Abstract
Safety monitoring in power stations is critical yet challenging due to harsh lighting, visual occlusions, and the impracticality of wearable devices. This paper presents PowerSkel, a novel device-free framework that leverages WiFi Channel State Information (CSI) to estimate human skeletal poses. By integrating low-cost ESP32-based mutual sensing networks with a specialized "CKDformer" neural network, PowerSkel achieves a high accuracy of 96.27% (PCK@50). The system is robust to electromagnetic interference through Sparse Adaptive Filtering (SAF) and maintains performance in total darkness, offering a cost-effective and privacy-preserving solution for industrial safety.
Introduction
Power stations represent high-stakes environments where real-time safety monitoring is essential. Traditional surveillance methods often fail: video cameras are limited by poor lighting and blind spots, while wearable sensors are often rejected by staff due to discomfort or safety risks. WiFi-based sensing has emerged as a promising alternative, using the perturbations in wireless signals caused by human movement to infer activity without requiring cameras or wearables. However, the high-interference electromagnetic environment of power stations poses significant challenges for existing WiFi sensing techniques.
The PowerSkel Framework
The authors propose PowerSkel, an end-to-end pipeline tailored for industrial power scenarios. The framework consists of three primary components: a mutual sensing network, a denoising algorithm, and a deep learning architecture for pose estimation.
Low-Cost Mutual Sensing Network
Instead of relying on expensive specialized hardware, PowerSkel utilizes custom-built sensors based on the ESP32 IoT chip. These nodes form a "mutual sniffing" network where each device functions as both a transmitter and a receiver. This configuration creates a dense web of signal paths, capturing a richer spatial representation of the environment compared to traditional one-to-one links.
Sparse Adaptive Filtering (SAF)
To mitigate the intense electromagnetic interference typical of power infrastructure, the Sparse Adaptive Filtering (SAF) algorithm was developed. SAF exploits signal sparsity to reconstruct clean CSI data, effectively removing noise before the signal is processed by the neural network.
CKDformer Architecture
The core of the pose estimation engine is the CKDformer. It utilizes a Conformer backbone—a hybrid architecture that integrates Convolutional Neural Networks (CNNs) for local feature extraction with Transformers for capturing global context. This dual capability is crucial for distinguishing fine-grained movements (e.g., operating a switch) from overall body posture.
To further enhance performance, the system employs Collaborative Knowledge Distillation (CKD). During training, multiple "student" models share information, preventing the system from converging on the limitations of a single model and resulting in a more generalized mapping from CSI signals to 17 human skeletal keypoints.
Experimental Evaluation
The framework was evaluated in a real-world 10 kV power distribution room. Six participants performed various maintenance tasks, such as operating cabinet doors and moving handcarts. Ground truth skeletal coordinates were captured using a Kinect V2.0 camera. The resulting dataset comprised over 30,000 samples.
Key Results
PowerSkel demonstrates state-of-the-art performance in CSI-based pose estimation:
- High Accuracy: The model achieved a 96.27% PCK@50, outperforming standard architectures like ResNet50 (92.45%) and ShuffleNet (89.99%).
- Operational Efficiency: The system operates effectively at a sampling rate of 30 Hz, making it suitable for real-time deployment on low-power IoT hardware.
- Cost-Effectiveness: By using ESP32-based nodes, the hardware cost is reduced to approximately $3 per node, significantly lower than industrial-grade alternatives.
- Environmental Robustness: Unlike vision-based systems, PowerSkel maintains high precision in pitch-black conditions, as WiFi signals are independent of the visible spectrum.
Conclusion and Future Work
PowerSkel provides a robust, low-cost solution for device-free human pose estimation in complex industrial environments. While highly effective, the authors note that the system's reliance on Kinect for training labels introduces potential biases due to Kinect's own line-of-sight limitations. Future research will focus on improving label accuracy through multi-view setups and developing cross-domain adaptation techniques to allow the model to generalize to new environments without extensive retraining.
Summary updated