Multi-source data fusion-based rape growth state accurate monitoring method and system

Abstract

The invention discloses a rape growth state accurate monitoring method and system based on multi-source data fusion, and relates to the technical field of smart agriculture, and the method comprises the steps: collecting multi-source heterogeneous data through an edge node, and packaging the multi-source heterogeneous data into a standard data packet; carrying out distributed model training by adopting a federated learning framework with privacy protection; updating the cloud dynamic aggregation parameters to generate a new generation of global model; fusing the multi-source features by using a cross-modal attention mechanism to generate a deep fusion feature vector; self-adaptive scheduling of computing resources is realized based on reinforcement learning; by continuously optimizing the system strategy through closed-loop feedback, the problems of insufficient data privacy protection, weak model generalization ability, resource scheduling rigidity and the like in the prior art are solved, and safe, accurate and real-time monitoring of the rape growth state is realized.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a method and system for precise monitoring of rapeseed growth status based on multi-source data fusion. Background Technology

[0002] In recent years, with the widespread application of the Internet of Things, remote sensing technology, and artificial intelligence in precision agriculture, significant progress has been made in crop growth status monitoring technology based on multi-source data fusion. This technology primarily involves deploying field sensor networks, drones, and satellite platforms to collect multimodal data such as vegetation indices and environmental parameters. Machine learning models are then used for growth status analysis and prediction, gradually achieving dynamic perception of field crops such as rapeseed from macro to micro levels, providing data support for agricultural management decisions.

[0003] However, existing technical solutions largely rely on centralized data processing or simple model ensembles, which have significant limitations. First, data privacy and data silo issues are prominent, making it difficult to securely share sensitive data between farms, thus restricting the model's generalization ability. Second, models are mostly trained on static data, making it difficult to adapt to the dynamic temporal characteristics of crop growth and the spatiotemporal heterogeneity of the field environment, resulting in insufficient prediction accuracy. Third, the allocation of computing resources is rigid, unable to adaptively schedule resources based on task urgency and system load, affecting the real-time performance and energy efficiency of the monitoring system. These inherent defects prevent existing methods from achieving secure cross-domain knowledge fusion and efficient resource collaboration, urgently requiring innovative solutions. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for accurate monitoring of rapeseed growth status based on multi-source data fusion to address the problems of existing rapeseed growth status monitoring methods, such as data privacy barriers leading to weak model generalization, static models being difficult to adapt to dynamic growth environments, and how to achieve adaptive and collaborative scheduling of computing resources to ensure real-time monitoring and energy efficiency.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for precise monitoring of rapeseed growth status based on multi-source data fusion, characterized by comprising the following steps:

[0008] Edge nodes distributed across multiple monitoring points collect local multi-source heterogeneous data and combine it with rapeseed variety and phenological context information to encapsulate it into standardized local structured data packets.

[0009] Each edge node uses its local structured data package to train a shared global model locally, and uploads the updated model parameters obtained from the training to the cloud after privacy protection processing.

[0010] The cloud dynamically weights and aggregates all received model parameter updates to generate and output an optimized global model.

[0011] A global model is used to process multi-source data, and a deep fusion feature vector representing the growth state of rapeseed is generated through its internal cross-modal attention fusion layer.

[0012] The adaptive decision engine dynamically generates computing resource scheduling strategies to allocate inference tasks to edge nodes or the cloud based on deeply fused feature vectors and real-time system status.

[0013] Based on the computing resource scheduling strategy, inference tasks are executed on designated computing nodes to obtain the rapeseed growth status monitoring results, and the results are fed back to the adaptive decision engine to continuously optimize its strategy.

[0014] As a preferred embodiment of the precise monitoring method for rapeseed growth status based on multi-source data fusion described in this invention, the edge nodes distributed at multiple monitoring points collect local multi-source heterogeneous data and, combined with rapeseed variety and phenological context information, encapsulate it into a standardized local structured data package. The specific steps are as follows:

[0015] The edge nodes simultaneously activate multispectral cameras, soil sensor arrays, and environmental meteorological stations to acquire images, physical and chemical data, and meteorological data in a parallel streaming manner.

[0016] Apply a unified timestamp and spatial geographic coordinate registration to the collected modal data streams to ensure that all data points are in the same spatiotemporal reference system;

[0017] Based on the preset rapeseed phenological knowledge base, the system automatically matches the current data with growth stage labels and associates them with the variety resistance information manually entered by farmers.

[0018] Spatiotemporally aligned and context-rich data is packaged into a standard data packet with self-describing metadata. The structure of this data packet supports lossless parsing of subsequent federated learning nodes.

[0019] As a preferred embodiment of the precise monitoring method for rapeseed growth status based on multi-source data fusion described in this invention, each edge node uses its local structured data packet to train a shared global model locally, and uploads the updated model parameters obtained from the training to the cloud after privacy protection processing. The specific steps are as follows:

[0020] Edge nodes download the latest global model parameters from the cloud and perform forward propagation to calculate the loss using local standardized data packets;

[0021] The model is trained locally multiple times using the stochastic gradient descent algorithm to minimize the loss function, thereby obtaining the local update amount of the model parameters.

[0022] Calculate the contribution of this local update amount to the improvement of global model performance, and dynamically adjust the intensity of the added differential privacy noise based on this contribution.

[0023] The updated model parameters with added noise are packaged together with the contribution estimate to form a secure update package, which is then transmitted to the cloud server.

[0024] As a preferred embodiment of the precise monitoring method for rapeseed growth status based on multi-source data fusion described in this invention, the cloud platform dynamically weights and aggregates all received model parameter updates to generate and output an optimized global model. The specific steps are as follows:

[0025] The cloud receives security update packets from each edge node and parses out the update amount and contribution estimate of the noise-adding parameters.

[0026] The contribution estimate of each node is used as the base weight, and a decay factor of the historical update quality of the node is introduced to perform dynamic weight calculation.

[0027] The calculated dynamic weights are used to perform a weighted average of all the noisy parameter updates to obtain the aggregated global parameter updates.

[0028] The aggregated update is applied to the previous generation global model, generating a global model and assigning it a version identifier.

[0029] As a preferred embodiment of the precise monitoring method for rapeseed growth status based on multi-source data fusion described in this invention, the step of processing multi-source data using a global model and generating a deep fusion feature vector representing the rapeseed growth status through its internal cross-modal attention fusion layer includes the following specific steps:

[0030] Multi-source data from the same monitoring unit are input into the corresponding data encoder branch in the global model to extract the initial features of each modality;

[0031] These initial features are fed into a cross-modal attention fusion layer, which learns the interaction weights between features by calculating the attention entropy between features of different modalities.

[0032] The learned interaction weights are used to perform weighted summation and nonlinear transformation on the initial features to generate a compact intermediate fused feature.

[0033] The intermediate fusion feature is subjected to layer normalization and dimensionality reduction to output a deep fusion feature vector with fixed dimensions and rich cross-modal information.

[0034] As a preferred embodiment of the precise monitoring method for rapeseed growth status based on multi-source data fusion described in this invention, the adaptive decision engine dynamically generates a computing resource scheduling strategy to allocate inference tasks to edge nodes or the cloud for execution based on the deep fusion feature vector and the real-time system status. The specific steps are as follows:

[0035] The adaptive decision engine monitors the computational load, available bandwidth, and task urgency indicated by the deep fusion feature vector of each edge node in real time.

[0036] The system state and feature vector are input together into a pre-trained reinforcement learning policy network, which outputs a preliminary decision on the priority of calculating the position.

[0037] Based on the current system global energy consumption target and task deadline constraints, the preliminary decision is optimized and corrected through multiple objectives to generate the final task allocation instruction.

[0038] The final allocation instructions are bound to lightweight model parameters or resource quotas required to execute the task, forming an executable computing resource scheduling strategy package.

[0039] As a preferred embodiment of the precise monitoring method for rapeseed growth status based on multi-source data fusion described in this invention, the steps include: executing inference tasks on designated computing nodes according to a computing resource scheduling strategy to obtain rapeseed growth status monitoring results, and feeding these results back to an adaptive decision engine to continuously optimize its strategy.

[0040] The designated computing node receives the scheduling strategy package, parses and loads the specified model parameters therein, and prepares to execute the inference task.

[0041] This node acquires real-time or near real-time local multi-source data, performs rapid preprocessing on it, and then inputs it into the loaded model for forward inference calculation.

[0042] The growth status indicators obtained from inference calculations are packaged together with their corresponding confidence levels to generate a structured monitoring result report, which is then sent to the user terminal.

[0043] Meanwhile, the execution latency, result confidence, and resource consumption of this task are sent as performance feedback data to the adaptive decision engine to update the training data of its policy network.

[0044] Secondly, this invention provides a precise monitoring system for rapeseed growth status based on multi-source data fusion, comprising:

[0045] The multi-source sensing module collects local multi-source heterogeneous data from edge nodes distributed across multiple monitoring points, and combines this data with rapeseed variety and phenological context information to encapsulate it into standardized local structured data packets.

[0046] The privacy training module allows each edge node to train a shared global model locally using its local structured data package, and then uploads the updated model parameters to the cloud after privacy protection processing.

[0047] The dynamic aggregation module performs dynamic weighted aggregation of all received model parameter updates in the cloud to generate and output an optimized global model.

[0048] The feature fusion module uses a global model to process multi-source data and generates a deep fusion feature vector representing the growth state of rapeseed through its internal cross-modal attention fusion layer.

[0049] The intelligent scheduling module and adaptive decision engine dynamically generate computing resource scheduling strategies to allocate inference tasks to edge nodes or the cloud for execution based on deeply fused feature vectors and real-time system status.

[0050] The closed-loop optimization module executes inference tasks on designated computing nodes according to the computing resource scheduling strategy to obtain the rapeseed growth status monitoring results, and feeds the results back to the adaptive decision engine to continuously optimize its strategy.

[0051] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the method for precise monitoring of rapeseed growth status based on multi-source data fusion as described in the first aspect of the present invention.

[0052] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for precise monitoring of rapeseed growth status based on multi-source data fusion as described in the first aspect of the present invention.

[0053] The beneficial effects of this invention are as follows: By introducing a privacy-preserving federated learning architecture, knowledge sharing and model co-evolution among multiple farms are achieved without the need for centralized raw data, effectively breaking down data silos and significantly improving the model's generalization ability; through cross-modal attention mechanisms and dynamic feature fusion, accurate mining of deep correlations in multi-source heterogeneous data is achieved, overcoming the dependence of traditional methods on static features and single data sources, and enhancing the accuracy and robustness of growth status perception; further, by leveraging an adaptive scheduling mechanism based on reinforcement learning, a dynamic allocation strategy for computing resources between the edge and cloud is constructed, significantly improving the system's real-time response capability and energy efficiency in complex field environments; finally, through a closed-loop feedback mechanism, decision-making strategies and model parameters are continuously optimized, forming a precise monitoring system with self-learning and self-evolution capabilities, achieving a synergistic improvement in the safety, accuracy, real-time performance, and adaptability of rapeseed growth status monitoring as a whole. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of the method for precise monitoring of rapeseed growth status based on multi-source data fusion in Example 1. Detailed Implementation

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0057] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0058] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0059] Example 1, referring to Figure 1This is the first embodiment of the present invention, which provides a method for precise monitoring of rapeseed growth status based on multi-source data fusion, characterized by including the following steps:

[0060] Edge nodes distributed across multiple monitoring points collect local multi-source heterogeneous data and combine it with rapeseed variety and phenological context information to encapsulate it into standardized local structured data packets.

[0061] Each edge node uses its local structured data package to train a shared global model locally, and uploads the updated model parameters obtained from the training to the cloud after privacy protection processing.

[0062] The cloud dynamically weights and aggregates all received model parameter updates to generate and output an optimized global model.

[0063] A global model is used to process multi-source data, and a deep fusion feature vector representing the growth state of rapeseed is generated through its internal cross-modal attention fusion layer.

[0064] The adaptive decision engine dynamically generates computing resource scheduling strategies to allocate inference tasks to edge nodes or the cloud based on deeply fused feature vectors and real-time system status.

[0065] Based on the computing resource scheduling strategy, inference tasks are executed on designated computing nodes to obtain the rapeseed growth status monitoring results, and the results are fed back to the adaptive decision engine to continuously optimize its strategy.

[0066] It should be noted that,

[0067] By synchronously triggering and parallelly acquiring data from multiple source sensors, the inherent consistency of heterogeneous data in the spatiotemporal dimensions is ensured, effectively eliminating data alignment errors caused by traditional asynchronous acquisition. With the help of spatiotemporal registration and phenological knowledge fusion, the original physical observation values ​​are transformed into structured information with clear agronomic semantics, significantly improving the usability and interpretability of the data in growth status analysis. The standardized encapsulation format provides a plug-and-play data interface for subsequent federated learning frameworks, fundamentally solving the integration obstacles caused by the heterogeneity of multi-node data.

[0068] The federated learning architecture enables distributed model training without centralized raw data, fundamentally solving the risks of agricultural data privacy protection and trade secret leakage. Through a contribution-adaptive differential privacy protection mechanism, useful information is preserved to the maximum extent while ensuring data security, overcoming the shortcomings of traditional fixed noise injection that leads to a significant decrease in model performance. Secure transmission protocols and encrypted encapsulation provide end-to-end security for the transmission of agricultural data in open network environments.

[0069] The dynamic aggregation mechanism, which uses both contribution and historical quality weighting, ensures that high-quality updates play a leading role in model evolution, effectively improving the convergence speed and generalization performance of the global model. The introduction of a quality decay factor suppresses the interference of random noise updates on model training, enhancing the robustness of the federated learning system in the face of data anomalies. Version management enables full traceability of the model iteration process, facilitating system maintenance and troubleshooting.

[0070] By utilizing a cross-modal attention mechanism, we can mine the intrinsic correlations of multi-source data, overcoming the problems of information redundancy and noise amplification caused by traditional simple feature splicing. By adaptively focusing on key modal features through attention entropy calculation, we can significantly improve the accuracy and robustness of growth state representation. Normalization and dimensionality reduction processing ensure the stability and processability of feature vectors, providing high-quality input features for subsequent decision analysis.

[0071] Dynamic scheduling of computing resources achieved through reinforcement learning effectively balances the contradiction between system response speed and energy consumption, realizing optimal resource allocation in complex field environments; the multi-objective optimization mechanism ensures that the system meets real-time requirements while taking into account long-term sustainability; the complete encapsulation of the policy package enables the scheduling instructions to be quickly parsed and executed by computing nodes, greatly improving the overall operating efficiency of the system.

[0072] By co-loading lightweight models and resource configurations, efficient execution of inference tasks on edge devices is ensured, achieving low-latency response of the monitoring system. Structured results and confidence assessments provide users with measurable decision support, enhancing the practical value of the system. The performance feedback mechanism forms a complete system optimization closed loop, enabling the entire monitoring system to continuously improve itself.

[0073] Specifically, the edge nodes distributed across multiple monitoring points collect local multi-source heterogeneous data and, combined with rapeseed variety and phenological context information, encapsulate it into standardized local structured data packets. The specific steps are as follows:

[0074] The edge nodes simultaneously activate multispectral cameras, soil sensor arrays, and environmental meteorological stations to acquire images, physical and chemical data, and meteorological data in a parallel streaming manner.

[0075] Apply a unified timestamp and spatial geographic coordinate registration to the collected modal data streams to ensure that all data points are in the same spatiotemporal reference system;

[0076] Based on the preset rapeseed phenological knowledge base, the system automatically matches the current data with growth stage labels and associates them with the variety resistance information manually entered by farmers.

[0077] Spatiotemporally aligned and context-rich data is packaged into a standard data packet with self-describing metadata. The structure of this data packet supports lossless parsing of subsequent federated learning nodes.

[0078] It should be noted that the edge monitoring node synchronously initiates the data acquisition process through an integrated multispectral imager, soil three-parameter sensing unit, and miniature meteorological monitoring device. The multispectral imager captures canopy reflectance images in the visible and near-infrared bands at a rate of 5 frames per second. The soil sensing unit records root zone temperature, moisture content, and conductivity values ​​at preset sampling intervals. The meteorological monitoring device continuously collects photosynthetically active radiation intensity and three-dimensional wind speed data. All sensing devices ensure precise synchronization at the start of acquisition through hardware trigger signals. Raw data is transmitted to the node's built-in preprocessing unit through a parallel interface. Subsequently, the spatiotemporal registration phase begins. The node's built-in GPS module provides centimeter-level spatial coordinate information, a high-precision clock chip provides millisecond-level time reference, and the data processing engine adds metadata headers containing latitude and longitude coordinates, altitude, and a precise timestamp to each data packet. It then uses a spatial interpolation algorithm to map the data from discrete monitoring points to a unified gridded geographic coordinate system. Next, semantic enhancement processing is initiated, calling the rapeseed growth model knowledge base stored in local flash memory. This knowledge base contains the complete phenological stage division standards from sowing to maturity and their corresponding growth characteristic descriptions. After matching the collection time with the phenological stage model, the system automatically generates growth stage identifiers. Simultaneously, it parses the variety genotype code input by farmers through a mobile application and extracts corresponding trait parameters such as disease resistance and drought tolerance from the variety characteristic database. Finally, standardized packaging is performed. The data packaging engine serializes the spatiotemporally aligned and semantically enhanced data according to the preset agricultural IoT data standards, generating a complete data package containing data source, collection parameters, quality control marks, and semantic tags. This data package is stored in a hierarchical structure to ensure that subsequent federated learning nodes can directly parse the core data and metadata.

[0079] By employing a hardware-level synchronous triggering and parallel acquisition mechanism for multi-source sensors, temporal consistency of heterogeneous data at the acquisition source is achieved, effectively solving the temporal misalignment problem caused by traditional sequential acquisition and providing an accurate time reference for multimodal data fusion. Utilizing high-precision positioning and timestamp registration technologies, a unified spatiotemporal reference framework is constructed, eliminating spatial matching errors caused by differences in sensor deployment locations and asynchronous acquisition times, significantly improving the spatial fusion accuracy of multi-source data. Through intelligent matching between the phenological period knowledge base and the variety characteristic database, raw physical observation data is transformed into structured information with clear agronomic semantics, enhancing the interpretability of data in growth status analysis and providing more discriminative feature inputs for subsequent machine learning models. Adopting a standardized layered encapsulation format ensures structural and semantic consistency of multi-node data, providing a plug-and-play data interface for the federated learning framework, fundamentally solving the common data heterogeneity problem in distributed monitoring systems, and laying a solid foundation for large-scale collaborative modeling.

[0080] Specifically, each edge node uses its local structured data packet to train a shared global model locally, and uploads the updated model parameters obtained from the training to the cloud after privacy protection processing. The specific steps are as follows:

[0081] Edge nodes download the latest global model parameters from the cloud and perform forward propagation to calculate the loss using local standardized data packets;

[0082] The model is trained locally multiple times using the stochastic gradient descent algorithm to minimize the loss function, thereby obtaining the local update amount of the model parameters.

[0083] Calculate the contribution of this local update amount to the improvement of global model performance, and dynamically adjust the intensity of the added differential privacy noise based on this contribution.

[0084] The updated model parameters with added noise are packaged together with the contribution estimate to form a secure update package, which is then transmitted to the cloud server.

[0085] It should be noted that the edge node establishes an encrypted connection with the cloud model repository through a secure transport layer protocol, downloads the latest global model parameter file verified by digital signature, which contains the complete neural network weight matrix and bias vector; after loading the model parameters in the local computing environment, the node calls the deep learning inference engine, inputs the multi-source feature data in the standardized data packet into the model computation graph, and performs forward propagation operations through convolutional layers, pooling layers and fully connected layers in sequence, finally obtaining the prediction result vector in the output layer; the system compares the prediction result with the true growth state label marked in the data packet, uses the cross-entropy loss function to calculate the prediction error of each sample, and calculates the arithmetic mean of the loss values ​​of all samples in the batch as the overall loss measure of the current model on the data of that node. Based on the calculated loss value, the node initiates a local training process, employing a stochastic gradient descent optimizer. The backpropagation algorithm is used to calculate the partial derivative of the loss function with respect to each model parameter, obtaining the direction of the parameter gradient. Local iteration count and learning rate parameters are set, and model weights are continuously adjusted over multiple training epochs. The loss function value is recalculated after each iteration. Training terminates when the loss decreases to a stable level or reaches a preset iteration limit. The cumulative change in model parameters at this point is the local update amount. After training, the system runs a contribution evaluation program. By comparing the performance improvement of the model on the validation dataset before and after local training, and considering the fit between the magnitude and direction of the update amount and the global model update history, the expected contribution score of the node's update to the global model optimization is calculated. Based on the contribution score, the node invokes a differential privacy protection engine, using a Laplace mechanism to generate random noise. The noise intensity is inversely proportional to the contribution; that is, weaker noise is added for high-contribution updates, and stronger noise is added for low-contribution updates. The calculated noise vector is then element-wise added to the local update amount to complete the privacy protection process. Finally, the node serializes and encapsulates the updated parameters with added noise, contribution evaluation value, node identifier, timestamp, and other information according to a predetermined format, encrypts the data packet using an asymmetric encryption algorithm, and uploads it to the designated access point of the cloud aggregation server via the Hypertext Transfer Security Protocol.

[0086] By establishing encrypted connections and digital signature verification mechanisms, the security and integrity of model parameter transmission are ensured, effectively preventing man-in-the-middle attacks and parameter tampering risks, and providing a trustworthy foundational environment for distributed learning. Employing complete forward propagation computation and cross-entropy loss evaluation, the adaptability of the global model to local data can be accurately measured, providing reliable gradient direction guidance for subsequent parameter optimization and ensuring the effectiveness of local training. Through multi-round iterative stochastic gradient descent training, model parameters can fully learn the distribution characteristics of local data. The generated updates retain both local specificity and relevance to the global model, providing high-quality parameter adjustment directions for model evolution. Introducing a contribution evaluation mechanism based on validation set performance improvement and update history fit objectively quantifies the value of each node's update to the global model, providing a scientific basis for subsequent weighted aggregation and avoiding interference from low-quality updates. A contribution-adaptive differential privacy protection strategy maximizes the retention of valuable information while ensuring data privacy and security, achieving the optimal balance between privacy protection and model performance, and overcoming the negative impact of traditional fixed noise injection methods on model convergence. By using serialization encapsulation and asymmetric encryption for transmission, the confidentiality and immutability of updated data after privacy protection are ensured during transmission, thus constructing an end-to-end secure data channel and providing a solid security guarantee for cross-domain collaborative learning.

[0087] Specifically, the cloud performs dynamic weighted aggregation of all received model parameter updates to generate and output an optimized global model. The specific steps are as follows:

[0088] The cloud receives security update packets from each edge node and parses out the update amount and contribution estimate of the noise-adding parameters.

[0089] The contribution estimate of each node is used as the base weight, and a decay factor of the historical update quality of the node is introduced to perform dynamic weight calculation.

[0090] The calculated dynamic weights are used to perform a weighted average of all the noisy parameter updates to obtain the aggregated global parameter updates.

[0091] The aggregated update is applied to the previous generation global model, generating a global model and assigning it a version identifier.

[0092] It should be noted that the cloud aggregation server receives security update packets from each edge node through a load balancer. First, it performs identity authentication and digital signature verification to ensure the legality and integrity of the data packets. After successful verification, an asymmetric decryption algorithm is used to decrypt the data packets, extracting the noisy parameter update amount and contribution estimate, and temporarily storing this data in a distributed cache. Then, the system enters the dynamic weight calculation stage. It retrieves the update quality records of each node from the model update history database for the most recent rounds, including the stability of historical contribution and the actual effect on improving the global model performance. The current contribution estimate is used as the base weight, multiplied by a decay coefficient calculated based on historical update quality. This decay coefficient is positively correlated with the stability of the node's historical updates; that is, the more stable the node's historical performance, the closer its decay coefficient is to one, while nodes with greater fluctuations will receive a lower decay coefficient. After multiplying the base weight by the decay coefficient and then normalizing the result, the final dynamic weight value for each node's update is obtained. Next, weighted aggregation calculation is performed. The system multiplies the noise parameter update amount of each node by its corresponding dynamic weight, and then sums all the weighted update amounts to obtain the aggregated global parameter update direction. During this process, the system checks for abnormal update values ​​and truncates update amounts that exceed a reasonable range to prevent individual abnormal nodes from negatively impacting the global model. Finally, model updates and version management are performed. The aggregated parameter update amounts are applied to the parameter space of the current global model, and the model weights are adjusted using an incremental update method. After the update is completed, the system generates a global model and assigns it a unique version identifier containing a timestamp and iteration round information. The new model parameters and version information are stored in the model repository, and the model version directory is updated to prepare for the next round of model downloads for edge nodes.

[0093] Through security verification and decryption extraction steps, only updated data from legitimate nodes can enter the aggregation process, effectively preventing malicious nodes from injecting false data and ensuring the security and reliability of the federated learning system. A dynamic weighting calculation mechanism based on historical update quality identifies and prioritizes updates from stable and reliable nodes, reducing the impact of individual node data anomalies or noise interference on the global model and improving the quality and stability of model aggregation. An aggregation algorithm combining weighted averaging and anomaly handling effectively suppresses the propagation of outliers while fully utilizing valuable information from each node, ensuring the accuracy and robustness of the global model update direction. Incremental updates are used to apply the aggregation results, enabling the model to evolve smoothly, avoiding performance fluctuations caused by parameter mutations, and maintaining the stability of model training. A comprehensive version identification and management system enables full traceability of the model iteration process, facilitating model performance analysis and system fault diagnosis, and providing technical support for comparing the effects of multiple model versions and rollback operations. These combined operations ensure that the cloud aggregation process not only effectively integrates the local knowledge of each node but also continuously optimizes the global model performance while ensuring security, providing stable and reliable model support for the entire monitoring system.

[0094] Specifically, the process of using a global model to process multi-source data and generating a deep fusion feature vector representing the growth state of rapeseed through its internal cross-modal attention fusion layer involves the following steps:

[0095] Multi-source data from the same monitoring unit are input into the corresponding data encoder branch in the global model to extract the initial features of each modality;

[0096] These initial features are fed into a cross-modal attention fusion layer, which learns the interaction weights between features by calculating the attention entropy between features of different modalities.

[0097] The learned interaction weights are used to perform weighted summation and nonlinear transformation on the initial features to generate a compact intermediate fused feature.

[0098] The intermediate fusion feature is subjected to layer normalization and dimensionality reduction to output a deep fusion feature vector with fixed dimensions and rich cross-modal information.

[0099] It should be noted that after receiving multi-source data from the same monitoring unit, the global model first routes the data to corresponding dedicated encoder branches through its internal parallel processing architecture: spectral image data is input to a visual feature extractor based on a convolutional neural network architecture, which progressively extracts visual features from low-level texture to high-level semantics through multiple convolutional and pooling layers; soil parameter data is input to a fully connected encoder network, which performs nonlinear transformations and feature enhancements on physicochemical parameters such as temperature, humidity, and conductivity; and environmental meteorological data is input to a time-series encoder, which captures dynamic change patterns of parameters such as illumination and wind speed through a gated recurrent unit. After each encoder outputs an initial feature vector of a unified dimension, these features are synchronously fed into a cross-modal attention fusion layer. This layer first calculates the correlation matrix between each pair of modal features, normalizes it using a softmax function to obtain the attention weight distribution, and then calculates the dispersion of each weight distribution based on information entropy theory, thereby evaluating the relative importance of different modalities in the judgment of a specific growth state. Next, the system recalibrates the original features using the learned attention weights, and fuses the multimodal features into a unified intermediate representation using a weighted summation method. Then, a nonlinear transformation is performed through a fully connected layer with an activation function to eliminate redundant information between features and enhance the representational power of discriminative features. Finally, the fused features undergo layer normalization to adjust the mean and variance of the feature distribution to improve training stability. Principal component analysis is then used for dimensionality reduction to remove noise and unimportant components, outputting a fixed-length deep fusion feature vector that fully preserves the key discriminative information from the multi-source data.

[0100] A multi-encoder parallel processing architecture was adopted to achieve targeted feature extraction for different modal data characteristics, fully leveraging the unique advantages of each data source in characterizing rapeseed growth status and providing a high-quality feature foundation for subsequent fusion. The attention entropy-based cross-modal fusion mechanism adaptively evaluates the contribution of different data sources in determining specific growth states, effectively solving the problem of inaccurate fixed weight allocation in traditional fusion methods and improving the intelligence and accuracy of feature fusion. Through nonlinear transformation and redundancy elimination, the feature dimensionality was significantly compressed while retaining key information from multiple data sources, avoiding the curse of dimensionality and improving subsequent computational efficiency, providing technical support for real-time monitoring. The combined application of layer normalization and principal component analysis ensured the stability and consistency of the output feature vector, reduced the model's sensitivity to small fluctuations in input data, and enhanced the robustness of the entire monitoring system under different environmental conditions. This series of processing steps together constructs an efficient and reliable multi-source information fusion channel, providing strong feature support for accurately identifying rapeseed growth status and fundamentally improving the perception capability and judgment accuracy of the monitoring system.

[0101] Specifically, the adaptive decision engine dynamically generates a computing resource scheduling strategy to allocate inference tasks to edge nodes or the cloud for execution based on the deep fusion feature vector and the real-time system state. The specific steps are as follows:

[0102] The adaptive decision engine monitors the computational load, available bandwidth, and task urgency indicated by the deep fusion feature vector of each edge node in real time.

[0103] The system state and feature vector are input together into a pre-trained reinforcement learning policy network, which outputs a preliminary decision on the priority of calculating the position.

[0104] Based on the current system global energy consumption target and task deadline constraints, the preliminary decision is optimized and corrected through multiple objectives to generate the final task allocation instruction.

[0105] The final allocation instructions are bound to lightweight model parameters or resource quotas required to execute the task, forming an executable computing resource scheduling strategy package.

[0106] It should be noted that the adaptive decision engine continuously collects computational load metrics such as CPU utilization, memory usage, and storage read / write speed through resource monitoring agents deployed on each edge node. Simultaneously, it utilizes network probes to detect uplink and downlink bandwidth availability and network latency data in real time. This is combined with a task priority identifier parsed from the deep fusion feature vector, which is automatically generated based on the degree of rapeseed growth anomaly and timeliness requirements reflected in the feature vector. The engine inputs the state description tuple, composed of these three types of dynamic system state parameters and the feature vector, into a pre-trained deep reinforcement learning policy network. This network employs a dual-deep Q-network architecture, containing multiple fully connected hidden layers and normalization processing layers. The output is computed through forward propagation. The system outputs the priority score probability distribution for all available computation locations. It then initiates a multi-objective optimizer to collaboratively optimize the initial decisions output by the policy network with preset global energy consumption thresholds and task response time limits. A constraint satisfaction algorithm is used to minimize the total system energy consumption while ensuring task timeliness, and a cost-benefit analysis is performed on edge computing and cloud computing solutions. Finally, the optimized task allocation instructions, along with corresponding lightweight model parameter configuration files, memory allocation schemes, processor core quotas, and other resource constraints, are serialized into a structured data packet. This data packet is encapsulated using a specific binary encoding format, containing a complete instruction header and checksum to ensure the integrity and parseability of the policy information during transmission.

[0107] By monitoring system status and assessing task urgency across multiple dimensions, a comprehensive understanding of the computing environment is achieved, providing accurate input information for intelligent decision-making and ensuring the matching degree between resource allocation schemes and actual needs. Employing a deep reinforcement learning strategy network for initial decision-making enables the system to learn optimal strategies from historical scheduling experience, adapting to complex changes in the field computing environment and improving the intelligence and adaptability of decision-making. A multi-objective optimization algorithm comprehensively considers energy consumption constraints and timeliness requirements, optimizing system energy efficiency while ensuring real-time monitoring and extending the continuous working time of edge devices in the field. The standardized encapsulation of structured strategy packages ensures that scheduling instructions can be quickly identified and executed by various computing nodes, reducing instruction parsing overhead and improving the response efficiency of the entire monitoring system. This complete intelligent decision-making process achieves dynamic optimal allocation of computing resources, providing efficient computing power support for rapeseed growth status monitoring and significantly improving the system's practicality and reliability in complex agricultural environments.

[0108] Specifically, based on the computing resource scheduling strategy, inference tasks are executed on designated computing nodes to obtain rapeseed growth status monitoring results, and these results are fed back to the adaptive decision engine to continuously optimize its strategy. The specific steps are as follows:

[0109] The designated computing node receives the scheduling strategy package, parses and loads the specified model parameters therein, and prepares to execute the inference task.

[0110] This node acquires real-time or near real-time local multi-source data, performs rapid preprocessing on it, and then inputs it into the loaded model for forward inference calculation.

[0111] The growth status indicators obtained from inference calculations are packaged together with their corresponding confidence levels to generate a structured monitoring result report, which is then sent to the user terminal.

[0112] Meanwhile, the execution latency, result confidence, and resource consumption of this task are sent as performance feedback data to the adaptive decision engine to update the training data of its policy network.

[0113] It should be noted that after the designated computing node receives the scheduling policy packet through the secure communication interface, it first performs digital signature verification and data integrity checks to confirm that the data packet has not been tampered with during transmission. After successful verification, the node calls the policy parser to decode the data packet, extracting the lightweight model parameter file, resource configuration parameters, and execution instruction set contained therein. The node dynamically allocates memory buffers and computing cores according to the resource configuration parameters, and simultaneously loads the corresponding lightweight inference model architecture from the local model library, loading the received model parameters into the model structure to complete the initialization preparation of the inference environment. Subsequently, the node initiates the data acquisition process, reading the latest image frames from the multispectral camera, real-time measurements from the soil sensor, and monitoring data from the meteorological station through the local interface. It performs standardized preprocessing on these multi-source data, including outlier filtering, unit unification, and format conversion, to ensure that the data meets the model input requirements. The preprocessed data is sent to the loaded inference model, which sequentially performs convolution operations, activation function processing, and fully connected layer calculations according to the forward propagation process. Finally, the output layer generates prediction results containing multiple growth status indicators such as leaf area index, biomass estimation, and stress level, while calculating the confidence score for each indicator. The node integrates the inference results with corresponding spatiotemporal information and data collection timestamps, generating a structured monitoring report according to a preset report template. This report includes data summaries, anomaly warnings, and decision recommendations, and is pushed to the user's terminal device via a message queue interface. Simultaneously, the node's internal performance monitoring module records the entire performance metrics of this inference task, including task queuing time, data preprocessing time, model inference latency, peak memory usage, and result confidence distribution. These metrics, along with the task identifier, are encapsulated into a performance feedback data packet and sent to the adaptive decision engine's performance database via a dedicated channel.

[0114] Through security verification and dynamic resource allocation mechanisms, the execution of inference tasks in a trusted environment is ensured, while refined management of computing resources is achieved, providing a fundamental guarantee for efficient task operation. Real-time data acquisition and standardized preprocessing processes guarantee the timeliness and quality of input data, effectively reducing the interference of data noise on inference results and improving the accuracy of monitoring results. A complete model inference process and confidence assessment generate interpretable growth state indicators, providing users with intuitive and reliable decision-making basis and enhancing the system's practicality. Structured report generation and multi-channel push mechanisms ensure that monitoring information is delivered to users in a timely and accurate manner, meeting the information acquisition needs of different application scenarios. Comprehensive collection and feedback of performance data provide rich optimization basis for the adaptive decision engine, enabling the system to continuously improve scheduling strategies based on actual operating status, forming a virtuous cycle of self-optimization. The synergistic effect of this series of operations ultimately achieves a complete closed loop for the monitoring system from task execution to result feedback, significantly improving the reliability and practicality of the entire system in practical applications.

[0115] This embodiment also provides a precise monitoring system for rapeseed growth status based on multi-source data fusion, including:

[0116] The multi-source sensing module collects local multi-source heterogeneous data from edge nodes distributed across multiple monitoring points, and combines this data with rapeseed variety and phenological context information to encapsulate it into standardized local structured data packets.

[0117] The privacy training module allows each edge node to train a shared global model locally using its local structured data package, and then uploads the updated model parameters to the cloud after privacy protection processing.

[0118] The dynamic aggregation module performs dynamic weighted aggregation of all received model parameter updates in the cloud to generate and output an optimized global model.

[0119] The feature fusion module uses a global model to process multi-source data and generates a deep fusion feature vector representing the growth state of rapeseed through its internal cross-modal attention fusion layer.

[0120] The intelligent scheduling module and adaptive decision engine dynamically generate computing resource scheduling strategies to allocate inference tasks to edge nodes or the cloud for execution based on deeply fused feature vectors and real-time system status.

[0121] The closed-loop optimization module executes inference tasks on designated computing nodes according to the computing resource scheduling strategy to obtain the rapeseed growth status monitoring results, and feeds the results back to the adaptive decision engine to continuously optimize its strategy.

[0122] This embodiment also provides a computer device applicable to the method for precise monitoring of rapeseed growth status based on multi-source data fusion, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the method for precise monitoring of rapeseed growth status based on multi-source data fusion as proposed in the above embodiment.

[0123] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0124] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for accurate monitoring of rapeseed growth status based on multi-source data fusion as proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0125] In summary, this invention achieves the following: by introducing a privacy-preserving federated learning architecture, it enables secure knowledge sharing and model co-evolution among multiple farms without the need for centralized raw data, effectively breaking down data silos and significantly improving the model's generalization ability; through cross-modal attention mechanisms and dynamic feature fusion, it achieves accurate mining of deep correlations in multi-source heterogeneous data, overcoming the dependence of traditional methods on static features and single data sources, and enhancing the accuracy and robustness of growth status perception; further, by leveraging a reinforcement learning-based adaptive scheduling mechanism, it constructs a dynamic allocation strategy for computing resources between the edge and cloud, significantly improving the system's real-time response capability and energy efficiency in complex field environments; finally, through a closed-loop feedback mechanism, it continuously optimizes decision-making strategies and model parameters, forming a precise monitoring system with self-learning and self-evolution capabilities, thus achieving a synergistic improvement in the safety, accuracy, real-time performance, and adaptability of rapeseed growth status monitoring.

[0126] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for precise monitoring of rapeseed growth status based on multi-source data fusion, characterized in that, Includes the following steps: Edge nodes distributed across multiple monitoring points collect local multi-source heterogeneous data and combine it with rapeseed variety and phenological context information to encapsulate it into standardized local structured data packets. Each edge node uses its local structured data package to train a shared global model locally, and uploads the updated model parameters obtained from the training to the cloud after privacy protection processing. The cloud dynamically weights and aggregates all received model parameter updates to generate and output an optimized global model. A global model is used to process multi-source data, and a deep fusion feature vector representing the growth state of rapeseed is generated through its internal cross-modal attention fusion layer. The adaptive decision engine dynamically generates computing resource scheduling strategies to allocate inference tasks to edge nodes or the cloud based on deeply fused feature vectors and real-time system status. Based on the computing resource scheduling strategy, inference tasks are executed on designated computing nodes to obtain the rapeseed growth status monitoring results, and the results are fed back to the adaptive decision engine to continuously optimize its strategy.

2. The method for precise monitoring of rapeseed growth status based on multi-source data fusion as described in claim 1, characterized in that: The edge nodes distributed across multiple monitoring points collect local multi-source heterogeneous data and, combined with rapeseed variety and phenological context information, encapsulate it into standardized local structured data packets. The specific steps are as follows: The edge nodes simultaneously activate multispectral cameras, soil sensor arrays, and environmental meteorological stations to acquire images, physical and chemical data, and meteorological data in a parallel streaming manner. Apply a unified timestamp and spatial geographic coordinate registration to the collected modal data streams to ensure that all data points are in the same spatiotemporal reference system; Based on the preset rapeseed phenological knowledge base, the system automatically matches the current data with growth stage labels and associates them with the variety resistance information manually entered by farmers. Spatiotemporally aligned and context-rich data is packaged into a standard data packet with self-describing metadata. The structure of this data packet supports lossless parsing of subsequent federated learning nodes.

3. The method for precise monitoring of rapeseed growth status based on multi-source data fusion as described in claim 2, characterized in that: Each edge node uses its local structured data package to train a shared global model locally, and uploads the updated model parameters to the cloud after privacy protection processing. The specific steps are as follows: Edge nodes download the latest global model parameters from the cloud and perform forward propagation to calculate the loss using local standardized data packets; The model is trained locally multiple times using the stochastic gradient descent algorithm to minimize the loss function, thereby obtaining the local update amount of the model parameters. Calculate the contribution of this local update amount to the improvement of global model performance, and dynamically adjust the intensity of the added differential privacy noise based on this contribution. The updated model parameters with added noise are packaged together with the contribution estimate to form a secure update package, which is then transmitted to the cloud server.

4. The method for precise monitoring of rapeseed growth status based on multi-source data fusion as described in claim 3, characterized in that: The cloud platform dynamically weights and aggregates all received model parameter updates to generate and output an optimized global model. The specific steps are as follows: The cloud receives security update packets from each edge node and parses out the update amount and contribution estimate of the noise-adding parameters. The contribution estimate of each node is used as the base weight, and a decay factor of the historical update quality of the node is introduced to perform dynamic weight calculation. The calculated dynamic weights are used to perform a weighted average of all the noisy parameter updates to obtain the aggregated global parameter updates. The aggregated update is applied to the previous generation global model, generating a global model and assigning it a version identifier.

5. The method for precise monitoring of rapeseed growth status based on multi-source data fusion as described in claim 4, characterized in that: The process of using a global model to process multi-source data and generating a deep fusion feature vector representing the growth state of rapeseed through its internal cross-modal attention fusion layer involves the following steps: Multi-source data from the same monitoring unit are input into the corresponding data encoder branch in the global model to extract the initial features of each modality; These initial features are fed into a cross-modal attention fusion layer, which learns the interaction weights between features by calculating the attention entropy between features of different modalities. The learned interaction weights are used to perform weighted summation and nonlinear transformation on the initial features to generate a compact intermediate fused feature. The intermediate fusion feature is subjected to layer normalization and dimensionality reduction to output a deep fusion feature vector with fixed dimensions and rich cross-modal information.

6. The method for precise monitoring of rapeseed growth status based on multi-source data fusion as described in claim 5, characterized in that: The adaptive decision engine dynamically generates a computing resource scheduling strategy based on the deep fusion feature vector and the real-time system state, allocating inference tasks to edge nodes or the cloud for execution. The specific steps are as follows: The adaptive decision engine monitors the computational load, available bandwidth, and task urgency indicated by the deep fusion feature vector of each edge node in real time. The system state and feature vector are input together into a pre-trained reinforcement learning policy network, which outputs a preliminary decision on the priority of calculating the position. Based on the current system global energy consumption target and task deadline constraints, the preliminary decision is optimized and corrected through multiple objectives to generate the final task allocation instruction. The final allocation instructions are bound to lightweight model parameters or resource quotas required to execute the task, forming an executable computing resource scheduling strategy package.

7. The method for precise monitoring of rapeseed growth status based on multi-source data fusion as described in claim 6, characterized in that: The process involves executing inference tasks on designated computing nodes according to a computing resource scheduling strategy to obtain rapeseed growth status monitoring results, and then feeding these results back to the adaptive decision engine to continuously optimize its strategy. The specific steps are as follows: The designated computing node receives the scheduling strategy package, parses and loads the specified model parameters therein, and prepares to execute the inference task. This node acquires real-time or near real-time local multi-source data, performs rapid preprocessing on it, and then inputs it into the loaded model for forward inference calculation. The growth status indicators obtained from inference calculations are packaged together with their corresponding confidence levels to generate a structured monitoring result report, which is then sent to the user terminal. Meanwhile, the execution latency, result confidence, and resource consumption of this task are sent as performance feedback data to the adaptive decision engine to update the training data of its policy network.

8. A precise monitoring system for rapeseed growth status based on multi-source data fusion, based on the precise monitoring method for rapeseed growth status based on multi-source data fusion as described in any one of claims 1 to 7, characterized in that: include, The multi-source sensing module collects local multi-source heterogeneous data from edge nodes distributed across multiple monitoring points, and combines this data with rapeseed variety and phenological context information to encapsulate it into standardized local structured data packets. The privacy training module allows each edge node to train a shared global model locally using its local structured data package, and then uploads the updated model parameters to the cloud after privacy protection processing. The dynamic aggregation module performs dynamic weighted aggregation of all received model parameter updates in the cloud to generate and output an optimized global model. The feature fusion module uses a global model to process multi-source data and generates a deep fusion feature vector representing the growth state of rapeseed through its internal cross-modal attention fusion layer. The intelligent scheduling module and adaptive decision engine dynamically generate computing resource scheduling strategies to allocate inference tasks to edge nodes or the cloud for execution based on deeply fused feature vectors and real-time system status. The closed-loop optimization module executes inference tasks on designated computing nodes according to the computing resource scheduling strategy to obtain the rapeseed growth status monitoring results, and feeds the results back to the adaptive decision engine to continuously optimize its strategy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method for precise monitoring of rapeseed growth status based on multi-source data fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method for precise monitoring of rapeseed growth status based on multi-source data fusion as described in any one of claims 1 to 7.

Images