A cell interference troubleshooting method based on AI and big data

Abstract

The application relates to the technical field of wireless communication network optimization, and discloses a cell interference investigation method based on AI and big data. The method comprises the following steps: a three-dimensional space-time index is established and is aggregated to form a space-time feature cube, the space-time feature cube is sliced along a time axis and is aggregated, and a dynamic graph snapshot sequence is output; based on the dynamic graph snapshot sequence and the space-time feature cube, an enhanced node feature vector is generated through multi-modal gating mechanism fusion, and an enhanced dynamic graph snapshot sequence is acquired; a causal direction bias of a structural causal model is injected into a message passing process of a graph neural network, a jamming probability of a candidate jamming cell is output based on the enhanced dynamic graph snapshot sequence, and after the confidence is corrected through counterfactual verification, a root cause positioning tuple is output. The application enhances features through multi-modal gating fusion, integrates a causal model to construct GNN causal attention, realizes interference tracing through counterfactual correction, and relies on digital twin deduction and SPC verification closed loop to avoid strategy issuing derived faults.

Description

Technical Field

[0001] This invention relates to the field of wireless communication network optimization technology, and in particular to a cell interference investigation method based on AI and big data. Background Technology

[0002] With the widespread application of high-density cellular deployment and broadband spectrum reuse technology in mobile communication networks, the electromagnetic environment between cells is becoming increasingly complex. Interference problems such as internal intermodulation and external blocking exhibit highly concealed, cross-site diffusion, and dynamic coupling characteristics.

[0003] Existing interference troubleshooting technologies primarily rely on expert manual drive testing or conventional machine learning pattern recognition. These solutions are mostly limited to slice analysis of isolated single sites. Even when introducing graph neural networks (GNNs) to model base station topology, their essence is merely node feature aggregation and probabilistic inference based on the strong correlation between underlying indicators. However, in real-world wireless interference scenarios, the underlying indicators of the affected cells often exhibit the most severe anomalies, making correlation-based AI models prone to falling into the logical trap of "reversing cause and effect." Furthermore, current automated network troubleshooting systems generally lack proactive security verification mechanisms before policy deployment. Directly executing optimization commands on existing physical network elements can easily trigger derivative faults such as regional network outages or widespread interference, making it difficult to meet the stringent requirements of modern large-scale communication networks for absolute accuracy in root cause localization and zero-risk closed-loop operation. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides a cell interference investigation method based on AI and big data. It enhances features through multimodal gating fusion, incorporates a causal model to construct GNN causal attention, and uses counterfactual correction to trace the cause of interference. Furthermore, it relies on digital twin inference and statistical process control (SPC) to verify the closed loop and avoid policy-induced faults.

[0005] The objective of this invention can be achieved through the following technical solutions: The first aspect of this disclosure provides a method for investigating cell interference based on AI and big data, including the following steps: S1. Construct a causal spatiotemporal data base: Access multi-source network data, establish a three-dimensional spatiotemporal index and aggregate it to form a spatiotemporal feature cube, construct a dynamic cell relationship graph based on the geographical and engineering parameter configuration relationships between cells, slice and aggregate the spatiotemporal feature cube along the time axis, and output a dynamic graph snapshot sequence containing the initial node feature vector; S2. Multimodal interference fingerprint recognition: Based on the dynamic graph snapshot sequence and the spatiotemporal feature cube, the basic statistical feature vector, the spectral texture embedding vector and the anomaly score are extracted respectively. The enhanced node feature vector is generated by fusing them through a multimodal gating mechanism, and then the dynamic graph snapshot sequence is updated to obtain the enhanced dynamic graph snapshot sequence. S3, Causal Enhancement Graph Neural Network Root Cause Localization: The causal direction bias of the structural causal model is injected into the message passing process of the graph neural network. Based on the enhanced dynamic graph snapshot sequence, the scrambling probability of the candidate scrambling cell is output. After the confidence is corrected by counterfactual verification, the root cause localization tuple is output. S4. Obtain the twin inference optimization strategy. Based on the root cause localization tuple, retrieve candidate strategies from the scenario strategy knowledge base. After performing comparative inference simulation verification in the digital twin model, encapsulate and output the target execution strategy. S5. Closed-loop self-healing orchestration and evolution: Based on the confidence level of the root cause localization tuple and the risk level of the target execution strategy, a decision matrix is ​​constructed to trigger the execution action. The execution result is verified by statistical process control. The knowledge base is updated based on the verification result and the relevant model is trained.

[0006] Furthermore, the construction of the causal spatiotemporal data base, outputting a dynamic graph snapshot sequence containing the initial node feature vectors, specifically includes: Geospatial data is discretized using Geohash encoding, and a three-dimensional index containing cell identifiers, Geohash grids, and time windows is established. The aggregated features of each dimension are extracted to form the spatiotemporal feature cube. Candidate edges are established with cells as nodes. Edge weights are calculated based on the geographical distance between two cells, frequency band overlap rate, and antenna azimuth coefficient, and weakly associated edges are filtered to generate a dynamic cell relationship graph. Slice along the time axis, and perform pooling aggregation on the grid features covered by each cell within each time window to generate an initial node feature vector. Combine this with the updated edge weights to generate the dynamic graph snapshot sequence.

[0007] Furthermore, the process of fusing and generating enhanced node feature vectors via a multimodal gating mechanism specifically includes: The basic statistical feature vector, the spectral texture embedding vector, and the anomaly score are each mapped to the same dimension through a linear transformation. Using the mapped modal vectors, spectral classification probabilities, and anomaly scores as gating inputs, attention scores for the corresponding modalities are generated through a weight matrix, and dynamic weights are obtained after normalization. The three mapping features are weighted and summed using the dynamic weights, and the enhanced node feature vector is output to update the node features of the corresponding cell.

[0008] Furthermore, the step of extracting the basic statistical feature vector, the spectral texture embedding vector, and the anomaly score respectively includes the step of automatically discovering novel interference patterns: The initial node feature vector is input into the trained stacked denoising autoencoder to calculate the scalar reconstruction error to obtain the anomaly score. When the scalar reconstruction error exceeds the adaptive threshold, it is marked as an abnormal cell. Clustering of the reconstructed residual vectors of abnormal cells reveals novel interference candidate clusters with similar deviation patterns. After confirmation, the corresponding spectral patterns are extracted as incremental samples to fine-tune the spectral classification model.

[0009] Furthermore, the process of injecting the causal direction bias of the structural causal model into the message passing process of the graph neural network specifically includes: The structural causal model is mapped onto the candidate edges of the dynamic cell relationship graph. For any candidate edge, if there is a causal path between the two nodes, a positive causal prior matrix term is assigned; otherwise, a zero or negative value is assigned. In the node feature aggregation stage of the graph attention network, a causal attention bias term is introduced when calculating the attention coefficients between nodes. This term enhances message passing weights when the propagation direction is consistent with the causal direction corresponding to the causal prior matrix term, and suppresses the opposite direction.

[0010] Furthermore, after correcting the confidence level through counterfactual verification, the root cause localization tuple is output, specifically including: For candidate disturbing cells, the changes in intervention parameters are simulated in the structural causal model to predict the expected changes in indicators of the disturbed cells; Simultaneously, the same feature intervention is performed on the causal augmentation graph neural network to predict the amount of change in the hidden layer response of the disturbed cell; The mean square error between the two variables is calculated to obtain a causal consistency score. The causal consistency score is used to correct the perturbation probability. When the final corrected confidence level is higher than a preset threshold, the root cause localization tuple containing the perturbation type, the list of perturbation cells, the causal evidence chain, and the confidence level is generated.

[0011] Furthermore, the acquisition of the twin inference optimization strategy specifically includes: The interference fingerprint vector of the current disturbed cell is concatenated with the root cause feature vector. Historical similar cases are recalled through the vector index library, and the structured index is used to filter by interference type and confidence range. The strategy set is extracted from the filtered cases, and a collaborative filtering algorithm combining historical success rate, impact range index and execution cost index is used to score and rank them, and recommended candidate strategies are output. In the digital twin base station model built based on historical network data, parameter changes are simulated for the recommended candidate strategies, and the expected improvement rate and a list of negatively impacted cells are output to determine the execution risk level.

[0012] Furthermore, the closed-loop self-healing orchestration and evolution specifically includes: The network metrics of multiple consecutive time windows after closed-loop execution are compared with the baseline data before execution using a statistical process control algorithm. If the policy execution is determined to fail, a rollback operation is triggered, and the interference characteristics and execution effect of the corresponding cell are written into the knowledge base as negative samples and the weight of the policy is reduced. If manual correction feedback is received, the corrected scrambling label is compared with the original output of the causal augmentation graph neural network to form a comparative learning sample pair, driving the model to fine-tune and retrain.

[0013] The second aspect of this disclosure provides a cell interference investigation system based on AI and big data, including: Memory, used to store computer programs; The processor is used to execute the computer program stored in the memory to implement the AI ​​and big data-based cell interference investigation method described above.

[0014] A third aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the AI ​​and big data-based cell interference investigation method described above.

[0015] The beneficial effects of this invention are: This invention utilizes multimodal gating fusion technology to align multi-source heterogeneous statistical features, spectral textures, and anomaly scoring depth into enhanced graph node features, significantly improving the representation and recognition sensitivity of unknown interference patterns in complex networks. Its core breakthrough lies in injecting prior knowledge from the Structural Causal Model (SCM) into the message passing mechanism of the Graph Neural Network (GNN) to construct causal bias attention, supplemented by counterfactual verification to correct the consistency of prediction results. This overcomes the technical bottleneck of traditional data-driven algorithms that rely solely on "appearance correlation" and are prone to misjudging the "disturbed cell" with the most severe feature distortion as the "disturbing source," achieving a leap from "correlation statistical inference" to "cross-site causal tracing." Simultaneously, by combining comparative simulation and SPC statistical verification closed loop with digital twin base stations, it effectively avoids the risk of secondary network degradation caused by the blind deployment of automated investigation strategies. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram illustrating the steps of a cell interference investigation method based on AI and big data provided in an embodiment of the present invention; Figure 2This is a schematic diagram of the root cause localization steps of the causal augmentation graph neural network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the steps for obtaining a twin inference optimization strategy provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 This embodiment provides a cell interference investigation method based on AI and big data, such as... Figure 1 As shown, it includes the following steps: S1. Construction of Causal Spatiotemporal Data Base: This involves accessing and cleaning multi-source data, establishing a three-dimensional spatiotemporal index and aggregating it to form a spatiotemporal feature cube, constructing a dynamic cell relationship graph, and then converting, slicing, and aggregating the spatiotemporal graph to output a dynamic graph snapshot sequence. This includes the following steps: A unified data bus is established to access performance management (PM) data, measurement reports / minimum drive test data (MR / CDT, where CDT refers to call detailed trace), alarm data, engineering parameter data, frequency sweep data, and complaint ticket data. The multi-source data undergoes deduplication, missing value imputation, outlier filtering, timestamp calibration, and field standardization. The cleaned data is assigned a unified spatiotemporal label, forming a global data lake. The global data lake stores the cleaned detailed data and its spatiotemporal labels, while the distributed time-series database stores aggregated features and time-series data organized based on a three-dimensional index.

[0020] Geohash encoding (a method of encoding two-dimensional latitude and longitude coordinates into one-dimensional strings) is used to discretize geospatial data. The Geohash accuracy level is adaptively selected based on the cell coverage radius. User measurement points, frequency scanning points, complaint locations, alarm locations, and cell coverage areas are mapped to Geohash grids. Using a 15-minute time window, various data types are aligned according to the Geohash grid, cell identifier, and time window to establish a (cell) structure. id Geohash grid time window A three-dimensional index is created and stored in a distributed time-series database. This three-dimensional index enables data from different data sources, such as PM, MR, and alarms, to be correlated at the same time point.

[0021] For each cell-grid-time window (cell, grid, time) cube cell, calculate the following characteristics at the PRB (Physical Resource Block) level: average interference power, interference fluctuation rate, average signal-to-interference-plus-noise ratio (SINR), overlap coverage, complaint density, alarm status, and service load. The grid corresponds to Geohash. grid The identified grid unit, cell corresponds to cell id The identified cell, time corresponds to time window The identified time window, this cube cell, is the smallest granularity data block located by the aforementioned 3D index. For each cell-grid-time window (cell, grid, time) cube cell, the corresponding aggregated features are calculated and stored, using the 3D index (cell...) id Geohash grid time window The data is arranged in rows and interference-related features in columns to form a spatiotemporal feature cube. This cube retains a fine-grained spatial structure of "cell-grid-time window". Subsequently, when constructing the graph snapshot, pooling aggregation is performed on multiple Geohash grids covered by the same cell within the same time window to generate initial feature vectors for nodes at the "cell-time window" granularity.

[0022] Simultaneously, using cells as nodes, candidate edges are established when two cells meet at least one of the following conditions: geographical distance less than twice the average inter-cell spacing, co-frequency or adjacent-frequency configuration, antenna azimuth angle less than 60°, and antenna pointing relative to each other. Weakly correlated edges are further filtered based on edge weight thresholds to generate a dynamic cell relationship graph G with cells as nodes. base = (V, E), where V represents the set of nodes, with each node corresponding to a unique cell; E represents the set of edges; the edge weights are dynamically calculated based on the geographical distance between two cells, the frequency band overlap rate, and the antenna azimuth coefficient, and the calculation formula is: w ij =α·(1 / D ij +ε)+β·F ij +γ·A ij ; In the formula, D ij For normalized geographic distance, ε is the smoothing constant, and F ij For frequency band overlap rate, A ij A is the azimuth coefficient (when the antennas of the two cells are directly facing each other (angle 0°)). ij =1; When the included angle reaches the preset threshold of 60°, A ij=0; a cosine attenuation function is used for smooth transition in the middle), α, β and γ are adjustable weight coefficients. By extracting positive samples of known interference events from historical data, gradient descent is used for fitting and optimization to ensure that the edge weights can distinguish edges with strong interference relationships to the greatest extent. The graph structure is dynamically refreshed as the operating parameters (such as antenna angle adjustment and frequency reconfiguration) are updated, reflecting changes in network topology and configuration. The resulting graph is a dynamic cell relationship library.

[0023] The spatiotemporal feature cube is sliced ​​along the time axis. Within each time window t, pooling aggregation is performed on all Geohash grid features covered by each cell (e.g., taking the mean and variance of PRB interference power, the mean of user SINR, etc.) to generate the initial node feature vector for the current time window. The latest engineering parameter data for the current time window is extracted from the dynamic cell relationship graph library to recalculate the edge weights. The node features and edge weights are combined into a graph structure data G. t = (V, E, X) t W t ), where X t W is the node feature matrix. t The edge weight matrix is ​​arranged in chronological order to obtain the dynamic graph snapshot sequence (G1, G2, ..., G...). t ).

[0024] It should be noted that in this step, the original 3D tensor indexed by "cell-grid-time window" is converted into a dynamic graph structure of "node-edge-time window" that can be processed by the graph neural network. Because the edge weights are updated according to changes in operating parameters and frequency configuration, the model can promptly reflect the interference propagation relationship after network parameter adjustments, thereby reducing the risk of misjudgment caused by static adjacency relationships. In the output dynamic graph snapshot sequence, the initial feature vector of each snapshot's node will serve as the basic input for step S2. In step S2, after multimodal fusion enhancement, it forms an enhanced node feature vector, which is then written back to the corresponding graph snapshot to form an enhanced dynamic graph snapshot sequence for consumption by the causal enhanced graph neural network in step S3, thus achieving standardized data flow between steps.

[0025] S2. Multimodal interference fingerprint recognition: Based on dynamic graph snapshot sequences and spatiotemporal feature cubes, basic statistical feature vectors, spectral texture embedding vectors, and anomaly scores are extracted using automated feature operators, CNN (Convolutional Neural Network), and SDAE (Stacked Denoising Autoencoder), respectively. These are then gated and fused to generate enhanced node feature vectors, resulting in an enhanced dynamic graph snapshot sequence. The process includes the following steps: Automated feature operators extract statistical features: Based on the grid aggregation features of spatiotemporal feature cubes, for each time window, cell-level pooling aggregation is first performed on multiple Geohash grids covered by the same cell to obtain the statistical input features of the corresponding cell within that time window. Then, the automated feature operator library is used to calculate time-domain features (noise floor abrupt change point, periodic autocorrelation coefficient, duration of continuous anomalies, volatility), frequency-domain features (specific PRB uplift index, spectral flatness, frequency band concentration, out-of-band leakage features), and spatial-domain features (interference source direction angle estimation, overlap coverage, density of neighboring cells with the same frequency), which are then concatenated into a 32-dimensional basic statistical feature vector, supplementing physically interpretable domain knowledge.

[0026] Spectrum Waterfall Plot Generation and CNN Interference Classification: For each cell within a time window t in the dynamic image snapshot sequence, extract the PRB interference power from multiple consecutive time windows (e.g., 6 time windows covering 90 minutes) in the global data lake. Using the PRB index as the horizontal axis and the time window sequence as the vertical axis, map the interference power values ​​to pixel grayscale to generate a spectrum waterfall grayscale map. The image size corresponds to the number of PRBs (e.g., 1200) and the number of time windows (e.g., 6).

[0027] An improved ResNet-18 (residual network with 18 layers and single-channel grayscale input) is used for classification, outputting probabilities for five classes: external interference, intermodulation interference, blocking interference, spurious interference, and no interference. The 128-dimensional vector after global average pooling is used as the spectral texture embedding vector. This embedding vector captures the texture pattern of interference in a two-dimensional time-frequency structure, and can distinguish the spectral morphology of different interferences.

[0028] Anomaly Detection and Novel Pattern Discovery: Initial node feature vectors extracted from dynamic graph snapshot sequences during normal periods (no complaints or alarms, stable key performance indicators (KPIs)) are used to train a stacked denoising autoencoder (SDAE) to learn a normal baseline. The reconstruction residual is calculated online for the initial node feature vectors of each time window. This involves inputting the initial node feature vectors into the SDAE and calculating the dimension-wise difference between the input vector and the reconstructed output vector to obtain the reconstruction residual vector. Simultaneously, the mean square error (MSE) of this reconstruction residual vector is calculated as the scalar reconstruction error. When the scalar reconstruction error exceeds a 3-sigma adaptive threshold, an abnormal cell is marked. The reconstruction residual vectors of abnormal cells are clustered using DBSCAN (density-based spatial clustering algorithm) to discover candidate clusters of novel interference patterns (automatically discovering abnormal clusters with similar deviation patterns; the cluster center of each cluster corresponds to a candidate novel interference pattern). The scalar reconstruction error (MSE) is normalized to obtain an SDAE anomaly score in the [0,1] interval, which serves as a continuous feature to measure the degree of deviation of the cell from the normal baseline.

[0029] Multimodal Feature Gated Fusion: The three pathways mentioned above produce heterogeneous features: a basic statistical feature vector (32-dimensional vector, carrying domain expert knowledge), a spectral texture embedding vector (128-dimensional continuous vector, capturing spectral morphology information), and an SDAE anomaly score (1-dimensional normalized value, representing the degree of deviation from the normal baseline). To fuse these heterogeneous features into a unified vector representation that can be directly consumed by a graph neural network, this invention sets up a multimodal feature encoder. This encoder first maps the three feature vectors to the same dimension through a learnable linear transformation, and then introduces a lightweight gating network: using the mapped modal vectors, the maximum value of the CNN classification probability or classification entropy, and the SDAE anomaly score as gating inputs, it generates the attention score of the corresponding modality through a learnable weight matrix, obtains dynamic weights through Softmax normalization, and performs a weighted summation of the three mapped features to output the final enhanced node feature vector h. cell The gating mechanism enables the model to adaptively adjust the contribution of each modality under different interference scenarios: since the gating network explicitly receives the CNN classification confidence and SDAE anomaly score as input, the weight of the spectral texture feature can automatically increase when the CNN classification confidence is high; when an anomaly deviation is detected, the weight of the SDAE anomaly score increases, thereby avoiding the inadequacy of a single model to express complex interference scenarios and enhancing the robustness of recognition.

[0030] Furthermore, this invention extracts typical spectral patterns from the novel interference clusters discovered by the stacked denoising autoencoder after expert confirmation and adds them to the incremental training samples of the CNN. Through transfer learning, the CNN classifier is fine-tuned, achieving a continuous evolution from unsupervised anomaly detection to supervised interference classification. This improvement enables the system not only to identify existing interference types but also to discover unknown patterns during operation and precipitate them as new identifiable categories, enhancing the model's adaptability in long-term operating scenarios.

[0031] Node feature update and output: For each time window t, the enhanced node feature vector is assigned to the dynamic graph snapshot G. t For the corresponding nodes in the cell, update the node feature matrix X. t After processing all time windows, an enhanced dynamic graph snapshot sequence is obtained. Each node in the graph snapshot sequence carries a high-dimensional feature representation that integrates spectral texture, anomaly degree, and statistical regularity, which is directly used as the input to the causal augmentation graph neural network in step S3. Meanwhile, intermediate products such as the spectral texture embedding vector and anomaly score are also used to construct the interference fingerprint vector of the knowledge base in step S4 and for subsequent anomaly tracing.

[0032] S3. Causal Augmented Graph Neural Network Root Cause Localization: Based on enhanced dynamic graph snapshots and dynamic cell relationship graphs, a structural causal model is introduced to impart a causal direction bias to message passing. The perturbation probability is calculated using a causal augmented graph neural network, and the confidence level is corrected through counterfactual verification before outputting the root cause localization tuple, such as... Figure 2 As shown, it includes the following steps: S31. Graph Structure Data and Causal Prior Preparation: For each time window t, take an enhanced dynamic graph snapshot from the output of step S2. Extract the node set V (each node corresponds to a cell), the edge set E (established based on the geographical distance, frequency band configuration, and antenna orientation angle of the dynamic cell relationship map), and the node feature matrix. (Based on the enhanced node feature vector h of each cell) cell Constructed by concatenating rows, where d is the enhanced node feature vector h. cell (dimensions) and edge weight matrix Simultaneously, a structural causal model (SCM) is introduced, using a directed acyclic graph. The SCM (Synchronous Causal Principle) is used to express the causal direction at the variable level and does not require the cell relationship graph itself to be an acyclic graph; when mapping to the cell node graph, only the causal direction of the variables in the SCM is transformed into a directional bias on the candidate cell edge. Nodes include power configuration, antenna parameters, user distribution, noise floor increase, SINR decrease, complaints, etc., and edges represent causal directions (e.g., an increase in the power of the interfering cell leads to an increase in the noise floor of the interfering cell). The SCM is mapped to the cell node graph, where the edge E in the dynamic cell relationship graph is used to limit the candidate adjacency range of potential interference propagation, and the SCM is used to assign a directed causal bias within the candidate adjacency range; for any candidate edge (i, j), if there is a causal path from cell i to cell j, a positive causal bias C is assigned. ij If the direction is opposite or irrelevant, a value of zero or negative is taken. This transforms the original candidate adjacency graph into a causal message-passing graph with directional bias.

[0033] S32. Causal Message Passing Mechanism: To inject the causal directionality of SCM into the message passing process of the graph neural network, this embodiment proposes a causal message passing mechanism, which calculates the attention coefficients e from node i to j in the standard graph attention network (GAT). ij = LeakyReLU(a [W h i || W h j Based on this, a causal attention bias term is added to enhance the attention distribution along the causal direction, where h i h j Let be the enhanced node feature vectors of nodes i and j, respectively; W be the learnable linear transformation matrix; a be the attention weight vector; and || denotes concatenation. The corrected attention coefficients are expressed as: ; In the formula, The edge weights between cells i and j in the dynamic graph snapshot output in step S1. C represents the learnable or preset edge weight influence coefficient. ij Let C be a causal prior matrix term, and β be the learnable causal bias strength; when C ij When the value is positive (i→j is the causal direction), the message passing weight from i to j is increased; in the opposite direction, it is suppressed or cleared to zero. This design transforms the directionality of the SCM causal graph into the directional constraint of message passing in a graph neural network (GNN), so that graph reasoning no longer depends solely on correlation, but proceeds along possible physical causal links.

[0034] S33. Calculation of Perturbation Probability Score: Construct a Causal-GNN (Causal Enhanced Graph Neural Network). The model uses a stacked 3-layer Graph Attention Network (GAT), with each layer using attention coefficients corrected for causal bias for message passing; the node representation update formula for the k-th layer is: ; In the formula, The graph attention coefficients of the k-th layer after causal bias correction are... For the k-th layer, a learnable linear transformation matrix, Let i be the set of neighbors of node i. The activation function is used; after three layers of propagation, each node obtains a final embedding representation containing three-hop neighborhood causal information. Finally, a fully connected layer maps the node embedding to a scalar output, which, after Sigmoid activation, yields the scrambling probability score P of the cell as a scrambling source within the time window t. interf .

[0035] S34. Counterfactual Verification and Confidence Correction: For high-probability candidate harassing cells located by GNN, the corresponding cell's transmit power is simulated to be reduced in SCM. dB, using the structural equation of SCM to predict the expected decrease in noise floor ΔNoise in disturbed cells. SCM Simultaneously, the same feature intervention is performed on Causal-GNN to obtain changes in hidden layer responses and predict the decrease in background noise ΔNoise. GNN The causal consistency score is expressed as: ; In the formula, c consistency A causal consistency score is assigned, ranging from (0,1]. A higher value indicates greater consistency between the causal reinforcement graph neural network and the structural causal model. ΔNoise SCM This represents the decrease in cell noise floor predicted by the structural causal model. ΔNoiseGNN This represents the reduction in background noise of the disturbed cell predicted by the causal augmentation graph neural network. τ is a scaling parameter that controls the sensitivity of the score to bias.

[0036] The final root cause confidence level is adjusted to: P final =P interf ×c consistency ; P final Cells exceeding the threshold are included in the list of disturbed cells. Simultaneously, the interference type label is determined based on the CNN classification probability of the spectral waterfall plot corresponding to the time window of the disturbed cell in step S2. Consistency verification is then performed by combining the spectral texture embedding vector of the candidate disturbing cell, the SDAE anomaly score, and the causal evidence chain to finally determine the interference type label corresponding to the root cause.

[0037] S35. Root Cause Location Tuple Output: For each cell identified as having interference and its root cause, output the root cause location tuple: Interference type, {list of interfering cells}, causal evidence chain, P final ; The causal evidence chain includes: the causal propagation path extracted from the GNN attention weights (such as cell A → cell B → cell C); and key evidence fragments from the SDAE anomaly score and spectral texture feature vector in step S2.

[0038] Training Objectives and Cold Start Strategy: The training loss function of Causal-GNN consists of two parts: ; in, Binary cross-entropy loss for locating interference sources (based on known interfering cell labels). The counterfactual consistency loss is calculated by minimizing ΔNoise for a batch of intervention samples. GNN With ΔNoise SCM Mean square error between, hyperparameter This is used to balance the two losses. Through joint optimization, Causal-GNN can accurately locate the source of interference while maintaining intrinsic consistency with physical causal laws.

[0039] To accelerate the deployment of real networks, simulation platforms (such as RayTracing+NS3) were used to generate more than 100,000 labeled interference scenarios for pre-training Causal-GNN. When migrating to the real network, only about 200 cells of on-site confirmation data need to be collected, and high-precision positioning capabilities can be obtained by fine-tuning the network.

[0040] S4. Obtaining Twin-Based Optimization Strategies: Based on root cause localization tuples, candidate strategies are obtained by retrieving the scenario strategy knowledge base through a dual-path index. After verification through digital twin inference, the strategy sequence, expected improvement rate, impact range, and risk level are output, such as... Figure 3 As shown, it includes the following steps: S41. Construct a scenario strategy knowledge base: Each case includes an interference fingerprint vector (directly concatenated from the spectrum texture embedding vector, basic statistical feature vector, and SDAE anomaly score of the disturbed or abnormal cell within the corresponding time window in step S2, which is different from the enhanced node feature vector after gated weighted fusion), a root cause feature vector (concatenated from the features of the disturbing cell in step S3 (from Causal-GNN node embedding), causal evidence chain, and confidence level), a set of strategies (such as power downscaling, tilt adjustment, frequency reconfiguration, filter installation requests, etc.), execution success rate, and side effect records (the deterioration of KPIs of surrounding cells and its range after execution). A dual-path index is used: one path is a FAISS (Facebook AI similarity search library / vector index library) vector index, which concatenates the interference fingerprint vector and the root cause feature vector and stores them in the vector database; the other path is an SQLite (lightweight relational database) structured index, which supports precise filtering by interference type, frequency band, and confidence level range. After each closed loop, the complete case data (environmental characteristics, reasoning path, execution strategy, and actual effect) is added to the database as a new sample and the success rate is updated. The weight of the corresponding strategy for failed cases is reduced accordingly.

[0041] S42. Strategy Retrieval and Collaborative Filtering Ranking: Based on the root cause localization tuple, the top-20 historical cases with the highest cosine similarity are first recalled using the concatenated vector of the current interference fingerprint vector and the root cause feature vector via FAISS; then, the top-20 cases are filtered by interference type label, interfering cell frequency band, and confidence range; a strategy set is extracted from the retained cases, and collaborative filtering weighted voting is used for ranking. Score= w 1·SR+ w 2·[1 / (Impact+ f )]+ w 3·[1 / (Cost+ f )]; In the formula, SR represents the historical success rate of the current strategy in similar cases, Impact represents the normalized strategy impact range indicator, and Cost represents the normalized execution cost indicator. f To prevent smoothing constants with denominators of zero, w 1 、w 2 、w 3 represents the weight. The scores are sorted in descending order, and the top-3 are selected as candidate recommendation strategies for subsequent digital twin simulation and verification.

[0042] S43. Digital Twin Simulation Verification: Based on historical measurement reports, engineering parameters, and service statistics of the target area, construct a digital twin base station model, including: ray tracing propagation model (characterizing signal attenuation and multipath effects), user spatiotemporal distribution model (fitted from historical MR data), and service load model (based on historical PRB utilization statistics).

[0043] For each recommendation strategy, perform the following comparative simulation in a digital twin environment: Pre-intervention simulation: Using the current network parameters as input, simulate the interference power distribution and user SINR of the target area, and compare it with the MR measurement data measured in step S1 to verify the model accuracy; when the error between the simulation result and the measured result exceeds the preset threshold, the propagation model parameters, user distribution parameters or service load parameters are calibrated. Strategies that still do not meet the accuracy requirements after calibration will not enter the automatic execution process and will only be used as manual reference.

[0044] Post-intervention simulation: Apply parameter changes corresponding to the strategy (such as reducing the power of the interfering cell by 3dB and raising the antenna tilt angle by 2°), and simulate the changes in interference power distribution, SINR of the affected cell, handover success rate, and KPI of surrounding cells after the change.

[0045] The simulation output evaluation metrics include: expected improvement rate ΔSINR (the expected percentage increase in SINR of disturbed cells), a list of negatively affected cells (cells whose KPI deterioration exceeds the threshold and the amount of deterioration), and an implementation risk level (comprehensively rated as low / medium / high based on the scope of negative impact and the magnitude of deterioration).

[0046] S44. Atomic Safety Operation Encapsulation: Encapsulates the executable operations corresponding to the policy into atomic operation APIs (Application Programming Interfaces). Each atomic operation includes: Precondition checks: such as verifying whether the current value is within the safe threshold before power adjustment, and verifying whether the target frequency is idle before frequency reconfiguration; Conflict detection: Detects whether there are resource conflicts with other optimization tasks being executed (such as the same cell being adjusted by two tasks at the same time). Rollback script: Records the original values ​​of current parameters, supports one-click recovery, and ensures quick rollback in case of execution failure.

[0047] Atomic operation types include at least: power adjustment (e.g., step 1dB), antenna tilt adjustment (e.g., step 1°), PCI (Physical Cell Identifier) / frequency reconfiguration, and filter installation request initiation.

[0048] S5. Closed-Loop Self-Healing Orchestration and Evolution: A decision matrix is ​​constructed based on root cause confidence and risk level to drive automatic strategy execution, manual task assignment, or alarm. After execution, the effect is verified by statistical process control. Based on the verification results, rollback, knowledge base update, and model feedback training are triggered, including the following steps: Constructing the confidence-risk decision matrix: The final root cause confidence P output from step S3... final The risk level of the corresponding candidate strategy output in step S4 is used for joint judgment: Automatic execution when confidence level is greater than or equal to 90% and risk level is low; automatic execution and generation of post-audit when confidence level is greater than or equal to 90% and risk level is medium; push for manual confirmation when confidence level is greater than or equal to 90% and risk level is high; push for manual confirmation and attach a twin simulation report when confidence level is greater than or equal to 70% and less than 90%; only alarm when confidence level is less than 70%.

[0049] Atomic operation execution: For automatic execution strategies, the orchestration engine generates an execution sequence based on priority (smaller impact operations take precedence) and dependencies (multiple operations within the same sub-area must be executed sequentially), calling the atomic operation API in step S4 in turn. Before each atomic operation is executed, precondition checks and conflict detection are performed one by one. If they pass, the original value is recorded as a rollback point before the changes are executed. If any operation fails or a conflict occurs, the operation is immediately interrupted and the executed operations are rolled back.

[0050] Closed-loop verification: After the strategy is executed, the system continuously monitors the MR data, PRB interference power, SINR, handover success rate, and complaints of the affected cell and surrounding cells. A Statistical Process Control (SPC) algorithm is used to compare data from multiple time windows after execution (e.g., 6 consecutive time windows) with data from adjacent baseline periods on the same day before execution, historical data from the same week and hour, or baseline periods with similar service loads. If the average SINR increases by more than a preset percentage (e.g., 5%) and the interference power decreases by more than a preset percentage (e.g., 10%), the strategy is considered successful; otherwise, a rollback process is triggered. In the event of failure, the interference fingerprint vector of the corresponding cell (combined directly from the spectral texture embedding vector, basic statistical feature vector, and SDAE anomaly score in step S2), the inference path log in step S3, the recommended strategy in step S4, and the execution effect are written as negative samples into the knowledge base and weighted down; in the event of success, they are solidified as positive samples and weighted up. Simultaneously, the actual KPI changes, actual interference power changes, engineer confirmation results, and rollback results after execution are fed back: the spectral waterfall plot with interference type confirmation label is fed back to the CNN training sample set in step S2; the initial node feature vectors corresponding to the confirmed normal or abnormal time periods are fed back to the SDAE training sample set; the enhanced dynamic graph snapshot with interference cell confirmation label and the causal evidence chain are fed back to the Causal-GNN localization training sample set in step S3; and the policy execution results, actual improvement rate, negative impact range, and rollback records are fed back to the scenario policy knowledge base in step S4 to achieve continuous updates of the model and policy base.

[0051] Human-in-the-loop interactive learning and model feedback training: Engineers' actions such as confirming and executing push work orders, rejecting them, modifying strategies, or correcting root causes are all used as strong feedback.

[0052] Root cause correction feedback: If the engineer corrects the scrambling cell located in step S3 (e.g., "the root cause is actually cell XX"), the corrected scrambling label is compared with the output of the original causal augmentation graph neural network (Causal-GNN) to form a contrastive learning sample pair. The feedback loss term drives the Causal-GNN to fine-tune and enhance its ability to discriminate the causal direction.

[0053] Strategy Correction Feedback: If the engineer replaces or modifies the recommended strategy, the new strategy and its execution effect are recorded. The strategy success rate of the corresponding case in the knowledge base of step S4 is directly updated, and the SR item in the collaborative filtering ranking is optimized accordingly.

[0054] A monthly model drift report is generated. When the root cause localization accuracy or strategy recommendation success rate is lower than the threshold, the causal augmentation graph neural network in step S3 and the CNN classifier and SDAE model in step S2 are retrained.

[0055] Example 2 This embodiment provides an end-to-end cell interference troubleshooting process. The system is deployed on the operator's wireless network operation and maintenance platform and includes a data acquisition layer, a spatiotemporal data base layer, an AI recognition layer, a causal root cause localization layer, a policy inference layer, and an automated orchestration layer.

[0056] First, the data acquisition layer accesses PM, MR / CDT, alarm, operating parameter, frequency scan, and complaint work order data through a unified data bus. The system performs field mapping on data from different sources, for example, unifying the cell identifier to "cell". id The sampling time will be unified as time. window Map latitude and longitude to Geohash grid For missing PRB interference power, the system can use interpolation of adjacent time windows in the same cell or weighted interpolation of neighboring cells in the same grid; for anomalous abrupt changes, the system combines box plot thresholds and historical distributions for filtering.

[0057] Subsequently, the system selects the Geohash accuracy level based on the cell coverage radius. Lower accuracy is used when the macro base station coverage radius is large, and higher accuracy is used when the indoor distributed antenna system (DAS) or dense urban area cell coverage radius is small. The time window is set to 15 minutes, aligned with the PRB measurement cycle. The system establishes Geohash... grid cell id and time window Three-dimensional indexing generates a spatiotemporal feature cube.

[0058] In this embodiment, the spatiotemporal feature cube is not directly input into the GNN, but first passes through a spatiotemporal graph transformation layer. The spatiotemporal graph transformation layer operates according to time... window Slicing, capturing the same cell within a time window id All Geohash covered grid Features are aggregated to obtain the node features of the cell within that time window; simultaneously, inter-cell edge weights are calculated based on engineering parameters and spatial relationships to form a dynamic graph snapshot G. t This process transforms data originally intended for spatial grid storage into a data structure oriented towards cell relationship reasoning, avoiding information loss caused by inconsistencies between the spatiotemporal grid model and the input structure of the graph neural network. Simultaneously, the GNN can perceive the inter-cell interference propagation relationships within each time window and dynamically update the graph structure as operating parameters change.

[0059] During the interference fingerprinting phase, the system constructs a spectrum waterfall plot for each cell. The spectrum waterfall plot uses the PRB number as the horizontal dimension, continuous time points as the vertical dimension, and interference power as grayscale values ​​or normalized intensity values. The CNN model outputs the probability of interference categories, such as external interference probability 0.82, intermodulation interference probability 0.06, blocking interference probability 0.04, spurious interference probability 0.03, and no interference probability 0.05.

[0060] Simultaneously, the system inputs the initial node feature vectors extracted from the dynamic image snapshot sequence during normal periods into the stacked denoising autoencoder for training. During real-time inference, if the reconstruction error of a cell in the current time window exceeds the dynamic 3-sigma threshold, it is marked as an anomaly. For multiple anomaly cells, the system uses DBSCAN to cluster their reconstruction error vectors. If the clustering result does not belong to an existing interference type, a candidate novel interference pattern is formed.

[0061] In this embodiment, the multimodal feature encoder compresses the spectral texture features output from the intermediate layers of the CNN into a 128-dimensional embedding vector, normalizes the autoencoder reconstruction error into a one-dimensional anomaly feature, and concatenates the outputs of the time-domain, frequency-domain, and spatial-domain feature operators into a basic feature vector. Subsequently, the gated fusion network learns the weights of each modality based on the current scene. For example, it increases the CNN embedding weights in scenarios with obvious spectral textures and increases the autoencoder anomaly scoring weights in scenarios with sudden anomalies but unstable spectral textures. This processing is based on the fact that different types of interference have different degrees of dependence on different modalities; its technical effect is to unify the node features h. cell It can more stably represent the cell interference state and provide continuous, learnable, and comparable inputs for subsequent Causal-GNN.

[0062] Furthermore, once DBSCAN discovers a stable new type of interference cluster, the system pushes the typical spectral pattern and contextual features corresponding to the cluster center to the expert annotation queue. After expert confirmation, the system uses these as incremental samples to fine-tune the CNN classifier. This process transforms unsupervised anomaly detection into supervised classification knowledge, solving the problem of the model's long-term inability to identify new interference samples due to their scarcity. The model can continuously evolve with changes in the network environment, reducing the cost of repeated manual checks.

[0063] During the root cause localization phase, the system will enhance the dynamic graph snapshot sequence and enhance the node feature vector h. cell The Causal-GNN is input with edge features. Edge features include the distance between two cells, frequency band overlap rate, azimuth coefficient, signal attenuation estimate, and MR time difference statistics. The Causal-GNN adopts a 3-layer graph aggregation structure, with each layer aggregating one-hop neighbor information to cover the cell interference propagation paths within a three-hop range.

[0064] The structural causal model predefines the causal relationships between power configuration, antenna parameters, user distribution, noise floor increase, SINR decrease, complaints, and KPI deterioration. For example, an increase in the power of the interfering cell may lead to an increase in the noise floor of neighboring co-channel interfering cells, which in turn leads to a decrease in SINR and an increase in complaints.

[0065] In this embodiment, Causal-GNN does not perform undirected aggregation solely based on adjacency relationships, but instead introduces SCM causal orientation bias in the message passing layer. When SCM determines that node i has a possible perturbation direction to node j, attention... ij Obtain a positive causal bias; when the propagation direction is opposite to the causal chain, attention ij The impact of the error is either reduced or not enhanced. During training, the system simultaneously calculates the localization loss and the counterfactual consistency loss. This is because ordinary GNNs are prone to misclassifying anomalies in disturbed cells as the cause of the disturbance; by introducing causal orientation, the model can distinguish between the "location of the anomaly" and the "source of the anomaly." This process improves the accuracy of root cause localization and generates an interpretable chain of causal evidence.

[0066] For candidate interfering cells, the system performs counterfactual verification. For example, the model identifies cell A as a candidate interfering cell, and cells B and C as the affected cells. The system intervenes in the SCM: if the transmit power of cell A decreases by 3dB, the predicted noise floor decreases by 2.1dB for cell B and 1.4dB for cell C. The system searches for similar historical cases; if the noise floor decrease caused by similar power reductions in the past is close to the predicted result, the causal consistency score is increased; otherwise, the confidence score is decreased. The final output root cause localization tuple includes: interference type, list of interfering cells or regions, causal evidence chain, and confidence score.

[0067] During the strategy recommendation phase, the system inputs the root cause localization tuple into the scenario-strategy knowledge base. Each case in the knowledge base stores the interference fingerprint vector, root cause feature vector, strategy set, success rate, and side effect records. The system performs vector similarity retrieval based on FAISS, and simultaneously performs filtering based on conditions such as frequency band, region, power range, equipment manufacturer, and risk level using SQLite or other structured databases.

[0068] In this embodiment, the system concatenates the interference fingerprint vector generated in the second step with the root cause feature vector generated in the third step to form a case retrieval vector, creating a dual-path index data foundation. Interference type serves as the primary key to quickly narrow down the candidate case range, while the engineering parameters of the interfering cell serve as a secondary screening condition. A causal evidence chain is used to reconstruct the digital twin verification scenario. This approach is necessary because retrieving based solely on interference type ignores the engineering conditions of the interfering cell, while retrieving based solely on engineering parameters fails to reflect the similarity of interference patterns. The technical advantage lies in the fact that the strategy recommendation can both match similar interference fingerprints and satisfy specific network parameter constraints.

[0069] After the system retrieves the top 20 similar cases, it ranks the strategies by weight. For example, strategies that reduce power by 3dB have a high historical success rate, a small impact range, and low execution cost are ranked higher; antenna tilt adjustments have a higher risk level if they may affect coverage; and on-site investigation work orders for external interference have high execution costs but are suitable for external interference scenarios such as privately installed amplifiers.

[0070] The digital twin model is used to extrapolate candidate strategies. If the simulation error between the current interference distribution and the actual measurement is less than 5%, the twin model is considered suitable for strategy verification. The system continues to simulate the changes in SINR, PRB interference power, and surrounding KPIs after the strategy is implemented, and outputs the expected improvement rate, a list of negatively impacted cells, and the risk level.

[0071] In this embodiment, the orchestration engine does not directly execute operations based on the policy recommendation results. Instead, it constructs a confidence-risk decision matrix based on root cause confidence and digital twin risk levels. If the confidence is high and the risk is low, the atomic operation library is invoked for automatic execution; if the confidence is high but the risk is moderate, automatic execution is performed with post-event auditing; if the risk is high or the confidence is moderate, manual confirmation is pushed; if the confidence is low, only an alarm is issued. The reason for this approach is that interference governance strategies may affect the coverage and KPIs of surrounding cells, and simply pursuing automation can bring network risks. Its technical effect is to control the risk of misoperation while improving the efficiency of automatic closed-loop operation.

[0072] After the strategy is executed, the system continuously collects MR and PM data of the disturbed cell, sampling every 5 minutes, and compares the data of the same period before execution with the data of the 6 time windows after execution. If the average SINR increases by more than 5% and the interference power decreases by more than 10%, the strategy is considered successful; otherwise, a rollback script is triggered to restore the original power, tilt, frequency, or PCI parameters.

[0073] The closed-loop results are written to the knowledge base. If the strategy is successful, the success rate weight of the strategy in similar scenarios is increased; if the strategy fails or has side effects, the strategy weight is decreased and the side effect sub-area is recorded. If engineers correct the root cause or strategy, the system will add the correction as strong feedback to the training set and the knowledge base, and periodically generate model drift reports.

[0074] Example 3 This embodiment provides an external interference scenario where, in a certain urban area, cell B experiences an abnormal increase in PRB interference power for three consecutive 15-minute time windows, resulting in a decrease in the average SINR among users and an increase in complaints. The system triggers a troubleshooting script.

[0075] The first step involves the data base mapping PM, MR / CDT, frequency sweep, and complaint data of cell B and surrounding cells onto Geohash gratings, generating a dynamic snapshot of the current time window. The spatiotemporal map transformation layer detects signs of co-frequency interference in multiple gratings surrounding cell B.

[0076] In the second step, the CNN identifies a broadband, continuously rising feature in the spectral waterfall plot, indicating the highest probability of external interference. The SDAE reconstruction error significantly exceeds the threshold, and the abnormal clustering is similar to the historical "private amplifier" pattern. The gated fusion network adaptively increases the weights of the spectral texture embedding vector, generating an enhanced node feature vector h. cell .

[0077] Third, Causal-GNN combines dynamic graph edge weights and causal message passing to locate a specific cell to the northwest of cell B (or associate it with the corresponding grid area through spatial mapping) as a suspected source of interference, instead of misjudging cell B, which is the most severely affected cell, as the source of interference. SCM counterfactual verification shows that if external interference sources in this direction are blocked, the noise floor of the affected cell is expected to decrease, consistent with historical cases of privately installed amplifiers, and the causal consistency score is high.

[0078] Fourth, the knowledge base uses "external interference - privately installed amplifiers" as the primary key to search for cases, and combines this with filtering strategies based on the affected frequency band, azimuth angle, and complaint density to recommend a combined strategy of "triggering on-site frequency scanning and location work orders" and "temporary neighbor cell parameter optimization." Digital twin verification shows that temporary optimization has a low impact on surrounding KPIs.

[0079] Fifth, because this strategy involves on-site work orders rather than direct parameter changes, the orchestration engine automatically generates work orders and pushes them to maintenance personnel, while continuously monitoring cell B indicators. After the governance is completed, the system detects an increase in SINR and a decrease in interference power, and uses this case as a positive sample.

[0080] This embodiment demonstrates that the present invention enables the system to further infer the "external disturbance area" from the "abnormality of the disturbed cell" through spatiotemporal graph transformation, gating fusion and causal message passing, avoiding misjudgment based solely on the intensity of the anomaly, and forming an executable governance closed loop through knowledge base and twin verification.

[0081] Example 4 This embodiment provides a scenario of interference between neighboring cells on the same frequency. Cell D and Cell E are configured on the same frequency, and their antenna azimuth angles have a relative pointing relationship. Cell E experiences an increase in noise floor and a decrease in SINR, but Cell D's own KPIs are normal.

[0082] The system uses dynamic graph edge weights to identify a high frequency band overlap and azimuth coefficient between cells D and E. CNN identification results show local PRB elevation in the spectrum of cell E, and SDAE classifies cell E as an anomaly, but cell D is not marked as a strong anomaly. Conventional anomaly detection might only identify E as a problem cell.

[0083] In causal message passing, Causal-GNN identifies the potential interference direction of cell D on cell E based on SCM priors and combines counterfactual verification to confirm the prediction: if the power of cell D decreases by 2dB, the noise floor of cell E decreases and the coverage loss of cell D is limited. Digital twin further verifies that the impact of the strategy is small and the risk level is low. With a root cause confidence greater than 90%, the orchestration engine calls the atomic operation library to automatically execute the power reduction of cell D and records the original power value and the rollback script.

[0084] After execution, the SPC closed-loop verification showed that, when comparing the six consecutive time windows after execution with the baseline data of the same period before execution, the average SINR of cell E increased by more than 5% and the interference power decreased by more than 10%. The system determined that the strategy was successful and wrote the case into the knowledge base.

[0085] This embodiment demonstrates that the causal enhancement GNN of the present invention can identify scenarios where "the interfering cell itself is not obviously abnormal but affects neighboring cells." Through counterfactual verification and digital twin inference, the system can automatically execute low-risk strategies while controlling coverage risks, thereby improving interference management efficiency.

[0086] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

Claims

1. A method for investigating interference in residential communities based on AI and big data, characterized in that: Includes the following steps: Constructing a causal spatiotemporal data foundation: Accessing multi-source network data, establishing a three-dimensional spatiotemporal index and aggregating it to form a spatiotemporal feature cube, constructing a dynamic cell relationship graph based on the geographical and engineering parameter configuration relationships between cells, slicing and aggregating the spatiotemporal feature cube along the time axis, and outputting a dynamic graph snapshot sequence containing the initial node feature vectors; Multimodal interference fingerprint recognition: Based on dynamic graph snapshot sequence and spatiotemporal feature cube, basic statistical feature vector, spectral texture embedding vector and anomaly score are extracted respectively. The enhanced node feature vector is generated by fusing through multimodal gating mechanism, and then the dynamic graph snapshot sequence is updated to obtain the enhanced dynamic graph snapshot sequence; Causal Enhancement Graph Neural Network Root Cause Localization: The causal direction bias of the structural causal model is injected into the message passing process of the graph neural network. Based on the enhanced dynamic graph snapshot sequence, the scrambling probability of the candidate scrambling cell is output. After the confidence is corrected by counterfactual verification, the root cause localization tuple is output. Obtain the twin inference optimization strategy, retrieve candidate strategies from the scenario strategy knowledge base based on the root cause localization tuple, perform comparative inference simulation verification in the digital twin model, and then encapsulate and output the target execution strategy. Closed-loop self-healing orchestration and evolution: Based on the confidence of the root cause localization tuple and the risk level of the target execution strategy, a decision matrix is ​​constructed to trigger the execution action. The execution result is verified by statistical process control. The knowledge base is updated and the relevant model is trained based on the verification result. The process of injecting the causal direction bias of the structural causal model into the message passing of the graph neural network specifically includes: The structural causal model is mapped onto the candidate edges of the dynamic cell relationship graph. For any candidate edge, if there is a causal path between the two nodes, a positive causal prior matrix term is assigned; otherwise, a zero or negative value is assigned. In the node feature aggregation stage of the graph attention network, a causal attention bias term is introduced when calculating the attention coefficients between nodes. When the propagation direction is consistent with the causal direction corresponding to the causal prior matrix term, the message passing weight is enhanced, while the opposite direction is suppressed. After the confidence level is corrected through counterfactual verification, the root cause localization tuple is output, specifically including: For candidate disturbing cells, the changes in intervention parameters are simulated in the structural causal model to predict the expected changes in indicators of the disturbed cells; Simultaneously, the same feature intervention is performed on the causal augmentation graph neural network to predict the amount of change in the hidden layer response of the disturbed cell; The mean square error between the two variables is calculated to obtain a causal consistency score. The causal consistency score is used to correct the perturbation probability. When the final corrected confidence level is higher than a preset threshold, the root cause localization tuple containing the perturbation type, the list of perturbation cells, the causal evidence chain, and the confidence level is generated.

2. The method of claim 1, wherein: The construction of the causal spatiotemporal data base, which outputs a dynamic graph snapshot sequence containing the initial node feature vectors, specifically includes: Geospatial data is discretized using Geohash encoding, and a three-dimensional index containing cell identifiers, Geohash grids, and time windows is established. The aggregated features of each dimension are extracted to form the spatiotemporal feature cube. Candidate edges are established with cells as nodes. Edge weights are calculated based on the geographical distance between two cells, frequency band overlap rate, and antenna azimuth coefficient, and weakly associated edges are filtered to generate a dynamic cell relationship graph. Slice along the time axis, and perform pooling aggregation on the grid features covered by each cell within each time window to generate an initial node feature vector. Combine this with the updated edge weights to generate the dynamic graph snapshot sequence.

3. The method of claim 1, wherein: The enhanced node feature vector generated by the multimodal gating mechanism specifically includes: The basic statistical feature vector, the spectral texture embedding vector, and the anomaly score are each mapped to the same dimension through a linear transformation. Using the mapped modal vectors, spectral classification probabilities, and anomaly scores as gating inputs, attention scores for the corresponding modalities are generated through a weight matrix, and dynamic weights are obtained after normalization. The three mapping features are weighted and summed using the dynamic weights, and the enhanced node feature vector is output to update the node features of the corresponding cell.

4. The method according to claim 1, characterized in that: The steps of extracting basic statistical feature vectors, spectral texture embedding vectors, and anomaly scores, including the automatic detection of novel interference patterns, are as follows: The initial node feature vector is input into the trained stacked denoising autoencoder to calculate the scalar reconstruction error to obtain the anomaly score. When the scalar reconstruction error exceeds the adaptive threshold, it is marked as an abnormal cell. Clustering of the reconstructed residual vectors of abnormal cells reveals novel interference candidate clusters with similar deviation patterns. After confirmation, the corresponding spectral patterns are extracted as incremental samples to fine-tune the spectral classification model.

5. The method according to claim 1, characterized in that: The optimization strategy for obtaining twin inference specifically includes: The interference fingerprint vector of the current disturbed cell is concatenated with the root cause feature vector. Historical similar cases are recalled through the vector index library, and the structured index is used to filter by interference type and confidence range. The strategy set is extracted from the filtered cases, and a collaborative filtering algorithm combining historical success rate, impact range index and execution cost index is used to score and rank them, and recommended candidate strategies are output. In the digital twin base station model built based on historical network data, parameter changes are simulated for the recommended candidate strategies, and the expected improvement rate and a list of negatively impacted cells are output to determine the execution risk level.

6. The method according to claim 1, characterized in that: The closed-loop self-healing orchestration and evolution specifically include: The network metrics of multiple consecutive time windows after closed-loop execution are compared with the baseline data before execution using a statistical process control algorithm. If the policy execution is determined to fail, a rollback operation is triggered, and the interference characteristics and execution effect of the corresponding cell are written into the knowledge base as negative samples and the weight of the policy is reduced. If manual correction feedback is received, the corrected scrambling label is compared with the original output of the causal augmentation graph neural network to form a comparative learning sample pair, driving the model to fine-tune and retrain.

7. A community interference investigation system based on AI and big data, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute a computer program stored in the memory to implement the AI ​​and big data-based cell interference investigation method as described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the cell interference investigation method based on AI and big data as described in any one of claims 1 to 6.

Images