Smart Buildings: Water Leakage Detection Using TinyML

Area of Science:

• Environmental Science
• Computer Science
• Engineering

Background:

• Global water scarcity necessitates efficient water management strategies.
• Water wastage from undetected leakages in building pipelines is a significant issue.
• Aging infrastructure and inefficient practices exacerbate water loss.

Purpose of the Study:

• To develop an effective, low-intervention water leak detection method for smart buildings.
• To explore the application of edge computing and TinyML for real-time water management.
• To enhance water utilization efficiency in water-stressed regions.

Main Methods:

• Utilized an acoustic dataset of water leakages in PVC pipelines.
• Preprocessed acoustic data into scalograms for analysis.
• Applied transfer learning with five Convolutional Neural Network (CNN) variants (EfficientNet, ResNet, AlexNet, MobileNet V1, MobileNet V2).
• Optimized the EfficientNet model for deployment on an Arduino Nano 33 BLE edge device using quantization.

Main Results:

• The EfficientNet model achieved maximum testing accuracy of 97.45%, recall of 98.57%, precision of 96.70%, and F1 score of 97.63%.
• The quantized EfficientNet model demonstrated low inference time (1932 ms), minimal RAM usage (255.3 KB), and small flash requirement (48.7 KB).
• The proposed TinyML solution enables efficient, localized decision-making for water leak detection.

Conclusions:

• TinyML and edge computing offer a viable solution for real-time water leak detection in smart buildings.
• The developed system can significantly reduce water wastage with minimal human intervention.
• The efficient model deployment on edge devices paves the way for cost-effective and scalable smart water management systems.