DeepSig

DeepSig Inc. is a venture-backed and product-centric technology company developing wireless processing software solutions using machine learning techniques to transform baseband processing, wireless sensing, and other key wireless applications. The company uses deep learning technologies to develop signal processing and radio systems by creating and employing new methodologies and software systems for the design and optimization of wireless communications.

The company claims that insufficient wireless models and algorithms are being used in wireless systems, and that more data-driven machine learning approaches can result in improvements to system performance, capacity, density, resilience, and efficiency of wireless systems.

Instead of relying on more traditional engineering approximations and simplifications, DeepSig's neural network-based signal processing solutions leverage complexity and data to improve key performance metrics, such as data throughput, multi-user capacity, latency minimization, and power consumption reduction.

According to DeepSig, this approach can increase the overall quality of service, reduce capital expenditures and operating expenses, and improve the performance of the radio access network across a wide range of components and effects.

The graph below illustrates DeepSig's position relative to other private companies.

The OmniSIG sensor is an AI-driven RF signal processing solution, which the company claims to have more capabilities than other spectrum monitoring solutions, due to being driven by machine learning. OmniSIG detects and classifies signals with low latency using data obtained directly from spectrum recordings.

OmniSIG Sensor can detect emitters across a wide range of bands and emitter types while on small or mobile platforms or while deployed on radio infrastructure devices. This enables coverage mapping, usage mapping, interference hunting, unauthorized emitter hunting, detection of cyber threats, and other mobile mapping applications.

DeepSig claims that the OmniSIG sensor offers an improvement of 4 to 10 dB higher sensitivity over other methods, and that in some instances, it can provide reductions of 10x or more in computational complexity and throughput. It achieves these reductions through algorithmic efficiency, increased parallelism, and reduced sample precision and dynamic range.

This sensor examines the spectrum environment to inform contextual analysis and decision making. According to DeepSig, OmniSIG provides higher sensitivity and accuracy and is more able to withstand impairments and dynamic spectrum environments than other solutions of this kind. In addition, it requires fewer computational resources and a lower dynamic range.

The OmniSIG software can be deployed and scaled on a wide range of target devices, including low-SWaP mobile and embedded systems, mobile personal computers, as well as cloud and datacenter environments. The product's web-based UI, open low-latency streaming interface, and control API aim to assist integration into customer systems and applications.

OmniSIG's detection and recognition capabilities have been validated across many signal types:

• Cellular and infrastructure signals such as GSM, LTE, CDMA2000, WCDMA, and WiMAX

• ISM-band signals such as WiFi and Bluetooth

• Mobile radio services such as P25, GMR, DMR

• IoT signals such as LoRa

• Commercial aircraft RADAR signals

The OmniSIG sensor publishes signal detection and parameter estimates in near real time to standard protocols using JSON (JavaScript Object Notation), which enables autonomous edge sensing and edge sensing with cloud fusion. The supported output methods include the following:

• SigMF-based files

• ZeroMQ sockets

• ElasticStack

• Web page

The OmniSIG SDK tool suite allows users to curate RF datasets, train deep learning inference models for custom wireless sensing applications, and deploy them to edge-sensing devices. OmniSIG SDK contains DeepSig’s baseline RF dataset for machine learning, including many consumer wireless signals.

This tool suite also allows customers to use their own custom data, signals, and signatures to train the AI sensor. DeepSig claims that it is the first such product on the market. The OmniSIG SDK contains tools for:

• sorting, labeling, and curating RF data;

• training the OmniSIG Sensor model with labeled data;

• evaluating its performance; and
• deploying the trained deep learning model into an OmniSIG runtime.

The OmniSIG SDK enables users to visualize signal captures and label them for use in AI systems using automated, semi-automated, and hand-tuning methods. The OmniSIG Sensor and OmniSIG SDK combine to enable a data engine that can improve signal detection and classification accuracy by identifying and capturing spurious data. The sensor’s ability to detect signals and signatures of interest is improved automatically through labeling the spurious data, retraining the network, and redeployment.

An OmniSIG model can be trained in a few hours using a single desktop-class GPU (Graphics Processing Unit), depending on the required level of performance. Models can be exported from a live SDK training process at any point for testing and validation.

Following training, the custom model is deployed with the OmniSIG Sensor for use in wireless sensing systems. The model's output is a metadata stream used to aid downstream systems and operators in the building of AI-based systems.

OmniPHY-5G is a communications system designed to significantly improve 5G RAN performance with machine learning by:

• reducing power consumption,

• exploiting channel information to support improved connections,

• exploiting Massive MIMO systems, and

• making better use of real-world feedback and responses to compensate for hardware, distortion, non-linearities, and effects.

DeepSig is developing these capabilities within its 5G-NR RAN L1 implementation. The purpose of 5G-NR RAN L1 is to provide case studies that allow VRAN and other 5G-NR base station integrators to understand and quantify the value of utilizing the data-centric RAN algorithms provided by OmniPHY-5G.

Time domain sampling simulations and over-the-air system tests have demonstrated that OmniPHY-5G's baseband processing performance offers improved bit error rate. This improvement provides 2x to 10x better signal reception and more efficient power-saving inference algorithms, which results in reduced computational costs and deployment of twice as many radio heads per baseband unit in a vRAN-centric front-haul architecture.

DeepSig first demonstrated this capability in an over-the-air system at the Brooklyn-5G Symposium in 2019. Using consumer-grade embedded and laptop NVIDIA GPUs, DeepSig demonstrated a 5G-NR downlink signal in which the receiver adapted to the environment. This adaptation continually improved the performance of channel estimation and equalization in the 900 MHz test.

DeepSig has demonstrated a machine learning-driven processing approach for Massive MIMO in simulation. The company also intends to build a test bed for real-world measurement of the data-centric method for Massive MIMO processing in UL-MIMO and DL-MIMO configurations in TDD (Time Division Duplex) and FDD (Frequency Division Duplex) systems.

DeepSig has developed custom C++ code for this application. The aim of this offering is to help save power and improve density and system performance in 5G RAN deployments.

OmniPHY is a communications system that learns from wireless channel measurements and is directly optimized for real-world hardware and channeling effects using end-to-end performance metrics and feedback.

By providing wireless systems that exploit imperfections and degrees of freedom, OmniPHY systems provide improved efficiency, resilience, and adaptivity in complex high-density, non-linear, hostile, or unique communications environments.

OmniPHY and machine learning approaches to physical layer modem optimization take several forms. They can be introduced in the form of small but key algorithms and subsystems inside complete communications suites, or they can be more extensive and radical in how they change the physical layer. In the case of certain systems where the ecosystem mainly needs to interoperate with itself, such as point-to-point backhaul, satellite communications, or single-vendor mesh networks, more radical changes can be implemented in the physical layer.

OmniPHY allows point-to-point and closed communications ecosystems to adapt completely across the modulation and coding dimensions of modems, while still providing standardized interfaces to networks and application layers. This allows modems to use the link and operate with existing best-practice methods such as FIPS 140-2 class AES256 link encryption, high-performance error correction using polar codes, message authentication, and error detection.

OmniPHY learned communications links focus on physical layer modulation and representation learning. The links adapt the representation of what is transmitted to optimize processing on both ends for key performance metrics such as bit error rate and energy efficiency. As a software modem capability, OmniPHY deploys on a range of off-the-shelf computer platforms and radio front-end devices, such as OmniSIG, to allow for the deployment and integration of communications links into unique and demanding applications.

OmniPHY can be applied in closed ecosystem wireless communications solutions where a high degree of adaptation is possible and learning-based physical layer techniques push algorithmic efficiency and performance to the extreme for specific hardware configurations, deployment environments, and wireless system constraints.

We continue to develop our standards-free communications system, learning and deployment software tools, and capabilities. We have conducted tests and deployments with partners including NASA and UAS vendors. We are working with new and existing partners to insert this technology into next-generation systems to save power, reduce parts costs, enhance system performance, and improve resilience and security.

DeepSig's solutions can enhance radio units (RU), distributed units (DU), and central units (CU) within Open RAN systems. Open RAN is making components, data, and algorithms within vRAN more accessible, which allows for algorithm innovation within commercial 5G stacks.

DeepSig's focus has been mainly within the DU in the 5G NR L1; however, the company sees numerous opportunities in L2 and L3 optimizations in scheduling, resource management, and closely coupled applications.

Hewlett Packard Enterprise and DeepSig Inc. have partnered to provide a machine learning-based signal identification solution for field applications. HPE provides the HPE Edgeline EL8000; a small multi-node server platform that runs DeepSig’s OmniSIG Sensor application. Edgeline and OmniSIG combine to create a platform for monitoring the wireless spectrum for a large number of signals.

The Edgeline (EL) 8000 is HPE’s converged edge compute system designed to provide data center computing in a compact form factor. It uses Intel processors, NVIDIA GPUs, NVMe SSD storage technology, networking options, HPE iLO security, and remote management features. HPE Edgeline 8000 can interface with a wide range of sensors, controllers, vehicle and system electronics, and industrial class systems. This system can run applications such as computer vision, AI, machine learning, and analytics.