Abstract
Nowadays, Internet of Things (IoT) has been becoming a hot research topic. Being an important part of Internet of Things, the wireless sensor networks collect various types of environmental data and construct the fundamental structure of the IoT applications. In order to find out the characteristics of the environmental data, in this paper, we focus on four types of these sensor data: temperature, humidity, light and voltage, and employ statistical methods to analyze and model these sensor data. The results of our research can be used to solve the missing sensor data estimation problem which is inevitable in the wireless sensor networks.
Introduction
The Internet of Things (IoT) relies heavily on Wireless Sensor Networks (WSNs) to monitor physical environments. However, these sensors are frequently deployed in harsh outdoor conditions where hardware failure and packet loss are common, leading to significant gaps in collected data. While various imputation algorithms exist to fill these gaps, they often fail to account for the unique statistical properties of different environmental variables. This study aims to identify the specific mathematical distributions that best fit temperature, humidity, light, and voltage data to improve the accuracy of missing data estimation.
Methodology
The research utilizes the Intel Berkeley Research Lab dataset, which includes readings from 54 sensors collected every 31 seconds. The analysis focuses on a 24-hour window to capture daily cycles. The methodology consists of two primary stages:
- Visual and Correlation Analysis: Plotting diurnal cycles to identify patterns and inter-variable relationships.
- Statistical Distribution Fitting: Using the Kolmogorov-Smirnov (K-S) goodness-of-fit test to determine the analytical distributions that most accurately represent each sensor data type.
Data Analysis and Insights
Initial visual analysis reveals distinct patterns for each environmental variable. As shown in Figure 1, temperature and light follow predictable diurnal cycles, while humidity exhibits a strong inverse correlation with temperature.
A key observation is the relative stability of temperature and light compared to the inherent volatility of voltage. While environmental factors are driven by solar cycles, sensor voltage fluctuates unpredictably, making it the most challenging variable to model. The inverse relationship between temperature and humidity suggests that missing values in one could potentially be estimated using the other as a proxy.
Statistical Results
The K-S goodness-of-fit tests identified the following best-fitting distributions for the sensor data:
- Temperature: Generalized Pareto distribution.
- Humidity: Johnson SB distribution.
- Light: Power Function distribution.
- Voltage: Frechet distribution.
While the Frechet distribution was the best fit for voltage, the fit was notably weaker than for the other variables due to voltage's high instability. This confirms that a "one size fits all" approach to sensor data modeling is insufficient.
Conclusion and Future Work
This study demonstrates that different environmental variables in WSNs possess unique statistical "fingerprints." Understanding these distributions is crucial for developing more accurate data reconstruction algorithms. Future research will focus on creating specialized missing-data estimation models that incorporate these identified distributions, with a particular emphasis on handling highly unstable variables like voltage.
Summary updated