A Deep Learning-Based Intelligent Scheduling Method and System for Gateway Data Transmission

Abstract

This application provides a deep learning-based intelligent scheduling method and system for gateway data transmission. The method includes: acquiring three-dimensional feature tensors corresponding to signaling streams, media streams, and user behavior data; performing spatiotemporal feature joint prediction on the three-dimensional feature tensors based on a pre-set dual-channel LSTM neural network to output a link bandwidth demand matrix and a risk warning level for burst traffic; establishing a bandwidth allocation decision tree based on the link bandwidth demand matrix, and determining a real-time bandwidth threshold based on the bandwidth allocation decision tree and the risk warning level; constructing a reinforcement learning state vector based on the current communication quality indicators, the link bandwidth demand matrix, and the real-time bandwidth threshold; parsing the reinforcement learning state vector through a deep Q-network to obtain path decision instructions, and sending the path decision instructions to the corresponding gateway routing nodes.

Description

Technical Field

[0001] This application relates to the field of data transmission technology, and in particular to a gateway data transmission intelligent scheduling method and system based on deep learning. Background Technology

[0002] Traditional gateway data transmission scheduling algorithms (such as round-robin and weighted fair queues) allocate bandwidth based on static rules or fixed weights, making it difficult to cope with the multidimensional challenges of dynamic network environments. Existing technologies rely on preset service priorities and cannot perceive the correlation characteristics of signaling streams, media streams, and user behavior in real time, leading to inaccurate prediction of burst traffic and causing link congestion or resource idleness. For example, in heterogeneous service mixed scenarios, the round-robin algorithm cannot identify the spatiotemporal burstiness of video streams, causing a surge in latency for high-priority voice services; while the weighted fair queue introduces weight allocation, the weight adjustment lags behind changes in network state, resulting in bandwidth utilization of less than 60%. Traditional methods also lack the ability to model the dynamic evolution of network topology. When node failures or link load changes abruptly, decision tree models based on historical data cannot generate adaptive routing strategies due to the fixed feature dimensions, resulting in a packet loss rate that is more than 30% higher than the optimal solution. These problems collectively lead to a significant degradation of QoS indicators in dynamic scenarios such as 5G edge computing and industrial IoT using existing gateway scheduling methods. Summary of the Invention

[0003] This application provides a deep learning-based intelligent scheduling method and system for gateway data transmission, which can improve the bandwidth utilization of gateway data.

[0004] In a first aspect, embodiments of this application provide a gateway data transmission intelligent scheduling method based on deep learning, the method comprising:

[0005] Obtain the three-dimensional feature tensors corresponding to signaling streams, media streams, and user behavior data;

[0006] Based on a pre-built dual-channel LSTM neural network, the spatiotemporal features of the three-dimensional feature tensor are jointly predicted, and the link bandwidth demand matrix and the risk warning level of burst traffic are output.

[0007] A bandwidth allocation decision tree is established based on the link bandwidth demand matrix, and a real-time bandwidth threshold is determined based on the bandwidth allocation decision tree and the risk warning level.

[0008] Construct a reinforcement learning state vector based on the current communication quality indicators, the link bandwidth requirement matrix, and the real-time bandwidth threshold;

[0009] The reinforcement learning state vector is analyzed using a deep Q-network to obtain path decision instructions.

[0010] Secondly, embodiments of this application provide a gateway data transmission intelligent scheduling system based on deep learning. This system is used to execute the gateway data transmission intelligent scheduling method based on deep learning as described in any one of the embodiments of this application. The system includes:

[0011] The data acquisition module is used to acquire the three-dimensional feature tensors corresponding to signaling streams, media streams, and user behavior data;

[0012] The data analysis module is used to perform spatiotemporal feature joint prediction on the three-dimensional feature tensor based on a pre-set dual-channel LSTM neural network, and output the link bandwidth demand matrix and the risk warning level of burst traffic.

[0013] The risk prediction module is used to establish a bandwidth allocation decision tree based on the link bandwidth demand matrix, and to determine a real-time bandwidth threshold based on the bandwidth allocation decision tree and the risk warning level.

[0014] The vector construction module is used to construct reinforcement learning state vectors based on the current communication quality indicators, the link bandwidth requirement matrix, and the real-time bandwidth threshold.

[0015] The instruction output module is used to parse the reinforcement learning state vector through a deep Q-network to obtain path decision instructions.

[0016] Thirdly, embodiments of this application provide an electronic device, which includes a memory and a processor;

[0017] The memory is used to store computer programs;

[0018] The processor is configured to execute the computer program and, when executing the computer program, implement the gateway data transmission intelligent scheduling method based on deep learning as described in any of the embodiments of this application.

[0019] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to implement the deep learning-based gateway data transmission intelligent scheduling method as described in any of the embodiments of this application.

[0020] This application provides a deep learning-based intelligent scheduling method for gateway data transmission. The method includes: acquiring three-dimensional feature tensors corresponding to signaling streams, media streams, and user behavior data; performing spatiotemporal feature joint prediction on the three-dimensional feature tensors based on a pre-set dual-channel LSTM neural network, outputting a link bandwidth demand matrix and a risk warning level for burst traffic; establishing a bandwidth allocation decision tree based on the link bandwidth demand matrix, and determining a real-time bandwidth threshold based on the bandwidth allocation decision tree and the risk warning level; constructing a reinforcement learning state vector based on current communication quality indicators, the link bandwidth demand matrix, and the real-time bandwidth threshold; and obtaining path decision instructions by parsing the reinforcement learning state vector through a deep Q-network. In this method, by integrating signaling streams, media streams, and user behavior data to obtain three-dimensional feature tensors, the dual-channel LSTM neural network employs a spatiotemporal feature joint prediction mechanism to reduce bandwidth demand prediction errors. Then, dynamic weighted fusion of spatiotemporal features is achieved through an attention weight matrix, enhancing the sensitivity of burst traffic warnings. Combining the reinforcement learning state vector with the deep Q-network, a closed-loop optimization system is constructed. Under dynamic bandwidth threshold control, this improves the generation speed of path decision instructions while reducing network latency fluctuations. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A schematic flowchart illustrating a gateway data transmission intelligent scheduling method based on deep learning, provided for an embodiment of this application;

[0023] Figure 2 This is a schematic block diagram of a gateway data transmission intelligent scheduling system based on deep learning, provided as an embodiment of this application. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0026] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0027] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0028] Please see Figure 1 , Figure 1 This is a schematic flowchart illustrating a deep learning-based intelligent scheduling method for gateway data transmission, provided in an embodiment of this application. Figure 1 As shown, the specific steps of this deep learning-based gateway data transmission intelligent scheduling method include: S101-S105.

[0029] S101. Obtain the three-dimensional feature tensors corresponding to the signaling stream, media stream, and user behavior data.

[0030] For example, a three-dimensional feature tensor can be constructed by fusing multi-source data to comprehensively characterize the dynamic features of the network. Signaling stream data (such as SIP protocol messages) reflects the connection status of the network control plane, media stream data (such as the bitrate and packet loss rate of RTP / RTCP packets) records the real-time transmission load, and user behavior data (such as terminal request type and frequency) provides the business layer context. The dimensions of the three-dimensional tensor are designed as time window × number of nodes × number of features. For example, using a 5-minute time window slice, 30-dimensional features (signaling interaction frequency, media stream bitrate, user request distribution, etc.) from 10 nodes can be aggregated. During implementation, data timestamps need to be aligned (such as NTP protocol synchronization), and normalized units need to be normalized (such as mapping latency to the 0, 1 interval), and missing values ​​need to be filled in using a sliding window. For example, after collecting data in real time from a Kafka message queue, Spark Structured Streaming is used to aggregate data in 5-second granular windows to generate a tensor with dimensions of 60×10×30 (60 time slices, 10 nodes, 30-dimensional features).

[0031] S102. Based on a pre-built dual-channel LSTM neural network, perform spatiotemporal feature joint prediction on the three-dimensional feature tensor, and output the link bandwidth demand matrix and the risk warning level of burst traffic.

[0032] For example, a dual-channel LSTM neural network consists of a temporal channel (stacked bidirectional LSTM) and a spatial channel (Graph Convolutional Network, GCN), which handle temporal dependencies and topological associations respectively. The temporal channel decomposes the three-dimensional tensor into the temporal features of each node (such as the bit rate change per second of a node), and captures long-short-term patterns (such as periodic burst traffic) through the bidirectional LSTM. The spatial channel constructs a node topology graph (the adjacency matrix is ​​based on physical link distance and logical coupling weights), and uses the GCN to aggregate the load features of neighboring nodes (such as the bandwidth utilization of adjacent nodes). The outputs of the two channels are dynamically fused through an attention mechanism: the temporal feature vector and the spatial feature vector are used to calculate the dot product attention score, which is normalized to generate a weight matrix, and then weighted and summed to obtain a joint feature matrix. The joint feature matrix is ​​processed through a fully connected layer to output a bandwidth demand matrix (the probability distribution of each node for different bandwidth levels) and a risk warning level (probability value of 0-1). For example, the bandwidth demand matrix of a core link predicts that it has an 85% probability of needing to allocate 100-120Mbps within 5 minutes, while detecting a burst traffic risk level of 0.8 (threshold 0.7 triggers an alarm). If the load of adjacent nodes on the link has reached 90%, the joint prediction model will dynamically increase its risk level and trigger the bandwidth preemption mechanism of S103.

[0033] S103. Establish a bandwidth allocation decision tree based on the link bandwidth demand matrix, and determine the real-time bandwidth threshold based on the bandwidth allocation decision tree and the risk warning level.

[0034] For example, dynamic resource allocation is achieved through a fuzzy logic controller and a decision tree, based on a bandwidth demand matrix and risk level. First, the basic bandwidth pool is divided according to service priority: voice services are allocated 40% of the bandwidth (preemptive queue), while video and data services are each allocated 30%. The fuzzy logic controller maps link load rates to bandwidth adjustment coefficients (e.g., a coefficient of 1.2 for 80% load rate), and calculates real-time thresholds based on risk levels (e.g., a static threshold of 100Mbps × coefficient 1.2 × risk attenuation factor 0.76). For high-risk links (risk level > 0.7), non-critical service bandwidth compression is initiated (e.g., video services release bandwidth exponentially), and the released resources are prioritized for allocation to the voice queue. Every 30 seconds, the decision tree adjusts the adjustment coefficient mapping rules based on the remaining bandwidth quota. For example, when the remaining quota for video services is below 20%, its bandwidth compression is limited to avoid service interruption.

[0035] S104. Construct reinforcement learning state vectors based on current communication quality indicators, link bandwidth demand matrix, and real-time bandwidth thresholds.

[0036] For example, the state vector needs to integrate real-time network metrics (latency, packet loss rate, jitter), predicted bandwidth requirements, and dynamic thresholds. Specific implementation includes: standardizing communication quality metrics (e.g., normalizing 50ms latency to 0.8), performing principal component analysis to reduce the dimensionality of the bandwidth requirement matrix (compressing it from 50 dimensions to 5 core features), and concatenating the normalized metrics (3 dimensions), the reduced requirement features (5 dimensions), and the real-time threshold (1 dimension) into a 9-dimensional state vector. For example, a state vector might be 0.8, -1.2, 1.5, 0.4, -0.3, 2.1, 0.7, -1.8, 83.6, where 83.6 is the real-time bandwidth threshold. This design ensures that the deep Q-network can simultaneously perceive real-time network status, predicted requirements, and resource constraints.

[0037] S105. The reinforcement learning state vector is analyzed through a deep Q-network to obtain the path decision instructions.

[0038] For example, Deep Q Network (DQN) employs a dual-network structure (online network + target network) to parse state vectors and generate path instructions. The online network calculates the Q-value of each candidate path through a fully connected layer (512 → 256 nodes). An ε-greedy strategy is used to balance exploration and utilization: a path is randomly selected with a 10% probability (e.g., trying a new low-load path C), and the path with the highest Q-value is selected with a 90% probability. The target network synchronizes online network parameters every 1000 steps to calculate expected rewards (e.g., current reward + future discount reward), and optimizes online network parameters using mean squared error loss. When the Q-value converges (standard deviation < 0.01 for 10 consecutive iterations), the path corresponding to the highest Q-value and the bandwidth allocation ratio are output to the gateway routing module. For example, if a state vector indicates core link congestion, DQN outputs that path D has the highest Q-value (allocating 70% bandwidth to voice services). Based on this, the routing module dynamically adjusts the OSPF protocol weights to redirect traffic to path D, ensuring service continuity.

[0039] This application provides a deep learning-based intelligent scheduling method for gateway data transmission. The method includes: acquiring three-dimensional feature tensors corresponding to signaling streams, media streams, and user behavior data; performing spatiotemporal feature joint prediction on the three-dimensional feature tensors based on a pre-set dual-channel LSTM neural network, outputting a link bandwidth demand matrix and a risk warning level for burst traffic; establishing a bandwidth allocation decision tree based on the link bandwidth demand matrix, and determining a real-time bandwidth threshold based on the bandwidth allocation decision tree and the risk warning level; constructing a reinforcement learning state vector based on current communication quality indicators, the link bandwidth demand matrix, and the real-time bandwidth threshold; and obtaining path decision instructions by parsing the reinforcement learning state vector through a deep Q-network. In this method, by integrating signaling streams, media streams, and user behavior data to obtain three-dimensional feature tensors, the dual-channel LSTM neural network employs a spatiotemporal feature joint prediction mechanism to reduce bandwidth demand prediction errors. Then, dynamic weighted fusion of spatiotemporal features is achieved through an attention weight matrix, enhancing the sensitivity of burst traffic warnings. Combining the reinforcement learning state vector with the deep Q-network, a closed-loop optimization system is constructed. Under dynamic bandwidth threshold control, this improves the generation speed of path decision instructions while reducing network latency fluctuations.

[0040] To more clearly illustrate the technical solution of this application, the technical solution of this application will be described below through specific embodiments. It should be noted that the specific embodiments are used to expand the description of the technical solution of this application, and are not intended to limit this application.

[0041] In some embodiments, the dual-channel LSTM neural network includes: a hidden layer and a fully connected layer. After constructing a reinforcement learning state vector based on the current communication quality index, the link bandwidth demand matrix and the real-time bandwidth threshold, and obtaining the path decision instruction by parsing the reinforcement learning state vector through a deep Q network, the method further includes: S106-S109.

[0042] S106. Distill and compress the parameters of the hidden layer to obtain the compressed hidden layer parameters. Sparsify the parameters of the last two fully connected layers to obtain the sparse fully connected layer parameters. Based on the compressed hidden layer parameters, the sparse fully connected layer parameters, and the dual-channel LSTM neural network, generate a lightweight LSTM sub-model.

[0043] For example, hidden layer parameter distillation compression is achieved through knowledge distillation, simplifying the multi-layered cascaded structure of the hidden layers in the original dual-channel LSTM neural network into a single-layer bidirectional structure while retaining the ability to extract key temporal features. A temperature scaling strategy is employed to soften the probability distribution of the hidden layer outputs. For instance, the 256-dimensional hidden states of the LSTM unit are smoothed using a temperature parameter T=2, and then the KL divergence loss is calculated with the 128-dimensional output of the lightweight sub-model, forcing the sub-model to mimic the temporal feature response pattern of the original model. Sparsity reduction of fully connected layer parameters uses an iterative L1 regularization method, applying sparse constraints to the weight matrices of the last two fully connected layers, gradually resetting 80% of redundant connection weights to zero, while retaining 20% ​​of the highly sensitive parameters.

[0044] S107. Federate and aggregate the lightweight LSTM sub-models of multiple gateway routing nodes to generate a global LSTM model. The global LSTM model is used to monitor the network links of the gateway routing nodes.

[0045] For example, lightweight LSTM sub-models from multiple gateway routing nodes are aggregated using a federated averaging algorithm. Each node trains its sub-model locally and only uploads the model parameters to the central server, avoiding privacy risks associated with raw data transmission. For instance, five nodes train their local sub-models, and the central server weights the hidden layer parameters according to the amount of data from each node (e.g., node 1 has 30% data, so its parameter weight is 0.3). The sparse parameters of the fully connected layers are filtered using a majority voting mechanism to select non-zero weight indices. The aggregated global LSTM model integrates the spatiotemporal feature distribution of the local data from each node to monitor the network links of the gateway routing nodes, such as the high-density traffic patterns of core urban nodes and the low-latency requirements of edge nodes. The federated aggregation process introduces a differential privacy mechanism, adding Gaussian noise (standard deviation σ = 0.01) to the uploaded parameters to ensure model security.

[0046] S108. Using a global LSTM model, network links exceeding a preset threshold are marked as faulty links. Ant colony optimization is performed on the topological adjacency matrix of the faulty links to generate transmission quality assessment results for alternative paths.

[0047] For example, when the global LSTM model detects that the packet loss rate of a link exceeds a threshold (e.g., 5% for 10 seconds), the ant colony algorithm is activated to explore paths in the topological adjacency matrix of the faulty link. The node connection weights in the adjacency matrix are dynamically updated based on historical transmission quality; for example, the weight of a high-latency link is reduced by 30%. When generating alternative paths using the ant colony algorithm, the transmission quality evaluation of each path includes three metrics: bandwidth utilization, hop count, and node load. For example, the alternative path B→C→D starting from the faulty node A has a bandwidth utilization of 85%, a hop count of 3, and an average node load of 70%, with an evaluation score of 0.85 × 0.7 (load-weighted) - 0.1 × 3 (hop penalty) = 0.515. During the optimization process, the subtrees of high-scoring paths are expanded, and after 1000 iterations, the top three alternative paths and their evaluation results are output. The evaluation results are linked to the current network state.

[0048] S109. Based on the transmission quality assessment results of the alternative paths, the path decision instructions are dynamically corrected to obtain the corrected state-action pair data, and the corrected state-action pair data is injected into the experience replay pool of the deep Q network.

[0049] For example, based on the transmission quality assessment results of alternative paths, the original path decision instructions generated by the deep Q network are prioritized. For instance, the original instruction selects path X (Q value 120), but the assessment shows that the transmission quality score of its alternative path Y (0.72) is higher than X's (0.65). Therefore, the priority of path Y is increased, and an additional 10% bandwidth quota is allocated. The corrected state-action pair data includes the original state vector, the corrected action, and the actual reward value (e.g., a reward of +0.5 corresponding to a 2% decrease in packet loss rate). After filtering outliers through the data verification module, it is injected into the deep Q network's experience replay pool. The replay pool uses a hierarchical sampling strategy, prioritizing the retention of corrected data with high reward values ​​to accelerate strategy convergence.

[0050] In some embodiments, the pre-configured dual-channel LSTM neural network includes: a stacked LSTM network of the first channel, a graph convolutional network of the second channel, a hidden layer and a fully connected layer. Based on the pre-configured dual-channel LSTM neural network, spatiotemporal feature joint prediction is performed on the three-dimensional feature tensor, and the link bandwidth demand matrix and the risk warning level of burst traffic are output. The method includes: S1021-S1025.

[0051] S1021. Perform spatiotemporal dimension segmentation on the three-dimensional feature tensor to generate temporal feature subsets and spatial feature subsets.

[0052] For example, the spatiotemporal segmentation of the 3D feature tensor is achieved based on time window slicing and node topology grouping. The temporal feature subset is generated by extracting the feature vectors of all nodes within each time slice. For instance, a 60×10×30 tensor is split into 60 groups of 10×30 matrices across 60 time slices, each group representing the instantaneous state of all gateway routing nodes at a specific moment. The spatial feature subset is partitioned based on node topology relationships. The 30-dimensional features of each node within the same time window are reorganized into graph-structured data according to physical connectivity (such as an OSPF adjacency list). For example, when 10 nodes form a star topology, the feature vector of the central node and the features of its nine connected edge nodes are combined into a 10×30×9 adjacency feature block. The segmentation process must preserve the mapping relationship between timestamps and node IDs to ensure the correspondence of spatiotemporal features in subsequent processing.

[0053] S1022. The time feature subset is propagated bidirectionally through the stacked LSTM network of the first channel to output the time feature vector.

[0054] For example, the stacked LSTM network consists of two layers of bidirectional LSTM units, with temporal feature subsets input to the network independently according to node dimensions. The first layer of bidirectional LSTM captures the causal and inverse dependencies of the time series through forward and backward propagation, respectively. For example, the forward LSTM processes time-slice data from t=1 to t=60 to identify the increasing trend of the broadcaster's bitrate; the backward LSTM performs reverse analysis from t=60 to t=1 to detect the lagging correlation of viewer request volume. The second layer of bidirectional LSTM performs high-order feature abstraction on the hidden states output by the first layer, such as extracting minute-level periodic patterns from second-level fluctuations (e.g., a peak every 5 minutes). The output of each time slice is a 256-dimensional vector, which generates a temporal feature vector by concatenating the hidden states from the forward and backward propagation.

[0055] S1023. The spatial feature subset is aggregated through a graph convolutional network with the second channel to output a spatial feature vector.

[0056] For example, a graph convolutional network constructs an adjacency matrix based on node topology, and a subset of spatial features is sliced ​​into the GCN according to a time window. The weights of the adjacency matrix are jointly determined by the physical link distance (e.g., fiber optic length) and the logical coupling strength (e.g., historical traffic correlation). For example, when the fiber optic distance between two nodes is 10km and the traffic correlation is 0.8, the adjacency weight is set to 0.8×e. {-0.1×10}=0.36. The first convolutional layer of GCN aggregates the features of first-order neighbor nodes, such as the weighted sum of the features of the center node and the features of its directly connected edge nodes, generating a 128-dimensional intermediate feature. The second convolutional layer expands to second-order neighbors, capturing the indirect influence of hop nodes (such as the coupling effect caused by edge nodes passing through the center node). After each convolutional layer, a nonlinearity is introduced through the ReLU activation function, outputting a spatial feature vector. The neighborhood aggregation capability provides spatial correlation for risk warning, for example, when a synchronous surge in the load of adjacent nodes is detected, a regional congestion warning is triggered.

[0057] S1024. In the hidden layer, a weighted Hadamard product operation is performed on the temporal feature vector and the spatial feature vector based on the preset attention weight matrix to generate a joint feature matrix.

[0058] For example, the temporal feature vector (256 dimensions) and spatial feature vector (128 dimensions) are multiplied by the node dimension to generate an attention score matrix. For instance, the score matrix for 10 nodes has a dimension of 10×10, reflecting the correlation between the spatiotemporal features of each node. After being normalized by Softmax, the score matrix is ​​multiplied element-wise with the spatial feature vector, and then concatenated with the temporal feature vector to generate a joint feature matrix. For example, when the temporal feature vector of node A is highly correlated with the spatial feature vector of node B (score 0.9), the joint feature of node A will strengthen the topological influence of node B. The dynamic weighting mechanism in this step can adaptively adjust the contribution of spatiotemporal features, for example, increasing the weight of temporal features in bursty traffic scenarios, and enhancing the salience of spatial features in topological congestion.

[0059] S1025. In the fully connected layer, perform fully connected mapping and Sigmoid probability distribution calculation on the joint feature matrix, and output the link bandwidth demand matrix and risk warning level.

[0060] For example, the joint feature matrix undergoes non-linear mapping through two fully connected layers. The first layer compresses the 256 + 128 = 384-dimensional input to 64 dimensions, and the second layer further maps it to the output dimension. The link bandwidth demand matrix is ​​generated using multi-class Softmax, converting the 64-dimensional vector into a probability distribution for each bandwidth level. For instance, the 5-level bandwidth demand matrix for 10 nodes has a dimension of 10 × 5, with each node corresponding to a probability value of [20Mbps, 50Mbps, 100Mbps, 200Mbps, 500Mbps]. The risk warning level is calculated using the Sigmoid function, mapping the 64-dimensional vector to scalar values ​​of 0-1.

[0061] In some embodiments, in the hidden layer, a weighted Hadamard product operation is performed on the temporal feature vector and the spatial feature vector based on a preset attention weight matrix to generate a joint feature matrix, including: S241-S244.

[0062] S241. Perform a dot product operation on the temporal feature vector and the spatial feature vector to generate the attention score matrix.

[0063] For example, the temporal feature vector and spatial feature vector are multiplied by a dot product to generate an attention score matrix, which quantifies the correlation strength between nodes at different time steps and topological nodes. The temporal feature vector contains the temporal pattern of a node within a historical window (such as periodic traffic fluctuations), while the spatial feature vector encodes the neighborhood dependency of a node in the topology (such as load transfer relationships). The dot product of the two is used to calculate similarity through the vector dot product. For example, the dot product score of the temporal feature vector of node A at t=60 and the spatial feature vector of node B is 0.92, reflecting a strong spatiotemporal coupling between the two during a traffic surge. The dimension of the attention score matrix is ​​the number of nodes × the number of nodes, and each element represents the matching degree of the spatiotemporal features between the source node and the target node, providing a quantitative basis for subsequent weighting.

[0064] S242. Normalize the attention score matrix to generate the attention weight matrix.

[0065] For example, the attention score matrix is ​​normalized using Softmax, generating a probability distribution attention weight matrix along the row direction to ensure that the weight sum of each target node is 1. The normalization process maps the original scores to the interval of 0 and 1. For instance, the attention score of node C for node D is compressed from 3.2 to 0.78, while the score for node E is reduced from 1.1 to 0.22, highlighting the topological influence of node D. The generation of the attention weight matrix needs to consider the dynamic balance between temporal and spatial features. For example, during bursts of traffic, the weight of temporal features increases to 70%, while in a stable state, the weight of spatial features dominates. This eliminates the difference in the dimensions of scores between different nodes, making the subsequent weighted sum comparable and preventing a few high-scoring nodes from monopolizing attention resources.

[0066] S243. The time feature vectors are weighted and summed according to the attention weight matrix to generate a weighted time feature vector.

[0067] For example, a weighted time feature vector is generated by weighting and summing the time feature vectors based on the attention weight matrix. The weighted summation operation replaces the time feature vector of each target node with a linear combination of the time features of all its associated nodes. Specifically, the weighted time feature of node F = 0.6 × original time feature of node F + 0.3 × time feature of node G + 0.1 × time feature of node H, reflecting the temporal influence of upstream nodes G and H on node F. The weighted time feature vector preserves global temporal dependencies; for example, historical traffic peaks of the main node are propagated to edge nodes after weighting, providing early warning of potential congestion risks. This enhances the feature representation of key nodes, such as spreading the burst traffic patterns of the core gateway to associated links, improving the model's ability to perceive cascading failures.

[0068] S244. Perform the Hadamard product operation on the weighted time eigenvector and spatial eigenvector to generate a joint feature matrix.

[0069] For example, the weighted temporal feature vector and spatial feature vector are multiplied element-wise using the Hadamard product to generate a joint feature matrix. The Hadamard product operation forces spatiotemporal features to interact locally along the same dimensions. For instance, multiplying the 128th dimension of the weighted temporal feature vector (encoding second-level traffic changes) with the 128th dimension of the spatial feature vector (encoding node load) amplifies instantaneous traffic anomalies at high-load nodes. Each element of the joint feature matrix represents a fused feature along a specific spatiotemporal dimension. For example, a value greater than 1 indicates synergistic enhancement of spatiotemporal features (e.g., node I experiences a synchronized surge in load with its neighboring nodes' traffic during peak periods), while a value less than 1 indicates feature conflict (e.g., node J experiences a decrease in traffic but its neighboring nodes experience an increase in load). This enables fine-grained fusion of spatiotemporal features, providing the fully connected layer with a joint representation that combines temporal dynamics and topological context, driving the accuracy of bandwidth demand and risk level predictions.

[0070] In some embodiments, a bandwidth allocation decision tree is established based on the link bandwidth demand matrix, and a real-time bandwidth threshold is determined based on the bandwidth allocation decision tree and the risk warning level, including: S1031-S1036.

[0071] S1031. Based on the preset service type priority, divide the basic bandwidth pools corresponding to voice services, video services and data services respectively, and set up a preemptive queue for voice services.

[0072] S1032. Construct a membership function based on a fuzzy logic controller to map the load rate in the link bandwidth demand matrix to a bandwidth adjustment coefficient.

[0073] S1033. Determine the rough bandwidth threshold based on the preset static threshold, bandwidth adjustment coefficient, and risk warning level.

[0074] S1034. Compress the non-critical service bandwidth quotas of high-risk links in the basic bandwidth pool to obtain the updated service bandwidth pool. High-risk links are those with a risk warning level higher than the threshold.

[0075] S1035. Calculate the remaining quota of the updated service bandwidth pool, and adjust the membership function according to the remaining quota to obtain the adjusted membership function.

[0076] S1036. Recalculate the real-time bandwidth threshold corresponding to the coarse bandwidth threshold based on the corrected membership function.

[0077] For example, the basic bandwidth pool is divided based on the priority of service type. Voice services exclusively occupy 40% of the bandwidth and a preemptive queue is enabled to ensure zero-interruption guarantee for high-priority calls. Video and data services are each allocated 30% of the bandwidth, forming the initial resource allocation framework. For instance, in a video conferencing scenario, the voice service pool reserves 100Mbps, the video pool 75Mbps, and the data pool 75Mbps. The preemptive queue allows the voice service to temporarily occupy idle bandwidth in other pools during bursts of traffic. The membership function constructed by the fuzzy logic controller converts the load rate in the link bandwidth demand matrix into a dynamic adjustment coefficient. The load rate is mapped to the range [0.8, 1.2] through a Gaussian membership function. For example, a link load rate of 75% corresponds to an adjustment coefficient of 1.15, and a load rate of 90% triggers a non-linear response, increasing the coefficient to 1.25. A rough bandwidth threshold is calculated by combining a static threshold (e.g., 100Mbps) with an adjustment coefficient and a risk warning level. The risk level is applied to the threshold through an exponential decay factor. When the risk level of a link is 0.8, the rough threshold = 100 × 1.15 × (1 - 0.8 × 0.3) = 100 × 1.15 × 0.76 = 87.4Mbps, reflecting the squeeze effect of risk on resources. For high-risk links (risk level > 0.7), an exponential compression strategy is initiated for non-critical business bandwidth quotas. Video service bandwidth is gradually released according to a time decay model. For example, the initial quota of 30Mbps is compressed to 18Mbps within 10 seconds, and the released 12Mbps is transferred to the voice preemption pool. After compression, the service bandwidth pool is updated, and the remaining quota is calculated in real time. For example, when the video pool has 18Mbps remaining, a quota correction mechanism is triggered. The membership function parameter of the fuzzy logic controller is dynamically adjusted according to the remaining amount. When the remaining amount of the video pool is less than 20%, the mapping range of load rate - adjustment coefficient is tightened from [0.8, 1.2] to [0.9, 1.1], limiting the bandwidth compression range to avoid service interruption. The corrected membership function recalculates the real-time bandwidth threshold. For example, if the load rate of a link drops to 70%, the corrected adjustment coefficient is 1.05. Combined with a risk level of 0.6, the real-time threshold = 100 × 1.05 × (1 - 0.6 × 0.3) = 100 × 1.05 × 0.82 = 86.1Mbps, achieving dynamic balancing of resource allocation.

[0078] A basic bandwidth pool is partitioned to construct a resource allocation framework. A fuzzy logic controller quantifies the load rate into an operable adjustment coefficient. A coarse threshold integrates static configuration, dynamic load, and risk factors. High-risk links compress and release redundant resources and update the bandwidth pool status. Remaining quota statistics drive the adjustment of membership function parameters, and the threshold is recalibrated through the corrected function. For example, in a sudden congestion scenario, if the load rate of a core link rises to 85% and the risk level is 0.9, the initial coarse threshold is compressed to 80Mbps, and the video service quota is compressed from 30Mbps to 15Mbps. The remaining quota triggers a membership function correction, limiting the upper limit of subsequent adjustment coefficients. After recalculation, the threshold is adjusted to 82Mbps, ensuring voice services while avoiding excessive compression that could cause video stuttering. The entire process forms a closed-loop control of "initial allocation → dynamic adjustment → compression and release → parameter correction → threshold recalculation," achieving elastic supply of network resources and risk hedging.

[0079] In some embodiments, a reinforcement learning state vector is constructed based on the current communication quality metric, the link bandwidth demand matrix, and the real-time bandwidth threshold, including: S1041-S1043.

[0080] S1041. Perform Z-score standardization on the communication quality indicators to generate a normalized indicator vector. The normalized indicator vector includes: delay vector, packet loss rate vector, and jitter indicator vector.

[0081] For example, Z-score standardization of communication quality indicators maps the original latency, packet loss rate, and jitter values ​​to a distribution range with zero mean and unit variance by calculating the mean and standard deviation of each indicator, resulting in a normalized indicator vector. This normalized vector preserves the relative deviation characteristics of the indicators and eliminates the interference of dimensional differences on model training. When standardizing the latency vector, if the average latency across the entire network is 50ms and the standard deviation is 10ms, a latency of 65ms on a certain link is normalized to 1.5, reflecting its deviation from the mean. The packet loss rate vector uses the network average of 2% and the standard deviation of 0.5% as a benchmark; a 3% packet loss rate corresponds to a Z-score of 2.0. The jitter indicator vector dynamically adjusts the benchmark based on historical data. For example, in a video conferencing scenario, the mean jitter is 5ms and the standard deviation is 2ms; a measured jitter of 9ms is normalized to 2.0. The normalized metric vector preserves the relative deviation characteristics of the metrics. For example, links with a latency Z-score of 1.5 should be optimized first, while links with a packet loss rate Z-score of 0.3 can remain as they are.

[0082] S1042. Perform principal component analysis to reduce the dimensionality of the link bandwidth demand matrix and extract the bandwidth demand feature vector.

[0083] For example, principal component analysis (PCA) dimensionality reduction of the link bandwidth demand matrix extracts key features through covariance matrix decomposition, compressing a high-dimensional matrix (such as 50-dimensional data of 10 nodes × 5 levels of bandwidth demand) into a low-dimensional orthogonal vector. The eigenvalue ranking of the covariance matrix determines the contribution rate of the principal components. For example, if the cumulative contribution rate of the first 5 eigenvalues ​​reaches 90%, it indicates that they can represent most of the variance information of the matrix. The bandwidth demand feature vector is generated by projecting it onto the principal component direction. For example, the probability distribution of a node's demand [20Mbps, 50Mbps, 100Mbps, 200Mbps, 500Mbps] is projected and transformed into a 5-dimensional vector [-0.3, 1.2, 0.8, -0.5, 0.1], where the first principal component 1.2 reflects the node's strong demand tendency for medium-to-high bandwidth (50-100Mbps).

[0084] S1043. The normalized index vector, bandwidth requirement feature vector, and real-time bandwidth threshold are concatenated into tensors to generate a reinforcement learning state vector.

[0085] For example, the tensor concatenation of the normalized index vector, bandwidth demand feature vector, and real-time bandwidth threshold employs a dimensionality expansion strategy to integrate three types of heterogeneous data into a unified state representation. The normalized index vector contains Z-score values ​​for latency, packet loss rate, and jitter (e.g., [1.5, 2.0, 2.0]), the bandwidth demand feature vector carries the core demand pattern after dimensionality reduction (e.g., a 5-dimensional vector [-0.3, 1.2, 0.8, -0.5, 0.1]), and the real-time bandwidth threshold, as a scalar parameter (e.g., 85Mbps), is expanded into a single-dimensional vector. The resulting reinforcement learning state vector has a dimension of 3+5+1=9, for example, [1.5, 2.0, 2.0, -0.3, 1.2, 0.8, -0.5, 0.1, 85], fully characterizing the current network's communication quality, bandwidth requirements, and resource constraints. The design of the state vector is compatible with both real-time performance and completeness. For example, when the real-time threshold drops from 85Mbps to 80Mbps, the change in the last value of the vector directly reflects the resource scarcity level, driving the deep Q network to prioritize low-bandwidth-consumption paths.

[0086] In some embodiments, a path decision instruction is obtained by parsing the reinforcement learning state vector through a deep Q-network, including: S1051-S1055.

[0087] S1051. The reinforcement learning state vector is mapped through a deep Q-network to generate the first Q-value matrix.

[0088] S1052. Determine the path index corresponding to the first Q-value matrix through the ε-greedy strategy, and generate the initial path decision instruction.

[0089] S1053. Calculate the expected reward value of the initial path decision instruction based on the deep Q-network, and update the first Q-value matrix according to the expected reward value to obtain the second Q-value matrix.

[0090] S1054. Calculate the mean squared error loss between the second Q-value matrix and the first Q-value matrix. Based on the mean squared error loss, update the first Q-value matrix through gradient backpropagation to obtain the third Q-value matrix.

[0091] S1055. Generate path decision instructions based on the third Q-value matrix and send the path decision instructions to the corresponding gateway routing node.

[0092] After receiving the reinforcement learning state vector, the deep Q-network maps it to the Q-value space of candidate paths through a multi-layer fully connected neural network, generating the first Q-value matrix. Each dimension of the state vector corresponds to a node in the network input layer. For example, after a 9-dimensional state vector undergoes a nonlinear transformation through a hidden layer (256 nodes), the output layer generates Q-values ​​equal to the number of candidate paths, forming the first Q-value matrix with a dimension of path number × 1. For example, when there are 5 candidate paths, the Q-value matrix might be [120, 95, 110, 80, 105], reflecting the expected long-term gains of each path under metrics such as latency and bandwidth utilization. The ε-greedy policy selects an action based on a preset exploration probability (e.g., ε = 0.1) in the first Q-value matrix, selecting the highest Q-value path with a 90% probability (e.g., path A with Q-value 120 corresponding to index 0) and randomly selecting other paths with a 10% probability (e.g., path D with Q-value 80 corresponding to index 3), generating an initial path decision instruction. The exploration mechanism avoids the model getting trapped in local optima; for example, randomly selecting a low Q-value path may discover hidden high-yield routes. The initial instruction carries the path index and Q-value confidence, for example, the index of path A is 0 and the confidence is 0.92, which guides the gateway routing node to establish preliminary forwarding rules.

[0093] The expected reward value of the initial path decision instruction is calculated through the target Q-network. The target network replicates the parameters of the main network with a delayed update to ensure training stability. The expected reward value consists of the immediate reward (e.g., a reward of +0.5 corresponding to a 15ms reduction in latency for path A) and the discounted maximum Q-value of the next state. For example, the current Q-value of 120 = immediate reward 0.5 + 0.9 × maximum Q-value of the next state 132. The first Q-value matrix is ​​updated temporally based on the expected reward value. For example, the Q-value of path A is adjusted from 120 to 120 + α × (0.5 + 0.9 × 132 - 120) = 120 + 0.1 × 8.8 = 120.88, generating the second Q-value matrix. This update process quantifies the long-term benefits of path selection. For example, although a high-bandwidth path consumes resources in the short term (reward -0.2), it gains a long-term positive benefit (discounted reward +1.5) by avoiding congestion. The mean squared error loss is used to calculate the difference between the second Q-value matrix and the first Q-value matrix. For example, the loss values ​​for the five paths are (0.88...). 2 The mean of the three Q-values ​​(0, 0, 0, 0) is 0.155. The loss function updates the main network parameters through gradient backpropagation, and the learning rate of 0.001 controls the update magnitude, making the third Q-value matrix approximate the target value. For example, the Q-value of path A converges to 123.5 after three iterations, while the Q-value of path D increases from 80 to 85 due to the increase in the number of explorations.

[0094] The third Q-value matrix, after softmax normalization, generates path decision instructions. For example, path A is designated as the primary path with a 62% probability, path B (Q-value 115) as a backup with a 30% probability, and the remaining paths share 8% of traffic for link health monitoring. These decision instructions are injected into the flow table of the gateway routing node, triggering the SDN controller to adjust OpenFlow rules, such as assigning a priority label to path A and matching rate limiting policies to real-time bandwidth thresholds. This closed-loop process enables dynamic optimization of network resources. For instance, in bursty traffic scenarios, the third Q-value matrix drives the gateway routing node to switch to a low-latency path within 10ms, ensuring SLA compliance.

[0095] Please see Figure 2 , Figure 2 This is a schematic block diagram of a deep learning-based intelligent scheduling system for gateway data transmission, provided in an embodiment of this application. The deep learning-based intelligent scheduling system 200 is used to execute the aforementioned deep learning-based intelligent scheduling method for gateway data transmission. The deep learning-based intelligent scheduling system 200 can be configured in a server.

[0096] The server can be a standalone server, a server cluster, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0097] like Figure 2 As shown, the gateway data transmission intelligent scheduling system 200 based on deep learning includes: a data acquisition module 201, a data analysis module 202, a risk prediction module 203, a vector construction module 204, and an instruction output module 205.

[0098] The data acquisition module 201 is used to acquire the three-dimensional feature tensors corresponding to the signaling stream, media stream, and user behavior data.

[0099] The data analysis module 202 is used to perform spatiotemporal feature joint prediction of three-dimensional feature tensors based on a pre-built dual-channel LSTM neural network, and output the link bandwidth demand matrix and the risk warning level of burst traffic.

[0100] The risk prediction module 203 is used to establish a bandwidth allocation decision tree based on the link bandwidth demand matrix, and to determine the real-time bandwidth threshold based on the bandwidth allocation decision tree and the risk warning level.

[0101] Vector construction module 204 is used to construct reinforcement learning state vectors based on the current communication quality indicators, link bandwidth demand matrix and real-time bandwidth threshold.

[0102] The instruction output module 205 is used to parse the reinforcement learning state vector through a deep Q-network to obtain path decision instructions.

[0103] This application provides an electronic device, which includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and, when executing the computer program, implement the deep learning-based gateway data transmission intelligent scheduling method as described in any of the embodiments of this application.

[0104] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it enables the processor to implement a deep learning-based gateway data transmission intelligent scheduling method as described in any of the embodiments of this application.

[0105] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A gateway data transmission intelligent scheduling method based on deep learning, characterized in that, The method includes: Obtain the three-dimensional feature tensors corresponding to signaling streams, media streams, and user behavior data; Based on a pre-built dual-channel LSTM neural network, the spatiotemporal features of the three-dimensional feature tensor are jointly predicted, and the link bandwidth demand matrix and the risk warning level of burst traffic are output. A bandwidth allocation decision tree is established based on the link bandwidth demand matrix, and a real-time bandwidth threshold is determined based on the bandwidth allocation decision tree and the risk warning level. Construct a reinforcement learning state vector based on the current communication quality indicators, the link bandwidth requirement matrix, and the real-time bandwidth threshold; The reinforcement learning state vector is parsed using a deep Q-network to obtain path decision instructions, which are then sent to the corresponding gateway routing node.

2. The intelligent scheduling method for gateway data transmission based on deep learning as described in claim 1, characterized in that, The dual-channel LSTM neural network includes a hidden layer and a fully connected layer. After constructing a reinforcement learning state vector based on the current communication quality index, the link bandwidth demand matrix, and the real-time bandwidth threshold, and parsing the reinforcement learning state vector through a deep Q-network to obtain the path decision instruction, the method further includes: The parameters of the hidden layer are distilled and compressed to obtain compressed hidden layer parameters. The parameters of the last two fully connected layers are sparsified to obtain sparse fully connected layer parameters. A lightweight LSTM sub-model is generated based on the compressed hidden layer parameters, the sparse fully connected layer parameters, and the dual-channel LSTM neural network. The lightweight LSTM sub-models of multiple gateway routing nodes are federated and aggregated to generate a global LSTM model, which is used to monitor the network links of the gateway routing nodes. The global LSTM model marks network links that exceed a preset threshold as faulty links, performs ant colony optimization on the topological adjacency matrix of the faulty links, and generates transmission quality evaluation results for alternative paths. Based on the transmission quality assessment results of the alternative paths, the path decision instructions are dynamically corrected to obtain corrected state-action pair data, which is then injected into the experience replay pool of the deep Q network.

3. The intelligent scheduling method for gateway data transmission based on deep learning as described in claim 1, characterized in that, The pre-configured dual-channel LSTM neural network includes: a stacked LSTM network in the first channel, a graph convolutional network in the second channel, hidden layers, and fully connected layers. Based on the pre-configured dual-channel LSTM neural network, the method performs joint spatiotemporal feature prediction on the three-dimensional feature tensor, outputting a link bandwidth demand matrix and a risk warning level for burst traffic. The three-dimensional feature tensor is segmented in terms of its spatiotemporal dimension to generate a temporal feature subset and a spatial feature subset; The time feature subset is propagated bidirectionally through a stacked LSTM network of the first channel to output a time feature vector; The spatial feature subset is aggregated by a graph convolutional network in the second channel to output a spatial feature vector. In the hidden layer, a weighted Hadamard product operation is performed on the temporal feature vector and the spatial feature vector based on a preset attention weight matrix to generate a joint feature matrix; In the fully connected layer, the joint feature matrix is ​​fully connected and Sigmoid probability distribution is calculated to output the link bandwidth demand matrix and risk warning level.

4. The intelligent scheduling method for gateway data transmission based on deep learning as described in claim 3, characterized in that, In the hidden layer, a weighted Hadamard product operation is performed on the temporal feature vector and the spatial feature vector based on a preset attention weight matrix to generate a joint feature matrix, including: Perform a dot product operation on the temporal feature vector and the spatial feature vector to generate an attention score matrix; The attention score matrix is ​​normalized to generate an attention weight matrix; The time feature vector is weighted and summed according to the attention weight matrix to generate a weighted time feature vector. Perform a Hadamard product operation on the weighted temporal feature vector and the spatial feature vector to generate a joint feature matrix.

5. The intelligent scheduling method for gateway data transmission based on deep learning as described in claim 1, characterized in that, The step of establishing a bandwidth allocation decision tree based on the link bandwidth demand matrix, and determining a real-time bandwidth threshold based on the bandwidth allocation decision tree and the risk warning level, includes: Based on the preset service type priority, the basic bandwidth pools corresponding to voice services, video services and data services are divided respectively, and a preemptive queue is set up for the voice services. Based on a fuzzy logic controller, a membership function is constructed to map the load rate in the link bandwidth demand matrix to a bandwidth adjustment coefficient. The rough bandwidth threshold is determined based on the preset static threshold, the bandwidth adjustment coefficient, and the risk warning level. Compress the non-critical service bandwidth quota of high-risk links in the basic bandwidth pool to obtain an updated service bandwidth pool. The high-risk links are those whose risk warning level is higher than the threshold. The remaining quota of the updated service bandwidth pool is calculated, and the membership function is corrected based on the remaining quota to obtain the corrected membership function. The real-time bandwidth threshold corresponding to the coarse bandwidth threshold is recalculated based on the modified membership function.

6. The intelligent scheduling method for gateway data transmission based on deep learning as described in claim 1, characterized in that, The step of constructing a reinforcement learning state vector based on the current communication quality metric, the link bandwidth requirement matrix, and the real-time bandwidth threshold includes: The communication quality indicators are Z-score standardized to generate a normalized indicator vector, which includes a delay vector, a packet loss rate vector, and a jitter indicator vector. Principal component analysis is performed on the link bandwidth demand matrix to reduce its dimensionality and extract the bandwidth demand feature vector. The normalized index vector, the bandwidth requirement feature vector, and the real-time bandwidth threshold are concatenated into tensors to generate a reinforcement learning state vector.

7. The intelligent scheduling method for gateway data transmission based on deep learning as described in claim 1, characterized in that, The step of parsing the reinforcement learning state vector through a deep Q-network to obtain path decision instructions and sending the path decision instructions to the corresponding gateway routing node includes: The first Q-value matrix is ​​generated by mapping the reinforcement learning state vector through the deep Q-network; The path index corresponding to the first Q-value matrix is ​​determined by the ε-greedy strategy, and an initial path decision instruction is generated. The expected reward value of the initial path decision instruction is calculated based on the deep Q-network, and the first Q-value matrix is ​​updated according to the expected reward value to obtain the second Q-value matrix; Calculate the mean squared error loss between the second Q-value matrix and the first Q-value matrix. Based on the mean squared error loss, update the first Q-value matrix through gradient backpropagation to obtain the third Q-value matrix. The path decision instruction is generated based on the third Q-value matrix and then sent to the corresponding gateway routing node.

8. A gateway data transmission intelligent scheduling system based on deep learning, characterized in that, The deep learning-based intelligent scheduling system for gateway data transmission is used to execute the deep learning-based intelligent scheduling method for gateway data transmission as described in any one of claims 1-7, wherein the deep learning-based intelligent scheduling system for gateway data transmission includes: The data acquisition module is used to acquire the three-dimensional feature tensors corresponding to signaling streams, media streams, and user behavior data; The data analysis module is used to perform spatiotemporal feature joint prediction on the three-dimensional feature tensor based on a pre-set dual-channel LSTM neural network, and output the link bandwidth demand matrix and the risk warning level of burst traffic. The risk prediction module is used to establish a bandwidth allocation decision tree based on the link bandwidth demand matrix, and to determine a real-time bandwidth threshold based on the bandwidth allocation decision tree and the risk warning level. The vector construction module is used to construct reinforcement learning state vectors based on the current communication quality indicators, the link bandwidth requirement matrix, and the real-time bandwidth threshold. The instruction output module is used to parse the reinforcement learning state vector through a deep Q-network to obtain path decision instructions.

Images