Signal processing can be divided into several categories:
Analog signal processing
Processing of continuous signals that have not been digitized, in which case the signal values are typically represented by a voltage, an electric current, or an electric charge around components in electronic devices. Analog signal processing is still relevant for many real world applications and is always the first step even when sampling and discretizing the signal for further digital processing.
Digital signal processing
Processing of digitized discrete-time sampled signals. Processing is done by general-purpose computers or by digital circuits such as applied specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or specialized digital signal processors (DSPs) or Advanced RISC Machines (ARM chips). Digital processing allows in many applications for several advantages over analog processing, such as error detection and correction in transmission as well as data compression. It also represents a fundamental technology for today's digital wireless communication and navigation systems.
Nonlinear signal processing
The classical approach to signal processing has always been to use linear methods and systems due to their interpretability and efficient implementation properties. However, for some applications it is beneficial to broaden the methodology to include non-linear processing methods. Denoising signals using wavelets and filterbanks, sparse sampling and fractional processes are examples of non-linear methods for signal processing that have been proven efficient in addressing many real world challenges.
Statistical signal processing
In many applications it is beneficial to have a model of the system under investigation. Unlike physical models, for example a swinging pendulum, most signals of interest are not deterministic in the sense that their behaviour can be perfectly predicted ahead of time. A model for such a signal will instead have to include broad properties, like the variation and correlation structure. This description is best done using mathematical statistics and the concept of stochastic processes. Using these models, optimality criteria can be expressed and achievable performance evaluated.
Machine learning is the study of computer algorithms that learn to do prediction and/or classification based on just a set of collected data, and without the strong assumption of an underlying model. It represents a subset of artificial intelligence, which refers to the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with sentient beings. The discipline of machine learning employs various approaches to teach computers to accomplish tasks where no satisfactory model is available.
We see machine learning as a natural extension of the classical signal processing paradigm, where the linear processing blocks are replaced by non-linear equivalents, enabling us to handle a much broader set of problems. Signal processing and machine learning can be used as orthogonal techniques, where domain knowledge is used with classical signal processing to obtain signal representations that are suited for machine learning. In modern approaches the machine learning techniques are integrated directly into the signal processing graph, performing non-linear prediction or dimensionality reduction as an integral part of the system. Approaches are traditionally divided into three broad categories, depending on the nature of the "signal" or "feedback" available to the learning system.
The computer is presented with example inputs and their desired outputs, and the goal is to learn a general rule that maps inputs to outputs. This is the most common application of machine learning and applies to both regression as well as classification/labelling tasks.
In this case no labels are given to the machine learning algorithm and the goal of the learning process will be to recover broad structures in the data. This can in turn be used to learn the underlying statistical distribution, how to generate artificial data points similar to those in the data set, or as a pre-processing step in a signal processing chain.
In a reinforcement learning setting the algorithm is directly interacting with the process it is observing. Examples includes playing a game, driving a car or controlling an industrial process. The goal of reinforcement learning is to find the optimal policy which attains the goal of the process, which may be winning the game, driving the car safely from A to B or controlling the process in within operating parameters.
Research and activities
Signal processing is a strong competence area in our department, and it serves as the fundamental tool in the majority of our R&D projects across all our research disciplines (wireless communications, navigation and acoustics). More recently, the rapid introduction and use of machine learning techniques has created an increase in new and innovative applications. This trend can be attributed to the Internet of things (IoT), where a large number of applications is based on analysing data from wireless sensors, delivered through a dedicated wireless connection. To reduce the amount of data being conveyed to a cloud-service for data processing, current trends attempt to shift the computational and storage tasks to involve computational elements residing on the edges of the network (edge computing). This introduces the concept of embedded machine learning in edge devices. Our view is that a strong know-how in signal processing is key to embedded machine learning. Hence, we try to combine our strong knowledge in signal processing with machine learning techniques in our R&D projects and use it as a powerful tool to extract information from signals and to implement intelligent algorithms to process, filter and classify data.
We are currently using our know-how in signal processing and/or machine learning in R&D projects within these technology and/or market domains:
- Intelligent transport systems (ITS)
- Water supply management
- Satellite communications (high-rate satellite modem)
- Powerline communications
- Subsea communications (acoustic communications)
- Speech technology
- Comprehensive privacy and security for resilient cyber-physical systems/IoT
- Non destructive testing (NDT)
- Ground penetrating radar applications
Typical projects for us are:
We conduct contract R&D as a partner for private industries and/or the public sector, and we deliver innovation by developing knowledge and technologies that are brought into practical use. We are usually involved in projects where R&D is used to increase the maturity level of a specific technology or a specific application of a technology from basic research to a more applicative solution. The maturity is commonly represented by a number from one to nine on a scale called the Technology Readiness Level (TRL), where one is used to represent pure basic research projects and nine is used to represent an actual system proven and ready for the market. We are mostly involved in projects trying to bridge the gap between TRL 2 – 6 on this scale. As a result, typical projects for us are the ones where we contribute with our know-how in activities such as:
- discuss and formulate technology concepts
- design electronic components or embedded systems for experimental proof of concepts
- validate a technology or an application in a controlled environment (laboratory testing)
- validate a technology or an application in a relevant environment (field testing)
- demonstrate a technology or an application in a relevant environment (field demonstration)
- being a trusted and independent technology advisor for future technology investments