Base station service withdrawal prediction method and device, electronic equipment and storage medium

Abstract

The application provides a base station withdrawal prediction method and device, electronic equipment and storage medium, relates to the network operation and maintenance technical field, and the method comprises the following steps: acquiring a plurality of alarm quantity accumulation sequences and environment characteristics of a target area in a historical time period, forming a plurality of feature vectors arranged along a time axis; inputting the plurality of feature vectors and the city one-hot encoding into a base station withdrawal quantity prediction model to obtain the base station withdrawal quantity of the target area in a future prediction time period output by the prediction model; the prediction model comprises a plurality of prediction units connected in series, and the prediction value output by a previous prediction unit is written back as part of the input features of a subsequent prediction unit, which is used to fill in unknown label information in the feature vector step by step. The application realizes accurate prediction of the batch base station withdrawal quantity of the county level through accumulation sequence feature construction, city one-hot encoding and iterative write-back type series prediction model; through the accumulation sequence input and the iterative write-back structure, the prediction accuracy and timeliness are improved.

Description

Technical Field

[0001] This application relates to the field of network operation and maintenance technology, and in particular to a method, apparatus, electronic device and storage medium for predicting base station outages. Background Technology

[0002] Base station outage prediction is an important technical direction in the field of network operation and maintenance. Its purpose is to identify base stations that may go out of service in advance so that operation and maintenance personnel can take preventive measures in a timely manner and reduce the impact of network failures on users' communication experience.

[0003] Currently, the focus is mainly on predicting the outage of individual base stations. For example, some solutions extract the alarm features of individual base stations and use heterogeneous base learners and stacked meta-learners for fusion prediction; other solutions are based on the historical alarm sequences of base stations and use recurrent neural network models such as bidirectional LSTM (Long Short-Term Memory) for training to predict the future outage risk of individual base stations.

[0004] However, the existing technical solutions mentioned above still have the following shortcomings in practical applications: the prediction target is only a single base station, which is difficult to cope with regional and batch base station outages; the prediction algorithms are mostly limited to traditional machine learning or recurrent neural networks, with limited prediction accuracy and timeliness; the existing solutions focus more on algorithm research and lack the ability to deploy automated and systematic solutions for production environments, resulting in insufficient practicality.

[0005] Therefore, how to achieve regional batch base station outage prediction has become an urgent problem to be solved. Summary of the Invention

[0006] This application provides a base station outage prediction method, apparatus, electronic device, and storage medium to address the shortcomings of existing technologies that only predict single base stations and are difficult to cope with regional batch base station outages, thereby achieving accurate prediction of regional batch base station outages.

[0007] This application provides a method for predicting base station outages, including the following steps: Obtain the cumulative sequence of multiple types of alarms and environmental characteristics of the target area within a historical time period, and form a multi-frame feature vector arranged along the time axis; The multi-frame feature vectors and the city-specific hot codes are input into the prediction model for the number of base station outages, and the number of base station outages in the target area in the future prediction time period is obtained from the prediction model. The prediction model comprises multiple cascaded prediction units. The predicted value output by the previous prediction unit is written back as part of the input features of the next prediction unit, which is used to fill in the unknown label information in the feature vector step by step.

[0008] According to the base station outage prediction method provided in this application, the prediction model includes a primary model and a secondary model; the primary model constitutes the prediction unit; the secondary model is composed of multiple primary models connected in series, and is used to write the predicted value output by the previous primary model back to the input feature vector of the next primary model.

[0009] According to the base station outage prediction method provided in this application, in the secondary model, the first primary model reads in several frames of historical feature vectors with complete known information, outputs the predicted value and writes it into the label position of the feature vector of the next frame; each subsequent primary model takes the feature vector matrix updated by the previous primary model as input, fills in the unknown label part step by step, and outputs the number of base stations out of service in the future prediction time period corresponding to the latest anchor time.

[0010] According to the base station outage prediction method provided in this application, the prediction unit includes an input module, a Transformer encoder module, a fully connected layer module, and an output layer module; The input module is used to perform data normalization and vector amplification on the input feature vector; The unique hot code of the city and the feature vector after vector amplification are concatenated and then input into the Transformer encoder module; The single-frame input of the Transformer encoder includes the alarm count accumulation sequence, which is directly used as the input information of the Transformer encoder in the time and feature dimensions in vector form.

[0011] According to the base station outage prediction method provided in this application, the city-specific hot code is determined based on the following method: Multiple prefecture-level administrative regions are sequentially coded; each prefecture-level administrative region corresponds to a unique region code. When performing base station outage prediction for a target area, a unique hot vector with a length equal to the total number of prefecture-level administrative regions is generated based on the regional code of the prefecture-level administrative region to which the target area belongs, and this vector is used as the unique hot code of the prefecture-level administrative region.

[0012] According to the base station outage prediction method provided in this application, the step of obtaining the cumulative sequence of multiple types of alarms and environmental characteristics of the target area within a historical time period to form a multi-frame feature vector arranged along the time axis includes: Using the current time as the anchor point, multiple historical moments are determined with a fixed time step. Each historical moment is used as a statistical baseline moment. The cumulative sequence of multiple alarm counts and environmental features corresponding to the statistical baseline moment are obtained and combined to form a single-frame feature vector corresponding to the statistical baseline moment. The single-frame feature vectors corresponding to each statistical benchmark time are arranged in chronological order to form the multi-frame feature vectors. The alarm quantity accumulation sequence includes at least one of the following: environmental alarm quantity accumulation sequence, base station outage alarm quantity accumulation sequence, and transmission alarm quantity accumulation sequence; the environmental characteristics include weather warning index and / or precipitation value.

[0013] According to the base station outage prediction method provided in this application, the prediction model is trained based on the following method: Collect historical data for the target area; the historical data includes cumulative sample sequences of multiple types of alarms, environmental sample characteristics, and the number of historical base station outages as labels. Construct an initial model containing multiple cascaded prediction units; The initial model is trained using the historical data. The error between the actual number of base station outages and the predicted value output by the model is used as the loss function. The model parameters are adjusted until the loss function value converges, and the trained prediction model is obtained.

[0014] This application also provides a base station outage prediction device, including the following modules: The acquisition module is used to acquire the cumulative sequence of multiple types of alarms and environmental features of the target area within a historical time period, forming a multi-frame feature vector arranged along the time axis. The base station outage prediction module is used to input the multi-frame feature vector and the city-specific hot code into the prediction model for the number of base station outages, and obtain the number of base stations outages in the target area in the future prediction time period output by the prediction model. The prediction model comprises multiple cascaded prediction units. The predicted value output by the previous prediction unit is written back as part of the input features of the next prediction unit, which is used to fill in the unknown label information in the feature vector step by step.

[0015] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the base station outage prediction methods described above.

[0016] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the base station outage prediction method as described above.

[0017] The base station outage prediction method, apparatus, electronic device, and storage medium provided in this application acquire a cumulative sequence of multiple alarm counts and environmental features of a target area over a historical time period, forming a multi-frame feature vector arranged along the time axis. The multi-frame feature vector and city-specific hot-coding are input into a prediction model for the number of base station outages, yielding the output of the prediction model for the number of base station outages in the target area within a future prediction time period. The prediction model contains multiple cascaded prediction units, with the predicted value output by the previous prediction unit being written back as part of the input features of the next prediction unit, used to progressively fill in unknown label information in the feature vector. This application, by acquiring a cumulative sequence of multiple alarm counts and environmental features of a target area and forming a multi-frame feature vector, and combining it with city-specific hot-coding input into a prediction model containing multiple cascaded prediction units, and using the predicted value output by the previous prediction unit written back as the input features of the next prediction unit to progressively fill in unknown label information, can predict the number of regional batch base station outages at the district / county level, effectively solving the problem that existing technologies can only predict single base station outages. Simultaneously, the cumulative sequence input and iterative write-back structure improves prediction accuracy and timeliness. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the base station outage prediction method provided in this application.

[0020] Figure 2 This is a schematic diagram of the secondary model structure provided in this application.

[0021] Figure 3 This is a schematic diagram of the primary model structure provided in this application.

[0022] Figure 4 This is the input / output schematic diagram of the district / county base station outage prediction model provided in this application.

[0023] Figure 5 This is a schematic diagram of the internal structure of a single-frame feature vector provided in this application.

[0024] Figure 6 This is a schematic diagram of the base station outage prediction system architecture provided in this application.

[0025] Figure 7 This is a schematic diagram of the base station outage prediction device provided in this application.

[0026] Figure 8 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] Currently, methods or systems for predicting communication base station outages mainly have the following drawbacks: ① It is difficult to predict regional and large-scale base station outages. Existing base station outage prediction methods all focus on individual base stations, and the input information for the prediction models is mainly based on a single base station. In actual network operation, regional and large-scale base station outages caused by severe weather, transmission interruptions, etc., can lead to large-scale wireless communication disruptions, resulting in significant economic losses and social impacts. Currently, there is a lack of risk assessment or prediction methods for regional and large-scale base station outages at the county or district level.

[0029] ② The prediction algorithms are relatively limited. Existing methods mainly use traditional algorithms such as manual rules, RNN (Recurrent Neural Network), LSTM, and decision trees for fault prediction, and their prediction accuracy and effectiveness are difficult to guarantee.

[0030] ③ The completion rate is low and it is not suitable for actual production environments. Existing base station outage prediction methods mainly focus on theoretical analysis and algorithm research. In actual production environments, base station outage prediction systems are required to have high timeliness and automation. Therefore, it is necessary to develop a comprehensive system that integrates feature information collection, model training, and predictive inference.

[0031] To address the aforementioned issues, this application proposes a method and system for predicting high-risk areas for batch base station outages in counties and districts based on an iterative write-back structure and a Transformer with time-series cumulative sequence input. This application designs the types and formats of input information required for batch base station outage prediction in counties and districts, and constructs an innovative prediction model algorithm structure, integrating it into an automated workflow for data acquisition, model training, and inference prediction. This system can predict the number of base stations out of service in various counties and districts nationwide within the next 24 hours, outputting a list of high-risk counties and districts based on the predicted values, alerting relevant network monitoring and maintenance departments to take preventative measures in advance and avoid large-scale failures.

[0032] The following is combined with Figures 1-8 This application describes the base station outage prediction method, apparatus, electronic device, and storage medium.

[0033] Figure 1 This is a flowchart illustrating the base station outage prediction method provided in this application, as shown below. Figure 1 As shown, the method includes the following: Step 101: Obtain the cumulative sequence of multiple alarms and environmental features of the target area within a historical time period, and form a multi-frame feature vector arranged along the time axis.

[0034] It should be understood that the target area refers to the district / county-level administrative region. Predicting at the district / county level aligns with the management units of network monitoring and maintenance departments, facilitating the direct use of prediction results to guide the scheduling of maintenance resources and the deployment of preventative measures.

[0035] It should be understood that the cumulative alarm count sequence refers to a set of numerical sequences arranged in chronological order, taking a predetermined statistical window (e.g., 24 hours) as the endpoint and calculating the cumulative alarm count from the beginning of the window to the current interval at predetermined time granularity (e.g., 1 hour). This sequence includes at least one of the following: environmental alarm count cumulative sequence, base station outage alarm count cumulative sequence, and transmission alarm count cumulative sequence, used to characterize the cumulative effect of faults over time.

[0036] It should be understood that environmental characteristics refer to external environmental factors that affect the operational stability of base stations, including weather warning indices and / or precipitation values. The weather warning index is obtained by quantifying and scoring the extreme weather warning levels of the target area within a preset time window (e.g., 1 point for blue warning, 2 points for yellow warning, 4 points for orange warning, and 8 points for red warning) and summing them up. It is used to characterize the impact of severe weather on the risk of base station outages.

[0037] It should be understood that multi-frame feature vectors refer to a vector sequence formed by arranging the single-frame feature vectors corresponding to multiple historical moments in chronological order. Each single-frame feature vector is composed of a cumulative sequence of alarm counts across multiple categories corresponding to that historical moment and environmental features. As input to the prediction model, multi-frame feature vectors contain temporally accumulated information in both the time and feature dimensions, enabling the model to perceive the gradual evolution of the fault.

[0038] Using the target area (e.g., a district or county-level administrative region) as the statistical unit and the current time as the anchor time, network operation data and environmental data for multiple historical moments are collected at fixed time steps (e.g., 2 hours). For each historical moment, using that moment as the statistical benchmark, a preset statistical window (e.g., 24 hours) is taken backward. Within the window, at preset time granularity (e.g., 1 hour) intervals, the cumulative number of alarms from the start of the window to the current interval point is calculated sequentially, resulting in cumulative sequences of environmental alarms, base station outage alarms, and transmission alarms. At the same time, the weather warning index and precipitation value corresponding to that moment are obtained as environmental features. The multiple cumulative sequences of that moment are combined with the environmental features to form a single-frame feature vector corresponding to that moment. The single-frame feature vectors corresponding to each historical moment are arranged in chronological order to form a multi-frame feature vector.

[0039] Step 102: Input the multi-frame feature vector and the city-specific hot code into the prediction model for the number of base station outages, and obtain the number of base station outages in the target area within the future prediction time period output by the prediction model.

[0040] It should be understood that city-level unique hot coding refers to sequentially numbering prefecture-level administrative regions nationwide, with each prefecture-level administrative region corresponding to a unique regional code. City-level unique hot coding is used to characterize the prefecture-level administrative region to which the target county belongs, enabling the prediction model to differentiate the input characteristics of different regions.

[0041] It should be understood that the prediction model comprises multiple cascaded prediction units. The predicted value output by the previous prediction unit is written back as part of the input features of the next prediction unit, used to progressively fill in the unknown label information in the feature vector. Here, the unknown label information refers to the label values ​​of the number of base stations out of service within future time periods corresponding to several frames close to the current anchor time in the multi-frame feature vector. At the prediction time, these label values ​​cannot be known in advance because the corresponding future time periods have not yet occurred, and therefore are missing in the feature vector, requiring model prediction to fill them in. Through an iterative write-back structure, using the label information of known frames, multiple cascaded prediction units progressively infer the label values ​​of unknown frames: the first prediction unit outputs a prediction of the first level of unknown labels based on frames with known labels and writes it back to fill in the missing labels; subsequent prediction units, based on the filled extended feature vector, progressively predict unknown labels closer to the current time. In this way, the model can obtain complete multi-frame feature vector input during the inference phase, thereby outputting an accurate prediction of the future number of outages.

[0042] The multi-frame feature vector containing information from multiple historical moments is concatenated and fused with the city-level hot code representing the prefecture-level administrative region to which the target area belongs, and used together as the input data of the prediction model. The prediction model performs forward calculation on the input data, and after multi-layer coding, feature transformation and regression mapping within the model, it outputs a value, which represents the total number of base stations expected to go out of service in the target area within the future prediction period (such as the next 24 hours).

[0043] In one embodiment, assume that the constructed multi-frame feature vector is S=[V1, V2, ..., Vm], where each Vk is a single-frame feature vector corresponding to a historical moment, with a length of L (e.g., L=124); let the city-level hot code be D, with a length of M (M is the total number of prefecture-level administrative regions in the country, e.g., M=300). First, the multi-frame feature vector S is flattened or its temporal structure is preserved, and then concatenated with the city-level hot code D to form a joint input feature vector X=[S, D] or X=[V1, V2, ..., Vm, D] (where D is concatenated with each frame or the whole, depending on the model design). The input X is fed into the prediction model. Internally, the model uses a Transformer encoder to perform multi-layer self-attention computation on the input features, capturing the temporal dependencies between frames and the modulating effect of city / district codes on the features. A fully connected layer maps the encoded high-dimensional features to a one-dimensional regression space, ultimately outputting a non-negative integer or floating-point number. This value represents the predicted number of base stations out of service in the target area within a future prediction period (e.g., the next 24 hours). For example, when the prediction model outputs a value of 15, it means that 15 base stations in that district / county are expected to go out of service within the next 24 hours. This predicted value can be directly used for subsequent screening of high-risk districts / counties and for early warning notifications.

[0044] The base station outage prediction method provided in this application obtains the cumulative sequence of multiple alarms and environmental features of a target area over a historical time period, forming a multi-frame feature vector arranged along the time axis. The multi-frame feature vector and city-specific hot-coding are input into a prediction model for the number of base station outages, yielding the number of base stations out of service in the target area within a future prediction time period. The prediction model contains multiple cascaded prediction units, with the predicted value output by the previous prediction unit being written back as part of the input features of the next prediction unit, used to progressively fill in unknown label information in the feature vector. This application, by obtaining the cumulative sequence of multiple alarms and environmental features of the target area and forming a multi-frame feature vector, and combining it with city-specific hot-coding input into a prediction model containing multiple cascaded prediction units, and using the predicted value output by the previous prediction unit written back as the input features of the next prediction unit to progressively fill in unknown label information, can predict the number of regional batch base station outages at the district / county level, effectively solving the problem that existing technologies can only predict single base station outages. Simultaneously, the cumulative sequence input and iterative write-back structure improves prediction accuracy and timeliness.

[0045] Based on the above embodiments, the prediction model includes a primary model and a secondary model; the primary model constitutes the prediction unit; the secondary model is composed of multiple primary models connected in series, and is used to write the predicted value output by the previous primary model back to the input feature vector of the next primary model.

[0046] like Figure 2 As shown, the prediction model adopts a two-level architecture. The primary model is the most basic prediction unit, and each primary model independently completes the mapping from input features to predicted values ​​of the number of service failures. The secondary model is composed of multiple primary models in a series. During the inference process of the secondary model, the first primary model receives the original input features and outputs the predicted value. This predicted value is written back and used as part of the input features of the second primary model. The second primary model outputs a new predicted value based on the updated features, and so on. Each subsequent primary model receives the predicted value written back from the previous level in turn, updates the input feature vector level by level, and finally the last primary model outputs the final prediction result.

[0047] The purpose of the above iterative write-back structure is to solve the problem of near-label agnosticity: in multi-frame feature vectors, the future number of service termination labels corresponding to frames closest to the current anchor time are unknown (because they have not yet occurred at the prediction time, corresponding to...). Figure 2 The orange portion (indicated in the image) cannot be directly used as effective input for training or inference.

[0048] This application embodiment achieves the step-by-step filling of unknown label information through a two-level architecture of primary and secondary models and an iterative write-back mechanism, making full use of all multi-frame feature information, while increasing the number of model parameters and nonlinear expression capabilities, effectively improving prediction accuracy.

[0049] Based on the above embodiments, in the secondary model, the first primary model reads in several frames of historical feature vectors with complete known information, outputs the predicted value and writes it into the label position of the feature vector of the next frame; each subsequent primary model takes the feature vector matrix updated by the previous primary model as input, fills in the unknown label part step by step, and outputs the number of base stations out of service in the future predicted time period corresponding to the latest anchor time.

[0050] refer to Figure 2In the secondary model, which consists of multiple primary models connected in series, the first primary model uses the feature vectors of several frames with fully known label information as input, outputs the predicted value of the unknown label in the next frame, and fills the label position of the corresponding frame with the predicted value. The second primary model takes the feature vector matrix updated by the first primary model (containing the filled label) as input, outputs the predicted value of the unknown label in the next frame after that, and writes it. And so on, each subsequent primary model receives the updated feature vector matrix of the previous level in turn, fills in the unknown label level by level, until the last primary model outputs the number of base stations out of service in the future prediction time period corresponding to the latest anchor time (i.e. the frame closest to the current time). This output is the final prediction result of the secondary model.

[0051] For example, suppose the secondary model consists of N primary models connected in series, denoted as M1, M2, ..., MN. The multi-frame feature vector contains a total of F frames, arranged from earliest to latest as Frame1, Frame2, ..., FrameF. The labels for the future service termination counts corresponding to the earlier frames (such as Frame1 to FrameK) are known (because the future time period corresponding to the historical moment of these frames has passed), while the labels for the later frames (such as FrameK+1 to FrameF) are unknown.

[0052] The input to the first primary model M1 is the feature vectors of K frames from Frame1 to FrameK (i.e., several frames with complete known information). M1 makes predictions based on this known information, outputs the predicted value P1 for the label of frame K+1, and writes P1 into the label position of Frame K+1. At this time, the feature vector matrix is ​​updated to Frame1 to FrameK+1 (where the label of Frame K+1 is P1).

[0053] The second primary model M2 takes the updated feature vectors of frames 1 to Frame K+1 as input. M2 makes predictions based on the extended feature matrix containing the first-level prediction values ​​and outputs the predicted value P2 for the label of frame K+2. P2 is then written to the label position of frame K+2, and the feature vector matrix is ​​further updated to Frame 1 to Frame K+2.

[0054] Repeat the above process. The input to the i-th primary model Mi is the feature vector matrix updated to FrameK+i-1, and the output is the predicted value Pi for the FrameK+i frame label, which is then written. This process continues until the N-th primary model MN outputs the predicted value PN for the future number of service outages for the frame corresponding to the latest anchor time (i.e., FrameF). PN is the final output of the secondary model.

[0055] In one embodiment, the secondary model consists of 11 primary models cascaded together, with a total of 23 frames (Frame 1 to Frame 23) of multi-frame feature vectors. The labels for the first 12 frames (Frame 1 to Frame 12) are known, while the labels for the last 11 frames (Frame 13 to Frame 23) are unknown. The first primary model reads in Frame 1 to Frame 12, outputs the predicted value for the label of Frame 13, and writes it; the second primary model reads in Frame 1 to Frame 13 (where the label of Frame 13 has been filled), outputs the predicted value for the label of Frame 14, and writes it; and so on, with the eleventh primary model reading in Frame 1 to Frame 22 and outputting the predicted value for the label of Frame 23. This predicted value is the final prediction of the number of base stations out of service within the next 24 hours corresponding to the latest anchor time.

[0056] Through the aforementioned step-by-step write-back and fill-in mechanism, the secondary model can use known label information to gradually infer unknown labels, ultimately achieving an accurate prediction of the number of future server outages at the current moment.

[0057] This application embodiment achieves autoregressive filling of unknown tags through the step-by-step concatenation and iterative write-back of the primary model, making full use of the temporal information of all frames, forming a progressive and refined prediction, and effectively improving the accuracy of the final prediction of the number of service outages.

[0058] Based on the above embodiments, the prediction unit includes an input module, a Transformer encoder module, a fully connected layer module, and an output layer module; The input module is used to perform data normalization and vector amplification on the input feature vector; The unique hot code of the city and the feature vector after vector amplification are concatenated and then input into the Transformer encoder module; The single-frame input of the Transformer encoder includes the alarm count accumulation sequence, which is directly used as the input information of the Transformer encoder in the time and feature dimensions in vector form.

[0059] refer to Figure 3 The initial model is divided into four functional modules according to the data processing flow. Specifically: Input module: Used to receive raw input data and perform preprocessing. For example... Figure 3As shown, the input module first performs adaptive normalization on the multi-frame feature vectors (e.g., 12 frames × 124 dimensions) to scale features of different dimensions to a uniform numerical range. Then, it expands the feature dimension from 124 dimensions to 248 dimensions (e.g., 12 frames × 248 dimensions) through linear interpolation. At the same time, the city code (1×1) is converted into a 1×340-dimensional one-hot vector through one-hot encoding and expanded into a 12×340-dimensional tensor through the time axis. Finally, the expanded feature vector (12×248 dimensions) and the city one-hot code (12×340 dimensions) expanded through the time axis are concatenated into a tensor to obtain a 12×588-dimensional joint input tensor, and position encoding is added to inject temporal position information.

[0060] The Transformer encoder module is used to perform deep encoding on the feature sequences output by the input module. For example... Figure 3 As shown, the Transformer encoder module consists of nine stacked Transformer encoders. Each Transformer encoder contains 12 parallel multi-head attention modules to capture long-distance dependencies between any two positions in the input sequence. The training process is then stabilized through residual connections and layer normalization, followed by nonlinear transformations through 16 feedforward networks, and finally, residual connections and layer normalization again. By stacking multiple encoders, the model can extract higher-level temporal features and inter-feature correlations layer by layer.

[0061] Fully connected layer module: Used to map the high-dimensional features output by the Transformer encoder module to the regression space. For example... Figure 3 As shown, the fully connected layer module includes multiple fully connected layers such as FC2, FC4, FC5, and FC6. Among them, FC2 concatenates the encoded features with important independent features (such as 1×8 dimensions) using tensors (8+8) to obtain fused features; FC4, FC5, and FC6 successively perform nonlinear transformations and dimensionality reduction on the fused features, gradually mapping high-dimensional features to low-dimensional representations, and finally outputting a feature vector that matches the dimension of the prediction target.

[0062] Output layer module: Used to convert the output of the fully connected layer module into the final predicted number of service outages. For example... Figure 3 As shown, the output layer module includes a Sigmoid activation function and denormalization processing. The Sigmoid function compresses the output of the fully connected layer FC6 to the (0,1) interval to obtain the normalized prediction value; the denormalization module restores the normalized prediction value to the original scale of the number of out-of-service base stations according to the normalization parameters used during training (for example, outputting a 1×2 dimensional vector, representing the predicted number of out-of-service base stations in the next 12 hours and the next 24 hours, respectively).

[0063] Through the collaborative work of the above four modules, the prediction unit can convert the original multi-frame feature vectors and city codes into accurate prediction values ​​of the number of service terminations, realizing end-to-end prediction from the original input to the final output.

[0064] The city-level one-hot encoded tensor, processed by the input module, is concatenated and fused with the multi-frame feature vector tensor, which has undergone vector augmentation, to form a new joint feature tensor. This joint tensor serves as the input to the Transformer encoder module, enabling the Transformer encoder to simultaneously perceive the temporal feature information and geographic affiliation information of the target region. For example, the feature vector after vector augmentation is 12 frames × 248 dimensions, and the city-level one-hot encoded tensor, after time-axis augmentation, is 12 frames × 340 dimensions. After concatenation, a 12-frame × 588-dimensional joint tensor is obtained. This tensor is then input into the encoder module, which consists of 9 Transformer encoder layers, for deep feature extraction.

[0065] In existing technologies, the single-frame input vector of a Transformer encoder is typically obtained by manually integrating information within a single time step or through model learning (word embedding process), while temporal sequence information is only assigned by positional encoding (such as sine and cosine positional encoding). Based on this, the embodiments of this application add temporal information to the single-frame input of the Transformer encoder. Moreover, this information is not a simple sequence of index changes, but a cumulative sequence of the number of fault alarms over time, namely, the cumulative sequence of the number of environmental alarms, the cumulative sequence of the number of 2G / 4G / 5G base station outages, and the cumulative sequence of access layer transmission ring interruption alarms. These cumulative sequences enable the model to perceive temporal information in both dimensions of the input matrix, thereby improving the performance of the prediction model.

[0066] When constructing the input tensor of the Transformer encoder, the cumulative alarm count sequence is used as part of the single-frame feature vector, participating directly in the encoding calculation in its original vector form without the need for additional word embeddings or feature transformations. This cumulative sequence not only provides the cumulative alarm information at each time point in the feature dimension, but also implicitly contains the accumulation process of alarms over time in the time dimension, since the cumulative sequence itself is a vector composed of cumulative values ​​at multiple time points. This allows the Transformer encoder to perceive temporal cumulative information simultaneously in two dimensions. For example, the cumulative alarm count sequence for environmental monitoring is 24 in length, the cumulative alarm count sequence for base station outage is 24 in length (3 sequences each for 2G, 4G, and 5G, totaling 72 dimensions), and the cumulative alarm count sequence for transmission is 24 in length. These cumulative sequences are directly concatenated into the single-frame feature vector, forming 120 dimensions (24+72+24) of a 124-dimensional vector, serving as the original input of the Transformer encoder.

[0067] This application embodiment achieves efficient fusion of multi-source features and dual-dimensional modeling of the fault accumulation process by feature normalization amplification, city-level coding splicing and fusion, and direct input of accumulated sequences into the Transformer encoder. This effectively improves the accuracy of batch base station outage prediction and the model's generalization ability to different regions.

[0068] Based on the above embodiments, the city-specific heat code is determined in the following way: Multiple prefecture-level administrative regions are sequentially coded; each prefecture-level administrative region corresponds to a unique region code. When performing base station outage prediction for a target area, a unique hot vector with a length equal to the total number of prefecture-level administrative regions is generated based on the regional code of the prefecture-level administrative region to which the target area belongs, and this vector is used as the unique hot code of the prefecture-level administrative region.

[0069] Due to the significant differences in geographical environment and network construction across the country, the patterns of base station outages vary from region to region. Therefore, the prediction model needs to have broad adaptability to the input characteristics of different regions.

[0070] All prefecture-level administrative regions (including prefecture-level cities, autonomous prefectures, and regions) that need to support forecasting nationwide are uniformly numbered, and each prefecture-level administrative region is assigned a unique positive integer identifier, forming a one-to-one mapping relationship between the identifier and the prefecture-level administrative region.

[0071] When it is necessary to predict the number of base stations out of service for a target county, the first step is to query the regional code of the prefecture-level administrative region to which the target county belongs (i.e., the unique number of the prefecture-level administrative region in the sequential code mapping table). Then, a zero vector with a length equal to the total number of prefecture-level administrative regions in the country is generated. The values ​​at the positions corresponding to the regional code in the vector are set to 1, and the remaining positions are kept to 0. The resulting vector is the unique hot code of the prefecture-level administrative region for the target county. This code serves as one of the input features of the prediction model and is used to characterize the geographical affiliation information of the target county.

[0072] This application's embodiments convert discrete regional information into a form that the model can process by using city-specific sequential encoding and one-hot vector generation. This enables the model to learn differentiated prediction strategies for different cities, effectively improving the model's adaptability to diverse regional characteristics and its prediction accuracy.

[0073] Based on the above embodiments, the step of obtaining the cumulative sequence of multiple types of alarms and environmental features of the target area within a historical time period, and forming a multi-frame feature vector arranged along the time axis, includes: Using the current time as the anchor point, multiple historical moments are determined with a fixed time step. Each historical moment is used as a statistical baseline moment. The cumulative sequence of multiple alarm counts and environmental features corresponding to the statistical baseline moment are obtained and combined to form a single-frame feature vector corresponding to the statistical baseline moment. The single-frame feature vectors corresponding to each statistical reference time are arranged in chronological order to form the multi-frame feature vectors.

[0074] It should be understood that anchor time refers to the current system time when performing prediction or feature construction, serving as a time reference point for collecting historical data.

[0075] The current system time used for prediction is taken as the time reference point (i.e., anchor time). A series of historical time points for collecting feature data are determined by moving along the historical direction of the time axis with a preset fixed time interval (e.g., 2 hours) as the step size. Each historical time point corresponds to one frame in the multi-frame feature vector. All historical time points together constitute the input time series window of the prediction model.

[0076] Each of the identified historical moments is used as an independent statistical benchmark. For each benchmark moment, the number of various alarms is accumulated backward from that moment to obtain the cumulative sequence of alarms for that moment, and the environmental features for that moment are also obtained. The cumulative sequence of alarms for that moment and the environmental features are concatenated and combined according to a preset dimensional order to form a single-frame feature vector that corresponds one-to-one with that moment. Each historical moment generates an independent single-frame feature vector, and all single-frame feature vectors together constitute a multi-frame feature vector sequence.

[0077] The single-frame feature vector generated at each historical moment (statistical baseline moment) is used as an element in the multi-frame feature vector. These elements are arranged according to the chronological order of the occurrence of each historical moment (e.g., from early to late or from late to early) to form a vector sequence with a temporal dimension. The positional encoding of each frame in this vector sequence implicitly contains the time interval information between the corresponding historical moment and the current anchor point time, enabling the prediction model to perceive the sequential dependency between frames when receiving input.

[0078] In one embodiment, the network operation information of the target county at a certain moment (anchor time) is integrated into a one-dimensional vector, referred to as the single-frame feature vector. For a single county, the overall workflow of the prediction model is as follows: using the current moment as a reference, inputting the multi-frame historical feature vector of the target county at a certain time step (e.g., 2 hours), combining it with the information of the prefecture-level city where the target county is located, and outputting the predicted value of the number of out-of-service base stations in the target county within the next 24 hours (e.g., ...). Figure 4 (As shown).

[0079] The single-frame feature vector is a one-dimensional vector of length 124 composed of integers. It is divided into multiple segments, namely the cumulative sequence of the number of dynamic environment alarms in the target county within 24 hours before the anchor time, the cumulative sequence of the number of 2G / 4G / 5G base stations out of service, the cumulative sequence of access layer transmission ring interruption alarms, the weather warning index, the precipitation value (mm), the total number of 2G / 4G / 5G base stations out of service within 12 hours after the anchor time, and the total number of 2G / 4G / 5G base stations out of service within 24 hours after the anchor time.

[0080] Taking the cumulative sequence of environmental alarms (i.e., power and environment type alarms) as an example, it is defined as the cumulative number of computer rooms in the target county that have environmental alarms (alarms are deduplicated based on computer room name, excluding engineering alarms, and the duration exceeds 10 minutes) within 24 hours before the anchor time. For example, if the sequence name is set to P and the anchor time is 02 / 05 00:00, then P[0] is the number of computer rooms that have alarms from 02 / 04 00:00 to 02 / 04 01:00, P[1] is the number of computer rooms that have alarms from 02 / 04 00:00 to 02 / 04 02:00, ..., P

[23] is the number of computer rooms that have alarms from 02 / 04 00:00 to 02 / 05 00:00.

[0081] Using the same method as above, the cumulative sequences of 2G, 4G, and 5G base station outage alarms and the cumulative sequences of access layer transmission ring interruption alarms were collected respectively, which constituted the timing information part of the single frame feature vector.

[0082] The weather warning index is calculated as follows: query the extreme weather warnings (excluding high temperature, drought, sandstorm, fog and haze warnings) for the target county within 24 hours before the anchor time. Blue warnings are worth 1 point, yellow warnings are worth 2 points, orange warnings are worth 4 points, and red warnings are worth 8 points. The sum of the scores is the weather warning index.

[0083] Number of base stations reporting outages in the next 12 / 24 hours: This is defined as the total number of 2G, 4G, and 5G base stations reporting outages in the target county within 12 / 24 hours after the anchor time (alarms are deduplicated based on alarm objects; only the latest alarm is retained for each alarm object, excluding engineering alarms, and only alarms lasting more than 30 minutes are counted). The number of base stations reporting outages in the next 24 hours also serves as the model output (label value).

[0084] The above information is assembled as follows: Figure 5 The single-frame feature vector shown is used to collect 23 frames of feature vectors according to the aforementioned time step. These feature vectors, together with the city codes, constitute the training and prediction data for the prediction model. The city codes are used to represent the prefecture-level administrative regions to which the target county belongs. More than 300 prefecture-level administrative regions across the country are coded in the form of consecutive positive integers.

[0085] This application embodiment forms a multi-frame feature vector that simultaneously contains temporal evolution information and fault accumulation effect by sampling multiple historical moments with a fixed step size, constructing cumulative sequences, and fusing environmental features. This provides high-quality structured input for the prediction model and improves the accuracy of batch base station outage prediction.

[0086] Based on the above embodiments, the prediction model is trained in the following manner: Collect historical data for the target area; the historical data includes cumulative sample sequences of multiple types of alarms, environmental sample characteristics, and the number of historical base station outages as labels. Construct an initial model containing multiple cascaded prediction units; The initial model is trained using the historical data. The error between the actual number of base station outages and the predicted value output by the model is used as the loss function. The model parameters are adjusted until the loss function value converges, and the trained prediction model is obtained.

[0087] Network operation records and environmental observation records of the target area over a past period are obtained from network management systems, operation and maintenance databases, and other external information systems according to preset time ranges and collection rules. These records serve as the raw sample data required for model training. Historical data includes cumulative sample sequences of multiple types of alarms, environmental sample characteristics, and the number of historical base station outages as labels. For example, with a time step of 2 hours, 12 anchor time points are collected per district / county per day (one every 2 hours), totaling approximately 24 × 30 × 12 = 8640 samples over the past 24 months. This represents approximately 3000 districts / counties across more than 300 prefecture-level administrative regions nationwide.

[0088] During the model training phase, an initial network structure is built, consisting of multiple prediction units (i.e., primary models) connected in a chain. Each prediction unit has the same network architecture and learnable parameters. The output of the previous prediction unit is designed as part of the input features of the next prediction unit, forming a chain-like transmission structure. The parameters of this initial model are randomly initialized or initialized using pre-trained weights before training begins, serving as the starting point for subsequent iterative training and parameter optimization. For example, a secondary model consisting of 11 primary models connected in a chain can be constructed. Each primary model contains: an input module (adaptive normalization, linear interpolation, tensor concatenation), a 9-layer Transformer encoder (each layer contains 12 attention modules and 16 feedforward networks), fully connected layers (FC2, FC4, FC5, FC6), and an output layer (Sigmoid and inverse normalization). This initial model serves as the training starting point, and subsequent training is performed end-to-end on it using historical data to optimize the parameters of each primary model, minimizing the final prediction error.

[0089] The collected historical dataset is used as training samples and input into the pre-built initial model. Forward computation is used to obtain the predicted number of service terminations output by the model. The error between this predicted value and the true label value is calculated as the loss function. The gradient of the loss function with respect to each model parameter is calculated using the backpropagation algorithm, and the optimizer is used to update the model parameters according to the gradient direction to reduce the loss value. This process is repeated for multiple rounds until the loss function value no longer decreases or a preset training termination condition is reached. At this point, the model parameters are saved, resulting in a trained prediction model. For example, the dataset is divided into training, validation, and test sets in a time dimension ratio of 80:19:1. The Adam optimizer is used, with an initial learning rate of 0.001, a batch size of 64, and a maximum training round of 100. An early stopping strategy is used, stopping training when the validation set loss does not decrease for 10 consecutive rounds. After training, the model is packaged into an ONNX format file for easy cross-platform deployment. Through the above training process, the model learns the mapping relationship from historical cumulative sequences, environmental features, and city codes to the number of future service terminations, and finally obtains a prediction model with good generalization ability.

[0090] This application embodiment uses historical data collection, serial prediction unit construction, and end-to-end joint training to enable the model to automatically learn the complex mapping relationship between alarm accumulation, environmental characteristics, and the number of service outages, forming an iterative optimization mechanism, and finally obtaining a prediction model with good generalization performance and robustness.

[0091] To further explain the base station outage prediction method proposed in this application, please refer to the following embodiments.

[0092] refer to Figure 6 The base station outage prediction system is divided into three parts: a training data acquisition module, a prediction model, and an inference prediction module. Its overall principle is as follows: Figure 6 As shown. Its overall workflow consists of the following three parts: Training data collection: By connecting with the network management system and other external systems, information such as the number of local environmental alarms, base station outage alarms, transmission alarms, weather warnings, and precipitation is collected from historical data at the city (district / county) level. This data is compiled into a dataset according to certain rules and used to train the base station outage prediction model.

[0093] Training the prediction model: Based on the dataset, a machine learning algorithm is used to build and train a model to predict the number of base stations out of service in the target district / county within the next 24 hours. This prediction model has several innovative designs to achieve better predictive capabilities, including a Transformer structure with time-series cumulative sequence input, an iterative write-back structure, and one-hot coding for each city / county.

[0094] Execute inference and prediction: After obtaining the prediction model, package it into an ONNX file and deploy it. At a fixed time every day, predict the total number of 2G, 4G and 5G communication base stations that will be out of service in all districts and counties across the country in the next 24 hours. Select the districts and counties with the highest predicted values ​​as high-risk districts and counties, issue early warning information to them, and remind the operation and maintenance personnel to pay more attention and take preventive measures against failures.

[0095] In one embodiment, the trained prediction model is deployed in a production environment to automatically perform predictive feature acquisition, inference prediction, and conclusion output. The specific execution flow is as follows: Step 1: Select a target county / district, set the current time as the initial anchor time, and the data acquisition module connects with the network management system and other external information systems via API (Application Programming Interface) to query the target county / district for the number of environmental alarms, base station outages, transmission alarms, weather warnings, precipitation, and other information in the 24 hours prior to the current anchor time. This information is then compiled into a single-frame feature vector according to predetermined rules. Simultaneously, based on the prefecture-level administrative region to which the target county / district belongs, a corresponding unique hot code for the prefecture-level city is generated. Step 1: Shift the anchor time one time step (2 hours) in the negative time axis direction, and generate a new frame of model input features and corresponding city-specific unique hot codes, similar to Step 1. Step 3: Repeat step 2 to generate multi-frame model input features (23 frames in total) arranged along the time axis and corresponding city-specific hot codes (the city-specific hot codes are the same). Step 4: Call the pre-trained ONNX model file, input the 23 frames of features and the city-specific hot codes into the model to perform inference, and output the predicted value of the number of base stations out of service in the target district / county within the next 24 hours; Step 5: Switch to the next target district / county and repeat steps 1-4 to predict the number of base stations out of service in the next 24 hours for all districts / counties across the country. Step 6: Select the districts and counties with the most predicted values, obtain a list of high-risk districts and counties, and issue an early warning notice.

[0096] This application embodiment achieves automated prediction of the number of base stations out of service in the next 24 hours across all counties and districts nationwide and screening of high-risk counties and districts through production environment deployment, automated feature collection and batch inference prediction. It has the ability to deploy with full-process automation, high timeliness and systematic deployment, and improves operation and maintenance response efficiency.

[0097] The base station outage prediction device provided in this application is described below. The base station outage prediction device described below can be referred to in correspondence with the base station outage prediction method described above.

[0098] refer to Figure 7 The base station outage prediction device provided in this application includes: The acquisition module 701 is used to acquire the cumulative sequence of multiple types of alarms and environmental features of the target area within a historical time period, forming a multi-frame feature vector arranged along the time axis. The base station outage prediction module 702 is used to input the multi-frame feature vector and the city-specific hot code into the prediction model for the number of base station outages, and obtain the number of base stations outages in the target area in the future prediction time period output by the prediction model. The prediction model comprises multiple cascaded prediction units. The predicted value output by the previous prediction unit is written back as part of the input features of the next prediction unit, which is used to fill in the unknown label information in the feature vector step by step.

[0099] The base station outage prediction device provided in this application obtains the cumulative sequence of multiple alarms and environmental features of a target area over a historical time period, forming a multi-frame feature vector arranged along the time axis. The multi-frame feature vector and city-specific hot-coding are input into a prediction model for the number of base station outages, yielding the number of base station outages in the target area within a future prediction time period. The prediction model contains multiple cascaded prediction units, with the predicted value output by the previous prediction unit being written back as part of the input features of the next prediction unit, used to progressively fill in unknown label information in the feature vector. This application, by obtaining the cumulative sequence of multiple alarms and environmental features of the target area and forming a multi-frame feature vector, and combining it with city-specific hot-coding input into a prediction model containing multiple cascaded prediction units, and using the predicted value output by the previous prediction unit written back as the input features of the next prediction unit to progressively fill in unknown label information, can predict the number of regional batch base station outages at the district / county level, effectively solving the problem that existing technologies can only predict single base station outages. Simultaneously, the cumulative sequence input and iterative write-back structure improves prediction accuracy and timeliness.

[0100] In one embodiment, the prediction model includes a primary model and a secondary model; the primary model constitutes the prediction unit; the secondary model is composed of multiple primary models connected in series, and is used to write the predicted value output by the previous primary model back to the input feature vector of the next primary model.

[0101] In one embodiment, in the secondary model, the first primary model reads in several frames of historical feature vectors with complete known information, outputs the predicted value, and writes it into the label position of the feature vector of the next frame; subsequent primary models take the feature vector matrix updated by the previous primary model as input, fill in the unknown label part step by step, and output the number of base stations out of service in the future predicted time period corresponding to the latest anchor time.

[0102] In one embodiment, the prediction unit includes an input module, a Transformer encoder module, a fully connected layer module, and an output layer module; The input module is used to perform data normalization and vector amplification on the input feature vector; The unique hot code of the city and the feature vector after vector amplification are concatenated and then input into the Transformer encoder module; The single-frame input of the Transformer encoder includes the alarm count accumulation sequence, which is directly used as the input information of the Transformer encoder in the time and feature dimensions in vector form.

[0103] In one embodiment, the base station outage prediction module 702 is further configured to: Multiple prefecture-level administrative regions are sequentially coded; each prefecture-level administrative region corresponds to a unique region code. When performing base station outage prediction for a target area, a unique hot vector with a length equal to the total number of prefecture-level administrative regions is generated based on the regional code of the prefecture-level administrative region to which the target area belongs, and this vector is used as the unique hot code of the prefecture-level administrative region.

[0104] In one embodiment, the acquisition module 701 is further configured to: Using the current time as the anchor point, multiple historical moments are determined with a fixed time step. Each historical moment is used as a statistical baseline moment. The cumulative sequence of multiple alarm counts and environmental features corresponding to the statistical baseline moment are obtained and combined to form a single-frame feature vector corresponding to the statistical baseline moment. The single-frame feature vectors corresponding to each statistical benchmark time are arranged in chronological order to form the multi-frame feature vectors. The alarm quantity accumulation sequence includes at least one of the following: environmental alarm quantity accumulation sequence, base station outage alarm quantity accumulation sequence, and transmission alarm quantity accumulation sequence; the environmental characteristics include weather warning index and / or precipitation value.

[0105] In one embodiment, the base station outage prediction device further includes a model training module, used for: Collect historical data for the target area; the historical data includes cumulative sample sequences of multiple types of alarms, environmental sample characteristics, and the number of historical base station outages as labels. Construct an initial model containing multiple cascaded prediction units; The initial model is trained using the historical data. The error between the actual number of base station outages and the predicted value output by the model is used as the loss function. The model parameters are adjusted until the loss function value converges, and the trained prediction model is obtained.

[0106] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a base station outage prediction method. The method includes: acquiring a cumulative sequence of multiple types of alarms and environmental features of a target area within a historical time period to form a multi-frame feature vector arranged along the time axis; inputting the multi-frame feature vector and the city-specific hot coding into a prediction model for the number of base station outages to obtain the number of base station outages in the target area within a future prediction time period output by the prediction model; wherein the prediction model contains multiple cascaded prediction units, and the predicted value output by the previous prediction unit is written back as part of the input features of the next prediction unit to progressively fill in the unknown label information in the feature vector.

[0107] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0108] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the base station outage prediction method provided by the above methods. The method includes: acquiring a cumulative sequence of multiple types of alarms and environmental features of a target area in a historical time period to form a multi-frame feature vector arranged along the time axis; inputting the multi-frame feature vector and the city-specific hot coding into a prediction model for the number of base station outages to obtain the number of base station outages in the target area in a future prediction time period output by the prediction model; wherein, the prediction model includes multiple cascaded prediction units, and the predicted value output by the previous prediction unit is written back as part of the input features of the next prediction unit to fill in the unknown label information in the feature vector step by step.

[0109] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the base station outage prediction method provided by the above methods. The method includes: acquiring a cumulative sequence of multiple types of alarms and environmental features of a target area within a historical time period to form a multi-frame feature vector arranged along a time axis; inputting the multi-frame feature vector and a city-specific hot-coded local area into a prediction model for the number of base station outages to obtain the number of base station outages in the target area within a future prediction time period output by the prediction model; wherein the prediction model includes multiple cascaded prediction units, and the predicted value output by the previous prediction unit is written back as part of the input features of the next prediction unit to progressively fill in the unknown label information in the feature vector.

[0110] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0111] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting base station outages, characterized in that, include: Obtain the cumulative sequence of multiple types of alarms and environmental characteristics of the target area within a historical time period, and form a multi-frame feature vector arranged along the time axis; The multi-frame feature vectors and the city-specific hot codes are input into the prediction model for the number of base station outages, and the number of base station outages in the target area in the future prediction time period is obtained from the prediction model. The prediction model comprises multiple cascaded prediction units. The predicted value output by the previous prediction unit is written back as part of the input features of the next prediction unit, which is used to fill in the unknown label information in the feature vector step by step.

2. The base station outage prediction method according to claim 1, characterized in that, The prediction model includes a primary model and a secondary model; the primary model constitutes the prediction unit; the secondary model is composed of multiple primary models connected in series, and is used to write the predicted value output by the previous primary model back to the input feature vector of the next primary model.

3. The base station outage prediction method according to claim 2, characterized in that, In the secondary model, the first primary model reads in several frames of historical feature vectors with complete known information, outputs the predicted value, and writes it into the label position of the feature vector of the next frame; each subsequent primary model takes the feature vector matrix updated by the previous primary model as input, fills in the unknown label part step by step, and outputs the number of base stations out of service in the future predicted time period corresponding to the latest anchor time.

4. The base station outage prediction method according to claim 1, characterized in that, The prediction unit includes an input module, a Transformer encoder module, a fully connected layer module, and an output layer module; The input module is used to perform data normalization and vector amplification on the input feature vector; The unique hot code of the city and the feature vector after vector amplification are concatenated and then input into the Transformer encoder module; The single-frame input of the Transformer encoder includes the alarm count accumulation sequence, which is directly used as the input information of the Transformer encoder in the time and feature dimensions in vector form.

5. The base station outage prediction method according to claim 1, characterized in that, The city-specific heat code is determined based on the following method: Multiple prefecture-level administrative regions are sequentially coded; each prefecture-level administrative region corresponds to a unique region code. When performing base station outage prediction for a target area, a unique hot vector with a length equal to the total number of prefecture-level administrative regions is generated based on the regional code of the prefecture-level administrative region to which the target area belongs, and this vector is used as the unique hot code of the prefecture-level administrative region.

6. The base station outage prediction method according to claim 1, characterized in that, The process involves acquiring the cumulative sequence of multiple alarm types and environmental features of the target area over a historical time period, forming a multi-frame feature vector arranged along the time axis, including: Using the current time as the anchor point, multiple historical moments are determined with a fixed time step. Each historical moment is used as a statistical baseline moment. The cumulative sequence of multiple alarm counts and environmental features corresponding to the statistical baseline moment are obtained and combined to form a single-frame feature vector corresponding to the statistical baseline moment. The single-frame feature vectors corresponding to each statistical benchmark time are arranged in chronological order to form the multi-frame feature vectors. The alarm quantity accumulation sequence includes at least one of the following: environmental alarm quantity accumulation sequence, base station outage alarm quantity accumulation sequence, and transmission alarm quantity accumulation sequence; the environmental characteristics include weather warning index and / or precipitation value.

7. The base station outage prediction method according to claim 1, characterized in that, The prediction model was trained in the following manner: Collect historical data for the target area; the historical data includes cumulative sample sequences of multiple types of alarms, environmental sample characteristics, and the number of historical base station outages as labels. Construct an initial model containing multiple cascaded prediction units; The initial model is trained using the historical data. The error between the actual number of base station outages and the predicted value output by the model is used as the loss function. The model parameters are adjusted until the loss function value converges, and the trained prediction model is obtained.

8. A base station outage prediction device, characterized in that, include: The acquisition module is used to acquire the cumulative sequence of multiple types of alarms and environmental features of the target area within a historical time period, forming a multi-frame feature vector arranged along the time axis. The base station outage prediction module is used to input the multi-frame feature vector and the city-specific hot code into the prediction model for the number of base station outages, and obtain the number of base stations outages in the target area in the future prediction time period output by the prediction model. The prediction model comprises multiple cascaded prediction units. The predicted value output by the previous prediction unit is written back as part of the input features of the next prediction unit, which is used to fill in the unknown label information in the feature vector step by step.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the base station outage prediction method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the base station outage prediction method as described in any one of claims 1 to 7.

Images