Energy Harvesting-Based Smart Transportation Mode Detection System via Attention-Based LSTM

Weitao Xu, Xingyu Feng, Jia Wang, Chengwen Luo, Jianqiang Li, Zhong Ming

Introduction: The Battery Bottleneck in Wearable Sensing

Continuous context sensing—the ability of a device to recognize whether a user is walking, driving, or riding a train—is a cornerstone of modern urban planning and personalized health tracking. However, this capability is currently limited by the "battery bottleneck." Traditional sensors like GPS are notoriously power-hungry, consuming up to 240mW, while even low-power accelerometers can significantly drain the limited battery capacity of smartwatches over a full day of continuous sampling.

Sensing Through Harvesting: Methodology

This research, published in IEEE Access, proposes a novel solution: using the Kinetic Energy Harvester (KEH) itself as a high-fidelity motion sensor. KEHs are typically used to convert human motion into electrical power. However, the voltage they produce is not just raw energy; it contains a unique "mechanical fingerprint" of the vibrations and movements experienced by the wearer.

Time-domain, frequency-domain, and spectrogram analysis of the KEH voltage signal, illustrating the distinct vibration signatures used for classification.

By treating the voltage output as a data source, the system achieves "sensing for free." The researchers developed a prototype using a Piezoelectric Energy Harvester (PEH) and a three-stage software pipeline:

  1. Signal Preprocessing: Noise reduction via a moving average filter and a specialized Stop Detection algorithm to handle stationary periods (e.g., a bus at a red light).
  2. Feature Engineering: Extraction of 27 window-based features, including statistical, time-domain, and frequency-domain metrics. The Minimum Redundancy Maximum Relevance (mRMR) algorithm was used to select the most informative features, such as spectral entropy.
  3. Attention-Based LSTM (AT-LSTM): A deep learning model that leverages the temporal strengths of LSTMs while using an attention mechanism to focus on the most distinctive vibration segments.

Experimental Results and Power Efficiency

The system was tested on a dataset of 38.6 hours of data collected from six volunteers in Sydney, Australia, covering seven modes: Walking, Running, Car, Bus, Train, Light Rail, and Ferry.

The AT-LSTM model achieved an overall accuracy of 97.05% with a 5-second sampling window, outperforming traditional machine learning baselines (SVM, KNN) by over 16%. Most significantly, the power consumption for sampling the KEH signal was measured at just 460µW. This is approximately 45 times more efficient than traditional accelerometer-based systems, which consume roughly 21mW for sampling alone.

Conclusion

The proposed system demonstrates that single-axis KEH signals, when processed with sophisticated deep learning models, can provide high-accuracy context sensing at a fraction of the power cost of traditional sensors. While future work is needed to miniaturize the hardware and test across diverse geographic locations, this "sensing through harvesting" approach offers a promising path toward truly autonomous, long-lasting wearable technology.

Summary updated

Abstract

Detecting the transportation mode of an individual’s everyday travel provides useful information in urban design, real-time journey planning, and activity monitoring. In existing systems, the accelerometer and GPS are predominant signal sources which quickly drain the limited battery life of the wearable devices. In this paper, we present an alternative approach for fine-grained transportation mode detection using kinetic energy harvester (KEH). We demonstrate the feasibility of using the output signal from the KEH device as the information source to achieve transportation mode detection. The proposed system is motivated by the fact that different transportation modalities produce distinctive motion patterns which are expected to leave distinctive patterns for context detection. To achieve fine-grained transportation mode detection, we design a transportation detection framework based on attention-based Long Short Term Memory (LSTM). We evaluate our approach using 38.6 hours of transportation data, which is collected from a total of six volunteers in three months’ time using our prototype. The evaluation results show that our approach is able to reach an overall accuracy of over 97% to detect fine-grained transportation modalities. In addition, our measurements show that the power consumption of the sampling KEH signal is only 460uW which significantly outperforms the existing transportation mode detection systems.
BibTeX
@article{xu2019energy,
  author = {Weitao Xu and Xingyu Feng and Jia Wang and Chengwen Luo and Jianqiang Li and Zhong Ming},
  title = {{Energy Harvesting-Based Smart Transportation Mode Detection System via Attention-Based LSTM}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2918555},
}
Figure