Alarm early warning method and device for base station equipment

Abstract

The invention discloses a base station equipment alarm early warning method and device. Comprising the following steps: acquiring a first multi-dimensional data sequence and a first base station adjacency matrix of first base station equipment in a first time period; performing feature extraction on the first multi-dimensional data sequence and the first base station adjacent matrix by using a multi-head attention network and a graph attention network of a base station equipment alarm prediction model to obtain a time sequence state vector and a space state vector, and performing feature fusion on the time sequence state vector and the space state vector by using a space-time cross attention network to obtain a time sequence state vector and a space state vector; obtaining a space-time state vector; analyzing the space-time state vector by using an output network, and determining a first prediction alarm type of the first base station equipment at the next moment of the first time period; and feeding back first early warning prompt information at least carrying the first prediction alarm type to the management terminal. The technical problem that the early warning result is poor in accuracy due to the fact that the future warning type is predicted and early warning is carried out only according to the time sequence data of the base station equipment in the prior art is solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a method and device for alarm warning of base station equipment. Background Art

[0002] In 4G / 5G wireless communication networks, efficient operation and maintenance of base station equipment is crucial for ensuring network stability and user experience. However, traditional base station equipment alarm and early warning methods are often limited to single time series analysis, such as monitoring the temporal evolution of equipment performance data, dynamic environment data, or historical alarm records to predict possible future failures. While this approach can capture the trend of equipment status changes over time, it cannot fully understand the spatial complexity of base station equipment, resulting in limited accuracy of prediction results. This is especially true when dealing with failures caused by interactions between base station equipment, which can easily lead to false positives or missed reports. This limits maintenance personnel's ability to proactively detect and prevent network problems, increases the delay and cost of network failure response, and poses a potential threat to network service quality.

[0003] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0004] The embodiments of the present application provide a base station equipment alarm warning method and device to at least solve the technical problem that the related technology only predicts the future alarm type of the base station equipment and issues a warning based on the time series data of the base station equipment, resulting in poor accuracy of the warning results.

[0005] According to one aspect of an embodiment of the present application, a base station device alarm warning method is provided, including: obtaining a first multidimensional data sequence and a first base station adjacency matrix of a first base station device in a first time period, wherein the first multidimensional data sequence includes a performance data set, a dynamic environment data set and an alarm data set of the first base station device at each time step in the first time period, and the first base station adjacency matrix is ​​used to reflect the adjacency relationship between the first base station device and other base station devices in the first time period; using a multi-head attention network and a graph attention network of a base station device alarm prediction model to perform feature extraction on the first multidimensional data sequence and the first base station adjacency matrix to obtain a time series state vector and a space state vector, and using a time-space cross attention network of the base station device alarm prediction model to perform feature fusion on the time series state vector and the space state vector to obtain a time-space state vector; using the output network of the base station device alarm prediction model to analyze the time-space state vector to determine a first predicted alarm type of the first base station device at the next moment in the first time period; and feeding back to the management terminal a first warning prompt information carrying at least the first predicted alarm type.

[0006] Optionally, the training process of the base station equipment alarm prediction model includes: obtaining multiple groups of training sample data, wherein each group of training sample data includes: a second multidimensional data sequence and a second base station adjacency matrix of the second base station device in a second time period as training samples, and a second alarm type of the second base station device at the next moment of the second time period as a training label, and the second time period is the time period before the first time period; constructing a neural network model, wherein the neural network model includes at least: a multimodal cross-attention network, a multi-head attention network, a graph attention network, a spatiotemporal cross-attention network, and an output network; and iteratively training the neural network model using multiple groups of training sample data to obtain a base station equipment alarm prediction model.

[0007] Optionally, multiple groups of training sample data are obtained, including: for each second base station device, respectively obtaining the initial multidimensional data sequence and initial spatial topology data of the second base station device in multiple second time periods, and obtaining the second alarm type of the second base station device at the next moment of each second time period, wherein the initial multidimensional data sequence includes: a performance data set, a dynamic loop data set and an alarm data set of the second base station device at each time step in the second time period, the performance data set includes at least one of the following: physical resource block utilization, number of wireless resource control connections, wireless connectivity rate, switching success rate, the dynamic loop data set includes at least one of the following: ambient temperature, ambient humidity, input voltage, standing wave ratio, optical attenuation, the alarm data set includes at least: alarm name, alarm type, alarm location, the initial spatial topology data at least Including one of the following: the number of switching between the subordinate cells of the second base station device and the subordinate cells of the adjacent base station device, whether the second base station device and the adjacent base station device share resources; performing preprocessing operations on multiple initial multidimensional data sequences and multiple initial spatial topology data respectively to obtain the second multidimensional data sequence and the second base station adjacency matrix of the second base station device in each second time period, wherein the preprocessing operation includes at least one of the following: time alignment, standardization processing, outlier processing, and construction of an adjacency matrix; using the second multidimensional data sequence and the second base station adjacency matrix of each second base station device in multiple second time periods as multiple training samples, and using the second alarm type of each second base station device at the next moment of the multiple second time periods as the sample label of the corresponding training sample to form multiple groups of training sample data.

[0008] Optionally, the neural network model is iteratively trained using multiple groups of training sample data to obtain a base station equipment alarm prediction model, including: for each training batch in the iterative training process, each training sample of the training batch is input into the neural network model to obtain each second predicted alarm type output by the neural network model, using the second predicted alarm type and the corresponding sample label to construct a multi-classification cross entropy loss function, and adjusting the model parameters of the neural network model based on the multi-classification cross entropy loss function.

[0009] Optionally, each training sample of the training batch is input into the neural network model to obtain each second prediction alarm type output by the neural network model, including: inputting each training sample of the training batch into the neural network model, and obtaining the second prediction alarm type corresponding to each training sample output by the neural network model according to the following process: using a multimodal cross-attention network to use the feature vector corresponding to the alarm data set of the second base station in the training sample at each time step in the second time period as the first query vector, and using the feature vectors corresponding to the dynamic environment data set and the performance data set of the second base station in the training sample at each time step in the second time period as the first key vector; for each time step in the second time period, calculating the inner product of the first query vector corresponding to the time step and the first key vectors corresponding to the time step to obtain the first attention coefficient of the dynamic environment data set and the performance data set at the time step, and using the first attention coefficient to perform weighted summation on the feature vectors corresponding to the dynamic environment data set and the performance data set at the time step, and adding the feature vector corresponding to the alarm data set at the time step containing the time step position vector to obtain the state vector of the second base station device at the time step; using a multi-head attention network to use the state vector of the second base station device at any time step in the second time period as the second query vector, and the state vectors of the second base station device at each other time step in the second time period are respectively used as the second key vectors, and the inner product of the second query vector and each second key vector is calculated to obtain the second attention coefficient of each other time step; the state vectors at each other time step are weighted summed using the second attention coefficient, and the state vector of the second base station device at the time step is added to obtain the weighted state vector of the second base station device at the time step; the weighted state vectors of the second base station device at each time step in the second time period are sequentially spliced ​​to obtain the time series state vector of the second base station device in the second time period; the graph attention network is used to take the feature vector corresponding to the base station spatial topology data of the second base station device in the second time period as the third query vector, and the feature vectors corresponding to the base station spatial topology data of each adjacent base station device of the second base station device in the second time period are respectively used as the third key vector, and the inner product of the third query vector and each third key vector is calculated to obtain the third attention coefficient of each adjacent base station device; the feature vectors corresponding to the base station spatial topology data of each adjacent base station device in the second time period are weighted summed using the third attention coefficient to obtain the spatial state vector of the second base station device in the second time period;Using a spatiotemporal cross-attention network, the temporal state vector of the second base station device in the second time period is used as a fourth query vector, and the spatial state vectors of each of the second base station device's neighboring base stations in a preset time period are used as a fourth key vector. By calculating the inner product of the fourth query vector and each of the fourth key vectors, a fourth attention coefficient for each of the neighboring base stations is obtained. The fourth attention coefficient is used to perform a weighted sum of the spatial state vectors of the multiple neighboring base stations in the second time period, and the sum is added to the temporal state vector of the second base station device in the second time period to obtain the spatiotemporal state vector of the second base station device in the second time period. The output network is used to analyze the spatiotemporal state vector to obtain a second predicted alarm type for the second base station device at the next moment in the second time period.

[0010] Optionally, the first early warning prompt information carrying at least the first predicted alarm type is fed back to the management terminal, including: analyzing the first attention coefficients of the dynamic environment data set and the performance data set of the first base station device at all time steps in the first time period based on the multimodal cross-attention network, and taking the time step corresponding to the largest first attention coefficient as the characteristic alarm root cause; analyzing the second attention coefficient of the time series state vector of the first base station device at each time step in the first time period based on the multi-head attention network, and taking the time step corresponding to the largest second attention coefficient as the timing alarm root cause; analyzing the fourth attention coefficients of each adjacent base station device of the first base station device based on the spatiotemporal cross-attention network, and taking the adjacent base station device corresponding to the largest fourth attention coefficient as the spatial alarm root cause; and feeding back the first early warning prompt information carrying the first predicted alarm type, the characteristic alarm root cause, the timing alarm root cause and the spatial alarm root cause to the management terminal.

[0011] Optionally, after determining the first predicted alarm type of the first base station device at the next moment in the first time period based on the space-time state vector, the method further includes: obtaining the first predicted alarm type of other base station devices in the target network including the first base station device at the next moment in the first time period; matching the first predicted alarm type of each base station device in the target network at the next moment in the first time period with a preset batch decommissioning alarm type set, and determining the number of base station devices whose first predicted alarm type is a batch decommissioning alarm type, wherein the batch decommissioning alarm type includes at least one of the following: remote control radio frequency unit link disconnection, input power disconnection, radio frequency unit link abnormality; taking the ratio of the number of base station devices to the total number of base station devices in the target network as the batch decommissioning probability of base station devices in the target network ; When the probability of batch deservice of base station equipment is not lower than the preset batch deservice probability threshold, a second early warning prompt information indicating that the target network has a high probability of batch deservice of base station equipment is fed back to the management terminal, wherein the form of the second early warning prompt information at least includes: the indicator light flashes at a first frequency, and the buzzer emits a prompt tone at a first frequency and a first volume; when the probability of batch deservice of base station equipment is lower than the batch deservice probability threshold, a third early warning prompt information indicating that the target network has a low probability of batch deservice of base station equipment is fed back to the management terminal, wherein the form of the third early warning prompt information at least includes: the indicator light flashes at a second frequency, and the buzzer emits a prompt tone at a second frequency and a second volume; wherein the first frequency is higher than the second frequency, and the first volume is higher than the second volume.

[0012] According to another aspect of an embodiment of the present application, a base station equipment alarm warning device is also provided, including: an acquisition module for acquiring a first multidimensional data sequence and a first base station adjacency matrix of a first base station device in a first time period, wherein the first multidimensional data sequence includes a performance data set, a dynamic environment data set and an alarm data set of the first base station device at each time step in the first time period, and the first base station adjacency matrix is ​​used to reflect the adjacency relationship between the first base station device and other base station devices in the first time period; a feature fusion module for using the multi-head attention network and the graph attention network of the base station equipment alarm prediction model to extract features from the first multidimensional data sequence and the first base station adjacency matrix to obtain a time series state vector and a spatial state vector, and using the time-space cross attention network of the base station equipment alarm prediction model to perform feature fusion on the time series state vector and the spatial state vector to obtain a time-space state vector; a prediction module for analyzing the time-space state vector using the output network of the base station equipment alarm prediction model to determine the first predicted alarm type of the first base station device at the next moment in the first time period; and an early warning module for feeding back to the management terminal a first early warning prompt information carrying at least the first predicted alarm type.

[0013] According to another aspect of an embodiment of the present application, a computer program product is further provided, the computer program product comprising: a computer program, wherein when the computer program is executed by a processor, the above-mentioned base station device alarm warning method is implemented.

[0014] According to another aspect of an embodiment of the present application, an electronic device is further provided, which includes: a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to execute the above-mentioned base station device alarm warning method through the computer program.

[0015] In an embodiment of the present application, a base station device alarm prediction model is used to perform spatiotemporal analysis on the first multidimensional data sequence and the first base station adjacency matrix of the first base station device within the first time period to obtain the first predicted alarm type of the first base station device at the next moment in the first time period, thereby improving the accuracy and reliability of the alarm prediction; finally, the first early warning prompt information carrying at least the first predicted alarm type is fed back to the management terminal, thereby achieving the purpose of early warning of potential alarms of the base station device in order to actively maintain and optimize network performance, thereby solving the technical problem that the related technology only predicts the future alarm type of the base station device and issues an early warning based on the time series data of the base station device, resulting in poor accuracy of the early warning results. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0017] Figure 1 This is a flow chart of an optional base station device alarm warning method according to an embodiment of the present application;

[0018] Figure 2 This is an optional principle diagram for obtaining training sample data according to an embodiment of the present application;

[0019] Figure 3 is a schematic diagram of an optional model analysis according to an embodiment of the present application;

[0020] Figure 4 is a schematic diagram of an optional root cause location according to an embodiment of the present application;

[0021] Figure 5 This is a schematic structural diagram of an optional base station equipment alarm warning device according to an embodiment of the present application;

[0022] Figure 6 It is a schematic structural diagram of an optional electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0024] It should be noted that the terms "first", "second", etc. in the specification, claims, and drawings of the present application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0025] Example 1

[0026] According to an embodiment of the present application, a base station equipment alarm warning method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0027] Figure 1 is a flow chart of a base station device alarm warning method provided according to an embodiment of the present application, such as Figure 1 As shown, the method includes the following steps:

[0028] Step S102: Obtain a first multidimensional data sequence and a first base station adjacency matrix of a first base station device within a first time period. The first multidimensional data sequence includes a performance dataset, a dynamic environment dataset, and an alarm dataset of the first base station device at each time step within the first time period. The first base station adjacency matrix is ​​used to reflect the adjacency relationship between the first base station device and other base stations within the first time period.

[0029] In step S104, the multi-head attention network and graph attention network of the base station equipment alarm prediction model are used to extract features of the first multidimensional data sequence and the first base station adjacency matrix to obtain a time series state vector and a spatial state vector, and the time-space cross attention network of the base station equipment alarm prediction model is used to perform feature fusion on the time series state vector and the spatial state vector to obtain a time-space state vector.

[0030] Step S106 : Analyze the spatiotemporal state vector using the output network of the base station device alarm prediction model to determine a first predicted alarm type of the first base station device at the next moment in the first time period.

[0031] Step S108: Feedback first warning prompt information carrying at least the first prediction alarm type to the management terminal.

[0032] Based on the scheme defined by the above steps S102 to S108, it can be known that in an embodiment of the present application, a base station device alarm prediction model is used to perform spatiotemporal analysis on the first multidimensional data sequence and the first base station adjacency matrix of the first base station device in the first time period, and the first predicted alarm type of the first base station device at the next moment in the first time period is obtained, thereby improving the accuracy and reliability of the alarm prediction; finally, the first early warning prompt information carrying at least the first predicted alarm type is fed back to the management terminal, thereby achieving the purpose of early warning of potential alarms of the base station device, so as to facilitate the active maintenance and optimization of network performance.

[0033] The following describes the steps of the base station equipment alarm warning method in conjunction with a specific implementation process.

[0034] First, the early warning system can collect an initial multidimensional data sequence of the first base station device within a first time period from a professional network management system, wherein the initial multidimensional data sequence includes a performance data set, a dynamic environment data set, and an alarm data set of the first base station device at each time step within the first time period. The performance data set includes, but is not limited to: physical resource block (PRB) utilization, number of radio resource control (RRC) connections, wireless connectivity rate, handover success rate, PDCP (Packet Data Convergence Protocol) layer traffic, etc. The dynamic environment data set includes, but is not limited to: ambient temperature, ambient humidity, input voltage, standing wave ratio, optical attenuation, etc. The alarm data set includes, but is not limited to: alarm name, alarm type, alarm location, etc. In addition, the first time period mentioned here is a predefined time window (such as one week, one month, etc.), and the time step is a fixed duration (such as 15 minutes, one hour, etc.).

[0035] At the same time, the early warning system can obtain the initial spatial topology data of the first base station device within the first time period, wherein the initial spatial topology data includes at least one of the following: the number of switches between the subordinate cells of the first base station device and the subordinate cells of the adjacent base station device, and whether resources are shared between the first base station device and the adjacent base station device.

[0036] To facilitate model input and analysis, the early warning system can perform preprocessing operations on the acquired initial multidimensional data sequence and initial spatial topology data, including but not limited to the following operations:

[0037] (1) Time alignment.

[0038] Time alignment is the process of aligning raw data of different frequencies and timestamps to the same time window to form standardized features that can be used for model input. This is mainly because: different data sources have different time granularities. For example, dynamic environment data is usually at the second or minute level, performance data collection cycles are usually 15 minutes or 30 minutes, and alarm data is event-triggered. If these time series data are not aligned, it will be impossible to perform comprehensive analysis on these time series data at the same time step. In addition, misalignment of time series data will also cause the model to be unable to correctly capture the correlation between different types of data, thereby affecting the accuracy of the model. For example, if these time series data are not aligned, the high-frequency temperature sampling corresponding to the dynamic environment data will overwhelm the key performance change points, and will also cause the timestamp of the alarm data to shift. Therefore, in the data preprocessing stage, the time series data can be time-aligned first to ensure the effectiveness and accuracy of the subsequent model.

[0039] Specifically, the strategy of time alignment is to interpolate data with different sampling frequencies (such as flow per second, temperature per minute) to a unified time granularity. In the embodiment of this application, taking all time series data as an example to be unified to a time step of 15 minutes, the time window for collecting multidimensional data series can be recorded as: T sw =[t1,t2,…t n ,…,t N ],and Δt=15min,where represents the starting time of the nth time step, Indicates the end time of the nth time step. If the first time period is 30 days, then the window length of the time window is 2880.

[0040] If each type of performance data in the performance data set is collected at the second level, the embodiment of the present application can use the average of the performance data in the nth time step as the performance data in that time step. Therefore, the performance data set in the nth time step can be expressed as:

[0041] p n=[average t∈Δt,n p prb (t),average t∈Δt,n p RRC (t),average t∈Δt,n p pdcp (t),…]

[0042] Among them, p n Represents the performance data set within the nth time step, average t∈Δt p prb (t) represents the average PRB utilization rate in the nth time step, average t∈Δt p RRC (t) represents the average number of RRC connections in the nth time step, average t∈Δt p pdcp (t) represents the mean PDCP layer traffic in the nth time step.

[0043] Since the dynamic ring data has a certain high frequency, the embodiment of the present application can use the maximum value of the dynamic ring data in the nth time step as the dynamic ring data in the time step, and then the dynamic ring data set in the nth time step can be expressed as:

[0044] s n =[max t∈Δt,n s temp (t),max t∈Δt,n s volt (t),max t∈Δt,n s vswr (t),…]

[0045] Among them, s n Represents the dynamic ring dataset in the nth time step, s temp (t) represents the maximum temperature in the nth time step, max t∈Δt s volt (t) represents the highest voltage in the nth time step, max t∈Δt s vswr (t) represents the highest SWR in the nth time step.

[0046] Since alarm data is event-type data, it is necessary to count the number of occurrences and durations of different types of alarms in each time step as the alarm data in the time step. Then the alarm data set in the nth time step can be expressed as:

[0047]

[0048] Among them, a n Represents the alarm data set in the nth time step, a hw (t),asw (t),a RRU (t),a volt (t) represents the indicator function of hardware alarm, standing wave alarm, RRU interruption alarm, and power supply voltage alarm respectively (1 if an alarm occurs, otherwise 0), max t∈Δt,n Duration(t) indicates the duration of the alarm in the nth time step (in minutes).

[0049] (2) Filling in time series data.

[0050] For missing values ​​of performance data and dynamic environment data in each time step, a filling method can be used to fill in the missing values. For example, the embodiment of the present application can use the mean of the seven time steps before the time step of the missing data to fill in the missing values.

[0051] (3) Outlier processing.

[0052] Outliers may be caused by equipment failure, temporary interference, or data collection errors. Therefore, the embodiments of the present application can process these outliers to improve the robustness and accuracy of the model. Taking any performance data in the nth time step as an example, it can be processed according to the following formula:

[0053]

[0054] Among them, median filling refers to filling the outlier with the median of the same type of performance data at each time step in the time window, μ represents the mean of the same type of performance data at each time step in the time window, and σ represents the standard deviation of the same type of performance data at each time step in the time window.

[0055] (4) Standardization processing.

[0056] The purpose of dynamic normalization is to eliminate dimensional differences between different features while adapting to changes in data distribution over time, such as baseline drift caused by seasonal changes or equipment aging. Taking the performance data in the nth time step as an example, it can be processed according to the following formula:

[0057]

[0058] Among them, p ′ n represents the normalized performance data at the nth time step, p n represents the performance data at the nth time step, μ represents the mean of the same type of performance data at each time step in the time window, and σ represents the standard deviation of the same type of performance data at each time step in the time window.

[0059] (5) Adjacency matrix construction.

[0060] When the initial spatial topology data is the number of switching between the subordinate cells of the RRU (Radio Remote Unit) of the first base station device and the subordinate cells of the RRU of the adjacent base station device, if the number of switching between the subordinate cells of the first base station device and the subordinate cells of the adjacent base station device within a preset time period exceeds a preset threshold, it is considered that the two base station devices have an adjacent relationship. When the initial spatial topology data is whether the BBU (Building Baseband Unit) in the first base station device and the BBU in the adjacent base station device share resources, if the first base station device and the adjacent base station device are connected to one device (or share other important resources), it is considered that the two base station devices have an adjacent relationship.

[0061] After the early warning system processes the initial multidimensional data sequence and the initial spatial topology data according to the above preprocessing operation, the first multidimensional data sequence and the first base station adjacency matrix can be obtained.

[0062] Furthermore, the early warning system can call a pre-trained base station device alarm prediction model to analyze the first multidimensional data sequence and the first base station adjacency matrix to obtain the first predicted alarm type of the first base station device at the next moment in the first time period.

[0063] The training process of the base station equipment alarm prediction model includes:

[0064] Step S1: Acquire multiple sets of training sample data. Each set of training sample data includes: a second multidimensional data sequence and a second base station adjacency matrix of a second base station device in a second time period as training samples, and a second alarm type of the second base station device at a next moment in the second time period as a training label, wherein the second time period is a time period before the first time period.

[0065] Specifically, in the technical solution provided in step S1 above, the method for obtaining the training sample data is as follows: Figure 2 Shown, including:

[0066] First, for each second base station device, an initial multidimensional data sequence and initial spatial topology data of the second base station device in multiple second time periods are obtained from a professional network management system, and a second alarm type of the second base station device at the next moment in each second time period is obtained. The initial multidimensional data sequence also includes: a performance data set, a dynamic environment data set, and an alarm data set of the second base station device at each time step in the second time period; the initial spatial topology data includes at least one of the following: the number of handovers between a subordinate cell of the second base station device and a subordinate cell of an adjacent base station device, and whether resources are shared between the second base station device and the adjacent base station device.

[0067] Next, preprocessing operations are performed on the multiple initial multidimensional data sequences and the multiple initial spatial topology data to obtain second multidimensional data sequences and second base station adjacency matrices for the second base station device in each second time period. The preprocessing operations include at least one of the following: time alignment (to ensure that time series data from different sources have consistent timestamps within the same time window), normalization (to eliminate dimensional differences between different data to adapt to changes in data distribution over time), outlier processing (to maintain data quality and model stability), and construction of an adjacency matrix (to quantify the spatial relationship between base station devices).

[0068] Finally, the second multidimensional data sequences and the second base station adjacency matrices of each second base station device in multiple second time periods are used as multiple training samples, and the second alarm types of each second base station device at the next moment in multiple second time periods are used as sample labels of corresponding training samples to form multiple groups of training sample data.

[0069] Through the above acquisition steps, rich information with both time series characteristics and spatial relationship characteristics can be extracted from historical data, forming multiple training samples and corresponding sample labels, providing a data foundation for subsequent iterative training of the model.

[0070] Step S2: construct a neural network model.

[0071] Specifically, the neural network model includes at least: a multimodal cross-attention network for fusing different types of time series data and evaluating the influence between modalities; a multi-head attention network for capturing the intrinsic dependencies of time series features; a graph attention network for encoding the spatial characteristics of base station equipment; a spatiotemporal cross-attention network for fusing temporal and spatial features; and an output network for converting the fused spatiotemporal features into predicted alarm types. Both the multimodal cross-attention network and the spatiotemporal cross-attention network are implemented using a Transform encoder and include positional encoding to ensure unique patterns across all time steps for each modality. The multi-head attention network uses multiple projection matrices to map inputs into different representation spaces, each representing a head. The outputs of multiple heads are concatenated, enabling parallel encoding of information across all time steps while also capturing the influencing factors of other modal features and the target modal features at each time step. It should be noted that the ultimate goal of the base station equipment alarm prediction model is to predict the alarm type. Therefore, the alarm features are defined as the target modal features, and other time series features are defined as other modal features.

[0072] Step S3: Iteratively train the neural network model using multiple sets of training sample data to obtain a base station equipment alarm prediction model.

[0073] In the technical solution provided in step S3 above, for each training batch in the iterative training process, each training sample of the training batch is input into the neural network model to obtain each second prediction alarm type output by the neural network model. Figure 3 The process shown obtains the second predicted alarm type corresponding to each training sample output by the neural network model:

[0074] Step 1: temporal coding.

[0075] Temporal encoding uses a multimodal cross-attention network to fuse the different modal features of the second base station device at each time step in the second time period to obtain the state vector of the second base station device at each time step in the second time period; then uses a multi-head attention network to capture the horizontal dependencies of the sequence and aggregate the embeddings of all time steps to obtain the final temporal state vector. The specific implementation method is as follows:

[0076] First, a multimodal cross-attention network is used to take the feature vector corresponding to the alarm dataset of the second base station in the training sample at each time step in the second time period as the first query vector, and the feature vectors corresponding to the dynamic environment dataset and performance dataset of the second base station in the training sample at each time step in the second time period as the first key vectors.

[0077] For each time step in the second time period, calculate the inner product of the first query vector corresponding to the time step (i.e., the feature vector corresponding to the alarm dataset at the current time step) and each first key vector corresponding to the time step (the feature vectors corresponding to the dynamic environment dataset and the performance dataset at the current time step), and obtain the first attention coefficient of the dynamic environment dataset and the performance dataset at the time step. The expression of the first attention coefficient can be written as:

[0078]

[0079] Among them, a represents the feature vector corresponding to the alarm dataset at the current time step, and x represents the feature vector corresponding to the dynamic environment dataset and performance dataset at the current time step. represents the first attention coefficient between the feature vector corresponding to the alarm dataset output by the h-th attention head at the n-th time step and the feature vector corresponding to the dynamic environment dataset s / performance dataset p. n represents the n-th time step in the second time period, and N represents the total number of time steps in the second time period. represents the unnormalized attention coefficient between the feature vector corresponding to the alarm dataset output by the h-th attention head at the n-th time step and the feature vector corresponding to the dynamic environment dataset / performance dataset, and POS nrepresents the position vector of the nth time step, a n Represents the feature vector corresponding to the alarm data set at the nth time step, represents the feature vector corresponding to the dynamic loop dataset and the performance dataset at the nth time step, H represents the number of attention heads, d represents the vector dimension, and W q,h represents the query matrix, W k,h Represents a matrix of values.

[0080] The first attention coefficient is used to perform a weighted summation of the eigenvectors corresponding to the dynamic environment dataset and the performance dataset at the time step, and the eigenvector corresponding to the alarm dataset at the time step, including the time step position vector, is added to obtain the state vector of the second base station device at the time step. Therefore, the state vector of the second base station device at the time step can be expressed as:

[0081]

[0082] Therefore, the state vector of the second base station device at the nth time step in the second time period is It uses the first attention coefficient All value vectors (i.e., the feature vectors corresponding to the dynamic ring dataset and the performance dataset at the nth time step) are weighted and then multi-head splicing and residual connection are performed to obtain the result.

[0083] Next, a multi-head attention network is used to take the state vector of the second base station device at any time step in the second time period as the second query vector, and the state vector of the second base station device at each other time step in the second time period as the second key vector. The inner product of the second query vector and each second key vector is calculated to obtain the second attention coefficient of each other time step.

[0084] Then, the state vectors at each other time step are weighted and summed using the second attention coefficient, and the state vector of the second base station device at the time step is added to obtain the weighted state vector of the second base station device at the time step.

[0085] Finally, the weighted state vectors of the second base station at each time step in the second time period are sequentially concatenated to obtain the time-series state vector of the second base station in the second time period. Therefore, the expression of the time-series state vector of the second base station in the second time period can be written as:

[0086]

[0087] Step 2: Spatial encoding.

[0088] Since the base station adjacency matrix of the second base station device in the second time period reveals the physical connection, environment sharing, and service dependency between base stations, the embodiment of the present application uses a graph attention network to encode the second base station adjacency matrix to obtain the spatial features of the second base station device in the second time period. The specific implementation method is as follows:

[0089] First, using the graph attention network, the feature vector corresponding to the base station spatial topology data of the second base station device in the second time period is used as the third query vector, and the feature vectors corresponding to the base station spatial topology data of each adjacent base station device in the second time period are used as third key vectors. The inner product of the third query vector and each third key vector is calculated to obtain the third attention coefficient of each adjacent base station device. Therefore, the expression of the third attention coefficient here can be written as:

[0090]

[0091] Among them, W s Represents a learnable parameter matrix, mainly used for linear transformation; LeakyReLU represents the activation function; represents the first-order neighbor set of the second base station device i; j and k both represent the numbers of the first-order neighbor base station devices of the second base station device i; a T represents the main attention parameter vector.

[0092] The third attention coefficient is used to perform a weighted summation of the eigenvectors corresponding to the base station spatial topology data of each adjacent base station device in the second time period to obtain the spatial state vector of the second base station device in the second time period. Therefore, the expression for the spatial state vector of the second base station device in the second time period can be written as:

[0093]

[0094] Where σ represents the activation function, v j Represents the spatial state vector corresponding to the base station adjacency matrix of the j-th base station device in the second time period.

[0095] Step 3: Space-time fusion.

[0096] In order to improve the accuracy of the model prediction results, the embodiment of the present application uses a spatiotemporal cross attention network to fuse the temporal state vector and spatial state vector of the second base station device in the second time period to obtain the spatiotemporal state vector of the second base station device in the second time period. The specific implementation method is as follows:

[0097] First, using the spatiotemporal cross attention network, the temporal state vector of the second base station device in the second time period is used as the fourth query vector, and the spatial state vectors of each adjacent base station device of the second base station device in the preset time period are used as the fourth key vector. By calculating the inner product of the fourth query vector and each fourth key vector, the fourth attention coefficient of multiple adjacent base stations is obtained. Therefore, the expression of the fourth attention coefficient can be written as:

[0098]

[0099] The fourth attention coefficient is then used to perform a weighted summation of the spatial state vectors of multiple adjacent base stations in the second time period, and the sum is added to the temporal state vector of the second base station in the second time period to obtain the spatiotemporal state vector of the second base station in the second time period. Therefore, the spatiotemporal state vector of the second base station in the second time period can be expressed as:

[0100]

[0101] Step 4: Output the results.

[0102] The spatiotemporal state vector is analyzed using the output network to obtain a second predicted alarm type of the second base station device at a next moment in the second time period.

[0103] It can be understood that the output network contains a learnable classification matrix, and each row of the classification matrix represents the space-time state vector corresponding to each type of alarm (such as standing wave alarm, RRU interruption alarm, BBU interruption alarm, power supply voltage alarm, etc.). Therefore, the output network obtains the correlation score between the space-time state vector of the second base station device in the second time period and the space-time state vector corresponding to each type of alarm by performing a dot product operation on the space-time state vector of the second base station device in the second time period and the classification matrix, which is recorded as Score i,a Represents the correlation score between the spatiotemporal state vector of the i-th second base station device in the second time period and the spatiotemporal state vector corresponding to the alarm type in a; the correlation score is then interpreted as a probability distribution through the Softmax function where p i It represents the probability distribution of the alarm type occurring in the i-th second base station device within the second time period, which reflects the relative likelihood of the occurrence of various alarm types in the second base station device at the next moment in the second time period; finally, the alarm type with the highest probability is output as the second predicted alarm type of the second base station device at the next moment in the second time period.

[0104] After obtaining each second prediction alarm type output by the neural network model through the above steps, the second prediction alarm type and the corresponding sample label can be used to construct a multi-classification cross entropy loss function, and the model parameters of the neural network model can be adjusted based on the multi-classification cross entropy loss function. The expression of the multi-classification cross entropy loss function can be written as:

[0105]

[0106] Among them, M represents the total number of multiple sets of training sample data, Q represents the number of alarm types, represents the second predicted alarm type q of the i-th second base station device, represents the probability of occurrence of the second predicted alarm type q, W q represents the weight of the second predicted alarm type q, and f q represents the frequency of occurrence of the second predicted alarm type q in multiple sets of training sample data, λ represents the L2 regularization coefficient, L2 regularization represents the L2 norm of the weight matrix. Its purpose is to prevent overfitting of the model. This is because large weight values ​​can make the model overly sensitive to noise in the training data, thus overcomplicating the model. By penalizing large weight values, L2 regularization helps the model learn smoother and more general decision boundaries, thereby improving the model's predictive performance on unseen data.

[0107] Therefore, in the base station equipment alarm prediction model obtained through the above training, the introduced multimodal cross-attention network allows different types of time series data (such as performance data, dynamic environment data, and alarm data) to pay attention to each other, which helps the base station equipment alarm prediction model understand which specific modal data is particularly important for alarm prediction in a specific period; and the introduced graph attention network can further analyze the spatial relationship of base station equipment, enabling the base station equipment alarm prediction model to identify key influencing factors in the spatial neighborhood. In addition, the introduced spatiotemporal cross-attention network achieves a deep fusion of time series features and spatial features in the field of base station equipment alarm warning. Compared with traditional alarm warning methods that often only focus on the analysis of time series data and ignore the spatial correlation between base station equipment, this fusion enables the model to capture the historical alarm patterns of a single base station equipment and understand the mutual influence between base station equipment, significantly enhancing the comprehensiveness and accuracy of the warning system.

[0108] It should be noted that after completing the model training, the hit rate can also be used to evaluate the performance of the model, where the denominator of the hit rate is the total number of base station devices in the data set, and the numerator is the number of base station devices whose predicted alarm type is the same as the actual alarm type.

[0109] In the technical solution provided in the above step S108, after the early warning system learns the first predicted alarm type of the first base station device at the next moment in the first time period, it can feedback the first early warning prompt information carrying at least the first predicted alarm type to the management terminal.

[0110] In addition, in traditional intelligent base station operation and maintenance, only the first warning prompt information carrying at least the first predicted alarm type is fed back to the management terminal. This means that maintenance personnel can only perform proactive maintenance based on the output first predicted alarm type, but lack maintenance direction, resulting in low maintenance efficiency. To this end, the embodiments of the present application analyze which features, which time steps, and which adjacent base stations have the greatest impact on the prediction results, thereby improving the interpretability of the prediction results and reducing the black box nature of deep learning methods.

[0111] As an optional implementation, in the technical solution provided in the above step S108, the early warning system can also first analyze the first attention coefficients of the dynamic environment data set and the performance data set of the first base station device at all time steps in the first time period based on the multimodal cross-attention network, and use the data set and time step corresponding to the largest first attention coefficient as the characteristic alarm root cause and the timing alarm root cause respectively, and analyze the fourth attention coefficients of each adjacent base station device of the first base station device based on the spatiotemporal cross-attention network, and use the adjacent base station device corresponding to the largest fourth attention coefficient as the spatial alarm root cause.

[0112] Specifically, the principle for determining the root cause of the alarm is as follows: Figure 4 As shown in the figure, the root cause of the characteristic alarm determined based on this principle can be written as:

[0113]

[0114] in, H represents the number of attention heads, N represents the total number of time steps in the first time period, Indicates the average influence of the feature vectors corresponding to the dynamic environment dataset / performance dataset on the feature vectors corresponding to the alarm dataset in the first time period. For example, if the feature vector x corresponding to the dynamic environment dataset in the first time period is the average attention coefficient of all attention heads H and all time steps N If it is higher, it means that the dynamic environment data set is generally important for alarm prediction in the first time period, and the characteristic alarm root cause of the first base station device in the first time period is the dynamic environment data set.

[0115] The root cause of a timing alarm can be expressed as:

[0116]

[0117] in, It represents the average influence of the feature vectors corresponding to the dynamic environment dataset / performance dataset at each time step in the first time period under all attention heads.

[0118] The expression of the spatial alarm root cause can be written as:

[0119] TOP=Max(α ij )

[0120] Next, first early warning prompt information carrying the first predicted alarm type, the characteristic alarm root cause, the temporal alarm root cause, and the spatial alarm root cause is fed back to the management terminal.

[0121] In addition, wireless networks are formed by the collaborative work of numerous base stations within a certain area, and there are complex relationships between them. Therefore, if only the alarm warning of a single base station device is considered, it is impossible to deeply analyze the potential problems at a deeper level of the network. To this end, the embodiments of this application propose a systematic and global perspective to prevent and respond to large-scale network failures. The specific implementation method is as follows:

[0122] Step 1: Obtain a first predicted alarm type of other base station devices in a target network including the first base station device at a next moment in a first time period.

[0123] Step 2: Match the first predicted alarm type of each base station device in the target network at the next moment in the first time period with the preset batch decommissioning alarm type set, and determine the number of base station devices whose first predicted alarm type is the batch decommissioning alarm type, where the batch decommissioning alarm type includes but is not limited to: remote control RF unit disconnection, input power disconnection, RF unit link abnormality, etc.

[0124] Step 3: The ratio of the number of base stations to the total number of base stations in the target network is used as the probability of batch decommissioning of base stations in the target network. This probability intuitively reflects the risk of batch failures that the target network may face.

[0125] Step 4: Determine the relationship between the probability of batch decommissioning of base station equipment and the preset batch decommissioning probability threshold. If the probability of batch decommissioning of base station equipment is not less than the batch decommissioning probability threshold, execute step 5; otherwise, execute step 6.

[0126] Step 5: Feedback a second warning message (i.e., high-risk warning message) indicating a high probability of mass decommissioning of base station equipment in the target network to the management terminal. The second warning message includes at least: an indicator light flashing at a first frequency and a buzzer sounding at a first frequency and a first volume. This is intended to quickly attract the attention of maintenance personnel and enable them to take emergency measures. For example, if the warning message indicates a high probability of RRU disconnection in the target network, maintenance personnel can first conduct a targeted inspection of the power generation and distribution facilities at the corresponding site.

[0127] Step 6: Feedback to the management terminal indicates a low probability of mass decommissioning of base station equipment in the target network (i.e., a low-risk warning). The third warning includes at least: a flashing indicator light at a second frequency and a buzzer sounding at a second frequency and a second volume. This serves to alert maintenance personnel that while immediate, high-intensity intervention is not necessary, they still need to monitor the relevant area and perform maintenance as appropriate.

[0128] The first frequency is higher than the second frequency, and the first volume is higher than the second volume.

[0129] In the above embodiment, by real-time monitoring and early warning of the risk of batch decommissioning of base station equipment in the target network, it is intended to help maintenance personnel respond to possible network failures more efficiently and accurately through early warning and refined early warning feedback mechanisms, thereby ensuring the stable operation and service quality of the network.

[0130] Example 2

[0131] According to an embodiment of the present application, a base station device alarm warning device for implementing the base station device alarm warning method in embodiment 1 is also provided. Figure 5 As shown, the base station equipment alarm warning device at least includes: an acquisition module 52, a feature fusion module 54, a prediction module 56 and an early warning module 58, wherein:

[0132] An acquisition module 52 is configured to acquire a first multidimensional data sequence and a first base station adjacency matrix of a first base station device within a first time period, wherein the first multidimensional data sequence includes a performance dataset, a dynamic environment dataset, and an alarm dataset of the first base station device at each time step within the first time period, and the first base station adjacency matrix is ​​configured to reflect the adjacency relationship between the first base station device and other base station devices within the first time period;

[0133] A feature fusion module 54 is configured to extract features from the first multidimensional data sequence and the first base station adjacency matrix using the multi-head attention network and the graph attention network of the base station equipment alarm prediction model to obtain a time series state vector and a spatial state vector, and to fuse features of the time series state vector and the spatial state vector using the spatiotemporal cross attention network of the base station equipment alarm prediction model to obtain a spatiotemporal state vector.

[0134] A prediction module 56 is configured to analyze the spatiotemporal state vector using an output network of the base station device alarm prediction model to determine a first predicted alarm type of the first base station device at a next moment in the first time period;

[0135] The early warning module 58 is configured to feed back first early warning prompt information carrying at least a first prediction alarm type to the management terminal.

[0136] The functions of each module of the base station equipment alarm warning device are described below in conjunction with a specific implementation process.

[0137] First, the acquisition module 52 can collect an initial multidimensional data sequence of the first base station device within a first time period from a professional network management system, wherein the initial multidimensional data sequence includes a performance data set, a dynamic environment data set, and an alarm data set of the first base station device at each time step within the first time period. The performance data set includes, but is not limited to: physical resource block (PRB) utilization, number of radio resource control (RRC) connections, wireless connectivity rate, handover success rate, PDCP (Packet Data Convergence Protocol) layer traffic, etc. The dynamic environment data set includes, but is not limited to: ambient temperature, ambient humidity, input voltage, standing wave ratio, optical attenuation, etc. The alarm data set includes, but is not limited to: alarm name, alarm type, alarm location, etc. In addition, the first time period referred to here is a predefined time window (such as one week, one month, etc.), and the time step is a fixed duration (such as 15 minutes, one hour, etc.).

[0138] At the same time, the acquisition module 52 can obtain the initial spatial topology data of the first base station device within the first time period, wherein the initial spatial topology data includes at least one of the following: the number of switches between the subordinate cells of the first base station device and the subordinate cells of the adjacent base station device, and whether resources are shared between the first base station device and the adjacent base station device.

[0139] To facilitate model input and analysis, the acquisition module 52 can also perform preprocessing operations on the acquired initial multidimensional data sequence and initial spatial topology data, including but not limited to the following operations: time alignment, time series data filling, outlier processing, standardization processing, adjacency matrix construction, etc.

[0140] Next, the feature fusion module 54 can call the pre-trained base station equipment alarm prediction model to extract the first multidimensional data sequence and the first base station adjacency matrix to obtain the corresponding space-time state vector; then the prediction module 56 can again call the output network of the base station equipment alarm prediction model to analyze the space-time state vector to determine the first predicted alarm type of the first base station equipment at the next moment in the first time period.

[0141] The training process of the base station equipment alarm prediction model includes:

[0142] Step S1: Acquire multiple sets of training sample data. Each set of training sample data includes: a second multidimensional data sequence and a second base station adjacency matrix of a second base station device in a second time period as training samples, and a second alarm type of the second base station device at a next moment in the second time period as a training label, wherein the second time period is a time period before the first time period.

[0143] Specifically, in the technical solution provided in step S1 above, the method for obtaining training sample data includes:

[0144] First, for each second base station device, an initial multidimensional data sequence and initial spatial topology data of the second base station device in multiple second time periods are obtained from a professional network management system, and a second alarm type of the second base station device at the next moment in each second time period is obtained. The initial multidimensional data sequence also includes: a performance data set, a dynamic environment data set, and an alarm data set of the second base station device at each time step in the second time period; the initial spatial topology data includes at least one of the following: the number of handovers between a subordinate cell of the second base station device and a subordinate cell of an adjacent base station device, and whether resources are shared between the second base station device and the adjacent base station device.

[0145] Next, preprocessing operations are performed on the multiple initial multidimensional data sequences and the multiple initial spatial topology data to obtain second multidimensional data sequences and second base station adjacency matrices for the second base station device in each second time period. The preprocessing operations include at least one of the following: time alignment (to ensure that time series data from different sources have consistent timestamps within the same time window), normalization (to eliminate dimensional differences between different data to adapt to changes in data distribution over time), outlier processing (to maintain data quality and model stability), and construction of an adjacency matrix (to quantify the spatial relationship between base station devices).

[0146] Finally, the second multidimensional data sequences and the second base station adjacency matrices of each second base station device in multiple second time periods are used as multiple training samples, and the second alarm types of each second base station device at the next moment in multiple second time periods are used as sample labels of corresponding training samples to form multiple groups of training sample data.

[0147] Through the above acquisition steps, rich information with both time series characteristics and spatial relationship characteristics can be extracted from historical data, forming multiple training samples and corresponding sample labels, providing a data foundation for subsequent iterative training of the model.

[0148] Step S2: construct a neural network model.

[0149] Specifically, the neural network model includes at least: a multimodal cross-attention network for fusing different types of time series data and evaluating the influence between modalities; a multi-head attention network for capturing the intrinsic dependencies of time series features; a graph attention network for encoding the spatial characteristics of base station equipment; a spatiotemporal cross-attention network for fusing temporal and spatial features; and an output network for converting the fused spatiotemporal features into predicted alarm types. Both the multimodal cross-attention network and the spatiotemporal cross-attention network are implemented using a Transform encoder and include positional encoding to ensure unique patterns across all time steps for each modality. The multi-head attention network uses multiple projection matrices to map inputs into different representation spaces, each representing a head. The outputs of multiple heads are concatenated, enabling parallel encoding of information across all time steps while also capturing the influencing factors of other modal features and the target modal features at each time step. It should be noted that the ultimate goal of the base station equipment alarm prediction model is to predict the alarm type. Therefore, the alarm features are defined as the target modal features, and other time series features are defined as other modal features.

[0150] Step S3: Iteratively train the neural network model using multiple sets of training sample data to obtain a base station equipment alarm prediction model.

[0151] In the technical solution provided in step S3 above, for each training batch in the iterative training process, each training sample in the training batch is input into the neural network model to obtain each second prediction alarm type output by the neural network model. Specifically, each training sample in the training batch is input into the neural network model, and the second prediction alarm type corresponding to each training sample output by the neural network model is obtained according to the following process:

[0152] Step 1: temporal coding.

[0153] Temporal encoding uses a multimodal cross-attention network to fuse the different modal features of the second base station device at each time step in the second time period to obtain the state vector of the second base station device at each time step in the second time period; then uses a multi-head attention network to capture the horizontal dependencies of the sequence and aggregate the embeddings of all time steps to obtain the final temporal state vector. The specific implementation method is as follows:

[0154] First, a multimodal cross-attention network is used to take the feature vector corresponding to the alarm dataset of the second base station in the training sample at each time step in the second time period as the first query vector, and the feature vectors corresponding to the dynamic environment dataset and performance dataset of the second base station in the training sample at each time step in the second time period as the first key vectors.

[0155] For each time step in the second time period, calculate the inner product of the first query vector corresponding to the time step (i.e., the feature vector corresponding to the alarm dataset at the current time step) and each first key vector corresponding to the time step (the feature vectors corresponding to the dynamic environment dataset and the performance dataset at the current time step), and obtain the first attention coefficient of the dynamic environment dataset and the performance dataset at the time step. The expression of the first attention coefficient can be written as:

[0156]

[0157] Among them, a represents the feature vector corresponding to the alarm dataset at the current time step, and x represents the feature vector corresponding to the dynamic environment dataset and performance dataset at the current time step. represents the first attention coefficient between the feature vector corresponding to the alarm dataset output by the h-th attention head at the n-th time step and the feature vector corresponding to the dynamic environment dataset / performance dataset. n represents the n-th time step in the second time period, and N represents the total number of time steps in the second time period. represents the unnormalized attention coefficient between the feature vector corresponding to the alarm dataset output by the h-th attention head at the n-th time step and the feature vector corresponding to the dynamic environment dataset / performance dataset, and POS n represents the position vector of the nth time step, a n Represents the feature vector corresponding to the alarm data set at the nth time step, represents the feature vector corresponding to the dynamic loop dataset and the performance dataset at the nth time step, H represents the number of attention heads, d represents the vector dimension, and W q,h represents the query matrix, W k,h Represents a matrix of values.

[0158] The first attention coefficient is used to perform a weighted summation of the eigenvectors corresponding to the dynamic environment dataset and the performance dataset at the time step, and the eigenvector corresponding to the alarm dataset at the time step, including the time step position vector, is added to obtain the state vector of the second base station device at the time step. Therefore, the state vector of the second base station device at the time step can be expressed as:

[0159]

[0160] Therefore, the state vector of the second base station device at the nth time step in the second time period is It uses the first attention coefficient All value vectors (i.e., the feature vectors corresponding to the dynamic ring dataset and the performance dataset at the nth time step) are weighted and then multi-head splicing and residual connection are performed to obtain the result.

[0161] Next, a multi-head attention network is used to take the state vector of the second base station device at any time step in the second time period as the second query vector, and the state vector of the second base station device at each other time step in the second time period as the second key vector. The inner product of the second query vector and each second key vector is calculated to obtain the second attention coefficient of each other time step.

[0162] Then, the state vectors at each other time step are weighted and summed using the second attention coefficient, and the state vector of the second base station device at the time step is added to obtain the weighted state vector of the second base station device at the time step.

[0163] Finally, the weighted state vectors of the second base station at each time step in the second time period are sequentially concatenated to obtain the time-series state vector of the second base station in the second time period. Therefore, the expression of the time-series state vector of the second base station in the second time period can be written as:

[0164]

[0165] Step 2: Spatial encoding.

[0166] Since the base station adjacency matrix of the second base station device in the second time period reveals the physical connection, environment sharing, and service dependency between base stations, the embodiment of the present application uses a graph attention network to encode the second base station adjacency matrix to obtain the spatial features of the second base station device in the second time period. The specific implementation method is as follows:

[0167] First, using the graph attention network, the feature vector corresponding to the base station spatial topology data of the second base station device in the second time period is used as the third query vector, and the feature vectors corresponding to the base station spatial topology data of each adjacent base station device in the second time period are used as third key vectors. The inner product of the third query vector and each third key vector is calculated to obtain the third attention coefficient of each adjacent base station device. Therefore, the expression of the third attention coefficient here can be written as:

[0168]

[0169] Among them, W s Represents a learnable parameter matrix, mainly used for linear transformation; LeakyReLU represents the activation function; represents the first-order neighbor set of the second base station device i; j and k both represent the numbers of the first-order neighbor base station devices of the second base station device i; a T represents the main attention parameter vector.

[0170] The third attention coefficient is used to perform a weighted summation of the eigenvectors corresponding to the base station spatial topology data of each adjacent base station device in the second time period to obtain the spatial state vector of the second base station device in the second time period. Therefore, the expression for the spatial state vector of the second base station device in the second time period can be written as:

[0171]

[0172] Where σ represents the activation function, v j Represents the spatial state vector corresponding to the base station adjacency matrix of the j-th base station device in the second time period.

[0173] Step 3: Space-time fusion.

[0174] In order to improve the accuracy of the model prediction results, the embodiment of the present application uses a spatiotemporal cross attention network to fuse the temporal state vector and spatial state vector of the second base station device in the second time period to obtain the spatiotemporal state vector of the second base station device in the second time period. The specific implementation method is as follows:

[0175] First, using the spatiotemporal cross attention network, the temporal state vector of the second base station device in the second time period is used as the fourth query vector, and the spatial state vectors of each adjacent base station device of the second base station device in the preset time period are used as the fourth key vector. By calculating the inner product of the fourth query vector and each fourth key vector, the fourth attention coefficient of multiple adjacent base stations is obtained. Therefore, the expression of the fourth attention coefficient can be written as:

[0176]

[0177] The fourth attention coefficient is then used to perform a weighted summation of the spatial state vectors of multiple adjacent base stations in the second time period, and the sum is added to the temporal state vector of the second base station in the second time period to obtain the spatiotemporal state vector of the second base station in the second time period. Therefore, the spatiotemporal state vector of the second base station in the second time period can be expressed as:

[0178]

[0179] Step 4: Output the results.

[0180] The spatiotemporal state vector is analyzed using the output network to obtain a second predicted alarm type of the second base station device at a next moment in the second time period.

[0181] It can be understood that the output network contains a learnable classification matrix, and each row of the classification matrix represents the space-time state vector corresponding to each type of alarm (such as standing wave alarm, RRU interruption alarm, BBU interruption alarm, power supply voltage alarm, etc.). Therefore, the output network obtains the correlation score between the space-time state vector of the second base station device in the second time period and the space-time state vector corresponding to each type of alarm by performing a dot product operation on the space-time state vector of the second base station device in the second time period and the classification matrix, which is recorded as The correlation score is then interpreted as a probability distribution through the Softmax function, which represents the relative probability of various alarm types occurring in the second base station device at the next moment in the second time period; finally, the alarm type with the highest probability is output as the second predicted alarm type of the second base station device at the next moment in the second time period.

[0182] After obtaining each second prediction alarm type output by the neural network model through the above steps, the second prediction alarm type and the corresponding sample label can be used to construct a multi-classification cross entropy loss function, and the model parameters of the neural network model can be adjusted based on the multi-classification cross entropy loss function. The expression of the multi-classification cross entropy loss function can be written as:

[0183]

[0184] Among them, M represents the total number of multiple sets of training sample data, Q represents the number of alarm types, represents the second predicted alarm type q of the i-th second base station device, represents the probability of occurrence of the second predicted alarm type q, W q represents the weight of the second predicted alarm type q, and f q represents the frequency of occurrence of the second predicted alarm type q in multiple sets of training sample data, λ represents the L2 regularization coefficient, L2 regularization represents the L2 norm of the weight matrix. Its purpose is to prevent overfitting of the model. This is because large weight values ​​can make the model overly sensitive to noise in the training data, thus overcomplicating the model. By penalizing large weight values, L2 regularization helps the model learn smoother and more general decision boundaries, thereby improving the model's predictive performance on unseen data.

[0185] Finally, the early warning module 58 may feed back the first early warning prompt information carrying at least the first prediction alarm type to the management terminal.

[0186] Furthermore, in traditional base station intelligent operation and maintenance, only the first warning information containing at least the first predicted alarm type is fed back to the management terminal. This means that maintenance personnel can only perform proactive maintenance based on the output first predicted alarm type, but lack maintenance direction, resulting in low maintenance efficiency. Therefore, the warning module 58 can also analyze the prediction results to determine which features, time steps, and adjacent base stations have the greatest impact on the prediction results, thereby improving the interpretability of the prediction results and reducing the black-box nature of deep learning methods.

[0187] As an optional implementation, the early warning module 58 can also first analyze the first attention coefficients of the dynamic environment data set and the performance data set of the first base station device at all time steps in the first time period based on the multimodal cross-attention network, and use the data set and time step corresponding to the largest first attention coefficient as the characteristic alarm root cause and the timing alarm root cause respectively, and analyze the fourth attention coefficients of each adjacent base station device of the first base station device based on the spatiotemporal cross-attention network, and use the adjacent base station device corresponding to the largest fourth attention coefficient as the spatial alarm root cause; and feed back the first early warning prompt information carrying the first predicted alarm type, characteristic alarm root cause, timing alarm root cause and spatial alarm root cause to the management terminal.

[0188] In addition, wireless networks are formed by the coordinated operation of numerous base stations within a certain area, and there are complex relationships between them. Therefore, if only the alarm warning of a single base station device is considered, it is impossible to deeply analyze the potential problems at a deeper level of the network. To this end, the early warning module 58 can also provide early warning of large-scale network failures through the following steps:

[0189] Step 1: Obtain a first predicted alarm type of other base station devices in a target network including the first base station device at a next moment in a first time period.

[0190] Step 2: Match the first predicted alarm type of each base station device in the target network at the next moment in the first time period with the preset batch decommissioning alarm type set, and determine the number of base station devices whose first predicted alarm type is the batch decommissioning alarm type, where the batch decommissioning alarm type includes but is not limited to: remote control RF unit disconnection, input power disconnection, RF unit link abnormality, etc.

[0191] Step 3: The ratio of the number of base stations to the total number of base stations in the target network is used as the probability of batch decommissioning of base stations in the target network. This probability intuitively reflects the risk of batch failures that the target network may face.

[0192] Step 4: Determine the relationship between the probability of batch decommissioning of base station equipment and the preset batch decommissioning probability threshold. If the probability of batch decommissioning of base station equipment is not less than the batch decommissioning probability threshold, execute step 5; otherwise, execute step 6.

[0193] Step 5: Feedback a second warning message (i.e., high-risk warning message) indicating a high probability of mass decommissioning of base station equipment in the target network to the management terminal. The second warning message includes at least: an indicator light flashing at a first frequency and a buzzer sounding at a first frequency and a first volume. This is intended to quickly attract the attention of maintenance personnel and enable them to take emergency measures. For example, if the warning message indicates a high probability of RRU disconnection in the target network, maintenance personnel can first conduct a targeted inspection of the power generation and distribution facilities at the corresponding site.

[0194] Step 6: Feedback to the management terminal indicates a low probability of mass decommissioning of base station equipment in the target network (i.e., a low-risk warning). The third warning includes at least: a flashing indicator light at a second frequency and a buzzer sounding at a second frequency and a second volume. This serves to alert maintenance personnel that while immediate, high-intensity intervention is not necessary, they still need to monitor the relevant area and perform maintenance as appropriate.

[0195] The first frequency is higher than the second frequency, and the first volume is higher than the second volume.

[0196] It should be noted that each module in the base station equipment alarm warning device in the embodiment of the present application corresponds one-to-one to each implementation step of the base station equipment alarm warning method in Example 1. Since a detailed description has been given in Example 1, some details not reflected in this embodiment can be referred to Example 1 and will not be elaborated here.

[0197] Example 3

[0198] According to an embodiment of the present application, a computer program product is also provided, which includes a computer program, wherein when the computer program is executed by a processor, the base station device alarm warning method in Example 1 is implemented.

[0199] According to an embodiment of the present application, a non-volatile storage medium is also provided, which includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the base station device alarm warning method in Example 1 by running the computer program.

[0200] According to an embodiment of the present application, a processor is further provided, which is used to run a computer program, wherein the base station device alarm warning method in Example 1 is executed when the computer program is running.

[0201] According to an embodiment of the present application, an electronic device is also provided, which includes: a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to execute the base station device alarm warning method in Example 1 through the computer program.

[0202] Specifically, when the computer program is running, the following steps are executed: obtaining a first multidimensional data sequence and a first base station adjacency matrix of the first base station device in a first time period, wherein the first multidimensional data sequence includes a performance data set, a dynamic environment data set and an alarm data set of the first base station device at each time step in the first time period, and the first base station adjacency matrix is ​​used to reflect the adjacency relationship between the first base station device and other base station devices in the first time period; using the multi-head attention network and the graph attention network of the base station device alarm prediction model to extract features of the first multidimensional data sequence and the first base station adjacency matrix to obtain a time series state vector and a space state vector, and using the time-space cross attention network of the base station device alarm prediction model to perform feature fusion on the time series state vector and the space state vector to obtain a time-space state vector; using the output network of the base station device alarm prediction model to analyze the time-space state vector to determine the first predicted alarm type of the first base station device at the next moment in the first time period; and feeding back to the management terminal a first early warning prompt information carrying at least the first predicted alarm type.

[0203] As an optional implementation, the electronic device may be in the form of a mobile terminal, a computer terminal or a similar computing device. Figure 6 The figure shows a hardware structure block diagram of an electronic device for implementing a base station device alarm warning method. Figure 6 As shown, the electronic device 60 may include one or more (illustrated by 602a, 602b, ..., 602n in the figure) processors 602 (the processor 602 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 604 for storing data, and a transmission device 606 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. It will be understood by those skilled in the art that Figure 6 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 6 More or fewer components than shown, or with Figure 6 Different configurations shown.

[0204] It should be noted that the one or more processors 602 and / or other data processing circuits described above may generally be referred to herein as "data processing circuitry". The data processing circuitry may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuitry may be a single independent processing module, or may be incorporated in whole or in part into any of the other components of the electronic device 60. As described in the embodiments of the present application, the data processing circuitry serves as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0205] The memory 604 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the base station equipment alarm warning method in the embodiment of the present application. The processor 602 executes various functional applications and data processing by running the software programs and modules stored in the memory 604, that is, implementing the vulnerability detection method of the above-mentioned application. The memory 604 may include a high-speed random access memory and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 604 may further include a memory remotely located relative to the processor 602, and these remote memories may be connected to the electronic device 60 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0206] The transmission device 606 is used to receive or send data via a network. Specific examples of the aforementioned network may include a wireless network provided by the communications provider of the electronic device 60. In one embodiment, the transmission device 606 includes a network interface controller (NIC), which can be connected to other network devices via a base station device to enable communication with the Internet. In one embodiment, the transmission device 606 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0207] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the electronic device 60 .

[0208] The serial numbers of the above embodiments are for description only and do not represent the advantages or disadvantages of the embodiments.

[0209] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0210] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only exemplary. For example, the division of units can be a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0211] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple units. Some or all of the units may be selected to achieve the purpose of the present embodiment according to actual needs.

[0212] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0213] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program code.

[0214] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A base station equipment alarm warning method, characterized in that: include: Obtaining a first multidimensional data sequence and a first base station adjacency matrix of a first base station device within a first time period, wherein the first multidimensional data sequence includes a performance dataset, a dynamic environment dataset, and an alarm dataset of the first base station device at each time step within the first time period, and the first base station adjacency matrix is ​​used to reflect an adjacency relationship between the first base station device and other base station devices within the first time period; Performing feature extraction on the first multidimensional data sequence and the first base station adjacency matrix using a multi-head attention network and a graph attention network of the base station equipment alarm prediction model to obtain a time series state vector and a spatial state vector, and performing feature fusion on the time series state vector and the spatial state vector using a spatiotemporal cross attention network of the base station equipment alarm prediction model to obtain a spatiotemporal state vector; Analyzing the spatiotemporal state vector using an output network of the base station device alarm prediction model to determine a first predicted alarm type of the first base station device at a next moment in a first time period; Feedback is given to the management terminal of the first early warning prompt information at least carrying the first prediction alarm type.

2. The method according to claim 1, characterized in that The training process of the base station equipment alarm prediction model includes: Acquire multiple sets of training sample data, wherein each set of the training sample data includes: a second multidimensional data sequence and a second base station adjacency matrix of a second base station device in a second time period as training samples, a second alarm type of the second base station device at a next moment in the second time period as a training label, and the second time period is a time period before the first time period; Constructing a neural network model, wherein the neural network model at least includes: a multimodal cross attention network, the multi-head attention network, the graph attention network, the spatiotemporal cross attention network, and an output network; The neural network model is iteratively trained using the multiple sets of training sample data to obtain the base station equipment alarm prediction model.

3. The method according to claim 2, characterized in that Obtain multiple sets of training sample data, including: For each second base station device, respectively obtain an initial multidimensional data sequence and initial spatial topology data of the second base station device in multiple second time periods, and obtain a second alarm type of the second base station device at the next moment of each second time period, wherein the initial multidimensional data sequence includes: a performance data set, a dynamic loop data set, and an alarm data set of the second base station device at each time step in the second time period, the performance data set including at least one of the following: physical resource block utilization, number of radio resource control connections, wireless connectivity rate, and switching success rate, the dynamic loop data set including at least one of the following: ambient temperature, ambient humidity, input voltage, standing wave ratio, and optical attenuation, the alarm data set including at least: alarm name, alarm type, and alarm location, and the initial spatial topology data including at least one of the following: the number of handovers between a subordinate cell of the second base station device and a subordinate cell of an adjacent base station device, and whether resources are shared between the second base station device and the adjacent base station device; Performing preprocessing operations on the plurality of the initial multidimensional data sequences and the plurality of the initial spatial topology data respectively to obtain a second multidimensional data sequence and a second base station adjacency matrix of the second base station device in each second time period, wherein the preprocessing operations include at least one of the following: time alignment, normalization processing, outlier processing, and constructing an adjacency matrix; The multiple groups of training sample data are composed of the second multidimensional data sequences and the second base station adjacency matrix of each second base station device in multiple second time periods as multiple training samples, and the second alarm type of each second base station device at the next moment in multiple second time periods as the sample label of the corresponding training sample.

4. The method according to claim 2, characterized in that Iteratively training the neural network model using the multiple sets of training sample data to obtain the base station equipment alarm prediction model includes: For each training batch in the iterative training process, each training sample of the training batch is input into the neural network model to obtain each second prediction alarm type output by the neural network model, and a multi-classification cross entropy loss function is constructed using the second prediction alarm type and the corresponding sample label, and the model parameters of the neural network model are adjusted according to the multi-classification cross entropy loss function.

5. The method according to claim 4, characterized in that Inputting each training sample of the training batch into the neural network model to obtain each second prediction alarm type output by the neural network model includes: Input each training sample of the training batch into the neural network model, and obtain the second prediction alarm type corresponding to each training sample output by the neural network model according to the following process: Using the multimodal cross-attention network, the feature vector corresponding to the alarm data set of the second base station in the training sample at each time step in the second time period is used as the first query vector, and the feature vectors corresponding to the dynamic environment data set and the performance data set of the second base station in the training sample at each time step in the second time period are respectively used as the first key vectors; for each time step in the second time period, the inner product of the first query vector corresponding to the time step and the first key vectors corresponding to the time step is calculated to obtain the first attention coefficient of the dynamic environment data set and the performance data set at the time step, and the first attention coefficient is used to perform weighted summation on the feature vectors corresponding to the dynamic environment data set and the performance data set at the time step, and the feature vector corresponding to the alarm data set at the time step, which includes the time step position vector, is added to obtain the state vector of the second base station device at the time step; Using the multi-head attention network, the state vector of the second base station device at any time step in the second time period is used as the second query vector, and the state vector of the second base station device at each other time step in the second time period is used as the second key vector, and the inner product of the second query vector and each second key vector is calculated to obtain the second attention coefficient of each other time step; using the second attention coefficient, weighted summation is performed on the state vectors at each other time step, and the state vector of the second base station device at the time step is added to obtain the weighted state vector of the second base station device at the time step; the weighted state vectors of the second base station device at each time step in the second time period are sequentially spliced ​​to obtain the time series state vector of the second base station device in the second time period; Using a graph attention network, the feature vector corresponding to the base station spatial topology data of the second base station device in the second time period is used as a third query vector, and the feature vectors corresponding to the base station spatial topology data of each adjacent base station device of the second base station device in the second time period are respectively used as third key vectors, and the inner product of the third query vector and each third key vector is calculated to obtain a third attention coefficient of each adjacent base station device; using the third attention coefficient, a weighted sum is performed on the feature vectors corresponding to the base station spatial topology data of each adjacent base station device in the second time period to obtain a spatial state vector of the second base station device in the second time period; Using the spatiotemporal cross attention network, the temporal state vector of the second base station device in the second time period is used as a fourth query vector, and the spatial state vectors of each adjacent base station device of the second base station device in a preset time period are used as a fourth key vector. By calculating the inner product of the fourth query vector and each of the fourth key vectors, a fourth attention coefficient of multiple adjacent base station devices is obtained, and the fourth attention coefficient is used to perform weighted summation on the spatial state vectors of the multiple adjacent base station devices in the second time period, and the sum is added to the temporal state vector of the second base station device in the second time period to obtain the spatiotemporal state vector of the second base station device in the second time period; The output network is used to analyze the spatiotemporal state vector to obtain a second predicted alarm type of the second base station device at a next moment in a second time period.

6. The method according to claim 2, characterized in that Feedback to the management terminal of first warning prompt information carrying at least the first prediction alarm type includes: Analyzing the first attention coefficients of the dynamic environment dataset and the performance dataset of the first base station device at all time steps in the first time period according to the multimodal cross-attention network, and taking the dataset and time step corresponding to the largest first attention coefficient as the feature alarm root cause and the timing alarm root cause, respectively; and analyzing the fourth attention coefficients of each adjacent base station device of the first base station device according to the spatiotemporal cross-attention network, and taking the adjacent base station device corresponding to the largest fourth attention coefficient as the spatial alarm root cause; Feedback is given to the management terminal of first warning prompt information carrying the first predicted alarm type, the characteristic alarm root cause, the temporal alarm root cause, and the spatial alarm root cause.

7. The method according to claim 1, characterized in that After determining, based on the spatiotemporal state vector, a first predicted alarm type of the first base station device at a next moment in the first time period, the method further includes: Obtaining a first predicted alarm type of other base station devices in a target network including the first base station device at a next moment in a first time period; Matching a first predicted alarm type of each base station device in the target network at a next moment in the first time period with a preset batch out-of-service alarm type set, and determining the number of base station devices for which the first predicted alarm type is a batch out-of-service alarm type, wherein the batch out-of-service alarm type includes at least one of the following: remote control radio frequency unit link disconnection, input power disconnection, and radio frequency unit link abnormality; The ratio of the number of base station devices to the total number of base station devices in the target network is used as the probability of batch decommissioning of base station devices in the target network; If the probability of batch decommissioning of the base station devices is not lower than a preset batch decommissioning probability threshold, feeding back to the management terminal second warning prompt information indicating that the target network has a high probability of batch decommissioning of base station devices, wherein the second warning prompt information is in the form of at least: an indicator light flashing at a first frequency, and a buzzer emitting a prompt tone at a first frequency and a first volume; When the probability of batch decommissioning of base station devices is lower than the batch decommissioning probability threshold, feeding back to the management terminal third warning prompt information indicating that the target network has a low probability of batch decommissioning of base station devices, wherein the third warning prompt information is in the form of at least: an indicator light flashing at a second frequency, and a buzzer emitting a prompt tone at a second frequency and a second volume; The first frequency is higher than the second frequency, and the first volume is higher than the second volume.

8. A base station equipment alarm warning device, characterized in that: include: an acquisition module, configured to acquire a first multidimensional data sequence and a first base station adjacency matrix of a first base station device within a first time period, wherein the first multidimensional data sequence includes a performance dataset, a dynamic environment dataset, and an alarm dataset of the first base station device at each time step within the first time period, and the first base station adjacency matrix is ​​used to reflect the adjacency relationship between the first base station device and other base station devices within the first time period; a feature fusion module, configured to extract features from the first multidimensional data sequence and the first base station adjacency matrix using a multi-head attention network and a graph attention network of the base station equipment alarm prediction model to obtain a time series state vector and a spatial state vector, and to fuse features of the time series state vector and the spatial state vector using a spatiotemporal cross attention network of the base station equipment alarm prediction model to obtain a spatiotemporal state vector; a prediction module, configured to analyze the spatiotemporal state vector using an output network of the base station device alarm prediction model to determine a first predicted alarm type of the first base station device at a next moment in a first time period; The early warning module is used to feed back first early warning prompt information carrying at least the first prediction alarm type to the management terminal.

9. A computer program product, characterized in that include: A computer program, wherein when the computer program is executed by a processor, the base station device alarm warning method according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: include: A memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the base station device alarm warning method according to any one of claims 1 to 7 through the computer program.

Images