Share
Embed
[CCNC2025] CSI Sampling for Room-by-Room Device Grouping in Practical Environments
Presented in IEEE CCNC 2025, Las Vegas, NV
S. Ishida, T. Murakami, and S. Otsuki
CSI Sampling for Room-by-Room Device Grouping in Practical Environments
IEEE Consumer Communications & Networking Conference (CCNC), Las Vegas, NV, Jan 2025
paper: N/A
pdf: N/A
Shigemi ISHIDA
January 12, 2025
Transcript
29 slides
-
CSI Sampling for Room-by-Room Device Grouping in Practical Environments Shigemi
ISHIDA*1, Tomoki MURAKAMI*2, Shinya OTSUKI*2 *1Future University Hakodate, JAPAN *2Nippon Telegraph and Telephone Corp., JAPAN -
Outline n Motivation n Previous work p CSI-based IoT device
grouping p Problem in practical environments n Proposed method p CSI sampling for practical environments n Evaluation n Summary Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025 2 -
Smart (?) House n Smart (?) speaker p Specify almost
everything to use services n IoT device coordination p Provides smart (?) life by manually configuring IoT devices 5 Which light? Turn off the light Restroom There is no place named restroom Bathroom! Livingroom Bedroom Bath- room Wash- room Dining/Kitchen Nature Remo Switchbot Switchbot Hub Echo Flex Echo Flex Smart power strip There is no device named bathroom Jan 12, 2025 Inamura-Ishida Lab, Future University Hakodate -
Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025 Put-and-Play (PnP)
IoT n Configures automatically after installation and becomes in service 6 I m in living room This TV is in the same room The light in this room is this one 4 people here (After taking shower) Turn off the light Bathroom becomes empty Okay! (Turn off the light in the bathroom) -
Goal 7 Network config Grouping Room-by- room grouping WPS Localize
Ask user the location when used Coordinate Turn of the light Must be bathroom light Okay! (Turn off the bathroom light) Learn from actual usage Where is the microwave installed? Jan 12, 2025 Inamura-Ishida Lab, Future University Hakodate -
Goal 8 Network config Grouping Room-by- room grouping WPS Localize
Ask user the location when used Coordinate Turn of the light Must be bathroom light Okay! (Turn off the bathroom light) Learn from actual usage Where is the microwave installed? Jan 12, 2025 Inamura-Ishida Lab, Future University Hakodate -
Goal 9 Network config Grouping Room-by- room grouping WPS Localize
Ask user the location when used Coordinate Turn of the light Must be bathroom light Okay! (Turn off the bathroom light) Learn from actual usage Where is the microwave installed? Jan 12, 2025 Inamura-Ishida Lab, Future University Hakodate -
Goal 10 Network config Grouping Room-by- room grouping WPS Localize
Ask user the location when used Coordinate Turn of the light Must be bathroom light Okay! (Turn off the bathroom light) Learn from actual usage Where is the microwave installed? Jan 12, 2025 Inamura-Ishida Lab, Future University Hakodate -
Goal 11 Network config Grouping Room-by- room grouping WPS Localize
Ask user the location when used Coordinate Turn of the light Must be bathroom light Okay! (Turn off the bathroom light) Learn from actual usage Where is the microwave installed? Jan 12, 2025 Inamura-Ishida Lab, Future University Hakodate Today s topic -
Related Work n Proximity-based device grouping based on p Sound
[jin 23], magnetism [jin 16], lighting [haus 20] → Requires special infrastructure or specific sensors n Radio-based device grouping p Amigo [varshavsky 07], PSP [cui 19] → Devices in a short range can be grouped Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025 12 [jin 23] Key agreement on IoT devices with echo profiling, IEEE/ACM Trans. Netw. 31(4) [jin 16] MagPairing: pairing smartphones in close proximity using magnetometers, IEEE Trans. Inf. Forensics Secur. 11(6) [haus 20] DevLoc: seamless device association using light bulb networks for indoor IoT environments, IEEE/ACM IoTDI [varshavsky 07] Amigo: proximity-based authentication of mobile devices, UbiComp [cui 19] PSP: proximity-based secure paring of mobile devices using WiFi signals, Wireless Netw., 25(2) -
CSI-Based IoT Device Grouping Our previous work, presented in IEEE
Globecom 2022 13 Jan 12, 2025 Inamura-Ishida Lab, Future University Hakodate -
Our Previous Work: Overview n CSI-based IoT device grouping [ishida
22] p Utilize CSI changes caused by human to group IoT device in the same room p Unsupervised learning (no labels required) l Grouped IoT devices with max ARI = 1.00 → Combined with room- layout estimator [joya 21] , location setup can be automated Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025 14 [ishida 22] Room-by-room device grouping for put-and-play IoT system, IEEE GLOBECOM [joya 21] Design of room-layout estimator using smart speaker, EAI MobiQuitous v v A B C D F E G G F E D C B A -
Our Previous Work: System Inamura-Ishida Lab, Future University Hakodate Jan
12, 2025 15 Device Grouping WLAN AP Data Retriever Out Room-by-room grouping using unsupervised learning Collect CSI from multiple devices (for several days) -
Feature Extraction Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025
16 1) Windowing 0.2 0.9 … 0.1 0.8 … … … … … … 0.2 0.9 … 0.1 0.8 … … … … … … <latexit sha1_base64="vYcOXklh+gWBP/5fdMQ1Vk+XYCU=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ab4uMY8OIxgnlIsoTZyWwyZGZ2mZkVwpKv8CQoiFc/x5N/4yTZgyYWNBRV3XR3hQln2njet1PY2Nza3inuunv7B4dHpeOTto5TRWiLxDxW3RBrypmkLcMMp91EUSxCTjvh5Hbud56o0iyWD2aa0EDgkWQRI9hY6bGfaDbIav5sUCp7VW8BtE78nJQhR3NQ+uoPY5IKKg3hWOue7yUmyLAyjHA6cyv9VNMEkwke0Z6lEguqg2xx8QxVrDJEUaxsSYMWqvtrIsNC66kIbafAZqxXvbn4n9dLTXQTZEwmqaGSLBdFKUcmRvP30ZApSgyfWoKJYvZYRMZYYWJsSK5NwV/9eZ20a1X/qnp5Xy836nkeRTiDc7gAH66hAXfQhBYQEPAMr/DmaOfFeXc+lq0FJ585hT9wPn8A7eqQXQ==</latexit> 21 Subcarrier –28 std p2p iqr … Subcarrier 28 std p2p iqr <latexit sha1_base64="pgBfoqmbH3eroXVc3hDQ/7TuQ2c=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ar8XEMePEYwTwkWcLsZDYZMzO7zMwKYclXeBIUxKuf48m/cZLsQRMLGoqqbrq7woQzbTzv2ymsrW9sbhW33Z3dvf2D0uFRS8epIrRJYh6rTog15UzSpmGG006iKBYhp+1wfDPz209UaRbLezNJaCDwULKIEWys9NBLNOtn/HHaL5W9qjcHWiV+TsqQo9EvffUGMUkFlYZwrHXX9xITZFgZRjidupVeqmmCyRgPaddSiQXVQTa/eIoqVhmgKFa2pEFz1f01kWGh9USEtlNgM9LL3kz8z+umJroOMiaT1FBJFouilCMTo9n7aMAUJYZPLMFEMXssIiOsMDE2JNem4C//vEpa51X/snpxVyvXa3keRTiBUzgDH66gDrfQgCYQEPAMr/DmaOfFeXc+Fq0FJ585hj9wPn8AnWWQ0A==</latexit> lj … Time … Time Time Subcarrier –28 Subcarrier 28 … … <latexit sha1_base64="vYcOXklh+gWBP/5fdMQ1Vk+XYCU=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ab4uMY8OIxgnlIsoTZyWwyZGZ2mZkVwpKv8CQoiFc/x5N/4yTZgyYWNBRV3XR3hQln2njet1PY2Nza3inuunv7B4dHpeOTto5TRWiLxDxW3RBrypmkLcMMp91EUSxCTjvh5Hbud56o0iyWD2aa0EDgkWQRI9hY6bGfaDbIav5sUCp7VW8BtE78nJQhR3NQ+uoPY5IKKg3hWOue7yUmyLAyjHA6cyv9VNMEkwke0Z6lEguqg2xx8QxVrDJEUaxsSYMWqvtrIsNC66kIbafAZqxXvbn4n9dLTXQTZEwmqaGSLBdFKUcmRvP30ZApSgyfWoKJYvZYRMZYYWJsSK5NwV/9eZ20a1X/qnp5Xy836nkeRTiDc7gAH66hAXfQhBYQEPAMr/DmaOfFeXc+lq0FJ585hT9wPn8A7eqQXQ==</latexit> 21 <latexit sha1_base64="pgBfoqmbH3eroXVc3hDQ/7TuQ2c=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ar8XEMePEYwTwkWcLsZDYZMzO7zMwKYclXeBIUxKuf48m/cZLsQRMLGoqqbrq7woQzbTzv2ymsrW9sbhW33Z3dvf2D0uFRS8epIrRJYh6rTog15UzSpmGG006iKBYhp+1wfDPz209UaRbLezNJaCDwULKIEWys9NBLNOtn/HHaL5W9qjcHWiV+TsqQo9EvffUGMUkFlYZwrHXX9xITZFgZRjidupVeqmmCyRgPaddSiQXVQTa/eIoqVhmgKFa2pEFz1f01kWGh9USEtlNgM9LL3kz8z+umJroOMiaT1FBJFouilCMTo9n7aMAUJYZPLMFEMXssIiOsMDE2JNem4C//vEpa51X/snpxVyvXa3keRTiBUzgDH66gDrfQgCYQEPAMr/DmaOfFeXc+Fq0FJ585hj9wPn8AnWWQ0A==</latexit> lj … Feature vector 2) Extract features, deriving CSI feature matrix 3) Pick 𝑁!"# rows, align them in a single row as a feature vector Obtain feature vector for each device Std deviation Peak-to-peak Inter-quartile range -
Feature Extraction Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025
17 1) Windowing 0.2 0.9 … 0.1 0.8 … … … … … … 0.2 0.9 … 0.1 0.8 … … … … … … <latexit sha1_base64="vYcOXklh+gWBP/5fdMQ1Vk+XYCU=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ab4uMY8OIxgnlIsoTZyWwyZGZ2mZkVwpKv8CQoiFc/x5N/4yTZgyYWNBRV3XR3hQln2njet1PY2Nza3inuunv7B4dHpeOTto5TRWiLxDxW3RBrypmkLcMMp91EUSxCTjvh5Hbud56o0iyWD2aa0EDgkWQRI9hY6bGfaDbIav5sUCp7VW8BtE78nJQhR3NQ+uoPY5IKKg3hWOue7yUmyLAyjHA6cyv9VNMEkwke0Z6lEguqg2xx8QxVrDJEUaxsSYMWqvtrIsNC66kIbafAZqxXvbn4n9dLTXQTZEwmqaGSLBdFKUcmRvP30ZApSgyfWoKJYvZYRMZYYWJsSK5NwV/9eZ20a1X/qnp5Xy836nkeRTiDc7gAH66hAXfQhBYQEPAMr/DmaOfFeXc+lq0FJ585hT9wPn8A7eqQXQ==</latexit> 21 Subcarrier –28 std p2p iqr … Subcarrier 28 std p2p iqr <latexit sha1_base64="pgBfoqmbH3eroXVc3hDQ/7TuQ2c=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ar8XEMePEYwTwkWcLsZDYZMzO7zMwKYclXeBIUxKuf48m/cZLsQRMLGoqqbrq7woQzbTzv2ymsrW9sbhW33Z3dvf2D0uFRS8epIrRJYh6rTog15UzSpmGG006iKBYhp+1wfDPz209UaRbLezNJaCDwULKIEWys9NBLNOtn/HHaL5W9qjcHWiV+TsqQo9EvffUGMUkFlYZwrHXX9xITZFgZRjidupVeqmmCyRgPaddSiQXVQTa/eIoqVhmgKFa2pEFz1f01kWGh9USEtlNgM9LL3kz8z+umJroOMiaT1FBJFouilCMTo9n7aMAUJYZPLMFEMXssIiOsMDE2JNem4C//vEpa51X/snpxVyvXa3keRTiBUzgDH66gDrfQgCYQEPAMr/DmaOfFeXc+Fq0FJ585hj9wPn8AnWWQ0A==</latexit> lj … Time … Time Time Subcarrier –28 Subcarrier 28 … … <latexit sha1_base64="vYcOXklh+gWBP/5fdMQ1Vk+XYCU=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ab4uMY8OIxgnlIsoTZyWwyZGZ2mZkVwpKv8CQoiFc/x5N/4yTZgyYWNBRV3XR3hQln2njet1PY2Nza3inuunv7B4dHpeOTto5TRWiLxDxW3RBrypmkLcMMp91EUSxCTjvh5Hbud56o0iyWD2aa0EDgkWQRI9hY6bGfaDbIav5sUCp7VW8BtE78nJQhR3NQ+uoPY5IKKg3hWOue7yUmyLAyjHA6cyv9VNMEkwke0Z6lEguqg2xx8QxVrDJEUaxsSYMWqvtrIsNC66kIbafAZqxXvbn4n9dLTXQTZEwmqaGSLBdFKUcmRvP30ZApSgyfWoKJYvZYRMZYYWJsSK5NwV/9eZ20a1X/qnp5Xy836nkeRTiDc7gAH66hAXfQhBYQEPAMr/DmaOfFeXc+lq0FJ585hT9wPn8A7eqQXQ==</latexit> 21 <latexit sha1_base64="pgBfoqmbH3eroXVc3hDQ/7TuQ2c=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ar8XEMePEYwTwkWcLsZDYZMzO7zMwKYclXeBIUxKuf48m/cZLsQRMLGoqqbrq7woQzbTzv2ymsrW9sbhW33Z3dvf2D0uFRS8epIrRJYh6rTog15UzSpmGG006iKBYhp+1wfDPz209UaRbLezNJaCDwULKIEWys9NBLNOtn/HHaL5W9qjcHWiV+TsqQo9EvffUGMUkFlYZwrHXX9xITZFgZRjidupVeqmmCyRgPaddSiQXVQTa/eIoqVhmgKFa2pEFz1f01kWGh9USEtlNgM9LL3kz8z+umJroOMiaT1FBJFouilCMTo9n7aMAUJYZPLMFEMXssIiOsMDE2JNem4C//vEpa51X/snpxVyvXa3keRTiBUzgDH66gDrfQgCYQEPAMr/DmaOfFeXc+Fq0FJ585hj9wPn8AnWWQ0A==</latexit> lj … Feature vector 2) Extract features, deriving CSI feature matrix 3) Pick 𝑁!"# rows, align them in a single row as a feature vector Obtain feature vector for each device Std deviation Peak-to-peak Inter-quartile range Group IoT devices by clustering based on feature vectors We use k-means in our paper Any clustering algorithm can be used -
Problem in Practical Env n Performance degradation with uneven human
distribution p Without human, our method cannot group IoT devices →We need to sample collect CSI with various people distribution Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025 18 Human location ARI 0.42 Dining room 0.27 Living room 1.00 Bedroom 0.29 Anywhere 0.95 Anywhere or no where 0.95 -
CSI Sampling in Feature Extraction Inamura-Ishida Lab, Future University Hakodate
Jan 12, 2025 19 1) Windowing 0.2 0.9 … 0.1 0.8 … … … … … … 0.2 0.9 … 0.1 0.8 … … … … … … <latexit sha1_base64="vYcOXklh+gWBP/5fdMQ1Vk+XYCU=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ab4uMY8OIxgnlIsoTZyWwyZGZ2mZkVwpKv8CQoiFc/x5N/4yTZgyYWNBRV3XR3hQln2njet1PY2Nza3inuunv7B4dHpeOTto5TRWiLxDxW3RBrypmkLcMMp91EUSxCTjvh5Hbud56o0iyWD2aa0EDgkWQRI9hY6bGfaDbIav5sUCp7VW8BtE78nJQhR3NQ+uoPY5IKKg3hWOue7yUmyLAyjHA6cyv9VNMEkwke0Z6lEguqg2xx8QxVrDJEUaxsSYMWqvtrIsNC66kIbafAZqxXvbn4n9dLTXQTZEwmqaGSLBdFKUcmRvP30ZApSgyfWoKJYvZYRMZYYWJsSK5NwV/9eZ20a1X/qnp5Xy836nkeRTiDc7gAH66hAXfQhBYQEPAMr/DmaOfFeXc+lq0FJ585hT9wPn8A7eqQXQ==</latexit> 21 Subcarrier –28 std p2p iqr … Subcarrier 28 std p2p iqr <latexit sha1_base64="pgBfoqmbH3eroXVc3hDQ/7TuQ2c=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ar8XEMePEYwTwkWcLsZDYZMzO7zMwKYclXeBIUxKuf48m/cZLsQRMLGoqqbrq7woQzbTzv2ymsrW9sbhW33Z3dvf2D0uFRS8epIrRJYh6rTog15UzSpmGG006iKBYhp+1wfDPz209UaRbLezNJaCDwULKIEWys9NBLNOtn/HHaL5W9qjcHWiV+TsqQo9EvffUGMUkFlYZwrHXX9xITZFgZRjidupVeqmmCyRgPaddSiQXVQTa/eIoqVhmgKFa2pEFz1f01kWGh9USEtlNgM9LL3kz8z+umJroOMiaT1FBJFouilCMTo9n7aMAUJYZPLMFEMXssIiOsMDE2JNem4C//vEpa51X/snpxVyvXa3keRTiBUzgDH66gDrfQgCYQEPAMr/DmaOfFeXc+Fq0FJ585hj9wPn8AnWWQ0A==</latexit> lj … Time … Time Time Subcarrier –28 Subcarrier 28 … … <latexit sha1_base64="vYcOXklh+gWBP/5fdMQ1Vk+XYCU=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ab4uMY8OIxgnlIsoTZyWwyZGZ2mZkVwpKv8CQoiFc/x5N/4yTZgyYWNBRV3XR3hQln2njet1PY2Nza3inuunv7B4dHpeOTto5TRWiLxDxW3RBrypmkLcMMp91EUSxCTjvh5Hbud56o0iyWD2aa0EDgkWQRI9hY6bGfaDbIav5sUCp7VW8BtE78nJQhR3NQ+uoPY5IKKg3hWOue7yUmyLAyjHA6cyv9VNMEkwke0Z6lEguqg2xx8QxVrDJEUaxsSYMWqvtrIsNC66kIbafAZqxXvbn4n9dLTXQTZEwmqaGSLBdFKUcmRvP30ZApSgyfWoKJYvZYRMZYYWJsSK5NwV/9eZ20a1X/qnp5Xy836nkeRTiDc7gAH66hAXfQhBYQEPAMr/DmaOfFeXc+lq0FJ585hT9wPn8A7eqQXQ==</latexit> 21 <latexit sha1_base64="pgBfoqmbH3eroXVc3hDQ/7TuQ2c=">AAAB8XicbVDLSgNBEOyNr7i+oh69DIaAp7Ar8XEMePEYwTwkWcLsZDYZMzO7zMwKYclXeBIUxKuf48m/cZLsQRMLGoqqbrq7woQzbTzv2ymsrW9sbhW33Z3dvf2D0uFRS8epIrRJYh6rTog15UzSpmGG006iKBYhp+1wfDPz209UaRbLezNJaCDwULKIEWys9NBLNOtn/HHaL5W9qjcHWiV+TsqQo9EvffUGMUkFlYZwrHXX9xITZFgZRjidupVeqmmCyRgPaddSiQXVQTa/eIoqVhmgKFa2pEFz1f01kWGh9USEtlNgM9LL3kz8z+umJroOMiaT1FBJFouilCMTo9n7aMAUJYZPLMFEMXssIiOsMDE2JNem4C//vEpa51X/snpxVyvXa3keRTiBUzgDH66gDrfQgCYQEPAMr/DmaOfFeXc+Fq0FJ585hj9wPn8AnWWQ0A==</latexit> lj … Feature vector 2) Extract features, deriving CSI feature matrix 3) Pick 𝑁!"# rows, align them in a single row as a feature vector Obtain feature vector for each device Std deviation Peak-to-peak Inter-quartile range Which row to pick? = CSI sampling -
CSI Sampling for Practical Environments Proposed Method 20 Jan 12,
2025 Inamura-Ishida Lab, Future University Hakodate -
Key Idea n Sample CSI when people are distributed in
many places p CSI changes affected by number and location of people n In practical environments... p Number and location of people are unknown → Assume CSI changes in different situations are independent → Perform ICA to extract CSI changes caused by different distribution of people Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025 21 … … -
Overview of CSI Sampling Inamura-Ishida Lab, Future University Hakodate Jan
12, 2025 22 ... Time Device 1 Device 2 ... Device n ICA ICA i ICA j Device 1 Device 2 Device n ... ... ... ... 1) Perform ICA on CSI feature matrices of all devices 2) Clustering in ICA space 3) Sample CSI feature in each cluster CSI feature matrix -
Experiment Setup n 1BD condominium p 1 WLAN AP: Buffalo
WXR-5700AX7S p 9 IoT devices: Raspberry Pi 3A+ (at height of 0 2m) p CSI collector: Intel Compute Stick computer p 1 in 40s, 1 in 30s, 2 under 10 were living in this condo n Collect CSI for 24h p Window length: 60s p 500 grouping trials l Derive mean ARI Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025 24 Bedroom Storeroom CL BR WC CL Living Dining Kitchen AP Data Retriever IoT Devices 2 1,3 4,6 5 7 8 9 5.3m 3.5m 2.5m 4.1m Dev ID # of packets # of valid windows 1 290663 2358 2 2968 0 3 198929 2357 4 268754 2880 5 308506 2880 6 347484 2880 7 369627 2880 8 247124 2880 -
Comparison 1. ICA sampling (proposed) 2. PCA sampling p PCA
instead of ICA on CSI feature matrices 3. Random sampling (our previous method) p Randomly sample rows of CSI feature matrix n # of ICA/PCA components 𝑁!"# = 12 p Contribution of pricipal components > 0.8 n # of clusters in ICA/PCA space 𝑁$%&' = 6 p # of selected windows 𝑁!"# = 𝑁$%&' ×𝑁()*$ Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025 25 -
IoT Device Grouping Result n ARI (as a function of
𝑁!"#$ ) p ICA (proposed): ARI=0.943 @𝑁!"#$ = 1 l ARI=0.991 @𝑁"#$% = 2 l ARI=1.00 @𝑁"#$% ≥ 4 p PCA also showed relatively high ARI p Random showed high ARI when 𝑁%&'( is big Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025 26 0 2 4 6 8 10 Nsamp 0.7 0.8 0.9 1.0 Mean ARI ICA PCA Random 44.9% up -
Summary n IoT device grouping for Put-n-Play IoT p Grouping
via CSI-based unsupervised learning p Performance degraded in practical environment n CSI sampling for practical environment p ICA on CSI to separate CSI changes caused by different people p Clustering in ICA space, sample in each cluster p Experimental evaluation demonstrated l Mean ARI = 0.944 (improved by 44.9%) Inamura-Ishida Lab, Future University Hakodate Jan 12, 2025 28