Robot-oriented AI-driven multi-network converged communication scheduling system

Abstract

The invention relates to the technical field of communication, in particular to a robot-oriented AI-driven multi-network converged communication scheduling system, which comprises a converged monitoring data flow generation module for acquiring network state, equipment state and task demand data in real time and generating a converged monitoring data flow; the AI dynamic scheduling strategy generation module inputs the data stream into an AI model, and outputs a strategy packet of a target network identifier, a bandwidth proportion and an abnormal threshold on line; and the scheduling strategy execution and abnormity diagnosis triggering module controls a multi-network interface to complete switching according to the strategy packet, marks a task data packet time sequence, reserves a bandwidth for a high-optimal task, and triggers traffic mirroring and fault diagnosis if a communication index exceeds a threshold value. According to the method, the dynamic strategy reasoning model driven by the graph convolutional network is constructed, and the priority-aware resource reservation and intelligent anomaly diagnosis mechanism is combined, so that efficient scheduling, stable communication and fault adaptive processing in the multi-network environment are realized.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and in particular, to an AI-driven multi-network fusion communication scheduling system for robots. Background Art

[0002] With the continuous growth of the number of mobile terminals and the wide deployment of multi-network environments (such as cellular networks, Wi-Fi, satellite communications), communication systems for task-driven scenarios (such as autonomous driving, industrial robots, emergency command) have put forward higher requirements for network handover stability, resource scheduling efficiency, and link anomaly perception capabilities. Especially in complex scenarios with frequent communication environment changes and different task priorities, how to achieve comprehensive perception of multi-source states, intelligent scheduling, and reliable communication guarantee has become an important research direction in the design of multi-network fusion communication systems.

[0003] Existing multi-network communication systems usually adopt static or rule-driven network selection mechanisms, which only make handover judgments based on partial network metrics (such as signal strength or latency), and it is difficult to comprehensively consider device status, task attributes, and network-wide resource loads, resulting in high network handover latency, low resource allocation efficiency, and lack of anomaly recognition and adaptive diagnosis capabilities. In addition, although some machine learning-based solutions introduce prediction models, they are mostly limited to time series modeling, and it is difficult to capture the structured relationships between multi-dimensional features and cannot effectively adapt to task dynamic changes and sudden link anomaly scenarios. Summary of the Invention

[0004] The present invention provides an AI-driven multi-network fusion communication scheduling system for robots, which improves the robustness, self-adaptability, and communication guarantee capabilities of the system in multi-network fusion scenarios.

[0005] An AI-driven multi-network fusion communication scheduling system for robots includes a fusion monitoring data stream generation module, an AI dynamic scheduling policy generation module, and a scheduling policy execution and anomaly diagnosis trigger module, wherein; The fusion monitoring data stream generation module collects multi-source data of the mobile terminal in real time, including network status data, device status data, and task requirement data, and generates a fusion monitoring data stream; The AI dynamic scheduling policy generation module inputs the fusion monitoring data stream into the AI dynamic scheduling model, and through offline learning of network handover rules and resource allocation patterns in the historical fusion monitoring data stream, outputs a dynamic scheduling policy packet including a target network identifier, a transmission bandwidth allocation ratio, and a group of anomaly detection thresholds online; The scheduling policy execution and exception diagnosis trigger module controls the multi-network interface chipset to perform network switching operations according to the dynamic scheduling policy package, and at the same time starts the task data packet timing marking engine to add timing tags to the task data packets to be transmitted, and allocates exclusive bandwidth resources for high-priority tasks through the elastic resource reservation channel. When it is detected that the actual communication metrics exceed the exception detection threshold group, a fault diagnosis process based on traffic mirroring is triggered.

[0006] Optionally, the fusion monitoring data stream generation module includes: Multi-source data collection: Through the multi-source sensor group deployed on the mobile terminal, multi-source data of the mobile terminal are collected in real time, including network status data, device status data, and task requirement data. The network status data includes the RSSI values of each operator's network, the channel error rate, and the base station load rate. The device status data includes the remaining battery power and the three-axis acceleration vector. The task requirement data includes the data volume size, the allowed maximum delay, and the priority identifier. Motion stability index calculation: Calculate the motion stability index based on the three-axis acceleration modulus and the battery power change rate. Timestamp alignment and data fusion: Attach a unified timestamp to the collected multi-source data, construct a structured monitoring data packet, and generate a check code based on the cyclic redundancy check, and output the fused monitoring data stream with check.

[0007] Optionally, the multi-source data collection includes: Network status data collection: Scan the operator network through the multi-band radio frequency probe array to obtain network status data in real time, including the RSSI values of each operator's network , the channel error rate , and the base station load rate ; Device status data collection: Collect device status data through the embedded motion co-processor, including the remaining battery power , and the three-axis acceleration vector ; Task requirement data collection: Extract task requirement data through the task parser, including the data volume size , the allowed maximum delay , and the priority identifier .

[0008] Optionally, the motion stability index is expressed as: ; Where is the motion stability index, is the three-axis acceleration vector, is the battery power change rate per unit time, , is the weighting coefficient.

[0009] Optionally, the AI dynamic scheduling policy generation module includes: Input data normalization: By parsing the structured data packets in the fusion monitoring data stream, multi-dimensional feature vectors at each moment are extracted, including network status parameters, device status parameters, and task requirement parameters, and Z-score normalization processing is performed on them to obtain normalized input data after eliminating the scale effect ; Policy inference: Input the normalized input data into the graph convolutional network (GCN) model to generate intermediate policies including the target network, bandwidth configuration, and anomaly threshold; Policy packet structuring: Decode the intermediate policies generated by policy inference, respectively determine the optimal target network identifier, the transmission bandwidth allocation ratios of the three operators, and the delay threshold and bit error rate threshold for communication status judgment, and encapsulate them into a standardized dynamic scheduling policy packet.

[0010] Optionally, the policy inference includes: Construct a feature map input matrix: Use the obtained normalized input data as the input of the graph convolutional network (GCN) model to construct a node feature matrix; Set the feature association graph structure: Construct a static adjacency matrix between features , which is used to describe the correlation between features, where represents that there is a structural dependence relationship between the th feature and the th feature, represents no connection; Graph convolutional feature fusion calculation: Perform one layer of graph convolutional operation to generate a fused hidden feature matrix ; Policy output vector generation: Flatten the fused hidden feature matrix into a vector , and input it into a multi-layer perceptron (MLP) to output intermediate policies, including the scores of three candidate target networks, the original bandwidth allocation scores of the three operators, and the benchmark value outputs of the delay threshold and bit error rate threshold.

[0011] Optionally, the policy packet structuring includes: Target network identifier recognition: Determine the number of the current optimal target network through Softmax normalization and maximum value position selection; Transmission bandwidth allocation ratio calculation: Normalize the bandwidth allocation scores of the three operators through Softmax to generate comparable ratio values; Abnormal detection threshold group calculation: Output the delay threshold and the bit error rate threshold by multiplying the scaling factor and adding the offset constant. Structured output of the policy package: Package the optimal target network identifier, the transmission bandwidth allocation ratio, and the delay threshold and the bit error rate threshold into a structured policy package as the final issued control instruction.

[0012] Optionally, the scheduling policy execution and abnormal diagnosis trigger module includes: Network switching and timing marking: According to the number of the target network specified in the dynamic scheduling policy package, control the multi-network interface chip to complete the network switching operation, and achieve seamless transfer by setting the radio frequency channel switching delay. At the same time, start the timing marking engine of the task data packet to add a unique identification header including the sending timestamp, the path identification hash value, and the global counter to each task data packet to be transmitted. Elastic resource reservation: Judge whether to enable the resource reservation mechanism according to the priority of the task, calculate the required bandwidth for the urgent task and combine it with the target network bandwidth ratio allocated in the policy to dynamically calculate the minimized reserved bandwidth value. At the same time, issue a flow table rule including the source address, the destination address, the reserved bandwidth, and the time-to-live parameter through the SDN controller. Abnormal diagnosis trigger: Continuously monitor the key performance indicators in the communication process, including the actual signaling delay and the bit error rate, and compare them with the issued delay threshold and the bit error rate threshold. When any index exceeds the delay threshold or the bit error rate threshold, trigger the traffic mirroring mechanism, calculate the mirror sampling rate according to the bit error rate deviation, and call the lightweight convolutional neural network to quickly diagnose the mirror traffic, and generate an abnormal type code for reporting and response.

[0013] Optionally, the elastic resource reservation includes: Priority determination and trigger condition judgment: Judge whether the current task is an urgent task. When the task priority meets the threshold condition start the elastic bandwidth reservation process, where is the urgent task threshold; Bandwidth demand and reserved bandwidth calculation: Calculate the required bandwidth according to the data volume and delay requirements of the task, and combine the allocation ratio and the maximum bandwidth of the target network to calculate the final reserved bandwidth. SDN flow table issuance and exclusive path configuration: Generate and issue a flow table through the software-defined network (SDN) controller to configure a bandwidth guarantee path for high-priority tasks.

[0014] Optionally, the abnormal diagnosis trigger includes: Abnormal condition judgment: Real-time collect the key performance indicators of the communication and compare them with the delay threshold and the bit error rate threshold issued by the policy package. When it indicates that there is a communication anomaly and trigger an abnormal response, where is the actual signaling delay, is the actual bit error rate; Traffic mirroring startup and mirroring rate calculation: When an abnormal trigger condition is met, the traffic mirroring mechanism is enabled and the mirroring sampling rate is dynamically adjusted based on the bit error rate deviation; Abnormal diagnosis execution and fault coding output: The collected mirrored traffic is input into a lightweight convolutional neural network for rapid feature recognition and generation of corresponding abnormality type codes.

[0015] Beneficial effects of the present invention: The present invention deploys multi-source sensors and task parsing components to build a fused monitoring data stream, achieving multi-dimensional collaborative perception of network status, device status, and task requirements. It can not only dynamically capture key network parameters such as RSSI value, channel bit error rate, and base station load rate, but also integrate factors such as remaining power, motion stability, and task delay sensitivity, greatly enhancing the real-time visibility of the communication environment and task characteristics.

[0016] The present invention inputs standardized features into a graph convolutional network (GCN) model and combines the topological correlation between features to model the communication state space. Without relying on time series, it can efficiently output the target network identification, bandwidth allocation ratio, and dynamic anomaly threshold group. Assisted by a strategy decoding module, structured scheduling instructions are generated to achieve accurate selection of the optimal communication path and resource allocation strategy in a multi-network environment.

[0017] The present invention, through the scheduling strategy execution and abnormal diagnosis triggering module, can dynamically calculate the required bandwidth of high-priority tasks and reserve physical channel resources when they are discovered, and realize end-to-end exclusive path construction based on the SDN controller. At the same time, it monitors the performance indicators of the communication link in real time and combines it with the dynamic threshold judgment mechanism. Once an abnormality is detected, it triggers mirror sampling and calls a lightweight convolutional neural network for online diagnosis, quickly locates the abnormality type and provides feedback, thereby improving the communication robustness and reliability of the system in complex network environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 Schematic diagram of system function modules according to an embodiment of the present invention; Figure 2 Schematic diagram of the AI dynamic scheduling strategy generation module according to an embodiment of the present invention. Detailed implementation mode

[0020] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art can also adopt other alternative methods for implementation; moreover, the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0021] As Figure 1 - Figure 2 shown, an AI-driven multi-network fusion communication scheduling system for robots includes a fusion monitoring data stream generation module, an AI dynamic scheduling strategy generation module, and a scheduling strategy execution and anomaly diagnosis trigger module, where; The fusion monitoring data stream generation module collects multi-source data of the mobile terminal in real time, including network status data, device status data, and task requirement data, and generates a fusion monitoring data stream; The AI dynamic scheduling strategy generation module inputs the fusion monitoring data stream into the AI dynamic scheduling model, and through offline learning of the network switching rules and resource allocation patterns in the historical fusion monitoring data stream, outputs a dynamic scheduling strategy packet including a target network identifier, a transmission bandwidth allocation ratio, and an anomaly detection threshold group online; The scheduling strategy execution and anomaly diagnosis trigger module controls the multi-network interface chipset to perform network switching operations according to the dynamic scheduling strategy packet, and at the same time starts the task data packet timing marking engine to add timing tags to the task data packets to be transmitted, and allocates exclusive bandwidth resources for high-priority tasks through the elastic resource reservation channel. When it is detected that the actual communication metrics exceed the anomaly detection threshold group, a fault diagnosis process based on traffic mirroring is triggered.

[0022] The fusion monitoring data stream generation module includes: Multi-source data collection: Scanning the operator network through a multi-band radio frequency probe array to obtain network status data in real time, including the RSSI values of each operator network , channel error rate , base station load rate , collecting device status data through an embedded motion coprocessor, including remaining battery power , three-axis acceleration vector , extracting task requirement data through a task parser, including data volume size , allowable maximum delay , priority identifier , expressed as: ; Among them, is the number of error bits, is the total number of transmitted bits; ; Among them, is the number of used resource blocks, and is the total number of resource blocks; Calculation of motion stability index: Calculate the motion stability index based on the triaxial acceleration modulus and the rate of change of power , expressed as: ; where, is the triaxial acceleration vector, is the rate of change of power within a unit time, is the weighting coefficient; Timestamp alignment and data fusion: Attach a unified timestamp to the multi-source data collected , construct a structured monitoring data packet , generate a check code based on cyclic redundancy check , and output a fused monitoring data stream with check , where, is the total number of sampling rounds, is the 32-bit cyclic redundancy check function.

[0023] The AI dynamic scheduling policy generation module includes: Input data standardization: By parsing the structured data packets in the fused monitoring data stream, extract the multi-dimensional feature vectors at each moment, including network status parameters, device status parameters, and task requirement parameters, and perform Z-score standardization processing on them to obtain the normalized input data after eliminating the scale effect , expressed as: ; ; where, is the mean vector of historical fused monitoring data, is the standard deviation vector of historical fused monitoring data; Policy inference: Input the normalized input data into the graph convolutional network (GCN) model to generate intermediate policies including the target network, bandwidth configuration, and anomaly threshold, specifically including: (1) Construct a feature map input matrix: Use the obtained normalized input data as the input of the graph convolutional network (GCN) model, and construct a node feature matrix, expressed as: ; where, is the initial node feature matrix; (2) Set the feature correlation graph structure: Construct a static adjacency matrix of features , used to describe the correlation between features, where, represents the There is a structural dependency between a feature and the feature, indicating no connection; (3) Graph Convolution Feature Fusion Calculation: Perform one layer of graph convolution operation to generate a fused hidden feature matrix, denoted as: ; where, is the adjacency matrix plus the identity matrix to form a graph with self - connection, is the diagonal matrix, , is the trainable weight parameter of the graph convolutional network, is the ReLU non - linear activation function, is the fused hidden feature matrix; (4) Policy Output Vector Generation: Flatten the fused hidden feature matrix into a vector , and input it into a multi - layer perceptron (MLP) to output an intermediate policy, denoted as: ; where, is the first - layer weight matrix, is the first - layer bias vector, is the first - layer ReLU activation function, is the first - layer output feature; ; where, is the second - layer weight matrix, is the second - layer bias vector, is the second - layer activation function, is the second - layer intermediate output; ; ; ; where, is the score of three types of candidate target networks, is the original bandwidth allocation score of three operators, is the benchmark value output of the delay threshold and the bit error rate threshold; Policy Package Structuring: Decode the intermediate policy generated by policy inference, respectively determine the optimal target network identifier, the transmission bandwidth allocation ratio of three operators, and the delay threshold and bit error rate threshold for communication status judgment, and encapsulate them into a standardized dynamic scheduling policy package, specifically including: (1)Target network identifier recognition: By means of Softmax normalization and maximum position selection, determine the number of the current optimal target network, expressed as: ; where is the number of the optimal target network; (2)Calculation of transmission bandwidth allocation ratio: Normalize the bandwidth allocation scores of the three operators through Softmax to generate comparable ratio values, expressed as: ; where is the original output of the bandwidth allocation of the th operator, is the normalized bandwidth ratio of the th operator; (3)Calculation of anomaly detection threshold group: Output the delay threshold and the bit error rate threshold by multiplying the scaling factor and adding the offset constant, expressed as: ; ; where is the anomaly detection reference value output by the policy inference, is the delay threshold, is the bit error rate threshold, , are the threshold scaling factors, , are the offset correction constants; (4)Structured output of policy package: Package the optimal target network identifier, the transmission bandwidth allocation ratio, and the delay threshold and the bit error rate threshold into a structured policy package as the final issued control instruction, expressed as: .

[0024] The scheduling policy execution and anomaly diagnosis trigger module includes: Network switching and timing marking: According to the number of the target network specified in the dynamic scheduling policy package, control the multi-network interface chip to complete the network switching operation, and achieve seamless transfer by setting the radio frequency channel switching delay. At the same time, start the timing marking engine of the task data packet to add a unique identification header including the sending timestamp, the path identification hash value, and the global counter to each task data packet to be transmitted, expressed as: ; where is the control signal for controlling the multi-network interface chip group, is the radio frequency channel mapping function, For switching the protection delay (set to 20 ms); ; Wherein, is the timestamp when the task data packet is sent, is the current path identifier, is the SHA-256 hash function, is the globally incrementing counter, is the identification header; Elastic resource reservation: Determine whether to enable the resource reservation mechanism according to the priority of the task. For urgent tasks, calculate the required bandwidth and combine it with the target network bandwidth ratio allocated in the policy to dynamically calculate the minimized reserved bandwidth value. At the same time, issue a flow table rule including source address, destination address, reserved bandwidth, and time-to-live parameters through the SDN controller. Specifically, it includes: (1) Priority determination and trigger condition judgment: Determine whether the current task is an urgent task. When the task priority meets the threshold condition start the elastic bandwidth reservation process, wherein, is the urgent task threshold; (2) Bandwidth requirement and reserved bandwidth calculation: Calculate the required bandwidth according to the data volume and delay requirements of the task, and combine it with the allocation ratio and maximum bandwidth of the target network to calculate the final reserved bandwidth, expressed as: ; ; Wherein, is the task data volume, is the maximum allowable delay, is the target network bandwidth ratio issued in the policy, is the maximum physical bandwidth of the target network, is the reserved margin coefficient; (3) SDN flow table issuance and dedicated path configuration: Generate and issue a flow table through the software-defined network (SDN) controller, and configure a bandwidth guarantee path for high-priority tasks, expressed as: ; Wherein, is the source IP address, is the destination IP address, is the reserved bandwidth value, is the flow table rule lifecycle; Abnormal diagnosis trigger: Continuously monitor the key performance indicators during the communication process, including the actual signaling delay and the bit error rate, and compare them with the issued delay threshold and bit error rate threshold. When any of the indicators exceeds the delay threshold or the bit error rate threshold, trigger the traffic mirroring mechanism, calculate the mirror sampling rate according to the bit error rate deviation, and call the lightweight convolutional neural network to quickly diagnose the mirrored traffic, generating an abnormal type code for reporting and response. Specifically, it includes: (1) Abnormal condition judgment: Real-time collect the key performance indicators of communication and compare them with the delay threshold and bit error rate threshold issued by the policy package. When holds, it indicates that there is a communication anomaly and an abnormal response is triggered. Among them, is the actual signaling delay, is the actual bit error rate; (2) Traffic mirroring start and mirror rate calculation: When the abnormal trigger condition is established, start the traffic mirroring mechanism and dynamically adjust the mirror sampling rate according to the bit error rate deviation , which is expressed as: Among them, is the mirror adjustment coefficient; (3) Abnormal diagnosis execution and fault code output: Input the collected mirrored traffic into the lightweight convolutional neural network for quick feature recognition and generate the corresponding abnormal type code, which is expressed as: ; Among them, is the abnormal diagnosis result code, is the abnormal diagnosis convolutional neural network model, is the sampled communication data stream; Abnormal diagnosis convolutional neural network model is expressed as: ; Among them, is the first layer of one-dimensional convolution, and the convolution kernel size is , is the second layer of one-dimensional convolution, and the convolution kernel size is , is the rectified linear unit activation function, is to flatten the convolution output into a one-dimensional vector, , are the fully connected weights and biases of the output layer respectively, is the index corresponding to the maximum probability of the output as the fault classification code, is the fault type code (such as 0: normal, 1: high delay, 2: high bit error, 3: link instability, etc.).

[0025] The present invention covers any alternatives, modifications, equivalent methods, and solutions made to the essence and scope of the present invention. For the public to have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, and those skilled in the art can fully understand the present invention without the description of these details. In addition, well-known methods, processes, procedures, components, and circuits are not described in detail to avoid unnecessary confusion to the essence of the present invention.

[0026] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An AI-driven multi-network fusion communication scheduling system for robots, characterized in that, It includes a fusion monitoring data stream generation module, an AI dynamic scheduling policy generation module, and a scheduling policy execution and anomaly diagnosis trigger module, where; The fusion monitoring data stream generation module collects multi-source data of the mobile terminal in real time, including network status data, device status data, and task requirement data, and generates a fusion monitoring data stream; The AI dynamic scheduling policy generation module inputs the fusion monitoring data stream into the AI dynamic scheduling model, and through offline learning of the network switching rules and resource allocation patterns in the historical fusion monitoring data stream, outputs a dynamic scheduling policy package including a target network identifier, a transmission bandwidth allocation ratio, and an anomaly detection threshold group online; The scheduling policy execution and anomaly diagnosis trigger module controls the multi-network interface chipset to perform network switching operations according to the dynamic scheduling policy package, and at the same time starts the task data packet timing marking engine to add timing tags to the task data packets to be transmitted, and allocates exclusive bandwidth resources for high-priority tasks through the elastic resource reservation channel. When it detects that the actual communication metrics exceed the anomaly detection threshold group, it triggers a traffic mirroring-based fault diagnosis process.

2. The AI-driven multi-network fusion communication scheduling system for robots according to claim 1, wherein The fusion monitoring data stream generation module includes: Multi-source data collection: Through a multi-source sensor group deployed on the mobile terminal, collect multi-source data of the mobile terminal in real time, including network status data, device status data, and task requirement data. The network status data includes the RSSI values of each operator's network, the channel error rate, and the base station load rate. The device status data includes the remaining battery power and the three-axis acceleration vector. The task requirement data includes the data volume size, the maximum allowable delay, and the priority identifier; Motion stability index calculation: Calculate the motion stability index based on the three-axis acceleration modulus and the battery power change rate; Timestamp alignment and data fusion: Attach a unified timestamp to the collected multi-source data, construct a structured monitoring data packet, and generate a check code based on the cyclic redundancy check, and output a fusion monitoring data stream with a check; 3. The AI-driven multi-network fusion communication scheduling system for robots according to claim 2, wherein The multi-source data collection includes: Network status data collection: Scanning the operator network through a multi-band RF probe array to obtain network status data in real time, including the RSSI values of each operator network , channel error rate , base station load rate ; Device status data acquisition: Collect device status data through an embedded motion co-processor, including remaining battery power , three-axis acceleration vector ; Task requirement data collection: Extract task requirement data through a task parser, including the amount of data , maximum allowable latency , priority identifier .

4. An AI-driven multi-network fusion communication scheduling system for robots according to claim 3, characterized in that, The motion stability index is expressed as: ； Among them, is the motion stability index, is the three-axis acceleration vector, is the rate of change of electric quantity per unit time, , are the weighting coefficients.

5. The AI-driven multi-network fusion communication scheduling system for robots according to claim 4, characterized in that, The AI dynamic scheduling policy generation module includes: Input data standardization: By parsing the structured data packets in the fusion monitoring data stream, multi-dimensional feature vectors at each moment are extracted, including network state parameters, device state parameters, and task requirement parameters, and Z-score standardization processing is performed on them to obtain the normalized input data after eliminating the scale effect. ; Policy inference: Input the normalized input data into the graph convolutional network model to generate an intermediate policy including the target network, bandwidth configuration, and anomaly threshold; Policy package structuring: Decode the intermediate policy generated by the policy inference, respectively determine the optimal target network identifier, the transmission bandwidth allocation ratios of the three operators, and the delay threshold and error rate threshold for communication status judgment, and encapsulate them into a standardized dynamic scheduling policy package.

6. An AI-driven multi-network fusion communication scheduling system for robots according to claim 5, characterized in that, The policy inference includes: Construct the input matrix of the feature map: Use the obtained normalized input data as the input of the graph convolutional network model to construct the node feature matrix; Set the feature association graph structure: construct the static adjacency matrix between features , which is used to describe the correlation between features. Among them, represents that there is a structural dependence relationship between the -th feature and the -th feature, represents no connection; Graph Convolution Feature Fusion Calculation: Perform one layer of graph convolution operation to generate a fused hidden feature matrix ; Policy output vector generation: Flatten the fused hidden feature matrix into a vector , and input it into a multi-layer perceptron to output an intermediate policy, including the scores of three types of candidate target networks, the original bandwidth allocation scores of three operators, and the baseline value outputs of the delay threshold and the bit error rate threshold.

7. An AI-driven multi-network fusion communication scheduling system for robots according to claim 6, characterized in that, The policy package structuring includes: Target network identifier identification: Determine the number of the current optimal target network through Softmax normalization and maximum value position selection; Transmission bandwidth allocation ratio calculation: Normalize the bandwidth allocation scores of the three operators through Softmax to generate comparable ratio values; Anomaly detection threshold group calculation: Output the delay threshold and error rate threshold by multiplying by a scaling factor and adding an offset constant; Structured output of the policy package: Package the optimal target network identifier, transmission bandwidth allocation ratio, and delay threshold and bit error rate threshold into a structured policy package as the final issued control instruction.

8. An AI-driven multi-network fusion communication scheduling system for robots according to claim 7, characterized in that, The scheduling policy execution and exception diagnosis trigger module includes: Network switching and timing marking: According to the number of the target network specified in the dynamic scheduling policy package, control the multi-network interface chip to complete the network switching operation, and achieve seamless transfer by setting the radio frequency channel switching delay. At the same time, start the timing marking engine of the task data packet to add a unique identification header including the sending timestamp, path identification hash value, and global counter to each task data packet to be transmitted; Elastic resource reservation: Determine whether to enable the resource reservation mechanism according to the priority of the task, calculate the required bandwidth for the urgent task, and combine the target network bandwidth ratio allocated in the policy to dynamically calculate the minimized reserved bandwidth value. At the same time, issue a flow table rule including source address, destination address, reserved bandwidth, and time-to-live parameters through the SDN controller; Exception diagnosis trigger: Continuously monitor the key performance indicators in the communication process, including the actual signaling delay and bit error rate, and compare them with the issued delay threshold and bit error rate threshold. When any index exceeds the delay threshold or bit error rate threshold, trigger the traffic mirroring mechanism, calculate the mirror sampling rate according to the bit error rate deviation, and call the lightweight convolutional neural network to quickly diagnose the mirrored traffic and generate an exception type code for reporting and response.

9. An AI-driven multi-network fusion communication scheduling system for robots according to claim 8, characterized in that The elastic resource reservation includes: Priority determination and trigger condition judgment: Determine whether the current task is an urgent task. When the task priority meets the threshold condition start the elastic bandwidth reservation process, where is the threshold for urgent tasks; Bandwidth demand and reserved bandwidth calculation: Calculate the required bandwidth according to the data volume and delay requirements of the task, and combine the allocation ratio of the target network and the maximum bandwidth to calculate the final reserved bandwidth; SDN flow table issuance and exclusive path configuration: Generate and issue a flow table through the software-defined network controller to configure a bandwidth guarantee path for high-priority tasks to use.

10. An AI-driven multi-network fusion communication scheduling system for robots according to claim 9, characterized in that, The exception diagnosis trigger includes: Abnormal condition judgment: Collect key communication performance indicators in real time and compare them with the latency threshold and bit error rate threshold issued by the policy package. When occurs, it indicates that there is a communication anomaly and an abnormal response is triggered. Among them, is the actual signaling delay, is the actual bit error rate; Traffic mirroring start and mirror rate calculation: When the exception trigger condition is established, start the traffic mirroring mechanism and dynamically adjust the mirror sampling rate according to the bit error rate deviation; Exception diagnosis execution and fault code output: Input the collected mirrored traffic into the lightweight convolutional neural network for rapid feature recognition and generate the corresponding exception type code.

Images