Probabilistic learning technique for improved performance of servosystems with incremental encoder feedback

The requirement of high resolution positional and digital velocity information from incremental, optical encoder sensors, has precipitated the widespread use of arctangent computations on sinusoidal/cosinusoidal encoder outputs. The accuracy of such resolution enhancement is often poor due to non-idealities in the encoder disk and in associated instrumentation electronics. Existing methods of determining an accurate mapping from the sensor output to the actual shaft position are cumbersome and require specialized high resolution equipment. In this paper, a new compensation technique is proposed. This technique can be utilized on any digitally based motion control system, without the use of any specialized calibration equipment. During an initial learning phase, the rotor/encoder combination is run at a near constant velocity, with some pseudo-random variation. The nominally sinusoidal and cosinusoidal sensor signals are processed at regular intervals using a two argument arctangent function. The relative frequency of occurrence of each resultant digital code is then utilized to yield a compensation function. This function provides the average compensation required over the mechanical cycle. It is shown that undersampling of individual waveforms can be tolerated, provided that they exhibit a periodicity and that pseudorandom sampling is employed. Low-velocity servosystem performance is shown to be greatly improved. Experimental results are included.