An on-line robust wavelet filtering is presented and applied to the dynamic data reconciliation problem via a constrained Kalman filter approach. The wavelet filtering is used to remove outliers and provide data smoothing prior to the reconciliation. Matrix computation is presented to facilitate the implementation of discrete wavelet transform (DWT) and inverse discrete wavelet transform (IDWT) for on-line filtering. An endpoint correction method is presented to overcome the endpoint effect that is caused by wavelet filtering on an on-line moving data window. The filtered outputs are treated as the output-measurements in the subsequent Kalman filter estimations. This latter filter is to estimate the state variables, subject to both dynamic and static equality constraints. Using accumulative balancing constraints, a method is proposed to detect and isolate the existence of single gross error in a dynamic system. A simulated example is used to illustrate the use and performance of this proposed dynamic data reconciliation method.