The DC resistivity method is an economical and efficient engineering geophysical detection method. It has been widely applied in engineering detection due to its high sensitivity to a water-bearing geological body. The linear inversion is now mainly used in the data imaging and interpretation of the DC resistivity method in real applications. However

its inversion results are easy to fall into the local optimum

which may lead to a wrong imaging result as well as the geological interpretation. The unsupervised inversion method can use both the physical laws and data mining. It could get rid of the dependence on the real model

and can retain the feasibility of global search ability in the real data. Based on the unsupervised inversion method

we developed a deep learning boundary characterization method based on multi-scale edge features. To solve the problem of the fuzzy boundary of inversion imaging

we proposed a multi-scale inversion method of resistivity based on convolution wavelet transform by drawing on the experience of multi-scale inversion in seismic and electromagnetic exploration. On this basis

we used the multi-scale inversion objective function as the loss function to correct the network gradient

and thus effectively improved its boundary characterization ability. We carried out field tests in the No. 1 air shaft project of the Shanghai regional airport connecting line. Taking the detection of the leakage point of the diaphragm wall of the 5th foundation pit as an example

we verified 15 low resistance anomalies to guide the reinforcement of the foundation pit

which proves that the method is feasible and effective.