Multi-modal large model-based sheldrake breeding method and system

Abstract

The present application relates to the field of muscovy duck breeding, and particularly relates to a muscovy duck breeding method and system based on a multi-modal large model. The method comprises the following steps: collecting multi-modal original data in the process of muscovy duck breeding; pre-processing the collected multi-modal original data; constructing a TwinsYOLO detection model, and training the TwinsYOLO detection model through the pre-processed visual data to output muscovy duck detection results; constructing a Conformer speech processing model, and training the Conformer speech processing model through the pre-processed audio data; constructing a V-Mallard multi-modal interactive decision model, and integrating the pre-processed visual, audio and environmental data and the outputs of the TwinsYOLO detection model and the Conformer speech processing model to generate breeding decisions. The present application is suitable for intelligent breeding of muscovy ducks.

Description

Technical Field

[0001] This invention relates to the field of Muscovy duck farming, specifically to a method and system for Muscovy duck farming based on a multimodal large model. Background Technology

[0002] The current duck farming industry still faces many prominent problems: In the traditional farming model, environmental control relies on manual experience, and key parameters such as temperature, humidity, and ammonia concentration fluctuate frequently, which can easily lead to group diseases; disease monitoring mainly relies on manual inspection, and it is difficult to identify minor lesions on the surface of sick ducks (such as feather damage caused by Salmonella infection) and abnormal behaviors in a timely manner, resulting in a high rate of missed detection and delays in treatment; farmers generally lack professional knowledge and find it difficult to make scientific decisions when faced with abnormal farming conditions. In remote areas, geographical limitations prevent them from obtaining on-site guidance from experts, forming a vicious cycle of extensive environmental control, high incidence of diseases, and inefficient response.

[0003] While some smart farming tools exist in the current technology, they have significant limitations: target detection technology is insufficient in recognizing highly similar individuals under conditions of dense duck flocks and complex lighting, making it difficult to accurately locate sick ducks; speech recognition technology has low accuracy in extracting human voices in noisy environments (duck calls, machinery noise, etc.), failing to guarantee reliable interaction; multimodal models mostly focus on image-text single-modal fusion, lacking full-modal integration of visual, auditory, and environmental data in farming scenarios, and the interaction mode requires a wake word to trigger and cannot interrupt the response, making it difficult to adapt to the needs of efficient operation in farming sites. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for raising Muscovy ducks based on a multimodal large model, thereby improving the intelligence capability of Muscovy duck farming.

[0005] The present invention achieves the above-mentioned objectives by adopting the following technical solution: Firstly, the present invention provides a method for raising Muscovy ducks based on a multimodal large model, comprising:

[0006] S1. Collect multimodal raw data during the duck farming process, including visual data, audio data, and environmental data;

[0007] S2. Preprocess the collected multimodal raw data, including cleaning, enhancement and standardization.

[0008] S3. Construct a TwinsYOLO detection model and train it using preprocessed visual data;

[0009] S4. Construct the Conformer speech processing model and train it using preprocessed audio data.

[0010] S5. Construct a V-Mallard multimodal interactive decision-making model and integrate the preprocessed visual, audio, and environmental data with the outputs of the TwinsYOLO detection model and the Conformer speech processing model to generate aquaculture decisions.

[0011] Furthermore, in step S1, visual data is collected through cameras. The cameras are deployed in the breeding area, feeding area and drinking area of ​​the duck farm to collect images and videos of ducks walking, eating, flapping their wings and resting. During the collection process, the shooting angle, distance and lighting conditions are adjusted to cover different background differences, light intensity and imaging angles to ensure the randomness and diversity of sample distribution.

[0012] Audio data is collected through an array of microphones, simultaneously capturing audio signals from human voice commands, duck calls, and mechanical equipment noise in the farm, and capturing changes in voiceprint characteristics;

[0013] Environmental data is collected in real time from the temperature, humidity and ammonia concentration in the duck farming environment using temperature and humidity sensors and ammonia concentration sensors.

[0014] Furthermore, in step S2, for visual data, firstly, redundant data with high repetition and those that were not fully captured due to being obscured by farm buildings are filtered and discarded. Then, dynamic random cropping, Gaussian blur, four-way flipping, random rotation, RandomColor data augmentation, and Mosaic methods are used to expand the diversity of the visual dataset. At the same time, illumination correction processing is performed to correct the color difference of the image under complex lighting conditions and improve the image clarity.

[0015] For audio data, noise reduction algorithms are used to remove environmental noise interference, and audio format conversion and preliminary feature extraction are performed.

[0016] For environmental data, perform missing value imputation and outlier correction.

[0017] Finally, all preprocessed data of different types are standardized to unify data format and dimensions.

[0018] Furthermore, step S3 specifically includes:

[0019] The TwinsYOLO detection model includes a feature extraction unit, an individual identification unit, an abnormal behavior capture unit, and a lesion identification unit;

[0020] The feature extraction unit adopts the YOLOv11Backbone architecture with residual networks, which includes an improved C2PSA module and a C3k2 conditional branch structure. It consists of multi-level convolutional layers, an adaptive feature enhancement module, and cross-layer residual connections, and is used to perform hierarchical modeling of the local detail features and overall structural features of the Muscovy duck target.

[0021] The individual identification unit is a twin neural network, which contains two convolutional network branches with identical structures and shared weights;

[0022] By embedding two convolutional network branches with identical structures and shared weights into a twin neural network, feature vectors of the input image are extracted, the L1 distance between the vectors is calculated and weighted, and a similarity score is output through the Sigmoid function to distinguish highly similar Muscovy duck individuals.

[0023] The abnormal behavior capture unit includes a cross-scale feature fusion subunit and a spatial pyramid pooling subunit, which are used to jointly analyze the morphological changes and behavioral states of individual Muscovy ducks at different spatial scales.

[0024] Feature maps are obtained from different levels of the backbone network through cross-scale feature fusion subunits. The shallow feature maps have high spatial resolution and are used to characterize the feather arrangement, local texture continuity and limb edge changes of Muscovy ducks. The deep feature maps have strong semantic expression capabilities and are used to characterize the overall posture structure, movement trend and behavioral stability of Muscovy ducks.

[0025] By aligning the size of feature maps at different levels and fusing them step by step, shallow local detail features and deep global behavioral features can be jointly expressed in the same feature space.

[0026] By setting multiple pooling windows of different scales on the fused feature map through spatial pyramid pooling sub-units, the feature map is partitioned and aggregated to obtain contextual information in local areas, medium ranges and overall ranges respectively, so that the model can analyze the behavioral state of individual ducks in different spatial sensing fields.

[0027] The lesion recognition unit includes a channel attention mechanism subunit;

[0028] The channel attention mechanism suppresses background noise and enhances the ability to identify small lesions on the surface of ducks. It outputs multi-scale prediction results through three parallel detection heads. The classification branch optimizes efficiency through depthwise separable convolution, and the regression branch improves the loss function to enhance positioning accuracy.

[0029] Furthermore, step S4 specifically includes:

[0030] The Conformer speech processing model includes a convolutional subsampling unit, a linear layer, a Dropout layer, multiple ConformerBlock stacked structures, and a voiceprint anomaly detection unit.

[0031] The input audio signal is convolutionally sampled by a convolutional subsampling unit, which reduces the length of the input sequence while preserving feature information.

[0032] The sampled audio data is subjected to linear transformation and random deactivation through linear layers and Dropout layers;

[0033] The ConformerBlock stacked structure includes a first feedforward network, a self-attention module, a convolution module, and a second feedforward network. The first feedforward network performs a nonlinear transformation on the input data. The self-attention module adopts the relative sinusoidal position coding scheme in Transformer-XL and captures the global dependencies of the audio sequence through a multi-head self-attention mechanism.

[0034] The convolution module adopts a depthwise separable convolution structure. First, it performs spatial filtering on each input channel separately, then introduces nonlinear transformation through the Swish activation function, and finally mixes information across channels through point convolution with 1×1 convolution kernels. While reducing the number of parameters and computation, it accurately captures local features and distinguishes human voice from environmental noise.

[0035] The second feedforward network has a similar structure to the first feedforward network but does not share parameters. It uses the Swish activation function to further transform and enhance the feature representation.

[0036] The voiceprint anomaly monitoring unit performs real-time analysis of the extracted human voice features and voiceprint data to identify abnormal voiceprint changes, providing reliable voice input for subsequent interactions.

[0037] Furthermore, step S5 specifically includes:

[0038] The V-Mallard multimodal interactive decision model includes a multimodal encoder, a large model inference unit, an interactive control unit, and a decision suggestion generation unit.

[0039] The multimodal encoder includes a visual encoder, an audio encoder, and a text encoder, which extract features from the preprocessed visual data, audio data, and text data input by farmers, respectively. By accumulating a large amount of multimodal data, the text feature space is aligned with the feature spaces of video, image, and audio. A status token is introduced to identify the input query type, thereby realizing end-to-end interaction.

[0040] The large model inference unit expands the Chinese vocabulary based on the Mixtral8x7B model, optimizes instructions through a bilingual text corpus, and has the ability to understand both Chinese and English. At the same time, it incorporates professional knowledge related to Muscovy duck farming and builds an expert knowledge base in the field of farming. The professional knowledge related to Muscovy duck farming includes disease prevention and control, environmental control, feed formulation and duck farm layout.

[0041] The large model inference unit adopts the SwitchTransformers architecture, replacing the traditional denseFFN with a sparse expert network in each feed-forward layer. All tokens are allocated to only one expert network through the router, simplifying the routing, allocation and merging process and reducing computing power requirements.

[0042] The interactive control unit adopts a duplex operation scheme, with two V-Mallard models working simultaneously. One model is responsible for generating the response to the current query, while the other continuously monitors the environmental input. When a new voice or text query is detected, the current generation process is interrupted, and the response to the new query is output first. It supports non-wake-up dialogue and interrupted command interaction, without the need for preset wake words or manual button triggering. It intelligently recognizes human voices and noise to achieve real-time interaction.

[0043] The decision suggestion generation unit combines multimodal feature analysis results with an expert knowledge base to generate corresponding aquaculture decisions based on farmers' queries and abnormal situations detected by the system.

[0044] Secondly, the present invention provides a Muscovy duck farming system based on a multimodal large model, used to implement the aforementioned Muscovy duck farming method based on a multimodal large model. The system includes:

[0045] The data acquisition module is used to collect multimodal raw data during the duck farming process, including visual data, audio data, and environmental data.

[0046] The data preprocessing module is used to clean, enhance, and standardize the collected multimodal raw data;

[0047] The detection module is used to build the TwinsYOLO detection model and train the TwinsYOLO detection model using preprocessed visual data.

[0048] The speech processing module is used to build the Conformer speech processing model and train the Conformer speech processing model using preprocessed audio data.

[0049] The interactive decision-making module is used to build the V-Mallard multimodal interactive decision-making model and integrate the preprocessed visual, audio, and environmental data, as well as the outputs of the TwinsYOLO detection model and the Conformer speech processing model, to generate aquaculture decisions.

[0050] Furthermore, the system also includes an edge computing deployment module, which is used to localize the functions of the detection module, voice processing module, and interactive decision-making module.

[0051] The edge computing deployment module uses lightweight optimization technology to compress each model, ensuring that the model is compatible with the hardware resources of the edge computing device and reducing running latency.

[0052] Furthermore, the system also includes a system management module, which is responsible for user management, parameter configuration, log recording, data backup and recovery functions;

[0053] The user management function of the system management module supports the creation of different role accounts such as farmers, administrators, and technical maintenance personnel, and the allocation of corresponding operation permissions. Farmers can view breeding data, interaction records and decision suggestions, while administrators are responsible for the overall maintenance and parameter adjustment of the system.

[0054] The parameter configuration function allows administrators to customize parameters such as data collection frequency, model detection threshold, and interactive response priority to adapt to the needs of different aquaculture scenarios.

[0055] The logging function is used to record the system's operating status, data collection status, user operation behavior, and model inference results, which facilitates subsequent traceability and problem investigation.

[0056] The data backup and recovery function regularly backs up the raw data, preprocessed data, analysis results, and interaction records collected by the system. When data loss or failure occurs in the system, it is used to quickly recover the data and ensure the stable operation of the system.

[0057] Furthermore, the system also includes a data security module, which is used to ensure the security of system data storage, transmission, and privacy protection.

[0058] The data security module uses the AES encryption algorithm to encrypt and store the collected duck images, videos, farmer voices, and personal information to prevent data leakage.

[0059] The data transmission process uses SSL encryption to ensure that the data is not tampered with or stolen during transmission. An access control mechanism is set up so that only authorized personnel can view the data of the corresponding level, and unauthorized access is restricted. Data security detection and vulnerability remediation are performed according to the set time, as well as security vulnerability scanning, to promptly discover and deal with potential security risks and ensure the security and stability of system data.

[0060] The beneficial effects of this invention are as follows:

[0061] This invention integrates multi-dimensional information such as vision, hearing, and environment through multimodal data acquisition and fusion processing, comprehensively capturing key data in the duck farming process and providing comprehensive support for intelligent decision-making.

[0062] The TwinsYOLO detection model provided by this invention breaks through the limitations of traditional target detection in dense occlusion and complex lighting environments, and realizes high similarity identification of individual ducks and accurate detection of sick ducks, promoting the transformation of disease prevention and control from passive treatment to proactive prevention and control.

[0063] The Conformer speech processing model provided by this invention effectively solves the speech recognition problem in the high-noise environment of a farm, ensuring the reliability of human-computer interaction.

[0064] The V-Mallard multimodal interactive decision-making model provided by this invention supports non-wake-up dialogue and interrupted commands, and achieves millisecond-level response by combining edge computing. It breaks through geographical limitations to provide farmers with "face-to-face" expert guidance, thus solving the industry pain point of scarce expert resources. Attached Figure Description

[0065] Figure 1 This is a flowchart of a method for raising Muscovy ducks based on a multimodal large model provided by the present invention;

[0066] Figure 2 This is a structural block diagram of a Muscovy duck farming system based on a multimodal large model provided by the present invention;

[0067] Figure 3 This is a block diagram of the Conformer speech processing model provided by the present invention;

[0068] Figure 4 This is a block diagram of the ConformerBlock stacked structure provided by the present invention;

[0069] Figure 5 This is a block diagram of the multi-head self-attention mechanism module provided by the present invention;

[0070] Figure 6 This is a block diagram of the convolution module structure provided by the present invention;

[0071] Figure 7This is a block diagram of the feedforward network structure provided by the present invention;

[0072] Figure 8 This is the V-Mallard dual-channel operation flowchart provided by the present invention;

[0073] Figure 9 This is the V-Mallard training flowchart provided by the present invention. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0075] This invention provides a method for raising Muscovy ducks based on a multimodal large model, such as... Figure 1 As shown, it specifically includes:

[0076] S1. Collect multimodal raw data during the duck farming process, including visual data, audio data, and environmental data;

[0077] Visual data acquisition uses high-definition cameras, deployed at key locations in the duck farm's breeding area, feeding area, and drinking area to collect images and videos of various physiological behaviors of the ducks, including walking, feeding, flapping their wings, and resting. During the acquisition process, the shooting angle, distance, and lighting conditions are adjusted to cover different background differences, light intensities, and imaging angles, ensuring the randomness and diversity of the sample distribution and avoiding interference from external variables.

[0078] Audio data acquisition uses an array microphone to simultaneously collect audio signals from human voice commands, duck calls, and mechanical equipment operating noise at the breeding site, capturing changes in voiceprint characteristics;

[0079] Environmental data acquisition uses temperature and humidity sensors and ammonia concentration sensors to collect key parameters such as temperature, humidity and ammonia concentration in the breeding environment in real time, providing environmental basis for breeding decisions.

[0080] All data acquisition devices transmit data via a combination of wired and wireless methods to ensure real-time data upload.

[0081] S2. Preprocess the collected multimodal raw data, including cleaning, enhancement and standardization.

[0082] The collected multimodal raw data is cleaned, enhanced, and standardized to generate structured data that can be analyzed by subsequent modules;

[0083] For visual data, redundant data with high repetition and those that were not fully captured due to being obscured by farm buildings were first screened and discarded. Then, dynamic random cropping, Gaussian blur, four-way flipping, random rotation, RandomColor data augmentation, and Mosaic methods were used to expand the diversity of the dataset. At the same time, illumination correction was performed to correct the color difference of the image under complex lighting conditions and improve the image clarity.

[0084] For audio data, noise reduction algorithms such as spectral subtraction are used to remove environmental noise interference, and audio format conversion and preliminary feature extraction are performed.

[0085] For environmental data, missing value imputation and outlier correction are performed to ensure the integrity and accuracy of the parameter data;

[0086] Finally, all preprocessed data of different types are standardized to unify data format and dimensions, providing a unified input standard for subsequent model training and inference.

[0087] S3. Construct a TwinsYOLO detection model and train it using preprocessed visual data to output duck detection results;

[0088] The TwinsYOLO detection model is built on an improved YOLOv11 algorithm and is used to achieve individual identification of Muscovy ducks and accurate detection of sick ducks in complex environments.

[0089] Specifically, the TwinsYOLO detection model includes a feature extraction unit, an individual recognition unit, an abnormal behavior capture unit, and a lesion recognition unit;

[0090] The feature extraction unit adopts the YOLOv11Backbone architecture with residual networks, which includes an improved C2PSA module and a C3k2 conditional branch structure. It consists of multi-level convolutional layers, an adaptive feature enhancement module, and cross-layer residual connections, and is used to perform hierarchical modeling of the local detail features and overall structural features of the Muscovy duck target.

[0091] In the initial stage of feature extraction, a shallow convolutional network with a small receptive field is used to extract local fine-grained features such as feather texture, lesion edges, and limb contours in the input image, forming a high-resolution low-level feature map.

[0092] As the number of network layers increases, the receptive field of features is gradually expanded by downsampling and channel expansion of multi-level convolutional features, enabling the network to perceive global semantic information such as the overall posture distribution, body shape and spatial relationship between individuals of Muscovy ducks.

[0093] By using the C3k2 conditional branching structure, multi-branch convolutional paths are set at different depths of the backbone network. Different branches correspond to different scales of feature extraction requirements, enabling the network to simultaneously acquire local texture features and medium-scale structural features within the same layer. Furthermore, the feature pathways are dynamically adjusted through the conditional branch selection mechanism, enhancing the adaptability of duck targets under different body sizes, poses, and occlusion states.

[0094] By introducing a feature weighting mechanism of channel and spatial dimensions on the basis of residual structure through the improved C2PSA module, the features output by multi-level convolution are adaptively recalibrated, the response of background regions and redundant information is suppressed, and the key regional features related to the duck target are highlighted, thereby improving the discriminativeness of feature expression while maintaining the lightweight nature of the network.

[0095] By using cross-layer residual connections, the local texture information extracted from the shallow layer is directly passed to the deep network and fused with the global semantic features extracted from the high layer. This avoids the weakening or loss of local details in the deep network and achieves the joint expression of local and global features.

[0096] Ultimately, a local-global feature representation is formed that takes into account both the ability to perceive details and the ability to understand the overall semantics. This provides a stable and highly discriminative feature input for subsequent individual duck identification, abnormal behavior capture, and lesion detection, and achieves a dynamic balance between computational complexity and detection accuracy.

[0097] The individual recognition unit is a Siamese neural network, containing two convolutional network branches with identical structures and shared weights;

[0098] By embedding two convolutional network branches with identical structures and shared weights into a twin neural network, feature vectors of the input image are extracted, the L1 distance between the vectors is calculated and weighted, and a similarity score is output through the Sigmoid function to distinguish highly similar Muscovy duck individuals.

[0099] The abnormal behavior capture unit includes a cross-scale feature fusion subunit and a spatial pyramid pooling subunit, which are used to jointly analyze the morphological changes and behavioral states of individual Muscovy ducks at different spatial scales.

[0100] The cross-scale feature fusion subunit obtains feature maps from different levels of the backbone network. The shallow feature maps have high spatial resolution and are used to characterize the feather arrangement, local texture continuity, and limb edge changes of the Muscovy duck. The deep feature maps have stronger semantic expressive power and are used to characterize the overall posture structure, movement trend, and behavioral stability of the Muscovy duck.

[0101] By aligning and fusing feature maps at different levels, shallow local detail features and deep global behavioral features are jointly expressed in the same feature space, enabling the model to make a comprehensive judgment by combining the overall behavioral pattern while detecting local anomalies.

[0102] The spatial pyramid pooling subunit sets up multiple pooling windows of different scales on the fused feature map to partition and aggregate the feature map, so as to obtain contextual information in local areas, medium ranges and overall ranges respectively, enabling the model to analyze the behavioral state of individual ducks in different spatial receptive fields.

[0103] Within smaller pooling regions, the model focuses on changes in the texture consistency and density of feather arrangement in local areas to identify subtle abnormalities such as ruffled or trembling feathers. Within medium-sized pooling regions, the model focuses on changes in limb swing amplitude and gait rhythm to identify behavioral characteristics such as unsteady walking and abnormal gait. Within larger pooling regions, the model focuses on the overall posture stability, activity range, and sustained movement ability of the Muscovy duck to identify overall behavioral abnormalities such as reduced activity and lethargy.

[0104] Through the synergistic effect of cross-scale feature fusion and spatial pyramid pooling sub-units, the abnormal behavior capture unit can fully integrate multi-level spatial context information in complex breeding environments, and realize dynamic perception and stable identification of abnormal duck behavior from local to overall.

[0105] The lesion recognition unit includes a channel attention mechanism subunit;

[0106] The channel attention mechanism suppresses background noise and enhances the ability to identify small lesions on the surface of ducks. It outputs multi-scale prediction results through three parallel detection heads. The classification branch optimizes efficiency through depthwise separable convolution, and the regression branch improves the loss function to enhance positioning accuracy.

[0107] S4. Construct a Conformer speech processing model and train it using preprocessed audio data to output the recognized speech features.

[0108] The Conformer speech processing model is used to accurately extract human voice features and monitor abnormal voiceprints in high-noise environments at aquaculture sites, ensuring the reliability of voice interaction.

[0109] Specifically, such as Figure 3 As shown, the Conformer speech processing model includes a convolutional subsampling unit, a linear layer, a Dropout layer, multiple ConformerBlock stacked structures, and a voiceprint anomaly detection unit.

[0110] The input audio signal is convolutionally sampled by a convolutional subsampling unit, which reduces the length of the input sequence while preserving feature information.

[0111] The sampled audio data is subjected to linear transformation and random deactivation through linear layers and Dropout layers;

[0112] like Figure 4 As shown, the ConformerBlock stacked structure sequentially includes a first feedforward network, a self-attention module, a convolutional module, and a second feedforward network. The first feedforward network performs a nonlinear transformation on the input data. The self-attention module adopts the relative sinusoidal position encoding scheme in Transformer-XL. The structure of the self-attention module is as follows: Figure 5 As shown, it includes layer normalization, multi-head attention with relative sinusoidal position encoding, and a Dropout layer, which captures the global dependencies of the audio sequence through a multi-head self-attention mechanism;

[0113] The convolution module employs a depthwise separable convolution structure, the structure of which is as follows: Figure 6 As shown, it sequentially includes layer normalization, pointwise convolution, gated linear unit, one-dimensional depthwise separable convolution, batch normalization, Swish activation function, pointwise convolution, and Dropout layer. First, spatial filtering is performed on each input channel separately, then nonlinear transformation is introduced through the Swish activation function, and finally, information is mixed across channels through pointwise convolution with 1×1 convolution kernel. While reducing the number of parameters and computation, it accurately captures local features and distinguishes human voice from environmental noise.

[0114] The second feedforward network has a similar structure to the first feedforward network but does not share parameters. It uses the Swish activation function to further transform and enhance the feature representation. The structures of the second and first feedforward networks are as follows: Figure 7 As shown, it includes, in sequence, layer normalization, linear layer, Swish activation function, Dropout layer, linear layer, and Dropout layer;

[0115] The voiceprint anomaly monitoring unit performs real-time analysis of the extracted human voice features and voiceprint data to identify abnormal voiceprint changes, providing reliable voice input for subsequent interactions.

[0116] S5. Construct a V-Mallard multimodal interactive decision-making model and integrate the preprocessed visual, audio, and environmental data with the outputs of the TwinsYOLO detection model and the Conformer speech processing model to generate aquaculture decisions.

[0117] The V-Mallard multimodal interactive decision-making model includes a multimodal encoder, a large model inference unit, an interactive control unit, and a decision suggestion generation unit.

[0118] like Figure 9 As shown, the multimodal encoder includes a visual encoder, an audio encoder, and a text encoder, which extract features from the preprocessed visual data, audio data, and text data input by farmers, respectively. By accumulating a large amount of multimodal data, the text feature space is aligned with the feature spaces of video, image, and audio. A status token is introduced to recognize the input query type, thereby realizing end-to-end interaction.

[0119] The large model inference unit expands the Chinese vocabulary based on the Mixtral8x7B model, optimizes instructions through a bilingual text corpus, and has the ability to understand both Chinese and English. At the same time, it incorporates professional knowledge related to Muscovy duck farming and builds an expert knowledge base in the field of farming. The professional knowledge related to Muscovy duck farming includes disease prevention and control, environmental control, feed formulation and duck farm layout.

[0120] The large model inference unit adopts the SwitchTransformers architecture, replacing the traditional denseFFN with a sparse expert network (SwitchFFN) in each feed-forward layer. All tokens are allocated to only one expert network through the router, simplifying the routing, allocation and merging process, reducing computing power requirements, and improving computing efficiency and adaptability.

[0121] The interactive control unit adopts a duplex operation scheme, such as Figure 8 As shown, the two V-Mallard models work simultaneously. One is responsible for generating the response to the current query, while the other continuously monitors the environmental input. When a new voice or text query is detected, the current generation process is interrupted, and the response to the new query is output first. It supports non-wake-up dialogue and interrupted command interaction, without the need for preset wake words or manual button triggering. It intelligently recognizes human voices and noise to achieve real-time interaction.

[0122] The decision suggestion generation unit combines multimodal feature analysis results with an expert knowledge base to generate accurate breeding suggestions based on farmers' queries (including issues such as sick duck treatment, environmental control, feed selection, and duck farm layout) and abnormal situations monitored by the system (such as sick duck test results and excessive environmental parameters). For complex and difficult problems, it supports and assists in contacting professional veterinarians.

[0123] like Figure 2 As shown, the present invention also provides a Muscovy duck farming system based on a multimodal large model, used to implement the Muscovy duck farming method based on a multimodal large model described in the present invention. The system includes:

[0124] The data acquisition module is responsible for collecting multimodal raw data during the duck farming process, including visual data, audio data, and environmental data.

[0125] Visual data acquisition uses high-definition cameras, deployed at key locations in the duck farm's breeding area, feeding area, and drinking area to collect images and videos of various physiological behaviors of the ducks, including walking, feeding, flapping their wings, and resting. During the acquisition process, the shooting angle, distance, and lighting conditions are adjusted to cover different background differences, light intensities, and imaging angles, ensuring the randomness and diversity of the sample distribution and avoiding interference from external variables.

[0126] Audio data acquisition uses an array microphone to simultaneously collect audio signals from human voice commands, duck calls, and mechanical equipment operating noise at the breeding site, capturing changes in voiceprint characteristics;

[0127] Environmental data acquisition uses temperature and humidity sensors and ammonia concentration sensors to collect key parameters such as temperature, humidity and ammonia concentration in the breeding environment in real time, providing environmental basis for breeding decisions.

[0128] All data acquisition devices transmit data via a combination of wired and wireless methods to ensure that data is uploaded to the data preprocessing module in real time.

[0129] The data preprocessing module is responsible for cleaning, enhancing, and standardizing the collected multimodal raw data to generate structured data that can be analyzed by subsequent modules.

[0130] In the data preprocessing module, for visual data, firstly, redundant data with extremely high repetition and those that were not fully captured due to being obscured by farm buildings are manually screened and discarded. Then, dynamic random cropping, Gaussian blur, four-way flipping, random rotation, RandomColor data augmentation, and Mosaic methods are used to expand the diversity of the dataset. At the same time, illumination correction processing is performed to correct the color difference of the image under complex lighting conditions and improve the image clarity.

[0131] For audio data, noise reduction algorithms such as spectral subtraction are used to remove environmental noise interference, and audio format conversion and preliminary feature extraction are performed.

[0132] For environmental data, missing value imputation and outlier correction are performed to ensure the integrity and accuracy of the parameter data;

[0133] Finally, all preprocessed data of different types are standardized to unify data format and dimensions, providing a unified input standard for subsequent model training and inference.

[0134] The detection module is used to build the TwinsYOLO detection model, train the TwinsYOLO detection model with preprocessed visual data, and output the duck detection results to the interactive decision module.

[0135] The speech processing module is used to build the Conformer speech processing model, train the Conformer speech processing model with preprocessed audio data, and output the recognized speech features to the interactive decision module.

[0136] The interactive decision-making module is used to build the V-Mallard multimodal interactive decision-making model and integrate the preprocessed visual, audio, and environmental data, as well as the outputs of the TwinsYOLO detection model and the Conformer speech processing model, to generate aquaculture decisions.

[0137] The edge computing deployment module is responsible for localizing the functions of the detection module, voice processing module, and interactive decision-making module. This module employs lightweight optimization techniques to compress each model, ensuring compatibility with the hardware resources of edge computing devices and reducing latency. The edge computing architecture achieves millisecond-level response times, avoiding the impact of network transmission delays on real-time decision-making at the farming site. This is particularly suitable for rural farming environments with limited network conditions, ensuring the system's basic functions operate normally under weak or no network conditions.

[0138] The system management module is responsible for user management, parameter configuration, log recording, data backup and recovery functions.

[0139] The user management function in the system management module supports the creation of different role accounts for farmers, administrators, and technical maintenance personnel, and the allocation of corresponding operation permissions. Farmers can view breeding data, interaction records, and decision suggestions, while administrators are responsible for the overall maintenance of the system and parameter adjustment.

[0140] The parameter configuration function allows administrators to customize system parameters such as data collection frequency, model detection threshold, and interactive response priority to adapt to the needs of different aquaculture scenarios.

[0141] The logging function records the system's operating status, data collection, user actions, and model inference results in detail, which facilitates subsequent traceability and problem investigation.

[0142] The data backup and recovery function regularly backs up the raw data, preprocessed data, analysis results, and interaction records collected by the system. When data loss or failure occurs, the system can quickly recover the data and ensure stable system operation.

[0143] The data security module is used to ensure the security of system data storage, transmission, and privacy protection.

[0144] The data security module uses the AES encryption algorithm to encrypt and store the collected duck images, videos, farmer voice recordings, and personal information to prevent data leakage. Data transmission is encrypted using the SSL protocol to ensure that data is not tampered with or stolen during transmission. A strict access control mechanism is in place, allowing only authorized personnel to view data at the appropriate level and restricting unauthorized access. Regular data security testing and vulnerability patching are performed, with a weekly security vulnerability scan to promptly identify and address potential security risks, ensuring system data security and stability.

[0145] The following will illustrate this with specific implementation examples.

[0146] Initialization and parameter configuration:

[0147] Administrators log in to the system through the management module to perform initial setup:

[0148] Equipment calibration: Install, debug, and calibrate the high-definition camera, array microphone, temperature and humidity sensor, ammonia concentration sensor, and other equipment in the data acquisition module to ensure that the equipment deployment location covers key areas such as the duck farm breeding area, feeding area, and drinking area, and that the equipment is operating normally;

[0149] Model parameter configuration: Set the learning rate, batch size and other training hyperparameters of the TwinsYOLO detection model, and adjust the detection threshold; configure the kernel size, number of network layers and other parameters of the Conformer speech processing model; set the response priority, expert knowledge base update cycle and other parameters of the V-Mallard multimodal interaction decision model.

[0150] Interaction rule configuration: Clarify the triggering conditions for non-wake-up dialogues and the response logic for interrupt commands to ensure that the duplex operation scheme of the interaction control unit is working properly;

[0151] User account creation and permission allocation: Create three types of role accounts: farmers, administrators, and technical maintenance personnel. Assign data viewing and interactive consultation permissions to farmers, system maintenance and parameter adjustment permissions to administrators, and equipment inspection and log viewing permissions to technical maintenance personnel.

[0152] Knowledge base update: Synchronize the professional knowledge of duck farming in the cloud knowledge base, including disease prevention and control, environmental control, feed formulation, duck farm layout, etc., to ensure that the large model reasoning unit has a comprehensive knowledge reserve in the field of farming.

[0153] Multimodal data acquisition:

[0154] After the data acquisition module is started, each acquisition device collects data in real time according to the preset frequency:

[0155] Visual data acquisition: High-definition cameras and video cameras continuously capture the physiological behaviors of Muscovy ducks, such as walking, feeding, flapping wings, and resting. By adjusting the shooting angle and distance, images and videos of Muscovy ducks under different lighting conditions are captured to ensure that the collected samples cover different background differences and imaging perspectives. The sample distribution is random to avoid interference from external variables.

[0156] Audio data acquisition: The array microphones synchronously record human voice commands, duck calls, and mechanical equipment operating noise at the breeding site to ensure the integrity of the audio signal and provide raw data for subsequent voiceprint feature extraction;

[0157] Environmental data acquisition: Temperature and humidity sensors and ammonia concentration sensors monitor parameters such as temperature, humidity and ammonia concentration in the breeding environment in real time, and the data is uploaded to the data preprocessing module in real time;

[0158] All collected data is transmitted via a combination of wired and wireless methods to ensure the stability and real-time nature of data transmission, and to guarantee the authenticity and reliability of the data.

[0159] Data preprocessing:

[0160] After receiving the raw data, the data preprocessing module processes it according to the following steps:

[0161] Visual data processing: First, redundant data with high repetition and those that were not fully captured due to building obstruction were manually screened and deleted. Then, the dataset was expanded using methods such as dynamic random cropping, Gaussian blur, four-way flipping, random rotation, RandomColor data augmentation, and Mosaic. At the same time, illumination correction was performed to correct the color difference of the images under complex lighting conditions. Finally, the processed images and videos were standardized to unify the data format and dimensions.

[0162] Audio data processing: The audio signal is denoised using spectral subtraction to remove environmental noise interference, the audio format is converted to a model-recognizable format, and preliminary feature extraction is performed.

[0163] Environmental data processing: Missing values ​​are filled and outliers are corrected for collected parameters such as temperature, humidity, and ammonia concentration to ensure data integrity and accuracy. Then, the data is standardized to be consistent with other modal data formats.

[0164] The preprocessed structured data is transmitted to the TwinsYOLO detection model, the Conformer speech processing model, and the V-Mallard multimodal interaction decision model to provide data support for subsequent analysis.

[0165] Diseased duck detection and identification:

[0166] The TwinsYOLO detection model receives the preprocessed visual data and performs the following operations:

[0167] Feature extraction: Using the YOLOv11 Backbone architecture, combined with the improved C2PSA module and C3k2 conditional branch structure of the residual network, local-global features of the duck are extracted through multi-level convolution and residual connections, dynamically balancing computational efficiency and detection accuracy.

[0168] Individual recognition: The embedded Siamese neural network extracts feature vectors from the input image through two convolutional network branches with identical structures and shared weights, calculates the L1 distance between the vectors and sums them with weights, and outputs a similarity score through the Sigmoid function to distinguish highly similar Muscovy duck individuals;

[0169] Abnormal Behavior Detection: The Neck module integrates multi-level contextual information through cross-scale feature fusion and spatial pyramid pooling to dynamically capture abnormal behaviors of Muscovy ducks, such as ruffled feathers, abnormal gait, and lethargy.

[0170] Lesion recognition: The channel attention mechanism suppresses background noise and enhances the ability to recognize small lesions on the surface of ducks. The three parallel detection heads of the Head module output multi-scale prediction results. The classification branch optimizes efficiency through depthwise separable convolution, and the regression branch improves the loss function to enhance positioning accuracy.

[0171] Output results: The TwinsYOLO detection model transmits the detection results of sick ducks (including duck farm number, duck number, disease type, etc.) to the V-Mallard multimodal interactive decision model and stores them in the system database.

[0172] Voice feature extraction and interaction preparation:

[0173] The Conformer speech processing model processes the preprocessed audio data:

[0174] Convolutional subsampling: Convolutional sampling is performed on audio signals to reduce sequence length while preserving key features;

[0175] Linear transformation and deactivation: Feature transformation is performed through linear layers, and some neurons are randomly deactivated through dropout layers to improve the model's generalization ability;

[0176] Multi-ConformerBlock Feature Enhancement: Audio data is sequentially processed through the nonlinear transformation of the feedforward network, the global dependency capture of the self-attention module (a multi-head self-attention mechanism based on relative sinusoidal positional encoding), the local feature extraction of the convolution module (depth-separable convolution + Swish activation function + 1×1 point convolution), and the feature enhancement of the second feedforward network to accurately extract human voice features.

[0177] Voiceprint anomaly detection: Real-time analysis of extracted voiceprint features to identify abnormal changes, and transmission of processed speech feature data to the V-Mallard multimodal interaction decision model to support voice interaction.

[0178] Multimodal interaction and decision guidance:

[0179] The V-Mallard multimodal interactive decision-making model integrates visual, audio, environmental data, and diseased duck detection results to achieve intelligent interaction and decision guidance with farmers.

[0180] Multimodal feature alignment: The visual encoder, audio encoder, and text encoder of the multimodal encoder encode visual features, speech features, and text input respectively. The input query type is identified by a status token to achieve multimodal feature space alignment.

[0181] Large-scale model reasoning: Based on the Mixtral8x7B, the reasoning unit expands the Chinese vocabulary and incorporates aquaculture knowledge. Combined with the SwitchTransformers architecture sparse expert network, it efficiently processes multimodal input and calls the cloud knowledge base for reasoning analysis.

[0182] Real-time interaction: The duplex scheme of the interaction control unit is in effect. One model generates the current query response, while the other model continuously monitors new inputs. It supports non-wake-up dialogue and interrupted commands. Farmers can ask questions at any time via voice or text without the need for wake words or manual keys. New queries can interrupt the current response and receive a reply first.

[0183] Decision suggestion generation: For farmers' problems (such as handling sick ducks, environmental control, feed selection, etc.) and abnormalities detected by the system (sick duck test results, environmental parameters exceeding the standard), the system generates precise farming suggestions and assists in contacting veterinarians for complex problems;

[0184] Results push: Decision suggestions are pushed to farmers' interactive interface in the form of voice or text, while interaction records and decision results are stored in the database.

[0185] Edge computing deployment and system maintenance:

[0186] Edge computing operation: The model lightweight optimization subunit of the edge computing deployment module compresses each model to adapt to the hardware resources of edge devices, and the local deployment subunit deploys the core inference function on the edge device to achieve millisecond-level response and ensure the normal operation of basic functions in weak network or network outage environments.

[0187] System monitoring and maintenance: The system management module monitors the equipment's operating status, data acquisition, and model inference results in real time, while the log recording unit records all operations and running information in detail; the data security module uses AES encrypted storage, SSL encrypted transmission, and strict access control, and regularly performs security vulnerability scans and repairs; administrators regularly use the system management module to back up data, view operating logs, and handle equipment failures and security alarms to ensure stable system operation.

[0188] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for raising Muscovy ducks based on a multimodal large model, characterized in that, include: S1. Collect multimodal raw data during the duck farming process, including visual data, audio data, and environmental data; Among them, visual data is collected by deploying cameras in the duck farm’s breeding area, feeding area and drinking area to collect images and videos of ducks walking, eating, flapping their wings and resting. During the collection process, the shooting angle, distance and lighting conditions are adjusted to cover different background differences, light intensity and imaging angle. Audio data is collected synchronously via array microphones, capturing audio signals of human voice commands, duck calls, and mechanical equipment noise from the farm, and capturing changes in voiceprint characteristics. Environmental data is collected in real time from the temperature, humidity and ammonia concentration in the duck farming environment using temperature and humidity sensors and ammonia concentration sensors. S2. Preprocess the collected multimodal raw data, including cleaning, enhancement and standardization. S3. Construct a TwinsYOLO detection model and train it using preprocessed visual data to output duck detection results; The TwinsYOLO detection model includes a feature extraction unit, an individual identification unit, an abnormal behavior capture unit, and a lesion recognition unit. The individual identification unit is a twin neural network containing two convolutional network branches with identical structures and shared weights. These branches extract feature vectors from the input image, calculate the L1 distance between the vectors, and perform a weighted summation. The similarity score is then output using the Sigmoid function to distinguish highly similar Muscovy duck individuals. The abnormal behavior capture unit includes a cross-scale feature fusion subunit and a spatial pyramid pooling subunit, used to jointly analyze the morphological changes and behavioral states of Muscovy duck individuals at different spatial scales. The cross-scale feature fusion subunit obtains feature maps from different levels of the backbone network. Shallow feature maps are used to characterize the feather arrangement, local texture continuity, and limb edge changes of the Muscovy ducks, while deep feature maps are used to characterize the overall posture structure, movement trends, and behavioral stability of the Muscovy ducks. By aligning the feature maps at different levels with their sizes and fusing them step by step, local detail features and global behavioral features are jointly expressed in the same feature space. The spatial pyramid pooling subunit sets multiple pooling windows of different scales on the fused feature maps to partition and aggregate the feature maps, so as to obtain contextual information in local areas, medium ranges, and the overall range, enabling the model to analyze the behavioral state of individual Muscovy ducks in different spatial receptive fields. S4. Construct a Conformer speech processing model and train it using preprocessed audio data to output the recognized speech features. The Conformer speech processing model includes a convolutional subsampling unit, a linear layer, a Dropout layer, multiple ConformerBlock stacked structures, and a voiceprint anomaly detection unit. The ConformerBlock stacked structure includes a first feedforward network, a self-attention module, a convolutional module, and a second feedforward network. The convolutional module adopts a depthwise separable convolutional structure, and the self-attention module adopts the relative sinusoidal position coding scheme in Transformer-XL. S5. Construct a V-Mallard multimodal interactive decision-making model and integrate the preprocessed visual, audio, and environmental data with the outputs of the TwinsYOLO detection model and the Conformer speech processing model to generate aquaculture decisions; The V-Mallard multimodal interactive decision model includes a multimodal encoder, a large model inference unit, an interaction control unit, and a decision suggestion generation unit. The interaction control unit adopts a duplex operation scheme, with two V-Mallard models working simultaneously. One model is responsible for generating the response to the current query, while the other continuously monitors the environmental input. When a new voice or text query is detected, the current generation process is interrupted, and the response to the new query is output first. It supports non-wake-up dialogue and interruptible command interaction.

2. The method for raising Muscovy ducks based on a multimodal large model according to claim 1, characterized in that, In step S2, For visual data, redundant data with high repetition and those that were not fully captured due to being obscured by farm buildings were first screened and discarded. Then, dynamic random cropping, Gaussian blur, four-way flipping, random rotation, RandomColor data augmentation and Mosaic methods were used to expand the diversity of the visual dataset. At the same time, illumination correction was performed to correct the color difference of the image under complex lighting conditions and improve the image clarity. For audio data, noise reduction algorithms are used to remove environmental noise interference, and audio format conversion and preliminary feature extraction are performed. For environmental data, perform missing value imputation and outlier correction. Finally, all preprocessed data of different types are standardized to unify data format and dimensions.

3. The method for raising Muscovy ducks based on a multimodal large model according to claim 1, characterized in that, The feature extraction unit adopts the YOLOv11Backbone architecture with residual networks, which includes an improved C2PSA module and a C3k2 conditional branch structure. It consists of multi-level convolutional layers, an adaptive feature enhancement module, and cross-layer residual connections, and is used to perform hierarchical modeling of the local detail features and overall structural features of the Muscovy duck target. The lesion recognition unit includes a channel attention mechanism subunit; The channel attention mechanism suppresses background noise and enhances the ability to identify small lesions on the surface of ducks. It outputs multi-scale prediction results through three parallel detection heads. The classification branch optimizes efficiency through depthwise separable convolution, and the regression branch improves the loss function to enhance positioning accuracy.

4. The method for raising Muscovy ducks based on a multimodal large model according to claim 1, characterized in that, In the Conformer speech processing model: The input audio signal is convolutionally sampled by a convolutional subsampling unit, which reduces the length of the input sequence while preserving feature information. The sampled audio data is subjected to linear transformation and random deactivation through linear layers and Dropout layers; The self-attention module adopts the relative sinusoidal position coding scheme in Transformer-XL, and captures the global dependencies of the audio sequence through a multi-head self-attention mechanism; The convolution module adopts a depth-separable convolution structure. First, it performs spatial filtering on each input channel separately. Then, it introduces nonlinear transformation through the Swish activation function. Finally, it mixes information across channels through point convolution with a 1×1 convolution kernel. While reducing the number of parameters and the amount of computation, it accurately captures local features and distinguishes between human voice and environmental noise. The second feedforward network has a similar structure to the first feedforward network but does not share parameters. It uses the Swish activation function to further transform and enhance the feature representation. The voiceprint anomaly monitoring unit performs real-time analysis of the extracted human voice features and voiceprint data to identify abnormal voiceprint changes, providing reliable voice input for subsequent interactions.

5. The method for raising Muscovy ducks based on a multimodal large model according to claim 1, characterized in that, In the V-Mallard multimodal interaction decision model, the multimodal encoder includes a visual encoder, an audio encoder, and a text encoder, which respectively extract features from the preprocessed visual data, audio data, and text data input by farmers. By accumulating a large amount of multimodal data, the text feature space is aligned with the feature spaces of video, image, and audio. A status token is introduced to identify the input query type, thereby realizing end-to-end interaction. The large model inference unit expands the Chinese vocabulary based on the Mixtral8x7B model, optimizes instructions through a bilingual text corpus, and has the ability to understand both Chinese and English. At the same time, it incorporates professional knowledge related to Muscovy duck farming and builds an expert knowledge base in the field of farming. The professional knowledge related to Muscovy duck farming includes disease prevention and control, environmental control, feed formulation and duck farm layout. The large model inference unit adopts the SwitchTransformers architecture, replacing the traditional denseFFN with a sparse expert network in each feed-forward layer. All tokens are allocated to only one expert network through the router, simplifying the routing, allocation and merging process and reducing computing power requirements. The interactive control unit adopts a duplex operation scheme, with two V-Mallard models working simultaneously. One model is responsible for generating the response to the current query, while the other continuously monitors the environmental input. When a new voice or text query is detected, the current generation process is interrupted, and the response to the new query is output first. It supports non-wake-up dialogue and interrupted command interaction, without the need for preset wake words or manual button triggering. It intelligently recognizes human voices and noise to achieve real-time interaction. The decision suggestion generation unit combines multimodal feature analysis results with an expert knowledge base to generate corresponding aquaculture decisions based on farmers' queries and abnormal situations detected by the system.

6. A Muscovy duck farming system based on a multimodal large model, used to implement the Muscovy duck farming method based on a multimodal large model as described in any one of claims 1-5, characterized in that, The system includes: The data acquisition module is used to collect multimodal raw data during the duck farming process, including visual data, audio data, and environmental data. The data preprocessing module is used to clean, enhance, and standardize the collected multimodal raw data; The detection module is used to build the TwinsYOLO detection model, train the TwinsYOLO detection model with preprocessed visual data, and output the duck detection results to the interactive decision module. The speech processing module is used to build the Conformer speech processing model, train the Conformer speech processing model with preprocessed audio data, and output the recognized speech features to the interactive decision module. The interactive decision-making module is used to build the V-Mallard multimodal interactive decision-making model and integrate the preprocessed visual, audio, and environmental data, as well as the outputs of the TwinsYOLO detection model and the Conformer speech processing model, to generate aquaculture decisions.

7. The Muscovy duck farming system based on a multimodal large model according to claim 6, characterized in that, The system also includes an edge computing deployment module, which is used to localize the functions of the detection module, voice processing module and interactive decision-making module. The edge computing deployment module uses lightweight optimization technology to compress each model, ensuring that the model is compatible with the hardware resources of the edge computing device and reducing running latency.

8. The Muscovy duck farming system based on a multimodal large model according to claim 6, characterized in that, The system also includes a system management module, which is responsible for user management, parameter configuration, log recording, data backup and recovery functions; The user management function of the system management module supports the creation of different role accounts such as farmers, administrators, and technical maintenance personnel, and the allocation of corresponding operation permissions. Farmers can view breeding data, interaction records and decision suggestions, while administrators are responsible for the overall maintenance and parameter adjustment of the system. The parameter configuration function allows administrators to customize parameters such as data collection frequency, model detection threshold, and interactive response priority to adapt to the needs of different aquaculture scenarios. The logging function is used to record the system's operating status, data collection, user actions, and model inference results, which facilitates subsequent traceability and problem investigation. The data backup and recovery function regularly backs up the raw data, preprocessed data, analysis results, and interaction records collected by the system. When data loss or failure occurs in the system, it is used to quickly recover the data and ensure the stable operation of the system.

9. The Muscovy duck farming system based on a multimodal large model according to claim 6, characterized in that, The system also includes a data security module, which is used to ensure the security of system data storage, transmission, and privacy protection. The data security module uses the AES encryption algorithm to encrypt and store the collected duck images, videos, farmers' voices, and personal information to prevent data leakage. The data transmission process uses SSL encryption to ensure that the data is not tampered with or stolen during transmission. An access control mechanism is set up so that only authorized personnel can view the data of the corresponding level, and unauthorized access is restricted. Data security detection and vulnerability remediation are performed according to the set time, as well as security vulnerability scanning, to promptly discover and deal with potential security risks and ensure the security and stability of system data.

Images