CN-122022276-A - Prawn breeding and feeding decision-making system based on multi-mode information fusion and AI intelligent agent

CN122022276ACN 122022276 ACN122022276 ACN 122022276ACN-122022276-A

Abstract

The invention discloses a prawn culture feeding decision-making system based on multi-mode information fusion and AI intelligent agent, relating to the technical field of intelligent aquaculture and artificial intelligence, which comprises an automatic material platform device, the system comprises a camera acquisition and visual analysis module, a prawn weight estimation module, a residual material identification and description module, a breeding production data analysis and feeding quantity prediction module, a feeding decision agent module and a man-machine interaction and execution module. The system can be used for realizing low-stress, high-precision automatic monitoring and accurate intelligent feeding decision deployed in a real commercial factory prawn culture workshop, realizing prawn weight estimation, residue sensing and daily feeding quantity intelligent decision under complex, high-density and turbid water conditions, being beneficial to improving the utilization rate of feed, reducing the water quality pollution and the labor cost and promoting the transformation of prawn culture from experience driving to data and intelligent driving.

Inventors

ZHAO HAOCHENG
WANG BAOJIE
LIU MEI
WANG LEI
JIANG KEYONG
WANG QI

Assignees

中国科学院海洋研究所

Dates

Publication Date: 20260512
Application Date: 20260109

Claims (8)

1. Shrimp culture feeding decision-making system based on multimode information fusion and AI intelligent agent, which is characterized by comprising: the automatic material platform device is used for bearing feed/prawns in the culture pond and realizing autonomous lifting; The shooting acquisition and visual analysis module is used for acquiring images of the prawns and the residual materials when the material platform is lifted to a preset water leaving height, and carrying out image analysis based on segmentation, classification and visual large models; The prawn weight estimation module is used for establishing a non-contact weight estimation model based on the projection area and the visual overall length of the prawn image; the residue identification and description module is used for judging whether the residue exists on the material table or not, and generating qualitative description by utilizing the visual large model when the residue exists; The breeding production data analysis and feeding amount prediction module is used for constructing a feeding amount prediction model based on multi-workshop multi-dimensional production data of the water body covering the supersquare; the feeding decision-making agent module is used for fusing multi-source information such as a data driving prediction result, an empirical formula, a visual analysis result, and remarks of breeding staff, and outputting a structured feeding suggestion under the reasoning of a large language model; And the man-machine interaction and execution module receives real-time culture data and user remarks through a Web interface or an application program interface, visually displays decision suggestions and compares the decision suggestions with actual feeding execution conditions for feedback, so that the dynamic adjustment of feeding strategies is realized.
2. The multi-modal information fusion and AI agent-based prawn culture feeding decision system as claimed in claim 1, wherein: The automatic material platform specifically comprises: The material table bearing structure is built by adopting a corrosion-resistant metal material, and is provided with a material table net disc with an edge baffle plate for bearing feed and prawns; The direct-current gear motor is connected with the material platform through a mechanical structure and driven by the control unit to realize the lifting of the material platform in the water body; The limiting device is arranged at the end point of the lifting stroke of the material table and used for preventing the structure from being damaged; An industrial camera unit which is arranged above the material table and is provided with a high-pixel imaging sensor and a dynamic wide mode; the lifting and collecting control unit is used for controlling the lifting of the material platform and triggering image collection; And And the AI interaction module integrates voice interaction and vision analysis, analyzes the instruction by utilizing a voice recognition model, analyzes the image by generating an artificial intelligent vision model, and feeds back the result by utilizing the voice synthesis model to realize man-machine interaction.
3. The multi-modal information fusion and AI agent-based prawn culture feeding decision system as claimed in claim 2, wherein: The visual analysis module and the weight estimation module are used for carrying out object detection and mask segmentation on the collected material table image by adopting a deep learning example segmentation model, and distinguishing at least two types of gestures of 'side shrimp' and 'back side shrimp'; calculating the projection area of the prawns based on the segmentation mask, and extracting the full-length visual field of the prawns from the mask by utilizing a skeleton refinement and principal axis extraction algorithm; adopting a regression model based on machine learning, taking the full visual length and the projection area as input features, realizing non-contact estimation of the weight of the prawns, and enabling an estimation result to be called by a subsequent module; And when the image is identified as the residue, calling a visual large model to analyze the image, and outputting natural language description comprising the residue position, the region size, the accumulation state, the relative severity and the water environment details.
4. The multi-modal information fusion and AI agent-based prawn culture feeding decision system as claimed in claim 3, wherein: The culture production data analysis module is used for carrying out missing value processing, abnormal value elimination and consistency verification on the obtained daily production record to obtain an effective daily production record, and carrying out scaling processing on input characteristics and target variables respectively by adopting a normalization method to divide a training set, a verification set and a test set for model training and evaluation, wherein the record comprises date, system ID, water temperature, dissolved oxygen, pH, daily water change amount, daily water change rate, ammonia nitrogen, nitrite, culture day age, weight, shrimp carrying amount, water volume, survival rate, average daily weight gain, daily meal number and daily feed amount.
5. The multi-modal information fusion and AI agent-based prawn culture feeding decision system as claimed in claim 4, wherein: The feeding quantity prediction module comprises: The model comparison unit is used for training and evaluating multiple regression models and time sequence models; The performance evaluation unit is used for comparing the performance of the model by adopting an error analysis index; The characteristic correlation analysis unit calculates the correlation between the input characteristic and the daily feeding quantity and identifies a high correlation characteristic subset; the model screening and optimizing unit selects an integrated tree model with optimal performance based on the evaluation result, and adopts a meta heuristic optimization algorithm to perform super-parameter search; And And the core model unit adopts a Bayesian optimization algorithm to carry out iterative optimization on the super parameters of the integrated tree model, and a feeding prediction core model is constructed.
6. The multi-modal information fusion and AI agent-based prawn culture feeding decision system as claimed in claim 5, wherein: the feeding decision agent module comprises: the data layer stores historical production records, model prediction results and feeding execution results by adopting a database, and configures a knowledge base formed by culture technical specifications and experience rules; The model layer integrates a Bayesian optimized core feeding prediction model and an experience feeding formula based on pond storage quantity, water body volume and prawn weight to form a double prediction mechanism; The decision layer calls a large language model, takes a core model prediction result, an empirical formula prediction result, a historical record fragment, a culture knowledge base entry and natural language remarks from a visual analysis module and a user as inputs, analyzes the difference and rationality of the double prediction result, and combines multi-modal information to generate a feeding suggestion; And And the report generating unit is used for sorting the result output by the decision layer into a structured decision report, wherein the structured decision report comprises a model predicted value, an experience predicted value, a final proposal feeding amount, an adjustment reason and an operation proposal.
7. The multi-modal information fusion and AI agent-based prawn culture feeding decision system as claimed in claim 6, wherein: the man-machine interaction and execution module comprises: the application programming interface is used for receiving real-time water quality indexes, biological parameters and user remark information from a workshop and sending the real-time water quality indexes, the biological parameters and the user remark information to the prediction and decision module; The Web front-end interface is used for displaying the environment and production data, model prediction results, residual material detection states, semantic descriptions generated by the large visual model and structured feeding decision reports; And And the interactive feedback unit is used for recording the adoption condition of the intelligent suggestion by the breeding personnel and the actual execution feeding amount and providing data support for the subsequent model iteration.
8. The multi-modal information fusion and AI agent-based prawn culture feeding decision system as claimed in claim 7, wherein: the accurate intelligent feeding decision method for the prawns, which is executed by the system, comprises the following steps: collecting the current day water quality parameter, the culture state parameter and the biological parameter from an industrial prawn culture workshop, and adopting the estimated weight output by the visual analysis module to the weight data in the biological parameter; Normalizing the parameters, and inputting the feature vectors into a core prediction model optimized by Bayes to obtain a first prediction value Q_ml; Estimating the number of prawns in the culture pond based on the weight, the shrimp carrying amount and the water volume, and calculating a second predicted value Q_rule by combining a preset unit weight feeding coefficient; storing the Q_ml and the Q_rule together with the history record into a data layer as candidate feeding quantity for the feeding decision agent to call; based on the first predicted value Q_ml, the second predicted value Q_rule and the corresponding history record, receiving a natural language remark R input by a user; if R is empty, generating recommended daily feed amount and brief description based on Q_ml; If R is not empty, inputting Q_ml, Q_rule, history record fragments, entries related to R in a culture knowledge base, residuals from a visual analysis module and environment description into a large language model together to obtain a comprehensive reasoning result comprising difference analysis, risk assessment and adjustment logic; Extracting a final recommended feeding quantity Q_final and a floating interval thereof from the comprehensive reasoning result, and comparing and displaying the final recommended feeding quantity Q_final and the floating interval with Q_ml and Q_rule to form a structured decision report; Record whether the user adopts Q final and the actual feeding amount performed, for closed loop optimization of the system.

Description

Prawn breeding and feeding decision-making system based on multi-mode information fusion and AI intelligent agent Technical Field The invention relates to the technical field of intelligent aquaculture and artificial intelligence, in particular to a prawn culture feeding decision-making system based on multi-mode information fusion and AI (advanced technology interface) intelligent bodies. Background Litopenaeus vannamei is one of the currently globally important aquaculture varieties, and the industrial high-density culture mode has obvious advantages in the aspects of stable supply, land saving, unit water yield improvement and the like. However, in a closed or semi-closed industrial culture system, the shrimp growth environment is complex, the turbidity of a water body is high, the shrimp carrying capacity of a single pond is large, and the traditional feeding management mode which depends on manual experience has the outstanding problems that 1, in the actual production at present, the shrimp weight is usually obtained by means of regular sampling, weighing and manual estimation, the frequency is limited, the operation has stress risks, the data timeliness and the representativeness are insufficient, the whole population weight level on the current day or the current moment is difficult to accurately reflect, 2, the residue monitoring depends on manual observation, the common material platform management mainly comprises the steps of pulling the material platform out of water at a fixed time point by culture personnel, judging whether the residue exists and how much of the residue exists through visual inspection, and adjusting the feeding amount on the next meal or the next day by experience. The method is strong in subjectivity and is greatly influenced by human experience and energy, and 3, daily feeding quantity decision is based on experience, namely, in a plurality of industrial shrimp farms, daily feeding quantity is roughly calculated through a mode of 'weight estimation, shrimp carrying quantity estimation and feed coefficient experience table', and is corrected by assistance of recent residue conditions. Because of lack of systematic utilization of historical production data and environmental data for modeling analysis, the above experience method is difficult to adapt to the condition differences of different workshops, water body environments, offspring seed batches and the like, and easily causes overfeeding or insufficient feeding to influence the utilization rate of feed, water quality stability and prawn health, and 4. The data utilization degree is insufficient and lacks closed loops, the factory culture often has a certain degree of data record, such as water temperature, dissolved oxygen, pH, ammonia nitrogen, nitrite, feeding amount and other daily degree data, but is scattered in a form of table, and systematic data analysis and intelligent decision capability are not formed. Meanwhile, automatic association and feedback are not available between the feeding execution result and the subsequent growth and water quality change, and an intelligent closed loop of feeding-response-adjustment cannot be formed. In summary, in the aspect of intelligent management of industrial prawn culture, the prior art has not yet lacked a comprehensive solution that takes into account non-contact estimation of prawn weight, automatic identification of residual materials, multi-dimensional production data modeling and intelligent accurate feeding decision on the premise of not significantly changing the existing production habit and reducing prawn stress as much as possible. Disclosure of Invention In order to solve the problems, the invention aims to provide a multi-mode information fusion and AI intelligent object-based prawn culture feeding decision system, which aims to realize low-stress accurate intelligent feeding for industrial culture of litopenaeus vannamei. In order to achieve the technical aim, the application provides a prawn culture feeding decision system based on multi-mode information fusion and AI intelligent agent, comprising: the automatic material platform device is used for bearing feed/prawns in the culture pond and realizing autonomous lifting; The shooting acquisition and visual analysis module is used for acquiring images of the prawns and the residual materials when the material platform is lifted to a preset water leaving height, and carrying out image analysis based on segmentation, classification and visual large models; The prawn weight estimation module is used for establishing a non-contact weight estimation model based on the projection area and the visual overall length of the prawn image; the residue identification and description module is used for judging whether the residue exists on the material table or not, and generating qualitative description by utilizing the visual large model when the residue exists; The breeding production data analysis and feeding amount p