Geomagnetic secular variation (SV) models for the epochs before the space era are based on magnetic observatory data, which represent relatively rough and noisy time series due to magnetic storms, anthropogenic spikes and gaps. These models are often strongly regularized in time, so that fast variations in the SV are smoothed out. However, recent studies show that at least some of the geomagnetic jerks observed at the Earth’s surface emanate from increasing and decreasing phases of secular acceleration (SA) pulses at the core surface. The latter ones are direct manifestation of the dynamic processes taking place in the liquid core. They were first detected from satellite data, which are both of higher quality and more homogeneous in terms of geographical coverage than ground data. Herein we attempt to carry out similar studies based on observatory data available for a longer period. The proposed method of SV modeling and recognition of SA pulses relies on a new technique of processing time series based on fuzzy mathematics. Comparison with the SV modeling results derived from satellite data shows their high conformity with the proposed method. Stability and reliability of the SA pulse recognition are demonstrated by the examples of well-studied SA pulses in 2006, 2009 and 2012. Moreover, several new SA pulses around 1996, 1999, 2002 and 2014 are discovered as a result of the new approach application to multi-observatory data analysis. The latter provides a basis for applying the method to older historical data and investigate SA pulses and geomagnetic jerks further back in time.