⚡ electrical engineering

A Deep Learning Approach to Multipath Component Detection in Power Delay Profiles

This paper proposes and evaluates a novel framework combining transformer-based autoencoders with DBSCAN clustering for robust multipath component detection in Power Delay Profiles, demonstrating superior performance over traditional methods through a relaxed F1 score evaluation.

Original authors: Ondrej Zeleny, Radek Zavorka, Ales Prokes, Tomas Fryza, Jaroslaw Wojtun, Jan M. Kelner, Cezary Ziolkowski, Aniruddha Chandra

Published 2026-03-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ondrej Zeleny, Radek Zavorka, Ales Prokes, Tomas Fryza, Jaroslaw Wojtun, Jan M. Kelner, Cezary Ziolkowski, Aniruddha Chandra

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 you are trying to hear a friend whispering in a crowded, noisy room. Your friend's voice is the signal, but the room is full of echoes bouncing off walls, windows, and furniture. These echoes are called Multipath Components (MPCs).

In the world of wireless communication (like your Wi-Fi or 5G), signals do the same thing. They bounce off buildings and cars, creating a messy mix of the original signal and its echoes. To make sense of this, engineers use a map called a Power Delay Profile (PDP). Think of the PDP as a topographic map of sound. The "mountains" on this map are the strong signals (your friend's voice), and the "valleys" are the silence or noise.

The problem? Sometimes the mountains are small, hidden in fog (noise), or the map is so blurry you can't tell a mountain from a hill. Traditional methods are like using a ruler to measure the hills; they often fail when the terrain gets messy.

This paper proposes a smarter way to find those "mountains" using Deep Learning, specifically a type of AI called an Autoencoder. Here is how it works, broken down into simple concepts:

1. The "Copycat" Artist (The Autoencoder)

Imagine you hire an artist who is very good at copying a drawing of a smooth, rolling hill. You show them thousands of pictures of smooth hills (the "normal" background noise of the signal). They learn to copy these perfectly.

Now, you show them a picture of a hill with a sudden, sharp spike (a Multipath Component). Because the artist has only ever seen smooth hills, they try to copy the spike as if it were a smooth hill. They fail. The copy looks wrong.

In the paper's language, this "failure to copy" is called Reconstruction Error. The AI realizes, "Hey, I couldn't copy this part perfectly, so this part must be special!" That "special part" is the signal peak we are looking for.

2. The "Noise Filter" (DBSCAN)

Once the AI points out all the places where it failed to copy the signal, you get a list of "suspicious spots." But sometimes, the AI gets a little jittery and marks a few random spots as suspicious just because of static.

To fix this, the authors use a clustering tool called DBSCAN. Think of this as a grouping game. If the AI marks three spots in a row as "suspicious," DBSCAN says, "Okay, these three are likely one big mountain, not three tiny pebbles." It groups the noise together and keeps only the biggest, most significant groups. This filters out the false alarms.

3. The "Super-Artist" (The Transformer)

The researchers tested four different types of AI artists:

CNN: Good at seeing local patterns (like a painter looking at one brushstroke).
LSTM/GRU: Good at remembering sequences (like a musician remembering a melody).
Transformer: The new superstar.

The Transformer is like an artist who doesn't just look at the brushstroke or the melody; it looks at the entire painting at once and understands how every part relates to every other part. It uses something called "Self-Attention" to focus on the most important details.

The Result: The Transformer-based artist was the best at spotting the "mountains" (peaks) while ignoring the "fog" (noise). It achieved the highest accuracy, finding more real signals and making fewer mistakes than the other methods.

Why Does This Matter?

Why do we care about finding these echoes?

Better Location Services: If your phone knows exactly where the signal bounces, it can pinpoint your location much more accurately (crucial for self-driving cars).
Faster Internet: By understanding the echoes, the network can adjust its settings to send data faster and with fewer errors.
6G and Beyond: As we move to faster, more complex networks, the "fog" gets thicker. We need smarter AI to cut through it.

The Bottom Line

The authors built a smart system that learns what a "normal" signal looks like, then spots the weird, sharp spikes that represent important data. By combining this "Copycat AI" with a smart "Grouping Game," they can find signal paths that traditional tools miss. Their best tool? A Transformer, which acts like a super-observant detective that sees the big picture, making wireless communication more reliable and precise.

1. Problem Statement

In wireless communication systems, multipath propagation causes signals to reflect, scatter, and diffract, leading to fading and interference. The Power Delay Profile (PDP) is a critical metric representing received signal power over time delay, where peaks correspond to dominant signal paths (Multipath Components or MPCs).

Accurately detecting these peaks is essential for:

Channel Estimation: Improving equalization and signal recovery.
High-Precision Positioning (5G/6G): Enhancing Time of Arrival (TOA) and Angle of Arrival (AOA) estimates.
Adaptive Modulation: Optimizing transmission parameters.

Challenges: Traditional detection methods (thresholding, matched filtering, maximum likelihood) struggle in complex environments due to:

Sensitivity to noise and low Signal-to-Noise Ratios (SNR).
Inability to distinguish weak MPCs from background interference.
The need for extensive manual tuning.
Increasing complexity in higher frequency bands (mmWave, THz).

2. Methodology

The authors propose a two-stage reconstruction-based anomaly detection framework that combines Deep Learning (DL) with unsupervised clustering.

A. Core Concept: Reconstruction-Based Anomaly Detection

The fundamental hypothesis is that an autoencoder trained to reconstruct "normal" PDP signals (smooth decay) will struggle to reconstruct sudden spikes (MPCs). Consequently, these spikes will result in high reconstruction errors, which serve as indicators for anomaly (peak) detection.

B. The Proposed Framework

Stage 1: Autoencoder Training & Reconstruction
- Four different deep learning architectures were evaluated as autoencoders:
  - CNN-AE: Convolutional layers to capture local spatial/temporal patterns.
  - LSTM-AE & GRU-AE: Recurrent networks to model sequential dependencies.
  - Transformer-AE: Utilizes self-attention mechanisms to capture long-range dependencies and global context without sequential processing constraints.
- The models are trained to minimize reconstruction error on PDP data.
- Optimization Note: For RNNs (LSTM/GRU), long sequences caused vanishing gradients. The authors found an optimal subsequence length of 85 samples (1/10th of original length) to balance learning capability and peak preservation.
Stage 2: DBSCAN Clustering
- The reconstruction errors are passed to the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm.
- Pre-processing: Negative reconstruction errors are filtered out to remove noise.
- First Pass: Groups high-error areas to identify potential peak regions.
- Second Pass: Refines the clusters to ensure only the single most significant peak per anomaly is retained (preventing multiple detections for one MPC).
- Final Selection: The peak with the highest power within each cluster is matched against the original data.

C. Evaluation Metric

To address the difficulty of precise peak labeling in noisy environments, the authors introduced a Relaxed F1-Score. This metric incorporates a tolerance window ( $N$ steps) around labeled anomalies, allowing for slight positional deviations between the detected peak and the ground truth.

3. Key Contributions

Novel Framework: Integration of reconstruction-based autoencoders with DBSCAN clustering specifically for MPC detection in PDPs.
Architecture Comparison: A comprehensive evaluation of CNN, LSTM, GRU, and Transformer architectures for this specific signal processing task.
Transformer Superiority: Demonstration that Transformer-based autoencoders outperform traditional RNNs and CNNs in this domain due to their ability to model long-range dependencies and global context.
Robust Metric: Proposal of a relaxed evaluation strategy (Relaxed F1) to better reflect real-world detection performance where exact peak alignment is difficult.

4. Experimental Results

Dataset: 80 PDP measurements from a 60 GHz channel between a moving vehicle and a fixed receiver (60 GHz band). Data was split into training/testing sets and augmented with rolling shifts and Gaussian noise.
Performance Comparison (Table II):
- Transformer-AE: Achieved the best performance with Recall: 61%, Precision: 73%, and F1-Score: 66%. It excelled at reconstructing noise, thereby minimizing noise in the error space and isolating true peaks.
- LSTM-AE: Balanced performance (Recall: 55%, F1: 48%).
- CNN-AE & GRU-AE: Showed slightly lower recall (53% and 44% respectively) but competitive precision.
Visual Analysis: The Transformer model produced the cleanest reconstruction of the background noise, making the MPC spikes (anomalies) stand out clearly in the error signal compared to other models.

5. Significance and Conclusion

Advancement over Heuristics: The study demonstrates that deep learning approaches can detect multipath components more robustly than traditional threshold-based heuristics, particularly in low-SNR and complex environments.
Transformer Potential: It highlights the underutilized potential of Transformer architectures in 1D time-series signal processing for wireless communications, showing they can outperform RNNs in capturing global signal context.
Practical Application: The proposed framework offers a data-driven, automated solution for channel estimation and positioning tasks without relying on manually tuned parameters.
Future Work: The authors suggest improving recall rates, exploring Reinforcement Learning (RL) for better pattern distinction, and testing the framework in more diverse and complex real-world scenarios.

Funding: The research was supported by the National Science Center (Poland), the Czech Science Foundation, and the Internal Grant Agency of Brno University of Technology.