An edge computing method and system for routers, gateways and cameras

Abstract

This invention provides an edge computing method and system for routers, gateways, and cameras in the field of Internet of Things (IoT) and edge computing technology. The method includes steps S1: the edge computing device collects multimodal data and preprocesses it to obtain preprocessed data; step S2: the input computing task is acquired, multidimensional indicators are collected in real time, a heuristic algorithm is used to solve the constructed multi-objective decision function, and an optimal task allocation strategy is dynamically generated based on the multidimensional indicators; step S3: the preprocessed data is synchronized based on the task type to execute the computing task locally or in the cloud; simultaneously, a two-level collaborative caching mechanism between the edge and cloud is established to cache frequently accessed preprocessed data locally on the edge computing device and synchronize all preprocessed data to the cloud when idle. The advantages of this invention are: it greatly improves the overall system performance, resource utilization efficiency, and collaborative intelligence level.

Description

Technical Field

[0001] This invention relates to the fields of Internet of Things and edge computing technology, and in particular to an edge computing method and system for routers, gateways and cameras. Background Technology

[0002] With the rapid development of IoT, AI, and big data technologies, smart home, smart security, and smart city applications are becoming increasingly widespread. Routers, gateways, and cameras, as key entry points and nodes in the IoT system, directly determine the performance and user experience of the entire system through their level of intelligence. These devices not only need to perform basic data collection and transmission functions, but are also expected to achieve local intelligent analysis, real-time response, and collaborative decision-making.

[0003] In traditional architectures, these devices typically employ a pure cloud computing model. This involves terminal devices (such as cameras) collecting raw data (e.g., video streams) and uploading it all over the network to a remote cloud server for storage, computation, and analysis. While this model can leverage the powerful computing and storage resources of the cloud to handle complex tasks, its inherent limitations become increasingly apparent when dealing with massive amounts of real-time IoT data. 1. High latency and network bandwidth pressure: Transmitting all data indiscriminately to the cloud leads to significant network latency, making it difficult to meet the millisecond-level response requirements of applications such as real-time facial recognition and abnormal behavior alarms. At the same time, high-bandwidth data such as high-definition video streams continuously consume large amounts of network bandwidth, increasing operating costs.

[0004] 2. Reliability depends on the network: When the network connection is unstable or interrupted, the cloud service will completely fail, causing the system to malfunction and lacking local autonomy.

[0005] 3. Data security and privacy risks: Sensitive data (such as home video) faces potential leakage risks during transmission and cloud storage, posing a challenge to user privacy protection.

[0006] To overcome the latency and bandwidth issues of cloud computing, edge computing technology has emerged. The core idea of ​​edge computing is to push computing and storage capabilities down to the network edge, processing data closer to the data source. Applying edge computing to devices such as routers, gateways, and cameras allows for preliminary data processing and decision-making locally, uploading only necessary results or summary information to the cloud, thereby effectively reducing latency, alleviating bandwidth pressure, and improving system reliability.

[0007] However, relying solely on edge computing also has limitations. Its computing power and storage space are limited, making it unsuitable for complex model training and big data analysis tasks. Therefore, combining the powerful capabilities of cloud computing with the low-latency advantages of edge computing to build a "cloud-edge" collaborative architecture has become the mainstream trend in current technological development.

[0008] Although this concept has been proposed, existing cloud-edge collaboration solutions still have many shortcomings in practical implementation, especially in applications involving heterogeneous devices such as routers, gateways, and cameras. 1. Rigid task allocation strategy: Existing systems often use preset static rules to divide tasks, which cannot be intelligently adjusted according to dynamic factors such as network status, real-time device load, and task urgency. This results in low system resource utilization efficiency and an inability to achieve the optimal balance between latency and computational accuracy.

[0009] 2. Low efficiency of cloud-edge data synchronization: The data synchronization mechanism between edge devices and the cloud is not efficient enough. Full synchronization or unoptimized transmission methods result in high synchronization latency and wasted bandwidth, affecting the real-time performance of cloud data analysis and the accuracy of edge-side decision-making.

[0010] 3. Lack of security collaboration mechanism: Existing solutions often focus on single-point security (such as transmission encryption or cloud security), lacking a unified security architecture that runs through the entire data lifecycle of cloud, edge, and terminal, and have weaknesses in identity authentication, access control, and privacy protection.

[0011] 4. Imperfect collaborative working mechanism: There is a lack of efficient collaborative mechanism between the cloud and the edge. For example, the cloud cannot guide the updating of the edge model in real time, and the edge side cannot effectively use the global knowledge of the cloud for decision-making, so the complementary advantages of "cloud-edge" collaboration cannot be fully utilized.

[0012] Therefore, how to provide an edge computing method and system for routers, gateways, and cameras to improve overall system performance, resource utilization efficiency, and collaborative intelligence has become an urgent technical problem to be solved. Summary of the Invention

[0013] The technical problem to be solved by the present invention is to provide an edge computing method and system for routers, gateways and cameras, so as to improve the overall system performance, resource utilization efficiency and collaborative intelligence level.

[0014] In a first aspect, the present invention provides an edge computing method for routers, gateways, and cameras, comprising the following steps: Step S1: The edge computing device collects heterogeneous multimodal data, identifies the data type of the multimodal data, and calls the corresponding preprocessing pipeline to preprocess the multimodal data based on the identified data type to obtain preprocessed data; the edge computing device is a router, gateway, or camera; Step S2: The edge computing device acquires the input computing task and collects multi-dimensional indicators in real time, including the latency sensitivity, computational complexity, data privacy level, current computing load of the edge computing device, available bandwidth, and cloud resource cost of the computing task. It constructs a multi-objective decision function with the total system latency, energy consumption, and cost as optimization objectives, and uses a heuristic algorithm to solve the multi-objective decision function. Based on the multi-dimensional indicators, it dynamically generates the optimal task allocation strategy to classify the computing task into edge-exclusive tasks, cloud-exclusive tasks, and cloud-edge collaborative tasks. Step S3: The edge computing device adopts a content-addressable differential synchronization mechanism to generate a data fingerprint from the preprocessed data using a hash algorithm. The preprocessed data is synchronized based on the task type to execute the computing task locally or in the cloud. At the same time, a two-level collaborative caching mechanism between the edge and the cloud is established to cache the frequently accessed preprocessed data locally on the edge computing device. All the preprocessed data is synchronized to the cloud when idle. Before synchronization, the compression algorithm and compression ratio are dynamically selected based on network bandwidth and data type, and only the data whose data fingerprint has changed is synchronized. Step S4: The edge computing device monitors historical data and uses a pre-deployed time series prediction model to predict the resource load fluctuations of the edge computing device and the cloud in future periods, and implements an elastic resource scheduling strategy based on the prediction results. Step S5: The edge computing device and the cloud are all based on the zero-trust principle, implementing continuous identity authentication and dynamic access control; end-to-end encryption using national cryptographic algorithms is adopted at the data transmission link layer, and blockchain technology is used to store key operation logs; for sensitive data in the stored preprocessed data, homomorphic encryption or differential privacy processing is performed, so that the sensitive data can still be calculated and analyzed in the encrypted state; Step S6: The cloud uses a federated learning framework to aggregate model parameter updates from multiple edge computing devices for the time series prediction model, generating a globally optimized model. The globally optimized model is then distributed to the edge computing devices in an incremental update manner, and the edge computing devices use local data to fine-tune the time series prediction model.

[0015] Furthermore, in step S1, the multimodal data includes at least network traffic data, video streams, audio streams, and sensor data; The preprocessing includes at least protocol parsing and feature extraction of network traffic data, keyframe extraction and background modeling of video streams, speech activity detection and noise reduction of audio streams, data calibration, filtering, smoothing and outlier detection of sensor data; The data calibration employs a linear or nonlinear compensation algorithm based on a standard reference value; the filtering and smoothing employs a Kalman filter or a moving average filter algorithm; and the outlier detection employs a detection algorithm based on statistical distribution or an isolated forest. In step S2, the multi-objective decision function transforms the multi-objective problem into a single-objective problem by using a weighted summation method or an ε-constraint method; the heuristic algorithm is a particle swarm optimization algorithm or a Q-learning reinforcement learning algorithm.

[0016] Furthermore, in step S3, the content-address-based differential synchronization mechanism specifically involves: using a Merkle tree structure to organize preprocessed data, and quickly locating differences by comparing the data fingerprints at the tree root; The secondary collaborative caching mechanism uses a least recently used strategy to replace data and dynamically adjusts the cached content based on a data popularity model.

[0017] Furthermore, in step S4, the time series prediction model is an ARIMA model or an LSTM neural network model; The elastic resource scheduling strategy specifically includes: preheating and migrating computing resources from the cloud or adjacent lightly loaded edge computing devices before the predicted peak edge computing load; staggering non-real-time computing tasks; and forming a resource pool among edge computing devices to achieve load balancing. The resource pool uses virtualization technology to create lightweight container instances on edge computing devices, enabling rapid migration and isolation of computing tasks.

[0018] Furthermore, in step S5, the continuous identity authentication includes multi-factor authentication based on device fingerprints and user behavior profiles; The dynamic access control is attribute-based access control, and its authorization policy is dynamically adjusted according to real-time risk scores. In step S6, the federated learning framework uses secure multi-party computation or homomorphic encryption to protect the privacy of the model parameter updates of the edge computing device before parameter aggregation. The incremental update employs model quantization and pruning techniques to reduce the amount of data distributed.

[0019] Secondly, the present invention provides an edge computing system for routers, gateways, and cameras, comprising the following modules: A multimodal data acquisition module is used to acquire heterogeneous multimodal data from edge computing devices, identify the data type of the multimodal data, and call the corresponding preprocessing pipeline based on the identified data type to preprocess the multimodal data to obtain preprocessed data; the edge computing device is a router, gateway, or camera; The task allocation strategy generation module is used to acquire the input computing tasks of the edge computing device, collect multi-dimensional indicators in real time, including the latency sensitivity, computational complexity, data privacy level, current computing load of the edge computing device, available bandwidth, and cloud resource cost of the computing tasks, construct a multi-objective decision function with the total system latency, energy consumption and cost as optimization objectives, and solve the multi-objective decision function using a heuristic algorithm. Based on the multi-dimensional indicators, the optimal task allocation strategy is dynamically generated to classify the computing tasks into edge-exclusive tasks, cloud-exclusive tasks and cloud-edge collaborative tasks. The computing task execution module is used by the edge computing device to adopt a content-addressable differential synchronization mechanism to generate a data fingerprint from the preprocessed data using a hash algorithm, and synchronize the preprocessed data based on the task type to execute the computing task locally or in the cloud; at the same time, a two-level collaborative caching mechanism between the edge and the cloud is established to cache the frequently accessed preprocessed data locally on the edge computing device, and synchronize all the preprocessed data to the cloud when idle. Before synchronization, the compression algorithm and compression ratio are dynamically selected according to network bandwidth and data type, and only the data whose data fingerprint has changed is synchronized. The elastic resource scheduling module is used by edge computing devices to monitor historical data, use pre-deployed time series prediction models to predict the resource load fluctuations of edge computing devices and the cloud in future periods, and implement elastic resource scheduling strategies based on the prediction results. The security management module implements continuous identity authentication and dynamic access control based on the zero-trust principle for both edge computing devices and the cloud; it uses end-to-end encryption with national cryptographic algorithms at the data transmission link layer and combines blockchain technology to store key operation logs; it performs homomorphic encryption or differential privacy processing on sensitive data in the stored preprocessed data, so that the sensitive data can still be calculated and analyzed in the encrypted state. The federated learning module is used in the cloud to aggregate model parameter updates of the time series prediction model from multiple edge computing devices using a federated learning framework, and generate a globally optimized model. The globally optimized model is then distributed to the edge computing devices in an incremental update manner, and the edge computing devices use local data to fine-tune the time series prediction model.

[0020] Furthermore, in the multimodal data acquisition module, the multimodal data includes at least network traffic data, video streams, audio streams, and sensor data; The preprocessing includes at least protocol parsing and feature extraction of network traffic data, keyframe extraction and background modeling of video streams, speech activity detection and noise reduction of audio streams, data calibration, filtering, smoothing and outlier detection of sensor data; The data calibration employs a linear or nonlinear compensation algorithm based on a standard reference value; the filtering and smoothing employs a Kalman filter or a moving average filter algorithm; and the outlier detection employs a detection algorithm based on statistical distribution or an isolated forest. In the task allocation strategy generation module, the multi-objective decision function transforms the multi-objective problem into a single-objective problem by using a weighted summation method or an ε-constraint method; the heuristic algorithm is a particle swarm optimization algorithm or a Q-learning reinforcement learning algorithm.

[0021] Furthermore, in the computation task execution module, the content-addressable differential synchronization mechanism specifically involves: using a Merkle tree structure to organize preprocessed data, and quickly locating differences by comparing the data fingerprints at the tree root; The secondary collaborative caching mechanism uses a least recently used strategy to replace data and dynamically adjusts the cached content based on a data popularity model.

[0022] Furthermore, in the elastic resource scheduling module, the time series prediction model is an ARIMA model or an LSTM neural network model; The elastic resource scheduling strategy specifically includes: preheating and migrating computing resources from the cloud or adjacent lightly loaded edge computing devices before the predicted peak edge computing load; staggering non-real-time computing tasks; and forming a resource pool among edge computing devices to achieve load balancing. The resource pool uses virtualization technology to create lightweight container instances on edge computing devices, enabling rapid migration and isolation of computing tasks.

[0023] Furthermore, in the security management module, the continuous identity authentication includes multi-factor authentication based on device fingerprints and user behavior profiles; The dynamic access control is attribute-based access control, and its authorization policy is dynamically adjusted according to real-time risk scores. In the federated learning module, the federated learning framework uses secure multi-party computation or homomorphic encryption to protect the privacy of the model parameter updates of the edge computing device before parameter aggregation; The incremental update employs model quantization and pruning techniques to reduce the amount of data distributed.

[0024] The advantages of this invention are: 1. Preprocessing of heterogeneous multimodal data collected by edge computing devices yields preprocessed data, which is then used to obtain input computing tasks. Multidimensional indicators, including those related to the computing tasks, are collected in real-time. A multi-objective decision function is constructed, with total system latency, energy consumption, and cost as optimization objectives. A heuristic algorithm is used to solve this multi-objective decision function, and an optimal task allocation strategy is dynamically generated based on the multidimensional indicators. This classifies computing tasks into edge-exclusive tasks, cloud-exclusive tasks, and cloud-edge collaborative tasks. Next, the edge computing device employs a content-addressed differential synchronization mechanism to generate data fingerprints from the preprocessed data using a hash algorithm. The preprocessed data is synchronized based on task type to execute the computing tasks locally or in the cloud. A two-level collaborative caching mechanism between the edge and cloud is also established. By monitoring historical data, the edge computing device uses a time-series prediction model to predict resource load fluctuations in the edge computing device and cloud over future periods. Based on the prediction results, an elastic resource scheduling strategy is implemented. Both the edge computing device and the cloud operate on a zero-trust principle, implementing continuous authentication and dynamic access control. National cryptographic algorithms are used at the data transmission link layer. The system employs end-to-end encryption and combines blockchain technology to store key operation logs. Sensitive data in the stored preprocessed data undergoes homomorphic encryption or differential privacy processing. A federated learning framework is used in the cloud to aggregate model parameter updates from multiple edge computing devices, generating a globally optimized model. This globally optimized model is then distributed to edge computing devices in an incremental update manner, allowing the devices to fine-tune the time series prediction model using local data. This dynamic multi-objective decision-making mechanism optimizes the allocation of computing tasks between the cloud and edge in real time, reducing latency and energy consumption from the source. Content-addressed differential synchronization and intelligent caching significantly improve data synchronization efficiency and save bandwidth resources. A security architecture integrating zero-trust principles, national cryptographic algorithms, and blockchain technologies provides comprehensive security throughout the entire data lifecycle. Furthermore, prediction-based elastic resource scheduling and federated learning-driven model collaborative updates enable proactive optimization of cloud and edge resources and continuous evolution of intelligence. These interconnected steps greatly improve overall system performance, resource utilization efficiency, and collaborative intelligence.

[0025] 2. By automatically identifying heterogeneous multimodal data (such as network traffic, video streams, audio streams, and sensor data) and calling the corresponding preprocessing pipelines (such as protocol parsing, keyframe extraction, and data calibration), the efficiency and accuracy of data processing are significantly improved. This data type-based adaptive preprocessing mechanism reduces manual intervention, ensures data quality, and provides standardized input for subsequent computing tasks, thereby reducing the computing load on edge devices and improving the overall system response speed. It is particularly suitable for resource-constrained edge environments such as routers, gateways, and cameras.

[0026] 3. By constructing a multi-objective decision function (optimizing latency, energy consumption, and cost) and using heuristic algorithms (such as particle swarm optimization or Q-learning) to dynamically generate task allocation strategies, computing tasks are divided into edge-exclusive, cloud-exclusive, or cloud-edge collaborative types. This decision mechanism based on real-time multi-dimensional indicators (such as latency sensitivity, computing load, and bandwidth) can adaptively balance computing resources, avoid resource bottlenecks, thereby maximizing system efficiency, reducing operating costs, and ensuring low-latency processing of critical tasks (such as real-time video analytics), thus improving the flexibility and economy of edge computing systems.

[0027] 4. By adopting a content-addressable differential synchronization mechanism (such as a Merkle tree structure) and a two-level collaborative caching mechanism, data fingerprints are generated and only changed data is synchronized, which significantly reduces the amount of data transmission and bandwidth consumption. At the same time, the cached content is dynamically adjusted according to the data popularity (such as using an LRU strategy), ensuring the local availability of frequently accessed data and reducing cloud dependence. This mechanism not only improves the efficiency of data synchronization, but also reduces network latency and energy consumption, making it particularly suitable for bandwidth-constrained edge scenarios, such as collaborative work between IoT devices.

[0028] 5. Predict resource load fluctuations using time series forecasting models (such as ARIMA or LSTM) and implement elastic resource scheduling strategies (such as resource preheating, off-peak scheduling, and resource pooling). This predictive approach allows the system to adjust resources in advance before load peaks to avoid performance degradation. It also enables rapid migration and isolation of computing tasks through virtualization technology, improving resource utilization and system reliability, reducing the risk of service interruption, and supporting the dynamic expansion of edge devices. It is suitable for environments with high volatility (such as smart homes or security monitoring).

[0029] 6. Based on the zero-trust principle, it implements continuous identity authentication, dynamic access control, end-to-end encryption (using national cryptographic algorithms), and blockchain notarization, and performs homomorphic encryption or differential privacy processing on sensitive data. This multi-layered security framework ensures the confidentiality and integrity of data during transmission, storage, and computation, preventing unauthorized access and tampering. At the same time, privacy protection technologies allow data analysis in an encrypted state, balancing data utility and security, complying with modern data regulations, and enhancing user trust. It is especially suitable for edge devices that process sensitive information (such as personal videos or network traffic).

[0030] 7. By adopting a federated learning framework, model parameter updates from multiple edge devices are aggregated (privacy is protected by secure multi-party computation or homomorphic encryption), generating a globally optimized model and distributing it to edge devices in an incremental update manner (such as model quantization and pruning). This method achieves distributed learning without exposing the original data, reduces communication overhead, and improves the adaptability and accuracy of the model through local fine-tuning. It supports continuous learning and optimization, improves the performance of edge intelligent systems, and reduces cloud dependence. It is suitable for scenarios that require real-time adaptation (such as adaptive network management or intelligent camera behavior analysis).

[0031] 8. By integrating multimodal data adaptive preprocessing, task allocation based on dynamic optimization of multidimensional indicators, efficient differential synchronization and caching mechanisms, predictive elastic resource scheduling, zero-trust security framework, and privacy-preserving federated learning, a high-efficiency, secure, and adaptive edge computing system was constructed. While significantly reducing latency, energy consumption, and cost, it maximized resource utilization, ensured data security and reliability, and continuously optimized system performance, comprehensively improving the intelligence level and collaborative efficiency of edge devices such as routers, gateways, and cameras. Attached Figure Description

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0033] Figure 1 This is a flowchart of an edge computing method for routers, gateways, and cameras according to the present invention.

[0034] Figure 2 This is a schematic diagram of the structure of an edge computing system for routers, gateways, and cameras according to the present invention. Detailed Implementation

[0035] The technical solution in this application embodiment follows the following general approach: A dynamic multi-objective decision-making mechanism optimizes the allocation of computing tasks between the cloud and edge in real time, reducing latency and energy consumption from the source; content-addressed differential synchronization and intelligent caching significantly improve data synchronization efficiency and save bandwidth resources; a security architecture integrating zero-trust principles, national cryptographic algorithms, and blockchain technologies provides comprehensive security throughout the entire data lifecycle; and proactive optimization of cloud-edge resources and continuous evolution of intelligence are achieved through prediction-based elastic resource scheduling and federated learning-driven model collaborative updates. These interconnected steps enhance overall system performance, resource utilization efficiency, and collaborative intelligence.

[0036] The implementation environment of this invention is a typical "cloud-edge-device" IoT architecture. Edge computing devices refer to devices such as routers, smart gateways, or smart cameras with certain computing, storage, and networking capabilities. Their hardware platforms are typically based on ARM or x86 architecture processors and run lightweight operating systems (such as OpenWRT, custom Linux-based systems, etc.). The cloud refers to one or more server clusters located in remote data centers, providing powerful computing, storage, and model training capabilities. The cloud and edge are connected via wired (such as Ethernet) or wireless (such as Wi-Fi, 5G) networks.

[0037] Please refer to Figures 1 to 2 As shown, a preferred embodiment of the edge computing method for routers, gateways, and cameras of the present invention includes the following steps: Step S1: The edge computing device collects heterogeneous multimodal data, identifies the data type of the multimodal data, and calls the corresponding preprocessing pipeline to preprocess the multimodal data based on the identified data type to obtain preprocessed data; the edge computing device is a router, gateway, or camera; Step S2: The edge computing device acquires the input computing task and collects multi-dimensional indicators in real time, including the latency sensitivity, computational complexity, data privacy level, current computing load of the edge computing device, available bandwidth, and cloud resource cost of the computing task. It constructs a multi-objective decision function with the total system latency, energy consumption, and cost as optimization objectives, and uses a heuristic algorithm to solve the multi-objective decision function. Based on the multi-dimensional indicators, it dynamically generates the optimal task allocation strategy to classify the computing task into edge-exclusive tasks, cloud-exclusive tasks, and cloud-edge collaborative tasks. Step S3: The edge computing device adopts a content-addressable differential synchronization mechanism to generate a data fingerprint from the preprocessed data using a hash algorithm. The preprocessed data is synchronized based on the task type to execute the computing task locally or in the cloud. At the same time, a two-level collaborative caching mechanism between the edge and the cloud is established to cache the frequently accessed preprocessed data locally on the edge computing device. All the preprocessed data is synchronized to the cloud when idle. Before synchronization, the compression algorithm and compression ratio are dynamically selected based on network bandwidth and data type, and only the data whose data fingerprint has changed is synchronized. Step S4: The edge computing device monitors historical data and uses a pre-deployed time series prediction model to predict the resource load fluctuations of the edge computing device and the cloud in future periods, and implements an elastic resource scheduling strategy based on the prediction results. Step S5: The edge computing device and the cloud are all based on the zero-trust principle, implementing continuous identity authentication and dynamic access control; end-to-end encryption using national cryptographic algorithms is adopted at the data transmission link layer, and blockchain technology is used to store key operation logs; for sensitive data in the stored preprocessed data, homomorphic encryption or differential privacy processing is performed, so that the sensitive data can still be calculated and analyzed in the encrypted state; Step S6: The cloud uses a federated learning framework to aggregate model parameter updates from multiple edge computing devices for the time series prediction model, generating a globally optimized model. The globally optimized model is then distributed to the edge computing devices in an incremental update manner, and the edge computing devices use local data to fine-tune the time series prediction model.

[0038] In step S1, the multimodal data includes at least network traffic data, video streams, audio streams, and sensor data; The preprocessing includes at least protocol parsing and feature extraction of network traffic data, keyframe extraction and background modeling of video streams, speech activity detection and noise reduction of audio streams, data calibration, filtering, smoothing and outlier detection of sensor data; The data calibration employs a linear or nonlinear compensation algorithm based on a standard reference value; the filtering and smoothing employs a Kalman filter or a moving average filter algorithm; and the outlier detection employs a detection algorithm based on statistical distribution or an isolated forest. The multimodal data specifically includes: Video stream: Real-time high-definition video data (such as 1080P resolution) acquired through image sensors.

[0039] Audio stream: Ambient audio data captured via the built-in microphone.

[0040] Sensor data: Connectable third-party sensor data, such as status data from door and window magnetic sensors and human infrared sensors accessed via Zigbee or Z-Wave protocols.

[0041] Edge computing devices are pre-installed with a data type recognition module. This module automatically identifies the type of input data by analyzing the protocol header of the data packet, the data format (such as video streams, which are usually in RTP / H.264 format, and sensor data, which are in a specific JSON format), or the hardware interface of the data source.

[0042] Based on the identified data type, the system invokes the corresponding preprocessing pipeline: For video streams, the preprocessing pipeline performs the following operations: 1. Keyframe extraction: Using inter-frame difference or deep learning-based object detection models (such as the lightweight YOLO model), keyframe images containing significant motion or specific targets (such as people or vehicles) are extracted from the video stream, greatly reducing the amount of data that needs to be processed.

[0043] 2. Background Modeling: A Gaussian Mixture Model (GMM) is used to model the scene background for subsequent moving target detection and abnormal behavior analysis.

[0044] For audio streams, the preprocessing pipeline executes as follows: 1. Voice Activity Detection (VAD): Uses energy and spectral entropy-based detection algorithms to distinguish between environmental noise and valid speech segments, and only performs subsequent processing on the speech segments.

[0045] 2. Noise reduction: Use spectral subtraction or deep learning-based speech enhancement algorithms to improve the signal-to-noise ratio of the speech signal.

[0046] For sensor data, the preprocessing pipeline performs the following: 1. Data calibration: For example, for temperature sensor data, linear compensation is performed using reference values ​​based on standard thermometers to eliminate system errors.

[0047] 2. Filtering and smoothing: The moving average filter is used to smooth short-term fluctuations in the data.

[0048] 3. Outlier detection: Statistical methods based on standard deviation (such as the 3σ criterion) are used to identify and remove outlier data points that deviate significantly from the normal range.

[0049] After preprocessing, the original multimodal data is transformed into standardized, high-quality preprocessed data, providing input for subsequent computational tasks.

[0050] In step S2, the multi-objective decision function transforms the multi-objective problem into a single-objective problem by using a weighted summation method or an ε-constraint method; the heuristic algorithm is a particle swarm optimization algorithm or a Q-learning reinforcement learning algorithm.

[0051] In practice, users set up a computing task through a mobile app to "identify strangers at the door in real time and issue an alarm," and this computing task is then sent to edge computing devices.

[0052] Edge computing devices collect multi-dimensional metrics related to the computing task in real time, including: Task attributes: latency sensitivity (high), computational complexity (medium, involving face recognition model inference), data privacy level (high, involving home internal images).

[0053] System status: Current CPU utilization of the camera (e.g., 60%), available memory, and current uplink network bandwidth (e.g., 5Mbps).

[0054] Cloud cost: The unit price for calling the cloud-based facial recognition API.

[0055] The system constructs a multi-objective decision function, and an example of its formal expression is as follows: Minimize F = w1 * Total_Delay + w2 * Total_Energy + w3 * Total_Cost Total_Delay includes edge processing latency and (if required) cloud transmission and processing latency; Total_Energy mainly represents the energy consumption of edge devices; Total_Cost is the cost incurred from using cloud services. w1, w2, and w3 are weighting coefficients that can be dynamically adjusted according to application scenario preferences (e.g., w1 has a higher weight in security scenarios).

[0056] A heuristic algorithm (such as Particle Swarm Optimization, PSO) is used to solve the above function. The PSO algorithm simulates the foraging behavior of a flock of birds and finds the optimal solution through iteration. In this embodiment, each "particle" represents a task allocation scheme (such as "processing all at the edge", "processing all in the cloud", "face detection at the edge, recognition in the cloud", etc.). The algorithm evaluates the objective function value F of each scheme and finally converges to the scheme with the smallest F value, which is the optimal task allocation strategy.

[0057] Based on this strategy, the "stranger recognition" task is classified as an edge-exclusive task (because of high latency and privacy requirements, and the camera has local inference capabilities), while the non-real-time large-scale face model training task is classified as a cloud-exclusive task.

[0058] In step S3, the content-addressable differential synchronization mechanism specifically involves: using a Merkle tree structure to organize preprocessed data and quickly locating differences by comparing the data fingerprints at the tree root; The secondary collaborative caching mechanism uses a least recently used strategy to replace data and dynamically adjusts the cached content based on a data popularity model.

[0059] To support computing tasks that may require cloud collaboration, edge computing devices employ a content-addressed differential synchronization mechanism. This mechanism preprocesses data, generating unique data fingerprints using the SHA-256 hash algorithm. Data is organized locally in a Merkle tree structure, where leaf nodes are data block fingerprints and parent nodes are hashes of their child node fingerprints. When synchronization with the cloud is needed, simply comparing the root hash values ​​of the local and cloud Merkle trees quickly determines data consistency. If inconsistencies are found, the system traverses the tree structure to precisely locate and synchronize the few changed data blocks, significantly improving efficiency.

[0060] At the same time, establish an edge-cloud two-level collaborative caching mechanism: Edge caching: Frequently accessed data (such as video clips of alarm events in the last 24 hours) are cached in the local storage of the smart camera and managed using the least recently used (LRU) strategy.

[0061] Cloud synchronization: During off-peak network hours at night, all pre-processed data is synchronized to the cloud for backup and macro-analysis. Before synchronization, a high compression ratio algorithm (such as H.265) is dynamically selected based on the current bandwidth (e.g., when bandwidth is low) and the data type (e.g., video data), and only data whose data fingerprints have changed is synchronized.

[0062] In step S4, the time series prediction model is an ARIMA model or an LSTM neural network model. The elastic resource scheduling strategy specifically includes: preheating and migrating computing resources from the cloud or adjacent lightly loaded edge computing devices before the predicted peak edge computing load; staggering non-real-time computing tasks; and forming a resource pool among edge computing devices to achieve load balancing. The resource pool uses virtualization technology to create lightweight container instances on edge computing devices, enabling rapid migration and isolation of computing tasks.

[0063] Edge computing devices continuously monitor historical data on their CPU, memory, and other resource usage. They then use pre-deployed time-series forecasting models (such as LSTM neural networks) to predict resource load fluctuations over a future period (e.g., the next hour). For example, they might predict a significant increase in computing load during peak household activity at 7 PM.

[0064] Based on the prediction results, implement an elastic resource scheduling strategy: Resource preheating and migration: Before the predicted load peak (7 pm), the system can preheat and migrate some computing tasks (such as audio event detection) from the cloud or from a less loaded smart speaker (as an auxiliary edge node) in the same home network to share the load of the camera.

[0065] Off-peak scheduling: Schedule non-urgent tasks (such as device log analysis) to be executed during periods of network downtime (such as 2 a.m.).

[0066] Resource pooling: By using virtualization technologies such as Docker containers, the computing resources of multiple edge devices (cameras, gateways, smart TVs) in the home are virtualized into a resource pool to achieve load balancing.

[0067] In step S5, the continuous identity authentication includes multi-factor authentication based on device fingerprints and user behavior profiles; The dynamic access control is attribute-based access control, and its authorization policy is dynamically adjusted according to real-time risk scores. This invention follows the zero-trust principle, which means that it never trusts any device or user by default.

[0068] Continuous authentication: Authentication between edge computing devices and the cloud is not just an initial connection, but a continuous process. For example, multi-factor authentication can be performed by verifying the device's unique hardware fingerprint (such as TPM chip information) and combining it with its behavioral profile (such as normal network access patterns).

[0069] Dynamic access control: Employs attribute-based access control (ABAC). For example, when an access request is detected to originate from an abnormal geographical location (attribute), the real-time risk score increases, and the authorization policy is dynamically adjusted, potentially denying access even if authentication is successful.

[0070] End-to-end encryption and evidence storage: Data transmission between edge computing devices and the cloud is encrypted end-to-end using national cryptographic algorithms (such as SM4). All critical operations (such as user login and configuration changes) are logged after hashing and recorded on a private blockchain to ensure immutability.

[0071] Privacy-preserving computation: Sensitive facial feature data stored in the cloud is homomorphically encrypted, allowing the cloud to perform matching calculations on the encrypted data directly without decryption, thereby protecting user privacy.

[0072] In step S6, the federated learning framework uses secure multi-party computation or homomorphic encryption to protect the privacy of the model parameter updates of the edge computing device before parameter aggregation. The incremental update employs model quantization and pruning techniques to reduce the amount of data distributed.

[0073] A federated learning framework (such as the FATE framework) runs in the cloud. Periodically (e.g., weekly), the cloud aggregates parameter updates (gradient information) from time series prediction models (LSTM models) from thousands of edge computing devices. Before aggregation, secure multi-party computation is used to protect the privacy of the parameter updates from each device, ensuring that the cloud cannot backtrack the original data.

[0074] After aggregation, a more accurate and better generalized globally optimized model is generated in the cloud. Then, the differences between this model and the old version (i.e., incremental updates) are quantized (converting 32-bit floating-point numbers to 8-bit integers) and pruned (removing unimportant neural connections). The model size is then compressed and distributed to various edge computing devices. These edge computing devices use newly generated local data to fine-tune the distributed model, making it better suited to their local environment. In this way, the intelligence level of all edge computing devices can continuously and collaboratively evolve while protecting data privacy.

[0075] A preferred embodiment of the edge computing system for routers, gateways, and cameras of the present invention includes the following modules: A multimodal data acquisition module is used to acquire heterogeneous multimodal data from edge computing devices, identify the data type of the multimodal data, and call the corresponding preprocessing pipeline based on the identified data type to preprocess the multimodal data to obtain preprocessed data; the edge computing device is a router, gateway, or camera; The task allocation strategy generation module is used to acquire the input computing tasks of the edge computing device, collect multi-dimensional indicators in real time, including the latency sensitivity, computational complexity, data privacy level, current computing load of the edge computing device, available bandwidth, and cloud resource cost of the computing tasks, construct a multi-objective decision function with the total system latency, energy consumption and cost as optimization objectives, and solve the multi-objective decision function using a heuristic algorithm. Based on the multi-dimensional indicators, the optimal task allocation strategy is dynamically generated to classify the computing tasks into edge-exclusive tasks, cloud-exclusive tasks and cloud-edge collaborative tasks. The computing task execution module is used by the edge computing device to adopt a content-addressable differential synchronization mechanism to generate a data fingerprint from the preprocessed data using a hash algorithm, and synchronize the preprocessed data based on the task type to execute the computing task locally or in the cloud; at the same time, a two-level collaborative caching mechanism between the edge and the cloud is established to cache the frequently accessed preprocessed data locally on the edge computing device, and synchronize all the preprocessed data to the cloud when idle. Before synchronization, the compression algorithm and compression ratio are dynamically selected according to network bandwidth and data type, and only the data whose data fingerprint has changed is synchronized. The elastic resource scheduling module is used by edge computing devices to monitor historical data, use pre-deployed time series prediction models to predict the resource load fluctuations of edge computing devices and the cloud in future periods, and implement elastic resource scheduling strategies based on the prediction results. The security management module implements continuous identity authentication and dynamic access control based on the zero-trust principle for both edge computing devices and the cloud; it uses end-to-end encryption with national cryptographic algorithms at the data transmission link layer and combines blockchain technology to store key operation logs; it performs homomorphic encryption or differential privacy processing on sensitive data in the stored preprocessed data, so that the sensitive data can still be calculated and analyzed in the encrypted state. The federated learning module is used in the cloud to aggregate model parameter updates of the time series prediction model from multiple edge computing devices using a federated learning framework, and generate a globally optimized model. The globally optimized model is then distributed to the edge computing devices in an incremental update manner, and the edge computing devices use local data to fine-tune the time series prediction model.

[0076] In the multimodal data acquisition module, the multimodal data includes at least network traffic data, video streams, audio streams, and sensor data; The preprocessing includes at least protocol parsing and feature extraction of network traffic data, keyframe extraction and background modeling of video streams, speech activity detection and noise reduction of audio streams, data calibration, filtering, smoothing and outlier detection of sensor data; The data calibration employs a linear or nonlinear compensation algorithm based on a standard reference value; the filtering and smoothing employs a Kalman filter or a moving average filter algorithm; and the outlier detection employs a detection algorithm based on statistical distribution or an isolated forest. The multimodal data specifically includes: Video stream: Real-time high-definition video data (such as 1080P resolution) acquired through image sensors.

[0077] Audio stream: Ambient audio data captured via the built-in microphone.

[0078] Sensor data: Connectable third-party sensor data, such as status data from door and window magnetic sensors and human infrared sensors accessed via Zigbee or Z-Wave protocols.

[0079] Edge computing devices are pre-installed with a data type recognition module. This module automatically identifies the type of input data by analyzing the protocol header of the data packet, the data format (such as video streams, which are usually in RTP / H.264 format, and sensor data, which are in a specific JSON format), or the hardware interface of the data source.

[0080] Based on the identified data type, the system invokes the corresponding preprocessing pipeline: For video streams, the preprocessing pipeline performs the following operations: 1. Keyframe extraction: Using inter-frame difference or deep learning-based object detection models (such as the lightweight YOLO model), keyframe images containing significant motion or specific targets (such as people or vehicles) are extracted from the video stream, greatly reducing the amount of data that needs to be processed.

[0081] 2. Background Modeling: A Gaussian Mixture Model (GMM) is used to model the scene background for subsequent moving target detection and abnormal behavior analysis.

[0082] For audio streams, the preprocessing pipeline executes as follows: 1. Voice Activity Detection (VAD): Uses energy and spectral entropy-based detection algorithms to distinguish between environmental noise and valid speech segments, and only performs subsequent processing on the speech segments.

[0083] 2. Noise reduction: Use spectral subtraction or deep learning-based speech enhancement algorithms to improve the signal-to-noise ratio of the speech signal.

[0084] For sensor data, the preprocessing pipeline performs the following: 1. Data calibration: For example, for temperature sensor data, linear compensation is performed using reference values ​​based on standard thermometers to eliminate system errors.

[0085] 2. Filtering and smoothing: The moving average filter is used to smooth short-term fluctuations in the data.

[0086] 3. Outlier detection: Statistical methods based on standard deviation (such as the 3σ criterion) are used to identify and remove outlier data points that deviate significantly from the normal range.

[0087] After preprocessing, the original multimodal data is transformed into standardized, high-quality preprocessed data, providing input for subsequent computational tasks.

[0088] In the task allocation strategy generation module, the multi-objective decision function transforms the multi-objective problem into a single-objective problem by using a weighted summation method or an ε-constraint method; the heuristic algorithm is a particle swarm optimization algorithm or a Q-learning reinforcement learning algorithm.

[0089] In practice, users set up a computing task through a mobile app to "identify strangers at the door in real time and issue an alarm," and this computing task is then sent to edge computing devices.

[0090] Edge computing devices collect multi-dimensional metrics related to the computing task in real time, including: Task attributes: latency sensitivity (high), computational complexity (medium, involving face recognition model inference), data privacy level (high, involving home internal images).

[0091] System status: Current CPU utilization of the camera (e.g., 60%), available memory, and current uplink network bandwidth (e.g., 5Mbps).

[0092] Cloud cost: The unit price for calling the cloud-based facial recognition API.

[0093] The system constructs a multi-objective decision function, and an example of its formal expression is as follows: Minimize F = w1 * Total_Delay + w2 * Total_Energy + w3 * Total_Cost Total_Delay includes edge processing latency and (if required) cloud transmission and processing latency; Total_Energy mainly represents the energy consumption of edge devices; Total_Cost is the cost incurred from using cloud services. w1, w2, and w3 are weighting coefficients that can be dynamically adjusted according to application scenario preferences (e.g., w1 has a higher weight in security scenarios).

[0094] A heuristic algorithm (such as Particle Swarm Optimization, PSO) is used to solve the above function. The PSO algorithm simulates the foraging behavior of a flock of birds and finds the optimal solution through iteration. In this embodiment, each "particle" represents a task allocation scheme (such as "processing all at the edge", "processing all in the cloud", "face detection at the edge, recognition in the cloud", etc.). The algorithm evaluates the objective function value F of each scheme and finally converges to the scheme with the smallest F value, which is the optimal task allocation strategy.

[0095] Based on this strategy, the "stranger recognition" task is classified as an edge-exclusive task (because of high latency and privacy requirements, and the camera has local inference capabilities), while the non-real-time large-scale face model training task is classified as a cloud-exclusive task.

[0096] In the computation task execution module, the content-addressable differential synchronization mechanism specifically involves: using a Merkle tree structure to organize preprocessed data and quickly locating differences by comparing the data fingerprints at the tree root; The secondary collaborative caching mechanism uses a least recently used strategy to replace data and dynamically adjusts the cached content based on a data popularity model.

[0097] To support computing tasks that may require cloud collaboration, edge computing devices employ a content-addressed differential synchronization mechanism. This mechanism preprocesses data, generating unique data fingerprints using the SHA-256 hash algorithm. Data is organized locally in a Merkle tree structure, where leaf nodes are data block fingerprints and parent nodes are hashes of their child node fingerprints. When synchronization with the cloud is needed, simply comparing the root hash values ​​of the local and cloud Merkle trees quickly determines data consistency. If inconsistencies are found, the system traverses the tree structure to precisely locate and synchronize the few changed data blocks, significantly improving efficiency.

[0098] At the same time, establish an edge-cloud two-level collaborative caching mechanism: Edge caching: Frequently accessed data (such as video clips of alarm events in the last 24 hours) are cached in the local storage of the smart camera and managed using the least recently used (LRU) strategy.

[0099] Cloud synchronization: During off-peak network hours at night, all pre-processed data is synchronized to the cloud for backup and macro-analysis. Before synchronization, a high compression ratio algorithm (such as H.265) is dynamically selected based on the current bandwidth (e.g., when bandwidth is low) and the data type (e.g., video data), and only data whose data fingerprints have changed is synchronized.

[0100] In the elastic resource scheduling module, the time series prediction model is an ARIMA model or an LSTM neural network model; The elastic resource scheduling strategy specifically includes: preheating and migrating computing resources from the cloud or adjacent lightly loaded edge computing devices before the predicted peak edge computing load; staggering non-real-time computing tasks; and forming a resource pool among edge computing devices to achieve load balancing. The resource pool uses virtualization technology to create lightweight container instances on edge computing devices, enabling rapid migration and isolation of computing tasks.

[0101] Edge computing devices continuously monitor historical data on their CPU, memory, and other resource usage. They then use pre-deployed time-series forecasting models (such as LSTM neural networks) to predict resource load fluctuations over a future period (e.g., the next hour). For example, they might predict a significant increase in computing load during peak household activity at 7 PM.

[0102] Based on the prediction results, implement an elastic resource scheduling strategy: Resource preheating and migration: Before the predicted load peak (7 pm), the system can preheat and migrate some computing tasks (such as audio event detection) from the cloud or from a less loaded smart speaker (as an auxiliary edge node) in the same home network to share the load of the camera.

[0103] Off-peak scheduling: Schedule non-urgent tasks (such as device log analysis) to be executed during periods of network downtime (such as 2 a.m.).

[0104] Resource pooling: By using virtualization technologies such as Docker containers, the computing resources of multiple edge devices (cameras, gateways, smart TVs) in the home are virtualized into a resource pool to achieve load balancing.

[0105] In the security management module, the continuous identity authentication includes multi-factor authentication based on device fingerprints and user behavior profiles; The dynamic access control is attribute-based access control, and its authorization policy is dynamically adjusted according to real-time risk scores. This invention follows the zero-trust principle, which means that it never trusts any device or user by default.

[0106] Continuous authentication: Authentication between edge computing devices and the cloud is not just an initial connection, but a continuous process. For example, multi-factor authentication can be performed by verifying the device's unique hardware fingerprint (such as TPM chip information) and combining it with its behavioral profile (such as normal network access patterns).

[0107] Dynamic access control: Employs attribute-based access control (ABAC). For example, when an access request is detected to originate from an abnormal geographical location (attribute), the real-time risk score increases, and the authorization policy is dynamically adjusted, potentially denying access even if authentication is successful.

[0108] End-to-end encryption and evidence storage: Data transmission between edge computing devices and the cloud is encrypted end-to-end using national cryptographic algorithms (such as SM4). All critical operations (such as user login and configuration changes) are logged after hashing and recorded on a private blockchain to ensure immutability.

[0109] Privacy-preserving computation: Sensitive facial feature data stored in the cloud is homomorphically encrypted, allowing the cloud to perform matching calculations on the encrypted data directly without decryption, thereby protecting user privacy.

[0110] In the federated learning module, the federated learning framework uses secure multi-party computation or homomorphic encryption to protect the privacy of the model parameter updates of the edge computing device before parameter aggregation; The incremental update employs model quantization and pruning techniques to reduce the amount of data distributed.

[0111] A federated learning framework (such as the FATE framework) runs in the cloud. Periodically (e.g., weekly), the cloud aggregates parameter updates (gradient information) from time series prediction models (LSTM models) from thousands of edge computing devices. Before aggregation, secure multi-party computation is used to protect the privacy of the parameter updates from each device, ensuring that the cloud cannot backtrack the original data.

[0112] After aggregation, a more accurate and better generalized globally optimized model is generated in the cloud. Then, the differences between this model and the old version (i.e., incremental updates) are quantized (converting 32-bit floating-point numbers to 8-bit integers) and pruned (removing unimportant neural connections). The model size is then compressed and distributed to various edge computing devices. These edge computing devices use newly generated local data to fine-tune the distributed model, making it better suited to their local environment. In this way, the intelligence level of all edge computing devices can continuously and collaboratively evolve while protecting data privacy.

[0113] In summary, the advantages of this invention are as follows: 1. Preprocessing of heterogeneous multimodal data collected by edge computing devices yields preprocessed data, which is then used to obtain input computing tasks. Multidimensional indicators, including those related to the computing tasks, are collected in real-time. A multi-objective decision function is constructed, with total system latency, energy consumption, and cost as optimization objectives. A heuristic algorithm is used to solve this multi-objective decision function, and an optimal task allocation strategy is dynamically generated based on the multidimensional indicators. This classifies computing tasks into edge-exclusive tasks, cloud-exclusive tasks, and cloud-edge collaborative tasks. Next, the edge computing device employs a content-addressed differential synchronization mechanism to generate data fingerprints from the preprocessed data using a hash algorithm. The preprocessed data is synchronized based on task type to execute the computing tasks locally or in the cloud. A two-level collaborative caching mechanism between the edge and cloud is also established. By monitoring historical data, the edge computing device uses a time-series prediction model to predict resource load fluctuations in the edge computing device and cloud over future periods. Based on the prediction results, an elastic resource scheduling strategy is implemented. Both the edge computing device and the cloud operate on a zero-trust principle, implementing continuous authentication and dynamic access control. National cryptographic algorithms are used at the data transmission link layer. The system employs end-to-end encryption and combines blockchain technology to store key operation logs. Sensitive data in the stored preprocessed data undergoes homomorphic encryption or differential privacy processing. A federated learning framework is used in the cloud to aggregate model parameter updates from multiple edge computing devices, generating a globally optimized model. This globally optimized model is then distributed to edge computing devices in an incremental update manner, allowing the devices to fine-tune the time series prediction model using local data. This dynamic multi-objective decision-making mechanism optimizes the allocation of computing tasks between the cloud and edge in real time, reducing latency and energy consumption from the source. Content-addressed differential synchronization and intelligent caching significantly improve data synchronization efficiency and save bandwidth resources. A security architecture integrating zero-trust principles, national cryptographic algorithms, and blockchain technologies provides comprehensive security throughout the entire data lifecycle. Furthermore, prediction-based elastic resource scheduling and federated learning-driven model collaborative updates enable proactive optimization of cloud and edge resources and continuous evolution of intelligence. These interconnected steps greatly improve overall system performance, resource utilization efficiency, and collaborative intelligence.

[0114] 2. By automatically identifying heterogeneous multimodal data (such as network traffic, video streams, audio streams, and sensor data) and calling the corresponding preprocessing pipelines (such as protocol parsing, keyframe extraction, and data calibration), the efficiency and accuracy of data processing are significantly improved. This data type-based adaptive preprocessing mechanism reduces manual intervention, ensures data quality, and provides standardized input for subsequent computing tasks, thereby reducing the computing load on edge devices and improving the overall system response speed. It is particularly suitable for resource-constrained edge environments such as routers, gateways, and cameras.

[0115] 3. By constructing a multi-objective decision function (optimizing latency, energy consumption, and cost) and using heuristic algorithms (such as particle swarm optimization or Q-learning) to dynamically generate task allocation strategies, computing tasks are divided into edge-exclusive, cloud-exclusive, or cloud-edge collaborative types. This decision mechanism based on real-time multi-dimensional indicators (such as latency sensitivity, computing load, and bandwidth) can adaptively balance computing resources, avoid resource bottlenecks, thereby maximizing system efficiency, reducing operating costs, and ensuring low-latency processing of critical tasks (such as real-time video analytics), thus improving the flexibility and economy of edge computing systems.

[0116] 4. By adopting a content-addressable differential synchronization mechanism (such as a Merkle tree structure) and a two-level collaborative caching mechanism, data fingerprints are generated and only changed data is synchronized, which significantly reduces the amount of data transmission and bandwidth consumption. At the same time, the cached content is dynamically adjusted according to the data popularity (such as using an LRU strategy), ensuring the local availability of frequently accessed data and reducing cloud dependence. This mechanism not only improves the efficiency of data synchronization, but also reduces network latency and energy consumption, making it particularly suitable for bandwidth-constrained edge scenarios, such as collaborative work between IoT devices.

[0117] 5. Predict resource load fluctuations using time series forecasting models (such as ARIMA or LSTM) and implement elastic resource scheduling strategies (such as resource preheating, off-peak scheduling, and resource pooling). This predictive approach allows the system to adjust resources in advance before load peaks to avoid performance degradation. It also enables rapid migration and isolation of computing tasks through virtualization technology, improving resource utilization and system reliability, reducing the risk of service interruption, and supporting the dynamic expansion of edge devices. It is suitable for environments with high volatility (such as smart homes or security monitoring).

[0118] 6. Based on the zero-trust principle, it implements continuous identity authentication, dynamic access control, end-to-end encryption (using national cryptographic algorithms), and blockchain notarization, and performs homomorphic encryption or differential privacy processing on sensitive data. This multi-layered security framework ensures the confidentiality and integrity of data during transmission, storage, and computation, preventing unauthorized access and tampering. At the same time, privacy protection technologies allow data analysis in an encrypted state, balancing data utility and security, complying with modern data regulations, and enhancing user trust. It is especially suitable for edge devices that process sensitive information (such as personal videos or network traffic).

[0119] 7. By adopting a federated learning framework, model parameter updates from multiple edge devices are aggregated (privacy is protected by secure multi-party computation or homomorphic encryption), generating a globally optimized model and distributing it to edge devices in an incremental update manner (such as model quantization and pruning). This method achieves distributed learning without exposing the original data, reduces communication overhead, and improves the adaptability and accuracy of the model through local fine-tuning. It supports continuous learning and optimization, improves the performance of edge intelligent systems, and reduces cloud dependence. It is suitable for scenarios that require real-time adaptation (such as adaptive network management or intelligent camera behavior analysis).

[0120] 8. By integrating multimodal data adaptive preprocessing, task allocation based on dynamic optimization of multidimensional indicators, efficient differential synchronization and caching mechanisms, predictive elastic resource scheduling, zero-trust security framework, and privacy-preserving federated learning, a high-efficiency, secure, and adaptive edge computing system was constructed. While significantly reducing latency, energy consumption, and cost, it maximized resource utilization, ensured data security and reliability, and continuously optimized system performance, comprehensively improving the intelligence level and collaborative efficiency of edge devices such as routers, gateways, and cameras.

[0121] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An edge computing method for routers, gateways, and cameras, characterized in that: Includes the following steps: Step S1: The edge computing device collects heterogeneous multimodal data, identifies the data type of the multimodal data, and calls the corresponding preprocessing pipeline to preprocess the multimodal data based on the identified data type to obtain preprocessed data; The edge computing device is a router, gateway, or camera; Step S2: The edge computing device acquires the input computing task and collects multi-dimensional indicators in real time, including the latency sensitivity, computational complexity, data privacy level, current computing load of the edge computing device, available bandwidth, and cloud resource cost of the computing task. It constructs a multi-objective decision function with the total system latency, energy consumption, and cost as optimization objectives, and uses a heuristic algorithm to solve the multi-objective decision function. Based on the multi-dimensional indicators, it dynamically generates the optimal task allocation strategy to classify the computing task into edge-exclusive tasks, cloud-exclusive tasks, and cloud-edge collaborative tasks. Step S3: The edge computing device adopts a content-addressable differential synchronization mechanism to generate a data fingerprint from the preprocessed data using a hash algorithm. The preprocessed data is synchronized based on the task type to execute the computing task locally or in the cloud. At the same time, a two-level collaborative caching mechanism between the edge and the cloud is established to cache the frequently accessed preprocessed data locally on the edge computing device. All the preprocessed data is synchronized to the cloud when idle. Before synchronization, the compression algorithm and compression ratio are dynamically selected based on network bandwidth and data type, and only the data whose data fingerprint has changed is synchronized. Step S4: The edge computing device monitors historical data and uses a pre-deployed time series prediction model to predict the resource load fluctuations of the edge computing device and the cloud in future periods, and implements an elastic resource scheduling strategy based on the prediction results. Step S5: The edge computing device and the cloud are all based on the zero-trust principle, implementing continuous identity authentication and dynamic access control; end-to-end encryption using national cryptographic algorithms is adopted at the data transmission link layer, and blockchain technology is used to store key operation logs; for sensitive data in the stored preprocessed data, homomorphic encryption or differential privacy processing is performed, so that the sensitive data can still be calculated and analyzed in the encrypted state; Step S6: The cloud uses a federated learning framework to aggregate model parameter updates from multiple edge computing devices for the time series prediction model, generating a globally optimized model. The globally optimized model is then distributed to the edge computing devices in an incremental update manner, and the edge computing devices use local data to fine-tune the time series prediction model.

2. The edge computing method for routers, gateways, and cameras as described in claim 1, characterized in that: In step S1, the multimodal data includes at least network traffic data, video streams, audio streams, and sensor data; The preprocessing includes at least protocol parsing and feature extraction of network traffic data, keyframe extraction and background modeling of video streams, speech activity detection and noise reduction of audio streams, data calibration, filtering, smoothing and outlier detection of sensor data; The data calibration employs a linear or nonlinear compensation algorithm based on a standard reference value; the filtering and smoothing employs a Kalman filter or a moving average filter algorithm; and the outlier detection employs a detection algorithm based on statistical distribution or an isolated forest. In step S2, the multi-objective decision function transforms the multi-objective problem into a single-objective problem by using a weighted summation method or an ε-constraint method; the heuristic algorithm is a particle swarm optimization algorithm or a Q-learning reinforcement learning algorithm.

3. The edge computing method for routers, gateways, and cameras as described in claim 1, characterized in that: In step S3, the content-addressable differential synchronization mechanism specifically involves: using a Merkle tree structure to organize preprocessed data and quickly locating differences by comparing the data fingerprints at the tree root; The secondary collaborative caching mechanism uses a least recently used strategy to replace data and dynamically adjusts the cached content based on a data popularity model.

4. The edge computing method for routers, gateways, and cameras as described in claim 1, characterized in that: In step S4, the time series prediction model is an ARIMA model or an LSTM neural network model. The elastic resource scheduling strategy specifically includes: preheating and migrating computing resources from the cloud or adjacent lightly loaded edge computing devices before the predicted peak edge computing load; staggering non-real-time computing tasks; and forming a resource pool among edge computing devices to achieve load balancing. The resource pool uses virtualization technology to create lightweight container instances on edge computing devices, enabling rapid migration and isolation of computing tasks.

5. The edge computing method for routers, gateways, and cameras as described in claim 1, characterized in that: In step S5, the continuous identity authentication includes multi-factor authentication based on device fingerprints and user behavior profiles; The dynamic access control is attribute-based access control, and its authorization policy is dynamically adjusted according to real-time risk scores. In step S6, the federated learning framework uses secure multi-party computation or homomorphic encryption to protect the privacy of the model parameter updates of the edge computing device before parameter aggregation. The incremental update employs model quantization and pruning techniques to reduce the amount of data distributed.

6. An edge computing system for routers, gateways, and cameras, characterized in that: Includes the following modules: The multimodal data acquisition module is used to collect heterogeneous multimodal data from edge computing devices, identify the data type of the multimodal data, and call the corresponding preprocessing pipeline based on the identified data type to preprocess the multimodal data to obtain preprocessed data. The edge computing device is a router, gateway, or camera; The task allocation strategy generation module is used to acquire the input computing tasks of the edge computing device, collect multi-dimensional indicators in real time, including the latency sensitivity, computational complexity, data privacy level, current computing load of the edge computing device, available bandwidth, and cloud resource cost of the computing tasks, construct a multi-objective decision function with the total system latency, energy consumption and cost as optimization objectives, and solve the multi-objective decision function using a heuristic algorithm. Based on the multi-dimensional indicators, the optimal task allocation strategy is dynamically generated to classify the computing tasks into edge-exclusive tasks, cloud-exclusive tasks and cloud-edge collaborative tasks. The computing task execution module is used by the edge computing device to adopt a content-addressable differential synchronization mechanism to generate a data fingerprint from the preprocessed data using a hash algorithm, and synchronize the preprocessed data based on the task type to execute the computing task locally or in the cloud; at the same time, a two-level collaborative caching mechanism between the edge and the cloud is established to cache the frequently accessed preprocessed data locally on the edge computing device, and synchronize all the preprocessed data to the cloud when idle. Before synchronization, the compression algorithm and compression ratio are dynamically selected according to network bandwidth and data type, and only the data whose data fingerprint has changed is synchronized. The elastic resource scheduling module is used by edge computing devices to monitor historical data, use pre-deployed time series prediction models to predict the resource load fluctuations of edge computing devices and the cloud in future periods, and implement elastic resource scheduling strategies based on the prediction results. The security management module implements continuous identity authentication and dynamic access control based on the zero-trust principle for both edge computing devices and the cloud; it uses end-to-end encryption with national cryptographic algorithms at the data transmission link layer and combines blockchain technology to store key operation logs; it performs homomorphic encryption or differential privacy processing on sensitive data in the stored preprocessed data, so that the sensitive data can still be calculated and analyzed in the encrypted state. The federated learning module is used in the cloud to aggregate model parameter updates of the time series prediction model from multiple edge computing devices using a federated learning framework, and generate a globally optimized model. The globally optimized model is then distributed to the edge computing devices in an incremental update manner, and the edge computing devices use local data to fine-tune the time series prediction model.

7. An edge computing system for routers, gateways, and cameras as described in claim 6, characterized in that: In the multimodal data acquisition module, the multimodal data includes at least network traffic data, video streams, audio streams, and sensor data; The preprocessing includes at least protocol parsing and feature extraction of network traffic data, keyframe extraction and background modeling of video streams, speech activity detection and noise reduction of audio streams, data calibration, filtering, smoothing and outlier detection of sensor data; The data calibration employs a linear or nonlinear compensation algorithm based on a standard reference value; the filtering and smoothing employs a Kalman filter or a moving average filter algorithm; and the outlier detection employs a detection algorithm based on statistical distribution or an isolated forest. In the task allocation strategy generation module, the multi-objective decision function transforms the multi-objective problem into a single-objective problem by using a weighted summation method or an ε-constraint method; the heuristic algorithm is a particle swarm optimization algorithm or a Q-learning reinforcement learning algorithm.

8. An edge computing system for routers, gateways, and cameras as described in claim 6, characterized in that: In the computation task execution module, the content-addressable differential synchronization mechanism specifically involves: using a Merkle tree structure to organize preprocessed data and quickly locating differences by comparing the data fingerprints at the tree root; The secondary collaborative caching mechanism uses a least recently used strategy to replace data and dynamically adjusts the cached content based on a data popularity model.

9. An edge computing system for routers, gateways, and cameras as described in claim 6, characterized in that: In the elastic resource scheduling module, the time series prediction model is an ARIMA model or an LSTM neural network model; The elastic resource scheduling strategy specifically includes: preheating and migrating computing resources from the cloud or adjacent lightly loaded edge computing devices before the predicted peak edge computing load; staggering non-real-time computing tasks; and forming a resource pool among edge computing devices to achieve load balancing. The resource pool uses virtualization technology to create lightweight container instances on edge computing devices, enabling rapid migration and isolation of computing tasks.

10. An edge computing system for routers, gateways, and cameras as described in claim 6, characterized in that: In the security management module, the continuous identity authentication includes multi-factor authentication based on device fingerprints and user behavior profiles; The dynamic access control is attribute-based access control, and its authorization policy is dynamically adjusted according to real-time risk scores. In the federated learning module, the federated learning framework uses secure multi-party computation or homomorphic encryption to protect the privacy of the model parameter updates of the edge computing device before parameter aggregation. The incremental update employs model quantization and pruning techniques to reduce the amount of data distributed.

Images