An AIoT motion classification system that performs real-time inference on simulated MPU6050 sensor data using an Edge Impulse model and publishes the results to ThingsBoard through MQTT.
The project demonstrates an end-to-end Edge AI workflow, from sensor data generation and feature extraction to model training, WebAssembly inference, real-time telemetry, dashboard visualization, and motion-based alerts.
The system classifies MPU6050 accelerometer and gyroscope readings into three motion states:
- Idle — no significant movement
- Move — moderate movement
- Shake — rapid or intense movement
Sensor data is generated using a simulated MPU6050 device in Velxio and published through MQTT.
The gateway receives the sensor readings, extracts both raw and differential motion features, performs inference locally using an Edge Impulse model exported as WebAssembly, and forwards the prediction results to ThingsBoard.
ThingsBoard is used to visualize the detected motion state and create an alarm whenever a Shake state is detected.
┌──────────────────────┐
│ Velxio MPU6050 │
│ Simulator │
└──────────┬───────────┘
│
│ MQTT
▼
┌──────────────────────┐
│ MQTT Broker │
│ Raw Sensor Data │
└──────────┬───────────┘
│
▼
┌──────────────────────────────┐
│ TypeScript Edge Gateway │
│ │
│ - Receive sensor readings │
│ - Extract features │
│ - Run Edge Impulse model │
│ - Generate prediction │
└──────────┬───────────────────┘
│
│ MQTT Telemetry
▼
┌──────────────────────┐
│ ThingsBoard │
│ │
│ - Live dashboard │
│ - Motion state │
│ - Confidence │
│ - Shake alarm │
└──────────────────────┘
The dataset was generated using a simulated MPU6050 sensor.
For each motion class, multiple data collection runs were performed while changing the simulated sensor behavior to represent the intended motion pattern.
The raw sensor features consist of three-axis accelerometer and gyroscope measurements:
accX
accY
accZ
gyroX
gyroY
gyroZ
In addition to the raw sensor values, differential features are calculated between consecutive samples. These features represent the change in each sensor measurement over time:
dAccX
dAccY
dAccZ
dGyroX
dGyroY
dGyroZ
For example:
dAccX = currentAccX - previousAccX
dGyroX = currentGyroX - previousGyroX
Using these differential features allows the model to capture how rapidly the sensor readings are changing, providing additional information about the dynamics of the movement rather than relying only on the absolute sensor values.
The three target labels are:
Idle
Move
Shake
The collected datasets were uploaded to Edge Impulse and used for training and evaluating the motion classification models.
The models were trained and evaluated using Edge Impulse.
Multiple model configurations were tested to compare their classification performance. The selected model was exported from Edge Impulse as a WebAssembly model, allowing inference to be performed locally inside the TypeScript gateway.
The exported model consists primarily of:
edge-impulse-standalone.js
edge-impulse-standalone.wasm
This allows motion classification to run locally without sending the raw sensor data to a cloud inference service.
The TypeScript gateway subscribes to the raw MPU6050 MQTT topic:
iot/mpu6050/raw
For each incoming sample, the gateway processes the raw accelerometer and gyroscope values and calculates the corresponding differential features.
The resulting feature vector contains both the current sensor readings and their changes relative to the previous sample.
The feature vector is then passed to the Edge Impulse model for inference.
The model returns the probability for each motion class, and the class with the highest confidence is selected as the current motion state.
Example prediction:
{
"motion": "shake",
"confidence": 0.97
}After classification, the gateway publishes telemetry to ThingsBoard through MQTT.
The telemetry contains the sensor measurements together with inference information such as:
motion
confidence
isShake
modelResults
ThingsBoard is used to:
- visualize the detected motion state in real time;
- monitor model predictions and sensor values;
- trigger an alarm when the detected state is
Shake.
.
├── data/
│ ├── idle/
│ ├── move/
│ └── shake/
│
├── models/
│ ├── model-1/
│ │ ├── edge-impulse-standalone.js
│ │ └── edge-impulse-standalone.wasm
│ └── model-2/
│ ├── edge-impulse-standalone.js
│ └── edge-impulse-standalone.wasm
│
├── src/
│ ├── gateway.ts
│ ├── gateway2.ts
│ ├── model.ts
│ └── model2.ts
│
├── simulator/
│ └── mpu6050.c
│
├── docs/
│ └── report.pdf
│
├── .gitignore
├── package.json
└── README.md
- Bun / Node.js
- MQTT broker
- ThingsBoard
- Velxio or another MPU6050 data source
- Edge Impulse exported WebAssembly model
Install the project dependencies:
bun installBefore running the gateway, configure the ThingsBoard connection and device credentials in the gateway according to your ThingsBoard installation.
Run the gateway using Bun:
bun run src/gateway.tsThe gateway will:
- initialize the Edge Impulse WebAssembly model;
- connect to the MQTT broker;
- subscribe to MPU6050 measurements;
- extract raw and differential motion features;
- perform local motion classification;
- publish the prediction results to ThingsBoard.
- TypeScript
- Bun / Node.js
- MQTT
- Edge Impulse
- WebAssembly
- ThingsBoard
- MPU6050
- Velxio
- TinyML / Edge AI
The complete coursework report, including model evaluation results, confusion matrices, system setup, ThingsBoard dashboard, and rule-chain configuration, is available in the docs directory.
This project was developed collaboratively by:
Developed as part of the Internet of Things course project, focusing on AIoT, TinyML, and Edge AI inference.