Abstract: Air conditioners are widely used in family homes all over the world. However, the side effects of using air conditioners and dehumidification can cause health problems if people remain in low-humidity environments. This paper traces the development of a software application and system to create an intelligent humidifier that automatically turns on or off for convenience or for those who cannot engage manual control. We applied our application to a humidifier for several days and conducted a qualitative evaluation of the approach. Results affirmed the usability and capacity of our automatic control system.
Introduction
Modern climate control often comes at a cost to respiratory health. While air conditioning (AC) units provide essential cooling and heating, they simultaneously strip moisture from the air. Prolonged exposure to low-humidity environments can lead to sore throats, chapped lips, and increased vulnerability to viruses like influenza. This paper introduces "HC" (Humidity Conditioner), an IoT-based system designed to transform standard humidifiers into intelligent, automated appliances that maintain optimal indoor environments without constant human intervention.
Problem Statement
Most consumers currently face a binary choice: manual humidifiers that require frequent monitoring to prevent over-humidification (which can lead to mold growth) or expensive "smart" units that often lack sophisticated predictive capabilities. There is a clear need for an affordable, retrofittable solution that can predict environmental changes and automate control logic based on real-time data.
System Architecture and Methodology
The "HC" system is built on a full-stack IoT architecture comprising three primary layers:
- Hardware Stack: The central hub is a Raspberry Pi 3, which interfaces with a DHT22 digital sensor to monitor ambient temperature and humidity. Control of the physical appliance is achieved via a power strip equipped with a built-in relay, allowing the Raspberry Pi to programmatically toggle power to any standard humidifier.
- Software & Cloud Integration: The mobile interface was developed using Thunkable, a low-code platform, and utilizes Firebase as a real-time database. This allows for seamless data synchronization between the sensors, the cloud, and the user's mobile device.
- Machine Learning Models: To move beyond simple threshold-based triggers, the system employs two distinct ML approaches:
- Polynomial Regression: Used to predict future humidity patterns by analyzing historical trends and weather data.
- Support Vector Machine (SVM): A classification algorithm that operates in "Auto" mode to decide whether the power relay should be triggered based on current and predicted environmental features.
Datasets and Training
The researchers utilized sixteen distinct datasets containing environmental variables such as temperature, wind speed, and humidity. These time-series datasets were used to train and validate the predictive models, specifically testing how different environmental "features" (e.g., the relationship between wind speed and indoor humidity) impacted the accuracy of the automated control logic.
Key Insights and Results
The study yielded several critical findings regarding the optimization of automated environmental control:
- Optimal Complexity: In predicting indoor humidity, the authors found that a polynomial degree of 3 provided the best balance between accuracy and generalization. While a linear model achieved an accuracy of 0.787, the degree 3 polynomial reached 0.942.
- Classification Reliability: The SVM models used for the binary decision-making (turning the device On vs. Off) achieved a perfect accuracy score of 1.0 across multiple feature sets. This suggests that SVMs are exceptionally robust for this specific IoT application.
- The Value of Retrofitting: A major advantage of this system is its ability to be applied to existing "dumb" hardware, significantly lowering the cost of entry for smart home automation.
Conclusion and Future Work
Qualitative testing over several days confirmed that the HC system effectively maintains comfort levels autonomously. However, limitations remain. The current system uses static thresholds rather than learning individual user preferences. Future iterations will focus on implementing reinforcement learning to study manual user overrides, allowing the system to adapt to specific personal comfort profiles over time. Additionally, the researchers aim to move from an external "add-on" relay system to direct hardware integration.
Summary updated