WifiSensing in Health care presentation.pptx

project proposal regarding wifi sensing

1. Introduction to Wifi - Sensing

2. Wifi Data Acquisition Wifi Data can be acquired using ESP-32 module and using Wifi.h Library the data can be processed and various important features like Channel State information (CSI) which gives amplitude and phases of the sub carrier frequencies.  

3. CSI Matrix   Amplitude of ith subcarrier Phase angle of ith subcarrier

4. Removing Noise by Rolling Mean Filter For example if we have an array of signal amplitudes: and we need to apply this filter at location to it then mathematically we write: -  

5. MODEL TRAINING SENDING MODEL Typical Models for Wifi -Sensing

6. Data Heterogenity Different locations have different working conditions. Therefore the model needs to be trained for each location separately and large amounts of annotated data required for each location.

7. Model Overfits to a specific location The model was trained in the living room and then tested for prediction in the other unseen rooms. The results showed that the model overfitted to the living room.

8. Overfitting Example The prediction accuracy was found highest for the living room as the model was trained in the living room. The other two unseen rooms donot show higher prediction accuracy due to overfitting of model to living room.

9. Too much Network Usage Sending too much training data from different locations to a centralized model will require a larger bandwidth which corresponds to higher network usage, thus increasing the communication Cost

10. Our Aim our goal here is to develop a collaborative training framework for WiFi sensing that shares knowledge learned from one set of locations to improve the prediction capability of models at other new locations.

11. SOLUTION: Federated Learning A machine learning technique in which the ML models are locally trained and then the updates in the parameters i.e weights are sent to the global central model instead of entire raw data.

12. Order: m-by-n No. of rows correspond to the no. of channels and no. of columns indicate no. of sub carriers in CSI.   k=m value depends on the number of activities to be predicted e.g. J=4 if 4 actions are predicted ( sitting, standing, lying falling)   Softmax function provides probability distribution of all 4 actions so that the total sum is 1.