5G network slice resource dynamic allocation method and system based on artificial intelligence

Abstract

The invention relates to a 5G network slice resource dynamic allocation method and system based on artificial intelligence. The method comprises the following steps: acquiring original data containing sub-band interference and beam error of a millimeter wave channel and preprocessing the original data; encoding the preprocessed data into channel hidden variables jointly representing interference statistical characteristics and an error range through a variational auto-encoder; generating an uncertainty score based on the reconstruction error and the hidden variable variance to quantify the channel model confidence; inputting the hidden variables and the network state into a deep reinforcement learning model to generate a resource allocation decision; and dynamically adjusting action space constraints according to the uncertainty scores, and adaptively switching between a conservative mode and a normal mode. By adopting the method, the problem of resource allocation failure caused by neglecting sub-band level interference and beam calibration errors in the prior art can be solved, the service quality of network slices can be ensured under an imperfect channel condition, and the robustness and reliability of a millimeter wave communication system are improved.

Description

Technical Field

[0001] This invention belongs to the field of data communication technology, and in particular relates to a method and system for dynamic allocation of 5G network slice resources based on artificial intelligence. Background Technology

[0002] In 5G and future 6G networks, network slicing technology is crucial for achieving multi-service coexistence. It allows the creation of isolated virtual networks on shared physical infrastructure for services with different needs, such as enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), and massive machine-type communication (mMTC). Millimeter wave (mmWave) and massive multiple-input multiple-output (MIMO) technologies, by utilizing high-frequency spectrum and large-scale antenna arrays, provide extremely high bandwidth and spectral efficiency, becoming core to achieving slicing performance goals. However, the performance of these technologies is highly dependent on precise beamforming and interference management. Existing AI-based resource allocation methods typically assume that channel state information (CSI) is ideal or that the noise is Gaussian white noise, ignoring non-ideal factors in real-world networks. Specifically, millimeter wave channels suffer from two major overlooked problems: first, sub-band level interference, where the intensity of interference in different sub-bands exhibits high randomness rather than uniform distribution due to frequency-selective fading; and second, beam calibration errors, stemming from physical defects in the antenna array (such as phase shift) and environmental changes (such as temperature fluctuations), leading to quasi-static deviations in beam pointing. These non-ideal factors, when coupled together, significantly reduce the effectiveness of AI resource allocation decisions (such as subband allocation and beam selection) based on inaccurate CSI, and may even cause slice service interruptions, seriously affecting network reliability and quality of service (QoS).

[0003] Traditional methods typically employ simplified channel models or optimization algorithms based on perfect CSI, failing to effectively capture the randomness of subband interference and the quasi-static characteristics of beam calibration errors. For example, reinforcement learning models directly use the original CSI as input, but due to CSI noise and errors, the model is prone to overfitting to inaccurate observations, causing allocation strategies to fail in real-world environments. Furthermore, existing methods lack explicit quantification of channel uncertainties, making it impossible to dynamically adjust strategies when channel information is partially observable or erroneous, thus struggling to achieve a balance between performance and reliability. Therefore, an innovative method is urgently needed that can jointly characterize non-ideal channel characteristics and achieve robust resource allocation, directly mapping non-ideal factors at the physical layer to the network layer decision-making process, thereby enhancing the adaptability and resilience of 5G network slicing. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and system for dynamic allocation of 5G network slicing resources based on artificial intelligence to address the above-mentioned technical issues.

[0005] Firstly, this application provides a method for dynamic allocation of 5G network slicing resources based on artificial intelligence, including:

[0006] S1. Acquire the raw channel data of the millimeter-wave channel and preprocess the raw channel data to obtain the preprocessed channel data; the raw channel data includes sub-band channel state information and beam calibration error data.

[0007] S2. Input the preprocessed channel data into the variational autoencoder model and generate channel latent variables through the encoder network.

[0008] S3. Based on the decoder output of the channel latent variables and variational autoencoder model, calculate the reconstruction error, and combine it with the variance of the latent variables of the encoder network output to generate an uncertainty score.

[0009] S4. Input the channel latent variables, the current network state, and the slice requirements into the deep reinforcement learning model, and output the first resource allocation decision through the policy network;

[0010] S5. When the uncertainty score exceeds a preset threshold, adjust the action space constraints of the deep reinforcement learning model to obtain the first resource allocation method instruction; the first resource allocation method instruction is used to execute the first resource allocation decision in a conservative mode.

[0011] S6. After executing the first resource allocation method instruction, monitor the network performance indicators in real time. When the network performance indicators are stable and the uncertainty score is lower than the preset threshold, restore the original action space constraints of the deep reinforcement learning model to obtain the second resource allocation decision.

[0012] S7. Based on the second resource allocation decision, generate a second resource allocation mode instruction; the second resource allocation mode instruction is used to switch to normal mode to execute the second resource allocation decision.

[0013] In one embodiment, the preprocessed channel data is input into a variational autoencoder model, and channel latent variables are generated through the encoder network, including:

[0014] S11. Input the preprocessed channel data into the encoder network of the pre-trained variational autoencoder model, and calculate the mean vector and variance vector of the latent variables through forward propagation.

[0015] S12. Based on the mean vector and variance vector, channel latent variables are generated through reparameterized sampling; the reparameterization formula is:

[0016]

[0017] in, This is the channel latent vector, used to characterize the probability distribution features of the original channel data in the latent space. It is the mean vector. It is the variance vector. A random noise vector sampled from a standard normal distribution. This represents element-wise multiplication.

[0018] In one embodiment, the reconstruction error is calculated based on the decoder output of the channel latent variables and the variational autoencoder model, and an uncertainty score is generated by combining the variance of the latent variables from the encoder network output, including:

[0019] S21. Input the channel latent variables into the decoder network of the variational autoencoder model, and obtain the reconstructed channel data through forward propagation;

[0020] S22. Based on the original channel data and the reconstructed channel data, the reconstruction error is obtained using the following formula:

[0021]

[0022] in, For reconstruction error, The raw channel data, To reconstruct channel data;

[0023] S23. Extract the variance values ​​of each dimension from the variance vector output by the encoder network, and obtain the average variance using the following formula:

[0024]

[0025] in, The mean variance is a measure used to characterize the uncertainty of the distribution of latent variables. For the dimension of latent variables, For the first Variance;

[0026] S24. The reconstruction error and the average variance are weighted and combined to generate an uncertainty score. The expression for the uncertainty score is:

[0027]

[0028] in, For uncertain fractions, The preset tradeoff parameters are verified using historical data to balance the contributions of reconstruction error and distribution uncertainty. For reconstruction error, The average variance is denoted as .

[0029] In one embodiment, the channel latent variables, the current network state, and the slice requirement are input into a deep reinforcement learning model, and a first resource allocation decision is output through a policy network, including:

[0030] S31. Construct the state space of the deep reinforcement learning model; the state space includes channel latent variables, slice demand vector, and resource availability vector; the slice demand vector includes the throughput threshold and latency limit of each slice; the resource availability vector includes the number of available sub-bands, the number of beams, and the power budget;

[0031] S32. Define the action space; the action space includes the subband assignment vector, the beamforming weight matrix, and the power assignment vector; the subband assignment vector is the set of subband indices assigned to each slice; the beamforming weight matrix is ​​the complex weight coefficient of each beam; the power assignment vector is the power level of each subband;

[0032] S33. Train a deep reinforcement learning model using the proximal policy optimization algorithm. The policy network outputs the action probability distribution and the value network evaluates the state value function to obtain the trained deep reinforcement learning model. Both the policy network and the value network are implemented by neural networks.

[0033] S34. Based on the trained deep reinforcement learning model, the current state is input into the policy network, and the first resource allocation decision is output through the sampling action probability distribution; the first resource allocation decision is used to directly control the base station resource allocation unit.

[0034] In one embodiment, training a deep reinforcement learning model using a proximal policy optimization algorithm includes:

[0035] S41. Initialize the policy network parameters and value network parameters of the deep reinforcement learning model, and construct the experience replay buffer; the experience replay buffer is used to store state transition data;

[0036] S42. Execute the current policy in a simulated network environment, collect state-action-reward sequence data and store it in the experience replay buffer; the action includes the first resource allocation decision; the reward is calculated based on the service quality satisfaction; the state includes channel latent variables, network state and slice requirements.

[0037] S43. Sample batch data from the experience replay buffer to obtain the dominance function estimate for each time step; the expression for the dominance function estimate is:

[0038]

[0039] in, This is the estimated value of the dominance function. For timing difference error, As a discount factor, The attenuation factor for generalized dominance estimation. This is the time step at which the trajectory terminates;

[0040] S44. Update the policy network parameters by maximizing the pruning objective function; the expression for the pruning objective function is:

[0041]

[0042] in, For the purpose of cropping, For policy network parameters, This is the estimated value of the dominance function. The probability ratio between the old and new strategies. These are pruning parameters used to limit the magnitude of policy updates;

[0043] S45. By minimizing the value function error, the value network parameters are updated. The expression for the value function error is:

[0044]

[0045] in, For value function error, For network parameters, This indicates the cumulative discount return. The state values ​​predicted by the value network;

[0046] S46. Repeat steps S42 to S45 until the policy network and value network converge, obtaining a trained deep reinforcement learning model.

[0047] In one embodiment, when the uncertainty score exceeds a preset threshold, the action space constraints of the deep reinforcement learning model are adjusted to obtain a first resource allocation method instruction, including:

[0048] S51. Compare the uncertainty score with a preset threshold to obtain the comparison result; the preset threshold is set by the quantile of the historical uncertainty score distribution to distinguish between normal and conservative modes.

[0049] S52. When the comparison result shows that the uncertainty score exceeds the preset threshold, modify the action space constraints to obtain the modified action space constraints. The action space constraints include setting the minimum number of sub-band allocations, the maximum angle of beamwidth, and the upper limit of power allocation.

[0050] S53. Based on the modified action space constraints, a first resource allocation decision is generated through the policy network of a deep reinforcement learning model, and a first resource allocation method instruction is generated based on the first resource allocation decision.

[0051] Secondly, this application also provides a dynamic allocation system for 5G network slicing resources based on artificial intelligence, including:

[0052] The data acquisition and preprocessing module is used to acquire the raw channel data of the millimeter-wave channel and preprocess the raw channel data to obtain the preprocessed channel data; the raw channel data includes sub-band channel state information and beam calibration error data.

[0053] The channel data processing module is used to input the preprocessed channel data into the variational autoencoder model and generate channel latent variables through the encoder network.

[0054] The uncertainty score module is used to calculate the reconstruction error based on the decoder output of the channel latent variables and variational autoencoder model, and to generate the uncertainty score by combining the variance of the latent variables of the encoder network output.

[0055] The first resource allocation decision module is used to input the channel latent variables, the current network state and the slice requirements into the deep reinforcement learning model, and output the first resource allocation decision through the policy network.

[0056] The first resource allocation method module is used to adjust the action space constraints of the deep reinforcement learning model when the uncertainty score exceeds a preset threshold, and obtain the first resource allocation method instruction; the first resource allocation method instruction is used to execute the first resource allocation decision in a conservative mode.

[0057] The second resource allocation decision module is used to monitor network performance indicators in real time after executing the first resource allocation method instruction. When the network performance indicators are stable and the uncertainty score is lower than the preset threshold, the original action space constraints of the deep reinforcement learning model are restored to obtain the second resource allocation decision.

[0058] The second resource allocation method module is used to generate a second resource allocation method instruction based on the second resource allocation decision; the second resource allocation method instruction is used to switch to normal mode to execute the second resource allocation decision.

[0059] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0060] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0061] The aforementioned AI-based dynamic allocation method and system for 5G network slicing resources acquires and preprocesses raw millimeter-wave channel data containing sub-band channel state information and beam calibration error data. This data is then input into a variational autoencoder to generate channel latent variables. The reconstruction error and uncertainty score are calculated by combining the decoder output with the variance of these latent variables. Subsequently, the channel latent variables, network state, and slicing requirements are input into a deep reinforcement learning model. The policy network outputs a first resource allocation decision containing sub-band allocation information. When the uncertainty score exceeds a threshold, the model's action space constraints are adjusted to generate a conservative mode instruction. After execution, network performance is monitored. Once the indicators stabilize and the score meets the threshold, the original constraints are restored to obtain a second decision and generate a normal mode instruction. These methods accurately capture non-ideal channel characteristics and quantify uncertainty, dynamically adapting the decision-making mode to achieve dynamic and robust allocation of slicing resources. This improves decision reliability in non-ideal scenarios and ensures network service quality and stability. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a flowchart illustrating a method for dynamically allocating 5G network slicing resources based on artificial intelligence in one embodiment.

[0064] Figure 2 This is a schematic diagram of the structure of an AI-based 5G network slice resource dynamic allocation system in one embodiment. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0066] refer to Figure 1 The document presents a flowchart illustrating the AI-based dynamic allocation method for 5G network slicing resources provided in this application. The method includes the following steps:

[0067] S1. Obtain the raw channel data of the millimeter-wave channel and preprocess the raw channel data to obtain the preprocessed channel data.

[0068] Specifically, acquiring raw channel data for millimeter-wave channels requires collaboration between a dedicated data acquisition module deployed on the base station side and the feedback mechanism of the user equipment. Subband channel state information is mainly acquired by the large-scale MIMO antenna array of the base station according to a set acquisition period. The specific acquisition process must follow the channel state information reference signal configuration specifications defined in the relevant protocols. The millimeter-wave frequency band is divided into continuous subbands, and the channel amplitude, phase, and delay information of each subband are sampled through the beam scanning mechanism of the antenna array. The sampled data is stored in the form of a complex matrix, and the matrix dimension is related to the number of antenna ports and the number of subbands.

[0069] To obtain beam calibration error data, a high-precision phase noise analyzer and a temperature sensor need to be integrated into the base station antenna array. The phase noise analyzer detects the phase offset of each antenna element in real time, and the temperature sensor collects temperature data at set intervals. The temperature fluctuation data and the phase offset data are timestamped to construct a beam calibration error dataset, which contains three dimensions: phase offset value, temperature change, and corresponding time information.

[0070] The preprocessing of the raw channel data needs to be performed in two steps. First, outlier cleaning is performed. For burst noise in the sub-band channel state information, an anomaly detection method based on the 3σ criterion is used to calculate the mean and standard deviation of the amplitude of each sub-band channel. Values ​​that exceed a certain range are identified as outliers and replaced using data from adjacent sampling points through linear interpolation. For outliers in the beam calibration error data, a sliding window median filtering method is used. The window size is set to a number of sampling points, and the outlier data in the center of the window is replaced with the median within the window.

[0071] Subsequently, data standardization processing is performed, converting the complex matrix of subband channel state information into two real matrices: amplitude and phase. The amplitude matrix is ​​normalized using a min-max normalization method, mapping the values ​​to a specific interval. The phase matrix is ​​normalized, transforming the phase values ​​to a specific interval. Simultaneously, the phase offset in the beam calibration error data is converted into a percentage deviation relative to the ideal phase, and the temperature data is converted into a temperature difference value relative to a reference temperature. Finally, preprocessed channel data with unified dimensions is formed, in tensor form. The dimensions are related to the sampling batch, the number of antenna ports, the number of subbands, and the number of features. The number of features includes four dimensions: amplitude, phase, phase offset percentage, and temperature difference value.

[0072] S2. Input the preprocessed channel data into the variational autoencoder model and generate channel latent variables through the encoder network.

[0073] Specifically, the variational autoencoder (VAE) model consists of two parts: an encoder network and a decoder network. The encoder network uses a structure combining convolutional neural networks and fully connected layers. The input layer receives the preprocessed channel data tensor, the dimension of which is related to the number of antenna ports, sub-bands, and features. The first convolutional layer of the encoder has several convolutional kernels of a specific size. By setting the stride and padding method, and in conjunction with the activation function, it extracts the local spatial features of the channel data. The second convolutional layer also has several convolutional kernels. By adjusting the stride, it achieves feature dimensionality reduction and outputs a feature map of a specific dimension. Then, a global average pooling layer converts the feature map into a fixed-length feature vector. Next, two parallel fully connected layers are connected, serving as a mean generation layer and a variance generation layer, respectively. The output of the mean generation layer is the mean vector of the channel latent variables, and the variance generation layer ensures that the output variance vector is positive through a specific activation function. The final generated channel latent variables are obtained through reparameterization techniques. This operation not only ensures the differentiability of the model but also achieves modeling of the distribution of latent variables, enabling efficient characterization of the core features and potential distribution information of the channel data.

[0074] S3. Based on the decoder output of the channel latent variables and variational autoencoder model, calculate the reconstruction error, and combine it with the variance of the latent variables of the encoder network output to generate an uncertainty score.

[0075] Specifically, the reconstruction error is calculated based on the difference between the decoder network output and the preprocessed channel data. The decoder network employs a structure combining deconvolutional and fully connected layers. First, the fully connected layers map the channel latent variables into high-dimensional feature vectors. Then, two deconvolutional layers gradually restore the data dimension, ultimately outputting a reconstructed channel data tensor whose dimension is completely consistent with the preprocessed channel data. The reconstruction error is calculated using mean squared error, which directly reflects the latent variables' ability to represent the original channel data. The generation of the uncertainty score combines the reconstruction error with the variance of the latent variables output by the encoder. First, the variance vector of the latent variables is averaged to obtain the global variance, which quantifies the dispersion of the channel data in the latent space. A larger global variance indicates higher uncertainty in the channel data. Subsequently, a weighted fusion method is used to calculate the uncertainty score. The uncertainty score ranges within a specific interval; a higher score indicates more significant non-ideal factors in the channel data. Based on this score, a quantitative assessment of the reliability of channel information can be achieved.

[0076] S4. Input the channel latent variables, the current network state, and the slice requirements into the deep reinforcement learning model, and output the first resource allocation decision through the policy network.

[0077] Specifically, the deep reinforcement learning model employs a deep deterministic policy gradient algorithm, which is suitable for decision-making problems in continuous action spaces and can meet the optimization requirements of continuous variables such as power allocation in resource allocation. The model's state space consists of three parts: channel latent variables, current network state, and slice requirements. The current network state includes indicators such as the number of allocated subbands for each slice, current beam usage, remaining power resources, and interference levels. Slice requirements are defined according to three service types: enhanced mobile broadband, ultra-reliable low-latency communication, and massive machine-type communication. Enhanced mobile broadband service uses throughput requirements as the core indicator, ultra-reliable low-latency communication service uses latency and reliability requirements as core indicators, and massive machine-type communication service uses connection number requirements and packet size as core indicators.

[0078] The action space corresponds to the first resource allocation decision, which includes three parts: subband allocation, beamforming, and power allocation. Subband allocation is represented by a binary vector with a length consistent with the total number of subbands. A specific value indicates that the subband has been allocated, while another value indicates that it has not been allocated. Beamforming is represented by a beam weight vector with a dimension related to the number of antenna ports. Each element is a complex number, representing the weighting coefficient of the corresponding antenna element. Power allocation is a continuous variable with a value range within a specific interval, representing the transmit power allocated to each subband.

[0079] The policy network of the deep reinforcement learning model adopts an Actor-Critic dual-network structure. The Actor network is responsible for outputting specific resource allocation decisions. Its input is a state space vector, which first passes through several fully connected layers and undergoes feature processing with activation functions. Then, it outputs action vectors through the output layer. The subband allocation part uses a specific activation function and is converted into a binary vector by a threshold. The beam weight vector part uses a specific activation function to map the values ​​to a specific interval and then converts them into complex numbers. The power allocation part uses a specific activation function to ensure that the output is non-negative. The Critic network is responsible for evaluating the value of the decisions. Its input is a state-action pair. It calculates the action value function through fully connected layers, which is used to guide the parameter updates of the Actor network.

[0080] During the training process of the model, interactive experience is stored in the experience replay pool, a specific batch size and learning rate are set, the target network updates parameters using a soft update method, and the policy network parameters are continuously optimized through interaction with the network simulation environment. Finally, the first resource allocation decision that maximizes long-term cumulative rewards is output. The reward function is designed as a weighted sum that comprehensively considers throughput improvement, latency reduction and resource utilization to ensure that the decision meets the service quality requirements of the slice.

[0081] S5. When the uncertainty score exceeds the preset threshold, adjust the action space constraints of the deep reinforcement learning model to obtain the first resource allocation method instruction.

[0082] Specifically, the determination of the preset threshold needs to be based on historical channel data and network operation experience. By statistically analyzing the effectiveness of resource allocation decisions under different uncertainty scores, the score corresponding to the decision failure probability being lower than a specific value is set as the preset threshold. When the uncertainty score exceeds the preset threshold, it indicates that the non-ideal factors of the current channel data are significant, and resource allocation decisions based on this data have a high risk. At this time, it is necessary to adjust the action space constraints of the deep reinforcement learning model and enter a conservative mode.

[0083] The adjustment of the action space constraints is specifically reflected in three aspects: In terms of subband allocation constraints, each slice is limited to a maximum number of subbands, with priority given to subbands with lower historical interference levels, while avoiding concentrated allocation of adjacent subbands to prevent interference superposition between subbands; In terms of beamforming constraints, the tolerance for beam pointing deviation is reduced, and the accuracy of beam pointing is ensured by increasing the compensation coefficient for beam calibration errors. The compensation formula is as follows:

[0084]

[0085] In the above formula, This represents the compensated beam weight. Represents the original beam weights. This represents the percentage of phase deviation. Regarding power allocation constraints, the maximum transmit power is reduced, and a power balancing strategy is adopted to ensure that the power deviation of each allocated sub-band does not exceed a specific value, avoiding increased interference due to excessive power. The adjusted action space is implemented by modifying the constraints of the policy network output layer, generating a first resource allocation method instruction. This instruction contains the adjusted action space parameters and the corresponding resource allocation decision, and is sent to the resource scheduling module through the base station's network management interface. The decision is executed in a conservative mode to ensure network stability.

[0086] S6. After executing the first resource allocation method instruction, monitor the network performance indicators in real time. When the network performance indicators are stable and the uncertainty score is lower than the preset threshold, restore the original action space constraints of the deep reinforcement learning model to obtain the second resource allocation decision.

[0087] Specifically, network performance monitoring relies on the performance monitoring module on the base station side and the feedback mechanism on the user equipment side. A distributed monitoring architecture is adopted, with the base station side collecting throughput, latency and packet loss rate data of each slice in real time, and the user equipment side uploading service experience-related performance data to the base station through a dedicated feedback channel to ensure the real-time and comprehensiveness of the data.

[0088] Throughput monitoring employs a packet statistics method, calculating the throughput value by statistically analyzing the total number and size of data packets passing through each slice per unit time. The specific formula is as follows:

[0089]

[0090] In the above formula, Represents throughput, Represents the size of a single data packet. This represents the sum of the sizes of all data packets passing through the slice within the statistical time period. This represents the statistical time interval. Delay monitoring uses a timestamp comparison method: a sending timestamp is marked when a data packet is sent, and a receiving timestamp is marked at the receiving end. The difference between the two is the end-to-end delay. Abnormal delay values ​​caused by network congestion are discarded. Packet loss rate monitoring uses a sequence number counting method: sent data packets are consecutively numbered, and the number of missing sequence numbers is counted at the receiving end. The calculation formula is as follows:

[0091]

[0092] In the above formula, Represents packet loss rate. The number of missing sequence numbers represents the total number of sequence numbers corresponding to data packets that were not detected by the receiver. The total number of data packets sent represents the total number of data packets sent by the sender within the statistical period.

[0093] The determination of stable network performance indicators requires meeting two conditions: First, within a number of consecutive collection periods, the fluctuation range of each performance indicator is controlled within a preset range, with throughput, latency, and packet loss rate each having corresponding fluctuation limits; second, each performance indicator meets the service quality requirements of the corresponding slice, with different types of slices having corresponding throughput, latency, and packet loss rate requirements.

[0094] When the above conditions are met simultaneously and the uncertainty score is below a preset threshold, it indicates that the reliability of the current channel data has been restored. At this point, the original action space constraints of the deep reinforcement learning model need to be restored. Specifically, the constraint parameters for subband allocation, beamforming, and power allocation are reset to their initial settings during model training. After the constraints are restored, the policy network will re-output a second resource allocation decision based on the current state space. This decision, compared to the first resource allocation decision, focuses more on improving resource utilization and service performance, thus transitioning from a conservative mode to an efficient mode.

[0095] S7. Based on the second resource allocation decision, generate a second resource allocation mode instruction; the second resource allocation mode instruction is used to switch to normal mode to execute the second resource allocation decision.

[0096] Specifically, a second resource allocation method instruction is generated based on the second resource allocation decision. The core function of this instruction is to trigger the switching of the resource allocation mode from conservative mode to normal mode. The generation of the instruction must comply with the base station's resource scheduling protocol and adopt a structured data format, including four parts: instruction identifier, handover time, details of the second resource allocation decision, and verification parameters. The instruction identifier is a unique string used to distinguish different resource allocation instructions; the handover time is set to a specific interval after the current time to ensure that the network has enough time to complete the smooth handover of resources and avoid service interruption during the handover process; the details of the second resource allocation decision specify the sub-band allocation list, beam weight vector, and power allocation values ​​for each slice; the verification parameters are the expected performance indicators corresponding to the decision, used for data verification after the handover.

[0097] After the second resource allocation instruction is generated, it is sent to the resource scheduling module through the base station's control channel. The scheduling module first verifies the completeness and legality of the instruction. After confirming that it is correct, it executes the second resource allocation decision according to the handover time. During the handover process, a gradual resource adjustment strategy is adopted to gradually replace the original resource allocation scheme. For example, subband allocation adopts the "allocate first, release later" approach to avoid temporary gaps in subband resources; power allocation adopts a linear adjustment approach to smoothly adjust the power from the conservative mode value to the normal mode value within a specific time. At the same time, the performance monitoring module tracks the network performance indicators after the handover in real time. If any abnormal indicators are found, a rollback mechanism is triggered to restore the conservative mode, ensuring the safety and reliability of the mode handover, and ultimately realizing the dynamic and robust allocation of 5G network slice resources.

[0098] In the aforementioned AI-based dynamic resource allocation method for 5G network slicing, raw channel data containing subband interference and beam errors is collected and preprocessed. A variational autoencoder is then used to encode this data into low-dimensional channel latent variables to jointly characterize the statistical distribution of non-ideal physical layer properties. Furthermore, an uncertainty score is generated based on the reconstruction error and the variance of the latent variables to quantify the confidence level of the channel model. This latent variable, along with the network state, is input into a deep reinforcement learning model to generate initial resource allocation decisions. When uncertainty is too high, a conservative mode is triggered by adjusting action space constraints to execute robust allocation. Once performance stabilizes and uncertainty decreases, the original constraints are restored, switching back to normal mode. This achieves dynamic adaptive adjustment of the resource allocation strategy under imperfect channel information conditions, effectively improving the service quality and system robustness of 5G network slicing in the millimeter-wave band when facing random interference and quasi-static errors.

[0099] In one embodiment, the preprocessed channel data is input into a variational autoencoder model, and channel latent variables are generated through the encoder network, including:

[0100] S11. Input the preprocessed channel data into the encoder network of the pre-trained variational autoencoder model, and calculate the mean vector and variance vector of the latent variables through forward propagation.

[0101] Specifically, the encoder network adopts a hierarchical structure combining convolutional neural networks (CNNs) and fully connected layers. After receiving preprocessed channel data tensors with dimensions matching the number of antenna ports, subbands, and features in the input layer, the data first enters the convolutional layers for feature extraction. During forward propagation, the first convolutional layer performs sliding convolution operations on the input data using convolutional kernels of a specific size, and, in conjunction with an activation function, preserves the local spatial features of the channel data, generating a feature map of the corresponding dimension. The second convolutional layer achieves feature dimensionality reduction by adjusting the stride, further refining the key information of the data. Subsequently, a global average pooling layer converts the two-dimensional feature map into a fixed-length one-dimensional feature vector, eliminating the impact of data dimensionality differences on subsequent calculations. After the feature vector is fed into two parallel fully connected layers, it undergoes linear transformations to generate the mean vector and variance vector of the latent variables. The variance generation layer ensures that the output is always positive through a specific activation function, providing a foundation for subsequent probability distribution modeling.

[0102] S12. Based on the mean vector and variance vector, channel latent variables are generated through reparameterized sampling; the reparameterization formula is:

[0103]

[0104] in, This is the channel latent vector, used to characterize the probability distribution features of the original channel data in the latent space. It is the mean vector. It is the variance vector. A random noise vector sampled from a standard normal distribution. This represents element-wise multiplication.

[0105] In the above formula, As a channel latent vector, its core function is to characterize the probability distribution features of the original channel data in the latent space, compressing high-dimensional channel data containing non-ideal factors into a low-dimensional vector form that retains the core information, which is convenient for subsequent reinforcement learning models to perform decision calculations. The mean vector is generated from the mean of the encoder network to produce the output of the fully connected layer, reflecting the central distribution location of the channel latent variables in the latent space; The variance vector is generated from the variance of the encoder network to produce the output of the fully connected layer. It is used to describe the degree of dispersion of the channel latent variables around the mean vector in the latent space. The larger the variance, the higher the uncertainty of the corresponding channel data. The random noise vector is sampled from a standard normal distribution. By introducing random factors, the model has the ability to perform probabilistic modeling, thus avoiding overfitting to single-channel data samples. This represents element-wise multiplication, where the mean vector, variance vector, and random noise vector are multiplied element-wise at corresponding positions, and the result is added to the corresponding element of the mean vector to obtain the channel latent variables. This reparameterization operation transforms the random sampling process into a differentiable operation, solving the problem of backpropagation in traditional probabilistic models. This ensures that the VAE model can achieve parameter optimization through gradient descent, and the generated channel latent variables retain the key features of the original channel data while incorporating quantitative information about data uncertainty.

[0106] In one embodiment, the reconstruction error is calculated based on the decoder output of the channel latent variables and the variational autoencoder model, and an uncertainty score is generated by combining the variance of the latent variables from the encoder network output, including:

[0107] S21. Input the channel latent variables into the decoder network of the variational autoencoder model, and obtain the reconstructed channel data through forward propagation.

[0108] Specifically, the decoder network adopts a hierarchical structure symmetrical to the encoder network to achieve the mapping of data from the latent space to the original space. At the beginning of forward propagation, the channel latent variables are first input into the fully connected layer, and the low-dimensional latent vectors are mapped to high-dimensional feature vectors that match the output dimension of the encoder convolutional layer through linear transformation. Then, the feature vectors enter the deconvolution layer. The first deconvolution layer upsamples the feature vectors through a convolution kernel of a specific size and recovers the spatial features of the channel data with the help of an activation function. The second deconvolution layer further adjusts the data dimension and finally outputs a reconstructed channel data tensor with the same dimension as the original channel data. This tensor retains the structural features of the original data, and its accuracy directly reflects the representational ability of the channel latent variables.

[0109] S22. Based on the original channel data and the reconstructed channel data, the reconstruction error is obtained using the following formula:

[0110]

[0111] in, For reconstruction error, The raw channel data, To reconstruct channel data.

[0112] In the above formula, The reconstruction error is used to quantitatively describe the difference between the original channel data and the reconstructed channel data. The smaller the error, the more completely the channel latent variables retain the features of the original data. The raw channel data, i.e. the millimeter-wave channel data before preprocessing, contains complete features including sub-band channel state information and beam calibration error data; To reconstruct the channel data, that is, the output of the decoder network forward propagation, is to restore the original channel data; The L2 norm square operation is used to calculate the sum of the squares of the differences between corresponding elements in the original data and the reconstructed data, thus obtaining the overall reconstruction error. This calculation method can amplify the impact of large differences and highlight key errors in the data restoration process.

[0113] S23. Extract the variance values ​​of each dimension from the variance vector output by the encoder network, and obtain the average variance using the following formula:

[0114]

[0115] in, The mean variance is a measure used to characterize the uncertainty of the distribution of latent variables. For the dimension of latent variables, For the first Variance.

[0116] In the above formula, The mean variance is the core function of which is to characterize the uncertainty of the distribution of latent variables. The larger the mean variance, the more dispersed the distribution of the latent variables in the latent space, and the more significant the non-ideal factors of the original channel data. The latent variable dimension, i.e., the number of elements in the channel latent variable vector, is determined by the structural design of the encoder network and is an important parameter reflecting the degree of channel data compression. For the first The variance in dimension represents the variance vector of the encoder output in dimension . The variance of each element value at each position corresponds to the degree of dispersion of the distribution of a certain feature dimension of the latent variable. After extracting the variance values ​​of all dimensions, the average is calculated to obtain the global average variance index.

[0117] S24. The reconstruction error and the average variance are weighted and combined to generate an uncertainty score. The expression for the uncertainty score is:

[0118]

[0119] in, For uncertain fractions, The preset tradeoff parameters are verified using historical data to balance the contributions of reconstruction error and distribution uncertainty. For reconstruction error, The average variance is denoted as .

[0120] In the above formula, Uncertainty score is a core indicator for comprehensively evaluating the reliability of channel data. The range of score values ​​is related to the quantification range of reconstruction error and average variance. The higher the score, the stronger the uncertainty of the current channel data, and the higher the risk of resource allocation decisions based on the data. To pre-set the trade-off parameters, the settings were verified using historical data, with values ​​ranging from 0 to 1. This applies when the accuracy of the original data reconstruction is more important than the distribution uncertainty. The value is close to 1, and vice versa. Its core function is to balance the contributions of reconstruction error and distribution uncertainty to uncertainty assessment. The reconstruction error calculated above reflects the uncertainty in the accuracy of data fitting; The average variance calculated earlier reflects the uncertainty at the level of latent variable distribution. This weighted combination method achieves a comprehensive quantification of channel uncertainty from two dimensions: data fit and distribution characteristics.

[0121] In an optional embodiment, the channel latent variables, the current network state, and the slice requirement are input into a deep reinforcement learning model, and a first resource allocation decision is output through a policy network, including:

[0122] S31. Construct the state space of the deep reinforcement learning model.

[0123] Specifically, a state space for the deep reinforcement learning model is constructed. This state space includes channel latent variables, slice demand vectors, and resource availability vectors. These three are concatenated dimensionally to form a unified state input vector, ensuring the model can comprehensively perceive the current network status and desired objectives. Channel latent variables, as the core feature components of the state space, are low-dimensional vectors compressed by a variational autoencoder, preserving the probability distribution characteristics and non-ideal factor information of the original channel data, providing a physical layer channel basis for decision-making. Slice demand vectors are constructed separately for three types of slices: enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (URLLC), and massive machine-type communication (mMTC). Each slice's demand vector includes two core indicators: throughput threshold and latency ceiling. eMBB slices emphasize high throughput thresholds, typically positively correlated with service bandwidth requirements; URLLC slices use extremely low latency ceilings as the core constraint, while matching the throughput floor corresponding to high reliability; and mMTC slices need to balance throughput thresholds and latency ceilings under massive connectivity. The resource availability vector directly reflects the current network's resource supply capacity. The number of available subbands is the total number of currently unoccupied millimeter-wave subbands, which is counted in real time by the base station resource management module. The number of beams is the number of independent beams that the massive MIMO antenna array can currently schedule, which is related to the hardware capability of beamforming. The power budget is the total transmit power that the base station can allocate within the current time window, which is determined by the base station energy management unit based on energy consumption constraints.

[0124] S32, Define the action space.

[0125] Specifically, an action space is defined, comprising a subband allocation vector, a beamforming weight matrix, and a power allocation vector. These three elements together constitute the complete output dimension of resource allocation decisions, and the actions in each dimension must match the actual network resource scheduling logic. The subband allocation vector is presented as an index set, with a length consistent with the number of slices. Each element is a list of indexes for the subbands allocated to the corresponding slice. The index values ​​in the list correspond one-to-one with the millimeter-wave subband numbers, ensuring the uniqueness and clarity of subband allocation and avoiding subband resource conflicts. The beamforming weight matrix is ​​a complex matrix, with a dimension equal to the product of the number of antenna ports and the number of beams. Each complex weight coefficient in the matrix corresponds to the weighted value of a single antenna port in a specific beam. By adjusting the amplitude and phase of the complex number, precise beam pointing is achieved, thereby matching the coverage requirements of different slices. The length of the power allocation vector is consistent with the total number of allocated subbands. Each element represents the power level of the corresponding subband. The range of power level values ​​is determined by the base station's power amplifier performance and energy consumption constraints, ensuring that power allocation meets both service requirements and hardware limitations.

[0126] S33. Train a deep reinforcement learning model using the proximal policy optimization algorithm, and obtain the trained deep reinforcement learning model by outputting the action probability distribution through the policy network and evaluating the state value function through the value network.

[0127] Specifically, both the policy network and the value network are implemented using neural networks. The PPO algorithm (Proximal Policy Optimization, a policy gradient algorithm) addresses the training instability caused by excessively large policy update steps in traditional policy gradient algorithms by introducing a pruning objective function, making it particularly suitable for high-dimensional action space scenarios such as network slice resource allocation. The policy network adopts a combined structure of "fully connected layers + batch normalization layers." The input is a state space vector, which extracts state features through multiple fully connected layers (using the ReLU activation function). The output layer then outputs the probability distribution of each action dimension: the sub-band allocation vector corresponds to the class distribution, the beamforming weight matrix corresponds to the Gaussian distribution in the complex domain, and the power allocation vector corresponds to the uniform distribution in a continuous interval. The value network shares the first half of the state feature extraction layer with the policy network. The second half outputs a state value function through an independent fully connected layer, which is used to evaluate the expected cumulative reward of all subsequent actions in the current state, providing a value benchmark for policy updates. During training, the model collects experience data by interacting with the network simulation environment. Each interaction includes a "state-action-reward-next state" quadruple. The experience data is stored in the experience replay pool for batch sampling. When updating parameters, the PPO algorithm controls the similarity between the new and old strategies within a preset range by pruning the policy update ratio, ensuring the stability of the training process until the model's cumulative reward converges to a preset threshold.

[0128] S34. Based on the trained deep reinforcement learning model, the current state is input into the policy network, and the first resource allocation decision is output by sampling the action probability distribution.

[0129] Specifically, the current state needs to undergo data preprocessing consistent with the training phase to ensure that the dimension and numerical range of the state vector match. After the preprocessed state vector is input into the policy network, the policy network calculates the probability distribution parameters of each action dimension through forward propagation. For the sub-band allocation vector, the sub-band index set of each slice is determined by sampling the category distribution; for the beamforming weight matrix, the complex weight coefficients of each antenna port are obtained based on sampling with a complex Gaussian distribution; for the power allocation vector, the power level of each sub-band is obtained through continuous distribution sampling. The sampled action combinations are converted into a standardized first resource allocation decision. The decision data is transmitted to the resource allocation unit through the base station's control interface. This unit drives the sub-band scheduling module, beamformer, and power amplifier to perform corresponding operations according to the decision content, realizing accurate sub-band allocation, dynamic adjustment of beam pointing, and on-demand power allocation, ultimately completing the implementation from model decision to actual resource scheduling.

[0130] In an optional embodiment, a deep reinforcement learning model is trained using a proximal policy optimization algorithm, including:

[0131] S41. Initialize the policy network parameters and value network parameters of the deep reinforcement learning model, and construct the experience replay buffer.

[0132] Specifically, policy network parameters Sum value network parameters All initialization methods employ randomization, with the initialization range controlled within a specific interval to avoid excessively large initial parameters that could lead to model output saturation. To improve initialization efficiency, the Xavier initialization method (a method for initializing neural network weights) can be used to ensure consistent variance between the input and output of each layer of the network, accelerating training convergence. The experience replay buffer adopts a first-in, first-out (FIFO) storage structure with a fixed storage capacity. When the buffer reaches its maximum capacity, the earliest stored state transition data is automatically deleted to ensure data timeliness. The state transition data stored in the buffer is ordered by... , , , , "It exists in the form of a quintuple, in which..." for Current state for Momentary action for Rewards at all times for Current state This is a termination flag (1 indicates the trajectory has ended, 0 indicates it continues).

[0133] S42. Execute the current policy in the simulated network environment, collect state-action-reward sequence data and store it in the experience replay buffer.

[0134] Specifically, the simulated network environment needs to accurately reproduce the core characteristics of 5G millimeter-wave networks, including subband interference randomness, beam calibration errors, and service characteristics of different slices, supporting dynamic adjustment of channel conditions to simulate the time-varying characteristics of real networks. When executing the current strategy, Moment State The input policy network samples the action a_t (i.e., resource allocation decision) and executes it in the environment. The environment updates the value to s_{t+1} based on the effect of the action and provides a reward. The reward calculation uses a weighted multi-objective function, the formula of which is: ,in, Service quality satisfaction (calculated based on throughput compliance rate and latency compliance rate). For resource utilization, Inter-subband interference level, , , The weighting coefficients are determined using the analytic hierarchy process (AHP) to match the priority requirements of the slices, ensuring that the rewards accurately guide the strategy optimization direction. The collected sequence data is stored in real-time in a quintuple format in the experience replay buffer, providing data support for subsequent batch training.

[0135] S43. Sample batch data from the experience replay buffer to obtain the dominance function estimate for each time step; the expression for the dominance function estimate is:

[0136]

[0137] in, This is the estimated value of the dominance function. For timing difference error, As a discount factor, The attenuation factor for generalized dominance estimation. This is the time step at which the trajectory terminates.

[0138] In the above formula, for The advantage function estimate at time step 1 is used for quantization. Moment Action The degree of advantage of a move relative to the average move; a positive value indicates that the move is better than the average level, while a negative value indicates the opposite. This is the core basis for strategy updates. for The timing difference error at time t is calculated using the following formula: This reflects the deviation between the predicted current state value and the actual return. The state value output by the value network; This is a discount factor, ranging from 0 to 1, used to balance the importance of immediate rewards versus future rewards. It is usually set to 0.95 to ensure that the model focuses more on short-term returns while also taking into account long-term benefits. This is the attenuation factor for the generalized advantage estimate (GAE), ranging from 0 to 1, with a typical value of 0.9. It is used to balance the bias and variance of the advantage estimate. The closer it is to 1, the smaller the bias of the dominance estimate, but the larger the variance. The trajectory termination time step is the end time of the current sampling sequence. At this point, the trajectory terminates, and the subsequent reward is 0. Through this cumulative summation method, the advantage function estimate integrates information from multi-step temporal difference errors, achieving an accurate assessment of the action's advantage.

[0139] S44. Update the policy network parameters by maximizing the pruning objective function; the expression for the pruning objective function is:

[0140]

[0141] in, For the purpose of cropping, For policy network parameters, This is the estimated value of the dominance function. The probability ratio between the old and new strategies. This is a pruning parameter used to limit the magnitude of policy updates.

[0142] In the above formula The pruning objective function is the optimization objective for updating the policy network parameters, and it is expressed in the expected form as the average target value of the batch data. These are the parameters that the policy network currently needs to update; The advantage function estimate calculated above is used to provide direction for the optimization of different actions; The probability ratio between the old and new strategies is calculated using the following formula: ,in Under the new strategy Status Selection The probability of an action. The corresponding probability of the old strategy before the update is used to measure the difference between the old and new strategies; This is the clipping parameter, used to limit the magnitude of policy updates. A typical value is 0.2 to prevent excessive policy changes in a single update from causing training failure. The `clip` function's purpose is to reduce the probability ratio... The constraint is in the interval [1-ε, 1+ε], when When the value exceeds this range, it is clipped to the interval boundary to ensure the robustness of the policy update. The min function is used for selection. and The smaller value in the equation is used as the objective term, which both guides policy optimization using the advantage function and prevents policy mutation through pruning.

[0143] S45. By minimizing the value function error, the value network parameters are updated. The expression for the value function error is:

[0144]

[0145] in, For value function error, For network parameters, This indicates the cumulative discount return. The state value is the predicted value of the value network.

[0146] in, The loss function of the value network is expressed in the form of mean squared error, which is used to measure the accuracy of the value network in predicting state values. The parameters of the value network; For value networks in parameters Below Moment State The projected value; express The cumulative discount return at any given time is calculated using the following formula: Reflecting from The present value of all future rewards starting from time point A is the state. An unbiased estimate of the true value; This represents the expected value calculation for batch sampled data. The value network parameter updates use gradient descent, and the loss function is calculated via backpropagation. The gradient is calculated, and the parameters are updated in the opposite direction of the gradient, so that the predicted value of the value network gradually approaches the true cumulative discount return, providing an accurate value benchmark for the estimation of the advantage function.

[0147] S46. Repeat steps S42 to S45 until the policy network and value network converge, obtaining a trained deep reinforcement learning model.

[0148] Specifically, convergence must satisfy two conditions: first, the pruning objective function of the policy network must be satisfied. If the fluctuation range is less than a preset threshold (e.g., 1e-4) for several consecutive training epochs (e.g., 10 epochs), it indicates that the policy update has stabilized and the target value cannot be significantly improved by parameter adjustment; secondly, the loss function of the value network. If the value network's prediction accuracy continues to decrease to a preset range (e.g., below 1e-3) and no longer decreases significantly, it indicates that the prediction accuracy has met the requirements. Furthermore, it needs to be verified using the average cumulative reward in the simulation environment. When the average reward for five consecutive rounds exceeds a preset performance threshold (set based on slice QoS requirements), model convergence can be confirmed, training can be stopped, and the final policy network and value network parameters can be saved.

[0149] In an optional embodiment, when the uncertainty score exceeds a preset threshold, the action space constraints of the deep reinforcement learning model are adjusted to obtain a first resource allocation method instruction, including the following steps:

[0150] S51. Compare the uncertainty score with the preset threshold to obtain the comparison result.

[0151] Specifically, the preset threshold is set by the quantile of the historical uncertainty score distribution and is used to distinguish between normal and conservative modes. The determination of the preset threshold relies on the correlation analysis of a large amount of historical channel data and the corresponding resource allocation effects. First, collect the uncertainty score samples under different channel scenarios over a period of time in the past. The samples should cover extreme scenarios such as strong sub-band interference and large beam calibration errors, as well as conventional scenarios with stable channels, ensuring the comprehensiveness of the sample distribution. Subsequently, statistically sort the historical score samples and calculate their quantiles. Usually, the score corresponding to a high percentile (such as the 90th percentile) is selected as the preset threshold. The selection of this quantile is based on the fact that the threshold can divide the scores of the vast majority of channel stable scenarios into "below the threshold" and only divide the scores of a few scenarios with extremely high uncertainty into "above the threshold", thereby reducing the excessive triggering of the conservative mode while ensuring network stability. The comparison process is executed in real time by the decision control unit of the base station, numerically comparing the currently calculated uncertainty score U with the preset threshold Th, and outputting a clear comparison result of "U>Th" or "U≤Th", providing a direct basis for subsequent mode switching.

[0152] S52. When the comparison result is that the uncertainty score exceeds the preset threshold, modify the action space constraint to obtain the modified action space constraint.

[0153] Specifically, the modification of the action space constraint needs to revolve around the core goals of "reducing decision-making risks and suppressing interference spread", and formulate differential constraint rules for different resource dimensions. The minimum number constraint for sub-band allocation is set according to the basic service requirements of various slices. For example, for the URLLC slice to ensure communication continuity, a minimum number of sub-band allocations not lower than a specific value needs to be set to avoid service interruption due to excessive contraction of sub-band resources caused by channel uncertainty; this minimum number is calculated by the ratio of the minimum bandwidth requirement of the slice to the bandwidth of a single sub-band, ensuring that the minimum rate requirement for service transmission is met. The maximum angle constraint for the beam width reduces interference by restricting the beam diffusion range. When the channel uncertainty is high, the risk of beam pointing deviation increases. The maximum angle of the beam width is reduced from θ1 in the normal mode to θ2 in the conservative mode (θ2 < θ1), reducing interference to adjacent slices by increasing the beam gain concentration. The setting of the angle value needs to consider the beamforming ability and coverage range requirements of the antenna array to avoid blind spots caused by excessive reduction. The upper limit constraint for power allocation is to lower the upper limit of the transmit power of each sub-band from P1 in the normal mode to P2 (P2 < P1). The value of P2 is based on the standard of "the inter-sub-band interference level does not exceed the preset interference threshold" and is calculated through the link budget formula, ensuring both the basic signal-to-noise ratio requirement of the service and preventing the negative impact brought by high power from exacerbating channel non-ideal factors.

[0154] S53. Based on the modified action space constraints, a first resource allocation decision is generated through the policy network of a deep reinforcement learning model, and a first resource allocation method instruction is generated based on the first resource allocation decision.

[0155] Specifically, the modified action space constraints are injected into the policy network of the deep reinforcement learning model through a parameter configuration interface. This is implemented by adding constraint judgment logic to the output layer of the policy network: when the number of sub-bands allocated by the network output is lower than the minimum, the idle sub-bands with the lowest interference levels are automatically allocated to meet the constraints; when the beamwidth exceeds the maximum angle, the beam range is compressed to meet the requirements by adjusting the beam weight matrix; when the power allocation value exceeds the upper limit, it is proportionally reduced to the upper limit and the power of each sub-band is rebalanced. The policy network samples actions under these constraints to generate a first resource allocation decision that balances constraint requirements and business needs. The decision includes a list of sub-band indices that meet the constraints, a beam weight matrix, and a power allocation vector. The first resource allocation method instruction is generated based on the decision. The instruction uses a standardized JSON format and includes the instruction type (conservative mode), constraint parameters (minimum number of sub-bands, maximum beam angle, power upper limit), decision details, and execution timestamp. It is then distributed to the resource allocation unit via the base station's resource scheduling bus to ensure accurate transmission and execution of the instruction.

[0156] In the aforementioned AI-based dynamic resource allocation method for 5G network slicing, raw channel data containing subband channel state information and beam calibration errors is collected and preprocessed. This data is then input into a variational autoencoder model to generate a channel latent variable that jointly characterizes the frequency-selective interference and beam error distribution features. Based on this latent variable and its decoding and reconstruction results, the reconstruction error is calculated, and an uncertainty score is generated by combining the variance output by the encoder. The channel latent variable, network state, and slice requirements are input into a deep reinforcement learning model to output a first resource allocation decision. When the uncertainty score exceeds a threshold, the decision is executed in a conservative mode by adjusting the action space constraints. In subsequent monitoring, when the performance is stable and the uncertainty decreases, the original constraints are restored to generate a second decision to switch back to the normal mode. This achieves adaptive adjustment of the resource allocation strategy under non-ideal channel conditions, effectively improving the robustness and quality of service assurance capabilities of 5G network slicing in the face of subband interference and beam calibration errors.

[0157] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0158] Based on the same inventive concept, this application also provides an AI-based 5G network slice resource dynamic allocation system for implementing the AI-based 5G network slice resource dynamic allocation method described above. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more AI-based 5G network slice resource dynamic allocation system embodiments provided below can be found in the limitations of the AI-based 5G network slice resource dynamic allocation method described above, and will not be repeated here.

[0159] In one exemplary embodiment, such as Figure 2 As shown, a dynamic allocation system 200 for 5G network slicing resources based on artificial intelligence is provided, including:

[0160] The data acquisition and preprocessing module 201 is used to acquire the raw channel data of the millimeter-wave channel and preprocess the raw channel data to obtain the preprocessed channel data; the raw channel data includes sub-band channel state information and beam calibration error data.

[0161] Channel data processing module 202 is used to input preprocessed channel data into variational autoencoder model and generate channel latent variables through encoder network;

[0162] Uncertainty score module 203 is used to calculate the reconstruction error based on the decoder output of the channel latent variables and variational autoencoder model, and generate uncertainty score by combining the latent variable variance of the encoder network output.

[0163] The first resource allocation decision module 204 is used to input channel latent variables, current network state and slice requirements into a deep reinforcement learning model, and output the first resource allocation decision through a policy network; the first resource allocation decision includes sub-band allocation, beamforming and power allocation.

[0164] The first resource allocation method module 205 is used to adjust the action space constraints of the deep reinforcement learning model when the uncertainty score exceeds a preset threshold, and obtain the first resource allocation method instruction; the first resource allocation method instruction is used to execute the first resource allocation decision in a conservative mode.

[0165] The second resource allocation decision module 206 is used to monitor network performance indicators in real time after executing the first resource allocation method instruction. When the network performance indicators are stable and the uncertainty score is lower than the preset threshold, the original action space constraints of the deep reinforcement learning model are restored to obtain the second resource allocation decision.

[0166] The second resource allocation method module 207 is used to generate a second resource allocation method instruction based on the second resource allocation decision; the second resource allocation method instruction is used to switch to normal mode to execute the second resource allocation decision.

[0167] Furthermore, the channel data processing module 202 is also used for:

[0168] S11. Input the preprocessed channel data into the encoder network of the pre-trained variational autoencoder model, and calculate the mean vector and variance vector of the latent variables through forward propagation.

[0169] S12. Based on the mean vector and variance vector, channel latent variables are generated through reparameterized sampling; the reparameterization formula is:

[0170]

[0171] in, This is the channel latent vector, used to characterize the probability distribution features of the original channel data in the latent space. It is the mean vector. It is the variance vector. A random noise vector sampled from a standard normal distribution. This represents element-wise multiplication.

[0172] Furthermore, the uncertainty fraction module 203 is also used for:

[0173] S21. Input the channel latent variables into the decoder network of the variational autoencoder model, and obtain the reconstructed channel data through forward propagation;

[0174] S22. Based on the original channel data and the reconstructed channel data, the reconstruction error is obtained using the following formula:

[0175]

[0176] in, For reconstruction error, The raw channel data, To reconstruct channel data;

[0177] S23. Extract the variance values ​​of each dimension from the variance vector output by the encoder network, and obtain the average variance using the following formula:

[0178]

[0179] in, The mean variance is a measure used to characterize the uncertainty of the distribution of latent variables. For the dimension of latent variables, For the first Variance;

[0180] S24. The reconstruction error and the average variance are weighted and combined to generate an uncertainty score. The expression for the uncertainty score is:

[0181]

[0182] in, For uncertain fractions, The preset tradeoff parameters are verified using historical data to balance the contributions of reconstruction error and distribution uncertainty. For reconstruction error, The average variance is denoted as .

[0183] Furthermore, the first resource allocation decision module 204 is also used for:

[0184] S31. Construct the state space of the deep reinforcement learning model; the state space includes channel latent variables, slice demand vector, and resource availability vector; the slice demand vector includes the throughput threshold and latency limit of each slice; the resource availability vector includes the number of available sub-bands, the number of beams, and the power budget;

[0185] S32. Define the action space; the action space includes the subband assignment vector, the beamforming weight matrix, and the power assignment vector; the subband assignment vector is the set of subband indices assigned to each slice; the beamforming weight matrix is ​​the complex weight coefficient of each beam; the power assignment vector is the power level of each subband;

[0186] S33. Train a deep reinforcement learning model using the proximal policy optimization algorithm. The policy network outputs the action probability distribution and the value network evaluates the state value function to obtain the trained deep reinforcement learning model. Both the policy network and the value network are implemented by neural networks.

[0187] S34. Based on the trained deep reinforcement learning model, the current state is input into the policy network, and the first resource allocation decision is output through the sampling action probability distribution; the first resource allocation decision is used to directly control the base station resource allocation unit.

[0188] Furthermore, the first resource allocation decision module 204 is also used for:

[0189] S41. Initialize the policy network parameters and value network parameters of the deep reinforcement learning model, and construct the experience replay buffer; the experience replay buffer is used to store state transition data;

[0190] S42. Execute the current policy in a simulated network environment, collect state-action-reward sequence data and store it in the experience replay buffer; the action includes the first resource allocation decision; the reward is calculated based on the service quality satisfaction; the state includes channel latent variables, network state and slice requirements.

[0191] S43. Sample batch data from the experience replay buffer to obtain the dominance function estimate for each time step; the expression for the dominance function estimate is:

[0192]

[0193] in, This is the estimated value of the dominance function. For timing difference error, As a discount factor, The attenuation factor for generalized dominance estimation. This is the time step at which the trajectory terminates;

[0194] S44. Update the policy network parameters by maximizing the pruning objective function; the expression for the pruning objective function is:

[0195]

[0196] in, For the purpose of cropping, For policy network parameters, This is the estimated value of the dominance function. The probability ratio between the old and new strategies. These are pruning parameters used to limit the magnitude of policy updates;

[0197] S45. By minimizing the value function error, the value network parameters are updated. The expression for the value function error is:

[0198]

[0199] in, For value function error, For network parameters, This indicates the cumulative discount return. The state values ​​predicted by the value network;

[0200] S46. Repeat steps S42 to S45 until the policy network and value network converge, obtaining a trained deep reinforcement learning model.

[0201] Furthermore, the first resource allocation module 205 is also used for:

[0202] S51, compare the uncertainty score with a preset threshold to obtain the comparison result; the preset threshold is set by the quantile of the historical uncertainty score distribution to distinguish between normal and conservative modes;

[0203] S52, when the comparison result shows that the uncertainty score exceeds the preset threshold, the action space constraint is modified to obtain the modified action space constraint; the action space constraint includes setting the minimum number of sub-band allocations, the maximum angle of beamwidth, and the upper limit of power allocation;

[0204] S53, based on the modified action space constraints, generates a first resource allocation decision through the policy network of a deep reinforcement learning model, and generates a first resource allocation method instruction based on the first resource allocation decision.

[0205] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the AI-based 5G network slice resource dynamic allocation method as described above.

[0206] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0207] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0208] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for dynamic allocation of 5G network slice resources based on artificial intelligence, characterized in that, The method includes: S1. Acquire the raw channel data of the millimeter-wave channel and preprocess the raw channel data to obtain preprocessed channel data; the raw channel data includes sub-band channel state information and beam calibration error data; S2. Input the preprocessed channel data into the variational autoencoder model, and generate channel latent variables through the encoder network; S3. Based on the channel latent variables and the decoder output of the variational autoencoder model, calculate the reconstruction error, and combine it with the variance of the latent variables output by the encoder network to generate an uncertainty score. S4. Input the channel latent variables, current network state and slice requirements into the deep reinforcement learning model, and output the first resource allocation decision through the policy network; S5. When the uncertainty score exceeds a preset threshold, the action space constraints of the deep reinforcement learning model are adjusted to obtain a first resource allocation method instruction; the first resource allocation method instruction is used to execute the first resource allocation decision in a conservative mode. S6. After executing the first resource allocation method instruction, monitor the network performance indicators in real time. When the network performance indicators are stable and the uncertainty score is lower than the preset threshold, restore the original action space constraints of the deep reinforcement learning model to obtain the second resource allocation decision. S7. Based on the second resource allocation decision, generate a second resource allocation mode instruction; the second resource allocation mode instruction is used to switch to normal mode to execute the second resource allocation decision.

2. The method according to claim 1, characterized in that, The step of inputting the preprocessed channel data into the variational autoencoder model and generating channel latent variables through the encoder network includes: S11. Input the preprocessed channel data into the encoder network of the pre-trained variational autoencoder model, and calculate the mean vector and variance vector of the latent variables through forward propagation. S12. Based on the mean vector and variance vector, the channel latent variable is generated through reparameterized sampling; the reparameterized formula is: in, This is the channel latent vector, used to characterize the probability distribution features of the original channel data in the latent space. It is the mean vector. It is the variance vector. A random noise vector sampled from a standard normal distribution. This represents element-wise multiplication.

3. The method according to claim 1, characterized in that, The decoder output based on the channel latent variables and the variational autoencoder model calculates the reconstruction error, and combines it with the variance of the latent variables output by the encoder network to generate an uncertainty score, including: S21. Input the channel latent variables into the decoder network of the variational autoencoder model, and obtain the reconstructed channel data through forward propagation; S22. Based on the original channel data and the reconstructed channel data, the reconstruction error is obtained using the following formula: in, For reconstruction error, This is the raw channel data. To reconstruct channel data; S23. Extract the variance values ​​of each dimension from the variance vector output by the encoder network, and obtain the average variance using the following formula: in, The mean variance is a measure used to characterize the uncertainty of the distribution of latent variables. For the dimension of latent variables, For the first Variance; S24. The reconstruction error and the average variance are weighted and combined to generate an uncertainty score, the expression of which is: in, For uncertain fractions, The preset tradeoff parameters are verified using historical data to balance the contributions of reconstruction error and distribution uncertainty. For reconstruction error, The average variance is denoted as .

4. The method according to claim 1, characterized in that, The step of inputting the channel latent variables, current network state, and slice requirements into a deep reinforcement learning model, and outputting a first resource allocation decision through a policy network, includes: S31. Construct the state space of the deep reinforcement learning model; the state space includes the channel latent variables, slice demand vector, and resource availability vector; the slice demand vector includes the throughput threshold and latency limit of each slice; the resource availability vector includes the number of available sub-bands, the number of beams, and the power budget; S32. Define the action space; the action space includes a subband allocation vector, a beamforming weight matrix, and a power allocation vector; the subband allocation vector is the set of subband indices assigned to each slice; the beamforming weight matrix is ​​the complex weight coefficient of each beam; the power allocation vector is the power level of each subband; S33. The deep reinforcement learning model is trained using the proximal policy optimization algorithm. The action probability distribution is output by the policy network and the state value function is evaluated by the value network to obtain the trained deep reinforcement learning model. Both the policy network and the value network are implemented by neural networks. S34. Based on the trained deep reinforcement learning model, the current state is input into the policy network, and the first resource allocation decision is output by sampling the action probability distribution; the first resource allocation decision is used to directly control the base station resource allocation unit.

5. The method according to claim 4, characterized in that, The process of training the deep reinforcement learning model using the proximal policy optimization algorithm includes: S41. Initialize the policy network parameters and value network parameters of the deep reinforcement learning model, and construct an experience replay buffer; the experience replay buffer is used to store state transition data; S42. Execute the current policy in a simulated network environment, collect state-action-reward sequence data and store it in the experience replay buffer; the action includes the first resource allocation decision; the reward is calculated based on service quality satisfaction; the state includes channel latent variables, network state and slice requirements. S43. Sample batch data from the experience playback buffer to obtain the advantage function estimate for each time step; the expression for the advantage function estimate is: in, This is the estimated value of the dominance function. For timing difference error, As a discount factor, The attenuation factor for generalized dominance estimation. This is the time step at which the trajectory terminates; S44. Update the policy network parameters by maximizing the pruning objective function; the expression for the pruning objective function is: in, For the purpose of cropping, For policy network parameters, This is the estimated value of the dominance function. The probability ratio between the old and new strategies. These are pruning parameters used to limit the magnitude of policy updates; S45. Update the value network parameters by minimizing the value function error, whereby the expression for the value function error is: in, For value function error, For network parameters, This indicates the cumulative discount return. The state values ​​predicted by the value network; S46. Repeat steps S42 to S45 until the policy network and the value network converge to obtain a trained deep reinforcement learning model.

6. The method according to claim 1, characterized in that, When the uncertainty score exceeds a preset threshold, the action space constraints of the deep reinforcement learning model are adjusted to obtain a first resource allocation method instruction, including: S51. The uncertainty score is compared with a preset threshold to obtain a comparison result; the preset threshold is set by the quantiles of the historical uncertainty score distribution to distinguish between normal and conservative modes. S52. When the comparison result is that the uncertainty score exceeds the preset threshold, the action space constraint is modified to obtain the modified action space constraint; the action space constraint includes setting the minimum number of sub-band allocations, the maximum angle of beamwidth, and the upper limit of power allocation; S53. Based on the modified action space constraints, the first resource allocation decision is generated through the policy network of the deep reinforcement learning model, and the first resource allocation method instruction is generated based on the first resource allocation decision.

7. A dynamic allocation system for 5G network slicing resources based on artificial intelligence, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire the raw channel data of the millimeter-wave channel and preprocess the raw channel data to obtain preprocessed channel data; the raw channel data includes sub-band channel state information and beam calibration error data. The channel data processing module is used to input the preprocessed channel data into the variational autoencoder model and generate channel latent variables through the encoder network. The uncertainty score module is used to calculate the reconstruction error based on the channel latent variables and the decoder output of the variational autoencoder model, and to generate an uncertainty score by combining the variance of the latent variables output by the encoder network. The first resource allocation decision module is used to input the channel latent variables, the current network state and the slice requirements into the deep reinforcement learning model, and output the first resource allocation decision through the policy network. The first resource allocation method module is used to adjust the action space constraints of the deep reinforcement learning model when the uncertainty score exceeds a preset threshold, thereby obtaining a first resource allocation method instruction; the first resource allocation method instruction is used to execute the first resource allocation decision in a conservative mode. The second resource allocation decision module is used to monitor network performance indicators in real time after executing the first resource allocation method instruction. When the network performance indicators are stable and the uncertainty score is lower than the preset threshold, the original action space constraints of the deep reinforcement learning model are restored to obtain the second resource allocation decision. The network performance indicators include throughput, latency, and packet loss rate. The second resource allocation method module is used to generate a second resource allocation method instruction based on the second resource allocation decision; the second resource allocation method instruction is used to switch to normal mode to execute the second resource allocation decision.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images