Learnable Template Matching Approach for Micro-Deformation Monitoring based on Integrated Sensing and Communication Platform

This paper proposes a learnable template-matching approach integrated with an electromagnetic signal feature embedding and clutter filtering mechanism to effectively suppress environmental clutter and enhance the accuracy of micro-deformation monitoring in Integrated Sensing and Communication (ISAC) platforms.

The Gist

Here is an explanation of the paper, translated into simple language with creative analogies.

The Big Picture: The "Dual-Purpose" Radio

Imagine a cell tower (Base Station) as a busy post office.

• Its main job: Delivering letters (data) to your phone (Communication).
• Its new side job: Trying to listen to the wind to see if a bridge is shaking (Sensing).

This paper is about a new technology called ISAC (Integrated Sensing and Communication). It tries to do both jobs at the same time using the same radio waves.

The Problem:
The post office is very loud. When the tower tries to listen to a tiny, subtle vibration on a bridge (like a micro-deformation), the "noise" of the city is overwhelming.

• The Noise: Cars driving by, wind blowing through trees, and the tower itself vibrating slightly.
• The Signal: The tiny, dangerous crack or shift in the bridge structure.

Trying to hear the bridge's tiny creak over the roar of traffic is like trying to hear a whisper in a heavy metal concert. Existing methods are too messy; they get confused by the noise and give bad measurements.

The Solution: The "Smart Detective" (LTM)

The authors propose a new AI-powered method called Learnable Template Matching (LTM). Think of this as a super-smart detective who has a specific "mugshot" of what the bridge's vibration should look like.

Here is how the detective solves the case, step-by-step:

1. The "Unwrapping" Trick (Phase Unwrapping)

Radio waves bounce off things and come back. Because waves are circular (like a clock face), the data often gets "wrapped up" or confused. It's like trying to read a map where the North Pole is glued to the South Pole.

• The Fix: The AI uses a CNN (a type of deep learning brain) to "unwrap" the map. It straightens out the confusing circular data so the detective can see the true shape of the vibration.

2. The "Template" Match (The Mugshot)

The detective knows exactly what a "healthy bridge vibration" looks like. It's usually a smooth, rhythmic pattern (like a heartbeat).

• The Trick: The AI creates a Learnable Template. This isn't a fixed picture; it's a flexible mold that the AI learns and adjusts as it trains.
• The Action: The AI slides this mold over the messy data.
  • If the data fits the mold (it's the bridge), it amplifies it (turns the volume up).
  • If the data doesn't fit the mold (it's a car or wind), it suppresses it (turns the volume down to zero).

3. The "Sieve" (Clutter Suppression)

Imagine you have a bucket of mixed sand and gold nuggets. The sand is the city noise; the gold is the bridge vibration.

• The LTM network acts like a magic sieve. It shakes the bucket. The sand (noise) falls through the holes, but the gold (the specific bridge vibration) stays on top.
• The AI is smart enough to know that the bridge moves in a specific rhythm, while the noise is random. It filters out the random stuff and keeps the rhythmic stuff.

Why This Matters

Before this paper:
If you tried to monitor a bridge with a standard cell tower, the system would say, "I can't tell if the bridge is safe because the wind is too loud." It was too inaccurate to be useful for safety.

With this paper:
The system can now say, "I see a tiny 1-millimeter shift in the bridge happening right now, even though a truck just drove by."

The Real-World Test

The researchers didn't just run this on a computer; they tested it on a real cell tower in Nanjing, China, monitoring the Yangtze River Bridge.

• The Result: The AI successfully ignored the vibrations caused by the tower itself and the wind, and accurately tracked the bridge's movement when heavy trucks drove over it.
• The Comparison: It worked much better than standard AI methods (like CNNs or LSTMs) and traditional filtering techniques. It was the only one that could consistently spot the dangerous, large movements.

The Takeaway

This paper introduces a way to turn our everyday cell towers into high-precision health monitors for our infrastructure. By using a "smart detective" AI that knows exactly what to listen for, we can use existing 5G/6G networks to keep our bridges, roads, and buildings safe without needing expensive, dedicated radar equipment.

In short: They taught the cell tower to ignore the city's noise and listen only to the "heartbeat" of our bridges.

Technical Summary

Here is a detailed technical summary of the paper "Learnable Template Matching Approach for Micro-Deformation Monitoring based on Integrated Sensing and Communication Platform."

1. Problem Statement

The paper addresses the challenge of Micro-Deformation Monitoring (mDM) using Integrated Sensing and Communication (ISAC) platforms. While ISAC systems (a key enabler for 6G) share spectrum and aperture resources for both communication and sensing, they suffer from inferior sensing performance compared to dedicated radar systems.

• Specific Challenge: Detecting micro-deformations (small structural displacements, e.g., bridge vibrations) is a "weak target" problem. In ISAC systems, the sensing signal is often overwhelmed by:
  1. Clutter: Environmental interference from vehicles, pedestrians, and the Base Station (BS) itself.
  2. Low Resolution: Limited by the shared resources and bandwidth.
  3. Phase Wrapping: The non-linear nature of phase extraction in electromagnetic (EM) signals makes direct displacement estimation difficult.
• Goal: To accurately estimate the Micro-Deformation Displacement (mDD) and its vibration frequency from ISAC signals despite poor sensing quality and heavy clutter.

2. Methodology

The authors propose an AI-assisted, Learnable Template Matching (LTM) approach. The methodology is structured into three main stages:

A. System Modeling and Problem Formulation

• Signal Model: The received ISAC signal is modeled as a superposition of the target signal (infrastructure deformation), clutter signals (BS vibration, vehicle noise), and Additive White Gaussian Noise (AWGN).
• Clutter Characterization: The environment clutter is modeled as a combination of deterministic periodic signals (BS vibration) and stochastic noise.
• Reconstruction Problem: The task is formulated as an optimization problem to reconstruct the mDD signal by minimizing the error between the observed signal and the reconstructed signal while suppressing clutter.

B. The Learnable Template Matching (LTM) Network

The core contribution is a deep learning architecture designed to solve the reconstruction problem by splitting it into two sub-tasks:

1. Phase Unwrapping (via CNN):

   • Challenge: Raw phase data from the BS is wrapped (modulo $2\pi$), causing discontinuities.
   • Solution: A Convolutional Neural Network (CNN) with narrow-band kernels and LeakyReLU activation is used to map the wrapped phase to a continuous, unwrapped phase signal. This ensures the continuity required for accurate displacement calculation.

2. Signal Decoupling (via Learnable Template):

   • Challenge: Separating the target mDD signal from the coupled clutter.
   • Solution: Instead of using a fixed template, the network learns an adaptive template ($T_L$) during training.
     • Target Enhancement: The unwrapped phase is convolved with the learnable template to enhance features matching the target's periodic motion.
     • Clutter Suppression: A Max Pooling layer is applied to the output. Since the target signal aligns with the learned template (producing peaks) while clutter does not, the pooling operator effectively extracts the target peaks and suppresses unmatched clutter.
     • Frequency Estimation: A Fast Fourier Transform (FFT) module is embedded to capture the vibration frequency component, utilizing a spectrum power-based optimization strategy to handle non-differentiable frequency selection.

C. Loss Function Design

The network is trained using a composite loss function ($L$) comprising three terms:

1. Displacement Bias Loss ($L_r$): Minimizes the error between estimated and ground-truth displacement.
2. Frequency Estimation Loss ($L_f$): Minimizes the error in estimated vibration frequency.
3. Signal Reconstruction Loss ($L_d$): Ensures the reconstructed mDD signal matches the physical model of the deformation.

3. Key Contributions

1. First ISAC-based mDM System: The authors propose the first framework specifically for micro-deformation monitoring using ISAC platforms, addressing the unique constraints of shared resources.
2. Clutter-Target Feature Model: They reformulate clutter suppression as a target signal enhancement task, modeling clutter as a mix of deterministic vibration and stochastic noise.
3. LTM Network Architecture: A novel deep learning framework that integrates:
   • CNN-based phase unwrapping.
   • Learnable periodic motion templates for adaptive signal matching.
   • Max pooling for feature extraction and clutter filtering.
   • Embedded FFT for frequency analysis.
4. Interpretability: The network is designed to be interpretable, explicitly separating phase processing, template matching, and frequency analysis, rather than acting as a "black box."

4. Results

The proposed approach was evaluated through numerical simulations and real-world experiments on the Nanjing Qixiashan Yangtze River Bridge.

• Simulation Results:

  • Convergence: The network converges rapidly for displacement bias (within 1 epoch) and frequency estimation (within 25 epochs).
  • Generalization: The model accurately estimated displacements in the 0–25 mm range and frequencies in the 5–15 Hz range, even with test samples having parameters different from the training set.
  • Clutter Suppression: The LTM successfully separated the target deformation signal from fixed-frequency clutter (0.41 Hz) and random noise, demonstrating superior performance over standard regression baselines (CNN, MLP, LSTM).

• Real-World Experiments (Bridge Monitoring):

  • Accuracy: The LTM method achieved the closest match to ground truth for bridge vibration trends compared to baseline algorithms and ablated variants.
  • Large Deformation Detection: In detecting large deformation events (>1 mm and >2 mm), LTM correctly identified 27 out of 37 events (within 0.1mm error) and 4 out of 3 events (within 0.2mm error for >2mm), significantly outperforming baselines which missed nearly half of the events.
  • Robustness: The system maintained high accuracy even under real-world noise and BS vibration coupling.

5. Significance

• Infrastructure Safety: This work demonstrates that ISAC systems, previously considered insufficient for high-precision sensing, can be effectively repurposed for critical infrastructure monitoring (e.g., bridges, tunnels) using AI.
• Resource Efficiency: It proves that high-precision micro-deformation monitoring can be achieved without dedicated radar hardware, leveraging existing communication Base Stations.
• Methodological Advancement: The proposed Learnable Template Matching approach offers a flexible, physics-informed deep learning solution for weak target detection in cluttered environments, which can be extended to other periodic signal extraction problems in electromagnetic sensing.