Method and device for active antenna element-oriented implicit fault prediction and intelligent positioning

Abstract

The application provides a method and device for predicting and intelligently positioning hidden faults of an active antenna unit, and relates to the technical field of communication. The method comprises the following steps: obtaining performance index data, measurement report data and hardware log data of the active antenna unit; performing space-time alignment and feature construction on the performance index data, the measurement report data and the hardware log data to generate a multi-dimensional feature vector; inputting the multi-dimensional feature vector into a supervised learning model to obtain a health degree score of the active antenna unit; when the health degree score is lower than a preset threshold, determining a fault feature according to the contribution of each feature in the multi-dimensional feature vector to the health degree score, and matching the fault feature with fault types in a knowledge base to generate a structured work order containing suspected fault points and troubleshooting suggestions. The application can significantly improve the discovery efficiency, positioning accuracy and operation and maintenance automation level of hidden faults of the active antenna unit, and realize a leap from passive response to active prediction.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a method and apparatus for predicting and intelligently locating latent faults in active antenna elements. Background Technology

[0002] Currently, for the detection and localization of latent faults in single-line AAUs of active antennas, existing technologies can only passively detect network-wide performance problems through macroscopic KPI threshold monitoring, and cannot detect microscopic degradation of radio frequency hardware; while cross-system coverage comparison relies on shared coverage conditions and cannot independently locate specific hardware faults; however, general machine learning methods are difficult to learn the precursor patterns of latent faults.

[0003] Therefore, there is an urgent need for an intelligent operation and maintenance method that can deeply integrate multi-source heterogeneous data from AAU, proactively predict hidden faults, and accurately locate the root cause. Summary of the Invention

[0004] This invention provides a method and apparatus for predicting and intelligently locating latent faults in active antenna units (AAUs). This method can significantly improve the efficiency of latent fault detection, the accuracy of fault location, and the level of automation in operation and maintenance.

[0005] Firstly, a method for predicting and intelligently locating latent faults in active antenna elements is provided, including: Acquire performance data, measurement report data, and hardware log data of the active antenna unit; Spatiotemporal alignment and feature construction are performed on performance index data, measurement report data, and hardware log data to generate multidimensional feature vectors for active antenna elements. These multidimensional feature vectors include features that characterize the health status of the active antenna elements. The multidimensional feature vector is input into a pre-trained supervised learning model to obtain the health score of the active antenna element. The supervised learning model is pre-trained using historical fault data as labeled samples, historical multidimensional feature vectors as input, and the probability of the active antenna element experiencing a latent fault within a preset time window as output. When the health score is lower than a preset threshold, at least one fault feature is determined based on the contribution of each feature in the multidimensional feature vector to the health score, and the fault feature is matched with the fault type in the knowledge base to generate a structured work order.

[0006] In this way, by fusing performance indicators, measurement reports and hardware log data across domains, a multi-dimensional feature vector oriented towards hardware health status is constructed, and a health score is output using a supervised learning model. When the score is lower than the threshold, a structured work order containing fault type and troubleshooting suggestions is automatically generated by combining feature importance analysis and knowledge base matching. This enables proactive prediction and accurate location of latent faults in active antenna units, overcoming the shortcomings of existing technologies such as passive response, inability to locate root causes and lack of closed-loop handling, and significantly improving operation and maintenance efficiency and the timeliness of fault detection.

[0007] In some possible implementations, performance metrics data, measurement report data, and hardware log data of the active antenna element are obtained, including: Retrieve performance metrics data from the performance database, including at least one of the following: average uplink interference noise per physical resource block, modulation and coding scheme level, and rank indication; retrieve measurement report data from the measurement report database, including reference signal received power and signal-to-interference-plus-noise ratio reported by the user terminal; retrieve hardware log data from the hardware log database, including channel calibration results and non-fatal alarm events.

[0008] In this way, by acquiring data from three dimensions—performance, measurement reports, and hardware logs—a comprehensive and complementary information foundation is provided for subsequent multi-source data fusion and the construction of latent fault characteristics, avoiding the blind spots of perception from a single data source.

[0009] In some possible implementations, spatiotemporal alignment and feature construction are performed on performance index data, measurement report data, and hardware log data to generate multidimensional feature vectors for active antenna elements, including: Using timestamps and cell identifiers as primary keys, performance index data, measurement report data, and hardware log data are spatiotemporally aligned. Based on the aligned data, the dynamic deviation between the average uplink interference noise per physical resource block and the historical baseline, the duration of interference spikes on a specific physical resource block, the number of channel correction failures per unit time, and the correlation coefficient between the modulation and coding scheme excellence rate and the interference level are calculated. Based on the features obtained from the above calculations, a multidimensional feature vector is generated.

[0010] In this way, by aligning the three types of heterogeneous data in time and space, we can unify them into the same coordinate system and construct special features that can quantify latent fault precursors such as interference anomalies, channel stability and modulation efficiency degradation. This enables the multi-dimensional feature vectors to comprehensively characterize the RF hardware health status of the active antenna unit, providing high-quality input for high-precision prediction of subsequent supervised learning models.

[0011] In some possible implementations, at least one fault feature is determined based on the contribution of each feature in the multidimensional feature vector to the health score, including: The feature importance analysis module of the supervised learning model is invoked to obtain the contribution value of each feature in the multidimensional feature vector; the features are sorted from largest to smallest according to their contribution values, and the features with the highest contribution values ​​are identified as fault features.

[0012] In this way, by utilizing the feature importance analysis capabilities built into the supervised learning model, the black-box output of the model is transformed into interpretable root cause localization information. Without additional manual analysis, key abnormal dimensions that lead to a decrease in health score can be automatically identified, providing clear fault characteristic basis for subsequent knowledge base matching and accurate order dispatch, and significantly improving the efficiency and accuracy of fault localization.

[0013] In some possible implementations, fault characteristics are matched with fault types in a knowledge base to generate structured work orders, including: Based on the fault characteristics, a pre-set knowledge base is queried. The knowledge base stores the mapping relationship between characteristics and fault types, as well as standard troubleshooting suggestions for each fault type. The fault characteristics are matched with the feature entries in the knowledge base to determine the successfully matched fault types, and the troubleshooting suggestions corresponding to the fault types are extracted. Based on the cell identifier, health score, fault characteristics, fault types, and troubleshooting suggestions of the active antenna unit, a structured work order is generated.

[0014] In this way, by automatically matching the fault features identified by the model with the expert knowledge base, the abstract abnormal features are transformed into specific fault types and standardized troubleshooting suggestions that can be directly understood by operation and maintenance personnel. This achieves a seamless connection from problem discovery to guidance on handling, significantly shortens the fault location and repair time, lowers the threshold for manual analysis, and improves the level of operation and maintenance automation.

[0015] In some possible implementations, a supervised learning model is trained through the following steps: A historical fault database is acquired, which stores manually confirmed cases of latent faults in active antenna elements. Each case includes performance index data, measurement report data, and hardware log data within a preset time window before the fault occurred, as well as a corresponding fault label. The data in the historical fault database is spatiotemporally aligned and features are constructed to generate historical multidimensional feature vectors, and the fault labels are used as labeled samples. An initial classification model is constructed, which can be an XGBoost, LightGBM, or random forest algorithm model. The model parameters are optimized through cross-validation, using the historical multidimensional feature vectors as input and the probability of a latent fault occurring in the active antenna element within a preset time window as output, so that the initial classification model learns the nonlinear mapping relationship between the historical multidimensional feature vectors and latent faults. The trained initial classification model is then used as a pre-trained supervised learning model.

[0016] In this way, by using historical fault cases confirmed by humans as "golden labels", the supervised learning model is trained to learn the complex nonlinear mapping relationship between multidimensional features and latent faults. This overcomes the shortcomings of traditional unsupervised methods, such as lack of specificity and high false alarm rate. The model can actively capture the precursor patterns of hardware latent faults such as power amplifier aging and channel abnormalities, and achieve accurate prediction of the probability of fault occurrence within a preset time window in the future, providing a reliable decision basis for proactive operation and maintenance.

[0017] Among some possible implementations, the method also includes: The structured fault order is pushed to the operation and maintenance platform, and the platform receives processing results from operation and maintenance personnel. The processing results include the confirmed root cause of the fault and the remedial measures. The processing results are then added as new labeled samples to the historical fault data to trigger fine-tuning of the supervised learning model and update the model. In this way, by feeding the processing results back to the training data, a closed-loop self-evolution mechanism is formed to continuously improve the model's prediction accuracy and adaptability.

[0018] In this way, by feeding the processing results back to the training data, a closed-loop self-evolution mechanism is formed, continuously improving the model's predictive accuracy and adaptability. Specifically, this mechanism enables the model to continuously learn from newly identified fault cases, automatically adapting to changes in the network environment and new hardware defects without requiring manual redesign of features or adjustment of the model structure. At the same time, incremental fine-tuning avoids the high computational cost of full retraining, ensuring the stability of the online prediction service, realizing the automated accumulation and reuse of operational experience, and significantly enhancing the system's long-term self-evolution capability.

[0019] Secondly, a latent fault prediction and intelligent location device for active antenna elements is provided, comprising: The first processing module is used to acquire performance index data, measurement report data, and hardware log data of the active antenna unit; The second processing module is used to perform spatiotemporal alignment and feature construction on performance index data, measurement report data and hardware log data to generate a multidimensional feature vector of the active antenna element. The multidimensional feature vector includes features that characterize the health status of the active antenna element. The third processing module is used to input the multidimensional feature vector into a pre-trained supervised learning model to obtain the health score of the active antenna unit. The supervised learning model is pre-trained using historical fault data as labeled samples, historical multidimensional feature vectors as input, and the probability of the active antenna unit experiencing a latent fault within a preset time window as output. The fourth processing module is used to determine at least one fault feature based on the contribution of each feature in the multidimensional feature vector to the health score when the health score is lower than a preset threshold, and to match the fault feature with the fault type in the knowledge base to generate a structured work order.

[0020] In some possible implementations, the first processing module is specifically used to: obtain performance index data from a performance database, the performance index data including at least one of the following: average uplink interference noise per physical resource block, modulation and coding scheme level, and rank indication; obtain measurement report data from a measurement report database, the measurement report data including the reference signal received power and signal-to-interference-plus-noise ratio reported by the user terminal; and obtain hardware log data from a hardware log database, the hardware log data including channel calibration results and non-fatal alarm events.

[0021] In some possible implementations, the second processing module is specifically used to: perform spatiotemporal alignment of performance index data, measurement report data, and hardware log data using timestamps and cell identifiers as association primary keys; calculate, based on the aligned data, the dynamic deviation between the average uplink interference noise per physical resource block and the historical baseline, the duration of interference spikes on a specific physical resource block, the number of channel correction failures per unit time, and the correlation coefficient between the modulation and coding scheme excellence rate and the interference level; and generate a multidimensional feature vector based on the features calculated above.

[0022] In some possible implementations, the fourth processing module is specifically used for: calling the feature importance analysis module of the supervised learning model to obtain the contribution value of each feature in the multi-dimensional feature vector; sorting the features according to their contribution values ​​from largest to smallest, and identifying the features with the highest contribution values ​​as fault features; querying a preset knowledge base based on the fault features, which stores the mapping relationship between features and fault types and the standard troubleshooting suggestions corresponding to each fault type; matching the fault features with the feature entries in the knowledge base to determine the successfully matched fault types and extracting the troubleshooting suggestions corresponding to the fault types; and generating a structured work order based on the cell identifier, health score, fault features, fault types, and troubleshooting suggestions of the active antenna unit.

[0023] In some possible implementations, the device also includes: a model training module for acquiring a historical fault database, performing spatiotemporal alignment and feature construction on the data in the historical fault database, generating historical multidimensional feature vectors, using fault labels as labeled samples, and constructing and training a supervised learning model; and a feedback update module for pushing structured work orders to the operation and maintenance platform, receiving processing results, and using the processing results as newly labeled samples to trigger fine-tuning of the supervised learning model.

[0024] Thirdly, an electronic device is provided, comprising: one or more processors; one or more memories; and one or more programs, wherein the one or more programs are stored in the one or more memories, and the one or more programs include instructions that, when executed by the one or more processors, cause the configured device to perform the method as described in the first aspect.

[0025] Fourthly, a computer storage medium is provided, the computer storage medium storing instructions that, when executed by a computer, cause the computer to perform the method described in the first aspect or the second aspect.

[0026] Fifthly, a computer program product is provided, the computer program product storing instructions that, when executed by a computer, cause the computer to perform the method described in the first aspect or the second aspect.

[0027] The beneficial effects of the second to fifth aspects can be referred to the introduction of the beneficial effects of the first aspect above, and will not be repeated here. Attached Figure Description

[0028] Figure 1 This is an overall architecture block diagram of an active prediction and intelligent positioning system for latent faults in AAU provided by an embodiment of the present invention; Figure 2 A flowchart of data fusion and feature engineering provided for embodiments of the present invention; Figure 3 This is a flowchart of the intelligent early warning and work order generation process provided in an embodiment of the present invention; Figure 4 This is a diagram illustrating the model training and online prediction separation architecture provided in an embodiment of the present invention. Figure 5 A flowchart illustrating a method for predicting and intelligently locating latent faults in active antenna elements, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of a latent fault prediction and intelligent positioning device for active antenna elements that can be used to implement the method of the present invention. Figure 7 This is a schematic diagram of an electronic device provided by the present invention. Detailed Implementation

[0029] The solutions provided by the embodiments of the present invention will now be described with reference to the accompanying drawings. In the embodiments of the present invention, "multiple" refers to two or more objects, and "various kinds" refers to two or more types. Terms such as "first," "second," etc., are only used to distinguish similar objects and are not necessarily used to describe a specific order or number of objects.

[0030] Currently, the active antenna unit (AAU) in 5G networks is a core radio frequency component of base stations, and its health status directly affects network quality and user experience. However, there is a lack of effective means to predict latent faults in AAUs caused by power amplifier aging, channel performance degradation, etc., and existing technologies have the following significant limitations.

[0031] Threshold monitoring solutions based on macro KPIs use general indicators such as drop rate, throughput, and RRC establishment success rate to set fixed thresholds for alarms. However, these indicators reflect the overall network performance and are not sensitive to the micro-degradation of the AAU's internal radio frequency hardware. Furthermore, the threshold rules are rigid and cannot adapt to dynamically changing network environments. Problems are often only discovered after user complaints, which is a passive response.

[0032] Positioning solutions based on inter-system comparison or external sensors assess device status by comparing 5G signals with 4G coverage or relying on physical sensors such as voltage and temperature. However, inter-system comparison depends on shared coverage conditions and cannot work independently in non-shared coverage scenarios; physical sensor solutions require hardware modifications to existing AAUs in the network, resulting in high deployment costs, poor compatibility, and difficulty in directly correlating sensor data with air interface performance indicators.

[0033] Prediction methods based on general machine learning or data augmentation utilize models such as SARIMA and RNNs to predict communication load and log sequences, or employ generative adversarial networks to address sample imbalance. These methods focus on macro-level resource scheduling or core network element service failures, without deeply exploring the characteristics of the AAU RF hardware itself, such as uplink interference, channel correction logs, and MCS distribution. They lack a dedicated feature system for hidden hardware failures such as power amplifier aging and channel anomalies, resulting in low prediction accuracy and an inability to locate specific faulty components.

[0034] In summary, existing technologies suffer from limitations such as reliance on macro-level indicators, need for hardware modifications, lack of specific features, or absence of closed-loop solutions. Specifically, macro-level KPI monitoring cannot detect subtle degradation in RF hardware; solutions relying on external references or physical sensors have poor applicability and high costs; and general machine learning algorithms lack a feature set specific to AAU hardware, making proactive prediction difficult. These technologies fail to construct a closed-loop system for proactively predicting AAU latent faults based on cross-domain data fusion and supervised learning, resulting in frequent user complaints and high maintenance costs.

[0035] To address the technical problems of imprecise AAU latent fault detection, passive prediction, and inaccurate location in existing technologies, this invention proposes a proactive prediction and intelligent location method and system for AAU latent faults in 5G wireless access networks. By fusing multi-dimensional data such as uplink interference, measurement reports, and hardware logs, a feature profile reflecting the hardware health of the AAU, including power amplifier aging and channel anomalies, is constructed. Historical latent fault cases, verified manually, are used as key indicators to train a supervised learning model that outputs a health score, achieving a leap from post-event alarms to pre-event prediction. Furthermore, by combining model feature importance analysis and knowledge base matching, when the health score falls below a threshold, a structured work order containing suspected fault points and specific troubleshooting suggestions is automatically generated, forming an automated closed loop of prediction, location, and handling. This method avoids the shortcomings of traditional solutions, such as reliance on macroscopic indicators, passive response, and inability to pinpoint root causes, significantly improving the efficiency of AAU latent fault detection, location accuracy, and the level of operational automation.

[0036] Figure 1 This is a block diagram illustrating the overall architecture of an active prediction and intelligent positioning system for latent faults in Automatic Active Automated Units (AAUs) according to an embodiment of the present invention. Figure 1 As shown, the system consists of five main parts: data source layer, data acquisition layer, feature engineering and model training layer, intelligent diagnosis and early warning layer, and operation and maintenance platform.

[0037] The data source layer is used to collect raw input data, including performance database, measurement report (MR) database, hardware log library, and historical fault library.

[0038] The performance database provides key performance indicators at the cell level, such as the average uplink interference noise per PRB, RSRP / SINR distribution, user rate, modulation and coding scheme level, and rank indication.

[0039] The MR database contains a large amount of air interface quality information reported by user terminals, such as measured reference signal received power, signal-to-interference-plus-noise ratio, and neighbor cell relationships, which are used to accurately reflect the wireless environment.

[0040] The hardware log library collects the operating status logs generated inside the active antenna unit (AAU) device, including channel calibration results, internal temperature, power amplifier gain, and non-fatal event alarms.

[0041] The historical fault database stores manually confirmed AAU fault cases and multi-dimensional data snapshots from a period of time before their occurrence, which serve as labels for model training.

[0042] The data acquisition layer extracts raw data from the aforementioned heterogeneous data sources and performs preliminary cleaning and alignment. Using timestamps and cell identifiers as the primary keys, it unifies performance data, MR data, and hardware logs into the same spatiotemporal coordinate system.

[0043] The feature engineering and model training layer transforms the cleaned raw data into high-value feature vectors and trains a machine learning model with predictive capabilities.

[0044] Specifically, this layer constructs features for latent faults of AAU based on knowledge in the field of communications, such as the dynamic deviation between the mean uplink interference and the historical baseline, the duration of interference surges on a specific physical resource block, the number of channel correction failures per unit time, and the correlation coefficient between the goodness rate of modulation and coding schemes and the interference level, forming a comprehensive profile of the health status of AAU.

[0045] Subsequently, using labeled data from the historical fault database, a binary classification supervised learning model, such as XGBoost or Isolation Forest, is trained. The constructed feature vector is used as input to output the probability that the AAU will experience a latent fault within a preset time window in the future.

[0046] The intelligent diagnosis and early warning layer deploys the trained model to the production environment to monitor the existing AAUs in real time. This layer periodically calculates the health score of each AAU; when the health score is lower than a preset threshold, an early warning process is automatically triggered. Combining the feature importance analysis of the model, it initially judges the most likely faulty components, such as RF channel abnormalities or power amplifier nonlinear distortion. At the same time, it matches corresponding troubleshooting suggestions from the knowledge base and generates a structured early warning work order.

[0047] The operation and maintenance platform receives work orders from the intelligent diagnosis and early warning layer, presents the early warning information to maintenance personnel, and guides on-site fault handling.

[0048] The entire system forms an end-to-end automated closed loop through data acquisition, feature engineering, model training, online prediction, and work order processing, achieving a leap from passive operation and maintenance to proactive prediction.

[0049] The following describes an active prediction and intelligent localization method for latent faults in an Automatic Automated Unit (AAU) according to an embodiment of the present invention, specifically including the following steps: In step S1, multi-source heterogeneous data are collected and fused.

[0050] In this embodiment of the invention, raw data related to the health status of the AAU is extracted in real time from the performance database, MR database, hardware log database, and historical fault database through the data acquisition layer.

[0051] Specifically, the performance database provides key performance indicators at the cell level, including the average uplink interference noise per PRB (IQI), RSRP / SINR distribution, user rate, modulation and coding scheme MCS level, and rank indicator.

[0052] The MR database contains air interface measurement information reported by user terminals, such as RSRP, SINR, and neighbor cell relationships.

[0053] The hardware log library collects internal operating status logs of the AAU, including channel calibration results, internal temperature, power amplifier gain, and non-fatal alarm events such as abnormal operation alarms of RRU secondary devices.

[0054] The historical fault database stores manually confirmed AAU fault cases and multi-dimensional data snapshots from a period of time before their occurrence, which serve as labels for model training.

[0055] The data acquisition layer uses timestamps and cell identifiers as the primary keys to clean, align, and fuse the aforementioned multi-source data to form a raw dataset under a unified spatiotemporal coordinate system.

[0056] In step S2, a multidimensional feature vector for AAU latent faults is constructed.

[0057] In this embodiment of the invention, the feature engineering and model training layer, based on knowledge in the field of communication, transforms the fused raw data into feature vectors that can effectively characterize the health status of AAU.

[0058] Figure 2 This is a flowchart illustrating the data fusion and feature engineering process according to an embodiment of the present invention. Figure 2 As shown, firstly, the original data is calculated and statistically analyzed by the rule engine to generate primary features, including: the dynamic deviation between the mean uplink interference and the historical baseline, the duration of interference spikes on specific physical resource blocks, the number of channel correction failures per unit time, and the correlation coefficient between the MCS excellence rate and the interference level.

[0059] Subsequently, the primary features are standardized or normalized to form a multidimensional feature vector with uniform dimensions, which serves as the input for subsequent models.

[0060] Specifically, the dynamic deviation between the uplink interference mean and the historical baseline is calculated using a preset time window as the granularity, as the real-time mean of the uplink interference noise per physical resource block within the current window. The historical baseline is the median or average of the interference mean within the same time window in the past period. Dynamic deviation = current mean - historical baseline, used to characterize the abnormal fluctuation range of the interference level.

[0061] The duration of a sudden increase in interference on a specific physical resource block is determined by setting an interference noise threshold for each physical resource block. When the interference value of a physical resource block continuously exceeds the threshold, the duration is recorded, and the maximum duration among all physical resource blocks is taken as the feature.

[0062] The number of channel calibration failures per unit time is the number of channel calibration failure events recorded in the hardware log within the most recent preset time window. The number of failures for different channels can be counted separately.

[0063] The correlation coefficient between MCS excellence rate and interference level is calculated by taking the proportion of transmissions with MCS level greater than a preset threshold within the same time window to obtain the MCS excellence rate; then, the Pearson correlation coefficient between this excellence rate and the average interference noise per physical resource block in the uplink is calculated to assess whether the decrease in modulation efficiency is caused by interference degradation.

[0064] In addition, other characteristics characterizing the health status of AAU can be calculated from the above-mentioned raw data, which are not limited in this invention.

[0065] In this embodiment of the invention, after all the primary features have been calculated, Z-score normalization or Min-Max normalization is used to transform each feature to a uniform dimensional range, thereby eliminating the impact of differences in magnitude between different features on subsequent models. Finally, these features are combined to form a multi-dimensional feature vector, which serves as the input to the supervised learning model.

[0066] In step S3, a supervised learning model is trained and a health score is output.

[0067] In this embodiment of the invention, a supervised binary classification model is trained using multi-dimensional feature vectors and labeled data from a historical fault database. The model can employ algorithms such as XGBoost, LightGBM, or Random Forest. Using feature vectors as input and whether an AAU experiences a latent fault within a preset time window (e.g., 24 hours) as output, the model parameters are optimized through cross-validation. After training, the model can learn complex, non-linear fault precursor patterns. The trained model is deployed to a production environment, and a health score is periodically calculated for each AAU in the live network, for example, every 15 minutes. The score range is set from 0 to 100, with lower scores indicating higher fault risk.

[0068] The model training process begins by extracting manually confirmed latent fault cases of the AAU (Automatic Active User Unit) from the historical fault database as positive samples. Specifically, using the occurrence time of each fault case as a baseline, historical multidimensional feature vectors within a preset time window prior to the baseline are selected as positive samples and labeled with 1, indicating that a latent fault will occur within the next 24 hours. Simultaneously, an equal number of samples are randomly collected from the normal operation period of the same AAU, for example, a period beyond 30 days before the fault occurs, as negative samples and labeled with 0, indicating that a latent fault will not occur within the next 24 hours.

[0069] The positive and negative samples are divided into training set, validation set and test set in chronological order.

[0070] XGBoost, LightGBM, or Random Forest were chosen as the binary classification model. Historical multidimensional feature vectors were used as input, and the output objective was whether a latent fault occurred within a pre-defined future time window. Grid search or Bayesian optimization was used to fine-tune the model's hyperparameters, including the maximum tree depth, learning rate, subsampling ratio, feature sampling ratio, and regularization parameters. The optimization objective was to maximize the AUC value or F1 score on the validation set.

[0071] During training, five-fold cross-validation can be used to evaluate the model's stability. Simultaneously, feature importance ranking is calculated, features with extremely low contribution are removed, feature dimensions are reduced, and overfitting is prevented. Model evaluation metrics include: accuracy, recall, precision, F1 score, AUC value, and confusion matrix. A model is considered to meet the deployment requirements when the AUC value on the validation set reaches 0.85 or higher and the recall is greater than 0.8.

[0072] Furthermore, this embodiment of the invention employs an architecture that separates offline training from online prediction. The offline component retrains or incrementally updates the model every 24 hours or week using newly added historical fault data, manages model versioning, and stores the model in a model library. The online component, the prediction service in the production environment, periodically loads the latest stable version of the model and performs the following operations on each AAU in the live network: Collect performance metrics, measurement reports, and hardware log data for the current time window.

[0073] Perform the same spatiotemporal alignment and feature construction as in the training phase to generate multidimensional feature vectors.

[0074] Input the feature vector into the model to obtain the fault probability value p.

[0075] The failure probability p is mapped to a health score using the formula: Health Score = (1-p) × 100, with a score range of 0-100. A lower score indicates a higher risk of failure.

[0076] If the model outputs a probability p > 0.5, an early warning will be issued. However, depending on actual business needs, a more lenient threshold can be set to balance false negatives and false positives.

[0077] The prediction service employs multi-threading or asynchronous processing to ensure that the latency of a single prediction is within milliseconds. Prediction results are written to a cache or time-series database in real time for monitoring, display, and alert triggering.

[0078] Furthermore, after maintenance personnel complete fault handling and confirm the root cause based on the structured work orders generated by the system, the handling results, including fault labels and actual fault occurrence times, are added as new positive samples and fed back into the historical fault database. The system periodically triggers model retraining or incremental learning, using the new samples to update model parameters, enabling the model to continuously adapt to new fault modes and form a closed loop of continuous self-evolution. Incremental learning can utilize XGBoost's continued training function or LightGBM's incremental learning interface, eliminating the need for a full retraining.

[0079] Figure 4 This is a diagram illustrating the model training and online prediction separation architecture of an embodiment of the present invention. Figure 4 As shown, the architecture is divided into an offline training part and an online prediction part.

[0080] Offline training periodically extracts samples from the historical fault database and the full dataset. The extracted samples undergo the same feature engineering process as online prediction, generating historical multi-dimensional feature vectors and corresponding fault labels. This data is used to train or update the supervised learning model, and the model parameters are optimized through cross-validation. After training, different versions of the model are versioned and stored in a model repository for loading and use in the online portion.

[0081] The online prediction component pulls real-time performance metrics, measurement reports, and hardware logs from the production environment. It performs the same spatiotemporal alignment and feature construction on the real-time data as the offline component, generating multi-dimensional feature vectors. The online prediction service loads the latest stable version of the model from the model library, inputs the feature vectors into the model, and outputs a health score for the active antenna elements, enabling periodic health predictions. The online component does not participate in model training; it only handles inference, thus ensuring the stability and low latency of the online service.

[0082] This design, which separates offline training from online prediction, allows the model to continuously iterate and optimize using newly added historical fault data, without affecting real-time prediction services in the production environment, thus balancing the model's evolution capability and the system's stability.

[0083] In step S4, intelligent early warning and root cause localization are performed.

[0084] In this embodiment of the invention, the intelligent diagnosis and early warning layer monitors the health score in real time. When the health score of a certain AAU is lower than a preset threshold, such as 60 points, an early warning process is automatically triggered.

[0085] Figure 3 This is a flowchart illustrating the intelligent early warning and work order generation process according to an embodiment of the present invention. Figure 3As shown, firstly, the feature importance analysis module of the model is invoked to identify the features that contribute most to the current low score, thereby initially determining the most likely faulty component, such as RF channel abnormality, power amplifier nonlinear distortion, or local oscillator drift. Then, standard troubleshooting suggestions from the knowledge base are matched according to the fault type, such as prioritizing checking channel calibration logs, arranging on-site AAU replacement testing, or checking the power amplifier power supply voltage. Finally, the system integrates cell identifiers, health scores, suspected fault points, and specific troubleshooting suggestions to generate structured early warning information.

[0086] In step S5, a closed-loop processing work order is generated and feedback is provided for optimization.

[0087] In this embodiment of the invention, the system automatically pushes early warning information to the operation and maintenance platform, generating a work order to be processed. Operation and maintenance personnel then conduct on-site testing or equipment replacement based on the troubleshooting suggestions in the work order.

[0088] After the fault handling is completed, the results, including the confirmed root cause of the fault and the repair measures, are sent back to the historical fault database for incremental training or parameter adjustment of the model in the next round, forming a complete closed loop from data collection, feature construction, model prediction, early warning and handling to feedback optimization.

[0089] Figure 3 This is a flowchart illustrating the intelligent early warning and work order generation process according to an embodiment of the present invention. Figure 3 As shown, the process specifically includes the following steps: First, a pre-trained supervised learning model is periodically invoked, with the multidimensional feature vector of the current active antenna unit (AAU) as input, and the model outputs the health score of the AAU.

[0090] Next, the health score is evaluated to determine if it falls below a preset threshold. If the health score is higher than or equal to the preset threshold, the AAU is considered to be operating normally, the process ends, and no warning needs to be generated.

[0091] If the health score is lower than the preset threshold, an early warning process is triggered: the system calls the root cause analysis module, which identifies several features that contribute most to the current low score based on the feature importance analysis results of the model, thereby locating the most likely faulty component, such as RF channel abnormality or power amplifier nonlinear distortion; at the same time, the identified fault features are matched with the fault types in the knowledge base, and corresponding standard troubleshooting suggestions are extracted, such as prioritizing the checking of channel calibration logs and arranging on-site AAU replacement testing.

[0092] Finally, the AAU's cell identifier, health score, suspected fault points, and specific troubleshooting suggestions are integrated to generate a structured early warning work order, which is automatically pushed to the operation and maintenance work order system to notify maintenance personnel to handle the issue.

[0093] This process achieves a complete closed loop from health assessment, anomaly detection, root cause localization, knowledge base matching to work order generation, ensuring that actionable measures can be automatically triggered when hidden fault risks are predicted.

[0094] Furthermore, it should be noted that in other embodiments of the present invention, the technical solution can be implemented in the following alternative forms, which can also achieve the technical effects of predicting AAU latent faults, locating root causes, and triggering treatment.

[0095] First, the prediction model is not limited to XGBoost and can be replaced by other machine learning or deep learning algorithms, including but not limited to random forests, support vector machines, long short-term memory networks, or deep neural networks. As long as the algorithm uses the same multidimensional feature vector, i.e., features constructed by fusing performance metrics, measurement reports, and hardware logs, for training and outputs a health score or failure probability representing the health status of the active antenna unit, it should be considered an equivalent substitution of the present invention.

[0096] Secondly, although the present invention preferably uses the average interference noise per physical resource block, measurement report data, and channel correction log as the core data source, other relevant indicators can be introduced as supplements or replacements in alternative solutions, such as the energy consumption data of active antenna units, ambient temperature and humidity data, or user plane latency indicators on the core network side.

[0097] Furthermore, the feature construction method can also be adjusted accordingly. For example, the dynamic deviation calculation can be replaced with statistical features based on a sliding window, such as mean, variance, percentiles, or other mathematical transformations. As long as the purpose is to characterize the health status of the radio frequency hardware, they are all equivalent solutions of this invention.

[0098] The threshold determination logic for health scores is not limited to fixed thresholds. Dynamic threshold strategies can be adopted, such as adaptively adjusting the warning trigger threshold based on the cell's service load, time period, or geographical location.

[0099] The closed-loop handling method is not limited to generating a structured work order and pushing it to the operation and maintenance platform. It can be replaced by automatically triggering hardware reset commands, automatically adjusting beamforming parameters to avoid faulty channels, or notifying maintenance personnel via instant messaging methods such as SMS and email. These alternative methods also achieve an automated closed loop from prediction to handling, and are equivalent transformations of this invention.

[0100] Furthermore, the technical solution of this invention has extremely high engineering feasibility and application value. First, the data acquisition threshold is low, and no hardware modification is required. The required data sources include key performance indicators, measurement reports and hardware logs, all of which are existing network data provided by the standard northbound interface of the operator's network management system. No physical sensors need to be installed on the active antenna unit device, nor is any modification to the base station hardware required. It is entirely based on software implementation, and the deployment cost is extremely low.

[0101] Secondly, the technology stack is mature and highly compatible. The big data processing framework, machine learning library, and microservice architecture involved are all mature industry standard technologies. The entire system can be used as an independent microservice module and seamlessly integrated into existing 5G network management platforms or intelligent operation and maintenance platforms through standard application programming interfaces, demonstrating excellent compatibility and scalability.

[0102] Furthermore, this solution addresses existing network pain points and offers clear commercial value. Latent faults in active antenna elements are a major cause of decreased 5G user experience, and traditional maintenance methods struggle to proactively detect them. This solution directly improves network quality and user satisfaction, and operators have a clear need for intelligent maintenance tools. Currently, this solution is ready for pilot testing in some densely populated 5G network areas and can be applied in existing network maintenance centers to perform periodic health checks on all active antenna elements across the network.

[0103] More importantly, the embodiments of this invention have moderate requirements for computing resources. Model training can be performed on an offline server cluster, and the online prediction stage only requires matrix operations on the feature vectors. Real-time requirements are not stringent, and existing server resources can support it without the need to purchase additional dedicated acceleration hardware.

[0104] Figure 5 This is a flowchart illustrating a method for latent fault prediction and intelligent location of active antenna elements, as shown in an embodiment of the present invention. Figure 5 As shown, it includes the following steps: 510: Obtain performance data, measurement report data, and hardware log data of the active antenna unit.

[0105] 520: Spatiotemporal alignment and feature construction are performed on performance index data, measurement report data, and hardware log data to generate multidimensional feature vectors for active antenna elements. The multidimensional feature vectors include features that characterize the health status of the active antenna elements.

[0106] 530: Input the multidimensional feature vector into a pre-trained supervised learning model to obtain the health score of the active antenna element; wherein, the supervised learning model is a model pre-trained by using historical fault data as labeled samples, taking historical multidimensional feature vectors as input, and taking the probability of the active antenna element having a latent fault within a preset time window in the future as output.

[0107] 540: When the health score is lower than the preset threshold, at least one fault feature is determined based on the contribution of each feature in the multidimensional feature vector to the health score, and the fault feature is matched with the fault type in the knowledge base to generate a structured work order.

[0108] By integrating performance metrics, measurement reports, and hardware log data across domains, a multi-dimensional feature vector oriented towards hardware health status is constructed. A supervised learning model is then used to output a health score. When the score is below a threshold, feature importance analysis and knowledge base matching are combined to automatically generate a structured work order containing fault types and troubleshooting suggestions. This enables proactive prediction and precise location of latent faults in active antenna units, overcoming the shortcomings of existing technologies such as passive response, inability to locate root causes, and lack of closed-loop handling. This significantly improves operation and maintenance efficiency and the timeliness of fault detection.

[0109] In some embodiments, step 510 includes: Retrieve performance metrics data from the performance database, including at least one of the following: average uplink interference noise per physical resource block, modulation and coding scheme level, and rank indication; retrieve measurement report data from the measurement report database, including reference signal received power and signal-to-interference-plus-noise ratio reported by the user terminal; retrieve hardware log data from the hardware log database, including channel calibration results and non-fatal alarm events.

[0110] In some embodiments, step 520 includes: Using timestamps and cell identifiers as primary keys, performance index data, measurement report data, and hardware log data are spatiotemporally aligned. Based on the aligned data, the dynamic deviation between the average uplink interference noise per physical resource block and the historical baseline, the duration of interference spikes on a specific physical resource block, the number of channel correction failures per unit time, and the correlation coefficient between the modulation and coding scheme excellence rate and the interference level are calculated. Based on the features obtained from the above calculations, a multidimensional feature vector is generated.

[0111] In some embodiments, step 540, determining at least one fault feature based on the contribution of each feature in the multidimensional feature vector to the health score, includes: The feature importance analysis module of the supervised learning model is invoked to obtain the contribution value of each feature in the multidimensional feature vector; the features are sorted from largest to smallest according to their contribution values, and the features with the highest contribution values ​​are identified as fault features.

[0112] In some embodiments, step 540 involves matching fault characteristics with fault types in a knowledge base to generate a structured work order, including: Based on the fault characteristics, a pre-set knowledge base is queried. The knowledge base stores the mapping relationship between characteristics and fault types, as well as standard troubleshooting suggestions for each fault type. The fault characteristics are matched with the feature entries in the knowledge base to determine the successfully matched fault types, and the troubleshooting suggestions corresponding to the fault types are extracted. Based on the cell identifier, health score, fault characteristics, fault types, and troubleshooting suggestions of the active antenna unit, a structured work order is generated.

[0113] In some embodiments, the supervised learning model is trained through the following steps: A historical fault database is acquired, which stores manually confirmed cases of latent faults in active antenna elements. Each case includes performance index data, measurement report data, and hardware log data within a preset time window before the fault occurred, as well as a corresponding fault label. The data in the historical fault database is spatiotemporally aligned and features are constructed to generate historical multidimensional feature vectors, and the fault labels are used as labeled samples. An initial classification model is constructed, which can be an XGBoost, LightGBM, or random forest algorithm model. The model parameters are optimized through cross-validation, using the historical multidimensional feature vectors as input and the probability of a latent fault occurring in the active antenna element within a preset time window as output, so that the initial classification model learns the nonlinear mapping relationship between the historical multidimensional feature vectors and latent faults. The trained initial classification model is then used as a pre-trained supervised learning model.

[0114] In some embodiments, step 550 is also included: The structured fault order is pushed to the operation and maintenance platform, and the platform receives processing results from operation and maintenance personnel. The processing results include the confirmed root cause of the fault and the remedial measures. The processing results are then added as new labeled samples to the historical fault data to trigger fine-tuning of the supervised learning model and update the model. In this way, by feeding the processing results back to the training data, a closed-loop self-evolution mechanism is formed to continuously improve the model's prediction accuracy and adaptability.

[0115] Figure 6 This is a schematic diagram illustrating a latent fault prediction and intelligent positioning device for active antenna elements, as shown in an embodiment of the present invention. Figure 6 As shown, it includes: The first processing module is used to acquire performance index data, measurement report data, and hardware log data of the active antenna unit.

[0116] The second processing module is used to perform spatiotemporal alignment and feature construction on performance index data, measurement report data and hardware log data to generate multidimensional feature vectors of active antenna elements. The multidimensional feature vectors include features that characterize the health status of the active antenna elements.

[0117] The third processing module is used to input the multidimensional feature vector into a pre-trained supervised learning model to obtain the health score of the active antenna unit. The supervised learning model is pre-trained using historical fault data as labeled samples, historical multidimensional feature vectors as input, and the probability of the active antenna unit experiencing a latent fault within a preset time window as output.

[0118] The fourth processing module is used to determine at least one fault feature based on the contribution of each feature in the multidimensional feature vector to the health score when the health score is lower than a preset threshold, and to match the fault feature with the fault type in the knowledge base to generate a structured work order.

[0119] In some embodiments, the first processing module is specifically configured to: obtain performance index data from a performance database, the performance index data including at least one of uplink per physical resource block interference noise average, modulation and coding scheme level, and rank indication; obtain measurement report data from a measurement report database, the measurement report data including reference signal received power and signal-to-interference-plus-noise ratio reported by the user terminal; and obtain hardware log data from a hardware log database, the hardware log data including channel correction results and non-fatal alarm events.

[0120] In some embodiments, the second processing module is specifically used to: perform spatiotemporal alignment of performance index data, measurement report data, and hardware log data using timestamps and cell identifiers as association primary keys; calculate, based on the aligned data, the dynamic deviation between the average uplink interference noise per physical resource block and the historical baseline, the duration of interference spikes on a specific physical resource block, the number of channel correction failures per unit time, and the correlation coefficient between the modulation and coding scheme excellence rate and the interference level; and generate a multidimensional feature vector based on the features calculated above.

[0121] In some embodiments, the fourth processing module is specifically used to: call the feature importance analysis module of the supervised learning model to obtain the contribution value of each feature in the multidimensional feature vector; sort the features according to the contribution value from largest to smallest, and determine the features with the highest contribution value as fault features; query a preset knowledge base based on the fault features, the knowledge base storing the mapping relationship between features and fault types and the standard troubleshooting suggestions corresponding to each fault type; match the fault features with the feature entries in the knowledge base to determine the successfully matched fault types, and extract the troubleshooting suggestions corresponding to the fault types; generate a structured work order based on the cell identifier, health score, fault features, fault types, and troubleshooting suggestions of the active antenna unit.

[0122] In some embodiments, the apparatus further includes: a model training module, configured to acquire a historical fault database, perform spatiotemporal alignment and feature construction on the data in the historical fault database, generate historical multidimensional feature vectors, use fault labels as labeled samples, and construct and train a supervised learning model; and a feedback update module, configured to push structured work orders to the operation and maintenance platform, receive processing results, and use the processing results as newly labeled samples to trigger fine-tuning of the supervised learning model.

[0123] Those skilled in the art will readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is implemented in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the present invention.

[0124] It should be noted that, Figure 6 The division of modules / units is illustrative and represents only one logical functional division; in actual implementation, other division methods are possible. For example, two or more functions can be integrated into a single data acquisition module. The integrated modules described above can be implemented either in hardware or as software functional modules.

[0125] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the aforementioned methods for predicting and intelligently locating latent faults in active antenna elements. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the method for predicting and intelligently locating latent faults in active antenna elements shown in any embodiment of the present invention by calling the computer program.

[0126] In one alternative embodiment, an electronic device is provided, such as Figure 7 As shown, Figure 7 The illustrated electronic device 700 includes a processor 701 and a memory 703. The processor 701 and the memory 703 are connected, for example, via a bus 702. Optionally, the electronic device 700 may further include a transceiver 704, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 704 is not limited to one type, and the structure of the electronic device 700 does not constitute a limitation on the embodiments of the present invention.

[0127] Among them, electronic devices can also be terminal devices, which can be any device that can install applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0128] It should be noted that, Figure 7 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0129] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any one of the above-mentioned methods for latent fault prediction and intelligent location of active antenna elements.

[0130] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0131] It should be noted that the terms "first," "second," etc., used in the specification and claims of this invention are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of the invention described herein can be implemented in an order other than that shown or described.

[0132] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0133] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for latent fault prediction and intelligent location of active antenna elements, characterized in that, include: Acquire performance data, measurement report data, and hardware log data of the active antenna unit; The performance index data, measurement report data, and hardware log data are spatiotemporally aligned and feature constructed to generate a multidimensional feature vector of the active antenna element, wherein the multidimensional feature vector includes features characterizing the health status of the active antenna element. The multidimensional feature vector is input into a pre-trained supervised learning model to obtain the health score of the active antenna unit; wherein, the supervised learning model is a model trained by using historical fault data as labeled samples, taking historical multidimensional feature vectors as input, and taking the probability of the active antenna unit experiencing a latent fault within a preset time window in the future as output. When the health score is lower than a preset threshold, at least one fault feature is determined based on the contribution of each feature in the multidimensional feature vector to the health score, and the fault feature is matched with the fault type in the knowledge base to generate a structured work order.

2. The method according to claim 1, characterized in that, The acquisition of performance index data, measurement report data, and hardware log data of the active antenna element includes: The performance metrics data are obtained from the performance database, and the performance metrics data includes at least one of the following: average uplink interference noise per physical resource block, modulation and coding scheme level, and rank indicator. The measurement report data is obtained from the measurement report database, and the measurement report data includes the reference signal received power and signal-to-interference-plus-noise ratio reported by the user terminal; The hardware log data is obtained from the hardware log library, and the hardware log data includes channel calibration results and non-fatal alarm events.

3. The method according to claim 2, characterized in that, The process of performing spatiotemporal alignment and feature construction on the performance index data, measurement report data, and hardware log data to generate a multidimensional feature vector for the active antenna element includes: Using timestamps and cell identifiers as the primary keys for association, the performance index data, measurement report data, and hardware log data are spatiotemporally aligned; Based on the aligned data, the dynamic deviation between the average interference noise of each physical resource block in the uplink and the historical baseline, the duration of interference spikes on a specific physical resource block, the number of channel correction failures per unit time, and the correlation coefficient between the modulation and coding scheme excellence rate and the interference level are calculated. The multidimensional feature vector is obtained based on the dynamic deviation of the average interference noise per uplink physical resource block from the historical baseline, the duration of the interference surge on the specific physical resource block, the number of channel correction failures per unit time, and the correlation coefficient between the modulation and coding scheme excellence rate and the interference level.

4. The method according to claim 3, characterized in that, The step of determining at least one fault feature based on the contribution of each feature in the multidimensional feature vector to the health score includes: The feature importance analysis module of the supervised learning model is invoked to obtain the contribution value of each feature in the multidimensional feature vector; The features are sorted from largest to smallest according to their contribution values, and the features with the highest contribution values ​​in a predetermined number of positions are determined as the fault features.

5. The method according to claim 4, characterized in that, The step of matching the fault characteristics with fault types in the knowledge base to generate a structured work order includes: Based on the fault characteristics, a preset knowledge base is queried. The knowledge base stores the mapping relationship between characteristics and fault types, as well as standard troubleshooting suggestions corresponding to each fault type. The fault features are matched with feature entries in the knowledge base to determine the fault types that are successfully matched, and the troubleshooting suggestions corresponding to the fault types are extracted. A structured work order is generated based on the cell identifier of the active antenna unit, the health score, the fault characteristics, the fault type, and the troubleshooting suggestions.

6. The method according to any one of claims 1-5, characterized in that, The supervised learning model is trained through the following steps: Obtain a historical fault database, which stores manually confirmed cases of latent faults in active antenna elements. Each case includes performance index data, measurement report data, and hardware log data within a preset time window before the fault occurred, as well as the corresponding fault tag. Spatiotemporal alignment and feature construction are performed on the data in the historical fault database to generate historical multidimensional feature vectors, and the fault labels are used as labeled samples. Construct an initial classification model, which is an XGBoost, LightGBM, or Random Forest algorithm model; Using the historical multidimensional feature vector as input and the probability of the active antenna unit experiencing a latent fault within a future preset time window as output, the model parameters are optimized through cross-validation, enabling the initial classification model to learn the nonlinear mapping relationship between the historical multidimensional feature vector and the latent fault. The trained initial classification model is used as the pre-trained supervised learning model.

7. The method according to claim 1, characterized in that, Also includes: The structural work order is pushed to the operation and maintenance platform, and the processing results are received from the operation and maintenance personnel. The processing results include the confirmed root cause of the fault and the repair measures. The processing results are added as new labeled samples to the historical fault data to trigger fine-tuning of the supervised learning model and update the supervised learning model.

8. A device for predicting and intelligently locating latent faults in active antenna elements, characterized in that, include: The first processing module is used to acquire performance index data, measurement report data, and hardware log data of the active antenna unit; The second processing module is used to perform spatiotemporal alignment and feature construction on the performance index data, measurement report data and hardware log data to generate a multidimensional feature vector of the active antenna element, wherein the multidimensional feature vector includes features characterizing the health status of the active antenna element. The third processing module is used to input the multidimensional feature vector into a pre-trained supervised learning model to obtain the health score of the active antenna unit; wherein, the supervised learning model is a model trained by using historical fault data as labeled samples, taking historical multidimensional feature vectors as input, and taking the probability of the active antenna unit experiencing a latent fault within a preset time window in the future as output. The fourth processing module is used to determine at least one fault feature based on the contribution of each feature in the multidimensional feature vector to the health score when the health score is lower than a preset threshold, and to match the fault feature with the fault type in the knowledge base to generate a structured work order.

9. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the method for latent fault prediction and intelligent location of active antenna elements as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to enable the computer to implement the method for latent fault prediction and intelligent location of active antenna elements as described in any one of claims 1 to 7.

Images