Drilling leaks are a major hazard in oil extraction, seriously affecting industrial production, social life and the natural environment. When a well leaks, the pressure fluctuation of the pressurised fluid in the well forms a pressure wave signal, which can be used to diagnose the leak location. However, the transient pressure wave signal is often interfered by white noise, and the traditional frequency domain filtering method is difficult to effectively denoise, which limits the accurate extraction of leakage features. An improved LCD (local feature scale decomposition) method, combined with SVR (support vector regression), was proposed to solve the problems of modal aliasing and endpoint effects. Meanwhile, a component division method based on the Marginal distance and cumulative mean was proposed to improve the quality of signal reconstruction. In addition, the neighbourhood difference method was used to analyse the variation of signal characteristics to achieve the accurate positioning of the leakage location. The effectiveness of the proposed method was verified by establishing a model and building an indoor well leakage detection device. The experimental study shows that the experimental error of the method is less than 5% and the positioning accuracy is high, which significantly improves the precision and reliability of well leakage detection and provides a new technical support for well leakage diagnosis in oil extraction.