Latest / EDGE AI POD / Revolutionizing Wi-Fi Sensing with Machine Learning and Advanced Radio Frequency Techniques

Transcript

Revolutionizing Wi-Fi Sensing with Machine Learning and Advanced Radio Frequency Techniques · EDGE AI POD

• 0:04
  Speaker 1: Like a very interesting talk about the integration or kind of the mashup between Wi-Fi sensing and ML, and it's really, really interesting.
• 0:12
  Speaker 1: So we're gonna have Joseph Chui from National Tsinghua University come up and give us a talk.
• 0:22
  Speaker 2: Hi everybody.
• 0:22
  Speaker 2: Obviously, I'm not this person.
• 0:24
  Speaker 2: I'm replacing him.
• 0:25
  Speaker 2: So my name is Joseph Chu.
• 0:27
  Speaker 2: I'm replacing my co-worker, professor Tsai from the University of Tsinghua here in Taiwan.
• 0:36
  Speaker 2: The topic today is actually more about the radio frequency sensing, which will be the next generation CG and also Wi-Fi standard feature.
• 0:48
  Speaker 2: That is different from this morning's talk about more about the vision or the camera, but this is about radio frequency.
• 0:55
  Speaker 2: Okay, so my background I educated in Sydney University, australia, and for the last 25 years I've been in research and also business development in semiconductors.
• 1:13
  Speaker 2: So we recently, our company actually started for about three years.
• 1:20
  Speaker 2: We've recently been awarded for the Taiwan Semiconductor Grand Challenge.
• 1:26
  Speaker 2: Because of the AI topic, okay, talking about today.
• 1:30
  Speaker 2: So we'll go through this radio frequency sensing based on Wi-Fi.
• 1:36
  Speaker 2: Wi-fi every year accounts for about 6 billion units a year for the total addressable market every year Increasingly.
• 1:46
  Speaker 2: The AI was expecting the volume was like increased a double or triple.
• 1:51
  Speaker 2: So this will be very significant application user case for the tiny machine learning.
• 2:00
  Speaker 2: Okay, so Wi-Fi sensing is actually a new technology that is using existing radio frequencies based on, for example, 2.4g, 5g, even up to 60 GHz for the radio frequency and sensing.
• 2:19
  Speaker 2: So what it can be used is actually like human activity recognition, intruder detection, localization, that is, be able to be very accurate.
• 2:31
  Speaker 2: For example, in the past, the Wi-Fi we know that for the 30 years, past 30 years is based on signal processing from communication, so it cannot be very accurate.
• 2:41
  Speaker 2: For example, we see that localization is probably to several meters away.
• 2:46
  Speaker 2: But with machine learning we are able to be close, narrow to 30 centimeters, which is usable for robotics, for ID and also for, like external orbit drone.
• 3:04
  Speaker 2: Ok, and this research is based on the feature injection from the China State Information, which is called CSI, to collect the data and make some judgment.
• 3:18
  Speaker 2: And this is rather new because nowadays, if you see the chip supplier, like Mediatek, like Procom, like Realtek, this kind of feature is there but it's not optimized.
• 3:32
  Speaker 2: So it's very difficult to apply the AI, especially the machine learning, into the hardware.
• 3:40
  Speaker 2: So we try to resolve this kind of a problem by providing the genetic platform so people can apply the algorithm into it.
• 3:50
  Speaker 2: Okay, consider the Wi-Fi in the indoor.
• 3:57
  Speaker 2: Nowadays, if you go to US you will see that already deployment with AT&T and Verizon, for example, they already use a so-called Wi-Fi sensing.
• 4:07
  Speaker 2: But there's some problem of the cost because it requires at least, like three pairs, three devices of AP routers.
• 4:19
  Speaker 2: So that's the main cost because hardware is super expensive on that part.
• 4:23
  Speaker 2: So we are trying to resolve the problem by reducing the configuration for the hardware configuration so in the future everybody can put in their Orkut and develop the Orkut into the environment.
• 4:39
  Speaker 2: So CSI information for those of you the first time to learn about this is the channel state information is actually the information we use in the channel.
• 4:52
  Speaker 2: So in the channel we mainly will have several orders of magnitude so you can extract the information from the dataset and from the dataset you can extract the information from the data set and from the data set you can do the processing.
• 5:04
  Speaker 2: So we will go through this.
• 5:06
  Speaker 2: Number one will be the magnitude Magnitude is the one you can easily attract and the second one will be the phase.
• 5:15
  Speaker 2: However, phase is very important here because, just like several antenna, you see that there must be a phase difference.
• 5:22
  Speaker 2: So in order to do the algorithm and make correct decision, we need to have this phase information to be as accurate as possible, and that's basically the limitation.
• 5:34
  Speaker 2: So from this research we would actually recommend to have more degrees of freedoms, more degree of freedom in the channel and the receiver side, you will be able to make decision.
• 5:47
  Speaker 2: Okay, consider the issue.
• 5:49
  Speaker 2: So we improve the quality of the feature from the Wi-Fi sensing application and to have more of a propagation environment exam.
• 6:00
  Speaker 2: And also the transmission bandwidth, how the bandwidth is going to affect on the sample data you are going to take.
• 6:06
  Speaker 2: Okay, and the transmission model.
• 6:10
  Speaker 2: So for those of you, this is a two major transmission model in communication we call OFDM, ofdma.
• 6:19
  Speaker 2: Ofdma is applying Wi-Fi 6 and Wi-Fi 7, but OFDM was before Wi-Fi 6, wi-fi 4, and Wi-Fi 11ab.
• 6:28
  Speaker 2: And we also need to consider the CSI data aggregation issue and the interference between other systems, because that would be the major part.
• 6:39
  Speaker 2: If you apply to your algorithm into this kind of detection, you have to consider all the interference from the air, because this is not like the vision.
• 6:47
  Speaker 2: You can easily use camera to manipulate the camera image, but the interference from the air you cannot see.
• 6:53
  Speaker 2: So you have to take this out into consideration.
• 6:56
  Speaker 2: So, number one, inference of propagation environment.
• 7:02
  Speaker 2: Okay, the main takeaway is this one because in the channel you actually have multi-paths, so you have all the reflections of the signal and this will induce the time dispersion into the received signal.
• 7:16
  Speaker 2: So we need to consider the delay and into the environment size.
• 7:21
  Speaker 2: So in this kind of a suggestion we recommend three sub-cases Small, medium size of the room and large size of room.
• 7:34
  Speaker 2: And we will exact through this simulation Basically we will go through like a bandwidth and how the bandwidth is going to affect your data set.
• 7:43
  Speaker 2: So this is one small-sized room based on 20 megahertz of the bandwidth.
• 7:50
  Speaker 2: So if you see from these two main things CSI magnitude and CSI phase the magnitude is almost flat but the phase difference will be something that you can only pick up from this data set.
• 8:04
  Speaker 2: So this is another simulation that the CSI phase is.
• 8:12
  Speaker 2: You see that this is generally non-linear for the multi-path channel and for the larger room because you have more freedom of publication.
• 8:26
  Speaker 2: So you will see both of the CSI magnitude and the CSI phase as informative.
• 8:32
  Speaker 2: So basically, nonlinear CSI phase is more informative for the decision making.
• 8:38
  Speaker 2: And we also talk about the influence on the transmission bandwidth.
• 8:45
  Speaker 2: So basically, if you have a bigger or wider transmission bandwidth you can actually improve the frequency selected.
• 8:55
  Speaker 2: So that means you have more data you can pick up from.
• 9:03
  Speaker 2: So in the following we'll go through the simulation quickly, run through it.
• 9:06
  Speaker 2: So this is for 40 to 80 megahertz, which is a larger bandwidth.
• 9:12
  Speaker 2: So you will see the magnitude and phase difference.
• 9:14
  Speaker 2: Okay, this is for medium-sized room and large size of room.
• 9:24
  Speaker 2: Okay, so the inference on the transmission mode.
• 9:31
  Speaker 2: This is one key takeaway as well.
• 9:34
  Speaker 2: In Wi-Fi 4, you only have OFDM.
• 9:37
  Speaker 2: In Wi-Fi 6 or above you have more advance of core OFDM, including the subcarrier.
• 9:44
  Speaker 2: More subcarrier means more sample space, means more degree of freedom to make decision.
• 9:51
  Speaker 2: So we'll see that actually through the simulation, the OFDMA with more room of a subcarrier is actually better in performance.
• 10:01
  Speaker 2: So this is quickly example of what we do, compared of DMA.
• 10:07
  Speaker 2: So another thing is aggregation.
• 10:17
  Speaker 2: Nowadays, if you use a simply take away from core comp or broad comp Mediatab chipset, you will have some difficulty to take out the example of the data set because all the data will actually collect.
• 10:34
  Speaker 2: So you need to do the face coordination.
• 10:37
  Speaker 2: This is super important in order to have a clear data at the receiver side.
• 10:43
  Speaker 2: So in the following we'll do the field measurement OK, and we see a person in the room and without person in the room, what will be the sample?
• 10:55
  Speaker 2: Ok, so almost flat response for the magnitude and almost the same.
• 11:01
  Speaker 2: So the only thing you can see here is only the face different in these two and the most important will be the other radio signal interference.
• 11:14
  Speaker 2: So in this case we take Bluetooth.
• 11:16
  Speaker 2: So if you take Bluetooth it's sharing the same frequency response from Wi-Fi.
• 11:22
  Speaker 2: You see a lot of interference as a circle one here.
• 11:30
  Speaker 2: So how do we resolve this kind of problem?
• 11:32
  Speaker 2: We actually need to induce more degrees of freedom into the sample space, which means the radio frequency channel information.
• 11:44
  Speaker 2: So either more channel or more face coordination.
• 11:49
  Speaker 2: Actually, if you can have a hardware solution which introduces coding at the transmitter side which you decrease, actually increases the degree of freedom by one order, for example, in the Wi-Fi we will increase from degree of freedom for two into three, so you can actually have more order in processing in the decision side.
• 12:13
  Speaker 2: Okay, so the conclusion is that based on the CSI information, we can actually do this kind of sensing and AI decision-making using the radio frequency system.
• 12:33
  Speaker 2: For medium-large room size you actually have more degree of freedom because you have more sample space, but for the small room you actually need to have more processing power at the receiver side to make it more compatible.
• 12:50
  Speaker 2: Ok, so basically that's the conclusion of the new radio frequency AI, that's in application of Wi-Fi, which will be the same for 6G as well.
• 13:04
  Speaker 2: Any questions?
• 13:10
  Speaker 1: Hey, thank you so much Appreciate it.
• 13:13
  Speaker 1: Thank you All right, fantastic.