Netrca: an effective network fault cause localization

Original title ：NETRCA: AN EFFECTIVE NETWORK FAULT CAUSE LOCALIZATION ALGORITHM

Chinese translation ：Netrca: An effective algorithm for network fault location

Time of publication ：2022 year 2 month 23 Japan

platform ： ICASSP 2022

source ：DAMO Academy, Alibaba Group, Hangzhou, China

The article links ：https://arxiv.org/abs/2202.11269

Open source code ：

Abstract

         Locating the root cause of network failure is very important for network operation and maintenance . However , Due to the complex network architecture and wireless environment , And limited annotation data , It is challenging to pinpoint the real root cause . In this paper , We propose a new algorithm NetRCA To deal with this problem . First , We Extract effective derived features from original data , Think about time 、 Direction 、 Attributes and interaction characteristics . secondly , We use Multivariate time series similarity and label propagation Methods , Generate new training data from tagged and unlabeled data , To overcome the shortage of labeled samples . Again , We Designed a combination XGBoost、 Rule set learning 、 The integrated model of attribute model and graph algorithm , To take full advantage of all data information , Improve performance . Last , Yes ICASSP 2022 AIOps The real data set of the challenge is tested and analyzed , To prove the superiority and effectiveness of our method .

        Keywords— Root cause analysis , Data to enhance , The time series , Integrated model , Wireless network

4. Conclusion

         This paper presents a new algorithm NetRCA To locate the root cause of network failure . In addition to the carefully designed feature Engineering , Our algorithm also uses data enhancement to generate new training data , To overcome the problem of insufficient labeled samples . Besides , We designed an integration method , Effectively combine different models , Carry out accurate and reliable causal reasoning for network faults .

2. Proposed network RCA frame

2.1. The framework outlined

         Proposed NetRCA The algorithm mainly includes three steps :1) Feature Engineering ,2) Data to enhance ,3) Model integration .NetRCA The frame of is as shown in the figure 1 Shown , It will be elaborated in the following sections .

chart 1: Proposed NetRCA Algorithm framework .

2.2. Feature Engineering

         Because the number of timestamps in each sample is different , Using all the time stamps directly to train the model may cause deviation , That is, the model is likely to focus on samples with more timestamp indexes . therefore , We train our model based on the features extracted from each sample . The generated features can be roughly divided into four categories : Time characteristics 、 Direction dependent features 、 Attribute characteristics and interaction characteristics .

         Some used in our model Time characteristics are based on data statistics , The data in each timestamp is assumed to be independent , Include Average 、 minimum value 、 Maximum 、 Median 、 Deciles and skewness （ skewness）. We also include some features that represent the shape of time series , Including the use of public tsfresh Package generated The number of peaks and the average value of changes .

        5G Multi antenna and multi beam forming are used in the network （ beamforming） To enhance performance [14,15]. Such as [16,17] Described , The beamforming direction and the distance between each node play an important role in the network performance , This is also important for root localization , Especially in AIOps Challenge In the provided cause and effect diagram , Root cause 2 And root cause 3 Between . We think the characteristics 20 In detecting the root cause 2 and 3 An important feature in the model . Due to features 20 Each node is given from 0 To 31 The index of , Map to a 4x8 Position matrix of , We first convert the index of each node into two-dimensional coordinates , Then the distance between each pair of nodes is measured by Euclidean distance . after , We derive the characteristic X and Y The characteristics of their interrelationship , The characteristics are 61/69/77/85 And characteristics 28/36/44/52, In order to further improve . Last , The statistical characteristics are summarized from the distance distribution of each time slice sample ( Mean value 、 variance 、 Quantiles, etc. ) For model training .

         According to the cause and effect diagram , Division feature   0 Outside , All nodes export attribute characteristics . As described in the problem description , These root causes ultimately lead to characteristics 0 The value of is low . that , The real root cause and its progeny factors are characteristic 0 The current value will have a greater impact than other factors . therefore , We generate a new feature , As each feature in the prediction feature 0 An estimate of the importance score on , The detailed derivation is in 2.4 Section gives .

         Generated X and Y The second order interaction characteristics of . Due to features X It's equal to characteristic Y The ratio to some unknown factor , We generate features X / Y To measure the impact of these unknown factors . Specially , Let's start by describing the problem X and Y Group the features in . For every pair of , We calculated X / Y The ratio of the . Last , We calculate the statistics for these ratios , Just like we calculate the characteristics of time .

To be continued