WO2021101490A1

WO2021101490A1 - Network failure prediction module and the method performed by this module

Info

Publication number: WO2021101490A1
Application number: PCT/TR2020/051133
Authority: WO
Inventors: Ali GÖKÇE; İren Berk ÖZALP; Mehmet Yavuz ALPTÜRK; Halil ERTAN
Original assignee: Türk Telekomüni̇kasyon Anoni̇m Şi̇rketi̇; İnnova Bi̇li̇şi̇m Çözümleri̇ Anoni̇m Şi̇rketi̇
Priority date: 2019-11-20
Filing date: 2020-11-19
Publication date: 2021-05-27

Abstract

The invention relates to a prediction module that can be used by the operation units in all networks without distinguishing between fixed, mobile, and broadband, using machine learning methods for detecting failures in the networks of telecommunications operators in advance and the method implemented by this module.

Description

NETWORK FAILURE PREDICTION MODULE AND THE METHOD PERFORMED BY THIS MODULE

Technical Field The invention relates to a prediction module that can be used by the operation units in all networks without distinguishing between fixed, mobile, and broadband, using machine learning methods for detecting failures in the networks of telecommunications operators in advance and the method implemented by this module. Prior Art

Today, prediction modules include only the device information that the failure will occur or the cause of the failure following that. In this case, the operation units cannot benefit from the results of the prediction module sufficiently. Because many devices in the network have sub - diffractions such as hundreds of shelves / slots / ports. With the prediction results that do not contain this sub - diffraction information, the operation units cannot have the opportunity to intervene in the device beforehand, even if the device that the failure will occur can be predicted in advance.

In addition, existing prediction modules create a separate model for each device and create a device - specific prediction report. This situation poses a problem in terms of the operating performance of the module since there are generally hundreds / thousands of devices in the networks of the telecommunication operators.

Furthermore, current approaches embody the unsupervised classification methodology by analyzing only the hidden patterns within the historical alarm data as the data source or generate prediction reports by applying the supervised classification methodology using only the historical performance data of the devices and the alarm data for labeling. The fact that different types and structures of data sources such as alarm and performance data cannot be used together in the same study emerges as a factor that directly affects the success of the prediction module. As a result, due to the above - mentioned drawbacks and the inadequacy of the existing solutions, an improvement in the technical field has been required.

The Purpose of Invention

The invention is inspired by the existing circumstances and aims to solve the above - mentioned drawbacks.

By means of the invention, a two - layer artificial intelligence solution method has been applied to share the sub - diffraction information of the device such as shelf / slot / port with the operation units in addition to the device information. In the first layer, a device - based prediction model runs and calculates the probability of the failure type that the prediction module works on to occur on each device in the next time frame. In the second layer, an anomaly model is created through features that contain sub - diffraction information for devices that are expected to fail in the first layer. According to this anomaly model, the sub - diffraction information with the most outlier value is shared as the sub - diffraction information where the failure will occur. With this method, it is intended that the operations units will intervene in a situation in a shorter time before the failure occurs, with the sub - diffraction information on a device that predicts a direct failure to occur.

Thanks to the invention, a different prediction module is designed for each failure type, regardless of the number of devices. In this way, a single prediction module is created for predicting the type of failure in the network. This prediction module is designed to be trained and to predict for all devices in the network. This approach improves the operating performance of the prediction module and reduces the complexity of the solution. At the same time, it allows the prediction of the first failure that will occur on devices that have never failed before.

By means of the invention, the historical alarm data, which has an event - based structure, and the performance data containing the past performance metrics of the devices are used together as a data source and the supervised classification methodology is applied. The module is designed to expand horizontally via the addition of new data sources. With the alarm device feature extraction module, valuable attributes of the devices are extracted from the historical alarm data. The extracted features are blended together with statistical simulations of the metrics on the performance side of the same device at the time diffraction level of the solution and the training set of the device is created. In this way, the training set consists of features extracted from different type data sets, such as alarm and performance data. In this way, the prediction success percentage of the prediction module is increased. Again, thanks to the invention, it is possible to work with dynamic values for the estimation range and a detailed prediction result report with high prediction accuracy. In this way, it is aimed to increase customer satisfaction by ensuring that telecommunications operators manage their operational units more efficiently and anticipate and eliminate failures that will affect the customer in advance. The structural and characteristic features and all advantages of the invention will be understood clearly through the drawings below and the detailed description made by referring to these figures, therefore the evaluation should be made by taking these figures and detailed explanations into consideration.

Brief Description of the Figures Figure 1 is a schematic block diagram of the network failure prediction module of the invention.

Figure 2 is a schematic block diagram representing the connections of the network failure prediction module of the invention with the operator units.

Reference Numbers 1. Operations Unit

2. Alarm data

3. Performance data

4. Compacting module

5. Initial extraction module 6. Second extraction module

7. Labeling module

8. Classification module

9. Third extraction module 10. Fourth extraction module

11. Anomaly module

12. Training set

13. Prediction model

14. Anomaly model

15. Failure prediction report Detailed Description of the Invention

In this detailed description, the network failure prediction module of the invention and the preferred embodiments of the method performed by this module are explained only for a better understanding of the subject.

In the network failure prediction module of the invention, which is shown in Figures 1 and 2 as a schematic block diagram, by using the event - based historical alarm data (2) and the past performance data of the devices (3), the malfunctions that will occur in devices and sub - diffractions are predicted with a certain percentage of success together with the failure reasons. The generated failure prediction report (15) is also shared with the operation units (1).

Alarm data (2) and performance data (3) are monitored by telecommunications operators using standard methods. In these methods, alarm data (2) and performance data (3) are generally stored in relational databases. In the invention, the past alarm data (2) and performance data (3) are read, pre - cleaned and saved in the database.

In the alarm data (2), some alarms can be counted as noises, and some are repetitive. With the compacting module (4), alarm data (2) are deduplicated, freed from repeated alarms, and stored in a separate table. The self - repetition numbers of alarms continue to be kept as separate information in the converted new format. The first extraction module (5) extracts many features by working on the new deduplicated format of the alarm data (2). This initial extraction module (5) performs a number of analyses to determine the associated alarms and alarm types that will cause failure. In this way, a list of related alarms and certain alarm types for the predicted failure are determined and this information is used during feature extraction. These features include values such as the total number of alarms that occur on the same device, the total number of alarms that occur in the same region due to the same failure, the number of associated alarms, the number of occurrence of certain alarm types, and how often these alarms repeat themselves.

At the same time, the second extraction module (6) runs in the prediction module and extracts some values such as average, maximum, minimum and standard deviation of the metrics of the devices in the performance data (3) as features.

After the initial extraction module (5) and the second extraction module (6) that extract the features, the data labeling module (7) runs. With this labeling module (7), labels are created for the training set (12) according to the prediction range by using the past failure dates in the alarm data (2). A device fault prediction training set (12) is created along with the extracted alarm and performance features, and labels. The features extracted by the initial extraction module (5) and the second extraction module (6) are device - independent. For this reason, training sets (12) and models are created according to the failure types without the need for a separate model for each device.

On the created training set (12), the supervised device classification module (8) works. In addition to creating a prediction model (13), this classification module (8) also addresses processes such as tackling with categorical data, normalization of data, and eliminating the imbalance between class ratios in the data before creating a prediction model (13). At the end of the process, the classification module (8) creates a prediction model (13) for devices that are likely to experience failures using artificial intelligence algorithms.

The prediction of only the device that will experience a failure is not considered sufficient by the operation units (1 ). Each device may consist of sub - diffractions such as hundreds of shelves / slots / ports. Without such device sub - diffraction information, the possibility of intervening in failures in advance is quite limited. For this reason, the third extraction module (9) derives features from the alarm data (2) and the fourth extraction module (10) from the performance data (3) regarding the sub - diffraction of the devices.

While some alarms occur directly on the devices in the alarm data (2), some alarms occur on the sub - diffractions of the devices. The third extraction module (9) uses alarms occurring on the sub - diffraction of the devices. Performance data (3) similarly includes performance metrics for the sub - diffraction of the devices. The fourth extraction module (10) is also used to derive features using these performance metrics. The anomaly module (11 ), which uses extracted features to describe the behavior of the sub - diffraction of devices, creates the device sub - diffraction anomaly model (14). This anomaly model (14) works by modeling the behavior of sub - diffractions of the devices to determine the sub - diffractions of the device in case of anomaly.

In the prediction module of the invention, along with following similar steps during prediction, it uses the previously created device prediction model (13) and the device sub - diffraction anomaly model (14). The newly observed alarm data (2) are processed by the compacting module (4) and the initial extraction module (5), respectively. The newly observed performance data (3) is similarly processed by the second extraction module (6). The generated features are combined into extracted datasets and provided as input to the device prediction model (13). For devices that are expected to have a malfunction on the output of this prediction model (13), firstly, the third extraction module (9) and the fourth extraction module (10) are run and the device sub - diffraction features of the alarm data (2) and performance data (3) are created. Afterwards, the created features and outputs of the prediction model (13) are provided as input to the anomaly module (11 ) and the anomaly model (14) is created. Thus, in addition to the device and the failure cause prediction, the sub - diffraction of the device that exhibits anomaly behavior within the relevant time period is shared with the operation units (1 ). The failure prediction report (15) shared with the operation units (1 ) includes information such as the device where the failure will occur, the type of the failure, the estimated date of the failure, the sub - diffraction information of the device where the failure will occur, and the probability of the failure to occur.

Claims

1. The invention is a prediction module that can be used by the operation units in all networks without making the distinction between fixed, mobile and broadband, and enables the detection of malfunctions in the networks of telecommunication operators in advance, characterized by comprising;

• a compacting module (4), which deduplicates alarm data (2) by ensuring that it is free from repeated alarms,

• an initial extraction module (5) that identifies the alarm types of the deduplicated alarm data (2) and extracts the information on the total number of alarms, the total number of alarms occurring in the same area due to the same failure, the number of associated alarms, the number of occurrences of certain alarm types and how often these alarms repeat themselves as features,

• the second extraction module (6) that extracts the average, maximum, minimum and standard deviation values of the metrics of the devices in the performance data (3) as features,

• the labeling module (7) that creates labels according to the prediction interval using the failure dates occurred in the past from the alarm data (2),

• a training set (12) consisting of features created with the initial extraction module (5) and the second extraction module (6), and labels created with the labeling module (7),

• the classification module (8) that creates the prediction model (13) from the training set (12),

• the third extraction module (9) that performs feature extraction of alarm data (2) by using the alarms incoming to the sub - diffractions of the devices,

• the fourth extraction module (10), which extracts features from performance data (3) using performance metrics for the sub - diffraction of the devices, • the anomaly module (11) that creates the anomaly model (14) that models the behavior of the sub - diffractions of the devices using the sub - diffraction features created with the third extraction module (9) and the fourth extraction module (10) and the outputs of the prediction model (13).

2. The invention is a prediction method that can be used by the operation units in all networks without making the distinction between fixed, mobile and broadband, and enables the detection of malfunctions in the networks of telecommunication operators in advance, characterized by comprising of the following steps;

• deduplication of alarm data (2) by removing repeated alarms with the compacting module (4),

• identification of the alarm types of the deduplicated alarm data (2) and extracts the information on the total number of alarms, the total number of alarms occurring in the same area due to the same failure, the number of associated alarms, the number of occurrences of certain alarm types and how often these alarms repeat themselves as features by the initial extraction module (5),

• extraction of the average, maximum, minimum, and standard deviation values of the metrics of the devices in the performance data (3) as features by the second extraction module (6),

• the labeling module (7) creates labels according to the prediction interval using the failure dates occurred in the past from the alarm data (2),

• generation of a training set (12) consisting of features created with the initial extraction module (5) and the second extraction module (6), and labels created with the labeling module (7),

• the classification module (8) creates a prediction model (13) from the training set (12),

• the third extraction module (9) makes feature extraction of the alarm data (2) by using the alarms occurring on the sub - diffractions of the devices, • the fourth extraction module (10) makes feature extraction of the performance data (3) by using the performance metrics of the sub - diffractions of the devices,

• the anomaly module (11) creates the anomaly model (14) that models the behavior of the sub - diffractions of the devices using the sub - diffraction features created with the third extraction module (9) and the fourth extraction module (10) and the outputs of the prediction model (13).