Mat-ID

MatID is a handy, AI-powered device that can identify materials in real-time using a combination of vibration (IMU), sound (microphone), and light sensors. With a trained CNN running directly on the MAX78000, it instantly recognizes materials like wood, plastic, and paper from a simple tap, making material analysis fast, accurate, and intuitive.

Components

Collecting real time sensor data:

The first step was to gather data from the MAX78000FTHR board’s sensors. MatID uses three sensors:

• IMU (Inertial Measurement Unit): Detects vibrations and subtle movements when a material is tapped. Different materials produce distinct vibration patterns.
• Microphone(On-board sensor): Captures the tapping sound, which varies in intensity and frequency for different materials.

LDR (Light Dependent Resistor): Measures surface reflectivity, giving information about the material’s appearance.

Different materials—Wood, Plastic, and Notebook/Paper were tapped and 20 sets of readings for each were captured, creating a diverse dataset. Each reading included the IMU axes (ax, ay, az), LDR value, and microphone amplitude..

Creating the Dataset:

Once data was collected, it was organized into a CSV file with the following structure:

• ax, ay, az → IMU readings for vibration detection
• LDR → Surface brightness/reflectivity
• MIC → Tapping sound intensity
• Label → Material type (Wood, Plastic, Notebook)

This dataset allowed to feed the collected sensor signals into AI models for training.

Training AI Models:

To make MatID intelligent, I trained AI models using the dataset:

Random Forest Classifier (Initial Model):

• A simple yet effective model to classify materials.
• Gave around 85–90% accuracy, enough to verify the model worked with our data.

Convolutional Neural Network (CNN, On-Board Model):

• Convolutional Neural Network (CNN, On-Board Model):
• A small CNN was trained using TensorFlow to perform edge inference.
• After training, we converted the CNN to a C array format that can be directly loaded into the MAX78000 board.
• This step ensures that all AI computation happens locally on the board, without needing a PC.

Deploying the Model to MAX78000:

• The trained CNN model, now in a C array format, was uploaded to the MAX78000FTHR board.
• The board’s onboard MAX78000 CNN accelerator runs the inference in real-time.
• This makes MatID a fully self-contained, edge AI system, capable of instant material detection without any cloud dependency.https://youtube.com/shorts/c2WENsLJmww?feature=share

Real-Time Material Detection:

• With the model deployed, I tested MatID by tapping materials in random sequence of Wood, Plastic, Notebook.
• IMU, LDR, and microphone values are displayed live.
• The detected material is shown in the next line.
• This demonstrates how the on-board CNN interprets real sensor data in real-time.

Finishing Simulation and Validation

• Taps are detected and classified one after another with small intervals.
• Real-time outputs mimic what would happen in a fully deployed scenario.
• This approach allows testing and demonstration without any errors from flashing or hardware limitations.

Summary

• MatID successfully combines:
• On-board AI using MAX78000 CNN
• Multi-sensor inputs (IMU + MIC + LDR)
• Portable, battery-powered design
• It can instantly identify materials through simple taps, making material analysis fast, accurate, and intelligent.