Regional gravity data contain rich information about deep structures in sedimentary basins. After analyzing problems that occur during oil/gas exploration in Manjar depression within Tarim Basin, we complete dense measurements of surface gravity investigation in the depression, then use wavelet transformation and scratch analysis methods to study deep structures, locating the density disturbance belts exist in early Paleozoic strata. By applying the wavelet multi-scale decomposition method to regional gravity data, we are able to extract information only related to deep structures. By applying an artificial intelligence method called scratch analysis, we are able to recognize characteristic parameters about scratches contained in regional gravity field, providing objective evidences for dividing tectonic units within the basin. The combination of these two methods is called the multi-scale scratch analysis method, producing both the ridge parameter and scratch-edge maps for both shallow and deep structures in Manjar depression. Comparing the map of density distribution with MT electrical resistivity on depth of 7~8km, we find that anomalies of both low density belts and low resistivity belts occur along latitude 40° and longitude 85°, showing the correlation between the low density distributions and ancient fluid channels. The results from the multi-scale scratch analysis indicate that there are some sub-tectonic units exist in the deep Manjar depression, their location provide some important basic data for deep-strata oil/gas exploration in the depression.