⚡ electrical engineering

Near-Sensor Damage Localization in Multi-Story Buildings with Level-Crossing Spike Encoding and a Compact Spiking Neural Network

This paper proposes an energy-efficient, event-driven structural health monitoring pipeline that combines a level-crossing spike encoder with a compact spiking neural network to localize damage in multi-story buildings, achieving near-baseline accuracy while significantly reducing energy consumption and data transmission rates through a tunable threshold mechanism.

Original authors: Saher Elsayed, Mohamed Ali, Khairi Azhar Aziz

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Saher Elsayed, Mohamed Ali, Khairi Azhar Aziz

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

Imagine a tall building as a giant, silent piano. When the wind blows or traffic rumbles nearby, the building vibrates, playing a unique "song" made of different notes (frequencies). If a part of the building gets damaged—like a cracked beam or a loose bolt—it's like a piano key getting stuck or a string going slightly out of tune. The song changes, and if you listen closely enough, you can tell exactly which key is broken.

This paper presents a new, ultra-efficient way for tiny sensors to listen to that song and figure out where the damage is, without draining their batteries.

Here is the breakdown of their approach using simple analogies:

1. The Problem: The "Always-On" Microphone

Traditionally, to monitor a building, you stick accelerometers (motion sensors) on every floor. These sensors act like microphones that record everything 24/7, even when the building is perfectly still.

The Issue: Recording silence is a waste of battery. It's like leaving a high-definition camera recording a dark, empty room all night just in case someone walks in. The battery dies before the building ever needs repair.

2. The Solution: The "Motion-Activated" Sensor

The authors propose a system that only "wakes up" when something interesting happens. They use two main tricks:

Trick A: The "Tuning Fork" Filter (Modal Filterbank)
Before the sensor even looks at the raw vibration, it passes the signal through a set of four "tuning forks" (filters).

Analogy: Imagine you are trying to hear a specific instrument in an orchestra. Instead of listening to the whole chaotic noise, you put on headphones that only let in the sound of the violins, then the flutes, then the drums.
Why it helps: Damage changes specific "notes" (frequencies) of the building. By separating the sound into these specific bands first, the system knows exactly what to listen for, making the job much easier.

Trick B: The "Threshold" Alarm (Level-Crossing Encoding)
Instead of recording every tiny wiggle of the building, the sensor only sends a digital "blip" (a spike) when the vibration changes by a certain amount.

Analogy: Imagine a security guard who only calls the police if a door opens more than a few inches. If the wind just rattles the door slightly, the guard stays quiet.
The Control Knob: The researchers have a single dial (called $\theta$ $θ$ ).
- Turn the dial low: The guard is very sensitive, calling for even tiny movements. This gives high detail but uses lots of battery.
- Turn the dial high: The guard only calls for big movements. This saves massive amounts of battery but might miss very subtle changes.

3. The Brain: A "Spiking" Neural Network

Once the sensor sends these sparse "blips," a tiny computer brain (a Spiking Neural Network or SNN) processes them.

How it works: Unlike a standard computer brain that does math constantly, this brain only does math when it receives a "blip."
The Benefit: If the building is calm, the brain does almost nothing. If the building is shaking, the brain wakes up and works. This makes the energy usage proportional to the activity, not the time.

4. The Results: The "Sweet Spot"

The researchers tested this on a simulated 5-story building with random damage and noise.

Accuracy: Their system got about 87% accuracy in finding the damaged floor. This is very close to the "gold standard" heavy-duty systems (which got 89%), but the heavy-duty systems use much more energy.
Energy Savings: Their system used roughly 6.5 times less energy to make a decision than the standard system.
The Trade-off: By turning that "sensitivity dial" (the threshold), they created a "Pareto frontier." This is just a fancy way of saying they found the perfect balance point where you can save a lot of battery with only a tiny drop in accuracy.

5. What They Discovered (The "Ablations")

They tested what happens if they remove parts of their system to see what matters:

No Filters: If they removed the "tuning forks" and just fed raw noise into the system, accuracy crashed from 87% to 60%. The filters are the most important part.
No "Memory": They tried adding a "memory" feature (recurrence) to the brain, thinking it might help remember past vibrations. It didn't help and actually wasted energy. The filters already did the heavy lifting, so the brain didn't need to remember anything.
The "Old School" Champion: Interestingly, a very simple, old-school math method (using a Support Vector Machine on the filtered data) was actually the most accurate (94%).
- The Takeaway: The authors aren't claiming their AI is the smartest at finding damage. They are claiming it is the most efficient at listening. It gets "good enough" accuracy while using a fraction of the battery and sending much less data over the radio.

Summary

Think of this system as a smart, energy-saving security guard for a building.

Old Guard: Stays awake 24/7, writing down every footstep, even when no one is there. Runs out of battery quickly.
New Guard: Only wakes up when someone takes a big step. Uses a special set of ears to ignore the wind and focus on footsteps. If the building is quiet, the guard sleeps.

The paper proves that this "sleepy" approach can still find exactly where a building is hurt, making it perfect for wireless sensors that need to last for years on a single battery.

Technical Summary: Near-Sensor Damage Localization in Multi-Story Buildings with Level-Crossing Spike Encoding and a Compact Spiking Neural Network

Problem Statement
Wireless, battery-powered accelerometer nodes are standard for vibration-based Structural Health Monitoring (SHM). However, their operational lifespan is constrained by the energy costs of continuous analog-to-digital conversion (ADC), on-node inference, and radio transmission. Traditional data-driven methods, such as 1D Convolutional Neural Networks (1D-CNNs), perform fixed, dense computations regardless of whether the structure is active or quiescent, coupling energy consumption to wall-clock time rather than information content. The challenge is to localize stiffness-degrading damage in multi-story buildings using an event-driven approach where compute cost and data volume scale with signal activity, thereby extending battery life without sacrificing localization accuracy.

Methodology
The authors propose a near-sensor pipeline consisting of three stages: a modal filterbank, a level-crossing encoder, and a compact Spiking Neural Network (SNN).

Structural Model and Data Generation: The study simulates a five-story shear building under ambient excitation (0.5–20 Hz). Damage is modeled as a localized fractional reduction (15–40%) of inter-story stiffness. The classification task involves six classes: one healthy state and five damaged states (one per story).
Sensor Front-End:
- Modal Filterbank: Each floor's acceleration signal passes through a bank of four causal fourth-order band-pass filters (0.8–14 Hz). This separates the signal into sub-bands to expose modal content and mode-shape information, acting as an analog pre-processing stage.
- Level-Crossing Encoding: Each filtered sub-band is digitized by a level-crossing (temporal-contrast) encoder. An event (ON or OFF spike) is emitted only when the signal deviates from a running reference by a fixed threshold $\theta$ . This generates a sparse binary spike tensor where the threshold $\theta$ serves as a single control knob for the trade-off between accuracy, compute cost, and data rate.
Classifier: A compact, feed-forward SNN processes the spike stream. It consists of an input layer projecting to 64 Leaky Integrate-and-Fire (LIF) neurons, which project to a 6-unit leaky-integrator readout. The network is trained using backpropagation through time with a surrogate gradient to handle the non-differentiable spike function.
Energy Accounting: Energy estimates are based on operation counts (MACs for CNNs, synaptic operations for SNNs) at 45 nm CMOS technology. The study explicitly excludes analog filterbank power, sensing, memory movement, and radio transmission, focusing solely on the comparative compute energy of the inference algorithms.

Key Contributions

Event-Driven Pipeline: The paper introduces a specific architecture where a modal filterbank precedes level-crossing encoding and a tiny SNN. It demonstrates that a single threshold parameter ( $\theta$ ) jointly controls accuracy, compute cost, and data rate.
Accuracy-Energy Characterization: The authors systematically map the performance of the SNN against a dense 1D-CNN baseline and classical modal-feature classifiers (SVM and Random Forest). They trace a Pareto frontier showing that energy consumption can vary by an order of magnitude while maintaining competitive accuracy.
Ablation Analysis: The study isolates the critical components of the pipeline, proving that the modal filterbank is the essential enabler for performance, while recurrent connections are unnecessary and energetically costly for this specific task.
Mechanistic Interpretability: The authors link the learned spike activity to structural mechanics, showing that damage produces spatially structured changes in per-floor spike activity that the classifier utilizes, rather than relying on opaque black-box patterns.

Results

Accuracy: On a six-class localization task, the proposed feed-forward SNN achieved 87.2% test accuracy at an operating threshold ( $\theta=0.6$ ). This is within 2 percentage points of the dense 1D-CNN baseline (89.2%).
Energy Efficiency: At the operating point, the SNN required an estimated 49.1 nJ per inference (FP32 equivalent), compared to 317.5 nJ for the 1D-CNN, representing a 6.5× reduction in estimated compute energy.
Pareto Frontier: By sweeping $\theta$ , the SNN achieved a range of operating points. At $\theta=0.8$ , accuracy remained at 86.0% with energy dropping to 35.5 nJ (8.9× reduction). At $\theta=1.8$ , energy dropped to 9.5 nJ, though accuracy fell to 66.7%.
Data Reduction: The level-crossing front-end reduced the data volume leaving the sensor by up to 13.7× in bits compared to dense Nyquist sampling, depending on the threshold.
Baselines: A tuned classical SVM on modal features achieved the highest accuracy (94.3%), while a Random Forest achieved 86.3%. An INT8 quantized CNN achieved 89.2% accuracy at 13.8 nJ.
Ablations: Removing the filterbank caused accuracy to collapse from 87.2% to 59.7%. Adding recurrence to the SNN increased energy by 3.4× without improving accuracy (dropping to 83.8%). A dense CNN applied to the same filterbank features reached 93.5% accuracy but at 18.6× the energy of the proposed SNN.
Robustness: The SNN maintained 79–87% accuracy across signal-to-noise ratios of 15–40 dB. Localization was most difficult for incipient damage (15–20% stiffness loss) but reached near 100% for severe damage (35–40% loss).

Significance and Claims
The paper explicitly frames its contribution not as a breakthrough in raw accuracy, but as a demonstration of event-driven, low-data-rate sensing at competitive accuracy.

Value Proposition: The primary value of the pipeline is its activity-proportional operation. Unlike dense pipelines that consume energy continuously, this approach spends energy only when structural motion generates events. This is critical for always-on sensors where the structure is often quiescent.
Modesty on Claims: The authors state clearly that the proposed SNN does not surpass the accuracy of a tuned classical SVM (94.3%) or a dense 1D-CNN (89.2%). The advantage lies in the sensing modality (sparse event streams) and the resulting reduction in data transmission and compute load.
Scope: The study is a controlled simulation on a five-story shear frame. The authors do not claim field-ready generalization, noting that validation on experimental benchmarks (e.g., IASC-ASCE, Z24 bridge) and full-scale field data with environmental variability remains a necessary next step.
Future Work: The paper suggests that the single-knob trade-off ( $\theta$ ) enables adaptive monitoring strategies where nodes can dynamically adjust thresholds to balance energy and resolution based on real-time conditions.