A method for intelligent optimization and deployment of communication resources

Abstract

The application discloses a kind of communication resource intelligent optimization deployment methods, belong to wireless communication technical field, after this method gathers data, after Hampel filtering, Z-score normalization and Min-Max normalization, generate synchronous data frame;Respectively extract the characteristic vector of three kinds of modal, obtain joint state representation vector by cross-modal attention fusion and equidistant mapping manifold embedding;Through the information entropy value calculated by differentiable entropy estimator and dynamically adjust decision network layer number and neuron quantity according to this, execute proximal policy optimization algorithm on the network after adjustment and output resource allocation action;Predict future state by variational autoencoder time series prediction network, lock resource in advance to realize forward-looking pre-scheduling for high-confidence prediction result;Difference residual between actual service quality and prediction value is used as feedback signal to carry out online fine-tuning to perception model and correct reward function, form closed loop continuous optimization, significantly improve throughput, delay, resource utilization and service continuity.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication network resource management technology, and in particular relates to a method for intelligent optimization and allocation of communication resources. Background Technology

[0002] The structure of fifth-generation mobile communication networks exhibits heterogeneity across three levels: diverse access technologies, service types, and channel states. The core management challenge arising from this lies in the imbalance between limited spectrum, power, time slots, and computing resources and highly dynamic, multi-dimensional, and increasingly complex service requests.

[0003] Current mainstream resource management solutions can be broadly categorized as follows:

[0004] Allocation strategies based on fixed rules, such as the proportional fair scheduler, determine the scheduling order in each scheduling slot based on the ratio of the user's instantaneous channel quality to the historical average rate. Its limitation lies in considering only a single physical dimension of information, lacking the ability to predict future channel evolution, and failing to respond to differentiated quality of service for different service types.

[0005] Deep reinforcement learning methods based on a single data source: These methods take the current channel state matrix as the model input, output a resource block allocation vector, and use algorithms such as near-end policy optimization to solve for the optimal policy. Their shortcomings include: the state space only covers physical layer information, neglecting network topology and service semantic features; furthermore, their decision network structure is rigid, resulting in insufficient inference accuracy in highly dynamic environments, while causing unnecessary computational load in static scenarios.

[0006] Weighted fusion-based multi-source data methods: These schemes use fixed weighting coefficients to linearly combine data from different sources before participating in decision-making. Their fundamental flaw lies in the assumption that the data sources are linearly independent of each other, thus failing to capture the complex nonlinear coupling relationship between physical channel quality, network topology congestion, and business semantic requirements. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0008] Specifically, this invention provides the following technical solution: a method for intelligent optimization and allocation of communication resources, comprising the following steps: data acquisition and preprocessing: acquiring reference signal received power, signal-to-interference-plus-noise ratio, channel quality indicator, resource block utilization, and block error rate of the physical layer; acquiring node load rate, backhaul link utilization, handover frequency, number of associated users, and user spatial distribution density factor of the network layer, and constructing an adjacency matrix; acquiring service type, average packet size, arrival interval, and priority level of the service layer; performing Hampel filtering for noise reduction and linear interpolation for the physical layer data, performing Z-score normalization and adjacency matrix symmetric normalization for the network layer data, and performing one-hot encoding, Z-score normalization, and Min-Max normalization for the service layer data, and aligning the timestamps to generate synchronous data frames; multimodal feature extraction: stacking physical layer data from multiple recent time slots and performing one-dimensional convolution and global average pooling to obtain a spectral evolution feature vector; encoding the normalized adjacency matrix and node features through a multi-head graph attention network to obtain a topology association feature vector; and processing user service features through a Transformer... Semantic encoding yields a business intent feature vector; Cross-modal fusion and embedding: The three feature vectors are linearly projected to the same dimensional space, cross-attention is calculated pairwise, and then fused into a trimodal fusion vector. This vector is then embedded into a low-dimensional manifold via equidistant mapping to obtain a joint state representation vector; Adaptive decision-making: The joint state representation vector is input into a decision network containing a differentiable entropy estimator. Based on historical samples from a sliding window, the information entropy of the current sample is calculated using Gaussian kernel density estimation. The number of network layers and neurons is dynamically adjusted accordingly, and proximal policy optimization is performed, outputting resource allocation actions; Pre-decision and reservation: The joint state representation vector sequence is input into a variational autoencoder temporal prediction network to generate future period state prediction values ​​and standard deviations. Confidence is calculated, and virtual decisions are made for predicted states exceeding a threshold, locking reserved resources. When a future period arrives, the Mahalanobis distance is calculated. If the distance is less than the threshold, reservation is activated; otherwise, it is released and a re-decision is triggered; Online feedback learning: The Euclidean distance between the actual service quality index and the predicted service quality index is collected as a differential residual, which is used for gradient updates of feature extraction and fusion network parameters, as well as for correcting the proximal policy optimization reward.

[0009] Preferably, the Hampel filtering denoising is as follows: taking a preset number of time slots before and after the current time as the center to form a window, calculating the median and the corrected absolute median difference; if the absolute difference between the current data and the median exceeds three times the corrected absolute median difference, then it is replaced with the median; the adjacency matrix symmetric normalization is as follows: after adding an identity matrix to the adjacency matrix, multiply it by the left and right sides by the negative 1 / 2 power of the degree matrix.

[0010] Preferably, the single-layer operation of the graph attention network encoding is as follows: the learnable linear transformation matrix of each attention head is applied to the node features; the transformed features of the center node and the neighboring nodes are concatenated and then the inner product is performed with the learnable attention vector and activated by a linear rectified function with leakage; the activation values ​​of the neighboring nodes (including themselves) are normalized by softmax to obtain the attention coefficients and then weighted and summed; finally, the outputs of each attention head are averaged and activated by an exponential linear unit.

[0011] Preferably, the attention pooling is performed as follows: after the output features of each node are subjected to learnable linear transformation and hyperbolic tangent activation, the inner product is made with the learnable context query vector, and the pooling weights are obtained by performing softmax normalization on the inner product results of all nodes, and then the node features are summed in a weighted manner.

[0012] Preferably, the cross-attention matrix is ​​calculated as follows: the source modality projection vector is mapped to a query vector through a learnable query projection matrix, the target modality projection vector is mapped to a key vector through a learnable key projection matrix, the dot product of the query vector and the transpose of the key vector is calculated and divided by a scaling factor, and then softmax normalization is performed on each row of the matrix.

[0013] Preferably, the isometric mapping is as follows: construct a K nearest neighbor graph in the cache pool and calculate the geodesic distance between sample point pairs, and perform multidimensional scaling to obtain low-dimensional embedding coordinates; for new samples, only calculate the geodesic distance between them and a preset number of nearest neighbors in the cache pool, and use landmark multidimensional scaling to solve for low-dimensional embedding coordinates.

[0014] Preferably, the dynamic adjustment of network depth is as follows: the current number of layers equals the base number of layers plus the adjustable amplitude multiplied by the hyperbolic tangent function value after scaling the difference between the information entropy value and the medium entropy value; the dynamic adjustment of network width is as follows: the current number of neurons per layer equals the base number plus the adjustable amplitude multiplied by the value after mapping the difference between the information entropy value and the high entropy threshold using the Logistic Sigmoid function; wherein the medium entropy value is half of the sum of the high entropy threshold and the low entropy threshold.

[0015] Preferably, the confidence score is 1 divided by 1 plus the Euclidean norm of the predicted standard deviation vector; the Mahalanobis distance is the square root of the transpose of the difference between the actual state vector and the predicted state vector, multiplied by the inverse of the predicted covariance matrix, and then multiplied by the difference vector; the predicted covariance matrix is ​​a diagonal matrix with the squares of each component of the standard deviation vector as its diagonal elements; each time the reserved resources are released, the release event is recorded as a negative sample for online fine-tuning of the variational autoencoder temporal prediction network.

[0016] Preferably, the online gradient update saves a full model checkpoint after each preset number of executions. If the mean of the difference residuals increases by more than a preset percentage after multiple consecutive updates, the update is rolled back to the most recent checkpoint.

[0017] The beneficial effects of this invention are:

[0018] First, features of the physical layer, network layer, and business layer are extracted by three encoders: temporal convolution, graph attention, and Transformer, respectively. Then, the cross-modal attention mechanism is used to establish pairwise interactive relationships among the three, which overcomes the one-sidedness of single data source perception and the inability of weighted fusion method to capture nonlinear coupling relationships.

[0019] Second, by using an entropy-aware dynamic decision-making network to adjust the network depth and width in real time based on information entropy, it can automatically expand in highly dynamic scenarios to ensure accuracy and automatically shrink in static scenarios to reduce latency and computational consumption, thus resolving the contradiction between accuracy and efficiency that cannot be achieved by a fixed network structure.

[0020] Third, by using variational time-series prediction networks to predict the probability of future states and using confidence-driven virtual decision-making to lock high-confidence resources in advance, reactive scheduling is upgraded to proactive pre-scheduling, significantly reducing the probability of service interruption and switching latency.

[0021] Fourth, the model is continuously optimized online through differential residual closed-loop feedback: the residuals are backpropagated to fine-tune the parameters of the sensing module to adapt to environmental drift, and the policy update direction is corrected through reward reshaping, so that the system remains stable and continuously optimized in long-term operation. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0023] This invention provides a method for intelligent optimization and allocation of communication resources, specifically comprising the following steps:

[0024] Step S1: Acquisition and multi-granularity preprocessing of multi-source communication data.

[0025] The reference signal received power, signal-to-interference-plus-noise ratio, channel quality indicator, physical resource block utilization, and block error rate are collected for each physical resource block in the communication network and used as physical layer channel state timing data. This data is denoted as matrix X. phy (t)∈R M×5 M represents the total number of schedulable physical resource blocks in the system, and the five columns are, in order, the reference signal received power, the signal-to-interference-plus-noise ratio, the channel quality indicator, the physical resource block utilization rate, and the block error rate.

[0026] The node load rate, backhaul link utilization, handover frequency, number of associated users, and user space distribution density factor of each network node are collected as network layer topology state space data. This data consists of a graph structure G(t) = (V, E(t), X... topo (t) represents, where V is the set of network nodes, E(t) is the set of communication link edges at time t, and X topo (t)∈R ∣V∣×5 The five columns are, in order, node load rate, backhaul link utilization, handover frequency, number of associated users, and user spatial distribution density factor. Simultaneously, the communication link connections between nodes are recorded, forming an adjacency matrix A(t)∈R. ∣V∣×∣V∣ A ij (t)=1 when (v i v j )∈E(t), otherwise it is 0.

[0027] Collect each user's service type identifier, average packet size of the service flow, service arrival interval, and service priority level as semantic feature data for the service layer application. This data is denoted as s. u (t)∈R 6 The first three dimensions are the three-dimensional one-hot encoding of the service type: eMBB corresponds to (1, 0, 0), URLLC corresponds to (0, 1, 0), and mMTC corresponds to (0, 0, 1). The fourth dimension is the average packet size of the service flow. The fifth dimension is the service arrival interval. The sixth dimension is the service priority level, with values ​​ranging from 1 to 5.

[0028] Outliers are removed from the physical layer channel state time-series data using a Hampel filter. For each location, a window is formed by taking a preset number of time slots before and after the current time. The median and corrected absolute median difference (REM) within the window are calculated. The REM is calculated as follows: REM equals 1.4826 multiplied by the median of the absolute differences between each data point within the window and the median. If the absolute difference between the current data point and the median exceeds three times the REM, it is replaced with the median; otherwise, the original value is retained. For missing values, a linear interpolation method is used: if data is missing at the current time, the two nearest data points with available data are found, and the data from these two time points are linearly interpolated according to the time interval ratio. The interpolation result is used to fill in the missing value.

[0029] Z-score normalization is performed on the features of each node in the network layer topology state space data. For each feature, the mean and standard deviation of the feature value across all nodes are calculated. Then, the feature value of each node is subtracted from the mean and divided by the standard deviation. Symmetric normalization with self-loops is performed on the adjacency matrix: first, an identity matrix is ​​added to the adjacency matrix to obtain an adjacency matrix containing self-loops. Then, the degree matrix is ​​calculated. The degree matrix is ​​a diagonal matrix, and its diagonal elements are the sum of the elements in each row of the adjacency matrix containing self-loops. Finally, the degree matrix is ​​multiplied by the negative half power of the degree matrix on the left and right sides of the adjacency matrix containing self-loops.

[0030] The business type identifier in the semantic feature data of the business layer is kept unchanged by one-hot encoding. Z-score normalization is performed on the average packet size and arrival interval of the business flow, and Min-Max normalization is performed on the business priority level. The Min-Max normalization method is to subtract 1 from the priority level and then divide by 4, so that the original value range of 1 to 5 is mapped to the interval of 0 to 1.

[0031] The preprocessed physical layer data, network layer data, and service layer data are aligned along the timestamp dimension to generate a synchronized data frame. Using the timeslot number of the physical layer data as the base time axis, for network layer data and service layer data, if there are no new sampled values ​​in the current timeslot, the most recent historical sampled value is used.

[0032] Step S2: Three-modal feature extraction.

[0033] The physical layer data from the most recent predetermined number of time slots are stacked into a three-dimensional tensor along the time dimension and input into a temporal convolutional encoder consisting of multiple layers of one-dimensional convolutional kernels. Each convolutional kernel slides along both the frequency and time dimensions, and the convolution result is activated by a linear rectified function. The linear rectified function is defined as follows: when the independent variable is greater than 0, the output is equal to the independent variable; when the independent variable is less than or equal to 0, the output is 0. After all convolutional layers, global average pooling is performed along both the frequency and time dimensions, i.e., the average value of each channel is calculated at all frequency and time positions, outputting a spectral evolution feature vector.

[0034] The symmetrically normalized adjacency matrix and the Z-score-normalized node feature matrix from step S1 are input into the graph attention network encoder, which consists of multiple multi-head attention aggregation layers. The single-layer operation is as follows: the learnable linear transformation matrix of each attention head is applied to the feature vectors of each node in the current layer. The transformed feature vectors of the center node and its neighboring nodes are concatenated and then multiplied by the learnable attention vector. A leaky linear rectified function is then applied. The leaky linear rectified function is defined as follows: when the independent variable is greater than 0, the output equals the independent variable; when the independent variable is less than or equal to 0, the output equals 0.01 multiplied by the independent variable. Softmax normalization is performed on the activation values ​​of all neighboring nodes and the center node itself to obtain normalized attention coefficients. Softmax normalization is performed by calculating a power value with the natural constant e as the base and each activation value as the exponent, then dividing each power value by the sum of all power values. The transformed feature vectors of the neighboring nodes are weighted and summed using the normalized attention coefficients. The average of the weighted sums of all attention heads is then applied, followed by an exponential linear unit activation function. The exponential linear unit activation function is defined as follows: when the independent variable is greater than or equal to 0, the output is equal to the independent variable; when the independent variable is less than 0, the output is equal to the value of the exponential function with the natural constant e as the base minus 1.

[0035] After passing through all layers, the output feature vectors of each node are transformed by a learnable linear transformation and activated by hyperbolic tangent, then multiplied by the learnable context query vector. The hyperbolic tangent function is defined as the sum of the exponential function value with the natural constant e as the base and its reciprocal. Softmax normalization is performed on the inner product results of all nodes to obtain the pooling weights for each node. These weights are then used to perform a weighted sum of the feature vectors of each node, outputting the topological association feature vector.

[0036] The business feature vectors of all users are organized into a matrix, and a learnable global query vector is added as the first row to input a Transformer semantic encoder. This encoder consists of multiple layers of multi-head self-attention sublayers and feedforward neural network sublayers stacked alternately. In each layer, the input matrix is ​​first processed by three learnable projection matrices to generate a query matrix, a key matrix, and a value matrix. The query matrix, key matrix, and value matrix are then divided into a predetermined number of attention heads along the column direction. In each attention head, the dot product between the query submatrix and the transpose of the key submatrix is ​​calculated, divided by a scaling factor, and then softmax normalized for each row of the matrix. This is then multiplied by the value submatrix to obtain the output of that attention head. The outputs of all attention heads are concatenated along the column direction and then mapped through a learnable output projection matrix. Finally, residual connections and layer normalization are applied. The residual connection is achieved by adding the input and output of the sub-layer. Layer normalization involves independently calculating the mean and variance of each row of the summed result, then subtracting the mean of the row from each element, dividing by the square root of the sum of the square of the row's standard deviation and a minimal constant, and finally multiplying by a learnable scaling parameter and adding a learnable translation parameter. This is followed by a feedforward neural network sub-layer, consisting of two fully connected layers. The first fully connected layer is activated by a linear rectified function, and the second fully connected layer maps back to the original dimension. Residual connections and layer normalization are applied again. After all encoding layers, the vector in the first row of the output matrix is ​​taken as the business intent feature vector.

[0037] Step S3, cross-modal attention fusion and manifold embedding.

[0038] The spectrum evolution feature vector, topology association feature vector, and business intent feature vector output in step S2 are mapped to the same shared dimension space through independent linear projection matrices to obtain the physical layer projection vector, network layer projection vector, and business layer projection vector. The linear projection method is to multiply the feature vector by the projection matrix on the left and then add the bias vector.

[0039] The cross-attention matrix is ​​calculated pairwise between the three projection vectors. Specifically, the source modality's projection vector is mapped to a query vector using a learnable query projection matrix, and the target modality's projection vector is mapped to a key vector using a learnable key projection matrix. The query vector and key vector have the same dimension. The dot product between the query vector and the transpose of the key vector is calculated, resulting in a square matrix whose rows and columns are the same as the dimensions of the shared dimension space. The dot product result is divided by a scaling factor, and then softmax normalization is performed on each row of the square matrix to obtain the cross-modal attention matrix of the source modality to the target modality. The scaling factor is the square root of the query vector's dimension.

[0040] The cross-modal attention matrices from the physical layer to the network layer, from the physical layer to the service layer, and from the network layer to the service layer are calculated sequentially as described above. Each cross-modal attention matrix is ​​multiplied by the value projection vector of the corresponding modality to obtain three weighted information vectors. The three information vectors are concatenated along the row direction into a long vector, which is then mapped by the output linear transformation matrix and a bias vector is added to obtain the three-modal fusion vector.

[0041] The trimodal fusion vector is embedded into a low-dimensional manifold space using an isometric mapping algorithm to obtain a joint state representation vector. The isometric mapping algorithm constructs a K-nearest neighbor graph in the cache pool and calculates the geodesic distance between each pair of sample points, then performs multidimensional scaling to obtain the low-dimensional embedding coordinates. The geodesic distance is approximated by using the shortest path algorithm to calculate the shortest path distance between sample points on the K-nearest neighbor graph. The goal of multidimensional scaling is to find a set of low-dimensional coordinates such that the Euclidean distance between points in the low-dimensional space is as consistent as possible with the geodesic distance in the original high-dimensional space. For newly arrived samples, an incremental method is used: only the geodesic distance between the new sample and a predetermined number of nearest neighbors in the cache pool is calculated, and the landmark multidimensional scaling method is used to solve for the low-dimensional embedding coordinates, without recalculating the entire geodesic distance matrix for all historical samples.

[0042] Step S4: Entropy-aware dynamic decision-making.

[0043] The joint state representation vector output from step S3 is input into a decision network containing a differentiable entropy estimator. This entropy estimator maintains a sliding window storing a preset number of recent historical joint state representation vector samples. Based on these historical samples, the probability density of the current sample is calculated using a Gaussian kernel density estimation method. The Gaussian kernel density estimation method is as follows: Using the current sample as the center and a preset bandwidth as the standard deviation of the kernel function, calculate the Euclidean distance between each historical sample and the current sample within the window. Substitute this distance into the standard Gaussian kernel function, calculate the average, and then divide by the bandwidth raised to the power of d, where d is the dimension of the joint state representation vector. The standard Gaussian kernel function is defined as an exponential function with the natural constant e as the base and the negative half of the square of the independent variable as the exponent, then divided by the square root of 2π. After the probability density is calculated, the information entropy value is calculated using logarithmic integration. The numerical approximation of logarithmic integration is as follows: sample multiple points from the estimated probability density distribution, calculate the logarithm of the probability density at each point, take the average, and then take the negative value.

[0044] Based on the calculated information entropy value, the number of hidden layers and the number of neurons per layer in the decision network are dynamically adjusted. The adjustment method is as follows: the current number of layers equals the baseline number of layers plus the adjustable amplitude multiplied by the hyperbolic tangent function value (scaled from the difference between the information entropy value and the medium entropy value). The hyperbolic tangent function is defined as: the exponential function value with the natural constant e as the base minus its reciprocal, then divided by the sum of the two. The current number of neurons per layer equals the baseline number plus the adjustable amplitude multiplied by the difference between the information entropy value and the high entropy threshold, mapped by the Logistic Sigmoid function. The Logistic Sigmoid function is defined as: 1 divided by 1 plus the exponential function value with the natural constant e as the base, where the independent variable of the exponential function is the negative of the input value. The medium entropy value is half the sum of the high entropy threshold and the low entropy threshold. The adjustment rule is: the higher the information entropy value, the more layers and neurons are added to increase model capacity; the lower the information entropy value, the fewer layers and neurons are added to reduce inference latency.

[0045] A near-end policy optimization algorithm is executed on a dynamically adjusted decision network. This algorithm maintains a policy network and a value function network, both sharing the same dynamic network structure. However, the last layer of the policy network outputs probability distribution parameters for resource allocation actions, while the last layer of the value function network outputs a scalar as a state value estimate. The outputs of the policy network include: Bernoulli distribution probability parameters for each physical resource block, classification distribution probability parameters for transmit power adjustment levels, classification distribution probability parameters for time slot allocation schemes, and Gaussian distribution mean and standard deviation parameters for edge computing power allocation ratios. Samples are taken from these probability distributions to obtain the specific resource allocation action vector for the current decision cycle. The algorithm's objective function is a pruning substitution objective function, calculated as follows: the probability ratio of the new policy to the old policy on executed actions is calculated, and then this probability ratio is multiplied by the dominance estimate. To prevent excessive policy updates, the probability ratio is pruned to the range of 1 plus or minus a pruning coefficient before multiplying by the dominance estimate, and the minimum of the two is taken. The advantage estimate is calculated as follows: starting from the current moment, the sum of the instantaneous reward at each moment and the discount of the state value estimate at the next moment is subtracted from the state value estimate at the current moment, and then the sum is accumulated forward in an exponential decay manner.

[0046] Step S5, forward confidence-driven resource pre-scheduling.

[0047] The joint state representation vector sequence of the current time step and a predetermined number of historical decision cycles is input into a temporal prediction network based on a variational autoencoder framework. This network consists of an encoder and a decoder. The encoder's input is the historical state sequence, which, after processing by a long short-term memory network, maps the hidden state and cell state of the last time step through two independent linear layers to the logarithms of the mean and variance parameters of a Gaussian distribution in the latent space. Latent variables are sampled from this Gaussian distribution using a reparameterization technique: a random vector is first sampled from a standard normal distribution, multiplied by the standard deviation vector, and then the mean vector is added. The latent variables are input into the decoder, which is another long short-term memory network. Using the latent variables as initial input, it progressively generates state predictions for the next predetermined number of decision cycles and the standard deviation of each prediction. During training, the parameters of the encoder and decoder are optimized by maximizing the lower bound of evidence, which includes two terms: reconstruction log-likelihood and KL divergence regularization.

[0048] The confidence score for the predicted state in each future decision-making cycle is calculated based on the standard deviation. The calculation method is: 1 divided by 1 plus the Euclidean norm of the predicted standard deviation vector. The Euclidean norm of the predicted standard deviation vector is equal to the square root of the sum of the squares of its components. The confidence score ranges from 0 to 1; the larger the standard deviation, the lower the confidence score.

[0049] The predicted future state value with a confidence score exceeding a preset threshold is input into the entropy-aware dynamic decision-making network in step S4 to perform virtual forward propagation, obtaining the corresponding virtual resource allocation action. Virtual forward propagation only calculates the output of the policy network and does not interact with the environment. In the current decision cycle, a portion of resources is reserved according to the virtual resource allocation action. When reserving resources, the physical resource block corresponding to the resource block allocation identifier in the virtual action is marked as reserved; the power margin corresponding to the power adjustment amount in the virtual action is pre-stored at the physical layer; and the computing power resources corresponding to the computing power allocation ratio in the virtual action are pre-reserved at the edge nodes. A timestamp label for the future decision cycle is attached to the reserved resources. The number of reserved resources is multiplied by a safety factor of a sub-1 to retain some margin to cope with prediction errors.

[0050] When the actual future decision cycle is reached, the true joint state representation vector at that moment is obtained, and the Mahalanobis distance between it and the predicted state for that cycle is calculated. The Mahalanobis distance is calculated as follows: the transpose of the difference between the actual state vector and the predicted state vector is multiplied by the inverse of the prediction covariance matrix, then multiplied by the square root of the difference vector. The prediction covariance matrix is ​​a diagonal matrix with the squares of the components of the prediction standard deviation vector as its diagonal elements. The advantage of Mahalanobis distance is that it normalizes the bias through the prediction variance, allowing for larger biases without triggering release when the prediction variance is large.

[0051] If the Mahalanobis distance is less than a preset deviation threshold, the actual state is determined to be consistent with the predicted state. Reserved resources are activated, and previously reserved physical resource blocks, power margins, and computing resources are directly allocated to the current business flow without rerunning the decision network, thus saving inference time. If the Mahalanobis distance is greater than or equal to the preset deviation threshold, the prediction deviation is determined to be too large. Reserved resources are released, and the reserved physical resource blocks and power margins are returned to the resource pool. Real-time re-decision is triggered, i.e., step S4 is rerun to calculate resource allocation using the current actual state as input. Each time reserved resources are released, the release event is recorded as a negative sample in the experience replay buffer for subsequent online fine-tuning of the variational autoencoder temporal prediction network to gradually reduce future prediction deviations.

[0052] Step S6: Differential resource optimization closed-loop feedback.

[0053] After the t-th decision cycle is completed, the actual end-to-end latency, average throughput, and packet loss rate within that cycle are collected as the actual service quality index vector. The predicted service quality index vector for that cycle is obtained from the auxiliary output header of the median function network in step S4. Both the actual and predicted service quality index vectors contain three dimensions: latency, throughput, and packet loss rate. The Euclidean distance between them is calculated as the difference residual. The Euclidean distance is calculated as the square root of the sum of the squares of the differences between the two vectors in each dimension. Before calculation, each dimension needs to be normalized to eliminate dimensional differences: latency is divided by the maximum allowable latency, throughput is divided by the maximum theoretical throughput, and the packet loss rate itself is dimensionless and not normalized.

[0054] The difference residuals are used as the loss signal for online fine-tuning. The loss function is defined as half the square of the difference residuals. For all learnable parameters in the temporal convolutional encoder, graph attention network encoder, and Transformer semantic encoder in step S2, and in the cross-modal attention fusion module in step S3, the gradient of the loss function with respect to each parameter is calculated using automatic differentiation, and then the parameters are updated according to the gradient descent direction. The update magnitude is the gradient multiplied by the online learning rate. Each online fine-tuning performs only one gradient update on a single sample, without multiple iterations, to avoid catastrophic forgetting. A full model checkpoint is saved after a preset number of decision cycles. If the mean increase of the difference residuals exceeds a preset percentage after multiple consecutive online fine-tunings, the model is rolled back to the most recent checkpoint.

[0055] The differential residual is multiplied by a penalty coefficient and then subtracted from the original immediate reward of the near-end policy optimization algorithm in step S4 to obtain the corrected immediate reward. The original immediate reward consists of four terms: a positive throughput term, a negative latency term, a negative packet loss rate term, and a negative energy consumption term, each multiplied by its respective weight coefficient and then summed. The corrected immediate reward is used for the advantage estimation calculation in subsequent policy gradient updates; that is, the immediate reward at each time step is replaced by the corrected immediate reward.

[0056] To better understand the technical solutions described in this invention, the following specific embodiments are provided:

[0057] Specific Implementation Example: Single-Cycle Complete Decision-Making Process for Dense Urban Cells in 5G

[0058] I. Scene Setting

[0059] Cell: 5G NR macro cell Cell-001, 3.5GHz band, 100MHz bandwidth (273 PRBs)

[0060] Decision cycle number: t=1000, cycle length ΔT=100ms;

[0061] Currently online users: 3

[0062] User A: eMBB service (high-definition video downlink), priority level 4;

[0063] User B: URLLC service (downlink of autonomous driving control commands), priority level 5 (highest);

[0064] User C: mMTC service (environmental sensor uplink), priority level 2;

[0065] Neighboring cells: Cell-002, Cell-003, Cell-004 (a total of 4 nodes forming the topology);

[0066] II. Step A: Data Acquisition and Preprocessing

[0067] A.1 Physical Layer Data Acquisition

[0068] Physical layer data of PRB#1 to PRB#5 were collected in the most recent 5 time slots (t-4 to t).

[0069] Table A-1: ​​Raw Physical Layer Data (5 PRBs × 5 Time Slots × 5 Indicators)

[0070]

[0071] A.2 Hampel Filtering for Noise Reduction (Taking PRB#3 SINR as an Example)

[0072] Take the SINR sequence of PRB#3 from time slot t-4 to t: {6.8, 6.2, 5.9, 6.5, 7.2}.

[0073] Step 1: Sort → {5.9, 6.2, 6.5, 6.8, 7.2}, median m med =6.5;

[0074] Step 2: The absolute deviation of each value from the median: {0.6, 0.3, 0, 0.3, 0.7}, after sorting, take the median: MAD = 0.3;

[0075] Step 3: Correct for absolute median: σ MAD =1.4826 × 0.3 = 0.4448;

[0076] Step 4: Threshold 3×σ MAD =1.3344;

[0077] Step 5: The difference between the current value 7.2 and the median 6.5 = 0.7 < 1.3344 → Keep the original value 7.2;

[0078] Example: Outlier detection scenario: If SINR changes from 5.9 to 1.2 at time t-2 (burst interference), the sequence becomes {6.8, 6.2, 1.2, 6.5, 7.2};

[0079] The order is {1.2, 6.2, 6.5, 6.8, 7.2}, and the median remains 6.5.

[0080] Absolute deviation: {5.3, 0.3, 0, 0.3, 0.7}, MAD = 0.3;

[0081] |1.2-6.5|=5.3>1.3344 → This is an outlier and should be replaced with the median of 6.5.

[0082] A.3 Linear interpolation of missing values ​​(taking the CQI missing value of PRB#1 in time slot t as an example)

[0083] Suppose that PRB#1 has a CQI of 11 in time slot t-1 and a CQI of 10 in time slot t+1 (taken from the next batch of data).

[0084] Interpolation result: Round to 11;

[0085] A.4 Network Layer Topology Data Acquisition and Preprocessing

[0086] Original node feature data for the four communities:

[0087] Table A-2: Original Node Feature Data

[0088] node load rate Backhaul utilization Switching frequency Number of associated users Density factor (users / km²) Cell-001 0.76 0.68 45 215 860 Cell-002 0.52 0.45 28 138 552 Cell-003 0.88 0.82 62 287 1148 Cell-004 0.35 0.30 15 96 384

[0089] Adjacency matrix A (1 indicates connected, 0 indicates not connected):

[0090]

[0091] Z-score normalization (taking load rate as an example):

[0092] mean μ ρ =(0.76+0.52+0.88+0.35) / 4=0.6275;

[0093] Standard deviation:

[0094]

[0095] Normalization results:

[0096] Cell-001: (0.76-0.6275) / 0.2061=0.6429;

[0097] Cell-002: (0.52-0.6275) / 0.2061=-0.5216;

[0098] Cell-003: (0.88-0.6275) / 0.2061=1.2251;

[0099] Cell-004: (0.35-0.6275) / 0.2061=-1.3464;

[0100] Similarly, the other four dimensions are calculated to obtain the normalized node feature matrix:

[0101] Table A-3: Node Feature Matrix after Z-score Normalization

[0102]

[0103] Symmetric normalization of adjacency matrix:

[0104] First, add a self-loop:

[0105]

[0106] Calculate the degree matrix:

[0107]

[0108] Perform symmetric normalization:

[0109] ;

[0110] First calculate:

[0111] :

[0112]

[0113] Multiply by the left again:

[0114] :

[0115]

[0116] A.5 Semantic Data Acquisition and Preprocessing at the Business Layer

[0117] Table A-4: Raw Semantic Data of the Business Layer

[0118]

[0119] one-hot encoding:

[0120] User A (eMBB): [1, 0, 0];

[0121] User B (URLLC): [0, 1, 0];

[0122] User C (mMTC): [0, 0, 1];

[0123] Packet size Z-score normalization (assuming μ within the statistical window) p =600Byte, σ p =550Byte):

[0124] User A: (1500-600) / 550=1.6364;

[0125] User B: (200-600) / 550 = -0.7273;

[0126] User C: (50-600) / 550 = -1.0000;

[0127] Arrival interval Z-score normalization (let μ) τ =100ms, σ τ =150ms):

[0128] User A: (10-100) / 150 = -0.6000;

[0129] User B: (1-100) / 150 = -0.6600;

[0130] User C: (300-100) / 150=1.3333

[0131] Priority Min-Max normalization (original 1~5 mapped to [0,1]):

[0132] User A: (4-1) / 4=0.75;

[0133] User B: (5-1) / 4=1.00;

[0134] User C: (2-1) / 4=0.25;

[0135] Table A-5: Preprocessed Business Feature Vectors

[0136]

[0137] Step B: Three-modal feature extraction

[0138] B.1 Temporal Convolutional Encoder

[0139] The most recent T c =10 time slots of physical layer data stacked into a three-dimensional tensor X phy (t)∈R 10×273×5 .

[0140] A 5×3 input sub-block is constructed using the 5-dimensional data of PRB#1 over the most recent 3 time slots (t-2 to t):

[0141]

[0142] Let the weight of the first convolutional kernel in the first layer be W. (0) (1,c,i,j)={c=1:[0.1,-0.2,0.1;0.0,0.3,-0.1;-0.1,0.0,0.2],c=2 :[-0.2, 0.1, 0.0; 0.2, -0.1, 0.3; 0.1, 0.2, -0.1], c=3: [0.0, 0.2, -0.1 ;-0.1, 0.1, 0.0; 0.2, -0.2, 0.1], c=4: [0.1, 0.0, 0.2; 0.0, -0.2, 0.1;- 0.1, 0.3, 0.0], c=5: [-0.1, 0.1, 0.0; 0.2, 0.0, -0.1; 0.0, 0.2, -0.2]};

[0143] The kernel size is F m =3,F d =3. At position PRB#1 (m=1, d out Calculate the convolution at position i=3 (align the convolution kernel center, and only consider the contribution of PRB#1 itself, i.e., position i=0):

[0144] c=1 (RSRP): Input [-93, -90, -89], kernel [0.0, 0.3, -0.1] → (-93)×0.0+(-90)×0.3+(-89)×(-0.1)=0-27+8.9=-18.1;

[0145] c=2 (SINR): Input [12.2, 13.0, 13.5], kernel [0.2, -0.1, 0.3] → 12.2×0.2+13.0×(-0.1)+13.5×0.3=2.44-1.30+4.05=5.19;

[0146] c=3 (CQI): Input [9, 11, 11], kernel [-0.1, 0.1, 0.0] → 9×(-0.1)+11×0.1+11×0.0=-0.9+1.1+0=0.2;

[0147] c=4 (utilization): Input [0.50, 0.48, 0.46], kernel [0.0, -0.2, 0.1] → 0.50×0.0+0.48×(-0.2)+0.46×0.1=0-0.096+0.046=-0.05;

[0148] c=5 (BLER): Input [0.014, 0.010, 0.009], kernel [0.2, 0.0, -0.1] → 0.014×0.2+0.010×0.0+0.009×(-0.1)=0.0028+0-0.0009=0.0019;

[0149] Accumulate and add bias b (0) (1) = 0.05:

[0150] Output = max(0, -18.1 + 5.19 + 0.2 - 0.05 + 0.0019 + 0.05) = max(0, -12.7081) = 0;

[0151] The output is 0, indicating that this convolutional kernel did not detect any significant features at the PRB#1 position. When PRB#5 (the PRB with the best channel quality) is calculated in the same way, a positive output value is assumed to be obtained.

[0152] After three layers of convolution (number of channels: 5→32→64→128) and global average pooling, a 128-dimensional spectral evolution feature vector F is obtained. phy (t), whose first 3 dimensions are assumed to be: F phy (t)[:3]=[0.15,-0.08,0.22];

[0153] B.2 Graph Attention Network Encoder

[0154] Input: Symmetric normalized adjacency matrix (From A.4) and normalized node feature matrix ∈R 4×5 (From Table A-3).

[0155] Taking the first attention head in the first layer as an example, calculate the attention coefficients of node Cell-001 to each neighborhood.

[0156] Let W be the linear transformation matrix. (1，1) ∈R 32×5 (The first two lines are used for demonstration purposes), attention vector a (1，1) ∈R 64 (Take the first 4 dimensions):

[0157]

[0158] The eigenvector h1 of Cell-001 (0) Perform a linear transformation on the expression [0.6429, 0.5913, 0.5217, 0.5287, 0.5217] (taking the first two dimensions):

[0159]

[0160] Similarly, calculate the other nodes:

[0161] Cell-002: W (1，1) h2 (0) [:2] = [-0.2562, -0.0689];

[0162] Cell-003: W (1，1) h3 (0) [:2]=[0.6833, 0.2357];

[0163] Cell-004: W (1，1) h4 (0) [:2] = [-0.7052, -0.2749];

[0164] Calculate the attention score of Cell-001 on itself (by concatenating itself with itself):

[0165]

[0166] Cell-001 to Cell-002 (spliced ​​[0.2782, 0.1083, -0.2562, -0.0689]):

[0167]

[0168] Cell-001 to Cell-003 (spliced ​​[0.2782, 0.1083, 0.6833, 0.2357]):

[0169]

[0170] Cell-001 and Cell-004 are not connected, so no calculation is performed.

[0171] softmax normalization:

[0172]

[0173] Denominator: 1.0938 + 0.9993 + 1.2351 = 3.3282;

[0174] α 11 =1.0938 / 3.3282=0.3286, α 12 =0.9993 / 3.3282=0.3002, α 13 =1.2351 / 3.3282=0.3712;

[0175] Analysis: Cell-001 showed the highest attention to Cell-003 (0.3712) because Cell-003 had the highest load rate (normalized 1.2251), which is correlated with the high load of Cell-001.

[0176] The outputs of the eight attention heads are averaged, activated by an ELU, and then passed through a second GAT layer to obtain the final feature vectors of four nodes. Attention pooling is then performed.

[0177] Let the final feature vectors of each node be (first 3 dimensions):

[0178]

[0179] Let the pooling parameters q[:3] = [0.5, 0.3, 0.2], b pool [:3]=[0.1, 0.0, -0.1], W pool It is approximately an identity matrix.

[0180] Calculate the attention score for each node:

[0181]

[0182] softmax weights:

[0183]

[0184] Analysis: Cell-003 obtained the highest pooling weight of 0.2853 because it has the highest load and the most frequent switching, and is located in the most critical topology position.

[0185] Topological association feature vector (first 3 dimensions): F topo (t)[:3]=0.2622×[0.35, 0.21, -0.08]+0.2326×[0.12, 0.08, 0.03]+0.2853×[0.52, 0.33, -0 .15]+0.2199×[0.05, 0.02, 0.01]=[0.2790, 0.1722, -0.0546]=[0.2790, 0.1722, -0.0546];

[0186] B.3 Transformer Semantic Encoder

[0187] Input the business feature matrices of 3 users (from Table A-5):

[0188]

[0189] Add a learnable global query vector q global =[0.1, 0.2, 0.3, 0.0, -0.1, 0.2] as the 0th row, forming a 4×6 extended matrix. After linear projection to 64 dimensions, it is input into a 2-layer Transformer encoder.

[0190] After two layers of multi-head self-attention encoding, take the first three dimensions of the approximate value of the 0th row (the output corresponding to the global query vector): F biz (t)[:3]=[0.31,-0.12,0.05];

[0191] IV. Step C: Cross-modal attention fusion and manifold embedding

[0192] C.1 Dimensional Alignment Projection

[0193] The feature vectors of the three modes are linearly projected to the shared dimension d. shared =128.

[0194] Let W proj topo The first 3 rows and first 3 columns:

[0195]

[0196] Network layer projection vectors (first 3 dimensions):

[0197]

[0198] Similarly (omitting the specific projection matrix), assume:

[0199]

[0200] C.2 Cross-modal attention matrix calculation

[0201] Take the attention of the physical layer to the network layer as an example. Let W... Q phy [:3,:3] and W K topo [:3,:3]:

[0202]

[0203] Calculate query vector Q phy =W Q phy f phy The first 3 dimensions:

[0204]

[0205] Calculate the key vector K topo =W K topo f topo The first 3 dimensions:

[0206]

[0207] Calculate Q phy K topo T The dot product matrix of the first 3 dimensions. Since the full dimension is 128, the dot product of the first 3 dimensions needs to be magnified by 128 / 3 ≈ 42.67 times to approximate the complete result:

[0208] First, calculate the dot product matrix of the first 3 dimensions:

[0209]

[0210] After magnification by 42.67 times, divide by the scaling factor (d) k ) 0.5 =8:

[0211]

[0212] Analysis: The attention score matrix shows that the attention score between the third dimension of the physical layer (corresponding to CQI / utilization-related features) and the first dimension of the network layer (load rate) is 0.021251 (the largest positive number), indicating that when the physical layer channel quality is good, the model pays more attention to the load status of the network layer; the attention score between the third dimension of the physical layer and the third dimension of the network layer (handover frequency) is -0.028494 (the largest negative number), indicating that when the channel quality is good, the model reduces its attention to handover frequency—because when the channel quality is good, there is no need to trigger handover to improve connectivity.

[0213] Perform softmax normalization on each row (considering the complete 128 columns, assuming the exponent of the remaining 125 columns is approximately 1):

[0214] Line 0: exp(0.004497)=1.004507, exp(0.005019)=1.005031, exp(-0.006027)=0.993990 Denominator≈1.004507+1.005031+0.993990+125=128.003528;

[0215] α 00 =1.004507 / 128.003528=0.007848, α 01 =1.005031 / 128.003528=0.007852, α 02 =0.993990 / 128.003528=0.007766;

[0216] Line 1: exp(0.001024)=1.001024, exp(0.001142)=1.001142, exp(-0.001371)=0.998630 Denominator≈1.001024+1.001142+0.998630+125=128.000796;

[0217] α 10 =1.001024 / 128.000796=0.007820, α 11 =1.001142 / 128.000796=0.007821, α 12 =0.998630 / 128.000796=0.007802;

[0218] Line 2: exp(0.021251)=1.021479, exp(0.023720)=1.024006, exp(-0.028494)=0.971911. Denominator ≈ 1.021479+1.024006+0.971911+125=128.017396;

[0219] α 20 =1.021479 / 128.017396=0.007979, α 21 =1.024006 / 128.017396=0.007999, α 22 =0.971911 / 128.017396=0.007592;

[0220] Analysis: The second row (physical layer PRB utilization / CQI dimension) shows more attention to the first column (load rate, 0.007979) and the second column (backhaul utilization, 0.007999) of the network layer than to the third column (switching frequency, 0.007592), indicating that when physical layer resources are strained, the model pays more attention to the load and backhaul status of the network layer than to the switching frequency.

[0221] C.3 Cross-weighted fusion

[0222] Let V ph tp Approximately an identity matrix. Information obtained by the physical layer from the network layer (first 3 dimensions):

[0223] Take A ph→tp The first 3 rows and first 3 columns of the submatrix and f topo Multiplying the first 3 dimensions of (t) (ignoring the contributions of the other 125 dimensions):

[0224]

[0225] 0th dimension: 0.007848×0.1794+0.007852×0.0311+0.007766×(-0.0620)=0.001408+0.000244-0.000481=0.001171;

[0226] First dimension: 0.007820×0.1794+0.007821×0.0311+0.007802×(-0.0620)=0.001403+0.000243-0.000484=0.001162;

[0227] Second dimension: 0.007979×0.1794+0.007999×0.0311+0.007592×(-0.0620)=0.001431+0.000249-0.000471=0.001209;

[0228] Similarly, calculate

[0229] A ph→bz f biz (t) and A tp→bz f biz The first 3 dimensions (hypothetical values) of (t):

[0230] [0.00095, 0.00112, 0.00087], [0.00132, 0.00105, 0.00098]

[0231] Concatenating the three results yields a 9-dimensional vector:

[0232] Concat=[0.001171, 0.001162, 0.001209, 0.00095, 0.00112, 0.00087, 0.00132, 0.00105, 0.00098]

[0233] Let W out [:3,:9]:

[0234] Three-modal fusion vector (first 3 dimensions, ignoring bias):

[0235] f fused (t)[0]=0.5×0.001171+(-0.2)×0.001162+0.1×0.001209+0.3×0.00095+0.0×0.00112+(-0.1)×0.00087+0.2×0.00132+0.1×0 .00105+(-0.3)×0.00098=0.000586-0.000232+0.000121+0.000285+0-0.000087+0.000264+0.000105-0.000294=0.000748;

[0236] f fused (t)[1]=(-0.1)×0.001171+0.6×0.001162+0.2×0.001209+(-0.2)×0.00095+0.4×0.00112+0.1×0.00087+0.0×0.00132+(-0.2 )×0.00105+0.3×0.00098=-0.000117+0.000697+0.000242-0.000190+0.000448+0.000087+0-0.000210+0.000294=0.001251;

[0237] f fused (t)[2]=0.3×0.001171+0.1×0.001162+(-0.4)×0.001209+0.1×0.00095+(-0.3)×0.00112+0.5×0.00087+(-0.2)×0.00132+0.3×0 .00105+0.1×0.00098=0.000351+0.000116-0.000484+0.000095-0.000336+0.000435-0.000264+0.000315+0.000098=0.000326;

[0238] C.4 Manifold Embedding

[0239] f in 128 dimensions fused (t) is embedded into a 64-dimensional manifold space via an isometric mapping. PCA is used as a linear approximation of the Isomap. Assume the first three principal components of the 1000 historical samples in the cache pool are:

[0240]

[0241] Projection score (using the first 3 dimensions):

[0242] PC1:

[0243] 0.12×0.000748+(-0.08)×0.001251+0.05×0.000326=0.000090-0.000100+0.000016=0.000006;

[0244] PC2: 0.06×0.000748+0.15×0.001251+(-0.03)×0.000326=0.000045+0.000188-0.000010=0.000223;

[0245] PC3: (-0.04)×0.000748+0.02×0.001251+0.11×0.000326=-0.000030+0.000025+0.000036=0.000031;

[0246] Joint state representation vector (first 3 dimensions):

[0247]

[0248] Analysis: The second dimension (0.000223) is significantly larger than the first dimension (0.000006) and the third dimension (0.000031), indicating that the "topology load pressure" feature is dominant in the current state, corresponding to a high-load network state with a load rate of 0.76 for Cell-001 and 0.88 for Cell-003.

[0249] V. Step D: Entropy-aware dynamic decision-making

[0250] D.1 Information Entropy Calculation

[0251] Differentiable entropy estimator maintains the nearest L KDE =200 z-sliding windows. Statistical parameters of the first 3 dimensions of the samples within the window: μ1=0.000010, σ1=0.000080; μ2=0.000180, σ2=0.000060; μ3=0.000025, σ3=0.000050;

[0252] The current sample z(t)[:3]=[0.000006, 0.000223, 0.000031];

[0253] Calculate kernel bandwidth:

[0254]

[0255] The most recent sample within the window

[0256] Taking z(t-1)[:3]=[0.000015,0.000195,0.000028] as an example, calculate the kernel function value:

[0257] Euclidean distance:

[0258]

[0259] Gaussian kernel function value:

[0260]

[0261] Estimating the probability density (assuming approximately 30 samples within the window are within the effective range of the kernel function, and the average kernel function value is approximately 0.18):

[0262]

[0263] Due to h dz Minimal, using the normalized entropy value. Estimate the current normalized information entropy based on the dispersion of samples within the window:

[0264] H(t)≈0.68;

[0265] Analysis: H(t) = 0.68 is at a moderately high level, falling between the low entropy threshold H. low =0.25 and high entropy threshold H high A value between 0.85 and 0.85 indicates that the current state has some uncertainty but is not extreme, and the network needs to be expanded appropriately.

[0266] D.2 Dynamic Network Structure Adjustment

[0267] Depth adjustment:

[0268]

[0269] Width adjustment:

[0270]

[0271] Analysis: The network was expanded from the baseline 3 layers with 256 neurons to 4 layers with 315 neurons, increasing the capacity by approximately (4 / 3)×(315 / 256)-1=64% to cope with the medium-high information entropy of 0.68.

[0272] D.3 PPO Algorithm Execution

[0273] Forward propagation of z(t) is performed on a decision network with 4 layers and 315 neurons.

[0274] (1) Spectrum Resource Block Allocation: The policy network outputs the allocation probabilities of 273 PRBs. The probabilities of the first 5 PRBs are:

[0275] p RB [:5]=[0.12, 0.08, 0.73, 0.21, 0.05];

[0276] Analysis: PRB#3 has the highest allocation probability (0.73) because although its SINR (7.2dB) is not high, its current utilization rate has dropped from 0.85 to 0.76, and the adjacent PRB#2 (SINR 14.8dB) and PRB#4 (SINR 11.0dB) have good channel quality—cross-modal fusion allows the model to capture the release trend of PRB#3. PRB#5 has the lowest probability (0.05). Although it has the best channel quality (SINR 16.0dB), its utilization rate is only 0.32, and the model tends to reserve it for high-priority services that may arrive in the future.

[0277] The sampling yields the binary allocation vector (first 5 bits):

[0278] a RB [:5]=[0,0,1,0,0], which means assigning PRB#3 to the current highest priority user B (URLLC).

[0279] (2) Transmit power control: probability distribution of 5 discrete power levels:

[0280]

[0281] Sampling: a pwr =+1dB;

[0282] Analysis: The probability of +1dB is the highest (0.45). Under the current high load, a moderate increase in power can improve the reliability of URLLC. Meanwhile, the probability of +3dB is only 0.10, indicating that the model tends to avoid excessive power increments that could cause interference.

[0283] (3) Time slot allocation: Probabilities of the four allocations:

[0284]

[0285] Sampling: a slot =(4:1);

[0286] Analysis: The (4:1) ratio has the highest probability (0.55), which is suitable for the current business scenario where eMBB downlink and URLLC downlink are dominant.

[0287] (4) Edge computing power allocation: Gaussian distribution parameter μ comp =0.35, σ comp =0.08, sampled a comp =0.32;

[0288] Combined action vectors:

[0289]

[0290] The value function network outputs the current state value V(z(t)) = 2.35 and predicts QoS:

[0291]

[0292] VI. Step E: Proactive Resource Pre-scheduling

[0293] E.1 Variational Time Series Prediction

[0294] Input the historical M=50 cycles of {z(951), ..., z(1000)} into the VAE encoder.

[0295] Let h be the last hidden state of the LSTM encoder. last ∈R 128 The first 8 dimensions:

[0296] [0.12, -0.08, 0.23, 0.05, -0.15, 0.09, 0.18, -0.03];

[0297] The latent distribution parameters (taking the first two dimensions) are obtained through a linear layer. Let W... μ [:2,:8]:

[0298]

[0299] μ w，1 =0.2×0.12+0.1×(-0.08)+(-0.3)×0.23+0.0×0.05+0.2×(-0.15)+(-0.1)×0.09+0.1×0.18+0.0×(-0.03)+0.01=0.024-0.008-0.069+0-0.030-0.009+0.018+0+0.01=-0.064;

[0300] μ w，2=(-0.1)×0.12+0.2×(-0.08)+0.1×0.23+(-0.2)×0.05+0.0×(-0.15)+0.3×0.09+(-0.1)×0.18+0.2×(-0.03)+(-0.02)=-0.012-0.016+0.023-0.010+0+0.027-0.018-0.006-0.02=-0.032;

[0301] Let logσ w If [:2] = [-1.20, -1.50], then σ w [:2]=[0.3012, 0.2231];

[0302] Sampling ε = [0.52, -0.31], we obtain the latent variables:

[0303]

[0304] The decoder uses w as the initial input and gradually generates state predictions for the next N=10 cycles.

[0305] The predicted mean and standard deviation (first 3 dimensions) for the first future period (t+1, i.e., 100ms later):

[0306]

[0307] Confidence score:

[0308]

[0309] Analysis: c(t+1) = 0.467 < θ c =0.75, reservation is not triggered.

[0310] The 5th future period (t+5, i.e., after 500ms):

[0311]

[0312] Analysis: c(t+5)=0.613<0.75, so it still does not trigger.

[0313] The 8th future period (t+8, i.e., after 800ms):

[0314]

[0315] Analysis: c(t+8)=0.754>θ c =0.75, triggering resource reservation. The model has high confidence in the state prediction after 800ms, and the prediction shows that the load pressure will increase further (the second dimension increases from the current 0.000223 to 0.000268).

[0316] E.2 Virtual Decision-Making and Resource Reservation

[0317] Will Input entropy-aware dynamic decision network (at this point, the estimated information entropy H(t+8)≈0.72, expanded to approximately 4 layers and 330 neurons), output virtual action:

[0318]

[0319] Analysis: The virtual action has allocated more PRB#4 than the current one, the power requirement has increased from +1dB to +2dB, and the computing power has increased from 32% to 40%, indicating that the model predicts that more high-priority services will arrive in the future.

[0320] Resource reservation is performed in the current period t (safety factor κ=0.8):

[0321] Reserve 80% of the capacity for PRB#3 and PRB#4;

[0322] Reserve +2dB × 0.8 = +1.6dB power margin;

[0323] Reserve 40% × 0.8 = 32% of computing power resources;

[0324] Attach a timestamp t+8 to the reserved resources.

[0325] E.3 Activation of Reserved Resources

[0326] After 800ms, the 1008th cycle (t+8=1008) is reached, and the actual joint state (first 3 dimensions) is:

[0327] z(1008)[:3]=[0.000014, 0.000259, 0.000033];

[0328] Predicted status:

[0329]

[0330] Calculate the Mahalanobis distance (using the first 3 dimensions approximation, magnification factor 64 / 3≈21.33):

[0331] Deviation vector: [-0.000001, -0.000009, -0.000001];

[0332] Predict the inverse of the covariance matrix (the reciprocal of the diagonal elements):

[0333]

[0334] Preset deviation threshold:

[0335]

[0336] Analysis: The Mahalanobis distance is much smaller than the threshold, and the actual state is highly consistent with the predicted state. Reserved resources are directly activated, saving about 2.8ms of re-decision time.

[0337] VII. Step F: Differential Closed-Loop Feedback

[0338] F.1 Actual QoS Acquisition and Differential Residual Calculation

[0339] After the t=1000th cycle (100ms) is completed, the actual QoS metrics are collected:

[0340] Latency:

[0341] User A (eMBB): 12.3ms (affected by high load);

[0342] User B (URLLC): 0.8ms (guaranteed by PRB#3);

[0343] User C (mMTC): 45.6ms (low priority queuing);

[0344] System average: D actual =(12.3+0.8+45.6) / 3=19.6ms;

[0345] Throughput:

[0346] User A: 185Mbps, User B: 15Mbps, User C: 0.5Mbps;

[0347] Total system throughput: T actual =200.5Mbps;

[0348] Packet loss rate:

[0349] User A: 0.3%, User B: 0%, User C: 1.2%;

[0350] System average: L actual =(0.003+0+0.012) / 3=0.005;

[0351] Actual QoS vector: q actual =[19.6, 200.5, 0.005];

[0352] Predicted QoS vector: q predict =[8.5, 245, 0.008];

[0353] Normalization (latency divided by 30ms, throughput divided by 300Mbps, packet loss rate not normalized):

[0354]

[0355] Differential residuals:

[0356]

[0357] Analysis: δ(t)=0.3986 mainly comes from the delay prediction bias (actual 19.6ms vs predicted 8.5ms), indicating that the perception model systematically underestimates the queuing delay under the current load.

[0358] F.2 Online Fine-tuning of the Perception Model

[0359] W in the cross-modal attention fusion module Q phy Taking (1, 1) (original value 0.2) as an example, let the gradient of this parameter after a complete backpropagation chain calculation be: ∂L online / ∂w 11 =0.0235;

[0360] Online learning rate η online =1×10 -4 Parameter update:

[0361]

[0362] Analysis: The gradient is positive (0.0235), meaning that increasing this parameter would increase the loss, therefore it needs to be decreased. The update amount is only 2.35 × 10. -6 This meets the constraints of single-step online fine-tuning—the parameter changes are minimal, avoiding catastrophic amnesia.

[0363] Similarly, all learnable parameters (the kernel weights and biases of the temporal convolutional encoder, the linear transformation matrix and attention vector of the graph attention network encoder, all projection matrices and biases of the Transformer encoder, and all projection matrices and biases of the cross-modal attention fusion module) are slightly adjusted according to their respective gradient directions.

[0364] F.3 Reward Remodeling

[0365] Original Instant Rewards t It consists of four parts (with weights w1=0.4, w2=0.3, w3=0.2, w4=0.1):

[0366] Positive reward for throughput: 0.4 × (200.5 / 300) = 0.4 × 0.6683 = 0.2673;

[0367] Delay penalty: -0.3 × (19.6 / 30) = -0.3 × 0.6533 = -0.1960;

[0368] Packet loss penalty: -0.2 × 0.005 = -0.0010;

[0369] Energy consumption penalty (let E) consumed / E max =0.65): -0.1×0.65=-0.0650;

[0370] r t =0.2673-0.1960-0.0010-0.0650=0.0053;

[0371] Analysis: Original reward r t =0.0053 is close to zero, indicating that although the system guarantees URLLC services (latency 0.8ms), the latency of eMBB and mMTC services is relatively high, and the overall performance is only "passable".

[0372] Modified instant reward (penalty coefficient β=0.05):

[0373] r t =0.0053 - 0.05 × 0.3986 = 0.0053 - 0.01993 = -0.01463;

[0374] Analysis: Due to the large residual (0.3986), the corrected reward becomes negative (-0.0146). This means that the current decision sequence will be judged as "bad experience" by the PPO algorithm, and the probability of this action sequence being selected again in subsequent policy gradient updates will decrease. The system will tend to learn policies that make predictions more accurate and execution more stable—for example, making more conservative estimates of latency under high load, or reserving more resources for eMBB users to avoid excessively long queues.

[0375] The revised reward r t The expression -0.0146, along with the state z(t), the dynamic state a(t), and the next state z(t+1), forms an empirical tuple (z(t), a(t), r). t ′,z(t+1)),stored in the PPO experience replay buffer,wait for enough samples to accumulate before executing the policy network update.

[0376] VIII. Summary of the complete numerical process for this period

[0377] Table 8-1: Summary of the complete numerical flow for the decision period at t=1000

[0378]

[0379] Table 8-2: Decomposition of QoS Guarantee Effects for Three Types of Services

[0380]

[0381] Analysis: The URLLC service (User B) received exclusive allocation of PRB#3 and a +1dB power boost, reducing latency to 0.8ms, perfectly meeting the 1ms budget requirement defined by 3GPP. The eMBB service (User A), while having a slightly higher latency (12.3ms), remained within acceptable limits. The mMTC service (User C) had a latency of 45.6ms, far below the 100ms budget. Overall, the system identified User B's high-priority URLLC requirement through cross-modal fusion and made reasonable resource allocation decisions under moderate uncertainty using an entropy-aware decision network.

[0382] IX. Numerical comparison with the comparison method within the same period

[0383] To further verify the effectiveness of the method of the present invention, the same input data at the same time t=1000 was fed into three comparison methods respectively, and the following comparison results were obtained:

[0384] Table 9-1: Comparison of decision results of four methods in the same period

[0385]

[0386] Table 9-2: Comparative Analysis of Decision-Making Basis for Four Methods

[0387]

[0388] Analysis: Method A (PF) assigns PRB#5 to an eMBB user (due to its highest SINR), forcing URLLC user B to wait in the queue, resulting in a latency of 2.8ms, which does not meet the 1ms budget. Method B (S-DRL), while possessing some intelligence, lacks business semantic awareness and schedules the URLLC service to PRB#2, still resulting in a latency of 1.6ms, which also does not meet the budget. Method C (Weighted Fusion) uses fixed weights. Although it considers business priority, the fixed weights cannot dynamically adapt to changes in the scenario, and the URLLC latency of 1.3ms still exceeds the budget.

[0389] Only the method of this invention (method D) uses a cross-modal attention mechanism to dynamically and weight the decreasing utilization trend of physical layer PRB#3 with the high priority of business layer URLLC, thereby allocating PRB#3 to user B. Combined with a +1dB power boost, this reduces URLLC latency to 0.8ms, perfectly meeting the 1ms budget requirement. This numerical comparison clearly demonstrates the substantial progress of the method of this invention compared to existing technologies.

[0390] 10. Conclusion of this embodiment

[0391] This embodiment demonstrates in detail the entire process of an intelligent optimization and allocation method for communication resources through a complete single-cycle numerical simulation, and draws the following conclusions:

[0392] 1. Steps B and C successfully fused three types of heterogeneous data—physical layer (PRB#3 utilization decline trend), network layer (Cell-003 high load rate 1.2251), and service layer (URLLC priority 1.00)—into a joint state representation vector z(t). The second dimension (0.000223) represents the load pressure feature, which is dominant and accurately reflects the current network state.

[0393] 2. Information entropy H(t) = 0.68 triggered the network to expand from 3 layers with 256 neurons to 4 layers with 315 neurons, increasing the capacity by 64%. This provided sufficient decision accuracy without significantly increasing inference latency, and successfully allocated PRB#3 to the URLLC service.

[0394] 3. The variational prediction network successfully predicted that the second dimension of the state would increase from 0.000223 to 0.000268 after 800ms (increased load pressure). The confidence level of 0.754 exceeded the threshold to trigger resource reservation. The actual arrival time was only 0.001 Mahalanobis distance, and the reserved resources were successfully activated.

[0395] 4. The differential residual δ(t) = 0.3986 reveals the underestimation of queuing delay by the perception model. Online fine-tuning allows for a small correction of the parameters (Δw = -2.35 × 10⁻⁶). -6 By reshaping the reward, the corrected reward r is made t =-0.0146 guides the strategy towards more accurate time delay estimation.

[0396] 5. Under the same scenario and the same input data, the method of this invention is the only one that enables the URLLC service latency (0.8ms) to meet the 1ms budget requirement, while the proportional fairness method (2.8ms), the single-source DRL method (1.6ms), and the weighted fusion method (1.3ms) all fail to meet the target. This numerical comparison fully demonstrates that the method of this invention has significant and unexpected technical effects.

[0397] To further verify the beneficial effects of the present invention, the following verification experiments were conducted:

[0398] I. Experimental Objective

[0399] By constructing a simulation platform that is highly consistent with the real 5G network environment, the method of the present invention is quantitatively compared with three existing technical methods from seven dimensions: system average throughput, average end-to-end latency, URLLC service extreme percentile latency, average resource utilization, service interruption probability, average single decision latency, and energy efficiency, to verify the superiority of the method of the present invention in various performance indicators.

[0400] II. Test Environment and Parameter Configuration

[0401] 2.1 Simulation Platform

[0402] A hybrid simulation architecture combining MATLAB 5G Toolbox and Python PyTorch is employed. MATLAB handles the 5G physical layer link-level simulation, including 3GPP TR 38.901 channel modeling, resource block scheduling, power control, and link adaptation. Python is responsible for training and inference of the deep learning model, comprising a trimodal feature extraction network, a cross-modal attention fusion module, an entropy-aware dynamic decision-making network, and a variational temporal prediction network. Both communicate in real-time via TCP sockets, exchanging state data and decision commands every 100ms.

[0403] 2.2 Network Scenario Parameters

[0404]

[0405] 2.3 Configuration of Comparison Method

[0406]

[0407] 2.4 Key parameters of the method of the present invention

[0408]

[0409] III. Experimental Group Design

[0410]

[0411] IV. Experimental Results and Data Analysis

[0412] 4.1 Experiment 1: Overall Performance Comparison

[0413] Table 1: Comparison of the overall performance of the four methods in a standard scenario (run for 24 hours, repeated 5 times, mean ± standard deviation)

[0414]

[0415] Data Analysis:

[0416] (1) In terms of throughput: The average throughput of the system using the method of this invention is 261.3 Mbps, which is 14.7% higher than the optimal weighted fusion method (227.9 Mbps). Among them, the throughput of URLLC is improved by 35.5% (from 3.1 Mbps to 4.2 Mbps), and the throughput of mMTC is improved by 38.1% (from 42.3 Mbps to 58.4 Mbps). This is because cross-modal attention fusion enables the system to accurately identify the urgent needs of URLLC services and prioritize them when resources are scarce. At the same time, it reserves resources for mMTC services in advance through forward pre-scheduling, avoiding them from waiting indefinitely in the queue and causing timeouts and retransmissions.

[0417] (2) In terms of latency: The URLLC method of the present invention has an average latency of 0.7ms, which is far lower than the 1ms budget; its 99.999th percentile latency is 1.2ms, while the weighted fusion method is 2.5ms. This is due to the entropy-aware dynamic decision network expanding network capacity under high entropy conditions to handle emergency scheduling with high precision, and the forward-looking pre-scheduling locking of resources in advance for future high-load conditions.

[0418] (3) Regarding resource utilization: the method of this invention achieves 94.2%, which is much higher than the 81.5% of the weighted fusion method. This is because the weighted fusion method uses fixed weights and cannot dynamically adjust the importance of different data sources, while the cross-modal attention mechanism of this invention can be dynamically adjusted—when the load is light, more attention is allocated to service priority features, and when the load is close to saturation, attention is shifted to physical layer channel quality features to improve spectrum efficiency.

[0419] (4) Decision latency: The method of this invention has a latency of 2.8ms, which is much lower than the 8.2ms of the single-source deep reinforcement learning method (a reduction of 56.9%). This is because the entropy-aware dynamic network automatically shrinks to 1 to 2 layers during low-entropy periods, which greatly improves the inference speed.

[0420] (5) Energy efficiency: The method of this invention is 1.68 × 10⁻⁶. 6 The bit / J efficiency is 27.3% higher than that of the weighted fusion method. This is due to two reasons: first, proactive pre-scheduling avoids frequent retransmissions caused by insufficient resources; and second, the entropy-aware network has lower computational energy consumption during low-entropy periods.

[0421] 4.2 Experiment 2: Adaptability Verification in High Dynamic Scenes

[0422] This experiment increased the user's movement speed to a maximum of 500 km / h (high-speed rail scenario) to verify the adaptability of the method of the present invention in a highly dynamic environment.

[0423] Table 2: Performance comparison under high dynamic scenarios (500km / h, 12 hours of simulation)

[0424]

[0425] Data analysis: In a 500km / h high-speed rail scenario, the average information entropy rises to 0.82. The entropy-aware mechanism of this invention automatically expands the decision network to an average of 4.6 layers and 368 neurons. Compared to a fixed 4-layer network variant, throughput is increased by 19.4% (from 168.5Mbps to 201.2Mbps), and the outage probability decreases from 3.5% to 0.8%. The single-source deep reinforcement learning method with a fixed 3-layer network has an outage probability as high as 5.8%, which can no longer meet the quality of service requirements. Meanwhile, the average decision latency of this invention, 4.1ms, is significantly lower than the 10.5ms of the fixed 4-layer network because the network can automatically shrink during low-entropy periods.

[0426] 4.3 Experiment 3: Robustness Verification of Sudden Business Impact

[0427] In this experiment, 1000 bursty URLLC service requests were injected at t=6h to observe the change in the service interruption probability over time.

[0428] Table 3: Change in the probability of business interruption over time after a sudden business shock (unit: %)

[0429]

[0430] Table 4: Comparison of resource reservation at the moment of impact

[0431]

[0432] Data Analysis: Approximately 90 seconds before the impact, the variational timing prediction network detected the abnormally shortened arrival interval of URLLC services. Once the confidence level exceeded the threshold, the system began to preemptively lock PRBs as backup resources. Within 0-5 minutes after the impact, the outage probability of the proposed method only slightly increased from 0.28% to 0.45%, while the proportional fairness method surged from 3.1% to 8.9%. This is because the proportional fairness method lacks a priority mechanism, and bursty URLLC services compete with eMBB services in the same queue. While single-source deep reinforcement learning and weighted fusion methods have some priority differentiation capabilities, they lack foresight. During the time between the impact and the next 100ms decision cycle, URLLC services cannot be guaranteed—for services with a latency budget of only 1ms, 100ms means approximately 100 packets have timed out. The proposed method recovered to near pre-impact levels (0.31%) within 10-20 minutes after the impact, thanks to differential closed-loop feedback, which enabled the system to adapt to the burst traffic pattern within approximately 15 minutes.

[0433] 4.4 Experiment 4: Validation of the effect of prospective pre-scheduling

[0434] This experiment compares the complete method of the present invention with a variant that does not include step E to quantify the contribution of forward-looking pre-scheduling.

[0435] Table 5: Performance Comparison with and without Forward-Looking Pre-Scheduling

[0436]

[0437] Table 6: Decomposition of the effect of pre-scheduling under different forecast lead times

[0438]

[0439] Data analysis: Without step E, the 99.999th percentile latency of URLLC is 1.9ms (exceeding the 1ms budget), which decreases to 1.2ms with step E (within budget). The handover success rate is 98.2% without pre-scheduling, increasing to 99.7% with pre-scheduling. Table 6 shows that approximately 42% of pre-scheduling occurs 7–10 cycles in advance (700ms–1s), precisely covering the typical 5G handover preparation time (approximately 500–800ms), enabling the system to reserve resources for the target cell before handover is triggered. The reserved resource activation rate is 87.5%, and the release rate is 12.5%. The main reason for the release is that the randomness of user movement trajectories prevents handover from occurring.

[0440] 4.5 Experiment 5: Long-term self-optimization stability verification

[0441] This experiment was run continuously for 72 hours to observe the changes in the performance indicators of the method of the present invention over time, and to verify whether the closed-loop feedback in step F can prevent model drift.

[0442] Table 7: Changes in performance indicators over time during long-term operation

[0443]

[0444] Table 8: Statistics on model parameter changes during 72 hours of operation

[0445]

[0446] Data Analysis: During 72 hours of continuous operation, the system throughput remained stable between 258 and 263 Mbps (fluctuation <2%), with an average latency of 7.9–8.3 ms and an outage probability of 0.27%–0.32%, showing no significant performance degradation. The mean differential residual value slowly decreased from 0.385 in the first hour to 0.331 in the 72nd hour (a decrease of 14.0%), indicating a continuous improvement in prediction accuracy. No model rollback was triggered during the 72 hours, and the parameter cosine similarity was 0.9987, proving that the small learning rate and single-step update strategy effectively suppressed catastrophic forgetting.

[0447] 4.6 Ablation Experiment: Decomposition of Contributions of Each Core Module

[0448] To quantify the independent contribution of each core module, ablation experiments were conducted—one module was shut down at a time and replaced with the simplest alternative.

[0449] Table 9: Ablation Experiment Results of Each Core Module

[0450]

[0451] Data Analysis:

[0452] (1) Cross-modal attention contributes the most: after removing it, the throughput decreased by 7.3% and the interruption probability increased by 93.3%, indicating that the cross-modal attention mechanism is the core contributing module of this invention.

[0453] (2) Proactive pre-scheduling is the next best: after removing step E, the probability of interruption increased by 106.7% (from 0.30% to 0.62%), indicating that step E is the most critical to ensuring business continuity.

[0454] (3) Entropy-aware dynamic network: Removing it has the smallest impact on throughput (-1.5%), but an impact on interruption probability of +16.7%, and the impact is significantly amplified in high dynamic scenarios (Experiment 2).

[0455] (4) Closed-loop feedback: After removing it, the interruption probability increased by 36.7% (to 0.41%), and it can be seen from Experiment 5 that long-term operation without closed-loop feedback will result in slow degradation.

[0456] 4.7 Statistical significance test

[0457] A two-sample t-test was performed on the results of all five replicate experiments.

[0458] Table 10: Results of statistical significance test (significance level α = 0.05)

[0459]

[0460] All p-values ​​are much less than 0.05, proving that the performance advantage of the method of the present invention is statistically significant and not due to random fluctuations.

[0461] V. Experimental Conclusions

[0462] Based on the above five sets of comparative experiments and ablation experiments, the following conclusions can be drawn:

[0463] First, in standard scenarios, compared to the optimal existing method (weighted fusion method), the method of this invention achieves a 14.7% increase in system throughput (261.3 vs. 227.9 Mbps), a 52.0% reduction in URLLC 99.999% percentile latency (1.2 vs. 2.5 ms), a 15.6% increase in resource utilization (94.2% vs. 81.5%), an 80.0% reduction in service interruption probability (0.3% vs. 1.5%), and a 27.3% improvement in energy efficiency (1.68 × 10⁻⁶). 6 vs. 1.32×10 6 (bit / J). All performance advantages passed the statistical significance test.

[0464] Second, in the scenario of a 500km / h high-speed rail, the method of this invention automatically expands to an average of 4.6 layers and 368 neurons through an entropy-aware dynamic network, maintaining a low interruption probability of 0.8%, while the interruption probability of the fixed network method is as high as 5.8%. The dynamic network reduces the average decision latency by 61.0% (4.1 vs. 10.5ms) compared to the fixed large network.

[0465] Third, when faced with a sudden surge of 1,000 URLLC services, the interruption probability of the method described in this invention only briefly increased from 0.28% to 0.45% (recovering within 5 minutes after the surge), while the interruption probability of the proportional fairness method soared to 8.9%. Proactive pre-scheduling begins reserving resources approximately 90 seconds before the surge, with a reserved resource activation rate of 87.5%.

[0466] Fourth, during 72 hours of continuous operation, the system performance remained stable, with no performance degradation or catastrophic forgetting events. The mean difference residual decreased slowly from 0.385 to 0.331 (a decrease of 14.0%), and the cosine similarity of the model parameters after 72 hours was 0.9987, proving that the online fine-tuning amplitude was effectively controlled and the model continued to optimize.

[0467] Fifth, ablation experiments show that cross-modal attention mechanism contributes the most (throughput decreases by 7.3%, interruption probability increases by 93.3%), followed by forward pre-scheduling (interruption probability increases by 106.7%). Entropy-aware dynamic networks are crucial for high-dynamic scenarios, and closed-loop feedback is essential for long-term stable operation.

[0468] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent optimization and allocation of communication resources, characterized in that, Includes the following steps: Data Acquisition and Preprocessing: Acquire reference signal received power, signal-to-interference-plus-noise ratio, channel quality indicator, resource block utilization, and block error rate at the physical layer; acquire node load rate, backhaul link utilization, handover frequency, number of associated users, and user spatial distribution density factor at the network layer, and construct an adjacency matrix; acquire service type, average packet size, arrival interval, and priority level at the service layer. Hampel filtering and linear interpolation are performed on the physical layer data for noise reduction and missing data. Z-score normalization and adjacency matrix symmetric normalization are performed on the network layer data. One-hot encoding, Z-score normalization and Min-Max normalization are performed on the service layer data. Synchronous data frames are generated by aligning with timestamps. Multimodal feature extraction: Physical layer data from multiple recent time slots are stacked and then subjected to one-dimensional convolution and global average pooling to obtain a spectral evolution feature vector; the normalized adjacency matrix and node features are encoded by a multi-head graph attention network to obtain a topological association feature vector; user business features are semantically encoded by Transformer to obtain a business intent feature vector. Cross-modal fusion and embedding: The three feature vectors are mapped to the same dimensional space by linear projection, and after cross attention is calculated in pairs, they are fused into a three-modal fusion vector. Then, they are embedded into a low-dimensional manifold by equidistant mapping to obtain a joint state representation vector. Adaptive decision-making: The joint state representation vector is input into a decision network containing a differentiable entropy estimator. Based on the historical samples in the sliding window, the information entropy of the current sample is calculated by Gaussian kernel density estimation. The number of network layers and neurons is dynamically adjusted accordingly, and then proximal policy optimization is performed to output resource allocation actions. Pre-decision and reservation: Input the joint state representation vector sequence into the variational autoencoder time series prediction network to generate future periodic state prediction values ​​and standard deviations, calculate the confidence level, perform virtual decision-making for predicted states that exceed the threshold and lock reserved resources, calculate the Mahalanobis distance when the future period is reached, if it is less than the threshold, activate the reservation, otherwise release and trigger re-decision; Online feedback learning: The Euclidean distance between the actual service quality index and the predicted service quality index is collected as the difference residual, which is used to update the gradient of the feature extraction and fusion network parameters, as well as to correct the reward for the near-end policy optimization.

2. The intelligent optimization and allocation method for communication resources according to claim 1, characterized in that, The Hampel filtering denoising method is as follows: taking a preset number of time slots before and after the current time as the center to form a window, calculating the median and the corrected absolute median difference; if the absolute difference between the current data and the median exceeds three times the corrected absolute median difference, it is replaced with the median; the adjacency matrix symmetric normalization method is as follows: after adding an identity matrix to the adjacency matrix, multiply it by the left and right sides by the negative 1 / 2 power of the degree matrix.

3. The intelligent optimization and allocation method for communication resources according to claim 2, characterized in that, The single-layer operation of the graph attention network encoding is as follows: the learnable linear transformation matrix of each attention head is applied to the node features; the transformed features of the center node and the neighboring nodes are concatenated and then multiplied by the learnable attention vector and activated by a linear rectified function with leakage; the activation values ​​of the neighboring nodes (including themselves) are normalized by softmax to obtain the attention coefficients and then weighted and summed; finally, the outputs of each attention head are averaged and activated by an exponential linear unit.

4. The intelligent optimization and allocation method for communication resources according to claim 3, characterized in that, The attention pooling is performed as follows: after the output features of each node are subjected to learnable linear transformation and hyperbolic tangent activation, the inner product is made with the learnable context query vector. The inner product results of all nodes are then normalized with softmax to obtain pooling weights, and the node features are then summed with weights.

5. The intelligent optimization and allocation method for communication resources according to claim 4, characterized in that, The cross-attention matrix is ​​calculated as follows: the source modality projection vector is mapped to a query vector through a learnable query projection matrix, the target modality projection vector is mapped to a key vector through a learnable key projection matrix, the dot product of the query vector and the transpose of the key vector is calculated and divided by a scaling factor, and then softmax normalization is performed on each row of the matrix.

6. The intelligent optimization and allocation method for communication resources according to claim 5, characterized in that, The isometric mapping is as follows: construct a K nearest neighbor graph in the cache pool and calculate the geodesic distance between sample point pairs, and perform multidimensional scaling to obtain low-dimensional embedding coordinates; for new samples, only calculate the geodesic distance between them and a preset number of nearest neighbors in the cache pool, and use landmark multidimensional scaling to solve for low-dimensional embedding coordinates.

7. The intelligent optimization and allocation method for communication resources according to claim 6, characterized in that, The dynamically adjusted network depth is: the current number of layers equals the base number of layers plus the adjustable amplitude multiplied by the hyperbolic tangent function value scaled by the difference between the information entropy value and the medium entropy value; the dynamically adjusted network width is: the current number of neurons per layer equals the base number plus the adjustable amplitude multiplied by the difference between the information entropy value and the high entropy threshold mapped by the Logistic Sigmoid function; where the medium entropy value is half the sum of the high entropy threshold and the low entropy threshold.

8. The intelligent optimization and allocation method for communication resources according to claim 7, characterized in that, The confidence score is 1 divided by 1 plus the Euclidean norm of the predicted standard deviation vector; the Mahalanobis distance is the square root of the transpose of the difference between the actual state vector and the predicted state vector, multiplied by the inverse of the predicted covariance matrix, and then multiplied by the difference vector; the predicted covariance matrix is ​​a diagonal matrix with the squares of each component of the standard deviation vector as its diagonal elements; each time the reserved resources are released, the release event is recorded as a negative sample for online fine-tuning of the variational autoencoder temporal prediction network.

9. The intelligent optimization and allocation method for communication resources according to claim 8, characterized in that, The online gradient update saves a full model checkpoint after each preset number of executions. If the mean of the difference residuals increases by more than a preset percentage after multiple consecutive updates, it rolls back to the most recent checkpoint.

Images