Smart CSI Processing for Accruate Commodity WiFi-based Humidity Sensing

Yirui Deng,Deepak Mishra,Shaghik Atakaramians,Aruna Seneviratne

DOI: https://doi.org/10.1109/ICASSPW62465.2024.10626905

2024-09-12

Abstract:Indoor humidity is a crucial factor affecting people's health and well-being. Wireless humidity sensing techniques are scalable and low-cost, making them a promising solution for measuring humidity in indoor environments without requiring additional devices. Such, machine learning (ML) assisted WiFi sensing is being envisioned as the key enabler for integrated sensing and communication (ISAC). However, the current WiFi-based sensing systems, such as WiHumidity, suffer from low accuracy. We propose an enhanced WiFi-based humidity detection framework to address this issue that utilizes innovative filtering and data processing techniques to exploit humidity-specific channel state information (CSI) signatures during RF sensing. These signals are then fed into ML algorithms for detecting different humidity levels. Specifically, our improved de-noising solution for the CSI captured by commodity hardware for WiFi sensing, combined with the k-th nearest neighbour ML algorithm and resolution tuning technique, helps improve humidity sensing accuracy. Our commercially available hardware-based experiments provide insights into achievable sensing resolution. Our empirical investigation shows that our enhanced framework can improve the accuracy of humidity sensing to 97%.

Signal Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the insufficient prediction accuracy of existing WiFi - based humidity sensing systems (such as WiHumidity) in practical deployment. Specifically, the current WiFi - based humidity sensing systems face the following challenges:

1. **Low accuracy**: The average prediction accuracy of existing WiFi humidity sensing systems (such as WiHumidity) is only 79%, which is not ideal for practical applications.
2. **Noise problem**: There is noise in the received channel state information (CSI) signals, and this noise can significantly affect the judgment accuracy of the classifier.
3. **Improper resolution configuration**: Improper resolution settings of the humidity sensing system can also affect the prediction accuracy.

To address these problems, the author proposes an enhanced WiFi - based humidity detection framework, which improves the accuracy of humidity sensing by introducing innovative filtering and data processing techniques. Specific improvement measures include:

- **Denoising method**: A new denoising method is proposed, using the Hampel filter to remove outliers and combining the moving - average filter to remove high - frequency noise.
- **Selection of machine - learning algorithms**: Classifiers with non - linear kernel functions (such as quadratic support vector machine QSVM and k - nearest neighbor KNN) are selected to replace the linear support vector machine LSVM, thereby improving the prediction accuracy.
- **Resolution adjustment**: The influence of different resolutions on the prediction accuracy is studied, and it is found that appropriately reducing the resolution can improve the prediction accuracy and avoid model over - fitting.

Through these improvements, the experimental results show that this framework can increase the average accuracy of humidity prediction to 97%, which is significantly better than the existing WiHumidity system. This makes the framework suitable for monitoring and control applications in most indoor environments.