Base station out-of-service alarm prediction method, device, equipment and storage medium

Abstract

The present invention discloses a method, device, equipment and storage medium for predicting a base station out-of-service alarm. The method comprises: performing word vector encoding on the alarm data in the historical operation and maintenance data of the base station to be predicted within a preset time period to obtain an alarm feature sequence; performing sliding sampling and feature addition on the alarm feature sequence through a preset sliding window to obtain an alarm prediction sample; and performing base station out-of-service alarm prediction based on the alarm prediction sample through a preset out-of-service alarm prediction model. Since the alarm prediction sample is generated by sampling and feature addition on the alarm feature sequence generated based on the historical alarm data of the base station to be predicted, and then the base station out-of-service alarm prediction is performed based on the alarm prediction sample using a pre-trained out-of-service alarm prediction model, compared to the existing method of predicting a base station out-of-service based on the experience of technical personnel, the above method of the present invention effectively utilizes the historical alarm data of the base station to be predicted, thereby ensuring the accuracy and reliability of the prediction result.

Description

Technical Field

[0001] The present invention relates to the field of network communication technology, and in particular to a base station out-of-service alarm prediction method, device, equipment and storage medium. Background Art

[0002] Communication base stations are the most critical nodes in mobile communication systems. Once a communication base station is out of service, the signal within the coverage area of ​​the base station antenna will be interrupted or weakened, and wireless services will not be provided to the signal coverage area, seriously affecting the user's perception experience.

[0003] Currently, base station outages are handled passively after an outage alarm has occurred and the base station has ceased service, relying primarily on manual experience and internet-based technologies. When a base station ceases service due to antenna failure, transmission failure, carrier frequency failure, power outage, or high temperature, the device generates an outage alarm and reports it to the network management system, which then synchronizes it to the centralized fault management platform. Operations and maintenance personnel, drawing on the equipment maintenance manual and past experience, develop an operation and maintenance plan. These personnel perform remote device reboots, parameter reconfigurations, and on-site device and battery replacements to troubleshoot and restore the base station to service.

[0004] With the rapid iteration of wireless communication technologies and the continuous expansion of network scale, users' demands for network quality are becoming increasingly stringent. Passive response methods for handling base station alarms are increasingly unable to meet operators' management needs for network alarms, especially outage alarms. Predicting base station outages based solely on technical experience is often crude and simplistic, failing to achieve efficient and accurate predictions.

[0005] The above content is only used to assist in understanding the technical solution of the present invention and does not constitute an admission that the above content is prior art. Summary of the Invention

[0006] The main purpose of the present invention is to provide a base station out-of-service alarm prediction method, device, equipment and storage medium, aiming to solve the technical problem that most of the existing out-of-service alarm predictions rely on experience for coarse-grained estimation and the prediction accuracy is not high.

[0007] To achieve the above object, the present invention provides a base station out-of-service alarm prediction method, the method comprising the following steps:

[0008] Obtain historical operation and maintenance data of the base station to be predicted within a preset time period;

[0009] Perform word vector encoding on the alarm data in the historical operation and maintenance data to obtain an alarm feature sequence;

[0010] Sliding sampling is performed on the alarm feature sequence through a preset sliding window to obtain an initial alarm sample;

[0011] Adding features to the initial alarm prediction sample according to the historical operation and maintenance data to obtain an alarm prediction sample;

[0012] A base station out-of-service alarm prediction is performed based on the alarm prediction samples using a preset out-of-service alarm prediction model.

[0013] Preferably, the step of performing word vector encoding on the alarm data in the historical operation and maintenance data to obtain an alarm feature sequence includes:

[0014] Determining the alarm identifier of the base station to be predicted according to the alarm data in the historical operation and maintenance data;

[0015] Mapping the warning identifier to a corresponding warning word vector through a preset word2vec model;

[0016] Encode the warning word vector to obtain a warning feature sequence.

[0017] Preferably, the warning word vector includes a service termination warning word vector and a non-service termination warning word vector;

[0018] The step of encoding the warning word vector to obtain a warning feature sequence includes:

[0019] Obtaining the word vector similarity between the service withdrawal warning word vector and the non-service withdrawal warning word vector;

[0020] Calculating the warning code value of the base station to be predicted according to the word vector similarity;

[0021] An alarm feature sequence is generated according to the alarm code value.

[0022] Preferably, the step of generating an alarm feature sequence according to the alarm code value includes:

[0023] According to the alarm code value, a sum of the alarm code values ​​of the base station to be predicted within a unit time period is counted;

[0024] The sum of the alarm code values ​​is used as the alarm characteristic value corresponding to each unit time period;

[0025] An alarm feature sequence of the base station to be predicted within the preset time period is constructed according to the alarm feature value.

[0026] Preferably, the step of adding features to the initial alarm prediction sample according to the historical operation and maintenance data to obtain the alarm prediction sample includes:

[0027] Analyzing the historical operation and maintenance data to obtain historical service withdrawal characteristics and base station attribute characteristics corresponding to the base station to be predicted;

[0028] Features are added to the initial alarm prediction sample according to the historical service withdrawal features and the base station attribute features to obtain an alarm prediction sample.

[0029] Preferably, the preset out-of-service alarm prediction model includes: a preset XGBoost model and a preset Lightgbm model;

[0030] The step of performing base station out-of-service alarm prediction according to the alarm prediction sample by using a preset out-of-service alarm prediction model includes:

[0031] Grouping the alarm prediction samples to obtain a preset number of alarm prediction subsample sets;

[0032] Inputting the alarm prediction subsample set into the preset XGBoost model and the preset Lightgbm model respectively;

[0033] A weighted sum is performed on the first prediction result output by the preset XGBoost model and the second prediction result output by the preset Lightgbm model to obtain a base station out-of-service alarm prediction result.

[0034] Preferably, before the step of obtaining historical operation and maintenance data of the base station to be predicted within a preset time period, the method further includes:

[0035] Obtain training samples, and perform GroupKFold grouping on the training samples to obtain multiple subsample sets;

[0036] Selecting a subsample set from the plurality of subsample sets as a model test set, and using the subsample sets other than the model test set as a model training set;

[0037] The initial XGBoost model and the initial Lightgbm model are trained respectively according to the model training set to obtain a preset XGBoost model and a preset Lightgbm model.

[0038] Preferably, the binary cross entropy loss function in the initial XGBoost model and the initial Lightgbm model is:

[0039]

[0040] In the formula, Focal loss is the binary cross entropy loss function, P i is the prediction probability, α and γ are weighting factors.

[0041] Preferably, before the step of obtaining historical operation and maintenance data of the base station to be predicted within a preset time period, the method further includes:

[0042] Obtain historical operation and maintenance data of different base stations;

[0043] Building a base station alarm code corpus based on the historical operation and maintenance data of the different base stations;

[0044] An initial word2vec model is iteratively trained based on the base station alarm code corpus to obtain a preset word2vec model.

[0045] Preferably, the step of constructing a base station alarm code corpus based on the historical operation and maintenance data of the different base stations includes:

[0046] Obtaining original alarm data of each base station from the historical operation and maintenance data of the different base stations;

[0047] Preprocessing the original alarm data to obtain valid alarm data;

[0048] Aggregating and sorting the valid alarm data, and generating an alarm sequence based on the aggregated and sorted alarm data;

[0049] A base station alarm code corpus is constructed according to the alarm sequence.

[0050] Preferably, the step of preprocessing the original alarm data to obtain valid alarm data includes:

[0051] Deduplication of the original alarm data is performed according to the alarm fields of the preset dimensions to obtain the alarm data to be processed;

[0052] Abnormal alarm data in the alarm data to be processed is eliminated to obtain valid alarm data.

[0053] Preferably, the step of aggregating and sorting the valid alarm data and generating an alarm sequence according to the aggregated and sorted alarm data includes:

[0054] Aggregating the valid alarm data into different alarm data groups according to base station identifiers;

[0055] Sorting the alarm data in each of the alarm data groups in the order of alarm occurrence time to obtain a sorted alarm data group;

[0056] An alarm sequence is generated according to the sorted alarm data group.

[0057] Preferably, the step of generating an alarm sequence according to the sorted alarm data group includes:

[0058] Taking each out-of-service alarm in the sorted alarm data group as the center, the alarm data within a preset time range is used as the alarm group corresponding to the out-of-service alarm;

[0059] Combine the alarm groups corresponding to all out-of-service alarms to obtain an alarm sequence.

[0060] In addition, to achieve the above-mentioned purpose, the present invention further proposes a base station out-of-service alarm prediction device, the base station out-of-service alarm prediction device comprising:

[0061] A data acquisition module is used to obtain historical operation and maintenance data of the base station to be predicted within a preset time period;

[0062] A vector encoding module, configured to perform word vector encoding on the alarm data in the historical operation and maintenance data to obtain an alarm feature sequence;

[0063] A window sampling module is used to perform sliding sampling on the alarm feature sequence through a preset sliding window to obtain an initial alarm sample;

[0064] A feature adding module, configured to add features to the initial alarm prediction sample according to the historical operation and maintenance data to obtain an alarm prediction sample;

[0065] The alarm prediction module is used to predict the base station out-of-service alarm according to the alarm prediction samples using a preset out-of-service alarm prediction model.

[0066] In addition, to achieve the above-mentioned purpose, the present invention also proposes a base station out-of-service alarm prediction device, which includes: a memory, a processor, and a base station out-of-service alarm prediction program stored on the memory and runnable on the processor, and the base station out-of-service alarm prediction program is configured to implement the steps of the base station out-of-service alarm prediction method as described above.

[0067] In addition, to achieve the above-mentioned purpose, the present invention also proposes a storage medium, on which a base station out-of-service alarm prediction program is stored. When the base station out-of-service alarm prediction program is executed by a processor, the steps of the base station out-of-service alarm prediction method as described above are implemented.

[0068] The present invention obtains the historical operation and maintenance data of the base station to be predicted within a preset time period, performs word vector encoding based on the alarm data in the historical operation and maintenance data to obtain an alarm feature sequence; then performs sliding sampling on the alarm feature sequence through a preset sliding window to obtain an initial alarm sample; then adds features to the initial alarm prediction sample based on the historical operation and maintenance data to obtain an alarm prediction sample; finally, performs base station decommissioning alarm prediction based on the alarm prediction sample through a preset decommissioning alarm prediction model. Since the alarm prediction sample is generated by sampling and adding features to the alarm feature sequence generated based on the historical alarm data of the base station to be predicted, and then the pre-trained decommissioning alarm prediction model is used to perform base station decommissioning alarm prediction based on the generated alarm prediction sample, compared to the existing method of predicting base station decommissioning based on the experience of technical personnel, the above method of the present invention effectively utilizes the historical alarm data of the base station to be predicted, ensuring the accuracy and reliability of the prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a structural diagram of a base station out-of-service alarm prediction device in a hardware operating environment involved in an embodiment of the present invention;

[0070] Figure 2 This is a flow chart of a first embodiment of a base station out-of-service alarm prediction method according to the present invention;

[0071] Figure 3 This is a schematic diagram of sliding window sampling in the first embodiment of the base station out-of-service alarm prediction method of the present invention;

[0072] Figure 4 This is a flow chart of a second embodiment of a base station out-of-service alarm prediction method according to the present invention;

[0073] Figure 5 This is a schematic diagram of the model structure of the word2vec model in the second embodiment of the base station out-of-service alarm prediction method of the present invention;

[0074] Figure 6 This is a flow chart of a third embodiment of a base station out-of-service alarm prediction method according to the present invention;

[0075] Figure 7 This is a flow chart of a fourth embodiment of a base station out-of-service alarm prediction method according to the present invention;

[0076] Figure 8 This is a schematic diagram of the base station out-of-service alarm prediction process based on the XGBoost model and Lightgbm model;

[0077] Figure 9 This is a flow chart of a fifth embodiment of a base station out-of-service alarm prediction method according to the present invention;

[0078] Figure 10This is a structural block diagram of the first embodiment of the base station out-of-service alarm prediction device of the present invention.

[0079] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0080] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0081] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a base station out-of-service alarm prediction device in the hardware operating environment involved in the embodiment of the present invention.

[0082] like Figure 1 As shown, the base station out-of-service alarm prediction device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the user interface 1003 may optionally include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (WIreless-FIdelity, WI-FI) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM) memory, or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk memory. The memory 1005 may optionally also be a storage device independent of the aforementioned processor 1001.

[0083] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation on the base station out-of-service alarm prediction device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0084] like Figure 1 As shown, the memory 1005 as a storage medium may include an operating system, a data storage module, a network communication module, a user interface module, and a base station out-of-service alarm prediction program.

[0085] exist Figure 1In the base station out-of-service alarm prediction device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the base station out-of-service alarm prediction device of the present invention can be set in the base station out-of-service alarm prediction device, and the base station out-of-service alarm prediction device calls the base station out-of-service alarm prediction program stored in the memory 1005 through the processor 1001, and executes the base station out-of-service alarm prediction method provided by the embodiment of the present invention.

[0086] The embodiment of the present invention provides a base station out-of-service alarm prediction method, referring to Figure 2 , Figure 2 This is a flow chart of the first embodiment of the base station out-of-service alarm prediction method of the present invention.

[0087] In this embodiment, the base station out-of-service alarm prediction method includes the following steps:

[0088] Step S10: Obtain historical operation and maintenance data of the base station to be predicted within a preset time period.

[0089] It should be noted that the execution subject of the method of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a smart phone, tablet computer and personal computer, or other electronic devices that can achieve the same or similar functions. This embodiment does not limit this.

[0090] It should be understood that the preset time period can be determined according to actual forecast needs, such as nearly one month (30 days), three months (90 days) or half a year (180 days), etc., and this embodiment does not limit it. The historical operation and maintenance data in this embodiment may include the working parameter data and alarm data of the base station. Among them, the alarm data may include various types of alarm records such as fault alarms, dynamic environment alarms, and performance alarms. These alarm records describe the time when the base station equipment alarm occurs, the alarm category, the base station name and other information, which can be used to construct a training sample set for the decommissioning alarm prediction model. The working parameter data may be data that describes the physical properties of the base station, including the latitude and longitude of the base station location, the covered area, the computer room to which it belongs, the base station name, the base station ID and other information.

[0091] In a specific implementation, the historical operation and maintenance data of the base station to be predicted within a preset time period can be first obtained from the network management system to which the base station to be predicted belongs.

[0092] Step S20: performing word vector encoding based on the alarm data in the historical operation and maintenance data to obtain an alarm feature sequence.

[0093] It should be emphasized that in order to achieve accurate prediction of base station decommissioning alarms, this embodiment preferably uses the performance alarms, environmental alarms, wireless alarms and other alarm data contained in the historical operation and maintenance data of the base station to be predicted, or the data after a series of processing based on these alarm data as the basis for prediction to predict the possibility of decommissioning alarms in the future for the base station to be predicted.

[0094] It should be noted that the above-mentioned alarm feature sequence can be a data object composed of alarm features, arranged in the order of the alarm occurrence time corresponding to the alarm features. The alarm features can be represented by specific numerical values ​​or symbols. For example, alarm feature 1 (10:00-11:00) - alarm feature 2 (11:00-12:00) - alarm feature 3 (12:00-13:00), etc. This embodiment does not impose any specific restrictions on the representation of the alarm features.

[0095] In actual applications, technicians have found the following characteristics of base station alarms after analyzing the (historical) alarm data of base stations: 1. Most out-of-service alarms are sudden alarms, that is, they have low correlation with previous alarms, which makes direct prediction more difficult; 2. There are many types of alarms, and there is no obvious pattern in the time when the alarms occur and the base stations where they occur; 3. The number of out-of-service alarms is small and the proportion is low: Taking a certain province and city as an example, a total of 130,316 alarms occurred from July to August of that year, of which 5,854 were out-of-service alarms, accounting for only 4.5%, and there is a serious data sparsity problem. In combination with the above actual situation, this embodiment preferably adopts an alarm analysis method based on word vectors, that is, word vector encoding of alarm data, which can solve the problem of sparse out-of-service alarm data on the one hand, and improve the efficiency and accuracy of out-of-service alarm prediction on the other hand.

[0096] It should be understood that word vectors are a mathematical representation of words in a language. They possess excellent semantic properties and are a commonly used method for characterizing words. The basic idea behind word vectors is to map each word in a language into a vector of fixed length, typically much smaller than the size of the language's lexicon, typically ranging from tens to hundreds of dimensions, through training on a large corpus. These vectors together form a word vector space, and each vector can be considered a point in this space. By introducing a "distance" metric within this space, the syntactic and semantic similarity between corresponding words can be determined based on the distance between their word vectors.

[0097] Applied to this embodiment, the word vector encoding can be to first vectorize the data to be processed to generate a word vector space, and then vector encode the word vectors in the space based on the word vector space, thereby realizing the operation of feature extraction and mathematical expression of the features of the data to be processed.

[0098] As an implementation method, encoding based on the word vector space in this embodiment can be to first calculate the similarity between the service withdrawal alarm word vector and the non-service withdrawal alarm word vector in the word vector space, then calculate the encoding value based on the calculated similarity, and then determine the alarm feature based on the encoding value, and finally sort the alarm features to obtain an alarm feature sequence. Of course, the method of encoding the word vector in this embodiment is not limited to the above method. All methods that can extract and quantify the features of different types of alarms (such as service withdrawal alarms or non-service withdrawal alarms) in the alarm data can be applied to this embodiment to realize the generation of the alarm feature sequence.

[0099] In the specific implementation, the alarm data in the historical operation and maintenance data of the base station to be predicted can be extracted first, and then the alarm identifiers (such as the name or title of the alarm) of all alarms that occurred in the base station in the past period of time can be determined based on these alarm data. Then, word vector encoding is performed based on these alarm identifiers, and then the alarm feature sequence is obtained based on the encoding results.

[0100] Step S30: Sliding sampling is performed on the alarm feature sequence through a preset sliding window to obtain an initial alarm sample.

[0101] It should be understood that the sliding window method, also known as the ruler method, is used to solve problems such as finding the properties (length, etc.) of continuous intervals that meet certain conditions. In this embodiment, the window size of the preset sliding window can be set to a fixed duration (e.g., 168 hours, i.e., 7*24 hours a week), and the sliding interval for each sampling through the sliding window can also be set arbitrarily, such as 24 hours, 48 ​​hours, etc.

[0102] like Figure 3 As shown, Figure 3 This is a schematic diagram of the sliding window sampling in the first embodiment of the base station out-of-service alarm prediction method of the present invention. If the sliding interval is set to 24 hours, the number of sliding samples in each sampling period is an integer multiple of 24 hours.

[0103] In a specific implementation, once the alarm feature sequence is determined, the preset sliding window can be used to perform sliding sampling on the alarm feature sequence to obtain an initial alarm sample. In this embodiment, the sampling range can be determined based on the number of days to be predicted. For example, if the number of days to be predicted is 3 days (i.e., 3*24 hours), and the rolling interval is 24 hours, the sliding window needs to be sampled 3 times.

[0104] Step S40: adding features to the initial alarm prediction sample according to the historical operation and maintenance data to obtain an alarm prediction sample.

[0105] It should be noted that the addition of the above-mentioned features in this step can be a process of adding the past decommissioning features, attribute features and other data of the base station to be predicted to the initial alarm sample based on historical operation and maintenance data. This embodiment can enrich the data dimensions of the alarm prediction sample by adding features, so that the subsequent decommissioning alarm prediction results are more accurate.

[0106] It should be noted that the above-mentioned decommissioning features may be feature data related to the decommissioning of base stations, which are analyzed based on historical operation and maintenance data, such as the number of occurrences of each type of alarm in the historical operation and maintenance data of the base station, the proportion of days for which historical decommissioning alarms occurred in the base station, the proportion of days for which each type of decommissioning alarm occurred in the base station, and / or whether a decommissioning alarm occurred in the previous 1-3 days for the base station, whether a decommissioning alarm occurred in the previous 4-7 days, etc. In addition, the above-mentioned attribute features may include data obtained after labeling the coverage area, computer room, base station name, base station ID, cell name, etc. of each base station, as well as data such as the number of coverage areas of each base station, the number of associated base stations under the computer room, and the number of cell names. In this embodiment, the above-mentioned labeling processing may be implemented through label encoder technology.

[0107] Step S50: performing base station out-of-service alarm prediction based on the alarm prediction samples using a preset out-of-service alarm prediction model.

[0108] It should be understood that whether a base station will generate an outage alarm in the future is a binary classification problem, that is, the prediction result is either an occurrence or a non-occurrence. Therefore, the preset outage alarm prediction model described in this embodiment can be a binary classification model trained using outage alarm prediction samples constructed based on a large amount of historical base station operation and maintenance data, such as a logistic regression model or a gradient descent tree (GBDT).

[0109] Furthermore, considering that GBDT has many nonlinear transformations, strong expressive power, does not require complex feature engineering and feature transformation, and has high prediction accuracy, the preset outage alarm prediction model in this embodiment is preferably a model obtained by training GBDT.

[0110] It should be understood that GBDT consists of two main concepts: regression trees and gradient boosting. The essence of a regression tree is to classify data, with each node split into two sub-nodes. During the formation of the binary tree, the optimal split point is continuously searched for through iterations, ultimately forming a binary tree. Gradient boosting first assigns the same weight to each training sample. With each iteration of model training, the weights of incorrectly classified samples are increased, while the weights of correctly classified samples are decreased. This results in a new model trained in the direction of the gradient that reduces the residual error, resulting in a model with higher prediction accuracy.

[0111] In the specific implementation, after obtaining the alarm prediction sample of the base station to be predicted, the alarm prediction sample can be input into the above-mentioned preset decommissioning alarm prediction model, and then the probability of the decommissioning alarm of the base station to be predicted occurring within the predicted time is determined based on the result output by the model, so as to achieve accurate prediction of the decommissioning alarm of the base station to be predicted.

[0112] This embodiment obtains the historical operation and maintenance data of the base station to be predicted within a preset time period, and performs word vector encoding based on the alarm data in the historical operation and maintenance data to obtain an alarm feature sequence; then, the alarm feature sequence is slidingly sampled through a preset sliding window to obtain an initial alarm sample; then, features are added to the initial alarm prediction sample based on the historical operation and maintenance data to obtain an alarm prediction sample; finally, a base station decommissioning alarm is predicted based on the alarm prediction sample through a preset decommissioning alarm prediction model. Since the alarm prediction sample is generated by sampling and adding features to the alarm feature sequence generated based on the historical alarm data of the base station to be predicted, and then the base station decommissioning alarm is predicted based on the generated alarm prediction sample through a pre-trained decommissioning alarm prediction model, compared to the existing method of predicting base station decommissioning based on the experience of technical personnel, the above method of this embodiment fully considers the historical alarm data of the base station to be predicted, thereby ensuring the accuracy and reliability of the prediction result.

[0113] refer to Figure 4 , Figure 4 This is a flow chart of the second embodiment of the base station out-of-service alarm prediction method of the present invention.

[0114] Based on the above first embodiment, in this embodiment, step S20 includes:

[0115] Step S201: determining the alarm identifier of the base station to be predicted according to the alarm data in the historical operation and maintenance data.

[0116] It should be noted that the alarm identifier described in this step can be identification information that can distinguish different alarms, such as the name or title of the alarm, etc. In this embodiment, the above-mentioned alarm identifier can be determined by the alarm data in the historical operation and maintenance data of the base station to be predicted, and can also be manually input. In actual applications, after analyzing the historical operation and maintenance data of a large number of base stations, technical personnel found that there are about 100 types of alarm identifiers corresponding to base stations, and these alarm identifiers respectively represent more than 100 different types of base station alarms. In actual applications, in order to save prediction time, the method of obtaining the above-mentioned alarm identifier in this embodiment can also be that the technical personnel first analyze the above-mentioned 100 types of alarm identifiers based on the alarm data in the historical operation and maintenance data, and then when performing the base station decommissioning alarm prediction, directly input these more than 100 types of alarm identifiers obtained in advance into the trained preset word2vec model for word vector encoding.

[0117] Step S202: Mapping the warning identifier to a corresponding warning word vector through a preset word2vec model.

[0118] In practical applications, considering that word2vec is an algorithm (model) that can be used to learn word vectors, its core idea is to use a simplified neural network to train the context close to the word to obtain the vectorized representation of the word. The algorithm can learn high-quality word vectors from a large-scale corpus in a relatively short period of time, so that the semantic similarity between words can be easily calculated, and this similarity has multiple angles and supports linear operations. Therefore, in this embodiment, it is preferred to vectorize the warning sign by the word2vec model.

[0119] It should be understood that the word2vec model mainly includes two models: CBOW (Continuous Bag of Words Model) and Skip-gram (Continuous Skip-gram Model), which respectively predict the central word through the context and the context through the central word, mapping the word into a multi-dimensional dense vector. The model structure is as follows: Figure 5 As shown, the input layer (w t-c 、w t-c+1 、w t 、w t+c-1 、w t+c ), hidden layer (X wt ), output layer (w t ); the input layer in the Skip-gram model (w t ), hidden layer (X wt ), output layer (w t-c 、w t-c+1 、w t 、w t+c-1 、w t+c ).

[0120] In a specific implementation, the alarm identifier can be mapped to a one-to-one corresponding alarm word vector through the pre-trained preset word2vec model, that is, the corresponding alarm word vector is obtained by multiplying it with the weight matrix obtained by model training.

[0121] Step S203: Encode the warning word vector to obtain a warning feature sequence.

[0122] It should be noted that the alarm word vector is encoded in this step, that is, the process of converting the alarm word vector into a value represented by a specific numerical value (coded value). The specific encoding rules are not limited in this embodiment. In addition, in this embodiment, the above-mentioned alarm feature sequence can be composed of a numerical value (alarm feature value) calculated by the coded value of all alarms of the base station within a unit time. For example, the granularity can be hourly, and the coded values ​​of all alarms per hour can be accumulated, and then the accumulated value can be used as the alarm coded value corresponding to the hour, and the alarm feature sequence can be constructed by counting the alarm coded values ​​of several hours.

[0123] This embodiment determines the alarm identifier of the base station to be predicted based on the alarm data in the historical operation and maintenance data; then maps the alarm identifier to the corresponding alarm word vector through the preset word2vec model; then encodes the alarm word vector to obtain the alarm feature sequence, which can quickly and accurately extract features from the alarm data and ensure the accuracy of subsequent alarm predictions.

[0124] Furthermore, in order to solve the problem of sparse out-of-service alarm data, which leads to inaccurate out-of-service alarm prediction results, this embodiment preferably uses non-out-of-service alarms to characterize the characteristics of out-of-service base stations. To this end, this embodiment divides the alarm word vector into out-of-service alarm word vectors and non-out-of-service alarm word vectors. Accordingly, the above step S203 may include:

[0125] S2031: Obtain the word vector similarity between the service-outage warning word vector and the non-service-outage warning word vector.

[0126] It should be understood that there are many methods for calculating the similarity between vectors, such as the cosine similarity calculation method, the Euclidean distance calculation method, and the Pearson correlation coefficient calculation method, etc. The specific similarity calculation method is not specifically limited in this embodiment.

[0127] However, considering that cosine similarity is widely used in calculating the similarity of document data, this embodiment preferably uses cosine similarity as the method for calculating the word vector similarity in this step.

[0128] It should be understood that for cosine similarity, a cosine value close to 1 and an angle close to 0 indicate that the two vectors are more similar, while a cosine value close to 0 and an angle close to 90 degrees indicate that the two vectors are less similar. Therefore, by calculating the cosine distance between the non-decommissioning alarm vector and the decommissioning alarm vector, the similarity between the non-decommissioning alarm and the decommissioning alarm can be obtained.

[0129] S2032: Calculate the alarm code value of the base station to be predicted based on the word vector similarity.

[0130] After obtaining the above word vector similarities, these word vector similarities can be accumulated according to the following formula (1), and then the accumulated result is used as the alarm code value of the base station to be predicted.

[0131] Wherein, the formula (1) is:

[0132]

[0133] Where, Ecode i is the alarm code value, j is the jth type of out-of-service alarm, i is the ith type of non-out-of-service alarm, δ ij is the word vector similarity between the i-th type of non-decommissioning alarm and the j-th type of decommissioning alarm.

[0134] S2033: Generate an alarm feature sequence according to the alarm code value.

[0135] In a specific implementation, in order to achieve refined prediction of the out-of-service alarm, after this embodiment calculates the alarm code value of the base station to be predicted in the above manner, the alarm code value can be counted at an hourly granularity, and then all the alarm code values ​​of the base station to be predicted within each hour are integrated (which can be summed up) into the alarm feature value of the hour according to the statistical results, and then an alarm feature sequence is constructed based on these alarm feature values.

[0136] This embodiment fully considers the correlation between non-out-of-service alarms and out-of-service alarms through the above method, including the impact of other alarms and adjacent alarms around each out-of-service alarm on the out-of-service alarm. By establishing a word vector encoding model related to the out-of-service alarm context information (i.e., the above-mentioned word2vec model), a large number of non-out-of-service alarms are used to characterize the characteristics of the out-of-service base station, thereby solving the problem of sparse out-of-service alarms and insufficient data.

[0137] Furthermore, in order to achieve out-of-service alarm prediction more accurately and efficiently, in this embodiment, the above-mentioned step S2033 may include: counting the sum of the alarm code values ​​of the base station to be predicted within a unit time period according to the alarm code value; using the sum of the alarm code values ​​as the alarm feature value corresponding to each unit time period; and constructing an alarm feature sequence of the base station to be predicted within the preset time period according to the alarm feature value.

[0138] It should be noted that the above-mentioned unit time period can be set according to actual conditions. For example, with the granularity of hours, the period can be set to 1 hour, 2 hours...n hours, etc. The specific value of the preset time period can also be set according to actual conditions. For example, if the unit time period is 1 hour, then the alarm feature sequence of the base station to be predicted between 12:00-15:00 can be composed of the alarm feature value a corresponding to 12:00-13:00, the alarm feature value b corresponding to 13:00-14:00, and the alarm feature value c corresponding to 14:00-15:00: "alarm feature value a | alarm feature value b | alarm feature value c".

[0139] In a specific implementation, after the alarm feature sequence is constructed, a prediction sample for out-of-service alarm prediction can be generated based on the alarm feature sequence, and then subsequent prediction operations can be performed.

[0140] refer to Figure 6 , Figure 6 This is a flow chart of the third embodiment of the base station out-of-service alarm prediction method of the present invention.

[0141] Based on the above first embodiment, in this embodiment, step 430 includes:

[0142] Step S401: Analyze the historical operation and maintenance data to obtain historical decommissioning characteristics and base station attribute characteristics corresponding to the base station to be predicted.

[0143] It should be emphasized that in this embodiment, the historical decommissioning feature can be feature data related to the decommissioning of the base station, which is analyzed based on the historical operation and maintenance data of the base station to be predicted. For example, the number of occurrences of each type of alarm in the historical operation and maintenance data of the base station, the ratio of days for the occurrence of historical decommissioning alarms of the base station, the ratio of days for each type of decommissioning alarm of the base station, and / or whether a decommissioning alarm occurred in the previous 1-3 days for the base station, whether a decommissioning alarm occurred in the previous 4-7 days, etc. The base station attribute feature can include data obtained after labeling the coverage area, computer room, base station name, base station ID, cell name, etc. of each base station, as well as data such as the number of coverage areas of each base station, the number of associated base stations under the computer room, and the number of cell names.

[0144] In specific implementation, historical operation and maintenance data can be analyzed through data analysis software, or it can be analyzed manually.

[0145] Step S402: adding features to the initial alarm prediction sample according to the historical service withdrawal features and the base station attribute features to obtain an alarm prediction sample.

[0146] It should be noted that the so-called adding features is to add the above-mentioned historical service withdrawal features and base station attribute features to the initial alarm prediction sample to improve the prediction accuracy.

[0147] This embodiment analyzes historical operation and maintenance data to obtain historical decommissioning characteristics and base station attribute characteristics corresponding to the base station to be predicted; then, features are added to the initial alarm prediction sample based on the historical decommissioning characteristics and base station attribute characteristics to obtain an alarm prediction sample. This can fully consider the causes and characteristics of the historical decommissioning alarms of each base station, and use this as the basis for subsequent alarm predictions, thereby ensuring the accuracy of the prediction results.

[0148] refer to Figure 7 , Figure 7 This is a flow chart of a fourth embodiment of a base station out-of-service alarm prediction method according to the present invention.

[0149] It is worth noting that with the rapid development of artificial intelligence technology, GBDT in the binary classification model has been widely used in classification, sorting, regression and other problems, but it has problems such as difficulty in parallel processing, low training efficiency and high computational complexity.

[0150] Considering that the XGboost and Lightgbm algorithms can both efficiently implement the GBDT algorithm and have made many algorithmic and engineering improvements, they have obvious advantages in processing large-scale data. They also support parallel computing and have higher accuracy. Therefore, this embodiment preferably uses the binary classification model of XGboost and Lightgbm based on GBDT to predict outage alarms. Therefore, the fourth embodiment of the base station outage alarm prediction method of the present invention is proposed.

[0151] It should be noted that both the XGboost and Lightgbm algorithms are based on the Gradient Descent Tree (GBDT) boosting method. GBDT is a representative boosting algorithm in ensemble algorithms. Ensemble algorithms split the original dataset into multiple sub-datasets, randomly assigning each sub-dataset to a weak estimator. Each weak estimator is trained based on the residual error of the previous weak estimator. After multiple iterations, the modeling results of all weak estimators are gradually aggregated to achieve better regression or classification performance than a single model.

[0152] In order to accurately predict the outage alarm using the XGBoost model and the Lightgbm model, in this embodiment, corresponding model training is required before the prediction. Accordingly, the base station outage alarm prediction method further includes the following steps:

[0153] Step S01: obtaining training samples, and performing GroupKFold grouping on the training samples to obtain multiple subsample sets;

[0154] It should be noted that the training samples in this embodiment are all labeled samples. In a specific implementation, labels can be added to the samples based on the predicted number of days n. If the base station corresponding to the training sample experiences an out-of-service alarm within the predicted n days, n is set to 1; if no out-of-service alarm occurs, n is set to 0.

[0155] In a specific implementation, the training samples can be grouped according to the week they are in, and then GroupKFold is used for cross-validation to split the original samples into K sub-sample sets.

[0156] It should be noted that the number of weeks of training samples can be determined based on the time nodes at which various types of data in the sample are generated. For example, if the time span of the prediction sample is 1 month, the data in the training sample can be divided into the first week (data), the second week (data), the third week (data) and the fourth week (data) according to the time when the data is generated.

[0157] Considering that due to the sample construction method, samples with similar time periods have high data similarity, the trained model can flexibly learn features from these samples, but will perform poorly when encountering new data. Therefore, this embodiment uses GroupKFold for cross-validation to effectively prevent samples from the same week from appearing in both the training and test sets, thereby improving the generalization ability of the final trained model.

[0158] Step S02: selecting a subsample set from the plurality of subsample sets as a model test set, and using the subsample sets other than the model test set as a model training set;

[0159] In the specific implementation, for the grouped sub-sample sets, each individual sub-sample set is retained as data for validating the model, and the other K-1 sample sets are used to train the XGBoost and Lightgbm models, ensuring that samples from the same week do not appear in both the model training set and the model test set.

[0160] Step S03: The initial XGBoost model and the initial Lightgbm model are trained respectively according to the model training set to obtain a preset XGBoost model and a preset Lightgbm model.

[0161] After the training samples are grouped, in order to ensure that each sub-sample set is verified once and K models are generated at the same time, the above steps S02K need to be repeated. After completing the training of the K sub-models, the weights of each sub-model are redistributed according to their accuracy on the verification set.

[0162] It should be noted that, in order to solve the problem of serious imbalance in the ratio of positive and negative samples in the training samples, this embodiment uses the improved initial XGBoost model and the improved initial Lightgbm model to train the outage alarm prediction model.

[0163] Specifically, Focal loss can be used instead of the traditional cross entropy loss function, so that the binary cross entropy loss function in the above initial XGBoost model and the initial Lightgbm model is:

[0164]

[0165] Where, P i is the prediction probability, α and γ are weighting factors.

[0166] Focal loss penalizes negative samples and easy-to-classify samples by introducing weighting factors α and γ. Its function form is as follows:

[0167]

[0168] A value greater than 0 reduces the loss of easily classified samples, allowing the model to focus more on difficult, misclassified samples. For example, if γ = 2, a positive sample with a prediction probability of 0.95 is considered an easy sample. Therefore, the power of γ (1 - 0.95) is smaller, and the loss function value also becomes smaller. However, a sample with a predicted probability of 0.3 will experience a relatively large loss. Therefore, by adding a balancing weighting factor α, the weight of positive samples is increased, thereby balancing the uneven ratio of positive and negative samples.

[0169] like Figure 8 As shown, Figure 8 The following is a flow chart of the base station outage alarm prediction process based on the XGBoost and Lightgbm models. After obtaining the outage alarm samples (i.e., the training samples mentioned above), they are grouped according to the week they occurred in using GroupKFold. The XGBoost and Lightgbm models are then trained simultaneously based on the number of sub-sample sets. After training, each sub-model is weighted based on its accuracy on the validation set. The final prediction result is obtained by fusing the prediction results output by the sub-models using the model weights.

[0170] Accordingly, step S40 in the first embodiment may include: grouping the alarm prediction samples to obtain a preset number of alarm prediction subsample sets; inputting the alarm prediction subsample sets into the preset XGBoost model and the preset Lightgbm model respectively; and performing weighted summation on the first prediction result output by the preset XGBoost model and the second prediction result output by the preset Lightgbm model to obtain a base station out-of-service alarm prediction result. For example, if the first prediction result output by the XGBoost model is (Y: 70%, N: 30%), the second prediction result output by the Lightgbm model is (Y: 60%, N: 40%), the weight of the XGBoost model is 0.7, and the weight of the Lightgbm model is 0.3, then the final prediction result is (Y: 70%*0.7+60%*0.3, N: 30%*0.7+40%*0.3)=(Y: 67%, N: 33%).

[0171] This embodiment obtains training samples and performs GroupKFold grouping on the training samples to obtain multiple sub-sample sets; selects a sub-sample set from the multiple sub-sample sets as a model test set, and uses the sub-sample sets other than the model test set as a model training set; trains the initial XGBoost model and the initial Lightgbm model according to the model training set to obtain a preset XGBoost model and a preset Lightgbm model, thereby building an AI service outage alarm prediction model based on the XGBoost model and the Lightgbm model. At the same time, by improving the XGBoost model and the Lightgbm model, that is, introducing two weighting factors into the model loss function, the impact of sample imbalance on model training is reduced, and during the model training process, the model parameters are optimized through grid optimization and cross-validation techniques, and the generalization ability of the model is improved by fusion of multi-model prediction results.

[0172] refer to Figure 9 , Figure 9 This is a flow chart of a fifth embodiment of a base station out-of-service alarm prediction method according to the present invention;

[0173] Based on the above embodiments, a fifth embodiment of a base station out-of-service alarm prediction method of the present invention is proposed.

[0174] The base station out-of-service alarm prediction method proposed in this embodiment further includes, before step S10, the following steps:

[0175] Step S01': Obtain historical operation and maintenance data of different base stations;

[0176] It should be noted that in order to quickly and accurately implement word vector encoding, this embodiment will construct a base station alarm code corpus based on the historical operation and maintenance data of different base stations, and then train the word2vec model based on the corpus.

[0177] In this embodiment, the above-mentioned historical operation and maintenance data can be collected from the network management system, mainly including the working parameter data and alarm data of the base station. Among them, the alarm data can include various types of alarm records such as fault alarms, dynamic environment alarms, and performance alarms. These alarm records describe the time when the base station equipment alarm occurs, the alarm category, the base station name and other information, which can be used to construct a training sample set for the out-of-service alarm prediction model. The working parameter data can be data describing the physical properties of the base station, including the latitude and longitude of the base station location, the covered area, the affiliated machine room, the base station name, the base station ID and other information.

[0178] Step S02': constructing a base station alarm code corpus based on the historical operation and maintenance data of the different base stations;

[0179] In a specific implementation, the above historical operation and maintenance data of different base stations can be processed to a certain extent (such as data denoising, deduplication, missing value filling, etc.), and then a base station alarm code corpus can be constructed based on the processed data.

[0180] Furthermore, considering that there are many duplicate and abnormal data in the (historical) operation and maintenance data of the base station, these data will directly affect the effect of model training. Therefore, as another embodiment, the above step S02' may also include:

[0181] Step S021': obtaining original alarm data of each base station from the historical operation and maintenance data of the different base stations;

[0182] It should be noted that the above-mentioned raw alarm data can be data describing information such as the time when the base station equipment alarm occurred, the alarm type, the base station name, etc. Typically, the raw alarm data contains a large number of duplicate alarms and abnormal alarms, which need to be removed to reduce interference with model training.

[0183] Step S022': pre-process the original alarm data to obtain valid alarm data.

[0184] It should be noted that the pre-processing of the original alarm data in this step may be to remove duplicate data (ie, deduplication) and eliminate invalid data (ie, denoising), so as to obtain valid alarm data.

[0185] In order to ensure the accuracy and reliability of valid data, this embodiment can also deduplicate the original alarm data according to the alarm fields of preset dimensions to obtain the alarm data to be processed; then eliminate the abnormal alarm data in the alarm data to be processed to obtain valid alarm data.

[0186] The preset dimensions may include the alarm occurrence time, alarm end time, alarm name and / or alarm network element. The above-mentioned abnormal alarm data can be eliminated based on a large amount of historical alarm data and the expert experience summarized by relevant experts based on the historical alarm data. For example, after statistics on historical alarm data, it is found that the causes of service outage alarms are roughly divided into the following three categories:

[0187] ① Equipment level: For example, equipment failure, network element abnormality, etc. cause service outage;

[0188] ②Business level: For example, service outages caused by provincial testing, carrier scheduling, etc.

[0189] ③ Human factors: For example, power outages, property disputes, etc. may cause service withdrawal.

[0190] Outage alarm data caused by non-network-related reasons, such as business-level or human factors, can significantly interfere with model training and prediction. Therefore, to improve model effectiveness, alarm data that may affect the prediction model is defined as invalid or abnormal alarm data and removed.

[0191] Accordingly, in this embodiment, the abnormal alarm data includes but is not limited to: data on alarms caused by some provincial tests, data on alarms that do not affect network quality, data on alarms occurring in power-saving base stations at night, and data on alarms occurring in some base stations during the carrier scheduling period.

[0192] Step S023 ′: aggregate and sort the valid alarm data, and generate an alarm sequence according to the aggregated and sorted alarm data.

[0193] It should be noted that the so-called aggregation sorting can be the process of aggregating alarm data according to certain rules (such as base station name) and dividing it into different data groups; then sorting these data groups in a certain order (such as the time when the alarm occurred).

[0194] Furthermore, considering that the objects of out-of-service alarm prediction are all network base stations, and the alarm patterns of different network base stations may vary, predicting the future out-of-service alarm situation of each base station based on its historical alarm data can make the prediction results more accurate, objective, and targeted. Therefore, in this embodiment, the valid alarm data can be aggregated into different alarm data groups according to the base station identifier; the alarm data in each alarm data group is then sorted in chronological order of alarm occurrence to obtain sorted alarm data groups; and an alarm sequence is then generated based on the sorted alarm data groups. The base station identifier can be, for example, the base station name or base station ID.

[0195] It should be noted that in order to maintain data consistency, the alarm time range of each alarm data group needs to be unified. Specifically, the earliest date D of the alarm occurrence in each alarm data group can be counted. first and the latest date D last , then according to the earliest date D first and the latest date D last To determine the alarm time range. After the alarm time range is unified, the alarm data in each alarm data group can be sorted in the order of the alarm occurrence time to obtain the sorted alarm data group. Then, the sorted alarm data group is further divided into hourly granularity, and the hourly periods without alarm data are filled with zero values ​​(i.e., missing items are filled). Finally, the alarm sequence is generated based on the filled alarm data group.

[0196] Furthermore, to ensure that the time span of the alarm data in an alarm group is not too large, in this embodiment, each out-of-service alarm in the sorted alarm data group can be used as the center, and the alarm data within a preset time range can be used as the alarm group corresponding to the out-of-service alarm. The alarm groups corresponding to all out-of-service alarms can then be merged to obtain an alarm sequence. For example, each out-of-service alarm can be used as the center, and all alarm data from the seven days before and two days after the out-of-service alarm, totaling ten days, can be arranged in sequence as an alarm group. The alarm groups of all base stations can then be merged to generate an alarm sequence that includes the out-of-service alarm and its subsequent co-station alarms. Of course, the method for obtaining the alarm group here is merely an example and does not constitute a specific limitation.

[0197] Step S024': constructing a base station alarm code corpus according to the alarm sequence.

[0198] After obtaining the above alarm sequence, a base station alarm code corpus can be constructed according to the above alarm sequence.

[0199] Step S03 ′: iteratively training the initial word2vec model based on the base station alarm code corpus to obtain a preset word2vec model.

[0200] It should be understood that after the base station alarm code corpus is constructed, the initial word2vec model can be iteratively trained based on the alarm sequences in the base station alarm code corpus to obtain a trained word2vec model.

[0201] This embodiment uses historical alarm data such as base station-level faults, performance, and dynamic environment alarms to serialize and analyze the historical alarm data based on word2vec word vector modeling, comprehensively exploring the intrinsic correlation between secondary alarms and out-of-service alarm data, and improving the accuracy of out-of-service alarm prediction.

[0202] In addition, an embodiment of the present invention further provides a storage medium storing a base station out-of-service alarm prediction program. When the base station out-of-service alarm prediction program is executed by a processor, the steps of the base station out-of-service alarm prediction method described above are implemented.

[0203] Reference Figure 10 , Figure 10 This is a structural block diagram of the first embodiment of the base station out-of-service alarm prediction device of the present invention.

[0204] like Figure 10 As shown, the base station out-of-service alarm prediction device proposed in an embodiment of the present invention includes:

[0205] The data acquisition module 101 is used to obtain historical operation and maintenance data of the base station to be predicted within a preset time period;

[0206] A vector encoding module 102 is configured to perform word vector encoding on the alarm data in the historical operation and maintenance data to obtain an alarm feature sequence;

[0207] The window sampling module 103 is configured to perform sliding sampling on the alarm feature sequence through a preset sliding window to obtain an initial alarm sample;

[0208] A feature adding module 104 is configured to add features to the initial alarm prediction sample based on the historical operation and maintenance data to obtain an alarm prediction sample;

[0209] The alarm prediction module 105 is configured to perform base station out-of-service alarm prediction based on the alarm prediction samples using a preset out-of-service alarm prediction model.

[0210] This embodiment obtains the historical operation and maintenance data of the base station to be predicted within a preset time period, and performs word vector encoding based on the alarm data in the historical operation and maintenance data to obtain an alarm feature sequence; then, the alarm feature sequence is slidingly sampled through a preset sliding window to obtain an initial alarm sample; then, features are added to the initial alarm prediction sample based on the historical operation and maintenance data to obtain an alarm prediction sample; finally, a base station decommissioning alarm is predicted based on the alarm prediction sample through a preset decommissioning alarm prediction model. Since the alarm prediction sample is generated by sampling and adding features to the alarm feature sequence generated based on the historical alarm data of the base station to be predicted, and then the base station decommissioning alarm is predicted based on the generated alarm prediction sample through a pre-trained decommissioning alarm prediction model, compared to the existing method of predicting base station decommissioning based on the experience of technical personnel, the above method of this embodiment fully considers the historical alarm data of the base station to be predicted, thereby ensuring the accuracy and reliability of the prediction result.

[0211] Furthermore, other embodiments or specific implementations of the base station out-of-service alarm prediction device of the present invention can refer to the above-mentioned method embodiments and will not be repeated here.

[0212] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0213] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0214] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0215] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for predicting a base station out-of-service alarm, characterized in that: The base station out-of-service alarm prediction method includes: Obtain historical operation and maintenance data of the base station to be predicted within a preset time period; Perform word vector encoding on the alarm data in the historical operation and maintenance data to obtain an alarm feature sequence; Sliding sampling is performed on the alarm feature sequence through a preset sliding window to obtain an initial alarm prediction sample; Adding features to the initial alarm prediction sample according to the historical operation and maintenance data to obtain an alarm prediction sample; Performing base station out-of-service alarm prediction based on the alarm prediction samples using a preset out-of-service alarm prediction model; The step of performing word vector encoding on the alarm data in the historical operation and maintenance data to obtain an alarm feature sequence includes: Determining an alarm identifier of the base station to be predicted based on the alarm data in the historical operation and maintenance data, where the alarm identifier is identification information for distinguishing different alarms; Mapping the warning identifier to a corresponding warning word vector through a preset word2vec model, wherein the warning word vector includes a service withdrawal warning word vector and a non-service withdrawal warning word vector; Obtaining the word vector similarity between the service withdrawal warning word vector and the non-service withdrawal warning word vector; Calculating the warning code value of the base station to be predicted according to the word vector similarity; generating an alarm feature sequence according to the alarm code value; The step of adding features to the initial alarm prediction sample according to the historical operation and maintenance data to obtain the alarm prediction sample includes: Analyzing the historical operation and maintenance data to obtain historical service withdrawal characteristics and base station attribute characteristics corresponding to the base station to be predicted; Features are added to the initial alarm prediction sample according to the historical service withdrawal features and the base station attribute features to obtain an alarm prediction sample.

2. The base station out-of-service alarm prediction method according to claim 1, wherein: The step of generating an alarm feature sequence according to the alarm code value comprises: According to the alarm code value, a sum of the alarm code values ​​of the base station to be predicted within a unit time period is counted; The sum of the alarm code values ​​is used as the alarm characteristic value corresponding to each unit time period; An alarm feature sequence of the base station to be predicted within the preset time period is constructed according to the alarm feature value.

3. The base station out-of-service alarm prediction method according to any one of claims 1 to 2, characterized in that: The preset out-of-service alarm prediction model includes: a preset XGBoost model and a preset Lightgbm model; The step of performing base station out-of-service alarm prediction according to the alarm prediction sample by using a preset out-of-service alarm prediction model includes: Grouping the alarm prediction samples to obtain a preset number of alarm prediction subsample sets; Inputting the alarm prediction subsample set into the preset XGBoost model and the preset Lightgbm model respectively; A weighted sum is performed on the first prediction result output by the preset XGBoost model and the second prediction result output by the preset Lightgbm model to obtain a base station out-of-service alarm prediction result.

4. The base station out-of-service alarm prediction method according to claim 3, wherein: Before the step of obtaining historical operation and maintenance data of the base station to be predicted within a preset time period, the method further includes: Obtain training samples, and perform GroupKFold grouping on the training samples to obtain multiple subsample sets; Selecting a subsample set from the plurality of subsample sets as a model test set, and using the subsample sets other than the model test set as a model training set; The initial XGBoost model and the initial Lightgbm model are trained respectively according to the model training set to obtain a preset XGBoost model and a preset Lightgbm model.

5. The base station out-of-service alarm prediction method according to claim 4, characterized in that: The binary cross entropy loss function in the initial XGBoost model and the initial Lightgbm model is: Where, Focal loss is the binary cross entropy loss function, P i is the predicted probability, α and γ is the weighting factor.

6. The base station out-of-service alarm prediction method according to any one of claims 1 to 2, characterized in that: Before the step of obtaining historical operation and maintenance data of the base station to be predicted within a preset time period, the method further includes: Obtain historical operation and maintenance data of different base stations; Building a base station alarm code corpus based on the historical operation and maintenance data of the different base stations; An initial word2vec model is iteratively trained based on the base station alarm code corpus to obtain a preset word2vec model.

7. The base station out-of-service alarm prediction method according to claim 6, characterized in that: The step of constructing a base station alarm code corpus based on the historical operation and maintenance data of the different base stations includes: Obtaining original alarm data of each base station from the historical operation and maintenance data of the different base stations; Preprocessing the original alarm data to obtain valid alarm data; Aggregating and sorting the valid alarm data, and generating an alarm sequence based on the aggregated and sorted alarm data; A base station alarm code corpus is constructed according to the alarm sequence.

8. The base station out-of-service alarm prediction method according to claim 7, characterized in that: The step of preprocessing the original alarm data to obtain valid alarm data includes: Deduplication of the original alarm data is performed according to the alarm fields of the preset dimensions to obtain the alarm data to be processed; Abnormal alarm data in the alarm data to be processed is eliminated to obtain valid alarm data.

9. The base station out-of-service alarm prediction method according to claim 7, wherein: The step of aggregating and sorting the valid alarm data and generating an alarm sequence according to the aggregated and sorted alarm data includes: Aggregating the valid alarm data into different alarm data groups according to base station identifiers; Sorting the alarm data in each of the alarm data groups in the order of alarm occurrence time to obtain a sorted alarm data group; An alarm sequence is generated according to the sorted alarm data group.

10. The base station out-of-service alarm prediction method according to claim 9, characterized in that: The step of generating an alarm sequence according to the sorted alarm data group comprises: Taking each out-of-service alarm in the sorted alarm data group as the center, the alarm data within a preset time range is used as the alarm group corresponding to the out-of-service alarm; Combine the alarm groups corresponding to all out-of-service alarms to obtain an alarm sequence.

11. A base station out-of-service alarm prediction device, characterized in that: The base station out-of-service alarm prediction device includes: A data acquisition module is used to obtain historical operation and maintenance data of the base station to be predicted within a preset time period; A vector encoding module, configured to perform word vector encoding on the alarm data in the historical operation and maintenance data to obtain an alarm feature sequence; A window sampling module is used to perform sliding sampling on the alarm feature sequence through a preset sliding window to obtain an initial alarm prediction sample; A feature adding module, configured to add features to the initial alarm prediction sample according to the historical operation and maintenance data to obtain an alarm prediction sample; An alarm prediction module is used to predict a base station out-of-service alarm based on the alarm prediction samples using a preset out-of-service alarm prediction model; The vector encoding module is further configured to determine an alarm identifier of the base station to be predicted based on the alarm data in the historical operation and maintenance data, where the alarm identifier is identification information for distinguishing different alarms; map the alarm identifier to a corresponding alarm word vector using a preset word2vec model, where the alarm word vector includes a decommissioning alarm word vector and a non-decommissioning alarm word vector; obtain a word vector similarity between the decommissioning alarm word vector and the non-decommissioning alarm word vector; calculate an alarm code value of the base station to be predicted based on the word vector similarity; and generate an alarm feature sequence based on the alarm code value. The feature adding module is also used to analyze the historical operation and maintenance data to obtain the historical decommissioning features and base station attribute features corresponding to the base station to be predicted; add features to the initial alarm prediction sample based on the historical decommissioning features and the base station attribute features to obtain an alarm prediction sample.

12. A base station out-of-service alarm prediction device, characterized in that: The device includes: a memory, a processor, and a base station out-of-service alarm prediction program stored in the memory and executable on the processor, wherein the base station out-of-service alarm prediction program is configured to implement the steps of the base station out-of-service alarm prediction method according to any one of claims 1 to 10.

13. A storage medium, characterized in that: The storage medium stores a base station out-of-service alarm prediction program, which, when executed by a processor, implements the steps of the base station out-of-service alarm prediction method according to any one of claims 1 to 10.

Images