💻 computer science

Adaptive Alarm Threshold Prediction in 4G Mobile Networks: A Percentile-Guided Deep Learning Framework with Interpretable Outputs

This paper proposes a percentile-guided deep learning framework, specifically the PCTN model, which automatically predicts adaptive alarm thresholds for 4G mobile networks by learning from live traffic patterns, thereby outperforming state-of-the-art models with fewer parameters while providing interpretable outputs that dynamically adjust to network conditions.

Original authors: Ayon Roy, Sadman Sharif, Shiva Prasad Sarkar

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ayon Roy, Sadman Sharif, Shiva Prasad Sarkar

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 a mobile network as a giant city made up of thousands of tiny neighborhoods called "cells." Each cell is responsible for keeping everyone in its area connected to the internet and phone lines. Sometimes, a neighborhood gets quiet, or the lights flicker on and off too much. When this happens, the network needs to know: Is this a real emergency, or just a normal lull?

In the past, network operators used a very rigid rulebook to decide this. They set a fixed "alarm threshold," like a speed limit sign that never changes.

The Problem: A speed limit of 20 mph is great for a quiet school zone at midnight, but it's terrible for a busy highway at rush hour. Similarly, a fixed alarm setting might ignore a real crash during a busy morning (because the noise is too high) or send an engineer to fix a "broken" cell at 3 AM when it's actually just sleeping (causing unnecessary trips).

This paper introduces a smart, adaptive system that acts like a seasoned traffic cop who knows exactly when to be strict and when to be lenient, based on the time of day, the day of the week, and how busy the neighborhood usually is.

The Core Idea: Learning the "Normal" to Spot the "Abnormal"

The researchers built a machine learning framework called PCTN (Percentile Guided Contextual Threshold Network). Here is how it works, using simple analogies:

1. The "No Manual" Problem
Usually, to teach a computer, you need a teacher with the right answers (labels). But in this case, no one knows the "perfect" alarm setting for every single moment in history. The paper solves this by creating its own "teacher."

The Analogy: Imagine trying to teach a student what "too much traffic" looks like without a textbook. Instead, the system looks at the last 10 days of traffic data and says, "Okay, 90% of the time, traffic is this light. If it gets heavier than that, then we sound the alarm." It derives the rules directly from the data's own behavior.

2. The Four Rules the System Learns
The system predicts four specific "rules" for when to trigger an alarm:

How long to wait: If a cell goes silent, how many hours should we wait before panicking? (e.g., "Wait 4 hours on a Tuesday, but only 2 hours on a Friday night.")
How long it can be silent: What is the maximum time a cell can be off before it's a problem?
Total flickering: How many times can the signal flicker on and off in a whole day?
Hourly flickering: How many times can it flicker in just one hour?

3. The "Smart Detective" (The PCTN Model)
The researchers tested four different types of "detectives" (AI models) to see which one could figure out these rules best.

The Old School Detective (XGBoost): Good at simple patterns, but gets confused when the weather changes (e.g., a sudden, extreme traffic day it has never seen before).
The Heavyweight Detective (iTransformer): A very powerful, complex model with a huge brain (lots of parameters). It does a good job but is slow and expensive to run.
The Hybrid Detective (CNN-BiLSTM): A mix of techniques that tries to spot patterns in time, but struggles with the specific quirks of this data.
The New Star (PCTN): This is the paper's invention. Instead of just guessing a number, PCTN acts like a statistician.
- First, it asks: "What does this neighborhood usually look like right now?" (It learns the average behavior).
- Second, it asks: "How much does it usually vary?" (It learns the volatility).
- Finally, it sets the alarm threshold based on those two answers.
- The Superpower: Because it understands the "average" and the "variation," it can explain why it set a specific threshold. If an engineer asks, "Why is the limit so high today?" the system can say, "Because today is a holiday and traffic is naturally chaotic."

4. Handling the "Floor" Problem
One specific challenge was that most cells rarely flicker at all. It's like trying to teach a computer to predict "how many times a car honks" when 94% of the time, the answer is "zero."

The Solution: PCTN uses a special "Bernoulli gate." Think of it as a two-step check. First, it asks, "Is there any honking at all?" If the answer is "No," it stops there. If "Yes," then it tries to guess how many times. This simple two-step trick made it much better at spotting rare, weird events than the other models.

The Results

When tested on real data from 10,644 cell towers across three different equipment manufacturers and nine regions:

PCTN won. It was the most accurate at predicting the right alarm settings.
It was efficient. It achieved these results using 83% fewer "brain cells" (parameters) than the most powerful competitor (the iTransformer).
It was fair. The system was statistically proven to be better than the others, not just by a tiny bit, but significantly.
It adapts. The system is designed to be retrained every day with new data, meaning it gets smarter and adjusts to the network's changing habits automatically.

Why This Matters

This paper doesn't just say "AI is cool." It solves a specific, expensive problem: preventing engineers from making unnecessary trips to fix things that aren't broken, while ensuring they don't miss things that are.

By moving from a "one-size-fits-all" rulebook to a "smart, context-aware" system, the network can stay healthy without wasting resources. The best part? Because the PCTN model explains its reasoning (by showing the average and the variation), human operators can look at its decisions, understand them, and even tweak them manually if they need to, without having to retrain the whole AI.

Technical Summary: Adaptive Alarm Threshold Prediction in 4G Mobile Networks

1. Problem Definition

In mobile telecommunications, alarms serve as critical early warning signals for cell degradation or failure. Currently, operators rely on static, manually configured thresholds to trigger these alarms (e.g., maximum inactive time, fluctuation counts). These fixed thresholds fail to account for dynamic network conditions, such as time of day, day of the week, traffic volume, and vendor-specific behaviors. Consequently, this rigidity leads to two operational failures:

Missed Faults: Serious issues during high-traffic periods go undetected because static thresholds are too lenient for the noise.
False Alarms: Minor, normal fluctuations during quiet periods trigger unnecessary engineer callouts.

The core challenge addressed by this paper is the adaptive alarm threshold prediction problem: predicting the optimal decision boundary (threshold) dynamically based on live network behavior. A significant hurdle is the absence of ground truth labels; no network log explicitly states what the "correct" threshold should have been at a specific moment.

2. Methodology

2.1 Data and Label Derivation

The study utilizes an anonymized dataset from a real 4G network in South Asia, comprising 10,648 unique cells across three equipment vendors and nine geographic regions over ten days (14,054 cell-day records).

Feature Engineering: The input vector includes 123 features, covering temporal context (time of day, day of week), raw hourly inactive times and fluctuation counts, time-block aggregations, rolling statistics, and vendor/region indicators.
Percentile-Guided Label Derivation: To solve the lack of ground truth, the authors introduce a strategy to derive labels directly from data statistics:
- Window Length ( $t_1$ ): Based on the count of hours exceeding a 5-minute inactive threshold in the next 8 hours.
- Inactive Time ( $t_2$ ): Derived from the 75th percentile of window inactive time, adjusted by sensitivity factors for morning peaks, evening hours, and off-peak times.
- Fluctuation Thresholds ( $t_3, t_4$ ): Derived from the 90th percentile of fluctuation counts with a floor of 1.
- Validation: A Kolmogorov-Smirnov test confirmed that deriving labels from a subset of cells did not introduce bias, validating the use of the full dataset for training.

2.2 Model Architectures

Four models were evaluated to predict four targets simultaneously: audit window duration, inactive time limit, total fluctuation count, and per-hour fluctuation limit.

XGBoost (Baseline): A Gradient Boosted Tree baseline using pseudo-Huber loss. It serves as a classical tabular data benchmark.
CNN-BiLSTM with Attention: A hybrid sequential model using convolutional layers for local pattern extraction, Bidirectional LSTM for temporal dependencies, and multi-head attention to weight relevant hours based on context.
iTransformer: A state-of-the-art model adapted for this task. Unlike standard Transformers that attend across time steps, iTransformer treats each feature as a token and learns attention across the feature dimension, better suited for multivariate alarm data.
PCTN (Proposed): The Percentile Guided Contextual Threshold Network. This novel architecture decouples threshold prediction into two interpretable stages:
- Stage 1 (Distribution Learning): Learns the statistical distribution (mean $\mu$ and variance $\sigma$ ) of network behavior for a given context using a Transformer-based encoder with attention pooling.
- Stage 2 (Analytical Derivation): Derives thresholds analytically from the learned distribution rather than predicting them directly.
  - Dynamic Alpha: Uses a learned $\alpha$ parameter to scale the threshold ( $\theta = \mu + \alpha\sigma$ ), allowing operators to adjust conservatism without retraining.
  - Specialized Heads:
    - Categorical Head: For discrete window lengths (2–8 hours).
    - Gaussian Head: For continuous inactive time.
    - Bernoulli Gate: For fluctuation targets, which suffer from extreme class imbalance (94% of values are at the floor of 1.0). The gate first predicts if a value is above the floor, then predicts the magnitude.

2.3 Training Strategy

The framework supports daily incremental retraining. While XGBoost is fully retrained, the neural models (PCTN, iTransformer, CNN-BiLSTM) utilize fine-tuning (5–10 epochs) on new daily data, leveraging weights from previous days to capture evolving trends efficiently.

3. Key Results

The models were evaluated on a held-out test set (3 days) using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and $R^2$ .

Overall Performance: PCTN achieved the best overall average $R^2$ of 0.733 and the lowest average MAE of 0.290. It outperformed the iTransformer ( $R^2$ = 0.436), CNN-BiLSTM ( $R^2$ = 0.324), and XGBoost ( $R^2$ = -31.03).
Efficiency: PCTN uses 142,261 parameters, which is 83% fewer than the iTransformer (818,508 parameters), yet it outperforms it on three of the four targets.
Statistical Significance: Pairwise comparisons between PCTN and all baselines were statistically significant ( $p < 0.001$ ) via the Wilcoxon signed-rank test.
Target-Specific Insights:
- Fluctuation Targets: PCTN achieved $R^2 \approx 0.57$ on fluctuation targets, while other neural models failed ( $R^2 \approx 0$ or negative). This success is attributed entirely to the Bernoulli gate, which explicitly models the binary nature of the floor-concentration problem.
- XGBoost Failure: XGBoost failed on continuous targets ( $R^2 < -100$ ) due to feature extrapolation failure. When test data (a high-stress day) exceeded the training range, tree-based models could not generalize, whereas neural models interpolated successfully.
- Interpretability: PCTN's dynamic $\alpha$ values varied meaningfully across samples (e.g., $\alpha$ for inactive time ranged from 8.8 to 22.2), allowing operators to inspect and manually adjust the "conservatism" of the policy for specific cells without retraining.

4. Significance and Claims

The paper claims the following contributions and significance:

First Formalization: It is the first work to frame adaptive alarm threshold prediction as a supervised machine learning problem, moving beyond fault detection to predicting the decision boundary itself.
Label Derivation Strategy: It introduces a percentile-guided method to create training supervision from unlabeled operational data, validated to be free of circular bias.
PCTN Architecture: The proposed PCTN demonstrates that structural alignment with data characteristics (using categorical, Gaussian, and Bernoulli heads) is more effective than simply increasing model size. It achieves state-of-the-art accuracy with significantly lower computational cost.
Operational Utility: The framework enables daily retraining, allowing thresholds to adapt to changing network behaviors. Crucially, the interpretable outputs (mean, variance, dynamic alpha) allow network operators to inspect the model's reasoning and manually adjust thresholds for specific contexts, bridging the gap between black-box AI and operational control.
Robustness: The study highlights that neural models are more robust to novel stress events (extrapolation) compared to tree-based models, making them safer for production deployment in dynamic environments.

The authors conclude that while the current study is a proof-of-concept limited to 4G data and a 10-day window, the framework provides a scalable path toward maintaining Quality of Service (QoS) without proportional increases in engineering effort as networks evolve toward 5G.