The invention provides a pipeline defect size inversion method based on SL. The method comprises the steps that data interpolation preprocessing is carried out on a defect triaxial magnetic flux leakage signal sample set of known defect size; multi-dimensional feature extraction is carried out to extract a signal feature set [F(t), F(delta)] in time domain; in frequency domain, a method based on wavelet transform is used to construct a wavelet energy feature set F(w); a multi-dimensional feature set of nsample defects F=[F (t), F(delta), F(w)] is constructed; an iterative Stacking Leaning network is used to automatically determine the final network structure; and the size of a magnetic flux leakage signal to be measured is predicted to acquire the prediction result of the pipeline defect size. According to the invention, the time-frequency domain multi-dimensional feature extraction method is used to construct defect features; information contained in the magnetic flux leakage signal is comprehensively analyzed; the size prediction capability of the network for complex defects is improved; the method has strong robustness and realizes defect fault diagnosis in a short time; socialharm caused by pipeline leakage is reduced; adaptability to different fault diagnosis sample sets is realized; and the network is universal and portable in the industrial field.