A cloud-edge collaborative multi-channel fusion communication control method and system

Abstract

This invention discloses a cloud-edge collaborative multi-channel fusion communication control method and system, belonging to the field of wireless communication network technology, aiming to improve the efficiency, reliability, and real-time performance of multi-channel fusion communication in complex scenarios. Specifically, it includes: constructing a layered collaborative architecture between the cloud and the edge; the edge end collecting network status parameters of multiple communication channels, outputting channel quality assessment values ​​based on a time-series prediction model, and performing multi-channel data transmission through an extended communication protocol when channel switching is triggered; the edge end performing network parameter configuration, traffic priority scheduling, and operational status data collection operations; the cloud generating soft routing optimization instructions based on the operational status data and sending them to the edge end for execution to adjust network transmission performance parameters; real-time monitoring of the container operation status and device status at the edge end; when the monitoring results match preset anomaly rules, the edge end generating alarm information, sending it back to the cloud to generate a global security policy, and sending it to the edge end to update the preset anomaly rules.

Description

Technical Field

[0001] This invention relates to the field of wireless communication network technology, and in particular to a cloud-edge collaborative multi-channel fusion communication control method and system. Background Technology

[0002] With the widespread adoption of various communication technologies such as 5G, Wi-Fi, and satellite communication, single channels are no longer sufficient to meet the communication needs of complex scenarios. There is an urgent need to integrate the advantages of different technologies through multi-channel fusion to overcome the limitations of single-network transmission, reduce the risk of network outages, and optimize costs. Therefore, in the context of the current deep integration of communication technologies and industry applications, the development of multi-channel fusion technology is of significant necessity.

[0003] Currently, multi-channel convergence is mainly practiced around three core areas: intelligent handover, weak network optimization, and load balancing. Intelligent handover is achieved by dynamically selecting the optimal channel through real-time monitoring of parameters such as network latency, packet loss rate, and signal strength, thus avoiding single points of failure. In scenarios with unstable signals (such as complex sea areas), weak network optimization is accomplished by employing multi-channel redundant transmission or downgrading to a low-bandwidth mode to ensure basic communication. Load balancing is achieved by distributing data traffic across different channels to alleviate congestion on a single network. Furthermore, the industry has begun exploring technological upgrades, such as combining future 6G and Wi-Fi 7 terahertz band technologies with AI-driven scheduling, and developing satellite-ground converged networks based on satellite communications, aiming to achieve superior performance such as seamless handover and global coverage without dead zones.

[0004] However, existing multi-channel fusion methods still have significant shortcomings in specific scenarios, especially in the field of UAV applications in complex sea areas, where three key problems have been exposed: First, when faced with the massive video stream transmission of UAVs, existing fusion solutions are unable to cope with the surge in network bandwidth pressure, and transmission stuttering is likely to occur; second, the complex sea environment is harsh and signal interference is strong, and existing weak network optimization strategies cannot fully guarantee communication reliability, and there is still a risk of data loss or connection interruption; third, the computing power of UAVs at the edge is limited, and existing solutions have not been optimized in conjunction with edge computing capabilities, resulting in insufficient real-time performance for key tasks such as target recognition, which affects the operational efficiency of UAVs.

[0005] Therefore, there is an urgent need for a method to achieve high efficiency, reliability, and real-time performance of multi-channel fusion communication in complex scenarios. Summary of the Invention

[0006] To address at least one of the aforementioned technical problems in the existing technology, this invention provides a cloud-edge collaborative multi-channel fusion communication control method and system to achieve high efficiency, reliability, and real-time performance of multi-channel fusion communication in complex scenarios.

[0007] Specifically, the present invention is achieved through the following technical solution:

[0008] In a first aspect, the present invention provides a cloud-edge collaborative multi-channel fusion communication control method, comprising:

[0009] Construct a layered collaborative architecture between the cloud and the edge; the cloud performs cluster scheduling, model deployment, and data storage operations on the edge, while the edge performs bidirectional data transmission, status information synchronization, and command response operations with the cloud;

[0010] The edge terminal collects network status parameters of multiple communication channels, outputs channel quality assessment values ​​based on a time-series prediction model, and triggers channel switching when the channel quality assessment values ​​meet preset switching conditions, and performs multi-channel data transmission through an extended communication protocol.

[0011] The edge device loads a lightweight optimized target detection model, performs inference calculations on the multi-source perception data collected by the terminal device equipped with the edge device, outputs the result data and transmits it to the cloud;

[0012] The edge device performs network parameter configuration, traffic priority scheduling, and operation status data collection operations, and synchronizes the operation status data to the cloud. The cloud generates a soft router optimization instruction based on the operation status data and sends it to the edge device for execution to adjust the network transmission performance parameters of the edge device.

[0013] The operating status of the container and the status of the terminal device at the edge are monitored in real time. When the monitoring result matches the preset anomaly rule, the edge generates an alarm message and transmits it to the cloud. The cloud generates a global security policy based on the alarm message and sends it to the edge to update the preset anomaly rule.

[0014] Secondly, the present invention provides a cloud-edge collaborative multi-channel fusion communication control system, comprising a cloud terminal and an edge terminal; wherein...

[0015] The edge device performs bidirectional data transmission with the cloud, status information synchronization, and command response operations, including:

[0016] The system collects network status parameters of multiple communication channels, outputs channel quality assessment values ​​based on a time-series prediction model, and triggers channel switching when the channel quality assessment values ​​meet preset switching conditions. Multi-channel data transmission is then performed through an extended communication protocol.

[0017] A lightweight and optimized target detection model is loaded, and inference calculations are performed on the multi-source perception data collected by the terminal device equipped with the edge terminal. The result data is then output and transmitted to the cloud.

[0018] Perform network parameter configuration, traffic priority scheduling, and operation status data collection operations, and synchronize the operation status data to the cloud;

[0019] The system monitors the operational status of containers and the status of terminal devices at the edge in real time. When the monitoring results match the preset anomaly rules, an alarm message is generated and transmitted to the cloud.

[0020] The cloud performs cluster scheduling, model deployment, and data storage operations on the edge, including:

[0021] Generate multi-channel task allocation instructions and send them to the edge for execution, so as to adjust the channel usage strategy of the edge in real time;

[0022] Based on the running status data uploaded by the edge device, a soft router optimization command is generated and sent to the edge device for execution to adjust the network transmission performance parameters of the edge device;

[0023] A global security policy is generated based on the alarm information uploaded from the edge device, and then sent to the edge device to update the preset anomaly rules.

[0024] Compared with the prior art, the advantages and positive effects of the present invention are mainly reflected in:

[0025] This invention is based on a layered collaborative architecture of "global control in the cloud and local execution at the edge" to achieve efficient and stable multi-channel converged communication. The cloud provides global support to the edge through cluster scheduling, model deployment, and data storage, while the edge executes local operations through bidirectional data transmission, state synchronization, and command response. This architecture allows the cloud to coordinate edge cluster resources and the edge to quickly respond to the needs of real-world scenarios, solving the limitations of computing power and real-time performance of a single cloud or edge in complex environments, and laying a collaborative foundation for multi-channel convergence.

[0026] Under this architecture, the timing prediction model and the extended communication protocol work together to form the core mechanism for channel handover: Based on the network status parameters of 4G / 5G, WiFi and other channels collected at the edge, and the future channel quality assessment value predicted by the LSTM neural network, the timing of channel handover is determined in advance, avoiding the handover lag problem caused by traditional real-time judgment and reserving a buffer for channel handover; The extended communication protocol ensures the integrity of data retransmission and the session persistence mechanism maintains the connection without interruption during channel handover, solving problems such as data loss and connection interruption during handover. The combination of the two makes channel handover both accurate and smooth, ensuring stable communication quality. Attached Figure Description

[0027] Figure 1 This is a flowchart of an embodiment of the cloud-edge collaborative multi-channel fusion communication control method proposed in this invention;

[0028] Figure 2 This is a schematic diagram of the structure of an embodiment of the cloud-edge collaborative multi-channel fusion communication control system proposed in this invention;

[0029] Figure 3 for Figure 2 A specific module architecture for one embodiment of cloud and edge computing. Detailed Implementation

[0030] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention.

[0031] The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0032] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0033] In the description of this invention, the consecutive numbers of the method steps are for ease of review and understanding. Considering the overall technical solution of this invention and the logical relationship between each step, adjusting the implementation order of the steps will not affect the technical effect achieved by the technical solution of this invention.

[0034] The following specific embodiments are provided to illustrate the technical solution of the present invention in detail.

[0035] Figure 1 The flowchart of the cloud-edge collaborative multi-channel fusion communication control method in this embodiment may include the following processes:

[0036] S101. Construct a layered collaborative architecture between the cloud and the edge.

[0037] See Figure 2 In this embodiment, the cloud is configured to perform cluster scheduling, model deployment, and data storage operations on the edge, while the edge is configured to perform bidirectional data transmission, status information synchronization, and command response operations with the cloud.

[0038] Specifically, the cloud is the core control and resource hub in the layered collaborative architecture. Built on cloud computing technology, it has powerful computing power, massive data storage capacity and global scheduling capabilities, and is the "decision and resource center" of the entire system.

[0039] The cloud provides support and management for the edge, primarily undertaking three core tasks:

[0040] Cluster scheduling: As a global manager, it performs unified scheduling of edge clusters (such as UAV communication terminals distributed in different areas of complex sea areas), including dynamically allocating computing resources and task priorities based on the real-time load, task requirements, network status, etc. of the edge terminals.

[0041] Model Deployment: The key algorithm models required by the system (such as multimodal data fusion models, target recognition models, intelligent network scheduling models, etc.) are uniformly deployed to the edge. At the same time, the models at the edge can be uniformly updated and upgraded according to business iteration or performance optimization needs.

[0042] Data storage: Receives and stores various types of data uploaded from the edge (such as massive video streams collected by drones, sensor data, device status data, etc.).

[0043] The edge is a local execution and data interaction node in a layered collaborative architecture. It usually exists in the form of hardware terminals (such as multi-channel converged communication terminals installed on drones, edge computing modules), and is deployed at the "edge position" close to the data generation source (such as drone operation site, complex sea area mission area). It has lightweight computing power, real-time data processing capabilities and bidirectional communication capabilities with the cloud, and is the "local execution and data entry point" of the entire system.

[0044] The core role of the edge computing edge is to connect the cloud with actual business scenarios, acting as a bridge between "data interaction" and "command response," specifically including three types of tasks:

[0045] Two-way data transmission: On the one hand, it uploads locally collected real-time data (such as video streams taken by drones, equipment operating parameters, network quality data, etc.) to the cloud; on the other hand, it receives data such as resource configuration, model parameters, and task instructions issued by the cloud as the basis for executing tasks locally.

[0046] Status information synchronization: Real-time synchronization of its own operating status (such as terminal load rate, computing power usage, device fault information, current network connection status, etc.) to the cloud.

[0047] Command Response: Receive various commands from the cloud (such as task start / stop commands, model update commands, network switching commands, etc.) and quickly execute the response operations locally.

[0048] Optional, see Figure 3 The cloud mainly includes a cluster scheduling module, a cloud-edge communication module, a data storage module, and a model management module; the edge mainly includes an edge proxy component, an instruction execution unit, and a device twin module. The cloud generates task allocation instructions through the cluster scheduling module, which are then sent to the edge proxy component at the edge via the cloud-edge communication module. The edge proxy component distributes the task allocation instructions to the instruction execution unit, and the execution results are processed by the device twin module and then transmitted back to the cloud's data storage module via the edge proxy component and the cloud-edge communication module. The cloud's model management module optimizes the model based on the stored data and sends the updated model to the edge. The edge proxy component then forwards the updated model to the instruction execution unit for use in subsequent tasks.

[0049] Specifically, the cloud, as the "control and resource hub" of the layered collaborative architecture, supports and manages the edge through four core modules: the cluster scheduling module is the "global task commander" of the cloud, whose core function is to generate and manage task allocation instructions; the cloud-edge communication module is the "data transmission bridge" between the cloud and the edge, undertaking the channel function of instruction issuance and data feedback; the data storage module is the "data warehouse" of the cloud, whose core function is to receive, store and manage various types of data; and the model management module is the "model optimization and distribution center" of the cloud, whose core function is to optimize and update models.

[0050] As a "local execution and data interaction node," the edge terminal achieves collaboration with the cloud and local task execution through three major modules: the edge proxy component is the "information relay station and task distributor" of the edge terminal, and is the core hub connecting the cloud and the internal modules of the edge terminal. It is mainly responsible for information reception and distribution, data collection and uploading; the instruction execution unit is the "task executor" of the edge terminal, and its core function is to receive and execute task instructions issued by the cloud; the device twin module is the "status monitoring and data organizer" of the edge terminal, and its core function is to monitor the status of the edge terminal itself in real time and organize the data.

[0051] In practice, the cloud-based cluster scheduling module generates task allocation instructions, which are transmitted to the edge proxy component at the edge via the cloud-edge communication module. The edge proxy component then distributes the instructions to the instruction execution unit. The instruction execution unit executes the task and generates results. Simultaneously, the device twin module organizes the operational status data at the edge, and both types of information are aggregated to the edge proxy component. The edge proxy component sends the task execution results and operational status data back to the cloud-edge communication module, where they are stored by the data storage module. The cloud-based model management module optimizes the model based on the historical data stored in the data storage module, and distributes the updated model to the edge proxy component at the edge via the cloud-edge communication module. The edge proxy component then forwards the updated model to the instruction execution unit, updating the model used at the edge. The updated model supports more efficient task execution at the edge, and the new execution results and operational status data are again sent back to the cloud, forming a continuously optimized collaborative closed loop. This ensures that the entire system operates efficiently and reliably in complex scenarios (such as UAV communication in complex maritime areas).

[0052] This embodiment achieves high efficiency, reliability, and dynamic optimization of multi-channel fusion by setting up a layered collaborative architecture between the cloud and the edge, relying on the deep cooperation between the cloud's global control capabilities and the edge's local execution advantages, with "cloud-based overall scheduling and edge-based real-time response" as the core means.

[0053] Specifically, the cloud-based cluster scheduling module can issue precise multi-channel task allocation instructions to the edge proxy component via the cloud-edge communication module based on the real-time status returned by the device twin module at the edge, avoiding single-channel overload or resource waste. Simultaneously, the instruction execution unit at the edge can quickly respond to the instruction, adjusting channel usage strategies in real time, solving the problems of "scheduling lag and blind channel allocation" in traditional multi-channel fusion, and ensuring smooth communication in complex scenarios. The cloud-based data storage module can store massive amounts of multi-channel communication data uploaded by the edge. The model management module optimizes the multimodal data fusion algorithm and intelligent network scheduling model based on this data, and distributes the updated model to the edge, enabling the edge to continuously improve the accuracy of channel switching (e.g., by optimizing the model to predict in advance that a channel is about to be interrupted and automatically switch to a backup channel), addressing the shortcomings of "fixed algorithm models and inability to adapt to dynamic marine environments" in traditional multi-channel fusion, and enhancing communication reliability in weak network environments.

[0054] Furthermore, since the edge devices are located close to the data source (such as drones), they can perform some data preprocessing locally via multiple channels (such as compressing video streams) before uploading critical data to the cloud. This reduces the amount of data transmitted to the cloud and alleviates the bandwidth pressure caused by massive data transmission during multi-channel fusion by relying on "edge preprocessing and cloud storage." At the same time, the cloud synchronizes the channel status of the edge devices in real time through the cloud-edge communication module and combines global data to realize cross-edge channel resource allocation (such as allocating idle satellite channel resources of one edge device to adjacent high-load edge devices), further improving the utilization rate of multi-channel resources. Ultimately, this achieves the effect of "low latency, high reliability, and wide coverage" in multi-channel fusion communication in scenarios such as complex sea areas, meeting the stable communication requirements of application scenarios such as drone logistics inspection and emergency response.

[0055] S102. Collect network status parameters of multiple communication channels at the edge, output channel quality assessment values ​​based on the time-series prediction model, and trigger channel switching when the channel quality assessment values ​​meet the preset switching conditions, and perform multi-channel data transmission through extended communication protocols.

[0056] Specifically, multiple communication channels refer to the transmission channels corresponding to different types of communication technologies that can be accessed at the edge, such as 4G / 5G, WiFi, and satellite communication. These channels each have different characteristics (e.g., 5G has high bandwidth but is easily blocked in complex sea areas, satellite communication has wide coverage but limited bandwidth, and WiFi is suitable for short-range high-speed transmission). The edge device integrates these channels to achieve multi-path data transmission, providing a basic carrier for converged communication.

[0057] Network status parameters refer to key indicators that are collected in real time at the edge and reflect the current performance of each communication channel. These include, but are not limited to, latency (time difference in data transmission), packet loss rate (proportion of data lost during transmission), signal strength (signal coverage strength of the channel), and bandwidth utilization (proportion of currently used bandwidth to the total bandwidth of the channel). These parameters are the original basis for evaluating channel quality and directly reflect the real-time communication capability of the channel.

[0058] Temporal prediction models are algorithmic models (such as LSTM) that predict future trends based on historical data sequences. By inputting network state parameters of a channel over a period of time (such as the latency and packet loss rate changes over the past 10 minutes) at the edge, the model can mine the temporal variation patterns of the parameters, predict the channel state trend in the near future, and thus output a forward-looking assessment of channel quality.

[0059] The channel quality assessment value is an indicator that quantifies the overall performance of a channel, output by a time-series prediction model based on network state parameters. It forms a numerical value or level (e.g., 0-100 points, with 80 points or above considered a high-quality channel) that can be directly used for decision-making by weighting or trend analysis of parameters such as delay, packet loss rate, and signal strength (e.g., predicting whether the packet loss rate will exceed a threshold in the next 5 seconds). This value is used to determine whether the channel meets the current communication requirements.

[0060] Extended communication protocols are optimized versions of existing communication protocols that support multi-channel collaborative transmission. Their core functions include: uniformly encapsulating data formats for different channels to achieve compatible data transmission across multiple channels; defining data synchronization rules during channel switching (such as avoiding duplicate transmissions or data loss); and supporting load balancing strategies for parallel multi-channel transmission (such as splitting large files and transmitting them simultaneously through different channels) to ensure efficient and consistent data transmission when multiple channels are integrated.

[0061] In specific implementation, network status parameters of various communication channels are collected at the edge, and channel quality assessment values ​​are output based on a time-series prediction model. This includes: the edge collects network status parameters of 4G / 5G links and WiFi links at a preset sampling frequency through a network sensing module; the collected network status parameters are standardized and converted into feature vectors of a preset dimension; the feature vectors are input into a time-series prediction model deployed at the edge, and the time-series prediction model outputs channel quality assessment values ​​of 4G / 5G links and WiFi links within a preset time period in the future.

[0062] In this embodiment, the time-series prediction model can be an LSTM neural network model built based on TensorFlow Lite; the cellular mobile communication link preferably uses a 5G link; and the WiFi link preferably uses a WiFi 6 link. The following detailed description of the solution uses only 5G and WiFi 6 links as examples.

[0063] Optionally, the process of inputting the feature vector into a time-series prediction model deployed at the edge, and the time-series prediction model outputting channel quality assessment values ​​for 5G links and WiFi 6 links within a preset future time period, may include:

[0064] Initialize the time series prediction model; the time series prediction model includes an input layer, a hidden layer and an output layer. The hidden layer is equipped with three gating units: an input gate, a forget gate and an output gate, to control the flow of information.

[0065] The standardized feature vectors are input into the time series prediction model, which generates initial hidden states and cell states based on the input data of historical time steps, and learns dependencies through gating units.

[0066] During the prediction process, each time the feature vector fragment of the current time step and the hidden state of the previous time step are input, the output of the current time step is calculated, and the output of the current time step is used as the input of the next time step. The calculation process is iterated until the prediction time range is reached or the stopping condition is met.

[0067] The time-series prediction model outputs predicted values ​​of key network performance indicators within a preset time period in the future, and generates channel quality assessment values ​​for multiple future time slices based on the predicted values ​​of key network performance indicators.

[0068] In practice, the edge device uses a built-in network sensing module to collect real-time network status data for both 5G and WiFi 6 links at a pre-set sampling frequency (e.g., once per second). The collected parameters include, but are not limited to, latency (e.g., the time difference between data transmission and reception, in milliseconds), packet loss rate (e.g., the percentage of lost packets out of 100), signal strength (e.g., RSRP value for 5G links and RSSI value for WiFi 6 links), and bandwidth utilization (e.g., the percentage of total bandwidth currently used). The collected raw network status parameters are standardized and transformed using a Min-Max normalization method (mapping parameter values ​​to the 0-1 range) to form a fixed-dimensional feature vector (e.g., 5-dimensional if including latency, packet loss rate, etc.). Each vector corresponds to the link status at a specific sampling time point.

[0069] A time-series prediction model deployed at the edge is loaded. This model is an LSTM neural network model built on the TensorFlow Lite framework (lightweight and adapted to edge computing power). The model structure includes an input layer (receiving feature vectors of a preset dimension), a hidden layer (with three gating units: input gate, forget gate, and output gate), and an output layer (outputting the prediction result). The gating units control the flow of information through a sigmoid activation function (e.g., the forget gate determines which information from historical cell states is discarded, the input gate controls the storage of new information, and the output gate determines the output content of the current hidden state).

[0070] The standardized feature vectors are input into the LSTM model in chronological order. The model first calculates and generates initial hidden states (short-term memory) and cell states (long-term memory) based on input data from multiple historical time steps (e.g., the past 30 sampling points). Through the gating units of the hidden layers, it learns the dependencies between parameters at different time steps (e.g., the change in packet loss rate after a certain period of time due to increased latency). In the process of predicting the future preset duration (e.g., the next 60 seconds), each time the feature vector fragment of the current time step (e.g., the features of the most recent sampling point) and the hidden state calculated in the previous time step are input, the cell state and hidden state are updated through the gating units to obtain the output of the current time step (e.g., the network state prediction for the next second). The current output is then used as the input for the next time step (combined with the new hidden state), and the above calculation is repeated until the prediction of all time slices within the future preset duration (e.g., each second is a time slice) is completed.

[0071] The model outputs predicted values ​​of key network performance indicators (such as specific values ​​of latency and packet loss rate) for each time slot within a preset future duration. Based on these predicted values, a weighted summation is performed (e.g., latency weight 0.3, packet loss rate weight 0.5, signal strength weight 0.2, to calculate a score from 0 to 100) to generate the channel quality assessment value of the 5G link and WiFi 6 link for each time slot.

[0072] LSTM can predict values ​​at future time steps and iteratively adjust the model input based on the prediction results to achieve longer-term time series forecasts. It's important to note that prediction accuracy may decrease as the number of time steps increases because errors accumulate. Therefore, in practical applications, the prediction time range and number of iterations can be selected based on specific circumstances to balance prediction accuracy and computational efficiency.

[0073] When the channel quality assessment value meets the preset switching conditions, channel switching is triggered, and multi-channel data transmission is performed through an extended communication protocol, specifically including the following process:

[0074] When the channel quality assessment value is lower than the preset threshold, an extended communication protocol is used to establish a multi-channel parallel transmission link. The extended communication protocol adds a multi-path identifier field to the message queue telemetry transmission protocol, and different fragments of the same data session are transmitted in parallel on multiple channels.

[0075] During channel switching, a session context mapping table is maintained to record data sequence numbers and transmission status, and buffered data that has not been fully transmitted is retransmitted through the new channel.

[0076] By utilizing the extended session persistence mechanism of the protocol, the connection state with the message broker is maintained during channel switching;

[0077] If the quality of the new channel does not meet the standard, it will fall back to the original channel within a preset time period, and the protocol will be downgraded to single-path transmission mode.

[0078] In practice, the edge device monitors the channel quality assessment values ​​output by each channel (such as 5G and WiFi 6) in real time. When the assessment value of a channel is lower than a preset threshold (such as 60 points), it is determined that the current quality of the channel does not meet the communication requirements, triggering a channel switching mechanism. An extended communication protocol is then initiated to establish a multi-channel parallel transmission link.

[0079] In this embodiment, the extended communication protocol can be modified based on the Message Queuing Telemetry Transport Protocol (MQTT). A multipath identifier field (e.g., using 2 bytes to mark different segments of the same data session) is added to the message header of the original protocol. For example, a 10MB video data stream can be split into 10 1MB segments, each carrying an identifier of "session ID + segment sequence number," which are transmitted simultaneously via 5G and WiFi 6 links. During channel switching, the system automatically generates and maintains a session context mapping table, which records in detail the sequence number (e.g., 1 to 10), transmission status (e.g., segment 1 has been acknowledged, segment 2 is transmitting, segment 3 has not been transmitted), and corresponding transmission channel information for each data segment. For buffered data that has not been transmitted (e.g., segment 2 transmission interrupted, segment 3 not started), the edge device re-initiates transmission through the newly switched channel (e.g., a satellite channel added after switching from 5G) to ensure data integrity. When switching channels, a connection-keeping frame carrying a session identifier is sent to the message broker (such as the MQTTBroker in the cloud) through the protocol layer. The message broker maintains the connection state of the original session based on this identifier (e.g., without resetting the session timeout timer) to avoid connection loss due to channel switching. After switching to the new channel, its quality assessment value is evaluated in real time. If the threshold (e.g., 60 minutes) is not reached within a preset time (e.g., 3 seconds), a fallback mechanism is triggered: the original channel is immediately switched back, and the extended protocol is downgraded to single-path transmission mode (i.e., data is transmitted only through the original channel, without splitting and sending in parallel), until the channel quality recovers to the required level.

[0080] This embodiment achieves channel handover through the cooperation of a timing prediction model and an extended communication protocol. It takes "proactive prediction of handover timing" and "efficient execution of the handover process" as the core means to jointly ensure channel quality: The timing prediction model is based on historical network status parameters (such as latency and packet loss rate) of 5G, WiFi 6 and other links collected at the edge. It learns the parameter change pattern through the gating unit of the LSTM neural network and outputs the channel quality assessment value within a preset time period in the future (such as the quantization score of 1 second in the next 10 seconds). When it is predicted that the assessment value of a certain channel will be lower than the preset threshold, the handover mechanism is triggered in advance. This approach solves the handover lag problem caused by traditional real-time status judgment, ensuring that handover preparation is completed before channel quality deteriorates, providing a time window to ensure communication continuity. The extended communication protocol plays a key role in the handover execution phase. It adds a multi-path identifier field to the MQTT protocol, which can split the same data session into identified fragments and transmit them in parallel through multiple channels. At the same time, it records the fragment sequence number and transmission status through a session context mapping table, ensuring that buffered data that has not been transmitted can be retransmitted through the new channel. Combined with the session persistence mechanism, it maintains the connection with the message broker without interruption. If the quality of the new channel is not up to standard, it can quickly fall back to the original channel and degrade to single-path mode. These methods solve the problems of data loss and connection interruption during handover, ensuring the smoothness of handover and data integrity. The two work together to form a closed loop of "prediction-execution-guarantee": the timing prediction model accurately determines "when to switch" and reserves buffer time for the execution of the extended communication protocol, while the extended communication protocol efficiently solves "how to switch", ensuring that data transmission is uninterrupted and the quality is not degraded during handover. Ultimately, it achieves high reliability of multi-channel converged communication in complex environments. Even in scenarios with sudden changes in channel state, it can maintain stable communication quality through advance prediction and smooth handover.

[0081] S103. Load a lightweight optimized target detection model at the edge, perform inference calculations on the multi-source perception data collected by the terminal device equipped with the edge, output the result data and transmit it to the cloud.

[0082] Specifically, the result data is structured information output by the edge computing layer after performing inference calculations on multi-source perception data (such as video frames captured by drones, infrared images, radar scan data, etc.) collected by the terminal device through a lightweight target detection model. For drones as the terminal device, the result data is the target recognition result, which typically includes: the detected target category (such as ships, islands, obstacles, people, etc.), the target's position coordinates in the image or perception data (such as bounding box coordinates), the target's confidence score (the model's judgment on the accuracy of the recognition result, such as 0.9 indicating a 90% probability of it being the target), and in some scenarios, it may also include the target's motion state (such as speed, direction), etc.

[0083] In practice, a lightweight, optimized target detection model is loaded at the edge. The optimized model introduces a cross-modal attention (CMA) module into the feature fusion layer of the base detection model, adjusts sample weights using a focusing loss function, and reduces the model's precision format using model quantization tools (e.g., using the HailoModel Zoo toolchain to convert an FP32 model to INT8 format). The edge receives multi-source perception data collected by terminal devices. Its heterogeneous computing units load the optimized target detection model, perform inference calculations on the multi-source perception data, and output result data, such as target coordinates, category information, and confidence scores collected by the drone. The edge synchronizes the result data to the cloud. When it receives a model update command from the cloud, the edge updates the target detection model online.

[0084] Specifically, the edge device pre-acquires a lightweight, optimized target detection model. The optimization process involves introducing a cross-modal attention module into the feature fusion layer (neck layer) of the base detection model (such as the YOLOv5s model, a popular computer vision algorithm primarily used for target detection). This allows the model to focus on key features for target detection in multi-source perception data (e.g., prioritizing feature regions related to "ships" and "obstacles" from visible and infrared images from UAVs). Simultaneously, a focus loss function is used to assign higher weights to difficult-to-identify samples (such as small or blurry targets) and lower weights to easily identifiable samples, improving the model's accuracy in identifying targets in complex scenes. Finally, model quantization tools (such as the Hailo Model Zoo toolchain) reduce the model's precision format from 32-bit floating-point (FP32) to 8-bit integer (INT8). The edge device receives multi-source perception data (such as visible light video frames, infrared images, and radar point cloud data) collected by the UAV through a data transmission interface. At the edge, heterogeneous computing units (such as GPUs, NPUs, or dedicated AI acceleration chips) are activated to load a lightweight and optimized target detection model. Preprocessed multi-source perception data is input into the model for inference computation. The model first extracts basic features from each data source through a backbone network, then fuses and weights different modal features using a feature fusion layer combined with a cross-modal attention module, highlighting effective features and suppressing redundant information. Subsequently, a detection head network performs target localization and classification on the fused features, ultimately outputting specific information about the drone target, including the target's coordinates in the image (e.g., the top-left and bottom-right pixel coordinates of the bounding box), category information (e.g., "ship," "island," "bird," etc.), and confidence score (e.g., 0.85, 0.92, etc.). The edge then organizes the output target recognition results (coordinates, category, confidence score) according to a preset format and synchronously transmits them to the cloud data storage module via the cloud-edge communication module.

[0085] The edge device monitors cloud commands in real time. When it receives a model update command (including the updated model file or parameters) from the cloud model management module, it initiates an online upgrade process: first, it verifies the integrity and compatibility of the update file, then replaces the old local model, and automatically loads the new model after the update is completed for subsequent inference calculations of multi-source perception data.

[0086] S104. The edge device performs network parameter configuration, traffic priority scheduling and operation status data collection operations, and synchronizes the operation status data to the cloud. The cloud generates a soft router optimization instruction based on the operation status data and sends it to the edge device for execution to adjust the network transmission performance parameters.

[0087] Specifically, network parameter configuration refers to the operation of setting and adjusting the underlying network attributes of each communication channel (such as 5G, WiFi 6) at the edge end according to the multi-channel converged communication requirements. Traffic priority scheduling refers to the management mechanism at the edge end that prioritizes multi-source data traffic (such as drone video streams, device status data streams, and control command streams) and allocates channel resources based on priority. Operational status data refers to various data collected in real time by the edge end that reflects its own device operation status and multi-channel network status. It mainly includes two types of data: first, device operation data, such as the computing power utilization rate (CPU / GPU load), memory usage rate, battery power (if it is a mobile terminal), device temperature, etc. of the heterogeneous computing units at the edge end; second, network status data, such as the real-time latency, packet loss rate, bandwidth utilization rate, signal strength of each channel, as well as the transmission success rate and data throughput after traffic scheduling. Soft routing optimization commands refer to commands generated by the cloud based on the operational status data synchronized at the edge end, used to adjust the network transmission strategy of the edge end. Its essence is the dynamic optimization control of the edge end's soft routing function (routing decisions are implemented through software, rather than hardware routing) by the cloud. For example, if cloud analytics finds that the bandwidth utilization of a certain edge 5G link has reached 90% (operational status data) and video streaming is experiencing stuttering, it will generate an instruction to "switch some medium-priority video streams to WiFi 6 link transmission"; if it finds that the edge memory usage is too high, causing data processing delays, it will generate an instruction to "reduce the sampling frequency of non-critical data to reduce memory consumption".

[0088] In specific implementation, the edge device performs network parameter configuration, traffic priority scheduling, and operation status data collection operations, and synchronizes the operation status data to the cloud, specifically including:

[0089] The edge device performs network parameter configuration through a containerized software routing system; the network parameter configuration includes automatic allocation of Internet Protocol addresses, configuration of network address translation rules and firewall policies;

[0090] A queue management algorithm is used for traffic priority scheduling, marking the UAV video stream data and target recognition result data as first priority; the first priority data is allocated transmission bandwidth first.

[0091] Real-time collection of operational status data from edge networks, synchronization of the collected operational status data to the cloud, activation of local caching when the network is interrupted, and resumption of transmission after network recovery.

[0092] Specifically, the edge device relies on a containerized software router system (such as software router software based on Docker) to perform network parameter configuration: the system automatically assigns independent Internet Protocol (IP) addresses to each access communication channel, such as 5G and WiFi 6 (e.g., dynamically assigning 192.168.1.x network segment addresses via DHCP service), ensuring that each channel can be uniquely identified at the network layer. Simultaneously, Network Address Translation (NAT) rules are configured to translate private IP addresses within the edge device's local area network to public network IP addresses (e.g., when accessing the cloud). Furthermore, firewall policies are set, such as allowing access requests from cloud IP addresses and blocking packets from unauthorized ports.

[0093] A queue management algorithm is used for traffic priority scheduling: The system first classifies and labels various types of data at the edge, marking real-time video stream data captured by drones (such as high-definition images at 25 frames per second) and target recognition result data (such as structured information containing target coordinates and categories) as first priority, and other data (such as device logs and non-real-time status information) as low priority. During data transmission, the algorithm allocates a dedicated high-priority queue for first-priority data. When multiple types of data are transmitted simultaneously, high-priority queue data is prioritized to occupy transmission bandwidth (e.g., when the 5G link bandwidth is 100Mbps, 60Mbps of dedicated bandwidth is reserved for first-priority data) to ensure that its transmission is not crowded out by low-priority data.

[0094] The edge device collects operational status data in real time through built-in monitoring components (such as network traffic monitoring tools and device status sensors). This data is packaged once per second and synchronously sent to the cloud-edge communication module in the cloud. If a network interruption is detected (such as three consecutive data transmission failures), a local caching mechanism is immediately triggered to temporarily store the operational status data to be transmitted in the local storage module of the edge device. Once the network is restored (such as when a response signal is detected from the cloud), the cached data is automatically read and transmitted to the cloud in chronological order to ensure that no data is lost.

[0095] Based on the aforementioned operational status data, the cloud generates software route optimization instructions and sends them to the edge for execution to adjust network transmission performance parameters. This involves the following processes:

[0096] The cloud receives synchronized operational status data from the edge, including network throughput, packet forwarding rate, channel occupancy, and processor load.

[0097] The cloud analyzes the operational status data, matches it with the target range corresponding to the operational status data, and generates software router optimization instructions. These software router optimization instructions include hardware passthrough enable instructions, processor frequency adjustment thresholds, and bandwidth allocation ratio parameters. Specifically, the hardware passthrough enable instruction controls the edge to enable single-root I / O virtualization technology (Single-root I / O virtualization (SR-IOV) is a PCIe-based I / O virtualization technology that virtualizes physical devices (such as network interface cards) into multiple virtual functions (VFs), allowing multiple logical partitions or virtual machines to directly share physical device resources, thereby improving performance and flexibility); the processor frequency adjustment threshold is used to dynamically adjust the processor's operating frequency; and the bandwidth allocation ratio parameter is used to optimize the bandwidth proportion of first-priority data.

[0098] The cloud sends the soft routing optimization command to the edge terminal through the message queue telemetry transmission protocol. The edge terminal receives and executes the soft routing optimization command to adjust the network transmission performance parameters.

[0099] In practice, the cloud receives operational status data synchronously uploaded by the edge terminal through the cloud-edge communication module. This data includes network throughput (e.g., 80Mbps for the current 5G link and 50Mbps for the WiFi 6 link), packet forwarding rate (e.g., 1200 packets forwarded per second at the edge), channel occupancy (e.g., 75% for the 5G link and 60% for the WiFi 6 link), and edge processor load (e.g., 72% CPU utilization). The cloud analyzes this operational status data item by item, comparing each data point with preset target ranges (e.g., network throughput should be maintained between 60-100Mbps, channel occupancy should be below 80%, and processor load should be controlled between 60%-80%). If the cloud finds that the 5G link occupancy is close to the threshold at 75% and the processor load is within a reasonable range but shows an upward trend at 72%, the cloud generates corresponding soft router optimization instructions based on the matching results. The generated software routing optimization commands include specific parameters: the hardware passthrough enable command is set to "enabled," which controls the edge to enable single-root I / O virtualization technology (such as PCIe passthrough mode) to reduce data forwarding latency at the virtualization layer; the processor frequency adjustment threshold is set to "frequency drops to 1.6GHz when load ≥ 80%, and rises to 2.4GHz when load ≤ 50%" to dynamically balance computing power and energy consumption; the bandwidth allocation ratio parameter is adjusted to "first priority data (video stream, target recognition results) accounts for 70% of the 5G link and 65% of the WiFi 6 link," prioritizing the transmission of critical data. The cloud encapsulates these software routing optimization commands into messages (including command type, specific parameter values, and timestamps) using the Message Queuing Telemetry Transport Protocol (MQTT), and sends them to the MQTT broker node at the edge via the cloud's MQTT client. The edge's MQTT client receives the message, parses the software routing optimization commands, and then performs the following operations:

[0100] According to the hardware pass-through enable command, PCIe pass-through is enabled through system kernel configuration, which directly maps the 5G modem to the application layer, reducing intermediate data transmission links.

[0101] Based on the processor frequency adjustment threshold, the CPU load is monitored in real time. When the load reaches 80%, the frequency is reduced to 1.6GHz, and when it is below 50%, the frequency is increased to 2.4GHz.

[0102] According to the bandwidth allocation ratio parameters, the traffic scheduling queue is adjusted to reserve 70% of the available bandwidth of the 5G link and 65% of the available bandwidth of the WiFi 6 link for the first priority data, thus completing the adjustment of network transmission performance parameters.

[0103] S105. Real-time monitoring of the operating status of the edge container and the status of the terminal device equipped with the edge. When the monitoring result matches the preset anomaly rule, the edge generates alarm information and transmits it to the cloud. The cloud generates a global security policy based on the alarm information and sends it to the edge to update the preset anomaly rule.

[0104] Specifically, container runtime status refers to the real-time running status of containers (such as Docker containers) at the edge that carry core functions such as soft routing systems, target detection models, and data transmission services. This includes the container's runtime status (e.g., "running," "stopped," "restarting"), resource usage (e.g., CPU usage, memory usage, disk I / O speed), network connectivity status (e.g., whether communication between the container and other modules is normal, and whether ports are open), and log output (e.g., whether error logs or exception messages are generated). Terminal device status refers to the real-time hardware and core functional status of terminal devices (such as drones) that work in conjunction with the edge. This includes hardware status (e.g., drone battery level, motor speed, propeller operating status, and GPS positioning signal strength), sensing device status (e.g., whether the camera is capturing images normally, whether the infrared sensor is outputting data, and whether the radar is scanning normally), flight status (e.g., flight altitude, flight speed, whether the heading is stable, and whether it deviates from the preset flight path), and communication status (e.g., whether the communication link between the drone and the edge is smooth and whether the data transmission rate meets the standard).

[0105] Preset anomaly rules refer to the standards and conditions pre-configured at the edge to determine whether the "container running status" and "terminal device status" are abnormal. They are the basis for triggering alarms and are set by technical personnel according to actual business needs. They usually exist in the form of "indicator thresholds and judgment logic".

[0106] Alarm information refers to structured information generated and transmitted to the cloud by the edge device after detecting that the container's operating status or the terminal device's status matches preset anomaly rules. It serves as the core carrier for the cloud to obtain edge device anomalies. Alarm information typically includes the anomaly type (such as "container resource overload", "drone GPS signal loss", etc.), the anomaly object (such as the ID of a specific container, the device number of the drone), the time of the anomaly (timestamp accurate to the second), the anomaly details (such as the container's current CPU utilization of 95%, the drone's GPS signal strength of -120dBm), and the current associated status (such as whether the anomaly caused data transmission interruption, target detection paused), etc.

[0107] Global security strategy refers to the overall strategy formulated by the cloud based on alarm information reported by edge devices, combined with the operational data of all edge devices and terminal devices in the global scope (such as whether multiple edge devices have the same type of container anomaly, or whether drones in a certain sea area frequently lose GPS signals), to optimize anomaly judgment and risk prevention.

[0108] In practical implementation, a monitoring agent component can be deployed at the edge to collect container runtime status data and terminal device status data in real time. Simultaneously, an anomaly rule base is preset at the edge. The monitoring agent component is configured to match the collected status data with the preset anomaly rule base; when a match is successful, an alarm message is generated. In this embodiment, the alarm message may include the anomaly type, occurrence timestamp, edge node identifier, and associated device ID.

[0109] The edge devices transmit alarm information to the cloud via a message queue telemetry transmission protocol. The cloud receives and stores alarm information from multiple edge devices, identifies anomaly types and correlation patterns through statistical analysis of frequency occurrence, and generates a global security policy. In this embodiment, the global security policy may include adding new anomaly rules, optimizing rule matching parameters, and anomaly response triggering conditions. The cloud distributes the global security policy to each edge device, and the edge devices update their local preset anomaly rule library upon receiving it.

[0110] Specifically, upon startup, the edge device deploys a monitoring agent component. This component collects container runtime status data in real time by calling container management interfaces (such as Docker Engine's REST API). For example, it retrieves specific container information every 5 seconds. Simultaneously, it collects terminal device status data every 3 seconds through the terminal device's communication interface (such as AT commands from a 4G module in a drone).

[0111] The edge device locally stores a pre-defined exception rule base. This rule base exists as a structured file containing multiple specific judgment rules, such as: "Container CPU utilization ≥ 90% for 3 consecutive collection cycles (i.e., 15 seconds)," "Container restarts ≥ 3 times within 10 minutes," "Drone battery level ≤ 20%," "GPS signal strength ≤ -110dBm for 2 consecutive collection cycles (i.e., 6 seconds)," and "Camera frame rate ≤ 5fps for 3 consecutive collection cycles (i.e., 9 seconds)." The monitoring agent component converts the real-time collected container and terminal device status data into a format recognizable by the rule base (e.g., converting "CPU utilization 85%" to the value "85") once per second, and matches it against each rule in the rule base: if a container's CPU utilization is 92%, 93%, and 91% for 3 consecutive collection cycles, it matches the rule "≥ 90% for 3 consecutive cycles"; if a terminal device's battery level is 18%, it directly matches the rule "≤ 20%." Upon successful matching, an alarm generation process is immediately triggered.

[0112] When alarm information is generated, the content is filled in according to a fixed template. For example, the anomaly type is clearly marked as a specific category such as "container CPU overload" or "drone low battery"; the occurrence timestamp is accurate to milliseconds; the edge node identifier is a unique number of that edge end; the associated device ID corresponds to the specific anomaly object (such as container ID "container_001" and drone ID "UAV_005"), and finally a structured text is formed.

[0113] The edge MQTT client encapsulates alarm information into messages according to protocol specifications (including the topic "alarm / edge_node_07" and the payload being alarm JSON data), and sends them to the cloud MQTT broker (such as an EMQX server deployed in the cloud) through an established TCP connection to ensure real-time information upload. After receiving all alarm information from the edge, the cloud MQTT broker stores it in a time-series database (such as InfluxDB) and processes it using data analysis tools (such as Grafana Loki): it counts the number of occurrences of various anomalies within 24 hours (such as "container CPU overload" occurring 28 times and "drone GPS signal weak" occurring 15 times), and analyzes correlation patterns (such as finding that 12 out of the 15 instances of "GPS signal weak" were accompanied by "5G link packet loss rate ≥10%").

[0114] Based on the analysis results, the cloud generates a global security policy: adding new anomaly rules (such as "synchronously monitoring GPS signals when the 5G link packet loss rate is ≥10%)); optimizing rule parameters (such as increasing the "drone low battery threshold" from 20% to 25% to avoid delays in emergency response due to rapid battery depletion); and adjusting trigger conditions (such as "triggering proactive cloud inquiry when the same type of anomaly occurs twice within one hour," whereas the original condition was three times). The cloud sends the global security policy to all edge devices in the form of MQTT messages. After receiving the message, the edge device compares it with the local rule library through the rule parsing module, automatically replacing old rules (such as updating the low battery threshold) and adding new rules (such as 5G and GPS associated monitoring rules), completing the update of the local preset anomaly rule library, and returning a "successful update" confirmation message to the cloud.

[0115] The method provided in this embodiment, in its first aspect, achieves multi-channel converged communication by constructing a layered collaborative architecture between the cloud and the edge: the cloud relies on a cluster scheduling module to coordinate edge cluster resources, a model management module to optimize and deploy core models, and a data storage module to store massive amounts of data; the edge receives instructions through an edge proxy component, an instruction execution unit to implement tasks, and a device twin module to synchronize states. This approach of "global cloud control and local edge execution" not only solves the problem of insufficient real-time performance of a single cloud in complex scenarios but also compensates for the limited computing power and storage of the edge, enabling dynamic allocation of multi-channel resources (such as allocating idle satellite channels to high-load edge devices). Simultaneously, local preprocessing of data at the edge (such as video stream compression) reduces the amount of data transmitted to the cloud, alleviating bandwidth pressure, ultimately achieving the effect of "low latency, high reliability, and wide coverage" in multi-channel converged communication, adapting to the needs of scenarios such as UAV inspection in complex sea areas.

[0116] Secondly, channel switching is achieved through a combination of a timing prediction model and an extended communication protocol to ensure channel quality. The timing prediction model, based on an LSTM network built with TensorFlow Lite, learns the dependencies of historical network state parameters (latency, packet loss rate, etc.) of 5G and WiFi 6 links through a gating unit, predicting the channel quality assessment value for a preset duration in the future, and determining the switching timing in advance to avoid the lag of traditional real-time judgment. The extended communication protocol adds a multi-path identifier field to MQTT to achieve parallel transmission of data fragments. It ensures uninterrupted connection through a session context mapping table to guarantee retransmission of incomplete data and a session persistence mechanism. If the new channel is unsuitable, it can quickly fall back to the original channel. This method of "predicting the switching timing and smoothly executing the switching" solves the problems of data loss and connection interruption during channel switching, ensuring stable communication quality.

[0117] Thirdly, a lightweight target detection model is introduced to improve the real-time performance of business processing: This model introduces a cross-modal attention module to focus on key features in the feature fusion layer of the basic detection model, optimizes sample weights using a focus loss function to improve the recognition accuracy in complex scenes, and reduces accuracy through model quantization tools to adapt to edge computing power. At the edge, the model is loaded with heterogeneous computing units to perform local inference on the multi-source perception data of the UAV, outputting target coordinates, category, and confidence level. Only the recognition results, rather than the original data, are transmitted to the cloud. This approach of "lightweight model and local inference" can not only meet the computing power limitations of the edge, but also reduce the bandwidth consumption of data transmission, avoid the delay of centralized processing in the cloud, improve the real-time performance of UAV target recognition, and ensure efficient business operations.

[0118] In summary, the three approaches support each other. The layered collaborative architecture provides basic support for channel switching and target detection, channel switching ensures the stability of data transmission channels, and lightweight target detection improves business processing efficiency. Together, they achieve high efficiency, reliability, and real-time performance of multi-channel converged communication in complex scenarios.

[0119] Corresponding to the aforementioned embodiment of a cloud-edge collaborative multi-channel fusion communication control method, the present invention also provides an embodiment of a cloud-edge collaborative multi-channel fusion communication control system.

[0120] like Figure 2 As shown, the cloud-edge collaborative multi-channel fusion communication control system in this embodiment mainly includes a cloud and an edge. The cloud performs cluster scheduling, model deployment, and data storage operations on the edge; the edge performs bidirectional data transmission, status information synchronization, and command response operations with the cloud.

[0121] Specifically, the edge end mainly performs the following operations:

[0122] The system collects network status parameters of multiple communication channels, outputs channel quality assessment values ​​based on a time-series prediction model, and triggers channel switching when the channel quality assessment values ​​meet preset switching conditions. Multi-channel data transmission is then performed through an extended communication protocol.

[0123] A lightweight and optimized target detection model is loaded, and inference calculations are performed on the multi-source perception data collected by the terminal device equipped with the edge terminal. The target recognition result is output and transmitted to the cloud.

[0124] Perform network parameter configuration, traffic priority scheduling, and operation status data collection operations, and synchronize the operation status data to the cloud;

[0125] Real-time monitoring of the operating status of edge containers and terminal devices; when the monitoring results match preset anomaly rules, alarm information is generated and transmitted to the cloud.

[0126] The cloud primarily performs the following functions:

[0127] Based on the running status data uploaded from the edge, a soft router optimization command is generated and sent to the edge for execution to adjust network transmission performance parameters;

[0128] A global security policy is generated based on the alarm information uploaded from the edge device, and then sent to the edge device to update the preset anomaly rules.

[0129] The system in this embodiment can be used to execute Figure 1 The steps of the method embodiment shown are similar in principle and process, and will not be repeated here.

[0130] The implementation process of the functions and roles of each unit in the above system is detailed in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0131] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0132] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cloud-edge collaborative multi-channel converged communication control method, characterized in that, The application relates to a layered collaborative architecture of cloud and edge, wherein the cloud performs cluster scheduling, model deployment and data storage operations on the edge, and the edge performs bidirectional data transmission, state information synchronization and instruction response operations with the cloud. The edge collects network state parameters of multiple communication channels, which are time series data including network delay, packet loss rate, signal strength and bandwidth utilization; the collected network state parameters are standardized and converted into feature vectors of a preset dimension, each feature vector corresponding to a link state at a sampling time point; The edge inputs the standardized feature vectors into a time series prediction model, which generates initial hidden states and cell states based on input data at historical time steps and learns the dependency relationship between parameters at different time steps; in the prediction process, the feature vector segment at the current time step and the hidden state at the previous time step are input each time to calculate the output at the current time step, and the output at the current time step is taken as the input at the next time step, and the calculation process is iteratively executed until the prediction of all time slices within a preset time length in the future is completed, and the network key performance indicator prediction value of each time slice within the preset time length in the future is output, and the channel quality evaluation value of each communication channel corresponding to each time slice is generated by weighted summation based on the prediction value; When the channel quality evaluation value is lower than a preset threshold, the edge establishes a multi-channel parallel transmission link by using an extended communication protocol; the extended communication protocol adds a multi-path identification field based on the message queue telemetry transmission protocol, and different fragments of the same data session are transmitted in parallel on multiple channels; in the channel switching process, a session context mapping table is automatically generated and maintained, which records the serial number, transmission state and corresponding transmission channel information of each data fragment, and a connection maintenance frame carrying a session identifier is sent to the cloud to maintain the connection state with the cloud by using the session maintenance mechanism after the protocol is extended; the buffered data that has not been transmitted is retransmitted through a new channel, and if the quality of the new channel does not meet the standard, the original channel is returned within a preset time length, and the protocol is downgraded to a single-path transmission mode; The edge loads a lightweight optimized target detection model, performs inference calculation on multi-source perception data collected by a terminal device carrying the edge, outputs result data and transmits the result data to the cloud; The edge performs network parameter configuration, traffic priority scheduling and running state data collection operations, and synchronizes the running state data to the cloud; the cloud generates soft routing optimization instructions based on the running state data and sends the instructions to the edge for execution to adjust the network transmission performance parameters of the edge; The container running state of the edge and the state of the terminal device are monitored in real time, and when the monitoring result matches a preset abnormal rule, the edge generates alarm information and transmits the alarm information to the cloud, and the cloud generates a global security policy based on the alarm information and sends the policy to the edge to update the preset abnormal rule. ​ ​ 2. The method of claim 1, wherein, The edge end loads the lightweight-optimized target detection model, performs inference calculation on the multi-source perception data collected by the terminal device equipped with the edge end, outputs result data, and transmits the result data to the cloud end. The edge end loads the lightweight-optimized target detection model; wherein the lightweight-optimized target detection model is obtained by introducing a cross-modal attention module at a feature fusion layer of a basic detection model, adjusting sample weights by using a focal loss function, and reducing model precision format by using a model quantization tool. The edge end receives multi-source perception data collected by a terminal device, inputs the lightweight-optimized target detection model, performs inference calculation on the multi-source perception data, and outputs result data. The edge end synchronizes the result data to the cloud end, and when a model update instruction issued by the cloud end is received, the edge end updates the target detection model by using an online upgrade method.

3. The method of claim 1, wherein, The edge end performs network parameter configuration, traffic priority scheduling, and collection of running state data, and synchronizes the running state data to the cloud end, and the process includes: The edge end performs network parameter configuration by using a containerized soft routing system; the network parameter configuration includes automatic allocation of Internet Protocol addresses, configuration of network address translation rules, and firewall policies; A queue management algorithm is used for traffic priority scheduling, and video stream data and target recognition result data are marked as first priority; the data of the first priority is preferentially allocated transmission bandwidth; The running state data of the edge end network is collected in real time, and the collected running state data is synchronized to the cloud end; when the network is interrupted, local caching is started, and when the network is restored, resuming transmission is performed.

4. The method of claim 3, wherein, The cloud end generates a soft routing optimization instruction based on the running state data, and issues the soft routing optimization instruction to the edge end to adjust the network transmission performance parameters of the edge end, and the process includes: The cloud end receives the running state data synchronized by the edge end, and the running state data includes network throughput, data packet forwarding rate, channel occupancy rate, and processor load; The cloud end analyzes the running state data, matches the target interval corresponding to the running state data, and generates a soft routing optimization instruction; the soft routing optimization instruction includes a hardware pass-through enable instruction, a processor frequency adjustment threshold, and a bandwidth allocation proportion parameter; wherein the hardware pass-through enable instruction is used to control the edge end to enable single-root I / O virtualization technology; the processor frequency adjustment threshold is used to dynamically adjust the running frequency of the processor; and the bandwidth allocation proportion parameter is used to optimize the bandwidth proportion of the first priority data; The cloud end issues the soft routing optimization instruction to the edge end by using a message queue telemetry transmission protocol, and the edge end receives and executes the soft routing optimization instruction to adjust the network transmission performance parameters thereof.

5. The method of claim 1, wherein, Real-time monitoring is performed on the container running state of the edge end and the state of the terminal device, and when the monitoring result matches a preset abnormal rule, the edge end generates alarm information and transmits the alarm information to the cloud end, the cloud end generates a global security policy based on the alarm information, and issues the global security policy to the edge end to update the preset abnormal rule, and the process includes: A monitoring agent component is deployed on the edge end to collect container running state data and terminal device state data in real time; A preset abnormal rule library is stored in the edge end. The monitoring agent component matches the collected state data with a preset abnormal rule library, generates an alarm information when a match is successful, and the alarm information contains an abnormal type, a timestamp, an edge node identifier, and an associated device ID; The edge end transmits the alarm information to the cloud end through a message queue telemetry transport protocol, the cloud end receives and stores the alarm information of the edge end, identifies an abnormal type and an associated rule through statistical analysis of the occurrence frequency, and generates a global security policy; the global security policy includes an added abnormal rule, rule matching parameter optimization, and an abnormal response trigger condition; The cloud end sends the global security policy to each edge end, and the edge end updates the local preset abnormal rule library after receiving the global security policy.

6. A cloud-edge collaborative multi-channel converged communication control system, characterized in that, The system comprises a cloud end and an edge end; wherein, The edge end performs bidirectional data transmission, state information synchronization, and instruction response operations with the cloud end, including: Collecting network state parameters of multiple communication channels, the network state parameters being time series data including network delay, packet loss rate, signal strength, and bandwidth utilization; performing standardization processing on the collected network state parameters to convert them into feature vectors of a preset dimension, each feature vector corresponding to a link state at a sampling time point; Inputting the standardized feature vectors into a time series prediction model, the time series prediction model generating an initial hidden state and a cell state based on input data at historical time steps, and learning the dependency relationship between parameters at different time steps; in the prediction process, a feature vector segment at the current time step and the hidden state at the previous time step are input each time, and the output at the current time step is calculated, the output at the current time step is taken as the input at the next time step, and the calculation process is iteratively executed until the prediction of all time slices within a preset time length in the future is completed, the network key performance indicator prediction value of each time slice within the preset time length in the future is output, and the channel quality evaluation value of each communication channel corresponding to each time slice is generated by weighted summation based on the prediction value; When the channel quality evaluation value is lower than a preset threshold, a multi-channel parallel transmission link is established using an extended communication protocol; the extended communication protocol adds a multi-path identifier field based on the message queue telemetry transport protocol, and different fragments of the same data session are transmitted in parallel on multiple channels; in the channel switching process, a session context mapping table is automatically generated and maintained, which records the serial number, transmission state, and corresponding transmission channel information of each data fragment, and a connection maintenance frame carrying a session identifier is sent to the cloud end to maintain the connection state with the cloud end using the session maintenance mechanism after the protocol is extended; the buffered data that has not been transmitted is retransmitted through a new channel, and if the quality of the new channel does not meet the standard, the original channel is returned within a preset time length, and the protocol is downgraded to a single-path transmission mode; Loading a lightweight optimized target detection model to perform inference calculation on multi-source perception data collected by a terminal device equipped with the edge end, outputting result data and transmitting the result data to the cloud end; Performing network parameter configuration, traffic priority scheduling, and running state data collection operations, and synchronizing the running state data to the cloud end; The edge end container running state and the terminal equipment state are monitored in real time, and when the monitoring result matches the preset abnormal rule, alarm information is generated and transmitted to the cloud end; The cloud end performs cluster scheduling, model deployment and data storage operations on the edge end, including: Generating multi-channel task allocation instructions and issuing them to the edge end for execution to adjust the channel usage strategy of the edge end in real time; Based on the running state data uploaded by the edge end, a soft routing optimization instruction is generated and issued to the edge end for execution to adjust the network transmission performance parameters of the edge end; Based on the alarm information uploaded by the edge end, a global security strategy is generated and issued to the edge end to update the preset abnormal rule.

7. The system of claim 6, wherein, The cloud end includes a cluster scheduling module, a cloud-edge communication module, a data storage module, and a model management module; the edge end includes an edge agent component, an instruction execution unit, and a device twin module; wherein, The cloud end generates task allocation instructions through the cluster scheduling module, which are issued to the edge agent component of the edge end through the cloud-edge communication module; the edge agent component distributes the task allocation instructions to the instruction execution unit for execution, and the execution results and the running state data of the edge end are arranged by the device twin module and then transmitted back to the cloud-end cloud-edge communication module through the edge agent component, and transmitted to the data storage module through the cloud-edge communication module for storage; the model management module optimizes the model based on the historical data stored in the data storage module, issues the updated model to the edge end, and forwards it to the instruction execution unit through the edge agent component to update the model; the updated model supports the execution of subsequent tasks of the edge end, and the new execution results and running state data are transmitted back to the cloud end again, forming a continuous optimization and collaborative closed loop.

Images