Federal learning collaborative edge computing target recognition system and method

Abstract

The invention discloses a federated learning collaborative edge computing target recognition system and method, and relates to the technical field of computer artificial intelligence, and the method comprises the steps: cleaning data such as an image at an edge device, carrying out feature entropy enhancement and multi-level coding, training a local model through employing improved AdaM and meta learning, encrypting parameters through chaotic mapping, and uploading the encrypted parameters; metadata such as data complexity is attached, the server uses weighted attention aggregation, verification is carried out according to the block chain, model and update suggestions are distributed to the equipment, and the equipment processes and outputs target details after multi-modal fusion recognition and knowledge graph borrowing. According to the method, the data transmission amount is reduced through edge preprocessing, the model performance is cooperatively improved through federated learning, data safety and model aggregation accuracy are guaranteed through multiple means, adaptive optimization and abnormal data processing can be achieved, multi-modal fusion and knowledge graph assistance accurate recognition can be achieved, and efficient and stable operation of the system is promoted through an incentive mechanism.

Description

Technical Field

[0001] The present invention relates to the field of computer artificial intelligence technology, and in particular to an edge computing target recognition system and method for federated learning collaboration. Background Art

[0002] In today's digital age, target recognition technology is widely used in numerous fields, including security monitoring, autonomous driving, and industrial inspection. Traditional target recognition methods mostly rely on centralized data processing and model training. However, with the explosive growth of data volumes and the increasing demand for data privacy protection, this centralized model has exposed numerous drawbacks. On the one hand, the transmission of massive amounts of data to central servers not only incurs high network bandwidth costs but also easily causes transmission delays, which can have serious consequences in scenarios with high real-time requirements, such as autonomous driving. On the other hand, the centralized storage of large amounts of sensitive data poses significant security risks. Once attacked, data leaks pose a threat to personal privacy, corporate trade secrets, and even national security.

[0003] The emergence of edge computing has alleviated the pressure of data transmission and processing to some extent. It offloads some data processing tasks to edge devices close to the data source, reducing data transmission volume and latency. However, the data volume of a single edge device is limited, and object recognition models trained based on their own data often lack accuracy and generalization capabilities, making them difficult to cope with complex and changing real-world scenarios. For example, in security monitoring, surveillance cameras in different areas collect their own data. A model trained solely on local data may not accurately identify objects that are common in other areas but rare in the local area.

[0004] Federated learning, an emerging distributed machine learning technology, provides a new approach to addressing data silos and privacy protection issues. It allows different participants to collaboratively train models without sharing original data. However, in practical applications, the combination of federated learning and edge computing still faces challenges. Current federated learning-assisted edge computing target recognition systems have shortcomings in model training efficiency, parameter aggregation accuracy, and adaptability to complex scenarios. Existing parameter encryption and transmission methods can affect the accuracy of model aggregation, and there is a lack of effective mechanisms for personalized model optimization based on the characteristics of edge devices and data features. Therefore, the development of an efficient, accurate, and secure federated learning-assisted edge computing target recognition system and method is of great practical significance and urgent need. Summary of the Invention

[0005] The present invention proposes a federated learning collaborative edge computing target recognition system and method to solve the problems mentioned in the above-mentioned prior art.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A federated learning collaborative edge computing target recognition method includes the following steps:

[0008] Data preprocessing steps: On the edge device side, perform routine cleaning, normalization, cropping and scaling operations on the collected data, introduce adaptive enhancement processing based on feature entropy, and calculate the feature entropy of local areas in the image where p i It is the feature probability distribution. At the same time, the image data is hierarchically encoded according to spatial and semantic information;

[0009] Local model training steps: The edge device uses the preprocessed data to train the model locally and uses the improved adaptive momentum estimation optimization algorithm to update the model parameters. The formula is in

[0010] β1 and β2 are exponential decay rates. At the same time, a meta-learning mechanism is introduced to allow the model to adapt to different task subsets;

[0011] Model parameter encryption and upload steps: The edge device uses an encryption algorithm based on chaotic mapping to encrypt local parameters, and performs permutation and diffusion operations on the model parameters through the pseudo-random sequence generated by the chaotic system. The encryption process meets Among them C i is the encrypted parameter, P i is the original parameter, S i Generate the key for the chaotic sequence and upload the encrypted parameters to the federated learning server;

[0012] Global model aggregation step: The federated learning server receives the encrypted model parameters and aggregates them using a federated aggregation algorithm based on the weighted attention mechanism. The algorithm assigns attention weights w to the model parameters of the edge devices based on the data information in the metadata file. i , and then by the formula Aggregation, where

[0013] Global model decryption and distribution step: Use the inverse chaotic mapping operation corresponding to the encryption process to decrypt the aggregated model parameters, and distribute the decrypted parameters to the edge devices along with a model update suggestion file;

[0014] Target recognition steps: Use a multimodal fusion target recognition method to identify newly acquired target images, fuse image and sensor data features, and build a cross-modal attention mechanism to allow the model to automatically pay attention to modal information and output information based on the recognition results.

[0015] Furthermore, it also includes the model adaptive optimization step. After the global model is distributed, the edge device adaptively optimizes the local model according to the changes and needs of the local data. By introducing the optimization strategy based on dynamic programming, a multi-stage optimization problem is constructed, using the formula Adjust the model parameters, where α t is the adaptive adjustment coefficient that changes dynamically over time.

[0016] Furthermore, it also includes abnormal data detection and processing steps. In the data preprocessing stage, an abnormal data is detected using an anomaly detection algorithm based on a deep autoencoder. The deep autoencoder calculates the reconstruction error by learning the distribution characteristics of normal data. When the reconstruction error exceeds the threshold, it is judged as abnormal data, and the abnormal data is repaired using a data repair method based on a generative adversarial network.

[0017] Furthermore, in the local model training step, a method based on the fusion of transfer learning and meta-learning is adopted. The pre-trained object recognition model is used as the initial model and fine-tuned on the local data. At the same time, meta-learning meta-parameters are introduced to allow the model to adapt to different tasks and data distributions. By building a meta-learner, the model parameters are learned and adjusted.

[0018] Furthermore, in the global model aggregation step, a blockchain-based trust mechanism is introduced. Each edge device records the uploaded model parameters and metadata files on the blockchain. At the same time, an incentive mechanism based on game theory is used to encourage edge devices to provide model parameters and reward them according to their contribution.

[0019] Furthermore, in the target recognition step, a knowledge graph containing target categories, attributes, and relationship information is constructed. After the target recognition is completed, the knowledge graph is inferred and verified, and errors in the recognition results are corrected through the constraints and association information of the knowledge graph.

[0020] Furthermore, the system of the edge computing target recognition method based on federated learning collaboration includes the following modules:

[0021] Edge data acquisition and preprocessing module: This module is responsible for collecting raw target images and sensor data on edge devices, and performs cleaning, normalization, feature entropy-based adaptive enhancement processing, and multi-level encoding. This module integrates multiple sensors and improves data processing speed through hardware acceleration chips and parallel computing architecture.

[0022] Edge model training module: This module uses preprocessed data to locally train the object recognition model, employing an improved adaptive momentum estimation optimization algorithm and meta-learning mechanism. This module integrates a deep learning framework and meta-learning library, supporting a variety of model architectures and task settings. During training, it dynamically adjusts the learning rate based on the characteristics and distribution of local data.

[0023] Data encryption and transmission module: Uses a chaos map-based encryption algorithm to encrypt locally trained model parameters, and uploads the encrypted model parameters and metadata files containing data complexity indexes to the federated learning server, using communication protocols and blockchain technology to transmit data;

[0024] Federated Learning Server Module: Receives encrypted model parameters from multiple edge devices and aggregates them using a federated aggregation algorithm based on a weighted attention mechanism. This module integrates blockchain nodes and a deep learning computing cluster. During the aggregation process, attention weights are assigned to model parameters on different edge devices based on the information in the metadata file and dynamically adjusted.

[0025] Model decryption and distribution module: This module uses the inverse chaotic mapping operation corresponding to the encryption process to decrypt the model parameters, distributes the decrypted global model parameters to each edge device, and generates and distributes the model update suggestion file.

[0026] Target recognition module: Uses the updated local model to identify newly acquired target images and sensor data, adopts a multimodal fusion target recognition method and a knowledge graph-based post-processing method. This module integrates a multimodal data processing engine and a knowledge graph reasoning engine, and outputs target information based on the recognition results.

[0027] Furthermore, it also includes a model adaptive optimization module. After each global model distribution, this module adaptively optimizes the local model according to the changes in local data and the requirements of the recognition task, using an optimization strategy based on dynamic programming and an adaptive adjustment coefficient. This module integrates an optimization algorithm library and a data analysis engine to monitor model performance and data distribution changes in real time.

[0028] Furthermore, it also includes an abnormal data detection and processing module. During the data preprocessing stage, this module uses an anomaly detection algorithm based on a deep autoencoder to detect abnormal data. For the detected abnormal data, a data repair method based on a generative adversarial network is adopted. This module integrates a deep autoencoder model and a generative adversarial network model, and performs training and inference through a deep learning framework.

[0029] Furthermore, it also includes an incentive and trust management module, which introduces a blockchain-based trust mechanism and a game theory-based incentive mechanism. It records model parameters and metadata files through the blockchain, verifies the authenticity and integrity of the data, and gives corresponding rewards based on the contribution of edge devices. This module integrates blockchain nodes and game theory algorithm libraries.

[0030] Compared with the existing technology, the beneficial effects of the present invention are:

[0031] In terms of data processing and transmission, edge devices perform preliminary screening and preprocessing of raw data, significantly reducing the amount of data transmitted to the server, lowering network bandwidth requirements, alleviating data transmission pressure, and improving system response speed. Furthermore, the edge computing-based architecture enables preliminary data processing locally, reducing the risk of data exposure during network transmission and enhancing data security.

[0032] In terms of model training and optimization, federated learning enables multiple edge devices to collaboratively train models without sharing raw data, effectively integrating the data strengths of all parties and improving model accuracy and generalization. Improved encryption algorithms ensure the privacy and security of model parameters during transmission. An aggregation algorithm based on a weighted attention mechanism rationally assigns weights based on the data quality and training results of each device, improving the accuracy of global model aggregation. After receiving the global model, edge devices use adaptive optimization strategies to make personalized adjustments based on local data characteristics and task requirements, further improving the model's recognition performance in local scenarios.

[0033] For abnormal data, this application can accurately detect and effectively process it, improve data quality, avoid the interference of abnormal data on model training and recognition results, and thus improve the reliability of target recognition. In actual application scenarios, multimodal data fusion and knowledge graph-based post-processing methods make target recognition more accurate and comprehensive, not only identifying target categories, but also obtaining more attributes and relationship information of the target. In addition, the trust mechanism based on blockchain and the incentive mechanism of game theory ensure the authenticity and integrity of the data and models in the system, stimulate the enthusiasm of edge devices to participate in collaboration, ensure the stable and efficient operation of the entire system, and provide better quality and reliable technical support for target recognition applications in security, transportation, industry and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a schematic block diagram of the edge computing target recognition method for federated learning collaboration proposed in the present invention;

[0035] Figure 2 This is a schematic block diagram of the federated learning collaborative edge computing target recognition system proposed in the present invention;

[0036] Figure 3 This is a bar chart comparing target recognition accuracy in different scenarios for the federated learning collaborative edge computing target recognition system and method proposed in the present invention;

[0037] Figure 4 A line graph showing the change in model training time versus data volume for the federated learning collaborative edge computing target recognition system and method proposed in the present invention;

[0038] Figure 5A pie chart showing the data transmission volume ratio of the federated learning collaborative edge computing target recognition system and method proposed in the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention 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 making creative efforts are within the scope of protection of the present invention.

[0040] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying 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 should not be understood as limiting the present invention.

[0041] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the said features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, 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 a connection between the two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below with reference to the accompanying drawings.

[0042] Reference Figures 1 to 5 : A federated learning collaborative edge computing target recognition method includes the following steps:

[0043] Data preprocessing steps: At the edge device end, use high-definition cameras, lidar, millimeter-wave radar and other sensors to collect target images and related data. For the collected original target image data, first perform a cleaning operation to remove fuzzy, repeated, and incomplete data to improve data quality. Then normalize the cleaned data using the formula Where x is the original pixel value, x min and x max are the minimum and maximum values ​​of the pixel values ​​of this channel, respectively, norm is the normalized pixel value. In order to enhance the image details, an adaptive enhancement process based on feature entropy is introduced. For each local area in the image, its feature entropy is calculated. where p i is the probability distribution of the i-th feature within the region. For regions with low feature entropy, a combination of histogram equalization and Gaussian filtering is used to enhance image detail, improving the subsequent model's ability to perceive objects in different scenarios. Furthermore, image data is hierarchically encoded based on spatial and semantic information, forming a multi-level data structure. For example, different image regions and features are categorized and encoded to facilitate efficient processing by subsequent models.

[0044] Local model training step: The edge device uses the pre-processed data to train the target recognition model locally. The improved adaptive momentum estimation (AdaM) optimization algorithm is used to update the model parameters. The update formula is: in

[0045] β1 and β2 are exponential decay rates that are dynamically adjusted during training based on the model's convergence speed and changes in the loss function. Furthermore, a meta-learning mechanism is introduced to enable the model to rapidly learn and adapt across different task subsets. For example, data can be divided into multiple distinct task subsets, with the model trained and learning on each subset, improving the model's generalization and transferability.

[0046] Model parameter encryption and upload steps: To ensure data privacy, the edge device uses an encryption algorithm based on chaotic mapping to encrypt the locally trained model parameters. The model parameters are permuted and diffused through a pseudo-random sequence generated by the chaotic system. The encryption process satisfies Among them C i is the encrypted parameter, P i is the original parameter, S i The key generated for the chaotic sequence is uploaded to the federated learning server. A metadata file containing local data features, training effect evaluation indicators, and a data complexity index is uploaded. The data complexity index is obtained by calculating the information entropy and feature diversity of the data.

[0047] Global model aggregation step: The federated learning server receives encrypted model parameters from multiple edge devices and aggregates them using a federated aggregation algorithm based on a weighted attention mechanism. The algorithm first assigns an attention weight w to the model parameters of each edge device based on the data complexity index and training effect evaluation indicators in the metadata file. i , and then by the formula Aggregation is performed, wherein During the aggregation process, the server also dynamically adjusts the model parameters, introducing regularization terms to balance the model's complexity and generalization capabilities. At the same time, the server verifies the authenticity and integrity of the model parameters and metadata files based on the blockchain.

[0048] Global model decryption and distribution: The federated learning server decrypts the aggregated global model parameters using the inverse chaotic mapping operation corresponding to the encryption process. The decrypted global model parameters are distributed to each edge device along with a model update recommendation file that provides targeted model fine-tuning strategies based on the edge device's local data characteristics and historical training data.

[0049] Target recognition: The edge device uses the updated local model to identify newly acquired target images. This multimodal fusion target recognition method incorporates features from other sensors (such as lidar and millimeter-wave radar) in addition to image features. By building a cross-modal attention mechanism, the model automatically focuses on key information in data from different modalities, improving the accuracy and robustness of target recognition. Based on the recognition results, detailed information such as the target's category, location, and posture is output.

[0050] The present invention also includes a model adaptive optimization step. After each global model distribution, the edge device adaptively optimizes the local model according to the changes in local data and the requirements of the recognition task. By introducing an optimization strategy based on dynamic programming, a multi-stage optimization problem is constructed with the goal of maximizing recognition accuracy and minimizing computing resource consumption. Using the formula Adjust the model parameters, where α t It is an adaptive adjustment coefficient that changes dynamically over time and is adjusted in real time according to the distribution changes of local data and the performance indicators of the model.

[0051] The present invention also includes an abnormal data detection and processing step. During the data preprocessing stage, an abnormal data is detected using an anomaly detection algorithm based on a deep autoencoder. The deep autoencoder reconstructs the input data by learning the distribution characteristics of normal data and calculates the reconstruction error. When the reconstruction error exceeds a preset threshold, the data is determined to be abnormal. For the detected abnormal data, a data repair method based on a generative adversarial network is used to generate data with a distribution similar to the normal data to replace the abnormal data.

[0052] In this paper, a method based on the fusion of transfer learning and meta-learning is employed during the local model training step. A pre-trained object recognition model is used as the initial model and fine-tuned on local data. Furthermore, meta-learning meta-parameters are introduced to enable the model to quickly adapt to different tasks and data distributions. By constructing a meta-learner and performing meta-training on multiple tasks, the model parameters are learned to be quickly adjusted, improving the model's training efficiency and generalization capabilities.

[0053] In this paper, a blockchain-based trust mechanism is introduced during the global model aggregation step to ensure the authenticity and integrity of model parameters. Each edge device records the uploaded model parameters and metadata files on the blockchain. During the aggregation process, the federated learning server verifies the blockchain records to ensure data reliability. Furthermore, a game-theory-based incentive mechanism is used to encourage edge devices to provide high-quality model parameters, with rewards awarded based on their contribution.

[0054] In this invention, a knowledge graph-based post-processing method is employed during the target recognition step. A knowledge graph containing information such as target categories, attributes, and relationships is constructed. After target recognition is complete, reasoning and verification are performed within the knowledge graph based on the recognition results. By leveraging the constraints and associations of the knowledge graph, errors in the recognition results are corrected, improving the accuracy and reliability of target recognition.

[0055] In the present invention, the system of the edge computing target recognition method of federated learning collaboration includes the following modules:

[0056] Edge Data Acquisition and Preprocessing Module: This module collects raw target image data and other sensor data from edge devices, performing cleaning, normalization, adaptive enhancement processing based on feature entropy, and multi-level encoding. This module integrates multiple sensors, including high-definition cameras, lidar, and millimeter-wave radar, and accelerates data processing through hardware acceleration chips and a parallel computing architecture.

[0057] Edge Model Training Module: This module uses preprocessed data to locally train the object recognition model. Using an improved Adaptive Momentum Estimation (AdaM) optimization algorithm and meta-learning mechanisms, this module integrates deep learning frameworks and meta-learning libraries, supporting a variety of model architectures and task settings. During training, the learning rate and other hyperparameters are dynamically adjusted based on the characteristics and distribution of the local data.

[0058] Data Encryption and Transmission Module: Utilizing a chaotic mapping-based encryption algorithm, the locally trained model parameters are tightly encrypted. This encryption method effectively resists various attacks and ensures the confidentiality of the model parameters. After encryption, the module uploads the encrypted model parameters, along with a metadata file containing the data complexity index, to the federated learning server. During transmission, the module utilizes a secure and reliable communication protocol, creating a solid barrier to prevent data leakage and tampering. Furthermore, blockchain technology, leveraging its immutable and traceable properties, provides a reliable record of the data transmission process, ensuring data security and traceability.

[0059] The Federated Learning Server Module receives encrypted model parameters from multiple edge devices and meticulously aggregates them using a federated aggregation algorithm based on a weighted attention mechanism. Equipped with powerful computing and storage capabilities, it acts like a fortress, integrating blockchain nodes and a deep learning computing cluster. During the aggregation process, it acts as a discerning conductor, precisely assigning attention weights to model parameters on different edge devices based on information contained in metadata files and dynamically adjusting these weights based on actual conditions. This enables efficient and accurate model aggregation, facilitating the smooth implementation of federated learning.

[0060] Model Decryption and Distribution Module: Decrypts aggregated global model parameters using the same inverse chaotic mapping operation as the encryption process. The decrypted global model parameters are distributed to each edge device, while also generating and distributing model update suggestion files. This module features an efficient decryption algorithm and file distribution mechanism.

[0061] Target Recognition Module: This module uses the updated local model to identify newly acquired target images and other sensor data. Using a multimodal fusion target recognition method and knowledge graph-based post-processing, this module integrates a multimodal data processing engine and a knowledge graph inference engine to output detailed target information based on the recognition results.

[0062] The present invention also includes a model adaptive optimization module, which will immediately perform adaptive optimization on the local model according to the changes in local data and the actual needs of the recognition task after each global model distribution. This module uses an optimization strategy based on dynamic programming, combined with an adaptive adjustment coefficient, to achieve fine-grained adjustment of the model. It integrates a huge library of optimization algorithms and a high-performance data analysis engine, which can monitor changes in model performance indicators and changes in data distribution in real time. Based on the monitoring results, the module will adjust the local model in a timely and accurate manner to ensure that the model is highly adapted to the actual application scenario, and continuously improve the accuracy and applicability of the model in tasks such as target recognition.

[0063] The present invention also includes an abnormal data detection and processing module. During the data preprocessing phase, this module uses an anomaly detection algorithm based on a deep autoencoder to detect abnormal data. For detected abnormal data, a data repair method based on a generative adversarial network is used. This module integrates a deep autoencoder model and a generative adversarial network model, and performs training and inference through a deep learning framework.

[0064] This invention also includes an incentive and trust management module, which introduces a blockchain-based trust mechanism and a game theory-based incentive mechanism to build a fair and reliable operating environment. Leveraging the immutable nature of blockchain technology, detailed records of model parameters and metadata files enable accurate verification of data authenticity and integrity, ensuring system data quality. Furthermore, the module rewards edge devices based on their contribution to the federated learning process. It integrates blockchain nodes and a game theory algorithm library, ensuring comprehensive fairness and reliability through scientific calculation and judgment, and promoting active collaboration among all participating parties.

[0065] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for edge computing target recognition based on federated learning collaboration, characterized in that: The following steps are involved: Data preprocessing steps: On the edge device side, perform routine cleaning, normalization, cropping and scaling operations on the collected data, introduce adaptive enhancement processing based on feature entropy, and calculate the feature entropy of local areas in the image where p i It is the feature probability distribution. At the same time, the image data is hierarchically encoded according to spatial and semantic information; Local model training steps: The edge device uses the preprocessed data to train the model locally and uses the improved adaptive momentum estimation AdaM optimization algorithm to update the model parameters. The formula is in β1 and β2 are exponential decay rates. At the same time, a meta-learning mechanism is introduced to allow the model to adapt to different task subsets; Model parameter encryption and upload steps: The edge device uses an encryption algorithm based on chaotic mapping to encrypt local parameters, and performs permutation and diffusion operations on the model parameters through the pseudo-random sequence generated by the chaotic system. The encryption process meets Among them C i is the encrypted parameter, P i is the original parameter, S i Generate the key for the chaotic sequence and upload the encrypted parameters to the federated learning server; Global model aggregation step: The federated learning server receives the encrypted model parameters and aggregates them using a federated aggregation algorithm based on the weighted attention mechanism. The algorithm assigns attention weights w to the model parameters of the edge devices based on the data information in the metadata file. i , and then by the formula Aggregation, where Global model decryption and distribution step: Use the inverse chaotic mapping operation corresponding to the encryption process to decrypt the aggregated model parameters, and distribute the decrypted parameters to the edge devices along with a model update suggestion file; Target recognition steps: Use a multimodal fusion target recognition method to identify newly acquired target images, fuse image and sensor data features, and build a cross-modal attention mechanism to allow the model to automatically pay attention to modal information and output information based on the recognition results.

2. The method for edge computing target identification based on federated learning collaboration according to claim 1, characterized in that: It also includes the model adaptive optimization step. After the global model is distributed, the edge device adaptively optimizes the local model according to the changes and needs of the local data. By introducing the optimization strategy based on dynamic programming, a multi-stage optimization problem is constructed using the formula Adjust the model parameters, where α t is the adaptive adjustment coefficient that changes dynamically over time.

3. The method for edge computing target identification based on federated learning collaboration according to claim 1, characterized in that: It also includes abnormal data detection and processing steps. In the data preprocessing stage, an anomaly detection algorithm based on a deep autoencoder is used to detect abnormal data. The deep autoencoder calculates the reconstruction error by learning the distribution characteristics of normal data. When the reconstruction error exceeds the threshold, it is judged as abnormal data, and the data repair method based on the generative adversarial network is used to repair the abnormal data.

4. The method for edge computing target identification based on federated learning collaboration according to claim 1, characterized in that: In the local model training step, a method based on the fusion of transfer learning and meta-learning is adopted. The pre-trained object recognition model is used as the initial model and fine-tuned on local data. At the same time, meta-learning meta-parameters are introduced to allow the model to adapt to different tasks and data distributions. By building a meta-learner, the model parameters are learned and adjusted.

5. The method for edge computing target identification based on federated learning collaboration according to claim 1, characterized in that: In the global model aggregation step, a blockchain-based trust mechanism is introduced. Each edge device records the uploaded model parameters and metadata files on the blockchain. At the same time, an incentive mechanism based on game theory is used to encourage edge devices to provide model parameters and reward them according to their contribution.

6. The method for edge computing target identification based on federated learning collaboration according to claim 1, characterized in that: In the target recognition step, a knowledge graph containing target categories, attributes, and relationship information is constructed. After the target recognition is completed, the knowledge graph is inferred and verified, and errors in the recognition results are corrected through the constraints and association information of the knowledge graph.

7. A system using the method for edge computing target recognition using federated learning collaboration according to any one of claims 1 to 6, characterized in that: Includes the following modules: Edge data acquisition and preprocessing module: This module is responsible for collecting raw target images and sensor data on edge devices, and performs cleaning, normalization, feature entropy-based adaptive enhancement processing, and multi-level encoding. This module integrates multiple sensors and improves data processing speed through hardware acceleration chips and parallel computing architecture. Edge model training module: This module uses preprocessed data to locally train the object recognition model, employing an improved adaptive momentum estimation optimization algorithm and meta-learning mechanism. This module integrates a deep learning framework and meta-learning library, supporting a variety of model architectures and task settings. During training, it dynamically adjusts the learning rate based on the characteristics and distribution of local data. Data encryption and transmission module: Uses a chaos map-based encryption algorithm to encrypt locally trained model parameters, and uploads the encrypted model parameters and metadata files containing data complexity indexes to the federated learning server, using communication protocols and blockchain technology to transmit data; Federated Learning Server Module: Receives encrypted model parameters from multiple edge devices and aggregates them using a federated aggregation algorithm based on a weighted attention mechanism. This module integrates blockchain nodes and a deep learning computing cluster. During the aggregation process, attention weights are assigned to model parameters on different edge devices based on the information in the metadata file and dynamically adjusted. Model decryption and distribution module: This module uses the inverse chaotic mapping operation corresponding to the encryption process to decrypt the model parameters, distributes the decrypted global model parameters to each edge device, and generates and distributes the model update suggestion file. Target recognition module: Uses the updated local model to identify newly acquired target images and sensor data, adopts a multimodal fusion target recognition method and a knowledge graph-based post-processing method. This module integrates a multimodal data processing engine and a knowledge graph reasoning engine, and outputs target information based on the recognition results.

8. The system of the edge computing target recognition method of federated learning collaboration according to claim 7 is characterized in that: It also includes a model adaptive optimization module, which adaptively optimizes the local model after each global model distribution based on changes in local data and the requirements of the recognition task. It adopts an optimization strategy based on dynamic programming and an adaptive adjustment coefficient. This module integrates an optimization algorithm library and a data analysis engine to monitor model performance and data distribution changes in real time.

9. The system of the edge computing target recognition method of federated learning collaboration according to claim 7 is characterized in that: It also includes an abnormal data detection and processing module. During the data preprocessing stage, this module uses an anomaly detection algorithm based on a deep autoencoder to detect abnormal data. For the detected abnormal data, a data repair method based on a generative adversarial network is adopted. This module integrates a deep autoencoder model and a generative adversarial network model, and performs training and inference through a deep learning framework.

10. The system of the edge computing target recognition method of federated learning collaboration according to claim 7 is characterized in that: It also includes an incentive and trust management module, which introduces a blockchain-based trust mechanism and a game theory-based incentive mechanism. It records model parameters and metadata files through the blockchain, verifies the authenticity and integrity of the data, and gives corresponding rewards based on the contribution of edge devices. This module integrates blockchain nodes and game theory algorithm libraries.

Images