---
title: 'Vibe2Spike: Batteryless Vibration Sensing via Optical Spikes'
url: https://www.emergentmind.com/topics/vibe2spike
type: topic
---

# Vibe2Spike: Batteryless Vibration Sensing via Optical Spikes

Searching arXiv for Vibe2Spike and closely related neuromorphic/event-based sensing papers.
Vibe2Spike is a batteryless wireless vibration-sensing framework for smart environments that couples visible light communication (VLC), event cameras, and spiking neural networks (SNNs) into a single sensing-to-classification pipeline. In the reported system, ultra-low-cost tags composed only of a piezoelectric disc, a Zener diode, and an LED harvest vibration energy and emit sparse visible light spikes without requiring batteries or RF radios; these optical spikes are captured by an event camera and classified by SNN models optimized with the EONS framework. Across five device classes, the system is reported to achieve 94.9% average classification fitness while explicitly analyzing latency-accuracy trade-offs induced by temporal binning [2508.11640].

## 1. System concept and operating domain

Vibe2Spike is designed for dense, low-cost sensing in intelligent environments where conventional sensing stacks are constrained by battery maintenance, wireless transmission overhead, and data processing complexity. Its central premise is that device vibration signatures can be sensed without an onboard power source and transmitted without RF, provided that vibration energy is directly converted into sparse optical emissions and interpreted with event-driven computation.

The architecture has three layers. First, a batteryless vibration tag attached to a vibrating object generates LED flashes. Second, these flashes form a VLC link rather than a radio link. Third, a neuromorphic receiver records the optical activity as asynchronous events and classifies the resulting temporal pattern with an SNN. The system therefore unifies battery-free sensing, VLC, event-based vision, and neuromorphic computation into a single end-to-end pipeline.

A common misconception is to treat Vibe2Spike as a conventional wireless sensor node with an unconventional classifier. The reported implementation is structurally different: the tag does not use a battery, a microcontroller, a radio, or active electronics. The transmitted signal is also not a high-bandwidth optical stream in the usual VLC sense, but a sparse on/off signaling mechanism for vibration events.

## 2. Batteryless tag and vibration-to-light transduction

The batteryless tag contains a piezoelectric disc, a Zener diode, an LED, and a 3D-printed housing with a dual-spring mechanical structure. The paper specifies a 40 mm piezoelectric disk and a 0.5 W, 12 V Zener diode. The tag is described as a parallel circuit comprising the piezoelectric disk, Zener diode, and LED.

The piezoelectric disc serves simultaneously as energy harvester and transducer. When the attached device vibrates, the piezo deforms and generates an AC voltage/current due to mechanical strain. The Zener diode is placed in parallel with the LED and piezo in order to regulate voltage, protect the LED from surges, and help rectify or constrain the AC behavior so that current flows in the proper direction for LED emission. The LED emits visible flashes when the piezo-generated voltage exceeds the LED’s forward threshold. As vibration continues, the LED produces a pattern of visible spikes whose temporal structure reflects the vibration signature of the source device [2508.11640].

Mechanical amplification is achieved without an electronic amplifier. The housing includes an external spring that provides constant preload and keeps the piezo pressed into the housing, and an internal spring that is stiffer and located under the piezo center, enabling asymmetric deformation. This dual-spring design increases strain across the piezo disk when the device vibrates. In effect, the tag behaves as a vibration-powered optical spike generator.

This design choice is significant because it pushes signal generation into the mechanics of the tag rather than into powered mixed-signal circuitry. A plausible implication is that the system’s energy model is dominated less by local sensing electronics than by the physical coupling between vibration source, piezo deformation, and optical observability.

## 3. VLC link and event-camera reception

Vibe2Spike communicates through visible light communication. The LED blink pattern is the transmitted signal, and the receiver is an event camera rather than a frame-based imager. The paper uses the Prophesee EVK3HD event camera. This choice exploits the fact that event cameras generate outputs only when pixel brightness changes, making them naturally compatible with sparse LED flashes.

Each event is represented as
$$
\text{Event} = (x, y, \text{polarity}, \text{timestamp}),
$$
where \(x,y\) are the pixel coordinates, polarity indicates ON or OFF, and timestamp is given at microsecond resolution. When the LED flickers, multiple adjacent pixels register brightness changes and produce asynchronous ON/OFF events. The paper emphasizes microsecond temporal resolution, sub-millisecond latency, low power, and the suitability of event cameras for sparse optical spikes [2508.11640].

The use of VLC is motivated by several system-level properties stated in the paper: avoidance of RF spectrum congestion, immunity to RF interference, physical confinement by line-of-sight, a privacy advantage because light does not pass through walls, and direct compatibility with event cameras. These properties define the practical deployment envelope of Vibe2Spike. The same properties also delimit it: line-of-sight is not an incidental detail but part of the communication model.

During preprocessing, the system performs a 1-minute calibration phase to identify the pixel with the highest cumulative event count, assumed to best represent the LED signal. This is presented as a training heuristic intended to improve signal-to-noise ratio. The authors note that the trained model is intended to be more general and compatible with pixel-parallel processing in future versions, which indicates that the present implementation still relies on a selected-pixel abstraction rather than full-scene event processing.

## 4. Temporal binning, spike encoding, and SNN inference

The raw event stream is converted into a structured representation through fixed-size temporal binning. The paper uses time windows of 50 ms and counts total ON events and total OFF events in each bin. For \(N\) bins, this yields a feature vector
$$
\mathbf{f} \in \mathbb{R}^{2N},
$$
with the two dimensions per bin corresponding to ON and OFF counts. The reported examples are \(N=10\), corresponding to 0.5 s and 20 features, and \(N=50\), corresponding to 2.5 s and 100 features.

The classification pipeline is explicitly described as: event camera recording, temporal binning, Argyle-4 spike encoding, SNN processing, and Winner-Take-All decoding. Argyle-4 encoding normalizes input values to a range determined by the training data, processes each ON/OFF pair every 3 simulation timesteps, and produces two complementary spike magnitudes, \(x\) and \(1-x\), so that each pair yields four spikes total. For 50 bins, the paper states that 50 ON counts plus 50 OFF counts give 100 values, encoded over 150 timesteps, with exactly 200 spikes injected per sample.

The output layer uses Winner-Take-All decoding:
$$
\hat{c} = \arg\max_k S_k,
$$
where \(S_k\) is the spike count of output neuron \(k\). Each output neuron corresponds to one device class, and inference selects the class with the most spikes over the simulation window [2508.11640].

SNN optimization is performed with EONS, “Evolutionary Optimization of Neuromorphic Systems,” which uses a genetic algorithm. Candidate networks are evaluated for classification accuracy and architecture suitability, then improved via selection, mutation, and crossover. The search space includes network topology, number of neurons, recurrent connections or loops, sparsity, synaptic weights, and time constants. For each binning configuration, the paper states that a population of 999 randomly initialized SNNs was generated and evolved over multiple generations. This places the classifier design itself inside the optimization loop rather than treating network architecture as fixed.

## 5. Experimental configuration and reported performance

The reported evaluation is a five-class device-recognition problem. Four classes are explicitly shown in the figures: palm sander, cordless drill, handheld vacuum, and inflator blower. The fifth class is implied by the paper’s statement of five classes, although the visible figures in the provided summary do not name it. The setup places the event camera about 30 cm from the tagged device with line-of-sight maintained. The camera is reported to detect the LED even at oblique viewing angles up to 85° angular offset from the optical axis. The dataset contains five trials per device and 25 minutes of labeled data in total.

The paper reports F1 accuracy or validation accuracy, confusion matrices, and classification fitness across binning configurations. It notes only minor confusion between blower and vacuum, attributed to similar vibration profiles. The best overall result is given as 50 bins, 2,500 ms latency, and 94.88% validation score, reported in the text as approximately 94.9%. The system as a whole is summarized as achieving 94.9% average classification fitness across five classes [2508.11640].

| Latency | Bins | Validation |
|---|---:|---:|
| 250 ms | 5 | 81.16% |
| 500 ms | 10 | 88.92% |
| 2,500 ms | 50 | 94.88% |
| 5,000 ms | 100 | 93.52% |

These results establish a clear latency-accuracy trade-off. Short windows provide faster response but poorer stability for sparse signals, while medium windows provide stronger aggregation and the highest reported performance. Longer windows add delay and show diminishing returns. The paper therefore treats temporal binning not merely as preprocessing, but as a design parameter controlling speed, noise robustness, and temporal expressiveness.

## 6. Interpretation, applications, and relation to adjacent neuromorphic work

The main contribution of Vibe2Spike is not only the reported classification score, but the specific division of labor across mechanics, optics, sensing, and inference. Vibration is converted directly into visible optical spikes at the edge; event cameras preserve sparse timing structure; and classification is delegated to an SNN optimized for temporal data. This division avoids battery maintenance, avoids RF radios, and keeps the sensing modality event-driven throughout.

The paper identifies several advantages: low power, low hardware complexity, less than \$1 per tag according to the paper, wireless and scalable operation via VLC, event-driven efficiency, privacy and security benefits due to the spatial confinement of visible light, and suitability for workshop monitoring, appliance recognition, predictive maintenance, context-aware smart spaces, and industrial sensing. It also records several limitations: dependence on vibration strength and consistency, latency introduced by temporal binning, computationally intensive EONS training, sensitivity to hyperparameter tuning, a current pixel-selection heuristic, and open issues involving occlusion robustness, variable lighting, multiple tags or overlapping signals, and neuromorphic hardware deployment.

Within the broader literature, Vibe2Spike occupies a distinct place among spike-based sensing systems. An earlier neuromorphic vibration-anomaly pipeline used accelerometer recordings, a Gammatone filterbank, Asynchronous Delta Modulation, and a balanced spiking neural network for online unsupervised anomaly detection, including a proof-of-concept on the DYNAP-SE processor [2106.00687]. By contrast, Vibe2Spike performs batteryless vibration capture through a passive optical tag and event camera. A plausible implication is that these systems represent two complementary trajectories in neuromorphic vibration sensing: one built around direct electrical acquisition and online anomaly detection, the other around batteryless optical signaling and device-class recognition.

Vibe2Spike also aligns with a more general trend toward spike-native processing rather than early conversion to conventional image-like representations. In spike video-language alignment, SPKLIP argues that direct spike-text alignment is preferable to forcing spikes into an RGB-like pipeline because sparse, asynchronous streams contain continuous spatiotemporal cues that can be lost under naive conversion [2505.12656]. Although the task differs, the methodological affinity is evident: Vibe2Spike similarly preserves event-native structure through event-camera acquisition, temporal binning, spike encoding, and SNN inference instead of reconstructing a dense visual surrogate.

For this reason, Vibe2Spike is best understood as a batteryless neuromorphic sensing framework rather than merely a low-power classifier. Its defining feature is the end-to-end translation of mechanical vibration into a sparse optical spike stream that remains event-centric from transduction to decision.

Source: https://www.emergentmind.com/topics/vibe2spike