EWNA Hackathon Team 6

As the hackathon approaches its final stage, we focused on organizing our Hackaday project page and preparing our final presentation.

We reviewed our project logs, updated the project description, and organized the major features we developed throughout the hackathon, including sound classification, emergency email alerts, device disconnection monitoring, and risk prediction.

We also prepared the final demonstration flow so that our presentation clearly explains the complete system:

Microphone → YAMNet sound classification → risk/emergency analysis → Arduino UNO Q → alert

Our final goal is to demonstrate how edge AI can continuously monitor sounds in the home, recognize emergencies, identify developing risks, and notify caregivers when assistance may be needed.

Detecting an emergency is important, but preventing an emergency before it happens can be even more valuable.

We therefore added a risk detection feature based on sound patterns over time. Instead of reacting to a single sound event, the system continuously monitors repeated environmental sounds.

For example, YAMNet can recognize water-related sounds such as dripping or running water. A single water drip is usually normal, but if the system continuously detects water sounds over a certain period of time, it may indicate a possible water leak.

The system can then generate an early warning before the situation develops into a more serious problem.

This adds a second layer of home safety:

Emergency detection → respond to danger

Risk prediction → identify potential danger early

We added another safety feature to improve the reliability of our system.

Since the microphone and Arduino UNO Q are essential for detecting emergency sounds, a hardware disconnection could prevent the system from monitoring the environment correctly. To address this issue, we added a connection monitoring function.

If the microphone or Arduino UNO Q becomes disconnected, the system will detect the failure and automatically send an email notification to the emergency contact.

This allows the caregiver to know not only when an emergency sound occurs, but also when the monitoring system itself is no longer functioning properly.

USB Hub arrived! We tested our first basic pipeline on the hardware: Microphone audio input -> YamNET classification -> alert if classified label falls under danger classes. The hardware worked successfully which was really great.

We also added two new features 
1. Jump soundwave during idle: When the model isn't detecting danger, the LED will display the live audio spectrum, which makes the system more "lively"!. As soon as danger is detected, the screen will switch to icons describing the event

2. email notification: Sends email alerts when an event is detected, each message includes the event, confidence, device name, and local timestamp

Finished ML model selection and testing - A pretrained audio-classification model YAMnet https://huggingface.co/STMicroelectronics/yamnet

For now, we use the three classes - smoke_alarm, glass_break, and fall_thud as the "danger" labels, the others are normal

Model ran successfully on our laptops, but the arduino board is still setting up on the app lab. We're aiming to deploy on the board ASAP