Universal data processing system and method adaptive to various agricultural tasks

Abstract

The invention relates to the technical field of agricultural data processing, in particular to a universal data processing system and method adaptive to various agricultural tasks, and the system comprises an input layer which is used for receiving multi-modal data, constructing a multi-modal data training set, and being in communication connection with an agricultural field physical equation library, the data acquisition module is used for receiving an agricultural physical equation matched with the multi-modal data; the core processing module is a data processing module which is used for establishing a multi-modal data fusion model for fusing multi-modal data, carrying out model parameter optimization in combination with a loss function of a physical equation and migrating data across tasks on the basis of a meta-learning framework; and the output layer is a control output module used for receiving the data processing result, generating a corresponding decision and outputting a control signal. The invention breaks through the bottleneck of the existing agricultural modeling method, and provides a general architecture capable of fusing multi-source heterogeneous data and supporting cross-task migration.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural data processing, and in particular to a general data processing system and method suitable for multiple types of agricultural tasks. Background Art

[0002] In recent years, with the advancement of intelligent agricultural production, data-driven agricultural technology has gradually become the key to improving production efficiency and resource utilization. However, the data characteristics in the agricultural field have multi-source heterogeneity problems, covering multi-dimensional information data in many aspects such as meteorology, soil, and crop growth. This information often differs in scale, accuracy, format, and other aspects, which makes it difficult for traditional deep learning models to obtain accurate prediction results when processing these multi-source data; the reason is that traditional methods do not consider that agriculture is an industry that involves a variety of complex physical processes, with a variety of physical phenomena and interactions from soil to atmosphere. Therefore, the physical laws are not effectively integrated with the actual observation data, which ultimately leads to insufficient generalization of the model.

[0003] In addition, existing agricultural AI models are usually designed for a single task, such as pest and disease identification, crop pathology analysis, crop yield prediction, etc. These models are usually tailored for specific tasks and lack a general architecture to support cross-task knowledge transfer. This single-task design limits the versatility and scalability of the model, requires frequent repeated development, increases development costs, and makes it difficult to quickly adapt to new tasks in practical applications.

[0004] Physical processes in agriculture, especially those related to soil water and heat transfer and crop growth dynamics, usually need to be described through mathematical modeling. Although traditional numerical calculation methods, such as finite difference method (FDM) and finite element method (FEM), can accurately simulate these complex physical processes, these methods consume huge computing resources in practical applications and are difficult to meet real-time requirements.

[0005] Therefore, how to ensure model accuracy while reducing the consumption of computing resources and achieving real-time deployment suitable for multiple tasks has become a difficult problem in the agricultural field. Summary of the invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the defects existing in the above-mentioned prior art, thereby providing a general data processing system and method suitable for various agricultural tasks.

[0007] A general data processing system adapted to various agricultural tasks, comprising: an input layer, a core processing module and an output layer connected in sequence; Input layer: a data acquisition module that is connected to multimodal sensors to receive multimodal data and build a multimodal data training set, and is connected to the agricultural field physical equation library to receive agricultural physical equations that are compatible with multimodal data; Core processing module: connected with the data acquisition module, used to establish a multimodal data fusion model that fuses multimodal data, optimize model parameters by combining the loss function of physical equations, and transfer data across tasks based on the meta-learning framework; Output layer: connected to the data processing module, used to receive data processing results and generate corresponding decisions and output control signals; Among them, the data processing module uses the loss function to optimize the multimodal data fusion model to obtain the optimized parameters as the basic model parameters of the meta-learning framework, so that the meta-learning framework can continuously optimize the model shared parameters during task migration.

[0008] Preferably, the data processing module is composed of a model building module, a physical constraint embedding module and a task migration module; The model building module uses the multimodal feature pyramid network as the backbone network structure to build a multimodal data fusion model; the multimodal data fusion model is trained based on a pre-built multimodal data training set; the trained multimodal data fusion model uses real-time multimodal data as model input; The physical constraint embedding module is used to connect with the model building module, receive the multi-scale features output by the multimodal data fusion model, and transform the physical equation into a residual loss term, a data fitting term, and a model parameter term, and jointly optimize the parameters of the multimodal data fusion model according to the fusion results to obtain the model optimization parameters; The task adaptive migration module is used to connect with the physical constraint embedding module, receive model optimization parameters and multi-scale fusion features, and combine them with its own meta-learning framework to perform shared parameter optimization and data matching across agricultural tasks.

[0009] Preferably, the multimodal data fusion model is provided with three different encoders to adapt to different data types, namely: a spatial encoder, a temporal encoder and a structured encoder; The spatial encoder uses a lightweight convolutional network with a spatial attention mechanism as the backbone network structure, takes the image data of the data acquisition module as input, and uses the feature map as the input. As output; The time series encoder uses a bidirectional LSTM network with temporal attention as the backbone network structure, takes the sensor time series data of the data acquisition module as input, and uses the time series feature vector As output; The structured encoder uses a fully connected network with knowledge graph embedding as the backbone network structure, takes the structured parameters of the data acquisition module as input, and the physical feature vector as output.

[0010] Preferably, the gated attention mechanism used by the multimodal data fusion model is expressed as: ; Where: g is the gate weight, is the activation function, and is the learnable weight, output gate value ; The output fusion feature expression is: .

[0011] Where: It is a multi-scale fusion feature.

[0012] Preferably, the meta-learning framework uses the optimization parameters as the basic parameters of the model, and optimizes the shared parameters of the model adapted to multiple tasks based on a pre-built basic task set.

[0013] Preferably, the residual loss term is composed of the L2 norm of the physical equation residual and the product of the corresponding weight system.

[0014] Preferably, the physical equation library includes a user-defined equation interface; Import third-party agricultural models through the custom equation interface.

[0015] A general method for data processing adapted to multiple types of agricultural tasks is implemented by using a general system for data processing adapted to multiple types of agricultural tasks, and specifically includes the following steps: Acquire multimodal data based on data acquisition module; Use the model building module to build a multimodal data fusion model, and perform preliminary training based on the training set constructed from multimodal data; Based on the pre-built agricultural field physical equation library, the domain differential equation adapted to the task is used to initially optimize the multi-scale fusion features output by the multimodal data fusion model in the form of residual loss, and the optimization parameters are output; The task migration module receives the optimization parameters and further optimizes the model sharing parameters so that the multimodal data fusion model can adapt to the new task based on the historical task data; the initialization parameter expression of the new task is: ; In the formula Indicates the initialization parameters of the new task; Represents model shared parameters; represents the updated learning rate; Represents the loss function of the new task; According to the actual requirements of the adapted new task, the task migration module outputs corresponding decision and control signals.

[0016] The technical solution of the present invention has the following advantages: The present invention systematically proposes a general method to solve the above-mentioned prior art problems by introducing a ternary collaborative architecture of physical embedding, multimodal fusion and task graph migration. Physical embedding can combine physical laws with observation data, realize the organic fusion of data drive and physical model through physical information neural network, and effectively improve the accuracy and robustness of the model. Multimodal fusion can process data from different sources and different types, so that the model can fully obtain various information in the agricultural production process, and further improve the prediction accuracy of the model. The task graph migration mechanism builds a general task framework, so that the model can share knowledge between different tasks, support cross-task transfer learning, thereby greatly reducing the cost of repeated development and improving the versatility and scalability of the system. In addition, the architecture can also effectively reduce the consumption of computing resources, and through reasonable design, it can improve the real-time, deployable and scalable performance of agricultural physical process modeling. In summary, the present invention breaks through the bottleneck of existing agricultural modeling methods through innovative architecture design, and proposes a general architecture that can integrate multi-source heterogeneous data and support cross-task migration, which provides strong technical support for the intelligent development of the agricultural field. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 This is a principle block diagram of a general data processing system adapted to various agricultural tasks of the present invention; Figure 2 This is a practical operation flow chart of Example 3. DETAILED DESCRIPTION

[0019] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. 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] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0021] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0022] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0023] Example 1 like Figure 1 A general data processing system adapted to various agricultural tasks, comprising: an input layer, a core processing module and an output layer connected in sequence; Input layer: a data acquisition module that is connected to multimodal sensors to receive multimodal data and build a multimodal data training set, and is connected to the agricultural field physical equation library to receive agricultural physical equations that are compatible with multimodal data; Core processing module: connected with the data acquisition module, used to establish a multimodal data fusion model that fuses multimodal data, optimize model parameters by combining the loss function of physical equations, and transfer data across tasks based on the meta-learning framework; Output layer: connected to the data processing module, used to receive data processing results and generate corresponding decisions and output control signals; Among them, the data processing module uses the loss function to optimize the multimodal data fusion model to obtain the optimized parameters as the basic model parameters of the meta-learning framework, so that the meta-learning framework can continuously optimize the model shared parameters during task migration.

[0024] Specifically: First, the data processing module consists of a model building module, a physical constraint embedding module, and a task migration module; The model building module uses the multimodal feature pyramid network as the backbone network structure to build a multimodal data fusion model; the multimodal data fusion model is trained based on a pre-built multimodal data training set; the trained multimodal data fusion model uses real-time multimodal data as model input; The physical constraint embedding module is used to connect with the model building module, receive the multi-scale features output by the multimodal data fusion model, and transform the physical equation into a residual loss term, a data fitting term, and a model parameter term, and jointly optimize the parameters of the multimodal data fusion model according to the fusion results to obtain the model optimization parameters; The task adaptive migration module is used to connect with the physical constraint embedding module, receive model optimization parameters and multi-scale fusion features, and combine its own meta-learning framework to perform shared parameter optimization and data matching across agricultural tasks.

[0025] Secondly, in the multimodal data fusion model described in this embodiment, three different encoders are provided to adapt to different data types, namely: a spatial encoder, a temporal encoder, and a structural encoder; The spatial encoder uses a lightweight convolutional network with a spatial attention mechanism as the backbone network structure, takes the image data of the data acquisition module as input, and uses the feature map As output; The time series encoder uses a bidirectional LSTM network with temporal attention as the backbone network structure, takes the sensor time series data of the data acquisition module as input, and uses the time series feature vector As output; The structured encoder uses a fully connected network with knowledge graph embedding as the backbone network structure, takes the structured parameters of the data acquisition module as input, and the physical feature vector as output.

[0026] In addition, the gated attention mechanism expression used in the multimodal data fusion model in this embodiment is: ; Where: g is the gate weight, is the activation function, and is the learnable weight, output gate value ; The output fusion feature expression is: .

[0027] Where: It is a multi-scale fusion feature.

[0028] Finally, for the task transfer module: In this embodiment, the meta-learning framework uses the model optimization parameters as the model basic parameters , and further combines the shared parameters of the model adapted to multiple tasks with the optimized pre-built basic task set.

[0029] In this embodiment, the residual loss term is composed of the L2 norm of the physical equation residual and the corresponding weight system product. It should be noted that in this embodiment, the residual loss term is dynamically adjusted by fusion results of the data fitting term and the model parameter term, specifically, the physical constraint weight is adaptively adjusted according to the data quality.

[0030] The physical equation library in this embodiment includes a user-defined equation interface; Import third-party agricultural models through the custom equation interface.

[0031] Example 2 This embodiment discloses a general method for data processing adapted to multiple types of agricultural tasks, which is implemented by applying a general system for data processing adapted to multiple types of agricultural tasks in Embodiment 1, and specifically includes the following steps: Acquire multimodal data based on data acquisition module; Use the model building module to build a multimodal data fusion model, and perform preliminary training based on the training set constructed from multimodal data; Based on the pre-built agricultural field physical equation library, the domain differential equation adapted to the task is used to initially optimize the multi-scale fusion features output by the multimodal data fusion model in the form of residual loss, and the optimization parameters are output; The task migration module receives the optimization parameters and further optimizes the model sharing parameters so that the multimodal data fusion model can adapt to the new task based on the historical task data; the initialization parameter expression of the new task is: ; In the formula Indicates the initialization parameters of the new task; Represents model shared parameters; represents the updated learning rate; Represents the loss function of the new task; According to the actual requirements of the adapted new task, the task migration module outputs corresponding decision and control signals.

[0032] It should be noted that the execution steps of the task migration module include: Pre-training based on meta-learning strategy: building a shared parameter space on the basic task set; Calculate the similarity between different tasks: quantify the correlation between tasks and guide the selection of transfer learning paths; Progressive fine-tuning: Dynamically adjust the fine-tuning strategy based on the similarity.

[0033] Specifically: 1. Meta-learning framework: Pre-training phase: on a pre-built basic task set The model shared parameters are obtained by optimization.

[0034] The meta-objective function expression is: ; Fine-tuning phase: new tasks The initialization parameters are ; Where: Indicates the initialization parameters of the new task; Represents model shared parameters; represents the updated learning rate; Represents the loss function of the new task; Indicated by the parameter Parameterized machine learning models; represents the i-th task, ; , ; N represents the total number of tasks in the basic task set.

[0035] 2. Task similarity measurement: Task feature vector ; Similarity calculation: ; in: Represents the embedding of physical equations; Represents the i-th task and the jth task The similarities between.

[0036] 3. Migration strategy: High similarity tasks, task similarity>0.75: fine-tune all parameters involved and decay the learning rate to 10% before fine-tuning.

[0037] Medium similarity task, 0.4 Task Similarity 0.75: Fine-tune the model's last three layers and decay the learning rate to 30%.

[0038] Low similarity tasks, task similarity < 0.4: freeze the underlying encoder and only update the task head.

[0039] 4. Relationship between modules: Input dependency: receiving the model basic parameters output by the physical constraint module With multi-scale fusion features .

[0040] Output function: Generate parameters suitable for new tasks , agricultural intelligent systems that are directly deployed to actual applications.

[0041] Example 3 like Figure 2 Based on Example 2, a specific example is given, the task is precision irrigation and variable fertilization of corn, and the method of Example 2 is further disclosed: 1. Application scenarios In view of the spatiotemporal heterogeneity of corn water and fertilizer demand, soil-crop-atmosphere fusion scenario modeling is used to achieve dynamic regulation of water and fertilizer integration to solve the problems of resource waste and environmental pollution caused by traditional irrigation methods.

[0042] 2. Specific implementation process 2.1 Data Preparation Input data: soil data, meteorological data, crop parameter data and control targets; Data acquisition sources and related parameters: Soil data: Soil profile sensors are used to obtain data. The moisture sensor has a depth of 0-100 cm and monitors every 20 cm layer with an accuracy of ±2%. The conductivity sensor detects salt distribution with a range of 0-10 dS / m. The temperature sensor monitors the root zone temperature with a range of -10~50°C.

[0043] Meteorological data include: reference evapotranspiration calculated by the Penman-Monteith equation, and the probability of precipitation in the next 24 hours obtained from weather forecasts.

[0044] Crop parameter data: coded data of growth stages using the universal plant growth and development stage coding system, and root distribution density measured by the micro-root tube method.

[0045] Control objectives: The soil moisture content in the root zone is maintained at 60%~80% of the field water holding capacity, and the nitrate nitrogen concentration is controlled at 20~50 ppm to avoid leaching contamination.

[0046] 2.2 Physical Equations Coupled water-solute transport equation: Richards flow equation: ; Solute transport equation: ; Where: Indicates soil moisture content; represents hydraulic conductivity; t represents time; h represents water head; represents the root water absorption equation used to describe the increase or decrease of water in a unit volume per unit time; z represents the z-axis coordinate; C represents the solute concentration; D represents the hydrodynamic diffusion coefficient; q represents the Darcy flow rate; Indicates an increase or decrease in solute.

[0047] Root water absorption equation: ; represents the root density distribution function, which is obtained by fitting the Logistic model; Indicates root water potential, -150~-50 kPa, varies with crop stage; represents the hydraulic head equation.

[0048] 2.3 Model Training and Optimization Multimodal feature encoding: Spatial encoding pathway: Input: Soil moisture spatial distribution map, generated by sensor interpolation, with a resolution of 1m×1m; Processing: 3 layers of convolution, kernel size 3×3, number of channels 64→128→256+spatial attention; Output: water migration characteristic map, size 16×16×256; Timing encoding pathway: Input: sensor time series data from the data acquisition module: sampling every hour, time window 72 hours; Processing: Bidirectional LSTM, 64 hidden layers + temporal attention; Output: moisture dynamic trend vector, 64 dimensions; Physical parameter path: Inputs: Crop root parameters and soil texture, such as sand / silt / clay ratio; Processing: Fully connected network: 3 layers, 64→32→16 dimensions; Output: physical property feature vector: 16 dimensions; Joint loss optimization: The loss function of a general data processing system adapted to multiple types of agricultural tasks in Example 1 is adaptively adapted to the specific task scenario of this Example 3 to obtain: Loss function = 0.6*data loss term + 0.3*physical loss term + 0.1*model parameter term; Data loss terms: predicted mean square error of soil moisture and nitrogen concentration; Physical loss term: residual L2 norm of Richards flow equation and solute equation; Model parameter item: L2 regularization of model parameters; 0.6, 0.3 and 0.1 are the weight coefficients of the corresponding loss items designed according to the actual task.

[0049] 2.4 Decision output and control Irrigation volume calculation: ; Indicates the root system depth, which is adjusted dynamically according to the actual time, for example, 60cm is used during the jointing period; Indicates the area of ​​the field; Indicates the effective precipitation in the next 24 hours; represents the irrigation volume in time t; Indicates the target soil moisture content; Indicates the actual moisture content of the soil.

[0050] Variable rate fertilization strategy: PID control based on solute concentration deviation: PID control output: ; ; Where: u(t) represents the control output; e(t) represents the deviation; represents the target set value of solute concentration; : actual value of solute concentration; represents proportional gain; represents the integral gain; Derivative gain.

[0051] Specific control output: Nitrogen fertilizer application rate, adjusted every 6 hours.

[0052] Real-time control instructions: Irrigation system: pulse drip irrigation, single irrigation volume ≤ 5mm, anti-runoff; Fertilizer applicator: According to chemical reaction formula =4:1:3 ratio of dynamic fertilizer mixing.

[0053] 3. Implementation Effect The application of the technology in this embodiment in the precision irrigation and variable fertilization decision-making of corn has achieved corresponding economic and ecological benefits. Compared with traditional irrigation methods, the irrigation water consumption is reduced by 23%, effectively alleviating the pressure of water shortage; the nitrogen fertilizer utilization rate is increased from 35% to 49.4%, reducing fertilizer waste while ensuring the balanced supply of nitrogen throughout the growth period of crops; by dynamically adjusting the irrigation intensity and the single irrigation volume ≤5mm, the soil runoff incidence rate is reduced by 42%, protecting the farmland ecosystem.

[0054] This embodiment can realize precise agricultural modeling and decision-making through the physical-data dual-drive architecture, and can be promoted and applied in many major core agricultural scenarios, such as crop breeding and gene optimization, precision cultivation and resource management, disaster monitoring and early warning, remote sensing monitoring and diagnosis, intelligent agricultural machinery and agricultural robots, agricultural product quality and supply chain optimization, soil health and sustainable management, etc., reflecting the universality of the technical solution in multi-field applications in agriculture.

[0055] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A general data processing system adapted to various agricultural tasks, characterized in that: include: The input layer, core processing module and output layer are connected in sequence; Input layer: a data acquisition module that is connected to multimodal sensors to receive multimodal data and build a multimodal data training set, and is connected to the agricultural field physical equation library to receive agricultural physical equations that are compatible with multimodal data; Core processing module: connected with the data acquisition module, used to establish a multimodal data fusion model that fuses multimodal data, optimize model parameters by combining the loss function of physical equations, and transfer data across tasks based on the meta-learning framework; Output layer: connected to the data processing module, used to receive data processing results and generate corresponding decisions and output control signals; Among them, the data processing module uses the loss function to optimize the multimodal data fusion model to obtain the optimized parameters as the basic model parameters of the meta-learning framework, so that the meta-learning framework can continuously optimize the model shared parameters during task migration.

2. A general data processing system adapted to various agricultural tasks according to claim 1, characterized in that: The data processing module consists of a model building module, a physical constraint embedding module, and a task migration module; The model building module uses the multimodal feature pyramid network as the backbone network structure to build a multimodal data fusion model; the multimodal data fusion model is trained based on the pre-built multimodal data training set; The trained multimodal data fusion model uses real-time multimodal data as model input; The physical constraint embedding module is used to connect with the model building module, receive the multi-scale features output by the multimodal data fusion model, and transform the physical equation into a residual loss term, a data fitting term, and a model parameter term, and jointly optimize the parameters of the multimodal data fusion model according to the fusion results to obtain the model optimization parameters; The task adaptive migration module is used to connect with the physical constraint embedding module, receive model optimization parameters and multi-scale fusion features, and combine them with its own meta-learning framework to perform shared parameter optimization and data matching across agricultural tasks.

3. A general data processing system adapted to various agricultural tasks according to claim 2, characterized in that: The multimodal data fusion model is provided with three different encoders to adapt to different data types, namely: a spatial encoder, a temporal encoder and a structured encoder; The spatial encoder uses a lightweight convolutional network with a spatial attention mechanism as the backbone network structure, takes the image data of the data acquisition module as input, and uses the feature map As output; The time series encoder uses a bidirectional LSTM network with temporal attention as the backbone network structure, takes the sensor time series data of the data acquisition module as input, and uses the time series feature vector As output; The structured encoder uses a fully connected network with knowledge graph embedding as the backbone network structure, takes the structured parameters of the data acquisition module as input, and the physical feature vector as output.

4. A general data processing system adapted to various agricultural tasks according to claim 3, characterized in that: The gated attention mechanism used by the multimodal data fusion model is expressed as: ； Where: g is the gate weight, is the activation function, and is the learnable weight, output gate value ; The output fusion feature expression is: 。 Where: It is a multi-scale fusion feature.

5. A general data processing system adapted to various agricultural tasks according to claim 4, characterized in that: The meta-learning framework uses optimized parameters as the basic parameters of the model, and optimizes the shared parameters of the model that are adapted to multiple tasks based on a pre-built basic task set.

6. A general data processing system adapted to various agricultural tasks according to claim 5, characterized in that: The residual loss term consists of the L2 norm of the physical equation residual and the product of the corresponding weight system.

7. A general data processing system adapted to various agricultural tasks according to claim 6, characterized in that: The physical equation library includes a user-defined equation interface; Import third-party agricultural models through the custom equation interface.

8. A general method for data processing adapted to various agricultural tasks, characterized in that: The implementation of a general data processing system adapted to multiple types of agricultural tasks as described in claim 6 specifically includes the following steps: Acquire multimodal data based on data acquisition module; Use the model building module to build a multimodal data fusion model, and perform preliminary training based on the training set constructed from multimodal data; Based on the pre-built agricultural field physical equation library, the domain differential equation adapted to the task is used to initially optimize the multi-scale fusion features output by the multimodal data fusion model in the form of residual loss, and the optimization parameters are output; The task migration module receives the optimization parameters and further optimizes the model sharing parameters so that the multimodal data fusion model can adapt to the new task based on the historical task data; the initialization parameter expression of the new task is: ; In the formula Indicates the initialization parameters of the new task; Represents model shared parameters; represents the updated learning rate; Represents the loss function of the new task; According to the actual requirements of the adapted new task, the task migration module outputs corresponding decision and control signals.

Images