Abstract Adaptive signal processing requires an efficient non‐stationarity detector. Most of the known non‐stationarity detection algorithms are based on residual statistics. The study proposes a novel non‐stationarity detection algorithm based on finite differences analysis of the processed signal. It also includes a suitable procedure for the forgetting factor design in the adaptation process. The performance of the proposed algorithm is experimentally compared with other known algorithms with regard to slow changes in signal stationarity as well as the influence of free coefficient selection on the quality of estimation. The developed algorithm exhibits the ability to effectively track both slow and abrupt changes in signal stationarity, with a small steady‐state error. The coefficients that need to be set during the application of this algorithm are given intuitive physical meaning. Simulations show good resistance to low signal‐to‐noise ratio and abrupt changes in noise variance. The results also demonstrate estimation performance relative to the water level signal from a thermal power plant steam separator.