To address the challenge that continuous small leakage signals are easily disrupted by noise, resulting in a low recognition rate for urban pipeline leakage, we propose an improved multivariate variational mode decomposition (IMVMD) fusion machine learning method specifically for the recognition of continuous small leakages in urban pipelines. Building upon the preliminary time-frequency assessment of the original leakage signal, we enhance the MVMD by incorporating the correlation coefficient and normalized Shannon entropy, enabling adaptive decomposition and reconstruction of the leakage signals. We establish a BP neural network based on the IMVMD and a SVM leakage recognition model also based on IMVMD. Random forest (RF) evaluation is employed to identify the signal feature inputs. The results indicate that the signal-to-noise ratio of the reconstructed signal using IMVMD is 55.42% higher than that of the original signal, demonstrating a superior decomposition effect compared to traditional MVMD EMD and VMD. RF is utilized to reduce the dimensionality of signal characteristics under various leakage conditions, resulting in the selection of four representative features: root mean square, short-term energy, margin factor, and waveform factor, which serve as inputs for the BP neural network and SVM leakage recognition model based on IMVMD. The accuracy of signal recognition reaches 98.22% and 97.22%, respectively. Compared to the traditional MVMD decomposition recognition model, this method improves accuracy by 10.72% and 10.22%, respectively, thereby providing reliable support for the detection and precise localization of continuous small leakages in urban pipelines.