An intelligent monitoring method and system for abnormal base station facilities

Abstract

The present invention discloses a method and system for intelligent monitoring of base station facility anomalies. The method includes: constructing a dynamic signal perception topology based on real-time data from the base station physical layer using a multi-scale wavelet fusion pulse neural network to generate a time-space aligned full-dimensional feature tensor of the base station; inputting the full-dimensional feature tensor of the base station into a fault mode adversarial distillation module, reconstructing the normal operating manifold of the base station based on a variational autoencoder with physical constraints, and outputting an abnormal feature vector with a fault fingerprint identifier; modeling a fault propagation graph for the abnormal feature vector, and generating a base station fault root cause topology map in combination with backpropagation credibility verification; inputting the base station fault root cause topology map into a resilient self-healing strategy generator, and finally outputting a base station facility self-healing strategy set that meets real-time constraints. Utilizing the embodiments of the present invention, the accuracy and real-time performance of anomaly monitoring can be improved, and the operational reliability and communication service quality of the base station can be enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent monitoring technology, and in particular to a method and system for intelligent monitoring of abnormalities of base station facilities. Background Art

[0002] With the rapid development of modern communications technology, base station facilities, as a key component of mobile communication networks, shoulder the crucial tasks of signal coverage and data transmission. The proper operation of base stations is crucial to ensuring network reliability and service quality. However, over time, base stations may be affected by various factors, such as equipment aging, environmental changes, and sudden failures, leading to performance degradation or even downtime. This not only affects users' communication experience but can also lead to significant economic losses and safety risks.

[0003] Traditional base station monitoring methods rely primarily on static data analysis and log auditing, often suffering from long response times, insufficient accuracy, and an inability to adapt to complex operating conditions. These methods often fail to detect potential faults in a timely manner, leading to increased base station failure rates and increased maintenance costs. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for intelligently monitoring abnormalities of base station facilities to address the deficiencies in the prior art, thereby improving the accuracy and real-time performance of abnormality monitoring and enhancing the operational reliability and communication service quality of the base station.

[0005] An embodiment of the present application provides a method for intelligently monitoring abnormalities of base station facilities, the method comprising:

[0006] Based on real-time data from the base station physical layer, a dynamic signal perception topology is constructed using a multi-scale wavelet fusion spiking neural network. A phase-synchronized time-frequency decomposition algorithm is used to eliminate multi-source signal interference and generate a spatiotemporally aligned full-dimensional feature tensor for the base station. The spiking neural network introduces a pulse timing-dependent plasticity mechanism, achieving adaptive enhancement of signal characteristics through dynamic modification of synaptic weights.

[0007] The full-dimensional feature tensor of the base station is input into the fault mode adversarial distillation module. A variational autoencoder based on physical constraints reconstructs the normal operating manifold of the base station. A gradient-masked residual focusing algorithm is used to separate sudden faults from progressive aging features, and an abnormal feature vector with a fault fingerprint is output. The variational autoencoder uses Laplace noise injection and frequency-domain sparsity constraints to enforce the distinction between transient anomalies and steady-state degradation modes.

[0008] A fault propagation graph is modeled for the abnormal feature vector. A graph-enhanced causal reasoning model is used to simulate the fault diffusion path between hardware components. The propagation weight is dynamically modified through a gated graph attention mechanism. Combined with backpropagation credibility verification, a base station fault root cause topology map is generated. The model optimizes fault location accuracy by combining a time-series reverse truncation algorithm with propagation path entropy constraints.

[0009] The root cause topology map of the base station fault is input into the resilient self-healing strategy generator, and a multi-objective repair decision tree is constructed based on the dynamic game framework driven by digital twins. The invalid repair paths are compressed through the adversarial strategy distillation technology of physical constraints, and federated reinforcement learning is used to coordinate the allocation of repair resources among multiple base stations. Finally, a set of base station facility self-healing strategies that meet real-time constraints is output. The dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update.

[0010] Optionally, based on the real-time data of the base station physical layer, a multi-scale wavelet fusion pulse neural network is used to construct a dynamic signal perception topology, and a phase-synchronized time-frequency decomposition algorithm is used to eliminate multi-source signal interference to generate a time-space aligned full-dimensional feature tensor of the base station. The pulse neural network introduces a pulse timing-dependent plasticity mechanism and realizes adaptive enhancement of signal characteristics through dynamic correction of synaptic weights, including:

[0011] Based on the RF signal time-frequency matrix and the hardware vibration modal spectrum, a multi-scale wavelet ridge detection algorithm is used to perform time-frequency decomposition on the original signal. Multipath interference is eliminated by ridge line tracking to generate a phase-synchronized time-frequency feature map.

[0012] The time-frequency feature map, energy consumption fluctuation curve, and equipment operation log stream are input into a spiking neural network. The synaptic weights are dynamically adjusted based on the pulse timing-dependent plasticity mechanism. High-frequency abnormal signals are enhanced through an adaptive pulse trigger function, and a spatiotemporally correlated pulse sequence code is output.

[0013] Performing multimodal fusion on the pulse sequence encoding, selecting key frequency band features using a wavelet scale domain attention gating mechanism, and generating a multi-source signal joint representation matrix through tensor splicing;

[0014] The multi-source signal joint representation matrix is input into the spatiotemporal alignment compression module, and low-rank tensor decomposition and residual connection are used to eliminate sensor data drift, and finally the full-dimensional feature tensor of the base station is output.

[0015] Optionally, the full-dimensional feature tensor of the base station is input into the fault mode adversarial distillation module, a variational autoencoder based on physical constraints is used to reconstruct the normal operating manifold of the base station, sudden faults and progressive aging features are separated by a gradient masked residual focusing algorithm, and an abnormal feature vector with a fault fingerprint is output. The variational autoencoder uses Laplace noise injection and frequency domain sparse constraints to forcibly distinguish between transient anomalies and steady-state degradation modes, including:

[0016] Based on the full-dimensional feature tensor of the base station, a physically constrained variational autoencoder is constructed, and the latent space distribution is perturbed by Laplace noise injection to generate a normal operating manifold reconstruction vector that is resistant to overfitting;

[0017] Performing residual calculation on the normal working condition manifold reconstruction vector and the original feature tensor, separating the steady-state baseline features using a frequency domain sparse constraint filter, and outputting a frequency domain decoupled residual feature matrix;

[0018] The residual feature matrix is input into the gradient masking and focusing module, which masks the aging-related gradients through a dynamic threshold gating mechanism, strengthens the local gradient response of sudden faults, and generates abnormally sensitive feature vectors;

[0019] Perform fault fingerprint identification on the abnormal sensitive feature vector, use convolutional sparse coding to extract the time-frequency domain fault mode signature, and finally output the abnormal feature vector with a unique identifier.

[0020] Optionally, a fault propagation graph is modeled for the abnormal feature vector, a graph-enhanced causal reasoning model is used to simulate the fault diffusion path between hardware components, the propagation weight is dynamically corrected through a gated graph attention mechanism, and a base station fault root cause topology graph is generated in combination with backpropagation credibility verification. The model optimizes fault location accuracy by jointly optimizing a time series reverse truncation algorithm and a propagation path entropy constraint, including:

[0021] Based on the abnormal feature vector, a graph structure of the hardware component connection topology is constructed, and the fault propagation probability weights between nodes are dynamically calculated based on the gated graph attention mechanism to generate an initial fault diffusion map;

[0022] The initial fault diffusion graph is input into the time series reverse truncation module, the back propagation path length is constrained by the causal mask matrix, the propagation path confidence is optimized in combination with the path entropy loss function, and the time series corrected fault propagation subgraph is output;

[0023] Perform reinforcement learning optimization on the fault propagation subgraph, use the Q-learning algorithm to simulate the fault diffusion path selection strategy, and use Monte Carlo tree search to screen high-return root cause propagation links;

[0024] The credibility of the root cause propagation link is verified, and low-probability pseudo causal relationships are eliminated based on Bayesian hypothesis testing, and finally a base station fault root cause topology map is output.

[0025] Optionally, the base station fault root cause topology map is input into a resilient self-healing strategy generator, a multi-objective repair decision tree is constructed based on a dynamic game framework driven by digital twins, invalid repair paths are compressed through physical constraint adversarial strategy distillation technology, and federated reinforcement learning is used to coordinate the allocation of repair resources among multiple base stations, ultimately outputting a set of base station facility self-healing strategies that meet real-time constraints. The dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update, including:

[0026] Based on the base station fault root cause topology map, a digital twin-driven dynamic game framework is constructed. The initial repair strategy branches are generated through a multi-objective decision tree to simulate the feasibility of the strategy under the constraints of the base station's physical parameters.

[0027] The initial repair strategy branch is input into the adversarial strategy distillation module, the effectiveness of the strategy is evaluated using the generative adversarial network discriminator, invalid repair paths are compressed through gradient reversal, and a streamlined strategy candidate set is output;

[0028] Performing federated reinforcement learning collaborative optimization on the simplified strategy candidate set, using a distributed policy gradient algorithm to coordinate multi-base station resource allocation weights to generate a global optimal repair strategy vector;

[0029] The global optimal repair strategy vector is input into the twin simulation verification platform, and the strategy parameter deviation is corrected through the mirror update mechanism constrained by physical equations, and finally a base station self-healing strategy set that meets real-time constraints is output.

[0030] Another embodiment of the present application provides a base station facility abnormality intelligent monitoring system, the system comprising:

[0031] A construction module is used to construct a dynamic signal perception topology based on real-time data from the base station physical layer using a multi-scale wavelet fusion spiking neural network. This module eliminates multi-source signal interference through a phase-synchronized time-frequency decomposition algorithm to generate a spatiotemporally aligned full-dimensional feature tensor for the base station. The spiking neural network introduces a pulse timing-dependent plasticity mechanism and achieves adaptive enhancement of signal characteristics through dynamic modification of synaptic weights.

[0032] A reconstruction module is configured to input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module, reconstruct the normal operating manifold of the base station based on a variational autoencoder with physical constraints, separate sudden faults from progressive aging features through a gradient-masked residual focusing algorithm, and output an abnormal feature vector with a fault fingerprint. The variational autoencoder uses Laplace noise injection and frequency-domain sparsity constraints to enforce the distinction between transient anomalies and steady-state degradation modes.

[0033] A correction module is used to model a fault propagation graph for the abnormal feature vector, using a graph-enhanced causal reasoning model to simulate the fault diffusion path between hardware components. The module dynamically corrects the propagation weights through a gated graph attention mechanism and combines backpropagation credibility verification to generate a base station fault root cause topology graph. The model optimizes fault location accuracy by combining a time-series reverse truncation algorithm with propagation path entropy constraints.

[0034] The output module is used to input the base station fault root cause topology map into the resilient self-healing strategy generator, construct a multi-objective repair decision tree based on the dynamic game framework driven by digital twins, compress invalid repair paths through the adversarial strategy distillation technology of physical constraints, and use federated reinforcement learning to coordinate the allocation of repair resources among multiple base stations, and finally output a set of base station facility self-healing strategies that meet real-time constraints. The dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update.

[0035] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0036] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0037] Compared with the existing technology, the present invention provides a method for intelligent monitoring of base station facility anomalies. According to the real-time data of the base station physical layer, a multi-scale wavelet fusion pulse neural network is used to construct a dynamic signal perception topology to generate a time-space aligned full-dimensional feature tensor of the base station; the full-dimensional feature tensor of the base station is input into the fault mode adversarial distillation module, and the variational autoencoder based on physical constraints reconstructs the normal operating manifold of the base station, and outputs an abnormal feature vector with a fault fingerprint identifier; a fault propagation graph is modeled for the abnormal feature vector, and a base station fault root cause topology graph is generated in combination with back-propagation credibility verification; the base station fault root cause topology graph is input into a resilient self-healing strategy generator, and finally a base station facility self-healing strategy set that meets real-time constraints is output, thereby improving the accuracy and real-time performance of anomaly monitoring and enhancing the operational reliability and communication service quality of the base station. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A hardware structure block diagram of a computer terminal for a method for intelligently monitoring abnormalities of base station facilities provided by an embodiment of the present invention;

[0039] Figure 2 A flowchart of a method for intelligently monitoring abnormalities of base station facilities provided by an embodiment of the present invention;

[0040] Figure 3 A schematic structural diagram of a base station facility abnormality intelligent monitoring system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0042] The embodiment of the present invention first provides a method for intelligently monitoring abnormalities of base station facilities. The method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0043] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for a method for intelligently monitoring abnormalities of base station facilities provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0044] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which, when executed, can cause the processor to execute any one of the base station facility abnormality intelligent monitoring methods.

[0045] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0046] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any one of the base station facility abnormality intelligent monitoring methods.

[0047] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0048] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0049] See also Figure 2 The embodiment of the present invention provides a method for intelligently monitoring abnormalities of base station facilities, which may include the following steps:

[0050] S201: Based on real-time data from the base station physical layer, a dynamic signal perception topology is constructed using a multi-scale wavelet fusion spiking neural network. Multi-source signal interference is eliminated through a phase-synchronized time-frequency decomposition algorithm to generate a spatiotemporally aligned full-dimensional feature tensor for the base station. The spiking neural network introduces a pulse timing-dependent plasticity mechanism and achieves adaptive enhancement of signal characteristics through dynamic modification of synaptic weights.

[0051] Specifically, the original signal can be decomposed into time and frequency using a multi-scale wavelet ridge detection algorithm based on the RF signal time-frequency matrix and the hardware vibration modal spectrum. Multipath interference can be eliminated by ridge line tracking to generate a phase-synchronized time-frequency feature map.

[0052] Base station physical layer data includes the time-frequency matrix (sampling rate 1 GHz, time window length 10 ms) of the RF signal (e.g., 5G NR signal frequency band 3.5 GHz) and the hardware vibration modal spectrum (accelerometer sampling rate 5 kHz, bandwidth 0-2 kHz). The multi-scale wavelet ridge detection algorithm uses Morlet wavelet basis functions and adjusts the scale parameters (e.g., scale sequence 1, 2, 4, 8) to cover different frequency band characteristics. For example, scale 1 corresponds to high-frequency noise (e.g., switching power supply interference), scale 4 corresponds to medium-frequency vibration (e.g., abnormal fan vibration), and scale 8 corresponds to low-frequency baseband signal fluctuations.

[0053] Ridge tracing is achieved through the following steps:

[0054] Time-frequency energy focusing: Calculate the modulus maximum of the wavelet coefficients at each scale to form candidate ridges. For example, in the time-frequency matrix of an RF signal, the ridges of a normal signal are continuous and smooth, while multipath interference (such as building reflections) produces discontinuous ridges.

[0055] Phase consistency verification: Calculate the phase change rate (in radians per second) between adjacent time points. If the phase jump exceeds π / 2 (the threshold is adjustable), it is identified as an interference ridge and removed. For example, a test discovered a phase jump at 150Hz in the vibration signal. After tracking, it was confirmed that it was caused by a loose screw, and the ridge was marked as a valid feature.

[0056] Multipath interference elimination: Based on ridge spatial density clustering (DBSCAN algorithm, neighborhood radius ε = 0.1, minimum sample size 5), the main path signal is separated from multipath reflection signals. For example, the main path signal ridges form a continuous band structure in the time-frequency plane, while multipath interference appears as discrete point clusters.

[0057] The resulting time-frequency signature is a three-dimensional tensor (time × frequency × energy intensity), for example, 1000 × 256 × 3 (time point × frequency bin × energy / phase / coherence). A phase synchronization algorithm (such as a phase-locked loop (PLL)) is used to align the clock offsets of different sensors (with an accuracy of ±1 μs) to ensure the time-frequency correlation between the vibration spectrum and the RF signal.

[0058] The time-frequency feature map, energy consumption fluctuation curve, and equipment operation log stream are input into a spiking neural network. The synaptic weights are dynamically adjusted based on the pulse timing-dependent plasticity mechanism. High-frequency abnormal signals are enhanced through an adaptive pulse trigger function, and a spatiotemporally correlated pulse sequence code is output.

[0059] The Spiking Neural Network (SNN) uses the Leaky Integrate-and-Fire (LIF) neuron model. The network structure includes an input layer (time-frequency features, energy consumption curve, log event encoding), a hidden layer (256 neurons), and an output layer (spike encoding).

[0060] Data encoding:

[0061] Time-frequency feature map: The energy value of each frequency slot is mapped to the firing rate of the input neuron (0~200Hz).

[0062] Energy consumption fluctuation curve: The current value (unit A) is encoded into pulse intervals through Delta modulation (such as 10A→50ms interval, 20A→25ms interval).

[0063] Device log stream: Event types (such as "temperature alarm" and "CPU overload") are encoded as one-hot vectors, and pulse trains are generated through a Poisson process.

[0064] Spike Timing-Dependent Plasticity (STDP) Mechanism:

[0065] Synaptic weight update rule: If the presynaptic neuron spikes earlier than the postsynaptic neuron, the weight increases (long-term potentiation, LTP); otherwise, the weight decreases (long-term depression, LTD). The update amplitude Δw = 0.01 × e^(-Δt / 20ms), where Δt is the pulse time difference. For example, if an abnormal RF signal (a sudden increase in high-frequency energy) triggers an input layer neuron to spike prematurely, the corresponding hidden layer neuron weight will increase by 15%.

[0066] Adaptive pulse triggering: Dynamically adjust the neuron threshold (baseline threshold V_th = 30mV). When high-frequency abnormalities are detected (such as the energy of the vibration spectrum at 1kHz is continuously 3 times higher than the baseline), the threshold is reduced to 25mV to enhance the pulse firing rate.

[0067] Spatiotemporal correlation coding: The output layer uses a population coding strategy, counting the pulses of the 256 neurons in the hidden layer according to a time window (50ms) to generate a binary pulse sequence (for example, "1011" indicates pulses in time slots 1, 3, and 4). For example, if a fan failure causes a periodic pulse cluster at 800Hz in the vibration spectrum, the corresponding output code will be a periodic "1001" pattern.

[0068] Performing multimodal fusion on the pulse sequence encoding, selecting key frequency band features using a wavelet scale domain attention gating mechanism, and generating a multi-source signal joint representation matrix through tensor splicing;

[0069] The goal of multimodal fusion is to integrate the spatiotemporal characteristics of pulse coding with the frequency domain characteristics of the original signal. The specific process is as follows:

[0070] Wavelet scale domain attention gating:

[0071] Scale energy calculation: Perform a discrete wavelet transform (DWT) on the time-frequency feature map, decomposing it into four subbands (approximation coefficient A3, detail coefficients D1-D3). Calculate the energy contribution of each subband (for example, the D3 subband accounts for 60% of the energy in the 1-2 kHz frequency range).

[0072] Attention weight generation: The Gated Recurrent Unit (GRU) learns the correlation between each subband and the pulse code. For example, when the pulse code shows high-frequency anomalies, the GRU increases the attention weight of the D3 subband to 0.8 and suppresses the weight of the A3 subband to 0.2.

[0073] Gated filtering: Soft threshold filtering (threshold = 0.1 × maximum coefficient value) is performed on low-weight subbands to retain key features. For example, after one filtering, the noise coefficient (<0.05) of the D3 subband is removed, and the effective feature retention rate is >90%.

[0074] Tensor concatenation and dimensionality reduction:

[0075] The pulse code matrix (size 1000×256) and the time-frequency features after wavelet filtering (1000×256×4) are concatenated along the feature axis to form a joint matrix (1000×256×5).

[0076] Non-negative matrix factorization (NMF) was used to reduce the dimensionality to 1000 × 256 × 2, retaining the main modes. For example, after decomposition, the first mode reflects the RF-vibration coupling characteristics, and the second mode reflects the energy consumption-log event correlation.

[0077] Example scenario: When the aging of a base station power module causes increased energy consumption fluctuations, pulse encoding shows dense pulses in power-related neurons (such as numbers 50-70). Wavelet attention gating enhances the vibration characteristics of the D2 subband (500Hz-1kHz), and the joint matrix finally highlights power-vibration coupling anomalies.

[0078] The multi-source signal joint representation matrix is input into the spatiotemporal alignment compression module, and low-rank tensor decomposition and residual connection are used to eliminate sensor data drift, and finally the full-dimensional feature tensor of the base station is output.

[0079] The spatiotemporal alignment compression module solves the temporal asynchrony and spatial redundancy issues of multi-sensor data. The core steps are as follows:

[0080] Low-rank tensor decomposition:

[0081] Tucker decomposition is used to decompose the joint representation matrix (1000×256×2) into a core tensor (50×50×2) and a factor matrix (temporal factor 1000×50, spatial factor 256×50, modal factor 2×2).

[0082] The proposed algorithm is iteratively optimized using the alternating least squares (ALS) method, with the objective function including a low-rank constraint (rank = 50) and a sparsity regularization term (λ = 0.01). For example, the decomposed temporal factor matrix reveals the diurnal pattern of base station load, while the spatial factor matrix identifies the hotspots of the antenna array.

[0083] Residual connection compensation:

[0084] The residual between the original joint matrix and the reconstructed matrix (core tensor × factor matrix) is passed through a bidirectional LSTM network to model sensor drift. For example, the zero-point drift of the accelerometer (0.1g per month) is predicted and compensated by the LSTM.

[0085] The residual-compensated data is added to the low-rank reconstruction result to generate the drift-corrected feature tensor.

[0086] Full-dimensional feature integration:

[0087] Concatenate the time factor matrix (1000×50), spatial factor matrix (256×50), and modal factor matrix (2×2) into a full-dimensional tensor (1000×256×2×50).

[0088] A deep autoencoder (DAE) is used to further compress the data to 1000×256×8 (8-dimensional latent features), retaining the principal components with a variance contribution rate of >95%.

[0089] Output example: In a certain fault scenario, the full-dimensional tensor shows abnormally high values at the 500th point in the time dimension (corresponding to the time of the fault), the 120th-130th units in the spatial dimension (corresponding to the RF power amplifier module), and the 5th-6th channels in the modal dimension, accurately indicating signal distortion caused by power amplifier overheating.

[0090] S202: Input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module, reconstruct the normal operating manifold of the base station based on a variational autoencoder with physical constraints, separate sudden faults from progressive aging features using a gradient-masked residual focusing algorithm, and output an abnormal feature vector with a fault fingerprint. The variational autoencoder uses Laplace noise injection and frequency-domain sparsity constraints to forcibly distinguish between transient anomalies and steady-state degradation modes.

[0091] Specifically, a physically constrained variational autoencoder can be constructed based on the full-dimensional feature tensor of the base station, and the potential space distribution can be perturbed by Laplace noise injection to generate a normal working condition manifold reconstruction vector that is resistant to overfitting.

[0092] The physically constrained variational autoencoder (VAE) aims to learn the characteristic distribution of base station normal operating conditions. Its core is to reconstruct the input data through an encoder-decoder architecture and introduce physical prior knowledge to constrain the latent space. The encoder uses a four-layer convolutional neural network (CNN) as input, taking the full-dimensional feature tensor of the base station (dimensions: time × frequency × sensor channels, for example, 100 × 128 × 8) and outputting mean and variance vectors (dimension: 64) representing the latent space distribution. The decoder, consisting of a three-layer deconvolutional network, reconstructs the latent variables back to the original feature dimensions.

[0093] Physics constraints are implemented in two ways:

[0094] Laplace noise injection: When sampling the latent space, Laplace-distributed noise (with scale parameter β = 0.1) is added to the mean vector. The noise intensity is positively correlated with the standard deviation of the sensor measurement error (e.g., 0.05). For example, if the mean of a latent vector is [0.3, -0.2, ...], the noise becomes [0.3+0.07, -0.2-0.03, ...], forcing the model to learn more robust feature representations.

[0095] Frequency-domain sparse constraint: A frequency-domain L1 regularization term with a weight of λ = 0.5 is added to the loss function. In implementation, a fast Fourier transform (FFT) is performed on the reconstructed signal output by the decoder, and the absolute value sum of high-frequency components (>50 Hz) is calculated to penalize unnecessary high-frequency components. For example, under normal operating conditions, base station vibration signals are primarily concentrated in the 0-30 Hz range, while abnormal signals may contain transient impact components above 100 Hz. The sparsity constraint suppresses the impact of these components on the reconstruction.

[0096] During training, the Adam optimizer (learning rate 0.001, β1=0.9, β2=0.999) was used, with a batch size of 64, and convergence after 500 iterations. The normal operating manifold reconstruction vector (dimension: 64) will serve as the benchmark for subsequent anomaly detection. For example, the reconstruction vectors of a base station clustered in a spherical distribution in the latent space, while anomalies deviate from this distribution.

[0097] Performing residual calculation on the normal working condition manifold reconstruction vector and the original feature tensor, separating the steady-state baseline features using a frequency domain sparse constraint filter, and outputting a frequency domain decoupled residual feature matrix;

[0098] The purpose of residual calculation is to extract the difference between the original features and the normal reconstructed features, highlighting the abnormal components. The specific process is as follows:

[0099] Residual generation: The original feature tensor (100×128×8) is subtracted element-by-element from the reconstructed tensor output by the decoder to generate the residual tensor. For example, if the original RF signal power at a certain time point is 25dBm and the reconstructed value is 24.8dBm, the residual is +0.2dBm.

[0100] Frequency domain sparse filtering: Perform FFT on the residual tensor along the time dimension to generate a frequency domain residual spectrum (dimensions: frequency × time × channel). Design a sparse filter:

[0101] High-pass filter: cutoff frequency 10 Hz, attenuation slope 24 dB / octave, filters out low-frequency aging-related fluctuations (such as 0.1 Hz fluctuations caused by the slow decrease in cooling fan speed).

[0102] Transient pulse detection: Wavelet transform modulus maximum detection is used to identify burst pulses with a duration of less than 50ms (such as 100Hz transient interference caused by circuit short circuit).

[0103] The filtered residual feature matrix (dimensions: 100×128×8) retains only high-frequency burst components and transient pulses. For example, a channel detects a sudden increase in residual energy at 200Hz (amplitude 0.5dB) at time t=30s, while the residual of the steady-state aging feature is suppressed to below 0.02dB.

[0104] Frequency-domain decoupling is achieved through independent component analysis (ICA), which decomposes multi-channel residuals into statistically independent source signals. For example, independent components such as "power module voltage fluctuation," "RF amplifier harmonic distortion," and "abnormal fan vibration" are isolated. Each component corresponds to a subspace of the residual matrix (dimensions: 100 × 128 × 3).

[0105] The residual feature matrix is input into the gradient masking and focusing module, which masks the aging-related gradients through a dynamic threshold gating mechanism, strengthens the local gradient response of sudden faults, and generates abnormally sensitive feature vectors;

[0106] The goal of the gradient masking focusing module is to distinguish between gradual aging (such as capacitor capacity decay) and sudden failures (such as fuse blowing). Its core is to filter irrelevant gradients through dynamic thresholds and amplify abnormally sensitive areas.

[0107] Gradient calculation: The Sobel operator is used to calculate the spatial gradient of the residual matrix (along the frequency and channel dimensions), resulting in a gradient magnitude matrix (dimensions: 100 × 128 × 8). For example, the gradient magnitudes of a frequency bin at 10 consecutive time points are [0.1, 0.3, 0.8, 1.2, 0.7, ...], indicating fluctuations caused by workload changes during the day.

[0108] Dynamic Threshold Gating:

[0109] Aging gradient threshold: Calculate the moving average (MA) and standard deviation (STD) of the gradient within a sliding window (window size 10 time points), and set the threshold to MA + 3 × STD. For example, if MA = 0.2 and STD = 0.05 within the window, the threshold is 0.35. Gradients below this value are considered aging-related and are masked.

[0110] Sudden fault enhancement: Nonlinearly amplifies gradient regions exceeding the threshold using an exponential function: Amplification factor = 1 + 2 × (gradient value - threshold) / threshold. For example, when the gradient is 0.5, the amplification factor = 1 + 2 × (0.5-0.35) / 0.35 = 1.86, increasing the gradient amplitude to 0.5 × 1.86 = 0.93.

[0111] Local Response Normalization: Perform spatial local competitive normalization (LRN) on the amplified gradient with a window size of 3×3 to suppress gradient redundancy in adjacent regions. For example, a peak gradient of 0.93 is reduced to 0.85 after normalization, while the surrounding gradient is reduced from 0.3 to 0.1.

[0112] In the processed gradient matrix (dimensions: 100 × 128 × 8), only high-gradient regions related to sudden faults are retained. For example, a power module's gradient suddenly increases to 1.2 at t = 45 seconds (corresponding to a 5% voltage drop), while aging-related gradients (such as 0.25 at t = 10 seconds) are completely suppressed.

[0113] Perform fault fingerprint identification on the abnormal sensitive feature vector, use convolutional sparse coding to extract the time-frequency domain fault mode signature, and finally output the abnormal feature vector with a unique identifier.

[0114] Fault fingerprinting aims to assign a unique code to each type of anomaly to facilitate subsequent root cause analysis. Convolutional sparse coding (CSC) is used to extract interpretable fault mode signatures from the time-frequency domain residuals.

[0115] Dictionary learning: During the offline training phase, the dictionary is learned using historical fault data (including 50 known fault types):

[0116] The dictionary size is set to 100 atoms (atom length 128, corresponding to the frequency dimension);

[0117] The K-SVD algorithm was used for iterative optimization, with the sparse constraint term λ=0.1 and 1000 iterations.

[0118] For example, Atom 1 might capture the "RF power amplifier harmonic distortion" pattern (center frequency 2.4 GHz, bandwidth 50 MHz), and Atom 2 corresponds to the "power surge" pattern (transient pulse width 10 ms).

[0119] Online encoding: Sparse encoding of abnormality-sensitive feature vectors generated in real time:

[0120] The input is the residual matrix after gradient focusing (100×128×8), which is separated by channel and processed separately;

[0121] Use the OMP (Orthogonal Matching Pursuit) algorithm to solve the sparse coefficients, and limit the number of non-zero coefficients to 5 per sample;

[0122] For example, the RF channel residual at a certain moment is decomposed into atom 3 (weight 0.7) + atom 15 (weight 0.3), corresponding to the mixed characteristics of "antenna standing wave ratio abnormality" and "clock jitter".

[0123] Fingerprint generation: The sparse coefficient vector (dimension 100) is converted into a 128-bit binary fingerprint through hash coding. For example, atomic positions with coefficient absolute values greater than 0.2 are set to 1, and all others are set to 0, generating a hexadecimal identifier such as "0x1A3F...". Time-frequency domain statistical features (such as peak frequency and pulse duration) are also added to form the final anomaly feature vector (dimension 256).

[0124] Uniqueness Verification: The cosine similarity is used to evaluate the fingerprint differences between different faults. For example, the fingerprint similarity between two power supply faults is 0.3, while the similarity between the power supply fault and the RF fault is less than 0.1, ensuring distinguishability.

[0125] S203: Modeling a fault propagation graph for the abnormal feature vector. A graph-enhanced causal reasoning model is used to simulate the fault diffusion path between hardware components. Propagation weights are dynamically modified using a gated graph attention mechanism. Backpropagation credibility verification is combined with the generation of a base station fault root cause topology graph. The model optimizes fault location accuracy using a time-series reverse truncation algorithm combined with propagation path entropy constraints.

[0126] Specifically, a graph structure of the hardware component connection topology can be constructed based on the abnormal feature vector, and the fault propagation probability weights between nodes can be dynamically calculated based on the gated graph attention mechanism to generate an initial fault diffusion graph.

[0127] The graph structure of the hardware component connection topology is constructed based on the physical connections between base station devices. Nodes represent hardware units (such as radio modules, power supply units, and baseband processors), and edges represent physical or logical dependencies between components (such as a power supply unit supplying power to a radio module). The feature vector of each node is composed of the concatenation of the anomaly feature vector (dimension 128) and device operating parameters (temperature, voltage, and signal quality), forming a 256-dimensional node feature.

[0128] The core of the Gated Graph Attention Network (GGAT) is to dynamically adjust attention weights through gating units. The specific implementation is as follows:

[0129] Attention coefficient calculation: For each pair of adjacent nodes (such as the RF module and baseband processor), the node features are projected onto the query and key vectors using a shared weight matrix (dimension 256×64), and the dot product attention score is calculated. For example, the dot product score of the RF module's query vector and the baseband processor's key vector is 0.8, while the score with the power supply unit is 0.3, reflecting the former's stronger functional dependency.

[0130] Gating mechanism: A GRU (Gated Recurrent Unit) is used to dynamically filter attention scores. The GRU's update gate determines whether to retain or ignore the propagation probability of an edge based on the node's current state and aggregated neighbor features. For example, when the power supply unit temperature exceeds a threshold (such as 80°C), the update gate reduces the connection weight between it and the RF module from 0.3 to 0.1.

[0131] Weight normalization: Softmax normalization is performed on all outgoing edges of the same node to ensure that the total propagation probability is 1. For example, the weight of the RF module to the baseband processor is 0.6, and the weight to the power unit is 0.4.

[0132] The final generated initial fault diffusion graph is stored in the form of an adjacency matrix, where the matrix elements represent the probability of a fault propagating from node i to node j (e.g., the probability of the RF module → baseband processor is 0.7, and the probability of the RF module → power supply unit is 0.2).

[0133] The initial fault diffusion graph is input into the time series reverse truncation module, the back propagation path length is constrained by the causal mask matrix, the propagation path confidence is optimized in combination with the path entropy loss function, and the time series corrected fault propagation subgraph is output;

[0134] The goal of the time-reverse truncation module is to limit the temporal depth of fault propagation paths, avoiding computational explosion caused by infinitely tracing back historical states. The causal mask matrix is an upper triangular matrix (1s on and above the main diagonal, 0s elsewhere) that forces propagation paths to be traced back only a limited number of steps (e.g., 3). For example, if the current fault occurs at time step t, only propagation paths within the time window from t-3 to t are allowed to be analyzed.

[0135] The path entropy loss function is used to quantify the uncertainty of the propagation path. The specific calculation method is:

[0136] Calculate the probability distribution of each path (e.g., RF module → power supply unit → baseband processor) (e.g., path probabilities 0.4, 0.3, 0.3);

[0137] Calculate the Shannon Entropy of the distribution. A higher entropy value indicates greater path uncertainty.

[0138] Adding entropy as a loss term to the total loss function forces the model to converge to highly confident propagation paths. For example, if the entropy of a path drops from 1.2 (high uncertainty) to 0.5 (low uncertainty), the model is more confident in its causal relationship.

[0139] The optimization process uses the Adam optimizer (learning rate 0.001, β1 = 0.9, β2 = 0.999), updating the GGAT weight parameters through backpropagation. During training, the path length is limited to a maximum of three hops (for example, a fault path from the radio module to the baseband processor to the antenna array is two hops in total). Paths exceeding this length are filtered using a mask matrix. For example, a four-hop path (radio module → power supply → fan → chassis) is truncated to three hops (radio module → power supply → fan).

[0140] After optimization, the timing-corrected fault propagation subgraph retains only high-confidence paths (e.g., probability > 0.5). For example, 10 paths in the initial graph may be compressed into 3 critical paths, each with an entropy loss value below 0.3.

[0141] Perform reinforcement learning optimization on the fault propagation subgraph, use the Q-learning algorithm to simulate the fault diffusion path selection strategy, and use Monte Carlo tree search to screen high-return root cause propagation links;

[0142] The reinforcement learning framework models fault propagation as a Markov decision process (MDP):

[0143] State: The node where the current fault occurs (such as a power supply unit) and its historical path (such as the past two-hop path);

[0144] Action: Select the next propagation node (e.g., from the power supply unit to the baseband processor);

[0145] Reward: The weighted sum of the path's diagnostic accuracy (verified based on historical failure data) and the repair cost (such as the cost of replacing a component), with a weight of 7:3.

[0146] The parameters of the Q-learning algorithm are set as follows:

[0147] Q-table: Due to the large state space (possibly containing tens of thousands of node combinations), a deep Q-network (DQN) is used to approximate the Q function. The network structure is a 3-layer fully connected (256→128→64) network.

[0148] Exploration strategy: ε-greedy strategy, initial ε = 0.3 (30% probability of randomly selecting an action), decaying by 10% every 1000 training steps;

[0149] Discount factor γ: 0.9, emphasizing the importance of recent rewards.

[0150] Monte Carlo Tree Search (MCTS) is used to expand high-reward paths:

[0151] Selection: Starting from the root node (initial failure point), child nodes are selected based on the UCB1 formula (exploration coefficient c=1.4) until a leaf node is reached;

[0152] Expansion: If a leaf node is not fully explored, add a new child node (such as adding a new propagation edge);

[0153] Simulation: Randomly simulate the propagation of faults to a terminal state (such as complete repair or system crash) and calculate the cumulative reward;

[0154] Backpropagation: Use the simulation results to reversely update the number of visits and reward values of the path nodes.

[0155] For example, after MCTS expanded the power supply unit node, the simulation found that the cumulative reward for the path "power supply → baseband processor → radio module" was 85 points (diagnostic accuracy 90%, repair cost 500), while the reward for the path "power supply → fan → chassis" was 60 points (accuracy 70500), and the reward for the path "power supply → fan → chassis" was 60 points (accuracy 70200). Ultimately, the former was selected as the root cause link with the higher reward.

[0156] The credibility of the root cause propagation link is verified, and low-probability pseudo causal relationships are eliminated based on Bayesian hypothesis testing, and finally a base station fault root cause topology map is output.

[0157] Bayesian hypothesis testing verifies the statistical significance of causal relationships by calculating posterior probabilities. The specific steps include:

[0158] Prior probability setting: Based on the historical fault database, the prior occurrence frequency of each path is calculated. For example, the historical occurrence probability of the "power supply → baseband processor" path is 0.4;

[0159] Likelihood calculation: Calculates likelihood probability based on the matching degree between the current fault characteristics (such as voltage sag, temperature anomaly) and the path. For example, if the current fault is accompanied by a baseband processor log error, the likelihood probability of the path "power supply → baseband processor" increases to 0.7.

[0160] Posterior probability calculation: Apply the Bayesian formula to update the path probability, for example, posterior probability = (likelihood × prior) / evidence factor;

[0161] Significance test: Set the significance level α = 0.05. If the posterior probability of a path is lower than the threshold (e.g. <0.05), it is determined to be a pseudo-causal relationship and is eliminated.

[0162] For example, the posterior probability of the path "RF module → antenna array" is 0.03 (<0.05) and is rejected; while the posterior probability of "power supply → baseband processor" is 0.62 and is retained.

[0163] The resulting base station root cause topology is presented as a directed graph, with nodes labeled with failure probabilities (e.g., baseband processor: 85%) and edges labeled with propagation probabilities and timing information (e.g., power supply → baseband processor: probability 0.7, delay 3 seconds). This graph directly guides operations personnel to prioritize high-probability failure points and, in conjunction with the self-healing strategy generator, triggers automated repair processes.

[0164] S204: Input the base station fault root cause topology map into the resilient self-healing strategy generator, construct a multi-objective repair decision tree based on the dynamic game framework driven by digital twins, compress invalid repair paths through the adversarial strategy distillation technology of physical constraints, and use federated reinforcement learning to coordinate the allocation of repair resources among multiple base stations, and finally output a set of base station facility self-healing strategies that meet real-time constraints. The dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update.

[0165] Specifically, a digital twin-driven dynamic game framework can be constructed based on the base station fault root cause topology map, and the initial repair strategy branches can be generated through a multi-objective decision tree to simulate the feasibility of the strategy under the constraints of the base station's physical parameters.

[0166] The dynamic game framework driven by digital twins creates virtual mirrors of base station hardware components, mapping the physical world's state in real time. First, based on the root cause topology, faulty nodes (such as radio frequency units and power modules) are mapped to the digital twin's three-dimensional model and associated with their physical parameters (such as temperature thresholds and voltage fluctuation ranges). For example, if the root cause topology for a base station radio frequency unit failure shows "cooling fan failure → temperature exceeding the specified limit → signal distortion," the twin model simulates the dynamic process of the temperature rising from 25°C to 85°C after the fan stops, and predicts the time window for signal distortion (e.g., after 30 seconds).

[0167] The multi-objective decision tree is constructed using NSGA-III (third generation non-dominated sorting genetic algorithm), and the optimization objectives include:

[0168] Repair time (target < 5 minutes, weight 0.4);

[0169] Resource consumption (cost budget < 1,000 yuan, weight 0.3);

[0170] Fault recurrence probability (target <5%, weight 0.3).

[0171] The branching rules of decision trees are based on physical constraints:

[0172] Hardware constraints: If replacing a power module requires power outage, a 2-minute power outage time must be reserved when generating branches.

[0173] Logical constraint: If the fault propagation path contains multiple nodes (such as A→B→C), the repair order must follow the reverse path (repair C first, then B, and finally A).

[0174] Example of an initial repair strategy branch:

[0175] Branch 1: Restart the cooling fan immediately (cost 0 yuan, time 30 seconds, recurrence probability 40%)

[0176] Branch 2: Replace the fan and upgrade the firmware (cost 500 yuan, time 4 minutes, recurrence probability 5%)

[0177] Branch 3: Enable the backup radio unit and isolate the faulty module (cost 800 yuan, time 2 minutes, recurrence probability 10%).

[0178] The physical feasibility of each branch was verified through twin simulation. For example, branch 1 was marked as an invalid path in the simulation because the fan was unable to restart due to aging. However, the temperature control simulation of branch 2 showed that it could be stabilized at 45°C, meeting the physical constraints.

[0179] The initial repair strategy branch is input into the adversarial strategy distillation module, the effectiveness of the strategy is evaluated using the generative adversarial network discriminator, invalid repair paths are compressed through gradient reversal, and a streamlined strategy candidate set is output;

[0180] The adversarial strategy distillation module consists of a generator and a discriminator:

[0181] Generator: A 4-layer fully connected network (input dimension 256, hidden layers 512 / 256 / 128, output dimension 64), responsible for generating the potential representation of the repair strategy;

[0182] Discriminator: 3-layer convolutional network (kernel size 3×3, stride 1, number of channels 32 / 64 / 128), outputting a strategy effectiveness score (0-1).

[0183] Training process:

[0184] Generator input: feature vector of the initial strategy branch (including parameters such as repair time, cost, and recurrence probability);

[0185] Discriminator training: supervised learning using historically successful strategies (positive samples) and artificially constructed ineffective strategies (negative samples), with a learning rate of 0.001 and a batch size of 32;

[0186] Gradient reversal: During the generator training phase, the gradient of the invalid strategy is multiplied by a negative coefficient (such as -0.5) during back propagation, forcing the generator to avoid invalid areas.

[0187] For example, when the generator proposes "clean only the fan blades (cost 50 yuan)," the discriminator, based on historical data, determines that this strategy has a 60% probability of recurring in high-temperature environments and assigns a low score of 0.2. The gradient reversal mechanism causes the generator to reduce the probability of generating such strategies.

[0188] Policy compression:

[0189] Invalid path elimination: strategies with a score < 0.3 are directly deleted;

[0190] Merging similar strategies: Using the DBSCAN clustering algorithm (neighborhood radius 0.2, minimum sample number 3), we merge strategies with a cost difference of less than 100 yuan and a time difference of less than 1 minute.

[0191] Ultimately, the initial 20 strategies were narrowed down to five candidate strategies, such as:

[0192] Candidate strategy A: fan replacement + firmware upgrade (score 0.92);

[0193] Candidate strategy B: Enable backup module + remote diagnosis (score 0.85).

[0194] Performing federated reinforcement learning collaborative optimization on the simplified strategy candidate set, using a distributed policy gradient algorithm to coordinate multi-base station resource allocation weights to generate a global optimal repair strategy vector;

[0195] The Federated Reinforcement Learning (FRL) framework consists of a central server and multiple base station nodes (e.g., 10 base stations). Each base station trains a policy gradient model locally and periodically uploads model parameters to the central server for aggregation.

[0196] The distributed policy gradient algorithm uses PPO (Proximal Policy Optimization):

[0197] State space: base station fault type (encoded as a 4-dimensional vector), resource inventory (number of spare parts, budget remaining);

[0198] Action space: select candidate strategies (5 discrete actions);

[0199] Reward function:

[0200] R = 0.5 × (1 − repair time / 300 s) + 0.3 × (1 − cost / 1000 yuan) + 0.2 × (1 − recurrence probability).

[0201] Federal collaborative process:

[0202] Local training: Each base station trains the PPO model based on local historical data (e.g., 1000 fault records), with 10 iterations and a learning rate of 0.0002.

[0203] Parameter aggregation: The central server uses the FedAvg algorithm to perform a weighted average of the strategic network parameters of the 10 base stations (weights are assigned based on the base station failure frequency);

[0204] Strategy distribution: The updated global model is distributed to each base station for the next round of training.

[0205] Resource allocation optimization:

[0206] Conflict resolution: When multiple base stations simultaneously request scarce resources (such as a specific fan model), an auction mechanism (Vickrey-Clarke-Groves rule) is used to allocate them.

[0207] Dynamic priority: Adjust resource scheduling weights based on business criticality (e.g., 5G core network base stations have higher priority than ordinary base stations).

[0208] After 5 rounds of federated training, the average reward of the global strategy on the test set increased from 0.65 to 0.82. The optimal strategy vector is as follows:

[0209] Policy vector: [action index 2 (enable backup module), resource weight 0.7, delay constraint 150 seconds].

[0210] The global optimal repair strategy vector is input into the twin simulation verification platform, and the strategy parameter deviation is corrected through the mirror update mechanism constrained by physical equations, and finally a base station self-healing strategy set that meets real-time constraints is output.

[0211] The twin simulation platform integrates multi-physics simulation engines, including:

[0212] Thermodynamic simulation: Calculate the device temperature distribution based on Fourier's law;

[0213] Electromagnetic simulation: FDTD (finite-difference time-domain) method is used to simulate signal attenuation;

[0214] Mechanical stress simulation: Predict component life through finite element analysis.

[0215] The mirror update mechanism dynamically corrects policy parameters by comparing the deviation between simulation results and physical constraints:

[0216] Parameter deviation detection: For example, the strategy predicts that the temperature will be 45°C after replacing the fan, but the simulation shows that it is actually 50°C, a deviation of 5°C.

[0217] Backpropagation correction: adjust parameters along the direction of gradient descent of the policy network (such as increasing the heat sink area or extending the fan startup time);

[0218] Constraint reinforcement: Penalty gradients (learning rate amplified by 10 times) are applied to policies that violate hard constraints (such as maximum temperature > 85°C).

[0219] Real-time guarantee:

[0220] Lightweight simulation: Using reduced-order models (ROMs) to reduce full simulation time from 30 minutes to 2 minutes;

[0221] Edge computing deployment: Deploy simulation containers (Docker) locally at the base station and use GPU acceleration (NVIDIA T4) to achieve real-time feedback.

[0222] The final self-healing policy set output includes:

[0223] Emergency strategy: For high-priority faults (such as power short circuit), the response time is less than 1 minute;

[0224] Standard strategy: routine hardware replacement, time < 5 minutes;

[0225] Preventive strategy: Proactive maintenance based on aging prediction, triggered 24 hours in advance.

[0226] Example: Self-healing strategy execution process for a base station power module failure:

[0227] Step 1: The digital twin detects an abnormal input voltage (fluctuation ±15%).

[0228] Step 2: Root cause analysis identifies "voltage regulator aging";

[0229] Step 3: Select "Switch to backup power + issue a maintenance work order" for the federal policy;

[0230] Step 4: Simulate and verify the backup power supply switching delay (measured at 45 seconds, meeting the <1 minute constraint).

[0231] Step 5: The policy takes effect, the base station resumes normal operation, and the data is uploaded to the management platform.

[0232] It can be seen that according to the real-time data of the base station physical layer, a multi-scale wavelet fusion pulse neural network is used to construct a dynamic signal perception topology, and a time-space aligned full-dimensional feature tensor of the base station is generated; the full-dimensional feature tensor of the base station is input into the fault mode adversarial distillation module, and the variational autoencoder based on physical constraints reconstructs the normal operating manifold of the base station, and outputs an abnormal feature vector with a fault fingerprint identifier; the fault propagation graph of the abnormal feature vector is modeled, and the base station fault root cause topology map is generated in combination with back-propagation credibility verification; the base station fault root cause topology map is input into the resilient self-healing strategy generator, and finally a base station facility self-healing strategy set that meets the real-time constraints is output, thereby improving the accuracy and real-time performance of abnormality monitoring, and improving the operational reliability and communication service quality of the base station.

[0233] Another embodiment of the present invention provides a base station facility abnormality intelligent monitoring system, see Figure 3 , the system may include:

[0234] Construction module 301 is used to construct a dynamic signal perception topology based on real-time data from the base station physical layer using a multi-scale wavelet fusion spiking neural network, eliminate multi-source signal interference through a phase-synchronized time-frequency decomposition algorithm, and generate a spatiotemporally aligned full-dimensional feature tensor of the base station. The spiking neural network introduces a pulse timing-dependent plasticity mechanism and achieves adaptive enhancement of signal characteristics through dynamic modification of synaptic weights.

[0235] Reconstruction module 302 is configured to input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module, reconstruct the normal operating manifold of the base station based on a variational autoencoder with physical constraints, separate sudden faults from progressive aging features through a gradient-masked residual focusing algorithm, and output an abnormal feature vector with a fault fingerprint. The variational autoencoder uses Laplace noise injection and frequency-domain sparsity constraints to forcibly distinguish between transient anomalies and steady-state degradation modes.

[0236] Correction module 303 is used to model a fault propagation graph for the abnormal feature vector, using a graph-enhanced causal reasoning model to simulate the fault diffusion path between hardware components, dynamically correcting the propagation weights through a gated graph attention mechanism, and generating a base station fault root cause topology graph in combination with backpropagation credibility verification. The model jointly optimizes fault location accuracy through a time-series reverse truncation algorithm and propagation path entropy constraints.

[0237] Output module 304 is used to input the base station fault root cause topology map into the resilient self-healing strategy generator, construct a multi-objective repair decision tree based on the dynamic game framework driven by digital twins, compress invalid repair paths through the adversarial strategy distillation technology of physical constraints, and use federated reinforcement learning to coordinate the allocation of repair resources among multiple base stations, and finally output a set of base station facility self-healing strategies that meet real-time constraints. The dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update.

[0238] It can be seen that according to the real-time data of the base station physical layer, a multi-scale wavelet fusion pulse neural network is used to construct a dynamic signal perception topology, and a time-space aligned full-dimensional feature tensor of the base station is generated; the full-dimensional feature tensor of the base station is input into the fault mode adversarial distillation module, and the variational autoencoder based on physical constraints reconstructs the normal operating manifold of the base station, and outputs an abnormal feature vector with a fault fingerprint identifier; the fault propagation graph of the abnormal feature vector is modeled, and the base station fault root cause topology map is generated in combination with back-propagation credibility verification; the base station fault root cause topology map is input into the resilient self-healing strategy generator, and finally a base station facility self-healing strategy set that meets the real-time constraints is output, thereby improving the accuracy and real-time performance of abnormality monitoring, and improving the operational reliability and communication service quality of the base station.

[0239] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0240] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps:

[0241] S201: Based on real-time data from the base station physical layer, a dynamic signal perception topology is constructed using a multi-scale wavelet fusion spiking neural network. Multi-source signal interference is eliminated through a phase-synchronized time-frequency decomposition algorithm to generate a spatiotemporally aligned full-dimensional feature tensor for the base station. The spiking neural network introduces a pulse timing-dependent plasticity mechanism and achieves adaptive enhancement of signal characteristics through dynamic modification of synaptic weights.

[0242] S202: Input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module. A variational autoencoder based on physical constraints reconstructs the normal operating manifold of the base station. A gradient-masked residual focusing algorithm is used to separate sudden faults from progressive aging features, and an abnormal feature vector with a fault fingerprint is output. The variational autoencoder uses Laplace noise injection and frequency-domain sparsity constraints to enforce the distinction between transient anomalies and steady-state degradation modes.

[0243] S203: Modeling a fault propagation graph for the abnormal feature vector. A graph-enhanced causal reasoning model is used to simulate the fault diffusion path between hardware components. Propagation weights are dynamically modified using a gated graph attention mechanism. Backpropagation credibility verification is combined with the generation of a base station fault root cause topology graph. The model optimizes fault location accuracy using a time-series reverse truncation algorithm combined with propagation path entropy constraints.

[0244] S204: Input the base station fault root cause topology map into the resilient self-healing strategy generator, construct a multi-objective repair decision tree based on the dynamic game framework driven by digital twins, compress invalid repair paths through the adversarial strategy distillation technology of physical constraints, and use federated reinforcement learning to coordinate the allocation of repair resources among multiple base stations, and finally output a set of base station facility self-healing strategies that meet real-time constraints. The dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update.

[0245] It can be seen that according to the real-time data of the base station physical layer, a multi-scale wavelet fusion pulse neural network is used to construct a dynamic signal perception topology, and a time-space aligned full-dimensional feature tensor of the base station is generated; the full-dimensional feature tensor of the base station is input into the fault mode adversarial distillation module, and the variational autoencoder based on physical constraints reconstructs the normal operating manifold of the base station, and outputs an abnormal feature vector with a fault fingerprint identifier; the fault propagation graph of the abnormal feature vector is modeled, and the base station fault root cause topology map is generated in combination with back-propagation credibility verification; the base station fault root cause topology map is input into the resilient self-healing strategy generator, and finally a base station facility self-healing strategy set that meets the real-time constraints is output, thereby improving the accuracy and real-time performance of abnormality monitoring, and improving the operational reliability and communication service quality of the base station.

[0246] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0247] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0248] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0249] S201: Based on real-time data from the base station physical layer, a dynamic signal perception topology is constructed using a multi-scale wavelet fusion spiking neural network. Multi-source signal interference is eliminated through a phase-synchronized time-frequency decomposition algorithm to generate a spatiotemporally aligned full-dimensional feature tensor for the base station. The spiking neural network introduces a pulse timing-dependent plasticity mechanism and achieves adaptive enhancement of signal characteristics through dynamic modification of synaptic weights.

[0250] S202: Input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module. A variational autoencoder based on physical constraints reconstructs the normal operating manifold of the base station. A gradient-masked residual focusing algorithm is used to separate sudden faults from progressive aging features, and an abnormal feature vector with a fault fingerprint is output. The variational autoencoder uses Laplace noise injection and frequency-domain sparsity constraints to enforce the distinction between transient anomalies and steady-state degradation modes.

[0251] S203: Modeling a fault propagation graph for the abnormal feature vector. A graph-enhanced causal reasoning model is used to simulate the fault diffusion path between hardware components. Propagation weights are dynamically modified using a gated graph attention mechanism. Backpropagation credibility verification is combined with the generation of a base station fault root cause topology graph. The model optimizes fault location accuracy using a time-series reverse truncation algorithm combined with propagation path entropy constraints.

[0252] S204: Input the base station fault root cause topology map into the resilient self-healing strategy generator, construct a multi-objective repair decision tree based on the dynamic game framework driven by digital twins, compress invalid repair paths through the adversarial strategy distillation technology of physical constraints, and use federated reinforcement learning to coordinate the allocation of repair resources among multiple base stations, and finally output a set of base station facility self-healing strategies that meet real-time constraints. The dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update.

[0253] It can be seen that according to the real-time data of the base station physical layer, a multi-scale wavelet fusion pulse neural network is used to construct a dynamic signal perception topology, and a time-space aligned full-dimensional feature tensor of the base station is generated; the full-dimensional feature tensor of the base station is input into the fault mode adversarial distillation module, and the variational autoencoder based on physical constraints reconstructs the normal operating manifold of the base station, and outputs an abnormal feature vector with a fault fingerprint identifier; the fault propagation graph of the abnormal feature vector is modeled, and the base station fault root cause topology map is generated in combination with back-propagation credibility verification; the base station fault root cause topology map is input into the resilient self-healing strategy generator, and finally a base station facility self-healing strategy set that meets the real-time constraints is output, thereby improving the accuracy and real-time performance of abnormality monitoring, and improving the operational reliability and communication service quality of the base station.

[0254] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A method for intelligently monitoring abnormalities of base station facilities, characterized in that: The method comprises: Based on real-time data from the base station physical layer, a dynamic signal perception topology is constructed using a multi-scale wavelet fusion spiking neural network. A phase-synchronized time-frequency decomposition algorithm is used to eliminate multi-source signal interference and generate a spatiotemporally aligned full-dimensional feature tensor for the base station. The spiking neural network introduces a pulse timing-dependent plasticity mechanism, achieving adaptive enhancement of signal characteristics through dynamic modification of synaptic weights. The full-dimensional feature tensor of the base station is input into the fault mode adversarial distillation module. A variational autoencoder based on physical constraints reconstructs the normal operating manifold of the base station. A gradient-masked residual focusing algorithm is used to separate sudden faults from progressive aging features, and an abnormal feature vector with a fault fingerprint is output. The variational autoencoder uses Laplace noise injection and frequency-domain sparsity constraints to enforce the distinction between transient anomalies and steady-state degradation modes. A fault propagation graph is modeled for the abnormal feature vector. A graph-enhanced causal reasoning model is used to simulate the fault diffusion path between hardware components. The propagation weight is dynamically modified through a gated graph attention mechanism. Combined with backpropagation credibility verification, a base station fault root cause topology map is generated. The model optimizes fault location accuracy by combining a time-series reverse truncation algorithm with propagation path entropy constraints. The base station fault root cause topology map is input into the resilient self-healing strategy generator. A multi-objective repair decision tree is constructed based on a dynamic game framework driven by digital twins. Invalid repair paths are compressed using physically constrained adversarial strategy distillation technology. Federated reinforcement learning is used to coordinate the allocation of repair resources among multiple base stations. Ultimately, a set of base station facility self-healing strategies that meet real-time constraints is output. The dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update. The method uses a multi-scale wavelet fusion pulse neural network to construct a dynamic signal perception topology based on the real-time data of the base station physical layer, eliminates multi-source signal interference through a phase-synchronized time-frequency decomposition algorithm, and generates a time-space aligned full-dimensional feature tensor of the base station. The pulse neural network introduces a pulse timing-dependent plasticity mechanism and realizes adaptive enhancement of signal characteristics through dynamic correction of synaptic weights, including: Based on the RF signal time-frequency matrix and the hardware vibration modal spectrum, a multi-scale wavelet ridge detection algorithm is used to perform time-frequency decomposition on the original signal. Multipath interference is eliminated by ridge line tracking to generate a phase-synchronized time-frequency feature map. The time-frequency characteristic map, energy consumption fluctuation curve, and equipment operation log stream are input into the spiking neural network. The synaptic weights are dynamically adjusted based on the pulse timing-dependent plasticity mechanism. High-frequency abnormal signals are strengthened through an adaptive pulse trigger function. The spatiotemporally correlated pulse sequence code is output. The energy consumption fluctuation curve is the pulse interval generated by the current value through delta modulation coding. The pulse sequence coding is subjected to multimodal fusion, and the key frequency band features are selected using the wavelet scale domain attention gating mechanism. A multi-source signal joint representation matrix is generated by tensor splicing. The goal of multimodal fusion is to integrate the spatiotemporal characteristics of the pulse coding with the frequency domain characteristics of the original signal. The multi-source signal joint representation matrix is input into the spatiotemporal alignment compression module, low-rank tensor decomposition and residual connection are used to eliminate sensor data drift, and finally the full-dimensional feature tensor of the base station is output; The full-dimensional feature tensor of the base station is input into the fault mode adversarial distillation module, and the normal operating condition manifold of the base station is reconstructed based on the variational autoencoder of physical constraints. The sudden fault and progressive aging features are separated by the gradient masked residual focusing algorithm, and the abnormal feature vector with the fault fingerprint is output. The variational autoencoder uses Laplace noise injection and frequency domain sparse constraints to forcibly distinguish between transient anomalies and steady-state degradation modes, including: Based on the full-dimensional feature tensor of the base station, a physically constrained variational autoencoder is constructed, and the latent space distribution is perturbed by Laplace noise injection to generate a normal operating manifold reconstruction vector that is resistant to overfitting; Performing residual calculation on the normal working condition manifold reconstruction vector and the original feature tensor, separating the steady-state baseline features using a frequency domain sparse constraint filter, and outputting a frequency domain decoupled residual feature matrix; The residual feature matrix is input into the gradient masking and focusing module, which masks the aging-related gradients through a dynamic threshold gating mechanism, strengthens the local gradient response of sudden faults, and generates abnormally sensitive feature vectors; Perform fault fingerprint identification on the abnormal sensitive feature vector, extract the time-frequency domain fault mode signature using convolutional sparse coding, and finally output an abnormal feature vector with a unique identifier; The fault propagation graph modeling is performed on the abnormal feature vector, and a graph-enhanced causal reasoning model is used to simulate the fault diffusion path between hardware components. The propagation weight is dynamically corrected through a gated graph attention mechanism, and a base station fault root cause topology graph is generated in combination with backpropagation credibility verification. The model jointly optimizes fault location accuracy through a time series reverse truncation algorithm and propagation path entropy constraints, including: Based on the abnormal feature vector, a graph structure of the hardware component connection topology is constructed, and the fault propagation probability weights between nodes are dynamically calculated based on the gated graph attention mechanism to generate an initial fault diffusion map; The graph structure of the hardware component connection topology is constructed based on the physical connection relationships between base station devices. Nodes represent hardware units, and edges represent physical or logical dependencies between components. The feature vector of each node is composed of the concatenation of the abnormal feature vector and the device operating parameters. The initial fault diffusion graph is input into the time series reverse truncation module, the back propagation path length is constrained by the causal mask matrix, the propagation path confidence is optimized in combination with the path entropy loss function, and the time series corrected fault propagation subgraph is output; Perform reinforcement learning optimization on the fault propagation subgraph, use the Q-learning algorithm to simulate the fault diffusion path selection strategy, and use Monte Carlo tree search to screen high-return root cause propagation links; ‌ The root cause propagation link is verified for credibility, low-probability pseudo-causal relationships are eliminated based on Bayesian hypothesis testing, and finally a base station fault root cause topology map is output; The base station fault root cause topology map is input into the resilient self-healing strategy generator, and a multi-objective repair decision tree is constructed based on a dynamic game framework driven by digital twins. Invalid repair paths are compressed through physically constrained adversarial strategy distillation technology, and federated reinforcement learning is used to coordinate the allocation of repair resources among multiple base stations. Finally, a set of base station facility self-healing strategies that meet real-time constraints is output. The dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update, including: Based on the base station fault root cause topology map, a digital twin-driven dynamic game framework is constructed. The initial repair strategy branches are generated through a multi-objective decision tree to simulate the feasibility of the strategy under the constraints of the base station's physical parameters. The initial repair strategy branch is input into the adversarial strategy distillation module, the effectiveness of the strategy is evaluated using the generative adversarial network discriminator, invalid repair paths are compressed through gradient reversal, and a streamlined strategy candidate set is output; Performing federated reinforcement learning collaborative optimization on the simplified strategy candidate set, using a distributed policy gradient algorithm to coordinate multi-base station resource allocation weights to generate a global optimal repair strategy vector; The global optimal repair strategy vector is input into the twin simulation verification platform, and the strategy parameter deviation is corrected through the mirror update mechanism constrained by physical equations, and finally a base station self-healing strategy set that meets real-time constraints is output.

2. A base station facility abnormality intelligent monitoring system, characterized in that: The system comprises: A construction module is used to construct a dynamic signal perception topology based on real-time data from the base station physical layer using a multi-scale wavelet fusion spiking neural network. This module eliminates multi-source signal interference through a phase-synchronized time-frequency decomposition algorithm to generate a spatiotemporally aligned full-dimensional feature tensor for the base station. The spiking neural network introduces a pulse timing-dependent plasticity mechanism and achieves adaptive enhancement of signal characteristics through dynamic modification of synaptic weights. A reconstruction module is configured to input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module, reconstruct the normal operating manifold of the base station based on a variational autoencoder with physical constraints, separate sudden faults from progressive aging features through a gradient-masked residual focusing algorithm, and output an abnormal feature vector with a fault fingerprint. The variational autoencoder uses Laplace noise injection and frequency-domain sparsity constraints to enforce the distinction between transient anomalies and steady-state degradation modes. A correction module is used to model a fault propagation graph for the abnormal feature vector, using a graph-enhanced causal reasoning model to simulate the fault diffusion path between hardware components. The module dynamically corrects the propagation weights through a gated graph attention mechanism and combines backpropagation credibility verification to generate a base station fault root cause topology graph. The model optimizes fault location accuracy by combining a time-series reverse truncation algorithm with propagation path entropy constraints. An output module is configured to input the base station fault root cause topology map into a resilient self-healing strategy generator, construct a multi-objective repair decision tree based on a dynamic game framework driven by digital twins, compress invalid repair paths through physically constrained adversarial strategy distillation technology, and coordinate the allocation of repair resources among multiple base stations using federated reinforcement learning. Ultimately, the module outputs a set of base station facility self-healing strategies that meet real-time constraints. The dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirroring updates. The method uses a multi-scale wavelet fusion pulse neural network to construct a dynamic signal perception topology based on the real-time data of the base station physical layer, eliminates multi-source signal interference through a phase-synchronized time-frequency decomposition algorithm, and generates a time-space aligned full-dimensional feature tensor of the base station. The pulse neural network introduces a pulse timing-dependent plasticity mechanism and realizes adaptive enhancement of signal characteristics through dynamic correction of synaptic weights, including: Based on the RF signal time-frequency matrix and the hardware vibration modal spectrum, a multi-scale wavelet ridge detection algorithm is used to perform time-frequency decomposition on the original signal. Multipath interference is eliminated by ridge line tracking to generate a phase-synchronized time-frequency feature map. The time-frequency characteristic map, energy consumption fluctuation curve, and equipment operation log stream are input into the spiking neural network. The synaptic weights are dynamically adjusted based on the pulse timing-dependent plasticity mechanism. High-frequency abnormal signals are strengthened through an adaptive pulse trigger function. The spatiotemporally correlated pulse sequence code is output. The energy consumption fluctuation curve is the pulse interval generated by the current value through delta modulation coding. The pulse sequence coding is subjected to multimodal fusion, and the key frequency band features are selected using the wavelet scale domain attention gating mechanism. A multi-source signal joint representation matrix is generated by tensor splicing. The goal of multimodal fusion is to integrate the spatiotemporal characteristics of the pulse coding with the frequency domain characteristics of the original signal. The multi-source signal joint representation matrix is input into the spatiotemporal alignment compression module, low-rank tensor decomposition and residual connection are used to eliminate sensor data drift, and finally the full-dimensional feature tensor of the base station is output; The full-dimensional feature tensor of the base station is input into the fault mode adversarial distillation module, and the normal operating condition manifold of the base station is reconstructed based on the variational autoencoder of physical constraints. The sudden fault and progressive aging features are separated by the gradient masked residual focusing algorithm, and the abnormal feature vector with the fault fingerprint is output. The variational autoencoder uses Laplace noise injection and frequency domain sparse constraints to forcibly distinguish between transient anomalies and steady-state degradation modes, including: Based on the full-dimensional feature tensor of the base station, a physically constrained variational autoencoder is constructed, and the latent space distribution is perturbed by Laplace noise injection to generate a normal operating manifold reconstruction vector that is resistant to overfitting; Performing residual calculation on the normal working condition manifold reconstruction vector and the original feature tensor, separating the steady-state baseline features using a frequency domain sparse constraint filter, and outputting a frequency domain decoupled residual feature matrix; The residual feature matrix is input into the gradient masking and focusing module, which masks the aging-related gradients through a dynamic threshold gating mechanism, strengthens the local gradient response of sudden faults, and generates abnormally sensitive feature vectors; Perform fault fingerprint identification on the abnormal sensitive feature vector, extract the time-frequency domain fault mode signature using convolutional sparse coding, and finally output an abnormal feature vector with a unique identifier; The fault propagation graph modeling is performed on the abnormal feature vector, and a graph-enhanced causal reasoning model is used to simulate the fault diffusion path between hardware components. The propagation weight is dynamically corrected through a gated graph attention mechanism, and a base station fault root cause topology graph is generated in combination with backpropagation credibility verification. The model jointly optimizes fault location accuracy through a time series reverse truncation algorithm and propagation path entropy constraints, including: Based on the abnormal feature vector, a graph structure of the hardware component connection topology is constructed, and the fault propagation probability weights between nodes are dynamically calculated based on the gated graph attention mechanism to generate an initial fault diffusion map; The graph structure of the hardware component connection topology is constructed based on the physical connection relationships between base station devices. Nodes represent hardware units, and edges represent physical or logical dependencies between components. The feature vector of each node is composed of the concatenation of the abnormal feature vector and the device operating parameters. The initial fault diffusion graph is input into the time series reverse truncation module, the back propagation path length is constrained by the causal mask matrix, the propagation path confidence is optimized in combination with the path entropy loss function, and the time series corrected fault propagation subgraph is output; Perform reinforcement learning optimization on the fault propagation subgraph, use the Q-learning algorithm to simulate the fault diffusion path selection strategy, and use Monte Carlo tree search to screen high-return root cause propagation links; ‌ The root cause propagation link is verified for credibility, low-probability pseudo-causal relationships are eliminated based on Bayesian hypothesis testing, and finally a base station fault root cause topology map is output; The base station fault root cause topology map is input into the resilient self-healing strategy generator, and a multi-objective repair decision tree is constructed based on a dynamic game framework driven by digital twins. Invalid repair paths are compressed through physically constrained adversarial strategy distillation technology, and federated reinforcement learning is used to coordinate the allocation of repair resources among multiple base stations. Finally, a set of base station facility self-healing strategies that meet real-time constraints is output. The dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update, including: Based on the base station fault root cause topology map, a digital twin-driven dynamic game framework is constructed. The initial repair strategy branches are generated through a multi-objective decision tree to simulate the feasibility of the strategy under the constraints of the base station's physical parameters. The initial repair strategy branch is input into the adversarial strategy distillation module, the effectiveness of the strategy is evaluated using the generative adversarial network discriminator, invalid repair paths are compressed through gradient reversal, and a streamlined strategy candidate set is output; Performing federated reinforcement learning collaborative optimization on the simplified strategy candidate set, using a distributed policy gradient algorithm to coordinate multi-base station resource allocation weights to generate a global optimal repair strategy vector; The global optimal repair strategy vector is input into the twin simulation verification platform, and the strategy parameter deviation is corrected through the mirror update mechanism constrained by physical equations, and finally a base station self-healing strategy set that meets real-time constraints is output.

3. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to claim 1 when executed.

4. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to claim 1 .

Images