This paper proposes a three-dimensional quantitative intelligent recognition method for caves in multi-layer rock–soil environment based on bidirectional long short-term memory (BiLSTM) neural network. Firstly, referring to the principle of shallow seismic reflection wave method, a multi-layer rock–soil structure model containing a single unfilled spherical cavity is established, and the acceleration response signals of the ground surface under pile hammer excitation are calculated. Secondly, the response signals under different working conditions are obtained as data sets by lots of simulation models for different working conditions. Finally, a dual channel separated architecture network model is designed based on BiLSTM, which realizes the quantitative identification of different karst cave working conditions in multi-layer rock–soil environment. The results show that the method proposed in this paper can quantitatively identify the three-dimensional position and diameter of caves under different rock-soil structures, and the identification accuracy within the tolerance error of 3 m has reached more than 98%.