The shape and size characteristics of underground cavity are of great significance to the stability calculation and evaluation of geological structure in the region. However, the borehole sonar currently used lacks universality and it is difficult to meet the requirements of both detection accuracy and detection range. Therefore, on the basis of traditional borehole sonar method, this paper proposes a new method of detecting cavity based on multi-frequency borehole sonar. With considerations given to the contribution of different frequencies to the actual detection object, this method synthesizes the detection data of various frequencies and effectively solves the problem of the opposition between detection scale and detection accuracy. Firstly, the optimal frequency expression of the detection system is derived to provide basic parameters to reflect the influence of different frequencies on the analysis of detection results to a greater extent, and the frequency influence factor parameter is established to describe the deviation degree between the natural frequency and the optimal frequency. Then, on the basis of superposition of the frequency influence factor, the detection data of various detection frequencies are synthesised, and detection objects are accurately measured by using the improved spectrum ranging method. Finally, after adding depth and azimuth information, the three-dimensional reconstruction of cavity shape is realized. The feasibility and accuracy of the method are verified by applying it to practical engineering projects.