🤖 machine learning

Causal Intervention Sequence Analysis for Fault Tracking in Radio Access Networks

This paper presents an AI/ML pipeline for Radio Access Networks that proactively prevents Service-Level Agreement breaches by identifying root-cause indicators and their precise causal sequence through labeled data analysis and Monte Carlo validation.

Original authors: Chenhua Shi, Joji Philip, Subhadip Bandyopadhyay, Jayanta Choudhury

Published 2026-07-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chenhua Shi, Joji Philip, Subhadip Bandyopadhyay, Jayanta Choudhury

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the internet as a massive, invisible city of data, buzzing with billions of conversations happening every second. In this city, the "Radio Access Network" (RAN) is the bustling street corner where your phone connects to the internet. Sometimes, traffic jams happen, or a streetlight flickers out, causing a "Service-Level Agreement" (SLA) breach—basically, the promise that your internet will be fast and reliable gets broken. For a long time, fixing these jams was like trying to find a single dropped coin in a hurricane; engineers had to look at blurry, slow-motion snapshots of the city (data collected every 15 minutes) and guess what went wrong, often needing a human to label the mess. But what if we could watch the city in high-definition, second-by-second, and see exactly which car hit the first pothole, causing the chain reaction? This is the world of "causal inference," a way of figuring out not just what happened, but why it happened and in what order. It's the difference between knowing a cake burned and knowing that the oven was set too hot, the timer was ignored, and the door was left open, all in a specific sequence.

Enter a team of researchers from Ericsson who decided to build a super-smart detective to solve these network mysteries. They realized that old tools were too slow and too blurry to catch the real culprits before customers even noticed the internet was slow. So, they created a new AI pipeline that acts like a time-traveling investigator. Instead of waiting for the 15-minute "blurry photo" to show a problem, this system dives into the high-speed, high-definition data (milliseconds and seconds) to spot the very first sign of trouble.

Here's how their detective works: First, it learns what "normal" looks like by studying the network when everything is running smoothly. Then, when a problem starts, it doesn't just scream "Something is wrong!" It asks, "Which specific indicator changed first?" and "What did that change cause next?" The system uses a three-step process. It starts by finding the "intervention indicators"—the specific variables that got pushed out of their normal behavior, like a traffic light turning red when it should be green. Next, it builds a mini-map, or a "causal subgraph," showing how these specific variables are connected, ignoring the noise of the rest of the city. Finally, it uses statistical tests (like the Kolmogorov-Smirnov test and Z-score) to trace the exact timeline of events, pinpointing the very first domino that fell.

The researchers tested this idea using real network data from a common problem: cell load issues, where too many people try to download at once, causing speeds to drop below 500 kbps. They compared their new method against an older, well-known technique called PCMCI. The results were clear: the old method got confused by the noise and couldn't figure out the order of events, while the new system successfully identified the chain of cause-and-effect. For instance, it could tell that a specific type of resource utilization (like PDCCH CCE Utilization) spiked first, which then led to a drop in throughput.

To make sure their detective wasn't just getting lucky, the team ran thousands of "Monte Carlo simulations." Think of this as running the same mystery a hundred times with slightly different random conditions to see if the detective always finds the same culprit. They found that by tweaking a few settings (like how many variables to look at at once), the system became incredibly reliable at identifying the true root causes. They discovered that certain metrics, like "RRC Connected Users DL" and "CCE Utilization AVG," were the most likely suspects in these traffic jams.

The beauty of this approach is that it doesn't need supercomputers or massive amounts of memory; it's lightweight enough to run on standard equipment, making it cheap and energy-efficient. It turns a chaotic flood of data into a clear, step-by-step story that engineers can actually understand and trust. Instead of reacting to a broken internet connection after a customer complains, this system allows operators to see the trouble brewing and fix it before anyone even notices. It's a shift from playing catch-up to staying one step ahead, ensuring that the digital city keeps humming along smoothly.

Technical Summary: Causal Intervention Sequence Analysis for Fault Tracking in Radio Access Networks

Problem Statement
Modern Radio Access Networks (RAN) face increasing complexity in fault diagnosis, particularly regarding Service-Level Agreement (SLA) breaches. Traditional approaches rely on low-resolution, aggregated data (e.g., 15-minute granularity) and supervised machine learning models that require extensive human intervention for ground-truth annotation. This manual labeling process is a critical bottleneck; human experts cannot efficiently analyze the massive volume of high-resolution (millisecond to sub-millisecond) trace and error logs required to isolate root causes before an SLA violation occurs on the aggregated 15-minute data. Existing causal discovery algorithms, such as Differentiable Adjacency Test (DAT-Graph) and PCMCI variants, often struggle with "hard interventions" (where a KPI is pinned to a constant during a fault) or fail to explain the specific temporal order in which events unfold, providing either a list of candidates or a static graph but not the dynamic sequence of failure.

Methodology
The authors propose an automated, high-resolution analysis framework designed to identify the causal intervention sequences leading to SLA breaches. The approach integrates three functional components:

Root Cause Discovery (RCD): This component identifies variables that undergo intervention and are causally related to SLA indicators. It employs a modified Peter-Clark (PC) algorithm under the hood, utilizing "Initial Partitioning with Localized Learning" and "Hierarchical Refinement." The process involves:
- Defining a binary indicator ( $F_t$ ) for hard intervention failures.
- Performing causal conditional independence tests to identify direct effects.
- Refining the causal graph by identifying parent nodes and removing false positives via conditional independence tests.
- Checking for distributional consistency across observational and interventional data.
Causal Subgraph Construction: Once intervention indicators are identified, the system constructs a new causal subgraph. This subgraph emphasizes the causal connections between the reduced subset of intervention variables and the variables contributing to the SLA condition, focusing on the specific causal structure of the abnormal state.
Deviation Detection and Sequence Ordering: The final component applies univariate deviation detection on the time series data of the selected intervention variables. It uses the Kolmogorov-Smirnov (K-S) test to measure statistical inequality between normal and abnormal state distributions and Z-Score analysis to determine the direction of deviation. This allows the system to capture the temporal anomaly events, trace the sequence of intervention incidents, and identify the earliest event as the most likely root cause.

To optimize the RCD algorithm, the authors utilize Monte Carlo simulations. They treat the number of randomly selected nodes ( $g$ ) and the number of experimental iterations ( $n$ ) as parameters. By analyzing the convergence of the estimated probability ( $p$ ) that a Key Performance Indicator (KPI) is a causal source, they determine optimal values for $g$ and $n$ to ensure robust identification of root causes across different KPIs.

Key Contributions

Automated High-Resolution Analysis: The framework operates on seconds and milliseconds granularity data, enabling the isolation of root cause candidates prior to the aggregation of data into 15-minute intervals.
Causal Sequence Discovery: Unlike methods that only identify root cause candidates, this approach reveals the exact temporal order of events (the "causal intervention sequence") leading to an SLA breach.
Hybrid Algorithmic Approach: The integration of RCD (for candidate selection), causal subgraph construction, and statistical deviation detection (K-S test and Z-Score) addresses the limitations of existing tools that struggle with hard interventions and temporal ordering.
Parameter Optimization via Monte Carlo: The paper provides a systematic method for determining optimal algorithm parameters ( $g$ and $n$ ) to ensure the stability and reliability of root cause identification.

Experimental Results
The authors evaluated the proposed approach against the PCMCI benchmark using real-world LTE network data focused on "Cell Load Issues," where the target SLA breach was defined as Downlink Throughput dropping below 500 kbps.

Comparison with PCMCI: While PCMCI could identify leading indicators (e.g., PDCCH CCE Utilization, DL PRB Utilization), it struggled to capture the causal intervention sequence and was sensitive to noise, requiring significant dimensionality reduction and feature synthesis.
RCD Performance: The proposed RCD approach successfully identified the root cause nodes (e.g., RRC Connected Users DL, CCE Utilization AVG) and mapped the sequence of deviations.
Statistical Validation: Histograms and time-series analysis confirmed that the identified indicators experienced "hard" or "soft" interventions consistent with the abnormal state.
Monte Carlo Findings: The simulations demonstrated that as the number of iterations ( $n$ ) increased, the variance in the estimated probability of a KPI being a root cause converged. The study identified specific optimal $g$ and $n$ values for various KPIs, confirming that the approach can reliably distinguish prominent causal sources.

Significance and Claims
The paper claims that this framework shifts fault management from reactive troubleshooting to proactive prevention. By providing high-resolution, causally ordered insights, the system allows operators to spot real-world triggers well before customers experience service degradation. The authors emphasize the practical benefits of their approach:

Scalability and Efficiency: The method is computationally lightweight, requiring no GPUs or large memory footprints, making it suitable for edge deployment and micro-services.
Flexibility: It accepts any mix of KPIs without fixed input dimensions and allows engineers to define "abnormal" states using simple rules.
Explainability: The explicit chain of events and visual validation (e.g., histograms, sequence diagrams) enable domain experts to review and confirm root causes easily.
Minimal Tuning: The model requires only three hyperparameters, reducing the burden of configuration.

Ultimately, the authors assert that their approach drastically reduces the number of variables operators must monitor for early warning signs, transforming high-volume telemetry into actionable, trusted insights for faster fault resolution.

Technical Summary: Causal Intervention Sequence Analysis for Fault Tracking in Radio Access Networks

More like this