TL;DR
Smartwatch battery life is a major bottleneck for user adoption. This paper introduces a novel deep learning model that analyzes real-world smartwatch usage data to understand battery drain patterns, achieving 85.3% accuracy in predicting high battery consumption events. The research also proposes a method to index battery quality changes over time, offering valuable insights for device manufacturers and users.
What Problem Does This Paper Address?
Smartwatches are increasingly popular, but their limited battery capacity remains a significant hurdle to widespread adoption. While previous research has explored smartwatch battery usage, there's a lack of comprehensive studies analyzing diverse smartwatch brands and user behaviors in real-world settings. This paper aims to bridge this gap by providing a deeper understanding of how users interact with their smartwatches and how these interactions impact battery life, ultimately seeking to improve battery efficiency.
Methods
The authors employ a multi-pronged approach:
- Clustering for Pattern Identification: They use k-means clustering to identify common patterns in smartwatch battery utilization throughout the day, categorizing users based on their charging and discharging habits.
- Novel Deep Learning Model (FCNN): A low-parameter, fully convolutional neural network (FCNN) is developed. This model is designed to be transparent, allowing for the extraction of insights from its learned filters. It converts battery consumption into a binary classification problem (low vs. high consumption). Key innovations include using a limited number of filters in the first and last layers for interpretability and incorporating residual connections for improved performance.
- Battery Quality Indexing: A longitudinal study is conducted to quantify changes in smartwatch battery quality over time. This involves transforming time-series battery data into a two-dimensional representation and calculating a distance-based index to rank battery performance degradation.
Datasets
The study utilizes a real-world dataset collected via the "insight for wear" Android application. This dataset comprises data from 832 users across different geographic locations and smartwatch brands. The data includes information from eight sensors: screen usage, heart rate, Bluetooth connection, ambient light, activity state (from Google Fit), notification, and battery usage. For the deep learning model, data from 671 users with at least 10 days of battery data was used, and for the longitudinal study, data from 62 users with over six months of battery information was analyzed.
Key Insights and Ideas
- Real-world diversity matters: Analyzing a diverse set of users and brands is crucial for understanding battery usage, as lab settings and limited user groups can lead to biased conclusions.
- Deep learning for transparency: Developing a deep learning model that is not a "black box" is essential for extracting actionable insights. The FCNN's filter analysis allows researchers to pinpoint which sensors and activities most impact battery life.
- Battery quality degrades over time: The proposed indexing method reveals that smartwatch battery quality can change significantly over their lifetime, with variations observed across different brands and operating system versions.
- User behavior is key: Understanding common charging and discharging patterns can inform developers on how to optimize background processes and application behavior for better energy efficiency.
Key Results
- The proposed FCNN model achieves 85.3% accuracy in predicting high battery discharge events, outperforming other machine learning algorithms.
- Screen usage and notifications are identified as significant contributors to battery drain, as revealed by the analysis of the FCNN's learned filters.
- Newer Wear OS versions (than 7.1.1) tend to drain batteries faster, suggesting potential issues with software updates or increased application demands.
- Motorola and Sony smartwatches show improved battery life over time, indicating effective optimization strategies in their software.
- The highest battery drain occurs during working hours (9:00-17:00), while sleep hours (22:00-6:00) are identified as the best time for batch operations due to users typically plugging in their devices.
- A significant difference in charging behavior is observed between weekends and weekdays, with Saturday showing a notable difference in plug-in frequency.
Limitations and Future Work
- Brand generalizability: While the study includes multiple brands, the findings regarding battery quality changes over time are based on a limited set of brands (Sony, Motorola, LGE, Asus, Huawei) due to data availability. The authors acknowledge that these findings may not be generalizable to all smartwatch brands.
- Weekend bias: The analysis of weekend vs. weekday charging behavior might be biased by the fact that some countries have different weekend structures (e.g., Sunday as a working day), which could skew results towards Saturday.
- Future Work: The authors plan to integrate their findings into a battery manager application on smartwatches or other battery-powered devices to observe the impact on battery behavior. They also suggest further research into the interaction between sensors and battery utilization.
Summary updated