The work was done under contract with Samsung Research USA.

The purpose was to create a smart watch app to recognize gestures of a hand with watch on it. The app uses data from accelerometer and gyrometer (acceleration, rotation and rotation acceleration in 3 dimensions) to recognize snap, flick, flip, double snap and double flip.

The gestures have to be identified within normal hand movements for any user. One of the tasks was to distinguish the specified gestures from other normal activities like cooking, typing, adjusting watch and so on.

The gestures shall be recognized by watch's computer in real time. So, complex algorithms typical for signal processing could not be used.

The first task was to identify objects for recognition: "events", or segments of data where a gesture may be located. This was achieved by calculating of a parameter 'meta-acceleration' for every measurement, and finding its peaks and valleys.

Large number of informative features were developed to describe the events for classification. Some of the features formalize discovered patters, others were based on Haar wavelets.

The gesture recognition procedure uses combination of logical rules with linear SVMs. The accuracy ranges from 89% (for flicks) to 98% (for flips). The app is launched. Provisional patent application "USER INTERFACE CONTROL USING IMPACT GESTURES" was submitted.

When an engine works, its cylinders fire consequently. It is why the data from piezoelectric sensors can be used to characterize work of each cylinder and the differences between them.

The project was executed in Colorado company Biodesix.

Mass spectra are used to find biochemical properties of patients' tissues in medical research. Each spectrum is a curve with time on x-axis, and abundance of ions registers by the sensor at this time on y-axis. The time, when a ion is registered on the sensor, depends of its mass and charge (m/z).

The problem is that MS spectra are usually not aligned or synchronized between samples, so that, in different samples, a peak corresponding to a particular molecular weight and charge will appear in slightly different times, different places on x-axis.

So, before mass spectra from several samples can be analyzed, they require an "alignment" to find peaks with the same m/z across all spectra. The problem is known as spectra alignment or "warping". Most of the available software for this task requires manually defining "interesting" m/z values, around which "close" peaks are found and matched across spectra. Mismatched peaks lead to "misalignment", and the results of the procedure strongly depend on the original selected peaks. The proposed new method is based on selection for matching of large key intervals of signal with multiple peaks, rather than single peaks. Matching the large intervals with complex profile (as key and lock) resulted in significantly improved alignment. The proposed method solves the problem by maximizing cross-correlation between the aligned spectra. In addition, the method works many times faster than the previously available method. By company's request, the program was written in C# and is built in the company's spectra analysis software.