How to speed up indoor Radio Dot design with AI
Indoor Radio Dot design and deployment is typically time-consuming and laborious process. Now, a team of scientists from Ericsson’s Global Artificial Intellgence Accelerator (GAIA) are proposing a new method of radio design harnessing the power of artificial intelligence and machine learning. Discover more here.
Indoor wireless connectivity is becoming more essential. Recent studies on cellular network usage shows that indoor traffic takes 87 percent of usage time in the US and generates 70 percent of global mobile data traffic. Since 2014, Ericsson has offered an indoor solution called the Ericsson Radio Dot System (RDS). RDS is easy to install and has superior coverage, designed for large indoor areas such as offices, stadiums, shopping malls, and universities. It is pretty similar to how we deploy our outdoor macro-cell system, and has the flexibility to reconfigure the settings to satisfy different network coverage requirements.
Deployment of RDS inside a new building requires a site survey and initial dot layout design that is often pretty time consuming and costly process. A site survey is required to figure out the optimal number and layout of radio dots for a floor for optimal deployment and performance.
A site survey consists of the several steps: firstly, we need to identify the wall type (concrete, glass, metal, etc.) of all wall segments. Then we need to measure macro-cell interference (the signal transmitted from nearby outdoor macro base stations) in order to calculate their interference on indoor locations. Once we have all surveys completed, an experienced radio designer determines the best radio dot locations by designing, evaluating and reiterating the signal propagation heatmap generated by a radio frequency (RF) planner software, in order to maximize the coverage efficiency for each floor. The figure below shows a typical design process from a given raw floor plan to its RDS deployment. Once a building landlord provides a raw floor plan of a target building, the engineers are sent to the building for a site survey to mark wall types with different colors. Then, RF designers place radio dots manually while checking its simulated heatmap of reference signal received power (RSRP) level, which is a type of received power measurement from the radio dots. Note that in the heatmap, darker pixel indicates stronger RSRP. The whole design process is not only costly but tedious, taking up to tens of days to be completed for one large building. Due to those factors, the current RDS deployment methodology is not scalable and it limits indoor RDS business.
DA-cGAN: Generative Adversarial Networks for Radio Design
To address inefficiencies in current RDS deployment process, we proposed a solution based on machine learning which formulates the design process as an image-to-image translation problem using generative adversarial networks (GANs). Specifically, we propose a dimension-aware conditional GAN (DA-cGAN) to generate a heatmap of optimal dot layout from a given floor plan. GANs are well-known for their capability of generating artificial images. GANs consist of two deep neural networks: for example, a generator and a discriminator . Two of them compete and learn from each other, and in the end the generator can generate a fake image which cannot be distinguished from the true image by the discriminator, and hopefully also by the human eyes. Some interesting applications using GANs include colorizing cartoon characters from their sketch images, creating super-resolution images, transferring artist style to photos, repairing images by filling their missing parts, and even generating three-dimensional models given two-dimensional images of objects from multiple perspectives.
Among the various applications, colorization is chosen as the starting point since it is most relevant to our problem. Colorization of a given sketch of floor plan not only needs to preserve its border shape, but also needs to learn from its internal structure to generate the desirable signal heatmap since the structure of a floor plan highly correlates with its optimal dot layout and its resulting heatmap. For example, a concrete wall has much higher signal attenuation than a dry wall, and therefore, the radio signal strength decades quickly around a concrete wall. To solve our image translation problem, we adopted the conditional GAN (cGAN), which learns the mapping from an input image and a random noise to the corresponding output image. However, we drop the random noise in our architecture following the convention in cGANs, since we only focus on generating one heatmap with optimal radio dot layout.
The below figures are our proposed GANs consisting of a generator and a discriminator . In order to preserve most of the floor plan structure, we adopt the U-Net architecture as our generator. U-Net is a special encoder-decoder network such that it concatenates each layer in the encoder to the symmetric layer in the decoder. A unique modification in the proposed generator is that it considers physical dimension vector as an additional input. Physical dimension is usually not considered in the previous GAN-related works which assume input data as pixels without physical dimension. We found that assumption is not enough for modeling radio signal propagation because, in a traditional physics-based simulator, a heatmap is derived based on its path loss which is dominated by two factors: 1. the distance between the radio dot and the current location, and 2. the environmental parameter known as path loss exponent. The path loss increases logarithmically as the physical distance increases from a radio dot, while the path loss exponent is a fixed value measured based on different scenarios. Another important feature of this generator is the macro-cell interference . We propose to incorporate the measured RSRP heatmap from outdoor macro-cell into our model, so that it is guided to generate more radio dots to cover the highly interfered regions.
Radio Dot Layout by AI Designer
The below figure shows three qualitative results with the test set. The first row shows a regular-size 8,440 sqft. floor plan, while the last two rows are wide-sized: 63,056 and 85,998 sqft. respectively. For each row, the first column is the input floor plan image, the second column is the ground truth heatmap from human designers with dots in green, and the last column is the predicted heatmap from DA-cGAN with dots in red. We observed that DA-cGAN tends to produce dot layout and heatmap almost like human designers if a given floor plan has more open spaces and simple wall types, while complicated, blurred or noisy floor plans with unseen or unusual tags cause its predicted heatmap to diverge from the ground truth. For a very large floor plan, we segment such a wide floor plan into multiple parts with reasonable size (due to the physical limitation of GPU VRAM) and merge predictions. With this case, DA-cGAN tends to predict more dots than human designs, probably because a wide floor plan does not necessarily have large number of dots. However, we believe that such discrepancy in layouts between human designers and our model does not necessarily mean that the model predicted performance is inferior to the ground truth, because our model has the potential of exploring the search space that the human designers have not found.
Conclusively, the proposed method is expected to reduce expensive design time of radio experts, the biggest roadblock in the way toward more initial budget calls, from days to minutes and thereby allow indoor business to reach out to far more buildings.
For more technical details of this work, please refer to our paper. Special thanks to the team members of GAIA working on this project: Taesuh Park, Chun-Hao Liu, and Hun Chang.