Robust blind calibration via total least squares

This paper considers the problem of blindly calibrating large sensor networks to account for unknown gain and offset in each sensor. Under the assumption that the true signals measured by the sensors lie in a known lower dimensional subspace, previous work has shown that blind calibration is possible. In practical scenarios, perfect signal subspace knowledge is difficult to obtain. In this paper, we show that a solution robust to misspecification of the signal subspace can be obtained using total least squares (TLS) estimation. This formulation provides significant performance benefits over the standard least squares approach, as we show. Next, we extend this TLS algorithm for incorporating exact knowledge of a few sensor gains, termed partially-blind total least squares.