Labeling the observations collected from the deep manned submersible is time-consuming and difficult, and traditional water leakage detection method is supervised. The online and unsupervised water leakage detection methods are necessary but barely researched. In this paper, a multisensor fusion method based on the sparse representation is proposed, and the unsupervised Bayesian-based online diagnostic technique is adopted for water leakage detection in the deep manned submersible. (1) Construct a robust model: an improved sparse representation method is introduced for obtaining a coefficient matrix. (2) Obtain the variable relationship: a coefficient matrix fusion method using the correlation coefficients of the variables is developed. (3) Detect the water leakage: the online Bayesian change point detection method using the fusion data is attempted for water leakage detection. Finally, a numerical simulation and a real deep manned submersible are used to illustrate the power of the proposed method.