The task of the our signal-processing group is to extract useful information from sensor signals and to make sure the sensors operate optimally. By handling the signals in all steps from the physical sensor to real-time signal processing to the graphical user interface, we produce measurement systems that are very accurate, very fast and also easy to set up, operate and interpret.
Our tools for offline modeling, analysis and design include Matlab, Simulink, C/C++, and of course, old-fashioned pen-and-paper. For real-time embedded implementation, we use high-performance programmable digital hardware (FPGAs), DSP processors, microcontrollers, and customized integrated circuits — all adapted to the demands of the specific application.
Real-time signal processing
Our experience in real-time signal-processing and feedback-control systems is essential for achieving state-of-the art sensor performance. For instance, we stabilize our micro-mechanical inertial sensors using digital control systems, perform advanced filtering (phase-shaping filters, noise reduction, equalization etc.), and extract the useful sensor signals by digital coherent detection methods – all in real time.
Before the digital control system can operate, the signals have to be transferred from the analog to the digital domain. We understand the issues involved in the critical transition from A-to-D and vice versa. We are experienced in custom mixed-signal electronics, digital noise shaping, interpolation/decimation, as well as optical sampling and digitization at gigahertz sample rates.
System modeling and analysis
In order to extract useful information from the sensor signals, it is necessary to have accurate models of the sensors, the electronics, the noise, and the specific application dynamics. Our expertise in system identification allows us to design efficient models that enable us to characterize physical parameters from the sensor response. After thorough calibration, we use the sensor models to compensate for such factors as temperature dependence and nonlinearity.
With our experience in optical research, we can design, model and optimize imaging systems for specific measurement tasks, taking both optical aberrations and image-sensor deficiencies into account. After calibration and compensation, we achieve a positional resolution much better than one image sensor “pixel”, which we exploit for applications like 3D-triangulation measurements.
Estimation and pattern recognition
We use statistical estimation techniques (e.g. Kalman filtering) to extract physical quantities that cannot be directly observed, to estimate measurement uncertainties, and to fuse information from several sensors and/or points in time. Using our custom nonlinear estimation software, we can easily adapt to a specific application, whereby we combine knowledge of the system, noise, and the expected behavior to improve the estimation accuracy. We are particularly experienced in the field of inertial navigation, where we use Kalman estimation supported by application-specific kinematic models to follow accurately the position and orientation of the device from measurements of the acceleration and the rate of rotation.
To further enhance the estimation quality, we use pattern-recognition methods and machine learning to choose from different process models. In the case of navigation, these models correspond to different modes of motion (“at rest”, “constant velocity”, “walking”, “running”, etc.), with optimized signal processing for each type of motion. We are also using pattern recognition in non-navigation applications for the classification and identification of more general events.