01
Continuous Data
Sensors produce data all the time, even when nothing important is happening.
Adaptive Edge Intelligence
CAAD-Gate
Adaptive Edge Intelligence for Low-Power IoT Devices
CAAD-Gate is a reusable adaptive data-gating framework for ESP32-class IoT devices. It learns normal sensor behaviour and forwards only meaningful, abnormal, or contextually important data to machine learning, alerts, logging, or communication.
CAAD-Gate helps IoT devices know when not to process.
65.33%
Estimated processing reduction
ESP32
Designed for low-power devices
3 Modules
Sound, motion and environment
The problem
Small IoT devices cannot process everything.
IoT devices continuously receive sensor data, but most data is normal. Running machine learning, communication, alerts, or logging for every data window wastes resources.
02
Mostly Normal
Only a small part of the sensor stream may contain a useful or abnormal event.
03
Fixed Thresholds
One fixed threshold cannot fit every user, room, machine, or environment.
04
Expensive Processing
Unnecessary ML and communication can waste CPU, memory, battery, and bandwidth.
The solution
An adaptive decision layer before heavy processing.
CAAD-Gate learns normal sensor behaviour, measures change, adjusts its threshold using context, and decides whether each sensor window should be forwarded or rejected.
“CAAD-Gate does not replace ML. It protects ML from unnecessary data.”
1
Sensor Windowing
Continuous sensor data is divided into small windows.
2
Lightweight Features
Level, variation, rate of change, peak, and duration are extracted.
3
Normal Behaviour Learning
The device learns a baseline for the current sensor and environment.
4
Adaptive Threshold
The decision boundary changes according to learned normal behaviour.
5
Context-Aware Adjustment
Battery, urgency, user mode, and environment can change sensitivity.
6
Gate Decision
Normal data is rejected. Useful or abnormal data is forwarded.
Sound Sensor Module
Environment Sensor Module
Motion Sensor Module
CAAD-Gate Core Engine
Normal Model · Feature Scoring · Adaptive Threshold · Context Layer · Statistics
Normal Model · Feature Scoring · Adaptive Threshold · Context Layer · Statistics
Gate CLOSED
Reject normal data
Reject normal data
Gate OPEN
Forward useful data
Forward useful data
Applications
One reusable core for different IoT products.
The sensor adapter can change while the CAAD-Gate core remains the same.
S
Sound Awareness Node
Detects meaningful sound windows before classification. Suitable for alarms, knocks, assistive hearing, and smart-home sound monitoring.
Sound ModuleM
Eldercare Motion Node
Learns normal movement and forwards possible fall, abnormal motion, or long inactivity windows for further analysis.
Motion ModuleE
Machine & Environment Monitor
Opens the gate during sudden change, abnormal drift, or unusual operating conditions in industrial or environmental systems.
Environment ModuleExperimental validation
Measured processing reduction with an honest trade-off.
The first validation used the ESC-50 environmental sound dataset. The adaptive gate reduced unnecessary frame-level processing while keeping recognition results close to the fixed-threshold baseline.
2,000
Total WAV files
1,600
Training recordings
400
Testing recordings
65.33%
Estimated frame-level processing reduction
| Method | Coverage | Processing Reduction | MFBE Top-1 | MFBE Top-5 |
|---|---|---|---|---|
| Fixed baseline | 93.50% | — | 25.94% | 46.29% |
| Adaptive CAAD-Gate | 78.50% | 65.33% | 24.50% | 46.04% |
Interpretation: lower coverage shows a tunable trade-off. A product may use a high-sensitivity mode for better event coverage or a low-power mode for stronger processing reduction.
Live demo
ESP32 sound-event demonstration.
The demo learns background sound, calculates an adaptive threshold, and opens the gate only when a meaningful sound event occurs.
CAAD-Gate Sound Demo
LIVE
Learning background noise
Energy0.16
Noise Floor0.14
Threshold0.42
Gate StatusCALIBRATING
Processing Saved0.00%
ReasonLearning background
1. Learn Background
The device measures the normal sound level.
2. Gate Closed
Normal background sound is rejected to save processing.
3. Event Detected
A clap, knock, alarm, or other meaningful event occurs.
4. Gate Open
The event is forwarded to ML, alerting, logging, or communication.
5. Return to Normal
The gate closes again and the saved-processing percentage increases.
Downloads
Developer-ready framework files.
Explore the complete V1 framework, reusable core, LilyGO demonstration sketch, and technical documentation package.
V1.0
CAAD-Gate Full Framework
Core engine, sensor modules, examples, and documentation.
Download ZIPV1.0
CAAD-Gate Core
Reusable sensor-independent adaptive decision engine.
Download CoreDemo
LilyGO Sound Sketch
ESP32/LilyGO sound-gate display demonstration.
Download SketchV1.0
Documentation Package
Theory, module reference, architecture, and experiment summary.
Documentation
Simple integration for embedded developers.
Each sensor module extracts lightweight features. The core engine uses those features to decide whether the gate should open or close.
Basic usage example
CAADFeatures features = sensorModule.extractFeatures(sensorData);
CAADDecision decision = gate.update(features);
if (decision.gateOpen) {
// Run ML, alert, communication, or logging
} else {
// Save processing
}Required Hardware
ESP32 or ESP32-S3, optional LilyGO display, sensor hardware, USB cable, and a laptop.
Required Libraries
Arduino ESP32 board package, TFT_eSPI for the display demo, and sensor-specific libraries.
Installation
Extract the framework, copy it into Arduino/libraries, restart the IDE, open an example, and upload.
No. It is a lightweight adaptive decision layer that runs before ML or alert processing.
No. Sound is the first validation domain. The framework is designed for sound, motion, temperature, air quality, and other sensors.
No. It decides whether data should be sent to machine learning, reducing unnecessary processing.
No guaranteed claim is made. It reduces estimated downstream processing load. Direct battery measurement depends on the final hardware.
For NBQSA judges
A practical embedded technology platform with evidence.
CAAD-Gate includes a working ESP32-based sound-gate demo, reusable framework files, sensor modules, documentation, experimental validation, and product applications across sound, eldercare motion, and machine or environment monitoring.
About & roadmap
Built for adaptive, efficient, ML-ready IoT products.
CAAD-Gate was developed by H.M. Thisara N. Herath as part of a broader research and product-development direction focused on adaptive edge intelligence for resource-constrained embedded devices.
V1 — Current
Reusable Core
Core engine, sound, environment and motion modules, ESP32 examples, and initial validation.
V2 — Next
Smarter Modes
Improved display, battery-aware modes, confidence scoring, class-specific thresholds, and event history.
V3 — Future
Connected Platform
Cloud dashboard, OTA updates, Arduino library release, multi-device aggregation, and edge-ML integration.
Contact
Collaboration, testing, and product integration.
Use this form for research discussion, developer testing, or possible product integration.
Project
CAAD-Gate — Adaptive Edge Intelligence
Developer
H.M. Thisara N. Herath
Focus
ESP32, low-power IoT, adaptive data gating, and edge intelligence