Causal Intervention Sequence Analysis for Fault Tracking in Radio Access Networks

I. INTRODUCTION

• The framework targets high-resolution analysis to identify ordered root causes and support remedies before 15-minute SLA violations.Traditional supervised models often use 15-minute aggregates and require human ground-truth annotation, which is impractical for the volume of millisecond-scale data.Source: §I · p. 1

II. RELATED WORK

• Sub-second data are too voluminous for engineers to sift through, and building clean labeled ground truth at scale is unrealistic.Earlier RAN fault-diagnosis tools [2] rely on 15-minute KPI aggregates.Source: §II · p. 1
• DAT-Graph and related methods [6] assume time-stable variable distributions, limiting their handling of hard interventions during faults.Hard interventions occur when a KPI is pinned to a constant; structural-causal theory calls these do-interventions.Source: §II · p. 1
• Anomaly pattern-mining tools can flag events but rarely explain why they unfold in a particular order.Source: §II · p. 2

A. Summary

• Three components identify intervention variables, construct a causal subgraph, and order temporal deviations to select the earliest likely root cause.The pipeline compares data preceding and following an SLA violation; RCD uses a modified PC algorithm to identify intervention variables among indicators causally related to the SLA under normal operation.Source: A. Summary · p. 2

B. Approach

• The procedure defines abnormal and preceding normal data, finds intervention indicators, builds their normal-state causal subgraph, then orders events using K-S tests and Z-scores.Likely root causes are selected based on the reliability and lead time of observed intervention events.Source: B. Approach · p. 2
• RCD combines initial partitioning and localized learning with hierarchical refinement to identify and refine root-cause candidates.The described steps include F-NODE definition, conditional-independence testing with Ψ-PC, score-based graph learning, parent-node refinement, and conditional-independence testing to remove false positives.Source: B. Approach · p. 2
• Monte Carlo runs estimate each KPI’s causal-source probability from its frequency of identification across repeated experiments for a given g.The method checks whether estimated variance falls as the number of experiments n increases, using that decrease as evidence that the probability estimate converges.Source: B. Approach · p. 3
• For each KPI, the proposed optimal g is the median-associated value, while optimal n corresponds to maximum variance reduction across tested n values.The method then consolidates per-KPI choices into shared g and n values for a single experiment across KPIs.Source: B. Approach · p. 3

• RCD arranges offending indicators into a fault-tracing hierarchy that domain experts can review and confirm.The framework captures a snapshot of indicators whose timing shifts stray from the normal causal pattern.Source: C. Benefits · p. 3
• Engineers can define abnormal periods in simple terms, and model tuning uses only three hyperparameters.Two hyperparameters are for RCD and one is for the K-S test.Source: C. Benefits · p. 3
• RCD accepts any mix of KPIs and can scale as networks add metrics.Unlike deep-learning models that need a fixed input dimension, RCD accepts variable feature sets.Source: C. Benefits · p. 3
• By sampling a small subset of variables, RCD runs quickly on data from thousands of cells.Source: C. Benefits · p. 3
• Before-and-after histograms of causal indicators give engineers a visual sanity check and support statistical testing.Source: C. Benefits · p. 3
• The approach drastically reduces the variables operators need to observe for early warning of SLA violations.Source: C. Benefits · p. 3

• PCMCI alone performs much worse than the proposed approach using RCD with Z-score and K-S anomaly detection.Source: §IV · p. 3
• The experiments contrast PCMCI with RCD and anomaly detection to compare their performance.Source: §IV · p. 3
• PCMCI is run separately on normal- and abnormal-operation data, with structural disparity used to identify deviation.Source: §IV · p. 3

• The study uses real-life LTE cell-load data, with Downlink Throughput below 500 kbps defined as the target SLA breach.Source: A. Dataset Curation · pp. 3–4
• KPIs are calculated from high-resolution UE and cell events and aggregated at 15-second granularity.Source: A. Dataset Curation · p. 4
• Normal and abnormal periods each last approximately 30 minutes and yield 240 data points across 60 KPIs per breach event.Source: A. Dataset Curation · p. 4

• PCMCI struggles to capture the causal intervention sequence for cell-load issues because it does not model counterfactuals or interventional distributions.Source: B. PCMCI on Cell Load Issue · p. 4
• PDCCH CCE Utilization and DL PRB Utilization are identified as leading indicators for load-based DL throughput.Figure 1 is labeled as PCMCI on the cell-load issue with normal state.Source: B. PCMCI on Cell Load Issue · p. 4, Figure 1 · p. 4
• Telecom-data noise can confuse PCMCI outputs, requiring domain knowledge for dimensionality reduction and feature synthesis.Source: B. PCMCI on Cell Load Issue · p. 4
• Changing p-value thresholds, conditional-independence tests, or time lags can produce different indicator sets and hinder interpretation.Source: B. PCMCI on Cell Load Issue · p. 4

C. RCD + Deviation Detection on Cell Load Issue

• RCD identifies root-cause indicators, while the K-S test orders leading indicators and Z-score reports the direction of deviation.The sequence is illustrated with root-cause nodes in yellow and ordered events in red dashed arrows.Source: C. RCD + Deviation Detection on Cell… · p. 4, Figure 3 · p. 4
• PRB utilization is fixed in the abnormal state, while the remaining KPIs shift in the opposite direction from their normal-state distributions.The paper characterizes PRB utilization as a hard intervention and the other KPI shifts as soft interventions.Source: C. RCD + Deviation Detection on Cell… · p. 5, Figure 4 · p. 5
• Figure 5 distinguishes normal from abnormal raw-data behavior and displays anomaly-start shifts, with 0 for no anomaly, -1 for below-threshold decreases, and 1 for above-threshold increases.The right panel shows the time-delay shift from univariate deviation detection.Source: Figure 5 · p. 5

D. Choice of different p-value under Kolmogorov-Smirnov Test

• Reducing the CIS p-value from 0.1 to 0.05 produces a different intervention sequence; 0.1 was chosen to tolerate noisy high-resolution abnormal-state data.Bonferroni and FDR corrections are applied to mitigate the higher false-positive risk associated with the larger p-value.Source: D. Choice of different p-value under… · p. 5
• The higher CIS p-value is intended to detect subtle but consistent shifts in the ordering of causal leading indicators in telecom data.The authors state that Bonferroni and FDR corrections help preserve statistical significance despite increased false-positive risk.Source: D. Choice of different p-value under… · p. 5
• Bonferroni and FDR corrections are applied to mitigate the increased false-positive risk associated with the higher p-value.Source: D. Choice of different p-value under… · p. 5

E. Monte Carlo Simulation

• The intersection of KPI subsets from multiple independent RCD runs is expected to reveal the true causal sources.Source: E. Monte Carlo Simulation · p. 5
• The authors use Monte Carlo simulation to generate experimental results that complement and support the theoretical analysis of RCD.Source: E. Monte Carlo Simulation · p. 5
• RCD uses input parameter g to randomly select nodes for building causal graphs.Source: 1) Experimental Setup · p. 5

1) Experimental Setup:

• The Monte Carlo setup varies g from 3 to 23 and uses six experiment sets with 10, 15, 20, 25, 30, 40, or 50 repetitions.For each fixed g and experiment set, the number of times each KPI appears as a causal source is counted.Source: 1) Experimental Setup · p. 5
• Estimated Pg,n variability for prominent causal-source KPIs steadily decreases as experiments increase from 10 to 50, supporting convergence to p across g.The proposed reliability check fits a variance trend for each g and accepts KPIs when 90% of fitted-line slopes are below zero.Source: 2) Experimental Results · p. 5, Figure 7 · p. 6

2) Experimental Results:

• The proposed g is associated with the median estimated p across g values 3–23, while n is chosen for the greatest variance reduction.For even numbers of estimates, the smaller of the two median-associated g values is selected; n is evaluated over 10, 15, …, 30, 40, 50.Source: 2) Experimental Results · pp. 5–6
• RRC Connected Users DL and CCE Utilization AVG are prominent causal sources, with source selection configurable by threshold or rank.Examples include estimated p > 0.4 with n > 0, or selecting the top 5% by p among KPIs with n > 0.Source: 2) Experimental Results · p. 6, Table I · p. 6

V. CONCLUSION

• The workflow combines Root-Cause Discovery, causal sub-graph analysis, and deviation detection to trace events from an initial anomaly to an SLA breach.It is presented as an AI/ML framework for identifying SLA-violation root causes in Radio Access Networks without heavy manual effort.Source: §V · p. 6
• The pipeline runs on standard CPUs and samples only some input variables, which the conclusion describes as energy- and cost-efficient for monitoring thousands of cells.Source: §V · p. 6
• Monte Carlo simulations confirm accuracy as the network scales, and a histogram-based check gives engineers a visual way to validate flagged indicators.The conclusion links the resulting insights to faster fault resolution and keeping subscribers connected.Source: §V · p. 6