AI Research at Qualcomm Technologies is advancing wireless communications and RF location sensing.
Wireless communications and RF sensing technologies continue to advance with 5G rolling out now and 5G Advanced in the wings. In order to sustain a high quality-of-service and deliver a superior user experience, creators of wireless communication devices must anticipate a rapidly changing physical landscape around the transmitter and the receiver while also pinpointing a mobile user’s location outside the reach of satellite services. A 5G modem in a handset must be able to function in wireless communication channels with obstacles and reflections from objects like floors and walls. Failure to effectively navigate this complex world could result in dropped calls or degraded signal quality.
As a pioneer in advanced wireless technologies and a leader in low-power Artificial Intelligence (AI), Qualcomm has been researching the intersection of these technologies. As an example, the research team at this year’s Mobile World Congress in Barcelona showed off an end-to-end over-the-air (OTA) testbed for an AI-enabled air interface. Machine Learning (ML) and wireless communications together make quite a powerful technology since they have complementary strengths. ML can take wireless communications to the next level, providing not only fast and flexible models and algorithms for communication in a dynamic environment but also accurate RF-based sensing of the environment and a mobile user’s location, even in a complex indoor environment such as an office building. Let’s look at what Qualcomm is accomplishing by melding these two technologies (check out their webinar for even more details).
Using AI to Enhance Wireless Communications
Qualcomm has already introduced a 5G modem that includes an AI engine. Qualcomm Technologies Inc.
First, we should be clear that AI is already impacting wireless communications in the Snapdragon X70 Modem-RF System, which will ship later this year. The X70, announced in February 2022, is the first standalone modem to include an on-board AI engine. (The X70 is expected to be used by high-end handset manufacturers.) However, this modem is only the beginning of this journey; Qualcomm Technologies is currently exploring over a dozen research areas for 5G enhancements with AI including channel state feedbacks and mmWave beam management. AI can potentially impact many other research areas in connection to 5G including power savings, security, contextual awareness, and positioning.
One of the key use cases for applying ML to wireless is using generative modeling to provide a more accurate channel representation and thereby design better algorithms and improve communications. Wireless signals, of course, are rarely transmitted along a straight line-of-sight, but can be blocked or reflected off of multiple surfaces, such as walls, floor, ceiling, and objects. Classical channel models that deal with these factors do work, but require cumbersome field measurements, have hard-coded model assumptions, and are slow to prototype.
Qualcomm is researching how AI can learn channel distribution with a GAN. QTI
Qualcomm AI Research has shown that neural channel models can accurately match complex data distributions with fast prototyping. Using a Generative Adversarial Network, or GAN, Qualcomm AI Research has designed a neural channel model that learns to generate multiple-antenna (MIMO) channel. The generator learns the channel distribution jointly with a discriminator that teaches the generator to capture the most relevant wireless features in the model. The neural channel model is interpretable and can be used for modeling channels with different configurations. More accurate channel models are essential for better communication design and crucial for evaluation of gains brought by AI compared to classical methods.
Another example of applying AI to wireless is to improve communication design. Qualcomm AI Research is applying neural augmentation to enhance Kalman filters, providing a more accurate channel acquisition and improving signal quality. Neural augmentation is a high-level design principle that suggests classical algorithms designed based on domain knowledge can be enhanced by integrating machine learning algorithms to tune and adapt algorithm parameters. Communication channels are hard to accurately estimate as they vary in time with unknown dynamics, and the observations of the pilot signal are noisy. A more accurate channel estimate at all the time steps for different dynamics enables more efficient communications. Classical Kalman filters lose accuracy over different dynamics and require adapting their parameters to channel dynamics. A neural augmentation of Kalman filter adapts Kalman parameters based on the channel dynamics on the fly. It shows superior results over both classical Kalman and standalone LSTM ML approaches. By keeping the Kalman filter as the backbone of the model, neural augmentation of Kalman filters provides robust generalization to unseen scenarios better than a single Kalman while keeping interpretability of the model. The performance improvement with respect to standalone Kalman filters comes from expressive power of neural networks that enables adapting Kalman parameters to more complicated dynamics, for example for high Doppler.
Qualcomm is augmenting classical Kalman filtering to improve communications in dynamic environments. QTI
Using AI to Enable Accurate Location Sensing with RF
Location-based services can be challenging when global navigation satellite system (GNSS) signals are weak or not available. Consequently, positioning through AI-augmented RF sensing can be useful indoors and in locations without a clear line-of-sight to GNSS for use cases, such as indoor navigation, vehicular navigation, AGV tracking, and asset tracking.
Two types of RF sensing based positioning can be distinguished. In the active case, a mobile device is actively trying to position itself by emitting RF signals and measuring the channel. In the passive case, RF signals are used as a radar for sensing objects or persons that are not emitting RF signals themselves.
Current methods for active positioning have limitations. The classical method of time difference of arrival (TDOA) does not require labeled data, but it suffers from accuracy issues under non-line-of-sight and potentially does not use valuable multi-path information. The machine learning assisted RF fingerprinting is more accurate but requires collection of densely sampled positioning labels at deployment. Data collection might have to be repeated if the channel changes significantly.
The physical environment can present numerous challenges for the indoor positioning of a mobile device. QTI
Qualcomm AI Research has developed an ML-based approach for active positioning called Neural RF SLAM that achieves the best of the two worlds. Neural RF SLAM trains a neural network to predict the 3D device location from channel state information. The training objective is based on being able to reconstruct the channel state measurement from the predicted 3D location. The reconstruction of the channel state is done by a tunable RF propagation model. The parameters of the propagation model are set alongside the neural network weights during training. Multipath information is used both during mapping and reconstruction. Neural RF SLAM achieves 43.4 cm accuracy 90% of the time using only a single access point with a single antenna as an anchor.
Similarly, passive positioning via RF sensing, which allows positioning people with access points alone, may soon be possible at scale. Passive positioning and sensing can be used for a variety of use cases, like detecting presence, counting people and monitoring sleep. Qualcomm AI Research’s method, WiCluster, is a weakly-supervised ML technique that works well for positioning people in non-line-of-sight environments, such as across building floors. It is weakly-supervised in the sense that only a few room-level labels and a floor plan are required. WiCluster discovers the 3D manifold that represents the subject motion in the channel state and maps it to the floor plan. In extensive testing, WiCluster has demonstrated precise subject positioning with errors in the 1-2m range, results generally comparable to supervised training models that require expensive data labeling.
Conclusions
While applying AI to wireless still has challenges that are being addressed, such as generalization outside of the modeled domain, adaptability, and the impracticalities of supervised learning, neural-augmented wireless communications and RF-sensing show tremendous promise in improving the customer experiences. Combined with Qualcomm’s “AI Firsts” announcements, these recent disclosures about the combination of AI and 5G to improve communications demonstrate Qualcomm Technologies’ unique advantage in modems, application processors, and AI research.