💻 computer science

Machine Learning for PRACH Interference Detection in 5G NR Networks

This paper benchmarks multiple machine learning models for 5G NR PRACH interference detection, revealing that while Transformer Encoders offer high precision, CNN and XGBoost provide the optimal balance of high accuracy and computational efficiency, making them the most suitable candidates for real-time interference management.

Original authors: Arsène KABORE, Désiré GUEL, Tounwendyam Fréderic OUEDRAOGO

Published 2026-09-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arsène KABORE, Désiré GUEL, Tounwendyam Fréderic OUEDRAOGO

Original paper licensed under CC BY 4.0 (https://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

In the invisible landscape of modern wireless communication, billions of devices constantly strive to connect to the internet through a complex web of cell towers. For a smartphone to begin this conversation, it must first send a brief, specific signal to the tower to announce its presence and request a connection. This initial handshake happens on a dedicated pathway known as the physical random access channel. However, in crowded areas where many devices try to speak at once, these signals can collide, creating a chaotic mix of interference that drowns out the original message. When this happens, the connection fails, leading to dropped calls or slow data. To keep networks running smoothly, engineers need a way to instantly recognize these signals even when they are buried under noise and competing transmissions.

Researchers at universities in Burkina Faso have tackled this challenge by testing a variety of artificial intelligence tools to see which one is best at spotting these hidden signals. They focused on the 5G New Radio standard, the latest generation of mobile technology, and simulated a realistic environment where signals are distorted by noise and interference from neighboring cells. Instead of relying on traditional mathematical formulas that struggle in such messy conditions, the team trained several different types of machine learning models to act as digital listeners. These models were fed thousands of simulated signal samples, ranging from clear transmissions to those heavily corrupted by interference, and asked to identify whether a valid signal was present and, if so, what type it was.

The study compared five distinct approaches to this problem. Two of the models were based on deep learning, a method where computers learn by processing data through layers of artificial neurons. One of these, a convolutional neural network, is designed to spot patterns in data much like a human recognizes shapes in a picture. Another, called a long short-term memory network, is built to understand sequences of events over time. A third deep learning model, known as a transformer encoder, uses a mechanism that allows it to weigh different parts of the signal simultaneously to find connections that might be far apart. The researchers also tested a one-dimensional convolutional model, which is a streamlined version of the pattern-spotting network, and a completely different type of algorithm called XGBoost, which makes decisions by building a series of simple rules, similar to how a person might sort items into categories based on a checklist of features.

After running these models through rigorous testing, the results revealed a clear trade-off between how well a model performs and how much computing power it requires. The deep learning models that process sequences over time, specifically the long short-term memory network and the transformer encoder, proved to be very accurate at avoiding false alarms. However, they came with a heavy cost: they took an extremely long time to learn from the data, requiring more than 1,900 seconds of training time, and they were slow to make decisions when actually in use. In contrast, the pattern-spotting convolutional neural network and the rule-based XGBoost model achieved the best overall balance. Both of these models correctly identified signals about 89.6 percent of the time, a score that represents the highest success rate in the study.

The most striking difference emerged in the speed of operation. The XGBoost model was by far the most efficient, completing its training in just 19.53 seconds and making a decision on a new signal in only 0.06 seconds. This speed makes it an ideal candidate for real-time applications where delays are unacceptable. The convolutional neural network was also a strong contender, taking 282.73 seconds to train and 2.60 seconds to make a prediction, which is fast enough for many practical network management tasks. While the transformer model showed the highest precision, meaning it rarely made mistakes when it did speak up, it missed a significant number of actual signals, making it too conservative for a system that needs to catch every possible connection attempt.

The researchers concluded that for the immediate future of 5G networks, the most effective strategy involves using either the convolutional neural network or the XGBoost model. These two approaches offer the best combination of high accuracy and manageable computing costs, ensuring that networks can detect interference quickly without overloading the hardware. While the more complex deep learning models showed promise in specific areas, their high demand for time and processing power currently limits their use in systems that must react instantly. The study, which relied on computer simulations of 35,000 signal samples, suggests that by choosing the right tool for the job, network operators can significantly improve the reliability of mobile connections in crowded environments. Future work will need to confirm these findings in real-world tests, but the current results provide a clear roadmap for building smarter, more resilient wireless networks.

Technical Summary: Machine Learning for PRACH Interference Detection in 5G NR Networks

1. Problem Statement

In 5G New Radio (NR) networks, the Physical Random Access Channel (PRACH) is critical for User Equipment (UE) to establish initial connections with the base station (gNB). This process relies on the transmission of preamble sequences, typically based on Zadoﬀ–Chu (ZC) codes. However, the presence of cell interference—arising from simultaneous access attempts by multiple UEs (intra-cell) or neighboring cells (inter-cell)—significantly degrades the accuracy of PRACH signal detection.

Traditional detection methods rely on correlation-based techniques. While effective in ideal conditions, these methods are susceptible to noise and interference, often leading to false alarms and missed detections. Furthermore, setting optimal thresholds for these traditional detectors is challenging under varying signal conditions. The paper addresses the need for more robust detection mechanisms that can operate effectively in realistic, interference-laden environments while adhering to the strict latency and computational constraints of 5G systems.

2. Methodology

The study conducts a comprehensive benchmarking of five distinct machine learning paradigms applied to PRACH interference detection. The methodology involves generating a synthetic dataset compliant with 3GPP specifications and evaluating models across accuracy, generalization, and computational efficiency.

2.1 Data Generation and Modeling

Signal Model: The received signal is modeled as $y(t) = s(t) + n(t) + i(t)$ , comprising the transmitted signal, additive white Gaussian noise (AWGN), and interference.
Dataset: A synthetic dataset of 35,000 samples per split (train/validation/test) was generated using PRACH Format 0 (ZC sequence length $N_{ZC}=839$ ).
Conditions: The data covers a wide range of Signal-to-Noise Ratios (SNR) from -18 dB to -6 dB (noise only) and Signal-to-Interference-plus-Noise Ratios (SINR) down to -30 dB. Interference scenarios include both intra-cell and inter-cell types.
Channel Modeling: The Extended Typical Urban (ETU) channel model was used by default, with support for TDL-A and EVA.
Labels: Samples were labeled for binary (interference present/absent) and multiclass (none/intra/inter) tasks.

2.2 Models Evaluated

The study compares five models, each with distinct inductive biases:

Convolutional Neural Networks (CNN): Utilizing 2D convolutional layers on correlation maps derived from matched-filtering against ZC roots. The architecture includes convolution, ReLU, batch normalization, and pooling layers to learn spatial patterns and timing invariance.
Long Short-Term Memory (LSTM): A recurrent architecture (2 layers, 128 units, bidirectional) designed to capture long-range temporal dependencies in raw IQ sequences or correlation streams, mitigating vanishing gradients.
Transformer Encoder: A self-attention-based model using multi-head attention and positional encodings to aggregate non-local temporal evidence in parallel, aiming to suppress structured interference.
XGBoost: A gradient-boosted decision tree ensemble trained on engineered tabular features. Features include energy ratios (in/out of window), peak-to-sidelobe ratios (PSR), peak positions, and SINR proxies.
Conv1D: A lightweight 1D convolutional network operating directly on raw IQ streams or correlation sequences, offering an $O(T)$ complexity alternative to 2D CNNs.

2.3 Experimental Setup

Hardware: Experiments were conducted on a laptop with an Intel Core i5-1335U, 16 GB RAM, and Intel Iris Xe Graphics.
Training: Deep learning models used the Adam optimizer with categorical cross-entropy loss and early stopping (patience=5). XGBoost utilized multi-class softmax with regularization.
Metrics: Performance was evaluated using Accuracy, Precision, Recall, F1-Score, Training Time (Fit Time), and Inference Time (Prediction Time).

3. Key Results

The study reveals significant trade-offs between detection performance and computational cost.

3.1 Detection Performance

F1-Score: XGBoost achieved the highest F1-score of 89.66%, closely followed by CNN at 89.64%. Conv1D performed competitively with 89.27%.
Accuracy: All models maintained accuracy above 87%, with CNN leading at 88.44% and XGBoost at 88.36%.
Precision vs. Recall: The Transformer Encoder demonstrated the highest precision (93.46%) but suffered from the lowest recall (83.38%), indicating a conservative detection behavior that misses true preambles. Conversely, LSTM showed reasonable recall but lagged in overall F1-score.

3.2 Computational Efficiency

Training Time: XGBoost was the most efficient, requiring only 19.53 seconds. In contrast, deep learning models required significantly more time: CNN (282.73 s), Conv1D (448.11 s), and notably LSTM (1903.94 s) and Transformer (1917.33 s).
Inference Time: XGBoost maintained its lead with an inference time of 0.06 seconds. CNN (2.60 s) and Conv1D (1.96 s) offered moderate latency, while LSTM and Transformer exceeded 10 seconds per sample.

4. Key Contributions

Comprehensive Benchmarking: The paper provides a rigorous comparison of diverse ML paradigms (CNN, LSTM, Transformer, XGBoost, Conv1D) under a unified, 3GPP-compliant interference scenario, a gap identified in prior literature.
Efficiency-Accuracy Trade-off Analysis: It explicitly quantifies the trade-off between detection metrics and computational costs, highlighting that while deep sequence models (LSTM/Transformer) offer theoretical advantages in temporal modeling, their computational overhead is prohibitive for real-time 5G applications.
Feature Engineering vs. Deep Learning: The study demonstrates that well-engineered features fed into tree-based models (XGBoost) can outperform or match deep learning models in both accuracy and speed, challenging the assumption that deep learning is always superior for signal detection.
Practical Decision Framework: The authors propose a decision tree for model selection based on operational constraints (latency, data availability, and computational budget), guiding practitioners toward XGBoost for real-time deployment and CNN/Conv1D for scenarios requiring robust pattern learning with moderate latency.

5. Significance and Claims

The paper claims that CNN and XGBoost are the most suitable candidates for real-time interference management in 5G NR systems.

XGBoost is positioned as the optimal choice for scenarios demanding extreme computational efficiency and rapid retraining, offering the best balance of speed and accuracy.
CNN is highlighted for its robust generalization capabilities and ability to learn complex spatial-temporal patterns directly from signal data, making it a strong alternative when interpretability of features is less critical than raw performance.

The authors emphasize that while Transformer and LSTM architectures show promise in specific theoretical contexts (e.g., high precision or long-range dependency modeling), their current computational costs (training >1900s, inference >10s) render them less viable for the latency-sensitive requirements of PRACH detection in practical 5G deployments.

6. Limitations

The paper modestly acknowledges several limitations:

Synthetic Data: All results are based on simulations using 3GPP-compliant synthetic datasets; real-world over-the-air (OTA) validation has not yet been performed.
Hardware Specificity: Computational analysis is limited to a single hardware environment, and scalability across diverse platforms requires further study.
Scope: The analysis focuses on PRACH Format 0 and specific interference scenarios; generalizability to other formats and complex multi-cell environments needs further investigation.
Statistical Rigor: While a correlation-based baseline is included, the authors note that further statistical significance testing across multiple seeds and channel realizations could strengthen the findings.