Base station power supply fault intelligent diagnosis method and system based on anomaly detection

Abstract

The invention discloses a base station power supply fault intelligent diagnosis method and system based on anomaly detection, and the method comprises the following steps: S1, collecting and preprocessing multi-modal monitoring data, and generating a multi-dimensional time series data set; s2, inputting the multi-dimensional time sequence data set into an improved ResNet-18 network for feature extraction, and generating a power supply fault feature vector; s3, inputting the power supply fault feature vector into an improved bee colony algorithm for anomaly detection, and outputting a power supply abnormal state mark and a power supply fault mode; s4, outputting a power supply fault diagnosis label through the classification function; s5, calling a power supply fault diagnosis label mapping rule, and outputting a power supply fault diagnosis result; and S6, based on the correlation matching degree, adjusting the convolutional layer weight of the improved ResNet-18 network and the fitness function parameter of the improved bee colony algorithm. According to the method, deep residual network modeling and a swarm intelligence anomaly identification mechanism are fused, and accurate identification, adaptive optimization and closed-loop diagnosis of the power supply fault of the base station are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and power fault diagnosis, and in particular to a method and system for intelligent diagnosis of base station power supply faults based on anomaly detection. Background Art

[0002] With the large-scale deployment of 5G communication infrastructure and the increasing intelligence of telecommunications networks, base stations are playing an increasingly critical role in urban communications, power dispatching, traffic control, and the Industrial Internet of Things. Base station power supplies are the core guarantee for the stable operation of communication equipment, and their operating status directly impacts the continuity and reliability of the communication network. However, due to factors such as power fluctuations, equipment aging, and environmental interference, base station power supply equipment frequently faces a variety of potential faults in actual operation, including short circuits, undervoltage, overheating, and battery failure. Accurately identifying and promptly addressing base station power supply faults has become a critical issue in current network operations and maintenance, as well as intelligent inspection systems.

[0003] Currently, mainstream methods for diagnosing base station power supply faults rely primarily on periodic manual inspections, remote monitoring and alarm systems, and status assessment mechanisms based on threshold-based strategies. While manual inspections can be effective for small-scale regional operations and maintenance, they face increasing challenges in the large-scale, high-density deployment of 5G macro and small base stations, including heavy labor burdens, long response times, and high rates of subjective misjudgment. This makes it difficult to support rapid, closed-loop fault resolution. Remote monitoring systems based on static alarm rules typically rely on fixed thresholds to identify anomalies in various power supply indicators (such as voltage, current, and temperature), triggering alarms when indicators deviate from the set thresholds. These approaches have significant limitations in practical deployments: First, due to individual differences between different device models and deployment environments, standardized threshold settings are often inadequate and prone to missed or false positives. Second, static rules cannot effectively capture temporal trends in device status and cannot identify incipient hazards, intermittent fluctuations, or gradual failure processes, thus reducing fault prediction and early warning capabilities. Third, the system lacks the ability to model correlations between multi-dimensional indicators, making it difficult to integrate and identify complex fault patterns.

[0004] Furthermore, while some existing fault diagnosis systems have incorporated machine learning algorithms for feature modeling and classification, most rely on static supervised learning methods and rely heavily on well-labeled fault samples for training. However, in real-world application scenarios, power supply fault events are characterized by low frequency, high risk, and a diverse range of types. Fault samples are difficult to obtain and extremely unevenly distributed, making model training prone to overfitting or insufficient recognition of small-category samples. Furthermore, many methods employ traditional feature engineering techniques or shallow network models, which fail to fully exploit the high-order temporal correlations inherent in multimodal monitoring data, resulting in weak generalization and robustness. Diagnostic accuracy significantly decreases when dealing with complex noisy environments or scenarios involving the coupling of multiple indicators. Furthermore, in current systems, the diagnostic module often operates separately from the model optimization process, lacking continuous learning capabilities based on diagnostic feedback. This makes it difficult to automatically update the model and adjust parameters based on historical labels, limiting the system's long-term adaptability and online self-evolution capabilities.

[0005] To address the above issues, the present invention proposes an intelligent diagnosis method and system for base station power supply faults based on anomaly detection. This method integrates an improved deep residual network (DRN) with an optimized bee swarm intelligence algorithm to construct a closed-loop diagnostic mechanism encompassing time series feature extraction, anomaly detection, diagnostic classification, and model self-feedback optimization. In terms of data processing, the system uniformly time-aligns and normalizes multi-source power supply monitoring data to form a multidimensional time series dataset adapted to the neural network architecture. In terms of feature modeling, an improved ResNet-18 network architecture is introduced, combining multi-branch one-dimensional convolution with an attention residual connection mechanism to achieve multi-scale extraction of key features. In the anomaly detection phase, an improved bee swarm algorithm is integrated to enable global search of the feature space and identify anomaly boundaries, improving the detection capability of small-sample faults. In the classification and output phase, a diagnostic label mapping mechanism is designed to perform category mapping and structured annotation for different fault states. In terms of model updating, the system uses similarity matching between current diagnostic labels and historical labels to trigger adaptive updates of model parameters. This combined learning and updating of network weights and optimization strategies enhances the system's generalization and evolution capabilities.

[0006] Therefore, based on the traditional base station power supply fault diagnosis system, the present invention overcomes the limitations of the existing technology in data expression ability, anomaly discrimination accuracy, strong sample dependence, poor adaptability, etc. by introducing a deep learning feature extraction structure, an improved swarm intelligent anomaly recognition strategy and a feedback-driven parameter self-update mechanism, and provides a base station power supply fault intelligent diagnosis solution with high robustness, high precision and self-evolution capability. Summary of the Invention

[0007] One purpose of the present invention is to propose a method and system for intelligent diagnosis of base station power supply faults based on anomaly detection. This method integrates multimodal monitoring data modeling, improved deep residual network feature extraction, a swarm intelligence-driven anomaly identification mechanism, and a diagnostic feedback optimization structure. This method systematically implements dynamic modeling, anomaly identification, fault classification, and model adaptive updating of the base station power supply operating status, thereby constructing a closed-loop intelligent diagnostic system of "acquisition-extraction-detection-classification-optimization." This method has the advantages of strong adaptability to small samples, high anomaly detection accuracy, and fast diagnostic feedback loops. It overcomes the technical bottlenecks of existing methods in heterogeneous data fusion, complex anomaly boundary identification, and online learning capabilities. It is suitable for intelligent inspection and fault warning scenarios of telecommunications base station power supplies in complex environments, enabling accurate identification, dynamic diagnosis, and continuous optimization of fault states.

[0008] According to an embodiment of the present invention, a method for intelligent diagnosis of base station power supply failure based on anomaly detection includes the following steps:

[0009] S1. Collect multimodal monitoring data of the base station power supply, perform preprocessing and timestamp alignment processing, and generate a multidimensional time series data set;

[0010] S2. Input the multidimensional time series dataset into an improved ResNet-18 network for feature extraction. The improved ResNet-18 network includes a multi-branch one-dimensional convolutional input layer and a residual connection structure embedded with an attention mechanism to generate a power failure feature vector.

[0011] S3. Inputting the power failure feature vector into an improved bee colony algorithm for anomaly detection. The improved bee colony algorithm performs a global search using a dynamic search radius and an adaptive guiding factor, performs anomaly identification processing based on feature clustering and cluster boundary detection, and outputs a power abnormality status mark and a power failure mode.

[0012] S4. Based on the power abnormality status mark and the power failure mode, output the power failure diagnosis label through the classification function;

[0013] S5. Based on the power supply fault diagnosis tag, call the power supply fault diagnosis tag mapping rule and output the power supply fault diagnosis result;

[0014] S6. Compare the power supply fault diagnosis label with the historical fault diagnosis label, calculate the relevant matching degree, and adjust the convolutional layer weights of the improved ResNet-18 network and the fitness function parameters of the improved bee swarm algorithm.

[0015] By constructing a complete closed-loop process of "acquisition-extraction-detection-classification-output-feedback optimization," this paper significantly improves the accuracy, adaptability, and real-time performance of base station power supply fault identification. First, unified acquisition and timestamp alignment of multimodal monitoring data ensures the integrity and synchronization of input data, providing a stable data foundation for subsequent feature modeling. Second, an improved ResNet-18 network architecture is introduced to perform deep feature extraction on multidimensional time series data, effectively enhancing the system's ability to represent power supply operating state changes. This is then combined with an improved bee swarm algorithm for anomaly detection, utilizing a dynamic search radius and adaptive guidance factor to intelligently search the feature space, improving the ability to identify faults with fuzzy boundaries, small sample sizes, or nonlinearity. The diagnostic results are output as power supply fault diagnostic labels through a mapping mechanism, which are then converted into structured diagnostic results, achieving highly readable status output. Furthermore, this method incorporates a parameter update mechanism based on label similarity feedback, enabling online adaptive model optimization and effectively improving the system's generalization across diverse device scenarios and historical fault samples. The overall method has the advantages of high feature expression accuracy, strong anomaly detection capability, and fast feedback loop, which significantly improves the intelligent diagnosis efficiency and accuracy of base station power supplies in complex environments. It is suitable for the deployment of automated fault identification and early warning systems in scenarios such as 5G communication operation and maintenance and remote power monitoring.

[0016] Optionally, the improved ResNet-18 network specifically includes:

[0017] The multi-branch one-dimensional convolution input layer receives a multi-dimensional time series data set and inputs each channel data in the multi-dimensional time series data set into an independent one-dimensional convolution branch for preliminary feature extraction. Each one-dimensional convolution branch outputs a channel feature vector and concatenates and fuses them to form a power supply timing feature representation.

[0018] The backbone residual network module includes multiple residual blocks, each of which contains a set of one-dimensional convolutional layers, batch normalization layers, and nonlinear activation function layers. In each residual block, the power timing feature representation is used as input and passed to the main path and the identity mapping path respectively. The main path performs a convolution transformation on the power timing feature representation to obtain the main path output feature, while the identity mapping path keeps the power timing feature representation unchanged. Finally, at the end of the residual block, the main path output feature and the power timing feature representation are element-by-element added to output the residual enhanced intermediate feature.

[0019] The embedded attention mechanism module is set between multiple residual blocks to model the inter-channel correlation of the residual enhanced intermediate features, perform attention modulation according to the attention weights, and output the attention enhanced feature representation;

[0020] The global average pooling layer and the fully connected layer perform dimensionality reduction and feature vector mapping on the attention-enhanced feature representation to generate the power fault feature vector.

[0021] By designing an improved ResNet-18 network structure, the feature modeling capability of multi-dimensional time series data of base station power supplies has been effectively improved. The network introduces a multi-branch one-dimensional convolution input layer, which can process each type of monitoring indicator in parallel, thereby fully retaining the dynamic evolution information between each channel. At the same time, an attention mechanism module is embedded in the residual connection structure, so that the network can adaptively focus on key time series change features when extracting deep features, suppress redundant information interference, and thus improve the discrimination and robustness of the overall feature vector. While ensuring computational efficiency, the network significantly enhances the model's sensitivity to slight abnormal fluctuations and low-amplitude fault signals. Compared with traditional network structures, it has stronger diagnostic expression capabilities in low signal-to-noise ratio scenarios, improving the front-end perception accuracy of the entire fault detection system.

[0022] Optionally, the S2 specifically includes:

[0023] S21, multi-branch one-dimensional convolution input layer, receives the multi-dimensional time series data set X = [x1, x2, ..., x C ], where x c ∈R T , x c is the time series data of the cth channel, with a length of T and C being the number of channels;

[0024] S22, x c Input the corresponding one-dimensional convolution branches respectively, perform convolution, batch normalization and gated activation function processing in sequence, and calculate the channel gated feature vector:

[0025] g c =BN(x c *w c +b c )·Sigmoid(BN(x c *w c +b c ));

[0026] Among them, g c is the c-th channel gated feature vector, Sigmoid(·) is the Sigmoid activation function, BN(·) is the batch normalization operation, x c is the time series data of the cth channel, w c is the one-dimensional convolution branch convolution kernel corresponding to the c-th channel, b c is the convolution bias term of the one-dimensional convolution branch corresponding to the c-th channel;

[0027] S23, all channel gate feature vectors g1, g2, ..., g C Perform feature dimension splicing to form a fused feature representation:

[0028] F0=Concat(g1,g2,…,g C );

[0029] Among them, F0 is the power timing feature representation, Concat(·) is the splicing operation;

[0030] S24, input the power timing feature representation F0 into the trunk residual network module, and extract features through multiple residual blocks in sequence; in the i-th residual block, construct the main path and the identity mapping path respectively, and the main path inputs the feature F to the i-1th residual block i-1 Perform one-dimensional convolution, batch normalization and activation operations to obtain the main path output feature R i :

[0031] R i =ReLU(BN(F i-1 *W i +b i ));

[0032] Among them, ReLU(·) is the activation function, BN(·) is the batch normalization operation, W i is the one-dimensional convolution kernel of the i-th residual block, b i is the convolution bias term of the i-th residual block; the identity mapping path directly retains F i-1 :

[0033] I i =F i-1 ;

[0034] Among them, I i Output features for the identity mapping path; Output features for the main path R i Output feature I with the identity mapping path i Perform element-by-element addition to calculate the residual enhanced intermediate feature F i :

[0035] F i =R i +I i ;

[0036] S25, the embedded attention mechanism module receives the residual enhanced intermediate feature F i , calculate the channel attention vector and time attention vector respectively, and generate the channel-time joint attention matrix A through the outer product i , perform attention modulation according to the attention weight and output the attention enhanced feature representation F′ i :

[0037] F′ i =σ(A i )⊙F i ;

[0038] Where σ(·) is the activation function;

[0039] S26. Represent the attention-enhanced feature F′ i The global average pooling layer and the fully connected layer are input to perform dimensionality reduction and feature vector mapping, and the power fault feature vector is output.

[0040] The optimized ResNet-18 network performs deep feature extraction on preprocessed multi-dimensional time series data. The network utilizes a multi-branch one-dimensional convolutional channel structure to independently model each dimension of power monitoring indicators. An embedded attention mechanism enhances the features of residual connection paths, effectively capturing correlations and changing trends between indicators. Compared with traditional shallow feature extraction or static threshold analysis, this method significantly improves the model's ability to represent non-explicit fault modes such as minor power fluctuations and intermittent anomalies, and enhances its ability to model complex nonlinear relationships. The resulting power fault feature vector exhibits high discrimination and compressibility, providing a key diagnostically valuable representation for the anomaly detection stage.

[0041] Optionally, the improved bee colony algorithm performs global search through a dynamic search radius and an adaptive guide factor, and specifically includes the following steps based on feature clustering and cluster boundary detection:

[0042] The dynamic search radius is adjusted based on the feature space distance between the search swarm individual and the current optimal swarm individual to control the local search range of the swarm individual during the iteration process;

[0043] The adaptive guidance factor is dynamically calculated based on the relative difference between the individual fitness of the swarm and the global optimal swarm individual fitness, and adjusts the direction and intensity of the swarm individual's deviation toward the global optimal position and the local cluster center;

[0044] The abnormality identification mechanism is based on the feature density clustering strategy. During the search process, the power failure feature vectors whose distance from each swarm individual position is greater than the search radius are judged as abnormal states, and the corresponding power failure mode is output in combination with the spatial position distribution.

[0045] By introducing a dynamic search radius and an adaptive guidance factor mechanism, the algorithm's ability to search for abnormal power supply states in high-dimensional feature spaces is significantly enhanced. Traditional swarm intelligence algorithms are prone to falling into local optimality when fault samples are sparse and boundaries are fuzzy. However, the present invention uses a radius adjustment mechanism based on cluster density to allow the individual search range to be adaptively adjusted according to the feature distribution, thereby improving the coverage of abnormal boundary samples. At the same time, the guidance factor constructed based on fitness differences makes the search direction more precise and effectively shortens the convergence path. The overall algorithm takes into account both global search and local refinement, and is particularly suitable for processing power supply fault feature spaces with nonlinear coupling and asymmetric distribution, improving the accuracy of anomaly recognition and the model's ability to detect small sample anomalies.

[0046] Optionally, the S3 specifically includes:

[0047] S31. Initialize the individual positions of the swarm, map the power failure feature vector to the initial search point in the search space dimension, set the search space dimension to d, the population size to N, and record the individual position of each swarm as Where j = 1, 2, 3, ..., N;

[0048] S32, introduce a dynamic search radius update mechanism, and adaptively adjust the search radius of each round according to the Euclidean distance between the individual fitness value of the swarm and the optimal solution of the group.

[0049]

[0050] Among them, ρ0 is the initial radius, λ is the control factor, is the current optimal individual position of the swarm, exp(·) is the exponential function;

[0051] S33, fusion adaptive guidance factor According to the relative level of individual fitness values ​​of the swarm, guide the swarm individuals to move between the optimal swarm individual position and the feature cluster center:

[0052]

[0053] Among them, α and β are the guiding direction control factors, is the feature cluster center, is the individual position of the swarm after moving, and the adaptive guidance factor The calculation method is:

[0054]

[0055] in, is the optimal individual fitness value of the bee colony, is the fitness value of the j-th swarm individual in the t-th iteration, ∈ is a constant to prevent division by zero;

[0056] S34, based on the feature clustering method of bee colony density, performs cluster analysis on the search individual position set and performs anomaly detection at the distance boundary. For each input power fault feature vector z k , if the following conditions are met:

[0057]

[0058] The power failure feature vector is determined to be an abnormal state, and the power abnormal state mark and power failure mode are output in combination with the corresponding cluster position.

[0059] An improved bee colony algorithm is introduced to perform global anomaly detection on the extracted power supply fault feature vectors. A dynamic search radius and adaptive guidance factor mechanism enable adaptive exploration of the high-dimensional feature space, compensating for the inadequate ability of traditional clustering or threshold methods to identify boundary samples. The algorithm maintains high anomaly detection sensitivity and accuracy even in scenarios with sparse data distribution, unbalanced categories, or a very small number of fault samples. Combined with a cluster boundary determination mechanism, it not only labels abnormal states but also preliminarily classifies potential fault types, providing clear structural input for the subsequent classification stage. This detection method, based on intelligent search and dynamic clustering, effectively improves the diagnostic system's adaptability and anomaly detection capabilities in complex scenarios.

[0060] Optionally, the S4 specifically includes:

[0061] S41, receiving the power supply abnormal state mark and the power supply failure mode, constructing the abnormal mark-failure mode feature pair (m j ,p j ), where m j ∈{0,1} represents the jth power supply abnormality status mark, p j represents the jth power supply failure mode;

[0062] S42, for each set of abnormality mark-fault mode feature pair (m j ,p j ) performs power fault diagnosis label generation operation, using the classification function f class Output (m j ,p j ) corresponding to the power supply fault diagnosis label y j :

[0063] y j =f class (m j ,p j );

[0064] S43, aggregate all power supply fault diagnosis labels to generate a power supply fault diagnosis label set Y = {y1, y2, ..., y N}, where N is the total number of power fault diagnosis tags.

[0065] By taking anomaly detection results as input and combining them with power supply fault pattern information to perform operational status classification, a structured power supply fault diagnostic label is generated. By establishing a label mapping mechanism, an efficient transition from feature space to category space is achieved, enabling the original anomaly features and status tags to be clearly mapped to specific fault categories or operational states. This label has a standardized representation, facilitating system integration and visualization. Compared to classification methods that rely on manual experience or simple rules, this step improves classification consistency and automation, effectively reducing the rate of false positives and the cost of human intervention, and providing a clear classification basis for subsequent diagnostic output and model optimization.

[0066] Optionally, the S5 specifically includes:

[0067] S51, receiving a power supply fault diagnosis tag set Y = {y1, y2, ..., y N};

[0068] S52, construct power fault diagnosis label mapping rule function f map :Y→R, change the power fault diagnosis label to y j Mapped to power failure diagnosis result r j :

[0069] r j =f map (y j ).

[0070] Through the mapping rules between diagnostic labels and diagnostic results, the conversion of labels to final diagnostic results is completed, realizing a closed-loop transformation from the analysis results within the model to the system output layer. This mechanism supports multi-dimensional output formats, including category codes, status descriptions, anomaly levels, etc., and has good structural scalability and result expression capabilities. Compared with traditional alarm systems that only output a single identifier, this step can provide more refined fault information, providing targeted repair basis and operation suggestions for the operation and maintenance system. In addition, the mapping rules can be flexibly adjusted according to actual operational needs, with good configuration flexibility, which improves the system's applicability and management coordination in multiple operating scenarios.

[0071] Optionally, the S6 specifically includes:

[0072] S61, based on power supply fault diagnosis label y j and historical fault diagnosis label y′ k , build a matching calculation function, and calculate the power supply-historical fault label pair (yj ,y′ k ) performs the correlation matching calculation to generate the correlation matching matrix S:

[0073] S j,k =Sim(y j ,y′ k );

[0074] Among them, Sim(·) is the matching degree calculation function, S j,k is the power supply-history fault label pair (y j ,y′ k )’s matching value;

[0075] S62, extract S according to the set threshold δ j,k The power supply-historical fault label pairs with a value no less than δ are used to construct a diagnostic label matching set:

[0076] Y match ={(y j ,y′ k )∣S j,k ≥δ};

[0077] S63. Statistical diagnostic label matching set Y match The number of power supply-historical fault label pairs n is set. When n meets the set update threshold, the update operation is performed to adjust the convolutional layer weights of the improved ResNet-18 network:

[0078]

[0079] in, is the convolutional layer weight of the improved ResNet-18 network in the tth round, is the convolutional layer weight of the improved ResNet-18 network in the t+1th round, η θ The learning rate for the improved ResNet-18 network update, L match is the loss function;

[0080] S64. Adjust the fitness function parameters of the improved bee colony algorithm:

[0081]

[0082] in, is the fitness function parameter of the improved bee swarm algorithm in the tth round, is the fitness function parameter of the improved bee colony algorithm in the t+1th round, η φ The learning rate updated by the improved bee swarm algorithm, L match is the loss function.

[0083] By constructing a dynamic feedback mechanism based on historical diagnostic labels and a similarity function to calculate the match between the current label and the historical label, and triggering adaptive adjustment of the model parameters when the match meets the update threshold condition, this mechanism can perform targeted updates to the convolutional layer weights of the improved ResNet-18 network and the fitness function parameters of the swarm algorithm, thereby realizing a data-driven optimization process based on historical experience. Unlike traditional models that rely on offline training and static deployment, this step realizes the continuous evolution of the model, with online updates and dynamic adaptation characteristics, significantly improving the model's self-correction ability and anti-degradation performance during long-term operation.

[0084] According to an embodiment of the present invention, a base station power supply fault intelligent diagnosis system based on anomaly detection includes the following modules:

[0085] The data acquisition and preprocessing module is used to collect multimodal monitoring data of the base station power supply, standardize the multimodal monitoring data and align the timestamps to generate a multidimensional time series data set;

[0086] The abnormal feature extraction and detection module is used to input the multidimensional time series data set into the improved ResNet-18 network, extract the power failure feature vector, and input the power failure feature vector into the improved bee swarm algorithm to perform anomaly detection based on dynamic search radius and adaptive guidance factor, and output the power abnormal status mark and power failure mode;

[0087] The fault diagnosis and result generation module is used to generate a power supply fault diagnosis label based on the power supply abnormal state mark and the power supply fault mode; and input the power supply fault diagnosis label into the label mapping rule function to output the power supply fault diagnosis result;

[0088] The incremental update module compares the power supply fault diagnosis label with the historical fault diagnosis label and determines whether to perform the update operation based on the relevant matching value. When the update conditions are met, the module dynamically adjusts the convolutional layer weights of the improved ResNet-18 network and the fitness function parameters of the improved bee swarm algorithm.

[0089] The diagnostic result output module is used to output the power supply fault diagnosis results to the operation and maintenance terminal in the form of text description, status code or graphical visualization.

[0090] Through modularization, a clear-structured and closed-loop intelligent diagnosis system for base station power supply faults was constructed, covering the entire process of data collection, feature extraction, anomaly detection, fault classification, result output, and model adaptive update. Based on multimodal data preprocessing, the system integrates deep feature modeling and intelligent anomaly recognition modules to achieve high-precision perception of complex power supply operating states. At the same time, through parameter updates and incremental training mechanisms driven by diagnostic feedback, the system has online learning and model evolution capabilities to adapt to changes in fault types in different scenarios. The system supports standard output of structured diagnostic results, which is convenient for integration and deployment with the operation and maintenance platform. It has high robustness, scalability, and engineering implementation value, and can significantly improve the power supply operation safety assurance capabilities of communication base stations in complex environments.

[0091] The beneficial effects of the present invention are:

[0092] The present invention significantly improves the perception and fault response capabilities of complex power supply operating states by constructing a set of intelligent diagnosis methods and systems for base station power supply faults based on anomaly detection. In the feature extraction stage, the present invention introduces a structurally improved ResNet-18 network model, which, combined with a multi-branch one-dimensional convolution input channel and an embedded attention mechanism, can fully mine the key feature patterns in multi-dimensional time series monitoring data and achieve deep modeling of multi-index data such as voltage, current, and temperature at different time scales. Compared with traditional shallow models or static feature engineering methods, this network architecture can still have good feature expression and discrimination capabilities in the face of noise interference or equipment differences, providing a stable and reliable representation basis for subsequent anomaly detection.

[0093] During the anomaly detection phase, this paper introduces an improved bee colony algorithm. Through a dynamic search radius adjustment mechanism and adaptive guidance factor design, the search individuals possess stronger local focus and global exploration capabilities in the feature space. This algorithm, combining feature clustering with boundary recognition strategies, effectively compensates for the insufficient precision of traditional clustering methods when dealing with small samples or abnormal samples with blurred boundaries, thereby improving the accuracy and sensitivity of power supply fault pattern recognition. In particular, when faced with nonlinear feature changes or atypical fault patterns, the improved bee colony search mechanism can dynamically adapt the search direction and step size, enhancing the robustness and generalization of anomaly detection.

[0094] In the fault classification and result generation stage, the present invention designs a classification mapping mechanism jointly driven by abnormal state marking and fault mode. By constructing a diagnostic label set and a diagnostic result mapping rule, it realizes the structured expression from the underlying features to the final classification label, effectively avoiding the misjudgment problem caused by relying on a single threshold or static model in the traditional label judgment logic. At the same time, the present invention further introduces a diagnostic feedback-driven model update mechanism. Based on the similarity matching between the current diagnostic label and the historical label, the matching matrix is ​​calculated, and the convolutional layer weights of the ResNet-18 network and the fitness function parameters in the swarm algorithm are dynamically adjusted under the premise of meeting the update conditions. This process works in conjunction with gradient feedback and adaptive gating mechanisms to construct a closed-loop optimization path at the parameter level, so that the model can self-correct and dynamically adapt according to the historical evolution of the fault, effectively improving the long-term response capability of the diagnostic system to external disturbances such as environmental changes and equipment aging.

[0095] Furthermore, this invention integrates online sample collection and incremental training modules, continuously capturing new abnormal samples during the fault diagnosis process and incorporating them into the incremental learning process for real-time model replenishment and retraining. This overcomes the technical bottleneck of traditional models that rely on static training data and struggle to adapt to scenario migration. By combining online learning with heterogeneous model structure adjustments, this system possesses long-term evolution capabilities and fault migration adaptability, effectively extending the model lifecycle and improving full-cycle diagnostic performance.

[0096] At the output level, the present invention supports multi-format, multi-dimensional presentation of diagnostic information, including fault category identification, temporal change trends, abnormal indicator dimensions and their weighted contributions, etc., adapting to the needs of different operation and maintenance management systems and enhancing the system's integrated value in scheduling support, operation and maintenance decision-making, and intelligent early warning. Ultimately, the present invention constructs an intelligent base station power supply fault diagnosis framework that integrates perception, recognition, classification, and optimization. This framework breaks through the key limitations of existing technologies in low-sample learning, complex feature modeling, dynamic model updates, and real-time diagnostic feedback, providing stable, efficient, and self-evolving technical support for intelligent inspection of telecommunications infrastructure. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0098] Figure 1 This is an overall flow chart of a method for intelligent diagnosis of base station power supply failure based on anomaly detection proposed by the present invention;

[0099] Figure 2 This is a structural diagram of a base station power supply fault intelligent diagnosis system based on anomaly detection proposed by the present invention. DETAILED DESCRIPTION

[0100] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0101] refer to Figure 1 , a base station power supply fault intelligent diagnosis method based on anomaly detection, comprising the following steps:

[0102] S1. Collect multimodal monitoring data of the base station power supply, perform preprocessing and timestamp alignment processing, and generate a multidimensional time series data set;

[0103] S2. Input the multidimensional time series dataset into an improved ResNet-18 network for feature extraction. The improved ResNet-18 network includes a multi-branch one-dimensional convolutional input layer and a residual connection structure embedded with an attention mechanism to generate a power failure feature vector.

[0104] S3. Inputting the power failure feature vector into an improved bee colony algorithm for anomaly detection. The improved bee colony algorithm performs a global search using a dynamic search radius and an adaptive guiding factor, performs anomaly identification processing based on feature clustering and cluster boundary detection, and outputs a power abnormality status mark and a power failure mode.

[0105] S4. Based on the power abnormality status mark and the power failure mode, output the power failure diagnosis label through the classification function;

[0106] S5. Based on the power supply fault diagnosis tag, call the power supply fault diagnosis tag mapping rule and output the power supply fault diagnosis result;

[0107] S6. Compare the power supply fault diagnosis label with the historical fault diagnosis label, calculate the relevant matching degree, and adjust the convolutional layer weights of the improved ResNet-18 network and the fitness function parameters of the improved bee swarm algorithm.

[0108] In this embodiment, the improved ResNet-18 network specifically includes:

[0109] The multi-branch one-dimensional convolution input layer receives a multi-dimensional time series data set and inputs each channel data in the multi-dimensional time series data set into an independent one-dimensional convolution branch for preliminary feature extraction. Each one-dimensional convolution branch outputs a channel feature vector and concatenates and fuses them to form a power supply timing feature representation.

[0110] The backbone residual network module includes multiple residual blocks, each of which contains a set of one-dimensional convolutional layers, batch normalization layers, and nonlinear activation function layers. In each residual block, the power timing feature representation is used as input and passed to the main path and the identity mapping path respectively. The main path performs a convolution transformation on the power timing feature representation to obtain the main path output feature, while the identity mapping path keeps the power timing feature representation unchanged. Finally, at the end of the residual block, the main path output feature and the power timing feature representation are element-by-element added to output the residual enhanced intermediate feature.

[0111] The embedded attention mechanism module is set between multiple residual blocks to model the inter-channel correlation of the residual enhanced intermediate features, perform attention modulation according to the attention weights, and output the attention enhanced feature representation;

[0112] The global average pooling layer and the fully connected layer perform dimensionality reduction and feature vector mapping on the attention-enhanced feature representation to generate the power fault feature vector.

[0113] In this embodiment, S2 specifically includes:

[0114] S21, multi-branch one-dimensional convolution input layer, receives the multi-dimensional time series data set X = [x1, x2, ..., x C ], where x c ∈R T , x c is the time series data of the cth channel, with a length of T and C being the number of channels;

[0115] S22, x c Input the corresponding one-dimensional convolution branches respectively, perform convolution, batch normalization and gated activation function processing in sequence, and calculate the channel gated feature vector:

[0116] g c =BN(x c *w c +b c )·Sigmoid(BN(x c *w c +b c ));

[0117] Among them, g c is the c-th channel gated feature vector, Sigmoid(·) is the Sigmoid activation function, BN(·) is the batch normalization operation, x c is the time series data of the cth channel, w c is the one-dimensional convolution branch convolution kernel corresponding to the c-th channel, b c is the convolution bias term of the one-dimensional convolution branch corresponding to the c-th channel;

[0118] S23, all channel gate feature vectors g1, g2, ..., g C Perform feature dimension splicing to form a fused feature representation:

[0119] F0=Concat(g1,g2,…,g C );

[0120] Among them, F0 is the power timing feature representation, Concat(·) is the splicing operation;

[0121] S24, input the power timing feature representation F0 into the trunk residual network module, and extract features through multiple residual blocks in sequence; in the i-th residual block, construct the main path and the identity mapping path respectively, and the main path inputs the feature F to the i-1th residual block i-1 Perform one-dimensional convolution, batch normalization and activation operations to obtain the main path output feature R i :

[0122] R i =ReLU(BN(F i-1 *W i +b i ));

[0123] Among them, ReLU(·) is the activation function, BN(·) is the batch normalization operation, W i is the one-dimensional convolution kernel of the i-th residual block, b i is the convolution bias term of the i-th residual block; the identity mapping path directly retains F i-1 :

[0124] I i =F i-1 ;

[0125] Among them, I i Output features for the identity mapping path; Output features for the main path R i Output feature I with the identity mapping path i Perform element-by-element addition to calculate the residual enhanced intermediate feature F i :

[0126] F i =R i +I i ;

[0127] S25, the embedded attention mechanism module receives the residual enhanced intermediate feature F i , calculate the channel attention vector and time attention vector respectively, and generate the channel-time joint attention matrix A through the outer product i , perform attention modulation according to the attention weight and output the attention enhanced feature representation F′ i :

[0128] F′ i =σ(A i )⊙F i ;

[0129] Where σ(·) is the activation function;

[0130] S26. Represent the attention-enhanced feature F′ i The global average pooling layer and the fully connected layer are input to perform dimensionality reduction and feature vector mapping, and the power fault feature vector is output.

[0131] In this embodiment, the improved bee colony algorithm performs global search through a dynamic search radius and an adaptive guiding factor, and specifically includes the following steps based on feature clustering and cluster boundary detection:

[0132] The dynamic search radius is adjusted based on the feature space distance between the search swarm individual and the current optimal swarm individual to control the local search range of the swarm individual during the iteration process;

[0133] The adaptive guidance factor is dynamically calculated based on the relative difference between the individual fitness of the swarm and the global optimal swarm individual fitness, and adjusts the direction and intensity of the swarm individual's deviation toward the global optimal position and the local cluster center;

[0134] The abnormality identification mechanism is based on the feature density clustering strategy. During the search process, the power failure feature vectors whose distance from each swarm individual position is greater than the search radius are judged as abnormal states, and the corresponding power failure mode is output in combination with the spatial position distribution.

[0135] In this embodiment, S3 specifically includes:

[0136] S31. Initialize the individual positions of the swarm, map the power failure feature vector to the initial search point in the search space dimension, set the search space dimension to d, the population size to N, and record the individual position of each swarm as Where j = 1, 2, 3, ..., N;

[0137] S32, introduce a dynamic search radius update mechanism, and adaptively adjust the search radius of each round according to the Euclidean distance between the individual fitness value of the swarm and the optimal solution of the group.

[0138]

[0139] Among them, ρ0 is the initial radius, λ is the control factor, is the current optimal individual position of the swarm, exp(·) is the exponential function;

[0140] S33, fusion adaptive guidance factor According to the relative level of individual fitness values ​​of the swarm, guide the swarm individuals to move between the optimal swarm individual position and the feature cluster center:

[0141]

[0142] Among them, α and β are the guiding direction control factors, is the feature cluster center, is the individual position of the swarm after moving, and the adaptive guidance factor The calculation method is:

[0143]

[0144] in, is the optimal individual fitness value of the bee colony, is the fitness value of the j-th swarm individual in the t-th iteration, ∈ is a constant to prevent division by zero;

[0145] S34, based on the feature clustering method of bee colony density, performs cluster analysis on the search individual position set and performs anomaly detection at the distance boundary. For each input power fault feature vector z k , if the following conditions are met:

[0146]

[0147] The power failure feature vector is determined to be an abnormal state, and the power abnormal state mark and power failure mode are output in combination with the corresponding cluster position.

[0148] In this embodiment, the S4 specifically includes:

[0149] S41, receiving the power supply abnormal state mark and the power supply failure mode, constructing the abnormal mark-failure mode feature pair (m j ,p j ), where m j ∈{0,1} represents the jth power supply abnormality status mark, p j represents the jth power supply failure mode;

[0150] S42, for each set of abnormality mark-fault mode feature pair (m j ,p j ) performs power fault diagnosis label generation operation, using the classification function f class Output (m j ,p j ) corresponding to the power supply fault diagnosis label y j :

[0151] y j =f class(m j ,p j );

[0152] S43, aggregate all power supply fault diagnosis labels to generate a power supply fault diagnosis label set Y = {y1, y2, ..., y N}, where N is the total number of power fault diagnosis tags.

[0153] In this embodiment, the S5 specifically includes:

[0154] S51, receiving a power supply fault diagnosis tag set Y = {y1, y2, ..., y N};

[0155] S52, construct power fault diagnosis label mapping rule function f map :Y→R, change the power fault diagnosis label to y j Mapped to power failure diagnosis result r j :

[0156] r j =f map (y j ).

[0157] In this embodiment, S6 specifically includes:

[0158] S61, based on power supply fault diagnosis label y j and historical fault diagnosis label y′ k , build a matching calculation function, and calculate the power supply-historical fault label pair (y j ,y′ k ) performs the correlation matching calculation to generate the correlation matching matrix S:

[0159] S j,k =Sim(y j ,y′ k );

[0160] Among them, Sim(·) is the matching degree calculation function, S j,k is the power supply-history fault label pair (y j ,y′ k )’s matching value;

[0161] S62, extract S according to the set threshold δ j,k The power supply-historical fault label pairs with a value no less than δ are used to construct a diagnostic label matching set:

[0162] Y match ={(y j ,y′ k )∣S j,k ≥δ};

[0163] S63. Statistical diagnostic label matching set Y match The number of power supply-historical fault label pairs n is set. When n meets the set update threshold, the update operation is performed to adjust the convolutional layer weights of the improved ResNet-18 network:

[0164]

[0165] in, is the convolutional layer weight of the improved ResNet-18 network in the tth round, is the convolutional layer weight of the improved ResNet-18 network in the t+1th round, η θ The learning rate for the improved ResNet-18 network update, L match is the loss function;

[0166] S64. Adjust the fitness function parameters of the improved bee colony algorithm:

[0167]

[0168] in, is the fitness function parameter of the improved bee swarm algorithm in the tth round, is the fitness function parameter of the improved bee colony algorithm in the t+1th round, η φ The learning rate updated by the improved bee swarm algorithm, L match is the loss function.

[0169] refer to Figure 2 , an intelligent diagnosis system for base station power supply faults based on anomaly detection, including the following modules:

[0170] The data acquisition and preprocessing module is used to collect multimodal monitoring data of the base station power supply, standardize the multimodal monitoring data and align the timestamps to generate a multidimensional time series data set;

[0171] The abnormal feature extraction and detection module is used to input the multidimensional time series data set into the improved ResNet-18 network, extract the power failure feature vector, and input the power failure feature vector into the improved bee swarm algorithm to perform anomaly detection based on dynamic search radius and adaptive guidance factor, and output the power abnormal status mark and power failure mode;

[0172] The fault diagnosis and result generation module is used to generate a power supply fault diagnosis label based on the power supply abnormal state mark and the power supply fault mode; and input the power supply fault diagnosis label into the label mapping rule function to output the power supply fault diagnosis result;

[0173] The incremental update module compares the power supply fault diagnosis label with the historical fault diagnosis label and determines whether to perform the update operation based on the relevant matching value. When the update conditions are met, the module dynamically adjusts the convolutional layer weights of the improved ResNet-18 network and the fitness function parameters of the improved bee swarm algorithm.

[0174] The diagnostic result output module is used to output the power supply fault diagnosis results to the operation and maintenance terminal in the form of text description, status code or graphical visualization.

[0175] Example 1:

[0176] In order to verify the feasibility and application effect of the present invention in actual industrial scenarios, in June 2024, a large-scale telecommunications operator deployed and implemented the "A method and system for intelligent diagnosis of base station power failures based on anomaly detection" proposed in the present invention at 5 typical macro base stations (numbered 01 to 05) in South China. These base stations have long faced typical problems such as complex power failure types, frequent failure of traditional alarm rules, and delayed response to manual inspections. Especially during the high-load period in summer, the frequency of failures remains high, and the pressure on manual operation and maintenance is extremely high. The average monthly number of failures fluctuates between 4.8 and 6.1, while the fault identification accuracy rate has long hovered between 68% and 70%. Traditional systems find it difficult to promptly detect potential anomalies caused by battery aging, power fluctuation interference, cold standby system lags, etc., which affect the continuous power supply and signal stability of base stations.

[0177] After the system is deployed, existing monitoring sensors are first connected to each base station, including data channels such as voltage, current, frequency, temperature and humidity, and uniformly sent to the system's data acquisition and preprocessing module for standardization and time alignment. After processing with the improved ResNet-18 network, the system can automatically extract key fault features from multi-dimensional time series. Subsequently, an improved bee swarm algorithm is used for anomaly detection. By introducing a dynamic search radius and an adaptive guidance factor mechanism, highly sensitive recognition of weak abnormal signals is achieved. Each recognition process automatically outputs a diagnostic label and performs label mapping. Finally, the system generates a structured diagnostic result and synchronously pushes it to the operation and maintenance platform, realizing an intelligent flow from fault detection to feedback loop.

[0178] To enhance the model's adaptability, the system also incorporates a feedback optimization mechanism based on historical label matching. When the current diagnostic results deviate significantly from historical database samples, the system automatically fine-tunes the neural network weights and swarm algorithm parameters. Furthermore, the system collects new fault samples in real time during operation to form an incremental training set, continuously optimizing model performance through online learning. This mechanism overcomes the limitations of traditional static model training, enabling the system to continuously learn new fault characteristics, gradually improving generalization and scenario adaptability during actual operation.

[0179] In terms of implementation effects, the results of two consecutive months of operation monitoring of the above five base stations are summarized as follows:

[0180] Table 1 Comparative data of diagnosis before and after implementation of the present invention

[0181]

[0182] As can be seen from the data in Table 1, the present invention has demonstrated significant advantages in fault prevention, diagnostic accuracy, and response efficiency. The average monthly number of faults has been reduced from 5.38 times before deployment to 1.22 times, with a decrease of 77.3%, significantly reducing the risk of power outages and the frequency of manual inspections. In terms of recognition accuracy, it has increased from the original average of 68.5% to over 92.5%, with the maximum improvement reaching 25.1%, demonstrating the high expressiveness and recognition robustness of the improved network model and intelligent anomaly detection algorithm. The overall system response time has been shortened by 18 to 21 minutes, effectively compressing the "diagnosis-feedback-decision-making" link and improving operation and maintenance scheduling efficiency. At the same time, in the new fault identification dimension, the system has an average of 3 new identifiable fault types for each base station, covering complex situations such as intermittent battery failure, UPS inverter drift anomalies, and load switching lag, making up for the problem of insufficient coverage of traditional template matching systems.

[0183] In summary, this invention significantly improves the stability, operational efficiency, and adaptability of base station power systems by constructing a closed-loop intelligent diagnostic mechanism encompassing feature extraction, anomaly detection, label classification, model updating, and result output. This system not only addresses the accuracy, real-time performance, and coverage bottlenecks of traditional fault identification methods, but also enables continuous learning for new fault types. It demonstrates excellent engineering feasibility and versatility in multi-site deployments, and has the potential for large-scale application.

[0184] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for intelligent diagnosis of base station power supply failure based on anomaly detection, characterized in that: The steps include: S1. Collect multimodal monitoring data of the base station power supply, perform preprocessing and timestamp alignment processing, and generate a multidimensional time series data set; S2. Input the multidimensional time series dataset into an improved ResNet-18 network for feature extraction. The improved ResNet-18 network includes a multi-branch one-dimensional convolutional input layer and a residual connection structure embedded with an attention mechanism to generate a power failure feature vector. S3. Inputting the power failure feature vector into an improved bee colony algorithm for anomaly detection. The improved bee colony algorithm performs a global search using a dynamic search radius and an adaptive guiding factor, performs anomaly identification processing based on feature clustering and cluster boundary detection, and outputs a power abnormality status mark and a power failure mode. S4. Based on the power abnormality status mark and the power failure mode, output the power failure diagnosis label through the classification function; S5. Based on the power supply fault diagnosis tag, call the power supply fault diagnosis tag mapping rule and output the power supply fault diagnosis result; S6. Compare the power supply fault diagnosis label with the historical fault diagnosis label, calculate the relevant matching degree, and adjust the convolutional layer weights of the improved ResNet-18 network and the fitness function parameters of the improved bee swarm algorithm.

2. The method for intelligent diagnosis of base station power supply failure based on abnormality detection according to claim 1, characterized in that: The improved ResNet-18 network specifically includes: The multi-branch one-dimensional convolution input layer receives a multi-dimensional time series data set and inputs each channel data in the multi-dimensional time series data set into an independent one-dimensional convolution branch for preliminary feature extraction. Each one-dimensional convolution branch outputs a channel feature vector and concatenates and fuses them to form a power supply timing feature representation. The backbone residual network module includes multiple residual blocks, each of which contains a set of one-dimensional convolutional layers, batch normalization layers, and nonlinear activation function layers. In each residual block, the power timing feature representation is used as input and passed to the main path and the identity mapping path respectively. The main path performs a convolution transformation on the power timing feature representation to obtain the main path output feature, while the identity mapping path keeps the power timing feature representation unchanged. Finally, at the end of the residual block, the main path output feature and the power timing feature representation are element-by-element added to output the residual enhanced intermediate feature. The embedded attention mechanism module is set between multiple residual blocks to model the inter-channel correlation of the residual enhanced intermediate features, perform attention modulation according to the attention weights, and output the attention enhanced feature representation; The global average pooling layer and the fully connected layer perform dimensionality reduction and feature vector mapping on the attention-enhanced feature representation to generate the power fault feature vector.

3. The method for intelligent diagnosis of base station power supply failure based on abnormality detection according to claim 1, characterized in that: The S2 specifically includes: S21, multi-branch one-dimensional convolution input layer, receives the multi-dimensional time series data set X = [x1, x2, ..., x C ], where x c ∈R T , x c is the time series data of the cth channel, with a length of T and C being the number of channels; S22, x c Input the corresponding one-dimensional convolution branches respectively, perform convolution, batch normalization and gated activation function processing in sequence, and calculate the channel gated feature vector: g c =BN(x c *w c +b c )·Sigmoid(BN(x c *w c +b c )); Among them, g c is the c-th channel gated feature vector, Sigmoid(·) is the Sigmoid activation function, BN(·) is the batch normalization operation, x c is the time series data of the cth channel, w c is the one-dimensional convolution branch convolution kernel corresponding to the c-th channel, b c is the convolution bias term of the one-dimensional convolution branch corresponding to the c-th channel; S23, all channel gate feature vectors g1, g2, ..., g C Perform feature dimension splicing to form a fused feature representation: F0=Concat(g1,g2,…,g C ); Among them, F0 is the power timing feature representation, Concat(·) is the splicing operation; S24, input the power timing feature representation F0 into the trunk residual network module, and extract features through multiple residual blocks in sequence; in the i-th residual block, construct the main path and the identity mapping path respectively, and the main path inputs the feature F to the i-1th residual block i-1 Perform one-dimensional convolution, batch normalization and activation operations to obtain the main path output feature R i : R i =ReLU(BN(F i-1 *W i +b i )); Among them, ReLU(·) is the activation function, BN(·) is the batch normalization operation, W i is the one-dimensional convolution kernel of the i-th residual block, b i is the convolution bias term of the i-th residual block; the identity mapping path directly retains F i-1 : I i =F i-1 ; Among them, I i Output features for the identity mapping path; Output features for the main path R i Output feature I with the identity mapping path i Perform element-by-element addition to calculate the residual enhanced intermediate feature F i : F i =R i +I i ; S25, the embedded attention mechanism module receives the residual enhanced intermediate feature F i , calculate the channel attention vector and time attention vector respectively, and generate the channel-time joint attention matrix A through the outer product i , according to the note The attention weight performs attention modulation and outputs the attention-enhanced feature representation F′ i : F′ i =σ(A i )⊙F i ; Where σ(·) is the activation function; S26. Represent the attention-enhanced feature F′ i The global average pooling layer and the fully connected layer are input to perform dimensionality reduction and feature vector mapping, and the power fault feature vector is output.

4. The method for intelligent diagnosis of base station power supply failure based on abnormality detection according to claim 1, characterized in that: The improved bee colony algorithm performs global search through dynamic search radius and adaptive guide factor, and specifically includes the following steps based on feature clustering and cluster boundary detection: The dynamic search radius is adjusted based on the feature space distance between the search swarm individual and the current optimal swarm individual to control the local search range of the swarm individual during the iteration process; The adaptive guidance factor is dynamically calculated based on the relative difference between the individual fitness of the swarm and the global optimal swarm individual fitness, and adjusts the direction and intensity of the swarm individual's deviation toward the global optimal position and the local cluster center; The abnormality identification mechanism is based on the feature density clustering strategy. During the search process, the power failure feature vectors whose distance from each swarm individual position is greater than the search radius are judged as abnormal states, and the corresponding power failure mode is output in combination with the spatial position distribution.

5. The method for intelligent diagnosis of base station power supply failure based on anomaly detection according to claim 1, characterized in that: The S3 specifically includes: S31. Initialize the individual positions of the swarm, map the power failure feature vector to the initial search point in the search space dimension, set the search space dimension to d, the population size to N, and record the individual position of each swarm as Where j = 1, 2, 3, ..., N; S32, introduce a dynamic search radius update mechanism, and adaptively adjust the search radius of each round according to the Euclidean distance between the individual fitness value of the swarm and the optimal solution of the group. Among them, ρ0 is the initial radius, λ is the control factor, is the current optimal individual position of the swarm, exp(·) is the exponential function; S33, fusion adaptive guidance factor According to the relative level of individual fitness values ​​of the swarm, guide the swarm individuals to move between the optimal swarm individual position and the feature cluster center: Among them, α and β are the guiding direction control factors, is the feature cluster center, is the individual position of the swarm after moving, and the adaptive guidance factor The calculation method is: in, is the optimal individual fitness value of the bee colony, is the fitness value of the j-th swarm individual in the t-th iteration, ∈ is a constant to prevent division by zero; S34, based on the feature clustering method of bee colony density, performs cluster analysis on the search individual position set and performs anomaly detection at the distance boundary. For each input power fault feature vector z k , if the following conditions are met: The power failure feature vector is determined to be an abnormal state, and the power abnormal state mark and power failure mode are output in combination with the corresponding cluster position.

6. The method for intelligent diagnosis of base station power supply failure based on abnormality detection according to claim 1, characterized in that: The S4 specifically includes: S41, receiving the power supply abnormal state mark and the power supply failure mode, constructing the abnormal mark-failure mode feature pair (m j ,p j ), where m j ∈{0,1} represents the jth power supply abnormality status mark, p j represents the jth power supply failure mode; S42, for each set of abnormality mark-fault mode feature pair (m j ,p j ) performs power fault diagnosis label generation operation, using the classification function f class Output (m j ,p j ) corresponding to the power supply fault diagnosis label y j : y j =f class (m j ,p j ); S43, aggregate all power supply fault diagnosis labels to generate a power supply fault diagnosis label set Y = {y1, y2, ..., y N }, where N is the total number of power fault diagnosis tags.

7. The method for intelligent diagnosis of base station power supply failure based on anomaly detection according to claim 1, characterized in that: The S5 specifically includes: S51, receiving a power supply fault diagnosis tag set Y = {y1, y2, ..., y N }; S52, construct power fault diagnosis label mapping rule function f map :Y→R, change the power fault diagnosis label to y j Mapped to power failure diagnosis result r j : r j =f map (y j )。 8. The method for intelligent diagnosis of base station power supply failure based on anomaly detection according to claim 1, characterized in that: The S6 specifically includes: S61, based on power supply fault diagnosis label y j and historical fault diagnosis label y′ k , build a matching calculation function, and calculate the power supply-historical fault label pair (y j ,y′ k ) performs the correlation matching calculation to generate the correlation matching matrix S: S j,k =Sim(y j ,y′ k ); Among them, Sim(·) is the matching degree calculation function, S j,k is the power supply-history fault label pair (y j ,y′ k )’s matching value; S62, extract S according to the set threshold δ j,k The power supply-historical fault label pairs with a value no less than δ are used to construct a diagnostic label matching set: AND match ={(and j ,and' k )∣S j,k ≥δ}; S63. Statistical diagnostic label matching set Y match The number of power supply-historical fault label pairs n is set. When n meets the set update threshold, the update operation is performed to adjust the convolutional layer weights of the improved ResNet-18 network: in, is the convolutional layer weight of the improved ResNet-18 network in the tth round, is the convolutional layer weight of the improved ResNet-18 network in the t+1th round, η θ The learning rate for the improved ResNet-18 network update, L match is the loss function; S64. Adjust the fitness function parameters of the improved bee colony algorithm: in, is the fitness function parameter of the improved bee swarm algorithm in the tth round, is the fitness function parameter of the improved bee colony algorithm in the t+1th round, η φ The learning rate updated by the improved bee swarm algorithm, L match is the loss function.

9. A base station power supply fault intelligent diagnosis system based on anomaly detection, executing the base station power supply fault intelligent diagnosis method based on anomaly detection according to any one of claims 1 to 8, characterized in that: Includes the following modules: The data acquisition and preprocessing module is used to collect multimodal monitoring data of the base station power supply, standardize the multimodal monitoring data and align the timestamps to generate a multidimensional time series data set; The abnormal feature extraction and detection module is used to input the multidimensional time series data set into the improved ResNet-18 network, extract the power failure feature vector, and input the power failure feature vector into the improved bee swarm algorithm to perform anomaly detection based on dynamic search radius and adaptive guidance factor, and output the power abnormal status mark and power failure mode; A fault diagnosis and result generation module is used to generate a power supply fault diagnosis tag based on the power supply abnormal state mark and the power supply fault mode; The power supply fault diagnosis label is input into the label mapping rule function, and the power supply fault diagnosis result is output; The incremental update module compares the power supply fault diagnosis label with the historical fault diagnosis label and determines whether to perform the update operation based on the relevant matching value. When the update conditions are met, the module dynamically adjusts the convolutional layer weights of the improved ResNet-18 network and the fitness function parameters of the improved bee swarm algorithm. The diagnostic result output module is used to output the power supply fault diagnosis results to the operation and maintenance terminal in the form of text description, status code or graphical visualization.

Images