WaterSmart Leak Detection System

Story

.

Water is a crucial natural resource. Water leakage in a water distribution system is main cause of water wastage.

I am building a ML based solution for detecting potential water leakage using Edge Impulse based on the vibration detection method and Thingy:53's accelerometer sensor. It will detect a water leakage and make the proper classifiction. The classification has two labels: leaking and no leak.

I used the nRF Edge Impulse mobile app to collect the accelerometer sensor data. Both the train and test data were collection and uploaded to the Edge Impulse portal for model training and binary building.

.

Hardware setup

.

I setup with a plastic host with a hole at one end for the water leaking simulation. The Thingy:53 kit was attached to the host using the host clamp and double-side tape.

.

.

.

Edge Impulse portal

.

I created a porject at the Edge Impulse portal and acquired the accelerometer sensor data as per instructed on the dashboard:

.

.

.

Data acquisition

.

In this project, I used the nRF Edge Impulse mobile app to do the data collection. It was fast and easy to collect the data using the mobile app. I connected the device and went ahead to do the data collection.

.

.

I collected few sets of training data with water leaking and no leak. I did the same for the test data too.

.

.

.

Here are the training data and test data collected and uploaded from the mobile app to the Edge Impulse portal.

.

.

.

Impulse design

.

After that I moved on to design an Impulse and train the model. I created a simple Impulse with just the Spectral Analysis as the processing block and NN Classifier as the learning block.

.

.

Here at the Spectral Analysis tab, parameters were set and saved. And feature generation was carried out.

.

.

The next step was doing the model training at NN Classifier tab.

.

.

Model testing

.

After training the model, I did the model testing step and here is the result:

.

.

Deployment

.

At the Deployment tab, I built the binary by selecting the Thingy:53 option.

.

.

I did not deploy the binary via Desktop, but deployed using the nRF Edge Impulse mobile app because it is easier to setup and deploy.

.

Demo

.

.

.

.

Here is the outcome of inferencing after deployment.

.

Summary

.

The project was built using one processing and one learning blocks with very simple parameters. I did not manage to do additional explorations with other type of processing blocks and/or other complex learning blocks.

Looking at the outcome, the classification is not very accurate. This can be the result of only few training data were collected and resulted in overfitting and little generalization.

And I did not write any code to do all the above. The Edge Impulse portal and nRF Edge Impulse mobile app have all the functionalities and facilities from data acquisition thru deployment which ease the process of bringing AI to the edge devices.

.

.