A training method and device of a base station outage alarm model

Abstract

The application discloses a training method and device of a base station retreat service alarm model, to improve the accuracy of predicting whether the base station retreats service. The method comprises the following steps: obtaining first historical data; dividing the first historical data according to a time window, and obtaining alarm time and corresponding alarm type of base station retreat service alarm in the time window; generating a feature sequence with alarm type codes corresponding to the alarm type of N continuous time windows, and arranging the alarm type codes in the feature sequence based on the time sequence of the corresponding alarm time; taking the feature sequence as sample data, taking whether the base station retreats service alarm occurs in the first time window after N continuous time windows as a label, inputting a bidirectional long short-term memory network model to perform training, and obtaining a base station retreat service alarm model. The method generates multiple training samples and corresponding labels based on historical alarm data to train the model, which can effectively predict whether the base station retreats service alarm occurs and improve the prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of model training, and in particular to a training method and apparatus for a base station outage alarm model. Background Technology

[0002] When existing network base stations malfunction during operation, they generate alarms, some of which are out-of-service alarms. These require the issuance of a fault work order, which is then handled by maintenance personnel at the base station. Out-of-service alarms indicate that the base station has gone out of service; after this, users will be unable to connect to the network, affecting normal communication. Currently, base station maintenance is reactive; fault work orders are only issued after an out-of-service alarm occurs. This lag between alarm generation and maintenance makes it difficult to guarantee the stability of base station functionality.

[0003] The technical problem this application aims to solve is how to accurately predict whether a base station will go out of service. Summary of the Invention

[0004] The purpose of this application is to provide a training method and apparatus for a base station outage alarm model, so as to improve the accuracy of predicting whether a base station will go out of service.

[0005] Firstly, a training method for a base station outage alarm model is provided, including:

[0006] Obtain first historical data, which includes the alarm time and corresponding alarm type of the base station outage alarm;

[0007] The first historical data is divided into time windows, and the alarm time and corresponding alarm type of the base station outage alarm that occurred within the time window are obtained.

[0008] A feature sequence is generated by generating alarm type codes corresponding to alarm types in N consecutive time windows. The alarm type codes in the feature sequence are arranged in temporal order based on the corresponding alarm times, where N is an integer greater than 1.

[0009] Using the feature sequence as sample data, and labeling whether a base station outage alarm occurs within the first time window after the N consecutive time windows, the data is input into a bidirectional long short-term memory network model for training to obtain a base station outage alarm model.

[0010] Secondly, a training device for a base station outage alarm model is provided, comprising:

[0011] The first acquisition module acquires first historical data, which includes the alarm time when the base station outage alarm occurred and the corresponding alarm type.

[0012] The second acquisition module divides the first historical data into time windows and acquires the alarm time and corresponding alarm type of the base station outage alarm that occurred within the time window;

[0013] The generation module generates a feature sequence based on the alarm type codes corresponding to the alarm types of N consecutive time windows. The alarm type codes in the feature sequence are arranged in chronological order based on the corresponding alarm times, where N is an integer greater than 1.

[0014] The training module uses the feature sequence as sample data and the label of whether a base station outage alarm occurs within the first time window after the N consecutive time windows. It inputs the data into the bidirectional long short-term memory network model for training to obtain the base station outage alarm model.

[0015] Thirdly, an electronic device is provided, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method of the first aspect.

[0016] Fourthly, a computer-readable storage medium is provided on which a computer program is stored, which, when executed by a processor, implements the steps of the method of the first aspect.

[0017] In this embodiment, first historical data is acquired, including the alarm time and corresponding alarm type of a base station outage alarm. The first historical data is divided into time windows, and the alarm time and corresponding alarm type of a base station outage alarm within each time window are obtained. A feature sequence is generated using the alarm type codes corresponding to the alarm types of N consecutive time windows. The alarm type codes in the feature sequence are arranged chronologically based on the corresponding alarm times, where N is an integer greater than 1. Using the feature sequence as sample data, and using whether a base station outage alarm occurred within the first time window after N consecutive time windows as a label, the data is input into a bidirectional long short-term memory network model for training to obtain a base station outage alarm model. This solution generates multiple training samples and corresponding labels based on historical alarm data to train the model, effectively predicting whether a base station outage alarm will occur and improving prediction accuracy. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0019] Figure 1 This is one of the flowcharts illustrating a training method for a base station outage alarm model according to an embodiment of the present invention.

[0020] Figure 2 This is a second schematic flowchart of a training method for a base station outage alarm model according to an embodiment of the present invention.

[0021] Figure 3 This is the third flowchart illustrating a training method for a base station outage alarm model according to an embodiment of the present invention.

[0022] Figure 4 This is the fourth flowchart illustrating a training method for a base station outage alarm model according to an embodiment of the present invention.

[0023] Figure 5 This is the fifth flowchart illustrating a training method for a base station outage alarm model according to an embodiment of the present invention.

[0024] Figure 6 This is the sixth flowchart illustrating a training method for a base station outage alarm model according to an embodiment of the present invention.

[0025] Figure 7 This is the seventh flowchart illustrating a training method for a base station outage alarm model according to an embodiment of the present invention.

[0026] Figure 8 This is a schematic diagram of the structure of a training device for a base station outage alarm model according to an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The drawing numbers in this application are only used to distinguish the various steps in the solution and are not used to limit the execution order of the various steps. The specific execution order is as described in the specification.

[0028] To address the problems existing in the prior art, embodiments of this application provide a training method for a base station outage alarm model, such as... Figure 1 As shown, it includes:

[0029] S11: Obtain first historical data, which includes the alarm time when the base station outage alarm occurred and the corresponding alarm type.

[0030] The first historical data in this step specifically includes data related to base station outage alarms that occurred within a historical time period. For any base station outage alarm occurring within that historical time period, the corresponding alarm time and alarm type are recorded. The alarm time can be either the alarm discovery time or the alarm occurrence time. The alarm type indicates the cause of the alarm and can be used for source tracing analysis to facilitate rapid anomaly repair. In addition, the first historical data may also include information such as the name of the network element that triggered the base station outage alarm.

[0031] For example, the historical data of a certain base station from July 1st to July 9th is shown in Table 1 below:

[0032] Table 1

[0033]

[0034]

[0035] In practical applications, the alarm discovery time and the alarm occurrence time may be different. In this example, the alarm occurrence time is taken as the alarm time corresponding to the alarm type, and the alarm name is taken as the alarm type.

[0036] S12: Divide the first historical data into time windows and obtain the alarm time and corresponding alarm type of the base station outage alarm that occurred within the time window.

[0037] The time window can be preset according to actual needs. For example, if the first historical data consists of base station outage alarm data from multiple days, then the first historical data can be divided into one-day time windows to obtain the base station outage alarm data for each day. Specifically, the time window to which the corresponding alarm type belongs is determined based on the alarm time, so that any alarm can be assigned to a single time window.

[0038] Based on the historical data from July 1st to July 9th, this step divides the data into 9 sets based on the time window, corresponding to the base station outage alarm data for those 9 days. For ease of explanation, the data obtained from the above time window division is referred to as historical sub-data in this example, and 9 historical sub-data sets are obtained in this step.

[0039] S13: Generate a feature sequence using alarm type codes corresponding to alarm types in N consecutive time windows. The alarm type codes in the feature sequence are arranged in chronological order based on the corresponding alarm times, where N is an integer greater than 1.

[0040] In this step, N is less than the number of historical sub-data obtained from the previous steps. For example, N can be 5, then this step can generate 5 feature sequences: July 1-July 5, July 2-July 6, July 3-July 7, July 4-July 8, and July 5-July 9. It should be understood that the value of N can be adjusted according to actual needs.

[0041] The alarm type codes corresponding to the above alarm types can be pre-defined codes used to mark and distinguish alarm types. These alarm type codes can be numbers, letters, or other forms of codes; in this example, a numerical code will be used.

[0042] In this step, the alarm type encoding in the feature sequence is based on the temporal arrangement of the corresponding alarm times, and the generated feature sequence is used to characterize the type and timing of alarms occurring within N consecutive time windows.

[0043] Optionally, to facilitate subsequent model training, the length of the generated feature sequence can be a preset length. For example, a feature sequence of length 12 can be generated from historical sub-data of a certain day. Specifically, the 24 hours of a day can be divided into 2-hour time intervals, and then a corresponding alarm type code can be determined for each 2-hour interval to represent the key alarm information within those 2 hours. Specifically, the alarm type code can be selected based on the severity of the alarm type, the number of alarms of the same alarm type, or other methods.

[0044] Additionally, if no alarms occur within the defined 2-hour period, the data is padded with 0s to generate a fixed-length feature sequence. The feature sequence generated in this step can also be stored as a feature vector.

[0045] S14: Using the feature sequence as sample data, and using whether a base station outage alarm occurs within the first time window after the N consecutive time windows as a label, input the bidirectional long short-term memory network model for training to obtain the base station outage alarm model.

[0046] Training a Bidirectional Long Short-Term Memory (Bi-LSTM) model often requires multiple sample data points and their corresponding labels. In this step, the labels for each sample data point can be predetermined, and then the sample data and their labels can be input into the model for training.

[0047] In this example, the feature sequence is determined based on the alarm type of N consecutive time windows. Whether a base station outage alarm occurs in the first time window after these N consecutive time windows is used as the label of the feature sequence.

[0048] For example, for the feature sequence generated from historical sub-data from July 1st to July 5th, the label for the feature sequence is whether a base station outage alarm occurred on July 6th. For instance, if a base station outage alarm occurred on July 6th, the label for the feature sequence of the aforementioned historical sub-data from July 1st to July 5th can be "1"; if no base station outage alarm occurred on July 6th, the label can be "0".

[0049] This application's solution can utilize N days of historical alarm data from a base station to predict the probability of a significant outage alarm occurring in the next day. Here, N is a hyperparameter that can be adjusted based on the actual performance of the training prediction.

[0050] Furthermore, if the actual historical alarm data contains a large number of alarm types and alarm frequencies, resulting in an excessively large overall data volume, the source data can be filtered first. Unimportant alarms can be removed, and important base station outage alarm data can be retained as historical alarm data for implementing the solution provided in this embodiment. Specifically, important alarms can be alarms indicating base station or cell outages. For example, for 4G base stations, important alarms could be "network element connection interruption" and "cell unavailability alarm." Important alarms can be pre-defined manually or automatically filtered based on historical alarm data.

[0051] The solution provided in this embodiment trains a Bi-LSTM model based on historical alarm data. After the model converges, it can be used to predict whether a base station outage alarm will occur in the future. In practical applications, the output of the trained Bi-LSTM model represents the probability of a future base station outage alarm. Automatic judgment can be achieved based on the magnitude of a preset baseline probability and the probability output by the model. If the probability output by the model is greater than the preset baseline probability, it can be determined that a base station outage alarm will occur in the future. Subsequently, an alarm can be issued in advance according to the actual situation, and potential faults can be handled in advance through manual maintenance to avoid the occurrence of base station outage alarms, reduce the negative impact of base station outage alarms, and thus improve customer experience.

[0052] Based on the solutions provided in the above embodiments, optionally, such as Figure 2 As shown, step S13 above includes:

[0053] S21: Divide the total duration of the N consecutive time windows according to the preset time length to obtain multiple time intervals.

[0054] For example, assuming N is 7, based on historical alarm data, this step divides the total duration from July 1st to July 7th. Assuming a preset length of 2 hours, each day can be divided into 12 time intervals, resulting in a total of 12 × 7 = 84 time intervals over 7 days.

[0055] In this step, the total duration of N consecutive time windows is equally divided based on a preset time length, so that the base station alarm data of each time period can be analyzed in a balanced manner in subsequent steps. In addition, it can also ensure that the length of the feature sequence is uniform, which is convenient for model training.

[0056] S22: Determine the alarm types corresponding to the multiple time intervals based on the alarm times corresponding to the alarm types of the N consecutive time windows.

[0057] In this step, a corresponding alarm type is determined for each time interval. This alarm type can be a significant alarm among the base station alarms occurring within that time interval. Subsequently, a feature value corresponding to that time interval is determined based on the alarm type. If no base station alarms occur within that time interval, the value for that time interval can be determined to be 0.

[0058] S23: Based on the preset correspondence between alarm type and alarm type code, the alarm type codes corresponding to the multiple time intervals are arranged in chronological order to generate the feature sequence.

[0059] The correspondence between the alarm types and alarm type codes mentioned above can be preset manually, or it can be automatically generated based on historical data and the importance of the alarm type. Optionally, for unimportant alarms, the corresponding alarm type code can be set to "0", thereby reducing the amount of data and the system processing load.

[0060] For example, suppose a base station alarm occurs at 9:18 on July 1st, with the alarm type being "[Derivative Alarm] VVIP Cell LTE Wireless Call Drop Rate Sudden Abnormal Alarm", and a base station alarm occurs at 12:02, with the alarm type being "User Plane Bearer Link Failure Alarm".

[0061] Among them, the alarm type code corresponding to "[Derivative Alarm] VVIP Cell LTE Wireless Drop Rate Sudden Abnormal Alarm" is 3, and the alarm type code corresponding to "User Plane Bearer Link Failure Alarm" is 5.

[0062] Therefore, the feature sequence generated based on the historical alarm data from July 1st could be “0, 0, 0, 0, 3, 0, 5, 0, 0, 0, 0, 0”.

[0063] The solution provided by this application can generate a feature sequence that represents historical base station alarms. The feature values ​​in the feature sequence are alarm type codes determined based on a preset time period of equal division. This can effectively represent the alarm types that occur within the corresponding time period, thereby optimizing the model training effect.

[0064] Based on the solutions provided in the above embodiments, optionally, such as Figure 3 As shown, if there are multiple alarm types within the target time interval, step S22 above includes:

[0065] S31: Determine the weight values ​​corresponding to the various alarm types within the target time interval.

[0066] The weight values ​​corresponding to the above alarm types can be preset according to actual needs. These weight values ​​represent the severity of the corresponding alarm type. Specifically, if an alarm type may cause a large number of business anomalies, its corresponding weight value can be determined to be larger. If an alarm type will not cause business anomalies, its weight value can be determined based on the scope of its actual impact.

[0067] S32: The alarm type with the largest weight value is determined as the alarm type corresponding to the target time interval.

[0068] In this step, for multiple alarm types within the target time interval, the alarm types can be sorted according to their weight values, and then the alarm type with the largest weight value can be selected as the alarm type corresponding to the target time interval.

[0069] In the solution provided in this application embodiment, the selected alarm type is the alarm type with the largest weight value in the target time interval, which can effectively express the most important alarms that occur in the target time interval. This is used as a feature value to generate a feature sequence, so that the feature sequence can more accurately represent the important alarms that occur in each time interval and optimize the training effect of the model.

[0070] Based on the solutions provided in the above embodiments, optionally, such as Figure 4 As shown, before step S13 above, the following steps are also included:

[0071] S41: Obtain the second historical alarm data, which includes the alarm time when the base station outage alarm occurred and the corresponding alarm type.

[0072] The second historical alarm data can be the same as or different from the first historical alarm data. For example, the first and second historical alarm data can be the historical alarm data from July 1st to July 9th. Alternatively, the first historical alarm data can be the historical alarm data from July 1st to July 9th, and the second historical alarm data can be the historical alarm data before July 1st.

[0073] The second historical alarm data includes the alarm time and the corresponding alarm type, used to indicate when and what kind of alarm occurred. The recording format of the second historical alarm data can be the same as that of the first historical alarm data, for example, it can be presented in a table format. In this example, it is assumed that both the first and second historical alarm data are the historical alarm data from July 1st to July 9th mentioned above.

[0074] S42: Divide the second historical data according to the time window and obtain the alarm type of the base station outage alarm that occurred within the time window.

[0075] Similar to the method used to divide the first historical data, this step divides the second historical data according to the aforementioned time window. When the time window is one day, the aforementioned historical alarm data can be divided into 9 historical sub-data corresponding to July 1st to July 9th.

[0076] S43: Perform one-hot encoding on the alarm types of M consecutive time windows to obtain M alarm encoding vectors corresponding to different time windows, where M is an integer greater than 1.

[0077] Where M can be the same as or different from the value of N in the above embodiments. In this step, single-hotspot encoding is performed on the alarm types in each historical sub-data to obtain the alarm encoding vector corresponding to each historical sub-data. For example, based on the historical alarm data of the 7 days from July 1st to July 7th in the above example, the alarm vector for each day is determined, and the following results can be obtained:

[0078] The alarm vector V1 for July 1st is: 1, 1, 1, 1, 1, 1, 1, 0, 0, 0

[0079] The alarm vector V2 on July 2nd is: 0, 0, 1, 0, 0, 1, 1, 1, 0, 0

[0080] The alarm vector V3 on July 3rd is: 1, 1, 0, 0, 0, 1, 1, 0, 0, 0

[0081] The alarm vector V4 for July 4th is: 1, 0, 0, 0, 0, 1, 1, 1, 0, 0

[0082] The alarm vector V5 for July 5th is: 1, 0, 0, 1, 1, 1, 1, 0, 0, 0

[0083] The alarm vector V6 for July 6th is: 1, 0, 0, 1, 1, 1, 0, 0, 0, 0

[0084] The alarm vector V7 on July 7th was: 0, 1, 0, 0, 1, 1, 1, 1, 0, 0

[0085] S44: Determine the attenuation coefficients corresponding to the M alarm coding vectors according to the time order of the time window.

[0086] In this scheme, the historical alarm data from July 1st to July 7th (7 days) is used to generate sample data, and the historical alarm data from July 8th is used to generate corresponding tags. The impact of the alarm data from July 1st to July 7th on the base station alarms on July 8th varies; therefore, an attenuation coefficient is applied in this step to characterize the impact of alarm data from different dates on the base station alarms on July 8th.

[0087] For example, the attenuation coefficient can be 0.5.

[0088] S45: Determine the feature vector based on the product of the M alarm coding vectors and their corresponding attenuation coefficients.

[0089] In this step, different alarm coding vectors correspond to different attenuation coefficients. For example, the attenuation coefficient for July 1st could be 0.5. 6 The attenuation factor corresponding to July 2nd could be 0.5. 5 And so on. The sum of the products of each alarm coding vector in M ​​and its corresponding attenuation coefficient is used to determine the feature vector. In this example, the feature vector can specifically be V. all =V1×0.5 6 +V2×0.5 5 +……+V7.

[0090] S46: Using the feature vector as sample data, and using whether a base station outage alarm occurs within the first time window after the M consecutive time windows as a label, input the data into a machine learning classification model to perform classification.

[0091] Based on the above feature vector V all Based on whether a base station outage alarm occurred on July 8, corresponding labels are determined to generate sample data and corresponding labels, which are then input into a machine learning classification model for classification.

[0092] Specifically, machine learning classification models can employ models such as XGBoost, LightGBM, and logistic regression. For example, feature vectors are used as X, and the corresponding labels are used as Y. These are then input into an XGBoost model for training, resulting in a converged model.

[0093] This invention selects alarm types using the XGBoosting algorithm, performs label encoding after selection, and then designs time series encoding. The generated sequence is then fed into a Bi-LSTM network for training. This solves the problem of manually extracting too many features in XGBoost and also addresses the issue of incomplete association rules and poor performance in Fpgrowth.

[0094] S47: Determine the weight values ​​corresponding to each alarm type based on the output of the machine learning classification model.

[0095] The machine learning classification model described above can classify different alarm types and determine the weight values ​​corresponding to each alarm type based on the output of the machine learning model.

[0096] The solution provided in this application can automatically determine the weight values ​​corresponding to various alarm types based on historical alarm data using a machine learning model. These weight values ​​characterize the likelihood that a corresponding alarm type will lead to a subsequent base station alarm. Compared to manually setting weight values, the weight values ​​determined in this solution are based on objective historical alarm data, avoiding the problem of inaccurate weight values ​​due to human factors and improving the accuracy of model predictions.

[0097] Based on the solutions provided in the above embodiments, optionally, such as Figure 5 As shown, if there are multiple alarm types within the target time window, step S43 above includes:

[0098] S51: Perform single-hotspot coding on each alarm type within the target time window.

[0099] Single-hotspot coding, also known as one-bit valid coding, uses an N-bit status register to encode N states. Each state has an independent register bit, and only one bit is valid at any given time. In practical applications, multiple base station alarms may occur within a day, and the same type of alarm may also occur multiple times on the same day. In this step, single-hotspot coding can be performed on each alarm type within a day.

[0100] S52: Determine the alarm coding vector corresponding to the target time window based on the cumulative result of the single hotspot coding of multiple alarm types in the target time window.

[0101] After performing single-hotspot coding on multiple alarm types within a day, the resulting single-hotspot codes are summed to obtain the alarm type code for that day.

[0102] The scheme provided by the embodiments of this application determines the alarm type code corresponding to a time window based on single hotspot coding, which enables the alarm coding vector to accurately represent the base station alarm type that occurs within the corresponding target time window.

[0103] Based on the solutions provided in the above embodiments, optionally, such as Figure 6 As shown, before step S14 above, the following steps are also included:

[0104] S61: Select at least one key alarm type from the multiple alarm types according to the weight values ​​corresponding to the multiple alarm types, wherein the weight value of the selected key alarm type is greater than the weight value of the unselected alarm type.

[0105] In this step, multiple alarm types are filtered, and at least one key alarm type is selected from the multiple alarm types based on weight values. This allows for targeted characterization of key alarms in the subsequently generated sample data.

[0106] Step S14 above includes:

[0107] S62: Using the feature sequence as sample data, and using whether the base station outage alarm of the key alarm type occurs within the first time window after the N consecutive time windows as a label, input the bidirectional long short-term memory network model to perform training.

[0108] The following is an example of this application, taking a 4G base station as an example. Its important outage alarms include base station outage alarm - network element connection interruption and cell outage alarm - cell unavailability alarm:

[0109] Based on the pre-acquired historical alarm data, after traversing all alarm data, the alarm type numbers are obtained as follows:

[0110]

[0111]

[0112] The alarms from the base station from July 1st to July 7th were extracted, and the alarm vectors for each day were generated with a one-day time window as follows:

[0113] The alarm vector V1 for July 1st is: 1, 1, 1, 1, 1, 1, 1, 0, 0, 0

[0114] The alarm vector V2 on July 2nd is: 0, 0, 1, 0, 0, 1, 1, 1, 0, 0

[0115] The alarm vector V3 on July 3rd is: 1, 1, 0, 0, 0, 1, 1, 0, 0, 0

[0116] The alarm vector V4 for July 4th is: 1, 0, 0, 0, 0, 1, 1, 1, 0, 0

[0117] The alarm vector V5 for July 5th is: 1, 0, 0, 1, 1, 1, 1, 0, 0, 0

[0118] The alarm vector V6 for July 6th is: 1, 0, 0, 1, 1, 1, 0, 0, 0, 0

[0119] The alarm vector V7 on July 7th was: 0, 1, 0, 0, 1, 1, 1, 1, 0, 0

[0120] Based on the preset attenuation coefficient a = 0.5, the feature vector V corresponding to July 1st to July 7th is determined. all =V1×0.5 6 +V2×0.5 5 +……+V7. There were no network element connection interruptions or cell unavailability alarms on the following day, July 8th, so it is marked as 0.

[0121] Then, using a 1-day sliding window, we continue to acquire vectors and labels.

[0122] The alarm vector V8 on July 8th was: 1, 0, 1, 0, 0, 1, 1, 0, 0, 0

[0123] The feature vector V corresponding to July 2nd to July 8th all =V2×0.5 6 +V3×0.5 5 +……+V8

[0124] On July 9th, some communities had no alarms available, so they were marked as 1.

[0125] By analogy, all accumulated alarm vectors can be obtained, serving as sample data and corresponding labels. The accumulated alarm vectors V obtained above are then used as the basis for further analysis. all As sample data, the corresponding labels are fed into the XGboost model for training.

[0126] After training converges, the `get_score(importance_type='total_gain')` method can be used to obtain the top alarms with the highest importance. In this example, the top 5 alarms with the highest importance are selected, and the order of importance is assumed to be as follows:

[0127] Alarm Types serial number Network element connection interruption 9 Community unavailable alarms 10 [Derivative Alarm] VVIP Cell LTE Wireless Disconnection Rate Sudden Abnormality Alarm 2 [Derivative Alarm] VVIP Cell VoLTE Downlink Packet Loss Rate Sudden Abnormality Alarm 1 User plane bearer link failure alarm 4

[0128] Subsequently, the above-mentioned important alarm types were renumbered as follows, and this number can represent the importance of the corresponding alarm type:

[0129] Alarm Types serial number Network element connection interruption 1 Community unavailable alarms 2 [Derivative Alarm] VVIP Cell LTE Wireless Disconnection Rate Sudden Abnormality Alarm 3 [Derivative Alarm] VVIP Cell VoLTE Downlink Packet Loss Rate Sudden Abnormality Alarm 4 User plane bearer link failure alarm 5

[0130] Subsequently, an alarm sequence is generated with a 2-hour time slot width. The alarm sequence for July 1st is as follows: 0, 0, 0, 0, 3, 0, 5, 0, 0, 0, 0, 0

[0131] The daily vector length is 12, and so on, to obtain the alarm vector for 7 days, with a total length of 84.

[0132] Using a one-day sliding window, all alarm sequence vectors and their labels can be obtained. Feeding the obtained sequences into a Bi-LSTM network yields the final trained model.

[0133] Based on the solutions provided in the above embodiments, optionally, such as Figure 7 As shown, if there are multiple alarm types within the target time interval, step S22 above includes:

[0134] S71: Select an alarm type as the alarm type corresponding to the target time interval based on the weight values ​​of the various alarm types within the target time interval, wherein the weight value of the selected alarm type is greater than the weight value of the unselected alarm type.

[0135] Based on the alarm type weights determined in the above examples, in this step, when selecting the alarm type corresponding to the target time interval, one alarm type is selected as the alarm type corresponding to the target time interval based on the weight values ​​corresponding to the various alarm types. The selected alarm type can accurately characterize the relatively important base station alarms occurring within the corresponding time interval, thereby accurately expressing the characteristics of base station alarms occurring in the target time interval, and thus optimizing the effect of the training model.

[0136] In the scheme provided in this application embodiment, multiple alarm types obtained after alarm type selection are numbered according to their importance weight. Based on historical alarm data obtained from appropriate historical time periods, the length of the extracted historical time period is determined comprehensively based on the number of base stations and the number of alarms. Then, for a certain base station, all alarms are extracted in a time window of N days. Each day is further divided into multiple time intervals, with the length of each time interval set to T minutes. Therefore, there are N*1440 / T time intervals in N days. The length of the sequence is the number of time intervals. Subsequently, the time interval of each alarm is determined according to the alarm occurrence time of each alarm. Only important alarm types are selected. If there are multiple alarm types in a time interval, the alarm type number with the highest importance is used as the alarm number of that time interval. If there are no important alarms in that time interval, the alarm number of that time interval is 0. Thus, an alarm feature sequence Seq with a length of N*1440 / T is obtained.

[0137] Subsequently, alarms for the following day after obtaining the above alarm feature sequence Seq are selected and labeled. If a significant service outage alarm occurs within that day, it is marked as 1; otherwise, it is marked as 0. This completes the labeling of a Seq sequence. Using a 1-day sliding time window, all historical alarms from base stations are traversed based on the above steps, generating corresponding alarm feature sequences Seq and their labels. All alarm sequence Seq and their labels are generated. The generated alarm type sequences Seq are used as training samples, and the corresponding labels are used as tags, and fed into the Bi-LSTM model for training. After training converges, the corresponding model is obtained.

[0138] The scheme provided in this application embodiment enables base station outage alarm prediction using a Bi-LSTM network. Specifically, generating an alarm superposition vector by superimposing attenuation coefficients for alarm type selection improves the accuracy of feature representation, thereby optimizing model training performance. Dividing each day into time intervals to generate equal-length alarm feature sequences facilitates model training. Compared to schemes requiring significant manual intervention and the extraction of large amounts of alarm features, this method offers the following advantages:

[0139] 1) It requires less subjective feature extraction and can automatically extract features using neural networks.

[0140] 2) It has the ability to detect the timing characteristics in alarm events, which can improve the accuracy of base station outage prediction.

[0141] 3) It can eliminate some invalid alarms during training, thereby improving training efficiency.

[0142] To address the problems existing in the prior art, this application also provides a training device 80 for a base station outage alarm model, such as... Figure 8 As shown, it includes:

[0143] The first acquisition module 81 acquires first historical data, which includes the alarm time when the base station outage alarm occurred and the corresponding alarm type.

[0144] The second acquisition module 82 divides the first historical data into time windows and acquires the alarm time and corresponding alarm type of the base station outage alarm that occurred within the time window.

[0145] The generation module 83 generates a feature sequence based on the alarm type codes corresponding to the alarm types of N consecutive time windows. The alarm type codes in the feature sequence are arranged in chronological order based on the corresponding alarm times, where N is an integer greater than 1.

[0146] Training module 84 uses the feature sequence as sample data and the label of whether a base station outage alarm occurs within the first time window after the N consecutive time windows. It inputs the data into the bidirectional long short-term memory network model for training to obtain the base station outage alarm model.

[0147] The device provided in this embodiment trains a Bi-LSTM model based on historical alarm data. After the model converges, it can be used to predict whether a base station outage alarm will occur in the future. In practical applications, the output of the trained Bi-LSTM model represents the probability of a future base station outage alarm. Automatic judgment can be achieved based on the magnitude of a preset baseline probability and the probability output by the model. If the probability output by the model is greater than the preset baseline probability, it can be determined that a base station outage alarm will occur in the future. Subsequently, an alarm can be issued in advance according to the actual situation, and potential faults can be handled in advance through manual maintenance to avoid the occurrence of base station outage alarms, reduce the negative impact of base station outage alarms, and thus improve customer experience.

[0148] In this application, the modules in the apparatus provided can also implement the method steps provided in the method embodiments. Alternatively, the apparatus provided in this application may include other modules besides those described above to implement the method steps provided in the method embodiments. Furthermore, the apparatus provided in this application can achieve the technical effects achievable by the method embodiments.

[0149] Preferably, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the various processes of the above-described training method embodiment for a base station outage alarm model and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0150] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, this computer program implements the various processes of the training method embodiment for a base station outage alarm model described above, and achieves the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0151] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0152] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0153] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0154] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0155] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0156] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0157] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0158] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0159] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A training method for a base station outage alarm model, characterized in that, include: Obtain first historical data, which includes the alarm time and corresponding alarm type when the base station outage alarm occurred; The first historical data is divided into time windows, and the alarm time and corresponding alarm type of the base station outage alarm within the time window are obtained. A feature sequence is generated by encoding the alarm types corresponding to the alarm types of N consecutive time windows. The alarm type codes in the feature sequence are arranged in chronological order based on the corresponding alarm times. The feature sequence represents the type and timing of alarms occurring within the N consecutive time windows, where N is an integer greater than 1. Using the feature sequence as sample data, and labeling whether a base station outage alarm occurs within the first time window after the N consecutive time windows, the data is input into a bidirectional long short-term memory network model for training to obtain a base station outage alarm model.

2. The method as described in claim 1, characterized in that, A feature sequence is generated by encoding the alarm types corresponding to N consecutive time windows, including: The total duration of the N consecutive time windows is divided according to a preset time length to obtain multiple time intervals; The alarm types corresponding to the multiple time intervals are determined based on the alarm times corresponding to the alarm types of the N consecutive time windows. Based on the preset correspondence between alarm types and alarm type codes, the alarm type codes corresponding to the multiple time intervals are arranged in chronological order to generate the feature sequence.

3. The method as described in claim 2, characterized in that, If there are multiple alarm types within the target time interval, the alarm types corresponding to the multiple time intervals are determined based on the alarm times corresponding to the alarm types of the N consecutive time windows, including: Determine the weight values ​​corresponding to the various alarm types within the target time interval; The alarm type with the highest weight value is determined as the alarm type corresponding to the target time interval.

4. The method as described in claim 3, characterized in that, Before generating the feature sequence using alarm type encodings corresponding to alarm types across N consecutive time windows, the process also includes: Obtain second historical alarm data, which includes the alarm time when the base station outage alarm occurred and the corresponding alarm type; The second historical alarm data is divided according to the time window, and the alarm type of base station outage alarm that occurred within the time window is obtained; Single hotspot encoding is performed on the alarm types of M consecutive time windows to obtain M alarm encoding vectors corresponding to different time windows, where M is an integer greater than 1; Determine the attenuation coefficients corresponding to the M alarm coding vectors based on the time order of the time window; The feature vector is determined by multiplying the M alarm coding vectors and their corresponding attenuation coefficients; Using the feature vector as sample data, and the label of whether a base station outage alarm occurred within the first time window after the M consecutive time windows, the data is input into a machine learning classification model for classification. The weight values ​​corresponding to each alarm type are determined based on the output of the machine learning classification model.

5. The method as described in claim 4, characterized in that, If there are multiple alarm types within the target time window, single-hotspot encoding is performed on the alarm types of M consecutive time windows to obtain M alarm encoding vectors corresponding to different time windows, including: Single hotspot coding is performed on each alarm type within the target time window; The alarm coding vector corresponding to the target time window is determined based on the cumulative result of the single hotspot coding of various alarm types within the target time window.

6. The method as described in claim 4, characterized in that, Using the feature sequence as sample data, and labeling whether a base station outage alarm occurred within the first time window after the N consecutive time windows, before inputting it into the bidirectional long short-term memory network model for training, the following steps are also included: At least one key alarm type is selected from the multiple alarm types based on the weight values ​​corresponding to the multiple alarm types, wherein the weight value of the selected key alarm type is greater than the weight value of the unselected alarm type; The training process involves using the feature sequence as sample data and labeling whether a base station outage alarm occurred within the first time window after the N consecutive time windows. The training includes inputting the data into a bidirectional long short-term memory network model. Using the feature sequence as sample data, and labeling whether a base station outage alarm of the key alarm type occurs within the first time window after the N consecutive time windows, the bidirectional long short-term memory network model is input for training.

7. The method as described in claim 4, characterized in that, If there are multiple alarm types within the target time interval, the alarm types corresponding to the multiple time intervals are determined based on the alarm times corresponding to the alarm types of the N consecutive time windows, including: Based on the weight values ​​corresponding to the various alarm types within the target time interval, one alarm type is selected as the alarm type corresponding to the target time interval, wherein the weight value of the selected alarm type is greater than the weight value of the unselected alarm type.

8. A training device for a base station outage alarm model, characterized in that, include: The first acquisition module acquires first historical data, which includes the alarm time when the base station outage alarm occurred and the corresponding alarm type. The second acquisition module divides the first historical data into time windows and acquires the alarm time and corresponding alarm type of the base station outage alarm that occurred within the time window; The generation module generates a feature sequence based on the alarm type codes corresponding to the alarm types of N consecutive time windows. The alarm type codes in the feature sequence are arranged in chronological order based on the corresponding alarm times. The feature sequence represents the type and timing of alarms occurring within the N consecutive time windows, where N is an integer greater than 1. The training module uses the feature sequence as sample data and the label of whether a base station outage alarm occurs within the first time window after the N consecutive time windows. It inputs the data into the bidirectional long short-term memory network model for training to obtain the base station outage alarm model.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 7.

Images