IGADA-IoT: IoT Sensor Energy Optimization in Wireless Sensor Networks Driven by Automatic Data Augmentation

This article presents a research contribution focused on optimizing data augmentation for time-series data from IoT sensors. The core innovation is a systematic framework that moves beyond empirical, single-generator approaches.

Moving Beyond Single-Generator Limitations

The paper identifies a critical flaw in existing data augmentation methods for wireless sensor networks: their reliance on a single generator and a fixed, empirically set number of samples. This approach is ineffective because it cannot dynamically address varying information gaps—the difference between what the model currently knows and what it needs to learn from the data. A single generator cannot adapt to the heterogeneous nature of the data or the model's evolving learning requirements.

The Multi-Generator Hierarchical Strategy

The proposed solution, IGADA-IoT, introduces a hierarchical multi-generator collaboration and scheduling strategy (HMGCS). This is the framework's core mechanism for generating diverse and targeted samples.

• It treats data augmentation as a resource allocation problem, deploying different types of generators (which likely include various perturbation, synthesis, or crossover techniques) based on their assessed capabilities.
• The strategy explicitly maps different generators to specific "information gap" profiles, enhancing the targetedness of augmentation. Instead of applying all generators uniformly, it schedules and allocates their outputs across multiple rounds to systematically reduce model uncertainty.

Joint Evaluation to Prevent Augmentation Errors

A second key component is the information gap-model performance joint evaluation and closed-loop method (IGMP-EC). This acts as a feedback mechanism to guide the HMGCS.

• It continuously evaluates two interconnected factors: the remaining information gap and the model's performance on downstream tasks.
• This dual evaluation creates a closed-loop control system that determines whether, how much, and what kind of augmentation to apply next. The primary insight here is that augmentation decisions cannot be based solely on accuracy or solely on theoretical information measures; the two must be considered jointly to avoid the risks of under-augmentation (insufficient training data) and over-augmentation (introducing noise or bias that degrades model performance).

Experimental Validation and Impact

The results validate the framework's effectiveness, showing a 7.27% average accuracy improvement over multiple downstream models compared to baselines. The most significant claim is an 8.67% improvement over advanced data augmentation methods, suggesting that the structured, adaptive approach outperforms even recent sophisticated techniques. The use of both public benchmark datasets (UCR Archive) and real-world deployment data strengthens the argument for the method's generalizability beyond controlled experiments.