogonal; when When calculating the conflict coefficient: Orthogonalization correction is applied to the two gradients: , The final gradient is obtained by weighted synthesis.

6. The method for assessing fish feeding desire by coupling adaptive forgetting learning and deep learning according to claim 1, characterized in that, The adaptive forgetting learning technique also includes multi-round forgetting accumulation error analysis and control, specifically including: A cumulative error model is established, and the retention loss after multiple rounds of forgetting is predicted based on the cumulative error model; a dynamically adjusted learning rate is introduced to control the growth of the cumulative error; a cumulative error bound is defined, and the cumulative error is monitored in real time during the forgetting process. When the cumulative error exceeds the maximum allowable value, the forgetting process is stopped.

7. The method for assessing fish feeding desire by coupling adaptive forgetting learning and deep learning according to claim 6, characterized in that, A cumulative error model is established, in which the increase in loss on retained data after the i-th round of forgetting training is proportional to the norm of the forgetting gradient of the i-th round of forgetting, thus enabling the prediction of the retention loss after the n-th round of forgetting, in order to understand the impact of multiple rounds of forgetting on the retained data in advance; a dynamically adjusted learning rate is introduced in each round of forgetting training, and the learning rate gradually decreases as the number of forgetting rounds increases, thereby controlling the cumulative error.

8. A device for assessing the feeding desire of fish schools by coupling adaptive forgetting learning and deep learning, characterized in that, The apparatus for implementing the method as described in any one of claims 1-7 comprises: Video acquisition module: used to collect video data during fish farming and feeding; Deep learning module: used to train a recurrent neural network model for evaluating the feeding desire of fish using the video data; Adaptive forgetting learning module: used to optimize the model using adaptive forgetting learning technology. The adaptive forgetting learning module includes an adaptive label flipping strategy submodule, a dual gradient correction mechanism submodule, or, based on these, a multi-round forgetting cumulative error analysis and control submodule.

Images