Precise feeding system based on multi-source heterogeneous data driving and artificial intelligence

Abstract

The invention discloses an accurate feeding system based on multi-source heterogeneous data driving and artificial intelligence. The accurate feeding system comprises a control unit, a data acquisition unit, a feeding unit and a feeding monitoring unit. The data acquisition unit comprises an environmental parameter acquisition module and a water quality parameter acquisition module; the feeding monitoring unit comprises an intelligent camera and an underwater identification frame. The feeding method comprises the following steps: S1, data acquisition; s2, the control unit determines the appropriate feeding amount based on the environment parameters, the water quality parameters and the feeding rule data; s3, feeding according to the basic feeding amount, monitoring the feeding condition of the fishes through a feeding monitoring unit after feeding, and performing secondary feeding according to the feeding condition. By means of the accurate feeding system, the bait coefficient can be remarkably reduced, bait waste is reduced, accurate mastering and timely regulation and control of fish feeding requirements can be achieved by using multi-source heterogeneous data such as environment data, water quality data and feeding data, and the scientificity and accuracy of feeding can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of feeding devices for fish farming, and in particular to a precision feeding system based on multi-source heterogeneous data drive and artificial intelligence. Background Art

[0002] Feed is the main cost expenditure in aquaculture (accounting for about 70% of the total cost of aquaculture), and traditional feeding methods often lead to feed waste. In addition, residual feed is also one of the main sources of water pollution. With the intensification and scale of aquaculture, traditional artificial feeding methods can no longer meet the needs of precise and efficient aquaculture. Precision feeding systems have received more and more attention. Precision feeding systems can significantly improve the utilization efficiency of feed and reduce feed waste, which is crucial to reducing aquaculture costs and improving economic benefits. Secondly, precision feeding helps reduce water pollution. Excessive feed will lead to deterioration of water quality and increase the risk of disease, while precision feeding can reduce the amount of uneaten feed that sinks to the bottom of the pond, thereby reducing pressure on the environment. In addition, precision feeding can also improve the quality of aquatic products and promote the healthy growth of farmed animals, which plays an important role in ensuring food safety and enhancing product market competitiveness. Currently, most farmers have begun to use automatic feeders for feeding operations. Automatic feeders can feed fish with a fixed time and quantity. However, the specific feeding amount is currently based on the personal experience of the staff, and its accuracy is poor. It often cannot make timely changes based on environmental factors, bait eating conditions and fish growth plans. Summary of the invention

[0003] The purpose of the present invention is to provide a precise feeding system based on multi-source heterogeneous data drive and artificial intelligence to solve the problems existing in the prior art.

[0004] In order to achieve the above-mentioned purpose, the technical solution of the present invention provides a precision feeding system based on multi-source heterogeneous data drive and artificial intelligence, wherein the precision feeding system comprises a control unit, a data acquisition unit, a feeding unit, and a feeding monitoring unit; the control unit is electrically connected to the data acquisition unit, the feeding unit, and the feeding monitoring unit; the data acquisition unit comprises an environmental parameter acquisition module and a water quality parameter acquisition module; the feeding unit is an automatic feeding machine; and the feeding monitoring unit comprises an intelligent camera and an underwater identification frame.

[0005] Furthermore, the feeding monitoring unit also includes an infrared fill light and a connecting rod; the underwater identification frame is a rectangular frame, the underwater identification frame is connected to the connecting rod through a connecting cross bar, and the smart camera is arranged on the connecting rod at the lower part of the underwater identification frame, and its viewing direction is obliquely upward, and the underwater identification frame is within its viewing range.

[0006] Furthermore, the feeding unit includes a silo, a feeding screw assembly, and a feeding assembly. A discharge port is provided at the bottom of the silo. The feeding screw assembly includes an outer shell, a central screw, and a driving motor. The outer shell is tubular and is provided below the discharge port of the silo. The discharge port is connected to the interior of the outer shell through a connecting pipe. The central screw is provided in the outer shell. A quantitative discharge port is provided in the lower part of the outer shell away from the discharge port. One end of the central screw passes through the side wall of the outer shell and is connected to the driving motor. The feeding assembly is a throwing tray or a pneumatic feeding device.

[0007] The present application also discloses a feeding method based on the above-mentioned precision feeding system, comprising the following steps: step S1, a data acquisition unit collects and obtains environmental parameters and water quality parameters; a control unit presets a fish feeding rule; step S2, the control unit determines a suitable feeding amount based on the environmental parameters, water quality parameters obtained by the data acquisition unit and the preset fish feeding rule data; step S3, the control unit controls the feeding unit to feed according to 60% to 70% of the feeding amount obtained in step S2 as a basic feeding amount, monitors the fish feeding situation through the feeding monitoring unit after the preset feeding time, determines a variable feeding amount according to the feeding situation, and controls the feeding unit to perform secondary feeding.

[0008] Furthermore, the step S3 of "monitoring the feeding situation of fish through a feeding monitoring unit after feeding, and determining the change of feeding amount according to the feeding situation" specifically includes the following steps: step S31, residual bait recognition model construction, constructing a residual bait recognition model based on the YOLO model; step S32, baseline establishment, generalizing the model in the actual use environment, and obtaining the residual bait situation data in the sampling area under normal feeding conditions; step S33, residual bait situation evaluation, using the residual bait recognition application model combined with the baseline data to evaluate the fish feeding situation.

[0009] Furthermore, the step S31 includes the following steps: S311, data collection: collecting underwater residual bait images or video data for training and verifying the residual bait recognition model; S312, data preprocessing: preprocessing the collected image or video data; manually annotating the preprocessed data and dividing it into a training data set and a verification data set; S313, model construction and training, constructing a residual bait recognition model, using the annotated training data set to train the model, and optimizing the model performance by adjusting the model parameters. Afterwards, the verification data set is used to evaluate the accuracy of the residual bait recognition model, and the residual bait recognition model is optimized according to the evaluation results.

[0010] Furthermore, in the step S31, the data collection object is a video stream, and the data preprocessing includes key frame extraction, image cropping, CLAHE enhancement, and Mosaic data enhancement; the residual bait recognition model includes a key frame extraction and preprocessing module, a time series fusion module, and a YOLO detection module. After the data is processed by the time series fusion module, it is input into the YOLO detection module to output the number of residual baits.

[0011] Furthermore, the image cropping is performed based on the underwater identification frame in the key frame image and in accordance with the size and position of the preset sampling area in the image.

[0012] Furthermore, the step S32 includes the following steps: step S321, model generalization: collecting underwater residual bait images or video data in the actual use environment; preprocessing the collected image or video data; manually annotating the preprocessed data, and dividing it into a generalized training data set and a generalized verification data set, and obtaining a residual bait recognition application model after generalization training; step S322, baseline establishment: determining the baseline data for residual bait evaluation through the residual bait situation data in the sampling area under normal feeding conditions.

[0013] Furthermore, the step S33 includes the following steps: step S331, real-time data collection, after the preset feeding time, using the residual bait recognition application model trained in step S32 to obtain the residual bait data of the sampling area in real time analysis; step S332, data comparison and analysis: comparing and analyzing the residual bait data obtained in step S331 with the baseline data established in step S32 to evaluate the eating situation of the fish; determining the change in feeding amount based on the comparison and analysis results.

[0014] The technical solution of the present invention has the following technical effects: through the use of the precision feeding system, the feed coefficient can be significantly reduced, the waste of feed can be reduced, the eutrophication level of water bodies and the content of harmful substances can be reduced, the occurrence of fish diseases and the use of fishery drugs can be reduced, thereby improving economic benefits. In the specific technical solution, the precision feeding system can accurately grasp and timely regulate the feeding needs of fish through the use of multi-source heterogeneous data such as environmental data, water quality data, and feeding data, which is conducive to improving the scientificity and accuracy of feeding. In terms of feeding data, by adopting artificial intelligence-based machine vision recognition models such as YOLO, the remaining bait situation can be more accurately monitored to obtain more accurate feeding data. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the overall scheme of the present invention.

[0016] Figure 2 It is a partial schematic diagram of a feeding unit according to an embodiment of the present invention.

[0017] Figure 3Schematic diagram of a meal monitoring unit according to an embodiment of the present invention.

[0018] Figure 4 It is a schematic diagram of the feeding method involved in the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] As attached Figure 1 As shown, the present invention involves a precision feeding system based on multi-source heterogeneous data drive and artificial intelligence, which includes a control unit 1, a data acquisition unit 2, a feeding unit 3, and a feeding monitoring unit 4.

[0021] The control unit 1 is electrically connected to the data acquisition unit 2, the feeding unit 3, and the eating monitoring unit 4, and is responsible for processing data and controlling the working status of each unit. An embedded processor or a processing device such as a PC can be selected.

[0022] The data acquisition unit 2 includes an environmental parameter acquisition module 21 and a water quality parameter acquisition module 22. The environmental acquisition module is used to collect weather data, including but not limited to air temperature, air humidity, air pressure, solar illumination, and wind speed. The water quality parameter acquisition module collects relevant data in the water body, including but not limited to water temperature, dissolved oxygen, ammonia nitrogen content, nitrite content, turbidity, transparency, and pH value.

[0023] The specific environmental parameter acquisition module 21 and the water quality parameter acquisition module 22 can select mature products currently on the market. For example, the environmental parameter acquisition module 21 includes a temperature and humidity sensor, an air pressure sensor, an illuminance sensor, an anemometer, etc.; the water quality parameter acquisition module 22 includes a temperature sensor, a dissolved oxygen sensor, an ammonia nitrogen content sensor, a nitrite content sensor, a turbidity sensor, a pH sensor, etc., or you can directly select the currently mature portable meteorological instruments and water quality multi-parameter data acquisition equipment to obtain relevant environmental parameters and water quality parameters.

[0024] The control unit 1 determines the appropriate feeding amount under the current environmental parameters and water quality parameters based on the environmental parameters, water quality parameters and preset fish feeding regularity data obtained by the data acquisition unit.

[0025] Fish feeding rules generally target the feeding requirements of different fish species at different growth stages, as well as the impact of environmental parameters and water quality parameters on feeding amounts. Fish feeding rules are existing knowledge in the field, generally including the feeding requirements of a specific fish species at different growth stages, different times (seasons), and different weights, and the impact of environmental parameters such as air pressure and temperature, and water quality parameters such as water temperature on feeding amounts.

[0026] The feeding unit 3 performs timed and quantitative feeding according to the control unit 1 , and the feeding unit 3 may be a currently commonly used automatic feeding machine.

[0027] Furthermore, as attached Figure 2 As shown, in one embodiment of the present application, a screw is used to more accurately control the feeding amount. Specifically, the feeding unit 3 includes a silo 31, a feeding screw assembly 32, and a feeding assembly 33. A discharge port 311 is provided at the bottom of the silo 31. The feeding screw assembly 32 includes a shell 321, a central screw 322, and a driving motor 323. The shell 321 is tubular and is arranged below the discharge port 311 of the silo 31. The discharge port 311 is connected to the inside of the shell 321 through a connecting pipe. The central screw 322 is arranged in the shell 321. A quantitative discharge port 3211 is provided at the lower part of the shell 321 away from the discharge port 311. One end of the central screw 322 passes through the side wall of the shell 321 and is connected to the driving connection 323. The feeding assembly 33 can be a common throwing plate or a pneumatic feeding device. In one embodiment of the present application shown in the accompanying drawings, a pneumatic feeding device is used, specifically a turbine fan blows the quantitative discharge port 3211 to the discharge port in the fish pond.

[0028] The feeding monitoring unit 4 is used to detect the feeding situation after feeding, evaluate the remaining bait situation by regional sampling, and the control unit controls the feeding process according to the remaining bait situation. Specifically, the feeding monitoring unit 4 includes a smart camera 41, which is a smart camera integrated with an embedded machine vision processor and is required to have good waterproof performance.

[0029] There are many patent applications for using cameras to detect fish feeding conditions, such as patent CN118120683B A machine vision-based intelligent feeding system and method for fish pond farming. The patent discloses analyzing and processing image data based on an improved artificial bee colony-support vector machine algorithm, judging the number of fish feeding in the fish pond, and comparing it with the fish number data in the database, so as to accurately judge the number of fish feeding in the fish pond, and compare the feeding amount data in the database to realize the judgment of the feeding amount. This type of method of monitoring the feeding situation of fish schools through machine vision has certain problems in practice. In fish pond farming, due to clustering behavior, feeding habits and other reasons, some fish will follow the eating fish schools even if they have no appetite, but their food intake is obviously insufficient, and it is often difficult to obtain a more accurate effect from external behavior alone.

[0030] As attached Figure 3 As shown, in one embodiment of the present application, the method of sampling remaining bait monitoring is used to evaluate the feeding situation of fish. The feeding monitoring unit 4 includes an intelligent camera 41, an underwater identification frame 42, an infrared fill light 43, and a connecting rod 44.

[0031] The underwater identification frame 42 is a rectangular frame, and adopts a bright color that can be distinguished from the water body. Its size can be adaptively set according to the size of the fish pond. The underwater identification frame 42 is connected to the connecting rod 44 through a connecting crossbar 421. The smart camera 41 is set on the connecting rod 44 at the lower part of the underwater identification frame 42, and its viewing direction is obliquely upward, and the underwater identification frame 42 is within its viewing range. The connecting rod 44 is fixed to the bottom of the fish pond or the horizontal support rod at the upper part of the fish pond.

[0032] The underwater identification frame 42 is used to determine the sampling area, which is convenient for subsequent image processing, and is generally a water body at a preset distance below the rectangular frame. For example, if the size of the underwater identification frame 42 is 1m×1m, the sampling area is a cube of 1m×1m×1m below the underwater identification frame 42, or it can be a parallelepiped.

[0033] The smart camera 41 evaluates the remaining bait, i.e., the remaining feed, in the sampling area through machine vision. The sampling area can be flexibly set according to the fish pond environment, the location of the feeding unit 3, and the eating habits of the fish.

[0034] As attached Figure 4As shown, the feeding method based on the above-mentioned precise feeding system includes the following steps: step S1, the environmental parameters and water quality parameters collected by the data acquisition unit; the control unit presets the fish feeding law; step S2, the control unit determines the appropriate feeding amount based on the environmental parameters, water quality parameters and preset fish feeding law data obtained by the data acquisition unit; step S3, the control unit controls the feeding unit to feed according to 60% to 70% of the feeding amount obtained in step S2 as the basic feeding amount, and monitors the fish feeding situation through the feeding monitoring unit after the preset feeding time, determines the variable feeding amount according to the feeding situation, and controls the feeding unit to feed twice. The variable feeding amount is 50% to 40% of the feeding amount obtained in step S2, and the sum of the basic feeding amount and the variable feeding amount should be 80% to 110% of the feeding amount obtained in step S2.

[0035] Specifically, in step S3, "after feeding, monitoring the feeding situation of the fish through the feeding monitoring unit, and determining the change of feeding amount according to the feeding situation" specifically includes the following steps: Step S31: construct a residual bait recognition model, which is based on the YOLO model and specifically includes the following steps: S311, data collection: collecting underwater residual bait images or video data for training and verifying residual bait recognition models; S312, data preprocessing: preprocessing the collected image or video data; manually annotating the preprocessed data and dividing it into a training data set and a verification data set; Specifically, in an example of the present application, the analysis object of the residual bait recognition model is a video stream, and the output structure is the number of residual baits. The specific analysis object is a video stream of 5 to 15 seconds. The position of the residual bait in the video will change, which is conducive to improving the recognition accuracy. In order to improve the data processing efficiency, key frames are intercepted from the video stream, such as the start frame, the middle frame, the end frame or the frames are taken at uniform intervals.

[0036] The key frames in the collected video stream are preprocessed, and the preprocessing in this application includes image cropping, CLAHE enhancement (contrast limited adaptive histogram equalization), and Mosaic data enhancement. The image cropping is based on the underwater identification frame 42 in the key frame image and is cropped according to the size and position of the preset sampling area in the image.

[0037] The specific image cropping is performed using the OpenCV library, which automatically detects the position of the underwater identification frame 42 and determines the cropping area based on the position of the underwater identification frame 42 and the preset rules (the size and proportion of the fixed shape corresponding to the sampling area). Since the underwater identification frame 42 has a regular shape and a special color, it is easier and more accurate to identify.

[0038] During data labeling, in one embodiment of the present application, due to limited collected data, all key frames in the same video stream are labeled.

[0039] S313, model construction and training, construct a leftover bait recognition model based on the YOLO model, use the labeled training data set to train the model, and optimize the model performance by adjusting the model parameters. Then use the validation data set to evaluate the accuracy of the leftover bait recognition model, and optimize it according to the evaluation results to obtain the leftover bait recognition model.

[0040] The model includes a key frame extraction and preprocessing module, a time series fusion module, and a YOLO detection module. The key frame data of the video stream is processed by the time series fusion module and then input into the YOLO detection module to output the number of remaining baits.

[0041] More specifically, the key frame extraction module takes frames at uniform intervals, and performs image cropping, CLAHE enhancement, and Mosaic data enhancement on the extracted key frames.

[0042] The time series fusion module is an LSTM time series fusion module. The key frame extraction module obtains the key frame input to the LSTM layer to extract the time series features. The output of the LSTM layer is sent to the YOLO detection module for target detection. To improve processing efficiency, the LSTM time series fusion module includes a layer of LSTM.

[0043] In the YOLO detection module, CSPDarknet53 is used as the backbone network for feature extraction. The SPPF module and PAN structure are used for feature fusion, and the prediction layer is used to generate the final detection result.

[0044] The model workflow is as follows: 1. Video stream input: The model receives video stream input. 2. Key frame extraction: Extract key frames from the video stream by taking frames at uniform intervals. 3. Temporal fusion: Input the extracted key frames into the LSTM temporal fusion module to extract temporal features. 4. Feature extraction and fusion: The output of LSTM enters the YOLO detection module, where CSPDarknet53 is responsible for feature extraction, and SPPF and PAN structures perform feature fusion. 5. Object detection: The prediction layer generates the final detection results, including bounding boxes and category probabilities. 6. Output: The system outputs the number of residual bait detected in each frame, and accumulates these numbers to obtain the total number of residual bait in the video.

[0045] Step S32, baseline establishment, generalizing the model in the actual use environment, and obtaining the remaining bait situation data in the sampling area under normal feeding conditions; specifically includes the following steps: Step S321, model generalization: collect underwater bait residue images or video data in actual use environment; pre-process the collected image or video data; manually annotate the pre-processed data and divide it into generalized training data set and generalized verification data set, and obtain the bait residue recognition application model after generalized training. This process mainly solves the influence of external environment such as water quality and light on the model.

[0046] Step S322, baseline establishment: Determine the baseline data for residual bait evaluation through the residual bait situation data in the sampling area under normal feeding conditions. The specific process is to obtain the underwater residual bait image or video data in the sampling area under normal feeding conditions, and obtain the baseline data through the residual bait recognition application model.

[0047] Step S33, residual bait situation evaluation, using the residual bait recognition application model combined with the baseline data to evaluate the fish feeding situation; specifically includes the following steps: Step S331, real-time data collection, after the preset feeding time, using the residual bait recognition application model trained in step S32 to obtain the residual bait data of the sampling area in real time analysis; Step S332, data comparison and analysis: compare and analyze the remaining bait data obtained in step S331 with the baseline data established in step S32 to evaluate the fish's eating habits; determine the variable feeding amount based on the comparison and analysis results. If the remaining bait amount is higher than the baseline data, reduce the variable feeding amount so that the total feeding amount is less than the feeding amount obtained in step S2; if the remaining bait amount is lower than the baseline data, increase the variable feeding amount so that the total feeding amount is greater than the feeding amount obtained in step S2.

[0048] Different rules can be set according to the specific situation of fish, and a gradient setting can be made for the variable feeding amount. For example, if the remaining bait data are 70% to 80%, 80% to 90%, 90% to 110%, 110% to 120%, and above 150% of the baseline data, and the corresponding total feeding amount is 110%, 105%, 100%, and 90% of the feeding amount obtained in step S2, the feeding is stopped and a signal is issued to require human intervention.

[0049] It should be noted that, unless otherwise clearly specified and limited, the terms such as "install", "connect", "connect", "fix", "set" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

Claims

1. A precision feeding system based on multi-source heterogeneous data drive and artificial intelligence, characterized in that: The precision feeding system includes a control unit, a data acquisition unit, a feeding unit, and a feeding monitoring unit; the control unit is electrically connected to the data acquisition unit, the feeding unit, and the feeding monitoring unit; the data acquisition unit includes an environmental parameter acquisition module and a water quality parameter acquisition module; the feeding unit is an automatic feeding machine; the feeding monitoring unit includes an intelligent camera and an underwater identification frame.

2. The precision feeding system according to claim 1, characterized in that: The feeding monitoring unit also includes an infrared fill light and a connecting rod; the underwater identification frame is a rectangular frame, which is connected to the connecting rod through a connecting cross rod. The smart camera is arranged on the connecting rod at the lower part of the underwater identification frame, and its viewing direction is obliquely upward, and the underwater identification frame is within its viewing range.

3. The precision feeding system according to claim 1, characterized in that: The feeding unit includes a silo, a feeding screw assembly, and a feeding assembly. A discharge port is provided at the bottom of the silo. The feeding screw assembly includes an outer shell, a central screw, and a driving motor. The outer shell is tubular and is provided below the discharge port of the silo. The discharge port is connected to the interior of the outer shell through a connecting pipe. The central screw is provided in the outer shell. A quantitative discharge port is provided at the lower part of the outer shell away from the discharge port. One end of the central screw passes through the side wall of the outer shell and is connected to the driving motor. The feeding assembly is a throwing tray or a pneumatic feeding device.

4. A feeding method based on the precision feeding system according to any one of claims 1 to 3, characterized in that: The method comprises the following steps: step S1, a data acquisition unit acquires environmental parameters and water quality parameters; a control unit presets a fish feeding rule; step S2, a control unit determines a suitable feeding amount based on the environmental parameters and water quality parameters acquired by the data acquisition unit and preset fish feeding rule data; step S3, a control unit controls a feeding unit to feed fish according to 60% to 70% of the feeding amount acquired in step S2 as a basic feeding amount, monitors the feeding situation of fish through a feeding monitoring unit after a preset feeding time, determines a variable feeding amount according to the feeding situation, and controls the feeding unit to perform secondary feeding.

5. The feeding method according to claim 4, characterized in that: The step S3 of "monitoring the feeding situation of fish through the feeding monitoring unit after feeding, and determining the change of feeding amount according to the feeding situation" specifically includes the following steps: step S31, constructing a residual bait recognition model, constructing a residual bait recognition model based on the YOLO model; step S32, establishing a baseline, generalizing the model in the actual use environment, and obtaining the residual bait situation data in the sampling area under normal feeding conditions; step S33, evaluating the residual bait situation, using the residual bait recognition application model combined with the baseline data to evaluate the fish feeding situation.

6. The feeding method according to claim 5, characterized in that: The step S31 includes the following steps: S311, data collection: collecting underwater residual bait images or video data for training and verifying the residual bait recognition model; S312, data preprocessing: preprocessing the collected image or video data; The preprocessed data is manually labeled and divided into a training data set and a validation data set; S313, model construction and training, constructing a leftover bait recognition model, using the labeled training data set to train the model, optimizing the model performance by adjusting the model parameters, and then using the validation data set to evaluate the accuracy of the leftover bait recognition model, and optimizing the leftover bait recognition model based on the evaluation results.

7. The feeding method according to claim 6, characterized in that: The data collection object in step S31 is a video stream, and the data preprocessing includes key frame extraction, image cropping, CLAHE enhancement, and Mosaic data enhancement; The residual bait recognition model includes a key frame extraction and preprocessing module, a time series fusion module, and a YOLO detection module. After being processed by the time series fusion module, the data is input into the YOLO detection module, and the number of residual baits is output.

8. The feeding method according to claim 6, characterized in that: The image cropping is based on the underwater identification frame in the key frame image and is cropped according to the size and position of the preset sampling area in the image.

9. The feeding method according to claim 6, characterized in that: The step S32 includes the following steps: step S321, model generalization: collecting underwater residual bait images or video data in the actual use environment; preprocessing the collected image or video data; manually annotating the preprocessed data, and dividing it into a generalized training data set and a generalized verification data set, and obtaining a residual bait recognition application model after generalization training; step S322, baseline establishment: determining the baseline data for residual bait evaluation through the residual bait situation data in the sampling area under normal feeding conditions.

10. The feeding method according to claim 6, characterized in that: The step S33 includes the following steps: step S331, real-time data collection, after the preset feeding time, use the residual bait recognition application model trained in step S32 to obtain the residual bait data of the sampling area through real-time analysis; step S332, data comparison and analysis: compare and analyze the residual bait data obtained in step S331 with the baseline data established in step S32 to evaluate the eating conditions of the fish; determine the change in feeding amount based on the comparison and analysis results.

Images