Base station facility abnormity intelligent monitoring method and system

Abstract

The invention discloses a base station facility abnormity intelligent monitoring method and system, and the method comprises the steps: employing a multi-scale wavelet fusion pulse neural network to construct a dynamic signal perception topology according to the real-time data of a base station physical layer, and generating a space-time aligned base station full-dimension feature tensor; inputting the full-dimensional feature tensor of the base station into a fault mode confrontation distillation module, reconstructing a normal working condition manifold of the base station based on a physical constraint variational auto-encoder, and outputting an abnormal feature vector with a fault fingerprint identifier; fault propagation graph modeling is carried out on the abnormal feature vectors, and a base station fault root cause topological graph is generated in combination with back propagation credibility verification; and inputting the base station fault root cause topological graph into a toughness self-healing strategy generator, and finally outputting a base station facility self-healing strategy set meeting real-time constraint. By utilizing the embodiment of the invention, the accuracy and the real-time performance of abnormity monitoring can be improved, and the operation reliability and the communication service quality of the base station are improved.

Description

Technical Field

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

[0002] With the rapid development of modern communication technology, base station facilities, as a key component of the mobile communication network, undertake the important tasks of signal coverage and data transmission. The normal operation of the base station is crucial for ensuring the reliability of the network and the quality of service. However, during the long-term operation of the base station, it may be affected by various factors, such as equipment aging, external environmental changes, and sudden failures, resulting in a decline in its performance or even downtime. This not only affects the communication experience of users but also may lead to significant economic losses and potential safety hazards.

[0003] Traditional base station monitoring methods mainly rely on static data analysis and log auditing, and often suffer from problems such as long response time, insufficient accuracy, and inability to adapt to complex working conditions in real time. These methods usually cannot detect potential faults in a timely manner, resulting in an increase in the failure rate of the base station and an increase in maintenance costs. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent monitoring method and system for abnormal base station facilities to solve the deficiencies in the prior art, improve the accuracy and real-time performance of abnormal monitoring, and enhance the operational reliability of the base station and the quality of communication services.

[0005] An embodiment of the present application provides an intelligent monitoring method for abnormal base station facilities, and the method includes: 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 multi-source signal interference is eliminated through a phase synchronization time-frequency decomposition algorithm to generate a spatio-temporally aligned base station full-dimensional feature tensor. Among them, the pulse neural network introduces a pulse timing-dependent plasticity mechanism, and the signal features are adaptively enhanced through dynamic correction of synaptic weights; The base station full-dimensional feature tensor is input into a fault mode adversarial distillation module, and a variational autoencoder based on physical constraints is used to reconstruct the normal operating condition manifold of the base station. A sudden fault and a progressive aging feature are separated through a gradient masking residual focusing algorithm, and an abnormal feature vector with a fault fingerprint identifier is output. Among them, the variational autoencoder uses Laplace noise injection and frequency domain sparse constraints to force the distinction between transient anomalies and steady-state degradation modes; Model the fault propagation graph for the abnormal feature vector, use the graph-enhanced causal inference model to simulate the diffusion path of faults among hardware components, dynamically correct the propagation weights through the gated graph attention mechanism, and generate the base station fault root cause topology graph by combining backpropagation credibility verification. Among them, the model jointly optimizes the fault location accuracy through the temporal reverse truncation algorithm and the propagation path entropy constraint; Input the base station fault root cause topology graph into the resilience self-healing strategy generator, construct a multi-objective repair decision tree based on the digital twin-driven dynamic game framework, compress invalid repair paths through the adversarial strategy distillation technology with physical constraints, and use federated reinforcement learning to coordinate the repair resource allocation among multiple base stations. Finally, output the base station facility self-healing strategy set that meets the real-time constraint. Among them, the dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update.

[0006] Optionally, according to the real-time data of the base station physical layer, construct a dynamic signal perception topology using a multi-scale wavelet fusion pulse neural network, eliminate multi-source signal interference through the phase synchronization time-frequency decomposition algorithm, and generate a spatio-temporal aligned base station full-dimensional feature tensor. Among them, the pulse neural network introduces a pulse timing-dependent plasticity mechanism, and adaptively enhances signal features through dynamic correction of synaptic weights, including: According to the radio frequency signal time-frequency matrix and the hardware vibration mode spectrum, use the multi-scale wavelet ridge detection algorithm to perform time-frequency decomposition on the original signal, eliminate multipath effect interference through ridge line tracking, and generate a phase-synchronized time-frequency feature map; Input the time-frequency feature map, the energy consumption fluctuation curve, and the device operation log stream into the pulse neural network, dynamically adjust the synaptic weights based on the pulse timing-dependent plasticity mechanism, and enhance high-frequency abnormal signals through an adaptive pulse trigger function, and output a spatio-temporal correlated pulse sequence encoding; Perform multi-modal fusion on the pulse sequence encoding, use the wavelet scale domain attention gating mechanism to select key frequency band features, and generate a multi-source signal joint representation matrix through tensor splicing; Input the multi-source signal joint representation matrix into the spatio-temporal alignment compression module, and use low-rank tensor decomposition and residual connection to eliminate sensor data drift, and finally output the base station full-dimensional feature tensor.

[0007] Optionally, input the base station full-dimensional feature tensor into the fault mode adversarial distillation module, reconstruct the normal operating condition manifold of the base station based on the variational autoencoder with physical constraints, separate sudden faults and progressive aging features through the gradient masking residual focusing algorithm, and output an abnormal feature vector with a fault fingerprint identifier. Among them, the variational autoencoder uses Laplace noise injection and frequency domain sparsity constraints to force the distinction between transient anomalies and steady-state degradation modes, including: Construct a physically constrained variational autoencoder based on the full-dimensional feature tensor of the base station, and generate a manifold reconstruction vector of normal operating conditions that resists overfitting by injecting Laplace noise to perturb the latent space distribution; Calculate the residual between the manifold reconstruction vector of normal operating conditions and the original feature tensor, and use a frequency-domain sparse constraint filter to separate the steady-state baseline features, and output a frequency-domain decoupled residual feature matrix; Input the residual feature matrix into a gradient masking and focusing module, and shield the gradients related to aging through a dynamic threshold gating mechanism, and strengthen the local gradient response of sudden faults to generate an anomaly-sensitive feature vector; Perform fault fingerprint identification on the anomaly-sensitive feature vector, extract time-frequency domain fault mode signatures using convolutional sparse coding, and finally output an anomaly feature vector with a unique identifier.

[0008] Optionally, model the fault propagation graph for the anomaly feature vector, use a graph-enhanced causal inference model to simulate the diffusion path of faults between hardware components, dynamically correct the propagation weights through a gated graph attention mechanism, and generate a base station fault root cause topology graph by combining backpropagation credibility verification. Among them, the model jointly optimizes the fault location accuracy through a time-series reverse truncation algorithm and a propagation path entropy constraint, including: Construct a graph structure of the hardware component connection topology based on the anomaly feature vector, dynamically calculate the fault propagation probability weights between nodes based on the gated graph attention mechanism, and generate an initial fault diffusion map; Input the initial fault diffusion map into a time-series reverse truncation module, constrain the length of the backpropagation path through a causal mask matrix, and optimize the propagation path confidence by combining the path entropy loss function, and output a time-series corrected fault propagation subgraph; Perform reinforcement learning optimization on the fault propagation subgraph, use the Q-learning algorithm to simulate the fault diffusion path selection strategy, and screen high-return root cause propagation links through Monte Carlo tree search; Verify the credibility of the root cause propagation link, and eliminate low-probability pseudo-causal relationships based on Bayesian hypothesis testing, and finally output a base station fault root cause topology graph.

[0009] Optionally, input the base station fault root cause topology graph into a resilience self-healing strategy generator, construct a multi-objective repair decision tree based on a digital twin-driven dynamic game framework, compress invalid repair paths through a physically constrained adversarial strategy distillation technique, and use federated reinforcement learning to coordinate the repair resource allocation between multiple base stations, and finally output a set of base station facility self-healing strategies that meet real-time constraints. Among them, the dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update, including: Construct a digital twin-driven dynamic game framework according to the base station fault root cause topology graph, generate initial repair strategy branches through a multi-objective decision tree, and simulate the feasibility of strategies under the constraints of base station physical parameters; Input the initial repair strategy branches into an adversarial strategy distillation module, use a generative adversarial network discriminator to evaluate the effectiveness of the strategies, compress invalid repair paths through gradient reversal, and output a refined strategy candidate set; Perform collaborative optimization of the refined strategy candidate set through federated reinforcement learning, adopt a distributed policy gradient algorithm to coordinate the resource allocation weights of multiple base stations, and generate a globally optimal repair strategy vector; Input the globally optimal repair strategy vector into a twin simulation verification platform, correct the deviation of strategy parameters through a mirror update mechanism constrained by physical equations, and finally output a base station self-healing strategy set that meets real-time constraints.

[0010] Another embodiment of the present application provides an intelligent monitoring system for abnormal base station facilities, and the system includes: A construction module, configured to construct a dynamic signal perception topology using a multi-scale wavelet fusion pulse neural network according to real-time data of the base station physical layer, eliminate multi-source signal interference through a phase synchronization time-frequency decomposition algorithm, and generate a spatio-temporal aligned base station full-dimensional feature tensor. Among them, the pulse neural network introduces a pulse timing-dependent plasticity mechanism to adaptively enhance signal features through dynamic correction of synaptic weights; A reconstruction module, configured to input the base station full-dimensional feature tensor into a fault mode adversarial distillation module, reconstruct the normal operating condition manifold of the base station based on a variational autoencoder with physical constraints, separate sudden faults and progressive aging features through a gradient masking residual focusing algorithm, and output an abnormal feature vector with a fault fingerprint identifier. Among them, the variational autoencoder uses Laplace noise injection and frequency domain sparse constraints to force the distinction between transient anomalies and steady-state degradation modes; A correction module, configured to model the fault propagation graph for the abnormal feature vector, use a graph-enhanced causal inference model to simulate the diffusion path of faults between hardware components, dynamically correct the propagation weights through a gated graph attention mechanism, and generate a base station fault root cause topology graph in combination with backpropagation credibility verification. Among them, the model jointly optimizes the fault location accuracy through a temporal reverse truncation algorithm and a propagation path entropy constraint; An output module, configured to input the base station fault root cause topology graph into a resilient self-healing strategy generator, construct a multi-objective repair decision tree based on a digital twin-driven dynamic game framework, compress invalid repair paths through a physical constraint-based adversarial strategy distillation technique, and adopt federated reinforcement learning to coordinate the repair resource allocation between multiple base stations, and finally output a base station facility self-healing strategy set that meets real-time constraints. Among them, the dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update.

[0011] Another embodiment of the present application provides a storage medium in which a computer program is stored. Wherein, the computer program is configured to execute the method described in any one of the above when running.

[0012] Another embodiment of the present application provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the method described in any one of the above.

[0013] Compared with the prior art, an intelligent monitoring method for base station facility anomalies provided by the present invention uses multi-scale wavelet fusion pulse neural networks to construct a dynamic signal perception topology based on real-time data of the base station physical layer, generating a spatio-temporally 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 condition 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 graph by combining backpropagation credibility verification; inputting the base station fault root cause topology graph into a resilience self-healing strategy generator, and finally outputting a set of base station facility self-healing strategies that meet real-time constraints, thereby being able to improve the accuracy and real-time performance of anomaly monitoring, and enhancing the operating reliability of the base station and the quality of communication services. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a hardware structure block diagram of a computer terminal for an intelligent monitoring method for base station facility anomalies provided by an embodiment of the present invention; Figure 2 It is a schematic flow chart of an intelligent monitoring method for base station facility anomalies provided by an embodiment of the present invention; Figure 3 It is a schematic structural diagram of an intelligent monitoring system for base station facility anomalies provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] An embodiment of the present invention first provides an intelligent monitoring method for base station facility anomalies. This method can be applied to an electronic device, such as a computer terminal, specifically, a general computer, etc.

[0017] The following will take running on a computer terminal as an example for a detailed description. Figure 1 It is a hardware structure block diagram of a computer terminal for an intelligent monitoring method for base station facility anomalies provided by an embodiment of the present invention. As Figure 1As shown in the figure, the computer device includes a processor, a memory, and a network interface connected via a system bus. Among them, the memory may include a non-volatile storage medium and an internal memory.

[0018] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions. When the program instructions are executed, the processor can be made to execute any one of the abnormal intelligent monitoring methods for base station facilities.

[0019] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0020] 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 be made to execute any one of the abnormal intelligent monitoring methods for base station facilities.

[0021] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that Figure 1 the structure shown in the figure is only a block diagram of some structures 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 those shown in the figure, or combine some components, or have different component arrangements.

[0022] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0023] See Figure 2 , an embodiment of the present invention provides an abnormal intelligent monitoring method for base station facilities, which may include the following steps: S201. According to the real-time data of the base station physical layer, construct a dynamic signal perception topology using a multi-scale wavelet fusion pulse neural network, eliminate multi-source signal interference through a phase synchronization time-frequency decomposition algorithm, and generate a spatio-temporally aligned base station full-dimensional feature tensor. Among them, the pulse neural network introduces a pulse timing-dependent plasticity mechanism, and realizes the adaptive enhancement of signal features through dynamic correction of synaptic weights; Specifically, according to the time-frequency matrix of the radio frequency signal and the hardware vibration mode spectrum, the multi-scale wavelet ridge detection algorithm can be used to perform time-frequency decomposition on the original signal. By ridge line tracking, the multipath effect interference can be eliminated, and a time-frequency feature map with phase synchronization can be generated; The base station physical layer data includes the time-frequency matrix (sampling rate 1 GHz, time window length 10 ms) of the radio frequency signal (such as 5G NR signal frequency band 3.5 GHz) and the hardware vibration mode spectrum (acceleration sensor sampling rate 5 kHz, frequency bandwidth 0 - 2 kHz). The multi-scale wavelet ridge detection algorithm uses the Morlet wavelet basis function and covers different frequency band characteristics by adjusting the scale parameter (such as the scale sequence is 1, 2, 4, 8). For example, scale 1 corresponds to high-frequency noise (such as switching power supply interference), scale 4 corresponds to medium-frequency vibration (such as abnormal fan vibration), and scale 8 corresponds to low-frequency baseband signal fluctuations.

[0024] The ridge line tracking is realized through the following steps: Time-frequency energy focusing: Calculate the modulus maximum value of the wavelet coefficients for each scale to form candidate ridge lines. For example, in the time-frequency matrix of the radio frequency signal, the ridge lines of normal signals show continuous and smooth characteristics, while multipath interference (such as building reflection signals) will produce discontinuous ridge lines.

[0025] Phase consistency verification: Calculate the phase change rate (unit: radian / second) between adjacent time points. If the phase jump exceeds π / 2 (the threshold is adjustable), it is determined as an interference ridge line and eliminated. For example, during a certain detection, it is found that there is a phase mutation in the vibration signal at 150 Hz, which is confirmed by tracking to be caused by a loose screw, and this ridge line is marked as an effective feature.

[0026] Multipath interference elimination: Based on ridge line spatial density clustering (DBSCAN algorithm, neighborhood radius ε = 0.1, minimum sample number 5), separate the main path signal from the multipath reflection signal. For example, the ridge lines of the main path signal form a continuous banded structure in the time-frequency plane, while multipath interference appears as discrete point clusters.

[0027] The finally generated time-frequency feature map is a three-dimensional tensor (time × frequency × energy intensity), for example, with a size of 1000 × 256 × 3 (time points × frequency bins × energy / phase / coherence). By using a phase synchronization algorithm (such as PLL phase-locked loop technology) to align the clock deviations of different sensors (accuracy ±1 μs), the time-frequency correlation between the vibration spectrum and the radio frequency signal is ensured.

[0028] Input the time-frequency feature map, the energy consumption fluctuation curve, and the device operation log stream into the spiking neural network, dynamically adjust the synaptic weights based on the spiking timing-dependent plasticity mechanism, strengthen the high-frequency abnormal signals through the adaptive spiking trigger function, and output the spatio-temporal correlated spiking sequence encoding; 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).

[0029] Data Encoding: Time-Frequency Feature Map: The energy value of each frequency slot is mapped to the firing rate of the input neuron (0 - 200Hz).

[0030] Energy Consumption Fluctuation Curve: The current value (unit: A) is encoded into the pulse interval through Delta modulation (e.g., 10A → 50ms interval, 20A → 25ms interval).

[0031] Device Log Stream: The event type (such as "temperature alarm", "CPU overload") is encoded into a one-hot vector, and a pulse sequence is generated through a Poisson process.

[0032] Spike-Timing-Dependent Plasticity (STDP) Mechanism: Synaptic Weight Update Rule: If the presynaptic neuron spikes earlier than the postsynaptic neuron, the weight increases (Long-Term Potentiation, LTP); otherwise, it decreases (Long-Term Depression, LTD). The update amplitude Δw = 0.01×e^(-Δt / 20ms), where Δt is the spike time difference. For example, when a certain radio frequency signal anomaly (sudden increase in high-frequency energy) triggers the input layer neuron to spike earlier, the weight of the corresponding hidden layer neuron increases by 15%.

[0033] Adaptive Spike Triggering: Dynamically adjust the neuron threshold (baseline threshold V_th = 30mV). When detecting a high-frequency anomaly (such as the energy at 1kHz in the vibration spectrum continuously being higher than three times the baseline), the threshold drops to 25mV to enhance the spike firing rate.

[0034] Spatio-Temporal Association Encoding: The output layer adopts a population coding strategy. The spikes of the 256 neurons in the hidden layer are statistically counted within a time window (50ms) to generate a binary pulse sequence (e.g., "1011" indicates that there are spikes in the 1st, 3rd, and 4th time slots). For example, when a certain fan failure causes a periodic pulse cluster to appear at 800Hz in the vibration spectrum, the corresponding output encoding is a periodic "1001" pattern.

[0035] Perform multimodal fusion on the said pulse sequence encoding, use the wavelet scale domain attention gating mechanism to select key frequency band features, and generate a multi-source signal joint representation matrix through tensor splicing; The goal of multimodal fusion is to integrate the spatio-temporal characteristics of the pulse encoding and the frequency domain features of the original signal. The specific process is as follows: Wavelet Scale Domain Attention Gating: Scale Energy Calculation: Perform a discrete wavelet transform (DWT) on the time-frequency feature spectrogram, which is decomposed into 4 sub-bands (approximate coefficient A3, detail coefficients D1 - D3). Calculate the energy proportion of each sub-band (e.g., the D3 sub-band accounts for 60% of the energy in the 1 - 2 kHz frequency band).

[0036] Attention Weight Generation: Learn the correlation between each sub-band and the pulse code through a gated recurrent unit (GRU). For example, when the pulse code shows high-frequency anomalies, the GRU raises the attention weight of the D3 sub-band to 0.8 and suppresses the weight of the A3 sub-band to 0.2.

[0037] Gated Filtering: Perform soft-threshold filtering on the sub-bands with low weights (threshold = 0.1 × maximum coefficient value) to retain key features. For example, after a certain filtering, the noise coefficient of the D3 sub-band (<0.05) is removed, and the effective feature retention rate > 90%.

[0038] Tensor Concatenation and Dimensionality Reduction: 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).

[0039] Use non-negative matrix factorization (NMF) to reduce the dimension to 1000×256×2, retaining the main modes. For example, after decomposition, the first mode reflects the radio frequency - vibration coupling feature, and the second mode reflects the energy consumption - log event correlation.

[0040] Example Scenario: When the aging of the base station power module leads to an increase in energy consumption fluctuations, the pulse code shows dense pulses in the power-related neurons (e.g., numbers 50 - 70). The wavelet attention gate strengthens the vibration features of the D2 sub-band (500 Hz - 1 kHz), and finally the joint matrix highlights the power - vibration coupling anomaly.

[0041] Input the multi-source signal joint representation matrix into the spatio-temporal alignment compression module. Use low-rank tensor decomposition and residual connection to eliminate sensor data drift, and finally output the full-dimensional feature tensor of the base station.

[0042] The spatio-temporal alignment compression module solves the problems of time asynchrony and spatial redundancy of multi-sensor data. The core steps are as follows: Low-rank Tensor Decomposition: Use Tucker decomposition to decompose the joint representation matrix (1000×256×2) into a core tensor (50×50×2) and factor matrices (time factor 1000×50, spatial factor 256×50, modal factor 2×2).

[0043] Through iterative optimization by Alternating Least Squares (ALS), the objective function includes a low-rank constraint (rank = 50) and a sparse regularization term (λ = 0.01). For example, the decomposed temporal factor matrix reveals the diurnal cycle pattern of the base station load, and the spatial factor matrix identifies the hot spots of the antenna array.

[0044] Residual connection compensation: The residual between the original joint matrix and the reconstructed matrix (core tensor × factor matrix) is modeled by a bidirectional LSTM network for sensor drift. For example, the zero drift of the acceleration sensor (0.1g per month) is predicted and compensated by the LSTM.

[0045] The data after residual compensation is added to the low-rank reconstruction result to generate a drift-corrected feature tensor.

[0046] Full-dimensional feature integration: The temporal factor matrix (1000×50), spatial factor matrix (256×50), and modal factor matrix (2×2) are concatenated into a full-dimensional tensor (1000×256×2×50).

[0047] A deep autoencoder (DAE) is used to further compress it to 1000×256×8 (8-dimensional latent features), retaining the principal components with a variance contribution rate > 95%.

[0048] Output example: In a certain fault scenario, the full-dimensional tensor shows extremely high values at the 500th point in the temporal dimension (corresponding to the fault occurrence time), the 120 - 130th units in the spatial dimension (corresponding to the RF power amplifier module), and the 5 - 6th channels in the modal dimension, accurately characterizing the signal distortion caused by overheating of the power amplifier.

[0049] S202, input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module, reconstruct the normal operating condition manifold of the base station based on the variational autoencoder with physical constraints, separate the sudden fault and progressive aging features through the gradient masking residual focusing algorithm, and output an abnormal feature vector with fault fingerprint identification, where the variational autoencoder uses Laplace noise injection and frequency-domain sparse constraints to force the distinction between transient anomalies and steady-state degradation modes; Specifically, according to the full-dimensional feature tensor of the base station, a variational autoencoder with physical constraints can be constructed, and the potential space distribution is perturbed by Laplace noise injection to generate a reconstructed vector of the normal operating condition manifold that resists overfitting; The physically-constrained Variational Autoencoder (VAE) aims to learn the feature distribution of the normal operating conditions of the base station. 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 4-layer Convolutional Neural Network (CNN). The input is the full-dimensional feature tensor of the base station (dimension: time × frequency × sensor channels, e.g., 100×128×8), and the output is the mean and variance vectors (dimension: 64), which characterize the latent space distribution. The decoder consists of a 3-layer transposed convolution network that reconstructs the latent variables into the original feature dimensions.

[0050] The physical constraints are implemented in two ways: Laplace noise injection: When sampling in the latent space, Laplace-distributed noise (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 certain latent vector is [0.3, -0.2,...], after adding noise, it becomes [0.3 + 0.07, -0.2 - 0.03,...], forcing the model to learn a more robust feature representation.

[0051] Frequency-domain sparsity constraint: An L1 regularization term in the frequency domain with a weight λ = 0.5 is added to the loss function. Specifically, when implementing, the reconstructed signal output by the decoder is subjected to a Fast Fourier Transform (FFT), and the sum of the absolute values of the high-frequency components (>50Hz) is calculated to penalize unnecessary high-frequency components. For example, under normal operating conditions, the vibration signal of the base station mainly concentrates in the range of 0 - 30Hz, and abnormal signals may contain transient impact components above 100Hz. The sparse constraint suppresses the influence of the latter in the reconstruction.

[0052] During the training process, the Adam optimizer (learning rate 0.001, β1 = 0.9, β2 = 0.999) is used, and the batch size is set to 64. After 500 iterations, it converges. The reconstructed vector of the normal operating condition manifold (dimension: 64) will be used as the benchmark for subsequent anomaly detection. For example, the reconstructed vectors of a certain base station are clustered into a spherical distribution in the latent space, while the abnormal samples deviate from this distribution.

[0053] Calculate the residual between the reconstructed vector of the normal operating condition manifold and the original feature tensor, and use the frequency-domain sparse constraint filter to separate the steady-state baseline features, outputting a frequency-domain decoupled residual feature matrix; The residual calculation aims to extract the difference between the original features and the normal reconstructed features, highlighting the abnormal components. The specific process is as follows: Residual generation: Subtract the reconstructed tensor output by the decoder from the original feature tensor (100×128×8) element by element to obtain the residual tensor. For example, if the original value of the radio frequency signal power at a certain time point is 25dBm and the reconstructed value is 24.8dBm, then the residual is +0.2dBm.

[0054] Frequency-domain Sparse Filtering: Perform FFT on the residual tensor along the time dimension to generate a frequency-domain residual spectrum (dimension: frequency × time × channel). Design a sparse filter: High-pass Filter: Cutoff frequency is 10 Hz, attenuation slope is 24 dB / octave, filtering out low-frequency aging-related fluctuations (such as 0.1 Hz fluctuations caused by the slow decline of the cooling fan speed).

[0055] Transient Pulse Detection: Use wavelet transform modulus maxima detection to identify burst pulses with a duration < 50 ms (such as 100 Hz transient interference caused by a circuit short).

[0056] In the filtered residual feature matrix (dimension: 100 × 128 × 8), only the high-frequency burst components and transient pulses are retained. For example, at time point t = 30 s in a certain channel, a sudden increase in residual energy at 200 Hz (amplitude 0.5 dB) is detected, while the residuals of the steady-state aging characteristics are suppressed to below 0.02 dB.

[0057] Frequency-domain decoupling is achieved through independent component analysis (ICA), which decomposes the multi-channel residuals into statistically independent source signals. For example, independent components such as "power supply module voltage fluctuation", "RF power amplifier harmonic distortion", and "abnormal fan vibration" are separated, and each component corresponds to a subspace of the residual matrix (dimension: 100 × 128 × 3).

[0058] Input the residual feature matrix into the gradient masking focusing module, and mask the aging-related gradients through a dynamic threshold gating mechanism to strengthen the local gradient response of sudden faults and generate an anomaly-sensitive feature vector; The goal of the gradient masking focusing module is to distinguish between progressive aging (such as capacitor capacitance decay) and sudden faults (such as fuse blowing). Its core is to filter out irrelevant gradients through a dynamic threshold and amplify the anomaly-sensitive area.

[0059] Gradient Calculation: Use the Sobel operator to calculate the spatial gradient of the residual matrix (along the frequency and channel dimensions) to obtain a gradient magnitude matrix (dimension: 100 × 128 × 8). For example, the gradient magnitudes at a certain frequency bin for 10 consecutive time points are [0.1, 0.3, 0.8, 1.2, 0.7,...], indicating fluctuations caused by changes in the daytime workload.

[0060] Dynamic Threshold Gating: Aging Gradient Threshold: Calculate the moving average (MA) and standard deviation (STD) of the gradients within a sliding window (window size is 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, and gradients below this value are considered aging-related and masked.

[0061] Sudden failure enhancement: Non-linearly amplify the 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, and the gradient amplitude is increased to 0.5×1.86 = 0.93.

[0062] Local response normalization: Perform spatial local competition normalization (LRN) on the amplified gradients with a window size of 3×3 to suppress gradient redundancy in adjacent regions. For example, a peak gradient of 0.93 drops to 0.85 after normalization, while the surrounding gradients drop from 0.3 to 0.1.

[0063] In the processed gradient matrix (dimension: 100×128×8), only retain the high-gradient regions related to sudden failures. For example, it is detected that the gradient of a certain power module suddenly increases to 1.2 at t = 45s (corresponding to a 5% voltage drop), while the gradients related to aging (such as 0.25 at t = 10s) are completely suppressed.

[0064] Perform fault fingerprint identification on the anomaly-sensitive feature vectors, extract time-frequency domain fault mode signatures using convolutional sparse coding, and finally output anomaly feature vectors with unique identifiers.

[0065] Fault fingerprint identification aims to assign a unique code to each type of anomaly for subsequent root cause analysis. Use convolutional sparse coding (CSC) to extract interpretable fault mode signatures from time-frequency domain residuals.

[0066] Dictionary learning: In the offline training stage, use historical fault data (including 50 known fault types) to learn the dictionary: Set the dictionary size to 100 atoms (atom length 128, corresponding to the frequency dimension); Use the K-SVD algorithm for iterative optimization with a sparse constraint term λ = 0.1 and 1000 iterations.

[0067] For example, atom 1 may capture the "RF power amplifier harmonic distortion" mode (center frequency 2.4GHz, bandwidth 50MHz), and atom 2 corresponds to the "power surge" mode (transient pulse width 10ms).

[0068] Online coding: Perform sparse coding on the anomaly-sensitive feature vectors generated in real-time: The input is the residual matrix (100×128×8) after gradient focusing, which is processed separately after being separated by channels; Use the OMP (Orthogonal Matching Pursuit) algorithm to solve the sparse coefficients, with the non-zero coefficient limit set to 5 per sample; 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 VSWR anomaly" and "clock jitter".

[0069] Fingerprint generation: The sparse coefficient vector (dimension 100) is converted into a 128-bit binary fingerprint through hash encoding. For example, the positions of atoms with absolute coefficient values > 0.2 are set to 1, and the rest are set to 0, generating a hexadecimal identifier such as "0x1A3F...". At the same time, additional time-frequency domain statistical features (such as peak frequency, pulse duration) are added to form the final anomaly feature vector (dimension 256).

[0070] Uniqueness verification: The fingerprint differences of different faults are evaluated through cosine similarity. For example, the fingerprint similarity of two power supply faults is 0.3, while the similarity between a power supply fault and an RF fault is < 0.1, ensuring distinguishability.

[0071] S203, perform fault propagation graph modeling on the anomaly feature vector, use the graph-enhanced causal inference model to simulate the diffusion path of faults between hardware components, dynamically correct the propagation weights through the gated graph attention mechanism, and generate the root cause topology graph of base station faults in combination with backpropagation credibility verification. Among them, the model jointly optimizes the fault location accuracy through the temporal reverse truncation algorithm and the propagation path entropy constraint; Specifically, according to the anomaly feature vector, a graph structure of the hardware component connection topology can be 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 through the physical connection relationship of the base station equipment. The nodes represent hardware units (such as RF modules, power supply units, baseband processors), and the edges represent the physical or logical dependency relationships between components (such as the power supply unit supplying power to the RF module). The feature vector of each node is composed of the anomaly feature vector (dimension 128) and the device operation parameters (temperature, voltage, signal quality) spliced together to form a 256-dimensional node feature.

[0072] The core of the Gated Graph Attention Network (GGAT) is to dynamically adjust the attention weights through a gated unit. The specific implementation is as follows: Attention coefficient calculation: For each pair of adjacent nodes (such as an RF module and a baseband processor), the node features are projected into query (Query) and key (Key) vectors through a shared weight matrix (dimension 256×64), and the dot product attention score is calculated. For example, the dot product score between the query vector of the RF module and the key vector of the baseband processor is 0.8, while the score with the power supply unit is 0.3, reflecting a stronger functional dependency between the former.

[0073] Gating mechanism: The GRU (Gated Recurrent Unit) is used to dynamically filter the attention scores. The update gate of the GRU determines the propagation probability of retaining or ignoring an edge based on the current state of the node and the aggregated features of its neighbors. For example, when the temperature of the power supply unit exceeds the threshold (such as 80 °C), the update gate reduces the connection weight between it and the RF module from 0.3 to 0.1.

[0074] 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 from the RF module to the baseband processor is 0.6, and the weight to the power supply unit is 0.4.

[0075] The initially generated fault diffusion graph is stored in the form of an adjacency matrix, and the matrix elements represent the probability of the fault propagating from node i to node j (for example, the probability from the RF module to the baseband processor is 0.7, and the probability from the RF module to the power supply unit is 0.2).

[0076] Input the initially generated fault diffusion graph into the temporal reverse truncation module. Constrain the length of the reverse propagation path through the causal mask matrix, and optimize the confidence of the propagation path in combination with the path entropy loss function to output a temporally corrected fault propagation subgraph; The goal of the temporal reverse truncation module is to limit the temporal depth of the fault propagation path and avoid computational explosion caused by infinite tracing of historical states. The causal mask matrix is an upper triangular matrix (1 above and including the main diagonal, and 0 for the rest), which is used to force the propagation path to only trace back a limited number of steps forward (such as 3 steps). For example, if the current fault occurs at time step t, only the propagation paths within the time window from t - 3 to t are allowed to be analyzed.

[0077] The path entropy loss function is used to quantify the uncertainty of the propagation path. The specific calculation method is as follows: Calculate the probability distribution (such as path probabilities 0.4, 0.3, 0.3) for each path (such as RF module → power supply unit → baseband processor); Calculate the Shannon Entropy of this distribution. The higher the entropy value, the greater the uncertainty of the path; Add the entropy value as a loss term to the total loss function to force the model to converge to a high-confidence propagation path. For example, the entropy value of a certain path drops from 1.2 (high uncertainty) to 0.5 (low uncertainty), indicating that the model is more confident in its causal relationship.

[0078] The optimization process uses the Adam optimizer (learning rate 0.001, β1 = 0.9, β2 = 0.999) to update the weight parameters of GGAT through backpropagation. During training, the path length is limited to a maximum of 3 hops (e.g., for a fault from the RF module → baseband processor → antenna array, there are 2 hops in total), and paths exceeding this length are filtered by the masking matrix. For example, a 4-hop path (RF → power supply → fan → chassis) will be truncated to 3 hops (RF → power supply → fan).

[0079] After optimization, the fault propagation subgraph of timing correction only retains high-confidence paths (e.g., probability > 0.5). For example, 10 paths in the initial graph may be compressed into 3 critical paths, and the entropy loss value of each path is less than 0.3.

[0080] Perform reinforcement learning optimization on the fault propagation subgraph, use the Q-learning algorithm to simulate the fault diffusion path selection strategy, and screen high-return root cause propagation links through Monte Carlo tree search; The reinforcement learning framework models fault propagation as a Markov decision process (MDP): State: The current node where the fault is located (such as the power supply unit) and its historical path (such as the past 2-hop path); Action: Select the next propagation node (such as from the power supply unit to the baseband processor); Reward: The weighted sum of the diagnostic accuracy of the path (verified based on historical fault data) and the repair cost (such as the cost of replacing components), with a weight of 7:3.

[0081] The parameter settings of the Q-learning algorithm are as follows: Q-table: Since the state space is large (possibly containing tens of thousands of node combinations), use a deep Q-network (DQN) to approximate the Q function, and the network structure is a 3-layer fully connected network (256 → 128 → 64); Exploration strategy: ε-greedy strategy, initial ε = 0.3 (30% probability of randomly selecting an action), and it decays by 10% every 1000 training steps; Discount factor γ: 0.9, emphasizing the importance of recent rewards.

[0082] Monte Carlo tree search (MCTS) is used to expand high-return paths: Selection: Starting from the root node (the initial fault point), select child nodes based on the UCB1 formula (exploration coefficient c = 1.4) until reaching the leaf node; Expansion: If the leaf node is not fully explored, add new child nodes (such as adding a new propagation edge); Simulation: Randomly simulate the propagation of faults to the termination state (such as complete repair or system crash), and calculate the cumulative reward; Backpropagation: Update the access times and reward values of path nodes in the reverse direction based on the simulation results.

[0083] For example, after a certain MCTS expands at the power supply unit node, the simulation finds that the cumulative reward for the path "Power Supply → Baseband Processor → RF Module" is 85 points (diagnostic accuracy rate of 90%, repair cost of 500), while the reward for the path "Power Supply → Fan → Chassis" is 60 points (accuracy rate of 70%, cost of 500), and the reward for the path "Power Supply → Fan → Chassis" is 60 points (accuracy rate of 70%, cost of 200). Finally, the former is selected as the high-return root cause link.

[0084] Verify the credibility of the root cause propagation link, eliminate low-probability pseudo-causal relationships based on Bayesian hypothesis testing, and finally output the base station fault root cause topology map.

[0085] Bayesian hypothesis testing verifies the statistical significance of causal relationships by calculating the posterior probability. The specific steps include: Prior probability setting: Based on the historical fault database, statistically calculate the prior occurrence frequency of each path. For example, the historical occurrence probability of the path "Power Supply → Baseband Processor" is 0.4; Likelihood calculation: Calculate the likelihood probability according to the matching degree between the current fault characteristics (such as voltage drop, 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" is increased to 0.7; Posterior probability calculation: Apply Bayes' formula to update the path probability. For example, posterior probability = (likelihood × prior) / evidence factor; Significance test: Set the significance level α = 0.05. If the posterior probability of a certain path is lower than the threshold (such as <0.05), it is determined as a pseudo-causal relationship and eliminated.

[0086] 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 the path "Power Supply → Baseband Processor" is 0.62 and is retained.

[0087] The finally output base station fault root cause topology map is presented in the form of a directed graph. The nodes are marked with the fault probability (such as Baseband Processor: 85%), and the edges are marked with the propagation probability and timing information (such as Power Supply → Baseband Processor: probability 0.7, delay 3 seconds). This map can directly guide the maintenance personnel to prioritize the inspection of high-probability fault points and trigger the automated repair process in combination with the self-healing strategy generator.

[0088] S204. Input the base station fault root cause topology graph into the resilience self-healing strategy generator, construct a multi-objective repair decision tree based on the digital twin-driven dynamic game framework, compress invalid repair paths through adversarial policy distillation technology with physical constraints, and use federated reinforcement learning to coordinate the allocation of repair resources among multiple base stations. Finally, output a set of base station facility self-healing strategies that meet real-time constraints. Among them, the dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update.

[0089] Specifically, a digital twin-driven dynamic game framework can be constructed according to the base station fault root cause topology graph, and an initial repair strategy branch can be generated through a multi-objective decision tree to simulate the policy feasibility under the physical parameter constraints of the base station. The digital twin-driven dynamic game framework maps the physical world state in real time by establishing a virtual mirror of the base station hardware components. First, based on the fault root cause topology graph, map the fault nodes (such as radio frequency units, power modules) to the 3D model of the digital twin body and associate their physical parameters (such as temperature threshold, voltage fluctuation range). For example, the fault root cause topology of a certain base station radio frequency unit shows that "the cooling fan fails → the temperature exceeds the standard → the signal is distorted". The twin model will simulate the dynamic process of the temperature rising from 25°C to 85°C after the fan stops rotating and predict the signal distortion time window (such as 30 seconds later).

[0090] The construction of the multi-objective decision tree uses NSGA-III (the third-generation non-dominated sorting genetic algorithm), and the optimization objectives include: Repair time (target < 5 minutes, weight 0.4); Resource consumption (cost budget < 1000 yuan, weight 0.3); Probability of fault recurrence (target < 5%, weight 0.3).

[0091] The branch generation rule of the decision tree is based on physical constraints: Hardware constraints: For example, when replacing the power module, a power-off operation is required, and a 2-minute power-off time needs to be reserved when generating the branch. Logical constraints: 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).

[0092] Example of the initial repair strategy branch: Branch 1: Immediately restart the cooling fan (cost 0 yuan, time 30 seconds, recurrence probability 40%); Branch 2: Replace the fan and upgrade the firmware (cost 500 yuan, time 4 minutes, recurrence probability 5%); Branch 3: Enable the standby radio frequency unit and isolate the faulty module (cost 800 yuan, time 2 minutes, recurrence probability 10%).

[0093] Verify the physical feasibility of each branch through twin simulation. For example, Branch 1 could not be restarted due to fan aging in the simulation and was marked as an invalid path; the temperature control simulation of Branch 2 showed that it could be stabilized at 45 °C, meeting the physical constraints.

[0094] Input the initial repair strategy branches into the adversarial strategy distillation module, use the discriminator of the generative adversarial network to evaluate the effectiveness of the strategy, compress the invalid repair paths through gradient reversal, and output a refined strategy candidate set; The adversarial strategy distillation module consists of a generator and a discriminator: Generator: A 4-layer fully connected network (input dimension 256, hidden layers 512 / 256 / 128, output dimension 64), responsible for generating the latent representation of the repair strategy; Discriminator: A 3-layer convolutional network (kernel size 3×3, stride 1, number of channels 32 / 64 / 128), outputting a strategy effectiveness score (0 - 1).

[0095] Training process: Input to the generator: The feature vector of the initial strategy branch (including parameters such as repair time, cost, recurrence probability, etc.); Discriminator training: Use historical successful strategies (positive samples) and artificially constructed invalid strategies (negative samples) for supervised learning, learning rate 0.001, batch size 32; Gradient reversal: In the generator training stage, multiply the gradient of the invalid strategy by a negative coefficient (such as -0.5) during backpropagation to force the generator to avoid the invalid area.

[0096] For example, when the generator proposes "only clean the fan blades (cost 50 yuan)", the discriminator determines based on historical data that the recurrence probability of this strategy in a high-temperature environment reaches 60% and gives a low score of 0.2. The gradient reversal mechanism causes the generator to reduce the generation probability of such strategies.

[0097] Strategy compression: Elimination of invalid paths: Directly delete strategies with a score < 0.3; Merge of similar strategies: Use the DBSCAN clustering algorithm (neighborhood radius 0.2, minimum number of samples 3) to merge strategies with a cost difference < 100 yuan and a time difference < 1 minute.

[0098] Finally, the initial 20 strategies are compressed to 5 candidate strategies, for example: Candidate strategy A: Replace the fan + firmware upgrade (score 0.92); Candidate strategy B: Enable the standby module + remote diagnosis (score 0.85).

[0099] Perform collaborative optimization of the reduced policy candidate set through federated reinforcement learning, and use the distributed policy gradient algorithm to coordinate the resource allocation weights of multiple base stations to generate a globally optimal repair policy vector; The Federated Reinforcement Learning (FRL) framework consists of a central server and multiple base station nodes (such as 10 base stations). Each base station locally trains a policy gradient model and regularly uploads the model parameters to the central server for aggregation.

[0100] The distributed policy gradient algorithm uses PPO (Proximal Policy Optimization): State space: Base station failure type (encoded as a 4D vector), resource inventory (number of spare parts, remaining budget); Action space: Select candidate policies (5 discrete actions); Reward function: R = 0.5×(1 - repair time / 300s) + 0.3×(1 - cost / 1000 yuan) + 0.2×(1 - recurrence probability).

[0101] Federated collaboration process: Local training: Each base station trains a PPO model based on local historical data (such as 1000 fault records), iterates 10 rounds, and the learning rate is 0.0002; Parameter aggregation: The central server uses the FedAvg algorithm to perform weighted averaging on the policy network parameters of 10 base stations (weights are allocated according to the base station failure frequency); Policy distribution: The updated global model is sent to each base station for the next round of training.

[0102] Resource allocation optimization: Conflict resolution: When multiple base stations request scarce resources (such as a specific model of fan) simultaneously, use the auction mechanism (Vickrey-Clarke-Groves rule) for allocation; Dynamic priority: Adjust the resource scheduling weights according to business criticality (such as the priority of 5G core network base stations is higher than that of ordinary base stations).

[0103] After 5 rounds of federated training, the average reward of the global policy on the test set is increased from 0.65 to 0.82. An example of the optimal policy vector is as follows: Policy vector: [Action index 2 (enable standby module), resource weight 0.7, latency constraint 150 seconds].

[0104] Input the global optimal repair strategy vector into the twin simulation verification platform, correct the policy parameter deviation through the mirror update mechanism constrained by physical equations, and finally output a set of base station self-healing strategies that meet the real-time constraints.

[0105] The twin simulation platform integrates a multi-physics field simulation engine, including: Thermodynamic simulation: Calculate the device temperature distribution based on Fourier's law; Electromagnetic simulation: Use the FDTD (finite difference time domain) method to simulate signal attenuation; Mechanical stress simulation: Predict the component life through finite element analysis.

[0106] The mirror update mechanism dynamically corrects the policy parameters by comparing the deviation between the simulation results and physical constraints: Parameter deviation detection: For example, the policy expects the temperature to be 45°C after replacing the fan, but the simulation shows that it is actually 50°C, with a deviation of 5°C; Backpropagation correction: Adjust the parameters along the gradient descent direction of the policy network (such as increasing the heat sink area or extending the fan startup time); Constraint strengthening: Apply a penalty gradient (learning rate amplified by 10 times) to the policy that violates the hard constraint (such as the maximum temperature > 85°C).

[0107] Real-time guarantee: Lightweight simulation: Adopt a reduced-order model (ROM) to compress the full-scale simulation time from 30 minutes to 2 minutes; Edge computing deployment: Deploy a simulation container (Docker) locally at the base station and use GPU acceleration (NVIDIA T4) to achieve real-time feedback.

[0108] The finally output self-healing strategy set includes: Emergency strategy: For high-priority faults (such as power short circuit), the response time < 1 minute; Standard strategy: Conventional hardware replacement, time < 5 minutes; Preventive strategy: Proactive maintenance based on aging prediction, triggered 24 hours in advance.

[0109] Example: The execution process of the self-healing strategy for a power module failure of a base station: Step 1: The digital twin detects an abnormal input voltage (fluctuation ±15%); Step 2: Root cause analysis locks "voltage regulator aging"; Step 3: The federated policy selects "switch to the standby power supply + dispatch a repair work order"; Step 4: Simulate and verify the standby power supply switching delay (measured 45 seconds, meeting the < 1 minute constraint); Step 5: The strategy takes effect, the base station resumes normal operation, and the data is uploaded to the management platform.

[0110] It can be seen that, based on the real-time data of the base station physical layer, a multi-scale wavelet fusion spiking neural network is used to construct a dynamic signal perception topology, generating a spatio-temporally 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 a variational autoencoder based on physical constraints is used to reconstruct the normal operating condition manifold of the base station, outputting 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 by combining backpropagation credibility verification; the base station fault root cause topology graph is input into a resilience self-healing strategy generator, and finally a set of self-healing strategies for base station facilities that meet real-time constraints is output, so as to improve the accuracy and real-time performance of anomaly monitoring, and enhance the operational reliability of the base station and the quality of communication services.

[0111] Another embodiment of the present invention provides an intelligent anomaly monitoring system for base station facilities. Refer to Figure 3 , the system may include: A construction module 301, configured to construct a dynamic signal perception topology according to the real-time data of the base station physical layer by using a multi-scale wavelet fusion spiking neural network, eliminate multi-source signal interference through a phase synchronization time-frequency decomposition algorithm, and generate a spatio-temporally aligned full-dimensional feature tensor of the base station. Among them, the spiking neural network introduces a pulse timing-dependent plasticity mechanism, and realizes the adaptive enhancement of signal features through dynamic correction of synaptic weights; A reconstruction module 302, configured to input the full-dimensional feature tensor of the base station into a fault mode adversarial distillation module, reconstruct the normal operating condition manifold of the base station by using a variational autoencoder based on physical constraints, separate sudden faults and progressive aging features through a gradient masking residual focusing algorithm, and output an abnormal feature vector with a fault fingerprint identifier. Among them, the variational autoencoder uses Laplace noise injection and frequency-domain sparse constraints to force the distinction between transient anomalies and steady-state degradation modes; A correction module 303, configured to model a fault propagation graph for the abnormal feature vector, simulate the diffusion path of faults between hardware components by using a graph-enhanced causal inference model, dynamically correct the propagation weights through a gated graph attention mechanism, and generate a base station fault root cause topology graph by combining backpropagation credibility verification. Among them, the model jointly optimizes the fault location accuracy through a time-sequence reverse truncation algorithm and a propagation path entropy constraint; An output module 304, configured to input the base station fault root cause topology graph into a resilience self-healing strategy generator, construct a multi-objective repair decision tree based on a digital twin-driven dynamic game framework, compress invalid repair paths through a physically constrained adversarial policy distillation technique, and use federated reinforcement learning to coordinate the repair resource allocation among multiple base stations, and finally output a set of self-healing strategies for base station facilities that meet real-time constraints. Among them, the dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update.

[0112] It can be seen that, according to the real-time data of the base station physical layer, a dynamic signal perception topology is constructed by using a multi-scale wavelet fusion spiking neural network to generate a spatio-temporally 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 is used to reconstruct the normal operating condition manifold of the base station, and an abnormal feature vector with a fault fingerprint identifier is output; a fault propagation graph is modeled for the abnormal feature vector, and a base station fault root cause topology graph is generated by combining backpropagation credibility verification; the base station fault root cause topology graph is input into the resilient self-healing strategy generator, and finally a set of self-healing strategies for the base station facilities that meet the real-time constraint is output, so as to improve the accuracy and real-time performance of anomaly monitoring, and enhance the operating reliability and communication service quality of the base station.

[0113] An embodiment of the present invention also provides a storage medium, in which a computer program is stored, and the computer program is configured to execute the steps in any one of the above method embodiments when running.

[0114] Specifically, in this embodiment, the above storage medium may be configured to store a computer program for executing the following steps: S201, according to the real-time data of the base station physical layer, a dynamic signal perception topology is constructed by using a multi-scale wavelet fusion spiking neural network, and multi-source signal interference is eliminated through a phase synchronization time-frequency decomposition algorithm to generate a spatio-temporally aligned full-dimensional feature tensor of the base station. The spiking neural network introduces a pulse timing-dependent plasticity mechanism, and realizes the adaptive enhancement of signal features through dynamic correction of synaptic weights; ‌ S202, input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module, and reconstruct the normal operating condition manifold of the base station by using a variational autoencoder based on physical constraints. The sudden fault and progressive aging features are separated through a gradient masking residual focusing algorithm, and an abnormal feature vector with a fault fingerprint identifier is output. The variational autoencoder adopts Laplace noise injection and frequency domain sparsity constraints to force the distinction between transient anomalies and steady-state degradation modes; S203, model a fault propagation graph for the abnormal feature vector, use a graph-enhanced causal inference model to simulate the diffusion path of faults between hardware components, dynamically correct the propagation weights through a gated graph attention mechanism, and generate a base station fault root cause topology graph by combining backpropagation credibility verification. The model jointly optimizes the fault location accuracy through a time series reverse truncation algorithm and a propagation path entropy constraint; S204. Input the base station fault root cause topology graph into the resilient self-healing strategy generator, construct a multi-objective repair decision tree based on the digital twin-driven dynamic game framework, compress invalid repair paths through the adversarial strategy distillation technology with physical constraints, and use federated reinforcement learning to coordinate the repair resource allocation among multiple base stations. Finally, output a set of self-healing strategies for base station facilities that meet real-time constraints. Among them, the dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update.

[0115] 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, generating a spatio-temporally 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 with physical constraints is used to reconstruct the normal operating condition manifold of the base station, outputting an abnormal feature vector with a fault fingerprint identifier; the fault propagation graph is modeled for the abnormal feature vector, and the base station fault root cause topology graph is generated by combining the backpropagation credibility verification; the base station fault root cause topology graph is input into the resilient self-healing strategy generator, and finally a set of self-healing strategies for base station facilities that meet real-time constraints is output, thereby improving the accuracy and real-time performance of abnormal monitoring and enhancing the operational reliability and communication service quality of the base station.

[0116] An embodiment of the present invention also provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0117] Specifically, the above electronic device may further include a transmission device and an input / output device. Among them, the transmission device is connected to the above processor, and the input / output device is connected to the above processor.

[0118] Specifically, in this embodiment, the above processor may be configured to execute the following steps through a computer program: S201. According to the real-time data of the base station physical layer, use a multi-scale wavelet fusion pulse neural network to construct a dynamic signal perception topology, eliminate multi-source signal interference through a phase synchronization time-frequency decomposition algorithm, and generate a spatio-temporally aligned full-dimensional feature tensor of the base station. Among them, the pulse neural network introduces a pulse timing-dependent plasticity mechanism, and realizes the adaptive enhancement of signal features through dynamic correction of synaptic weights. ‌ S202. Input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module. Reconstruct the normal operating condition manifold of the base station based on the variational autoencoder with physical constraints, and separate the sudden fault and progressive aging features through the gradient masking residual focusing algorithm, and output an abnormal feature vector with a fault fingerprint identifier. Among them, the variational autoencoder adopts Laplace noise injection and frequency domain sparsity constraints to force the distinction between transient anomalies and steady-state degradation modes; S203. Model the fault propagation graph for the abnormal feature vector. Use the graph-enhanced causal inference model to simulate the diffusion path of faults among hardware components, dynamically correct the propagation weights through the gated graph attention mechanism, and generate the base station fault root cause topology graph in combination with the backpropagation credibility verification. Among them, the model jointly optimizes the fault location accuracy through the temporal reverse truncation algorithm and the propagation path entropy constraint; S204. Input the base station fault root cause topology graph into the resilient self-healing strategy generator. Construct a multi-objective repair decision tree based on the digital twin-driven dynamic game framework, compress the invalid repair paths through the adversarial strategy distillation technology with physical constraints, and use federated reinforcement learning to coordinate the repair resource allocation among multiple base stations. Finally, output a set of self-healing strategies for base station facilities that meet the real-time constraints. Among them, the dynamic game framework jointly guarantees the physical feasibility of the repair strategy through the fault scenario twin simulation and the policy gradient mirror update.

[0119] It can be seen that according to the real-time data of the base station physical layer, a dynamic signal perception topology is constructed by using a multi-scale wavelet fusion pulse neural network to generate a full-dimensional feature tensor of the base station with spatio-temporal alignment; input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module, reconstruct the normal operating condition manifold of the base station based on the variational autoencoder with physical constraints, and output an abnormal feature vector with a fault fingerprint identifier; model the fault propagation graph for the abnormal feature vector, and generate the base station fault root cause topology graph in combination with the backpropagation credibility verification; input the base station fault root cause topology graph into the resilient self-healing strategy generator, and finally output a set of self-healing strategies for base station facilities that meet the real-time constraints, so as to improve the accuracy and real-time performance of anomaly monitoring, and enhance the operating reliability and communication service quality of the base station.

[0120] The above has detailed the structure, features and function effects of the present invention according to the illustrated embodiments. The above is only the preferred embodiment of the present invention, but the present invention is not limited to the scope defined by the drawings. Any changes made according to the concept of the present invention, or modified into equivalent embodiments with equivalent changes, still within the spirit covered by the specification and drawings, shall be within the protection scope of the present invention.

Claims

1. An intelligent monitoring method for abnormal base station facilities, characterized in that, The method includes: According to the real-time data of the base station physical layer, a multi-scale wavelet fusion spiking neural network is used to construct a dynamic signal perception topology. The multi-source signal interference is eliminated through a phase synchronization time-frequency decomposition algorithm, and a spatio-temporal aligned full-dimensional feature tensor of the base station is generated. Among them, the spiking neural network introduces a spike timing-dependent plasticity mechanism, and the signal features are adaptively enhanced through dynamic synaptic weight correction; Input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module. The variational autoencoder based on physical constraints reconstructs the normal operating condition manifold of the base station. The sudden fault and progressive aging features are separated through a gradient masking residual focusing algorithm, and an abnormal feature vector with a fault fingerprint identifier is output. Among them, the variational autoencoder adopts Laplace noise injection and frequency-domain sparse constraints to force the distinction between transient anomalies and steady-state degradation modes; Model the fault propagation graph for the abnormal feature vector. The graph-enhanced causal inference model is used to simulate the diffusion path of faults between hardware components. The propagation weight is dynamically corrected through the gated graph attention mechanism, and the root cause topology graph of the base station fault is generated by combining the backpropagation credibility verification. Among them, the model jointly optimizes the fault location accuracy through the time-series reverse truncation algorithm and the propagation path entropy constraint; Input the root cause topology graph of the base station fault into the resilience self-healing strategy generator. A multi-objective repair decision tree is constructed based on the digital twin-driven dynamic game framework. The invalid repair paths are compressed through the physical constraint-based adversarial strategy distillation technology, and the federal reinforcement learning is used to coordinate the repair resource allocation among multiple base stations. Finally, a set of self-healing strategies for base station facilities that meet the real-time constraint is output. Among them, the dynamic game framework jointly guarantees the physical feasibility of the repair strategy through the fault scenario twin simulation and the policy gradient mirror update; 2. The method according to claim 1, wherein The step of using a multi-scale wavelet fusion spiking neural network to construct a dynamic signal perception topology according to the real-time data of the base station physical layer, eliminating multi-source signal interference through a phase synchronization time-frequency decomposition algorithm, and generating a spatio-temporal aligned full-dimensional feature tensor of the base station. Among them, the spiking neural network introduces a spike timing-dependent plasticity mechanism, and the signal features are adaptively enhanced through dynamic synaptic weight correction, includes: According to the time-frequency matrix of the radio frequency signal and the hardware vibration modal spectrum, the multi-scale wavelet ridge detection algorithm is used to perform time-frequency decomposition on the original signal, and the multipath effect interference is eliminated through ridge line tracking, and a phase synchronization time-frequency feature map is generated; Input the time-frequency feature map, the energy consumption fluctuation curve, and the device operation log stream into the spiking neural network. Based on the spike timing-dependent plasticity mechanism, the synaptic weight is dynamically adjusted, and the high-frequency abnormal signal is enhanced through the adaptive spike trigger function, and a spatio-temporal correlated pulse sequence encoding is output; Perform multi-modal fusion on the pulse sequence encoding. The key frequency band features are selected by using the wavelet scale domain attention gating mechanism, and a multi-source signal joint representation matrix is generated through tensor splicing; Input the multi-source signal joint representation matrix into the spatio-temporal alignment compression module. Low-rank tensor decomposition and residual connection are used to eliminate the sensor data drift, and finally the full-dimensional feature tensor of the base station is output.

3. The method according to claim 2, characterized in that, Input the full-dimensional feature tensor of the base station into the fault mode adversarial distillation module, reconstruct the normal operating condition manifold of the base station based on the variational autoencoder with physical constraints, separate the sudden fault and progressive aging features through the gradient masking residual focusing algorithm, and output an abnormal feature vector with a fault fingerprint identifier. Among them, the variational autoencoder adopts Laplace noise injection and frequency-domain sparse constraints to force the distinction between transient anomalies and steady-state degradation modes, including: According to the full-dimensional feature tensor of the base station, construct a variational autoencoder with physical constraints, and generate a reconstruction vector of the normal operating condition manifold that resists overfitting by injecting Laplace noise to perturb the latent space distribution; Calculate the residual between the reconstruction vector of the normal operating condition manifold and the original feature tensor, and use a frequency-domain sparse constraint filter to separate the steady-state baseline features, and output a frequency-domain decoupled residual feature matrix; Input the residual feature matrix into the gradient masking focusing module, shield the aging-related gradients through a dynamic threshold gating mechanism, strengthen the local gradient response of sudden faults, and generate an anomaly-sensitive feature vector; Perform fault fingerprint identification on the anomaly-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.

4. The method according to claim 3, wherein Model the fault propagation graph for the abnormal feature vector, adopt a graph-enhanced causal inference model to simulate the diffusion path of faults between hardware components, dynamically correct the propagation weights through a gated graph attention mechanism, and generate a base station fault root cause topology graph by combining backpropagation credibility verification. Among them, the model jointly optimizes the fault location accuracy through a temporal reverse truncation algorithm and a propagation path entropy constraint, including: According to the abnormal feature vector, construct a graph structure of the hardware component connection topology, and dynamically calculate the fault propagation probability weights between nodes based on the gated graph attention mechanism to generate an initial fault diffusion map; Input the initial fault diffusion map into the temporal reverse truncation module, constrain the length of the backpropagation path through a causal mask matrix, and optimize the propagation path confidence by combining the path entropy loss function, and output a temporally corrected fault propagation subgraph; Perform reinforcement learning optimization on the fault propagation subgraph, adopt the Q-learning algorithm to simulate the fault diffusion path selection strategy, and screen high-return root cause propagation links through Monte Carlo tree search; Verify the credibility of the root cause propagation link, and eliminate low-probability pseudo-causal relationships based on Bayesian hypothesis testing, and finally output the base station fault root cause topology graph.

5. The method according to claim 4, wherein Input the base station fault root cause topology graph into the resilience self-healing strategy generator, construct a multi-objective repair decision tree based on a digital twin-driven dynamic game framework, compress invalid repair paths through physically constrained adversarial strategy distillation technology, and use federated reinforcement learning to coordinate the repair resource allocation among multiple base stations, and finally output a set of base station facility self-healing strategies that meet real-time constraints. Among them, the dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and policy gradient mirror update, including: According to the base station fault root cause topology graph, construct a digital twin-driven dynamic game framework, generate initial repair strategy branches through a multi-objective decision tree, and simulate the feasibility of strategies under the constraints of base station physical parameters; Input the initial repair strategy branches into the adversarial strategy distillation module, use the generative adversarial network discriminator to evaluate the effectiveness of the strategies, compress the invalid repair paths through gradient reversal, and output a refined strategy candidate set; Perform federated reinforcement learning collaborative optimization on the refined strategy candidate set, adopt the distributed policy gradient algorithm to coordinate the resource allocation weights of multiple base stations, and generate a globally optimal repair strategy vector; Input the globally optimal repair strategy vector into the twin simulation verification platform, correct the strategy parameter deviation through the mirror update mechanism constrained by physical equations, and finally output a base station self-healing strategy set that meets the real-time constraints.

6. An intelligent monitoring system for abnormal base station facilities, characterized in that, The system includes: A construction module for constructing a dynamic signal perception topology using a multi-scale wavelet fusion pulse neural network according to the real-time data of the base station physical layer, eliminating multi-source signal interference through a phase synchronization time-frequency decomposition algorithm, and generating a spatio-temporally aligned base station full-dimensional feature tensor. Among them, the pulse neural network introduces a pulse timing-dependent plasticity mechanism and realizes the adaptive enhancement of signal features through dynamic correction of synaptic weights; A reconstruction module for inputting the base station full-dimensional feature tensor into the fault mode adversarial distillation module, reconstructing the normal operating condition manifold of the base station based on a variational autoencoder with physical constraints, separating sudden faults and progressive aging features through a gradient masking residual focusing algorithm, and outputting an abnormal feature vector with a fault fingerprint identifier. Among them, the variational autoencoder adopts Laplace noise injection and frequency domain sparse constraints to force the distinction between transient anomalies and steady-state degradation modes; A correction module for modeling the fault propagation graph of the abnormal feature vector, simulating the diffusion path of faults between hardware components using a graph-enhanced causal inference model, 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. Among them, the model jointly optimizes the fault location accuracy through a timing reverse truncation algorithm and a propagation path entropy constraint; An output module for inputting the base station fault root cause topology graph into a resilience self-healing strategy generator, constructing a multi-objective repair decision tree based on a digital twin-driven dynamic game framework, compressing invalid repair paths through a physical constraint-based adversarial strategy distillation technique, and using federated reinforcement learning to coordinate the repair resource allocation among multiple base stations, and finally outputting a base station facility self-healing strategy set that meets the real-time constraints. Among them, the dynamic game framework jointly ensures the physical feasibility of the repair strategy through fault scenario twin simulation and strategy gradient mirror update.

7. The system according to claim 6, wherein The construction module is specifically used for: According to the time-frequency matrix of the radio frequency signal and the hardware vibration modal spectrum, perform time-frequency decomposition on the original signal using a multi-scale wavelet ridge detection algorithm, eliminate multipath effect interference through ridge line tracking, and generate a phase-synchronized time-frequency feature map; Input the time-frequency feature map, energy consumption fluctuation curve, and device operation log stream into the spiking neural network, dynamically adjust the synaptic weights based on the spike-timing-dependent plasticity mechanism, strengthen the high-frequency abnormal signals through the adaptive spike trigger function, and output the spatio-temporal correlated spike sequence encoding; Perform multi-modal fusion on the spike sequence encoding, use the wavelet scale-domain attention gating mechanism to select the key frequency band features, and generate the multi-source signal joint representation matrix through tensor splicing; Input the multi-source signal joint representation matrix into the spatio-temporal alignment compression module, adopt low-rank tensor decomposition and residual connection to eliminate the sensor data drift, and finally output the base station full-dimensional feature tensor.

8. The system according to claim 7, wherein The reconstruction module is specifically used for: According to the base station full-dimensional feature tensor, construct a physically constrained variational autoencoder, inject Laplace noise to perturb the latent space distribution, and generate an overfitting-resistant normal operating condition manifold reconstruction vector; Calculate the residual between the normal operating condition manifold reconstruction vector and the original feature tensor, use the frequency-domain sparse constraint filter to separate the steady-state baseline features, and output the frequency-domain decoupled residual feature matrix; Input the residual feature matrix into the gradient masking focusing module, shield the aging-related gradients through the dynamic threshold gating mechanism, strengthen the local gradient response of the sudden fault, and generate the anomaly-sensitive feature vector; Perform fault fingerprint identification on the anomaly-sensitive feature vector, extract the time-frequency domain fault mode signature using convolutional sparse coding, and finally output the anomaly feature vector with a unique identifier.

9. A storage medium, characterized in that, A computer program is stored in the storage medium, wherein the computer program is set to execute the method according to any one of claims 1-5 when running.

10. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is set to run the computer program to execute the method according to any one of claims 1-5.

Images