Incipient faults often exhibit insignificant changes in data distribution, and the presence of variable collinearity in high-dimensional data can result in highly ill-conditioned or even singular sample covariance matrices. Consequently, traditional methods based on canonical variate analysis face challenges in inverting the covariance matrix and fail to detect incipient faults in a timely manner. To address these issues, this paper proposes a method for incipient fault detection by integrating probability information with sparse canonical variate analysis(PSCVA). First, an L₀ constraint is imposed when solving for canonical variates, and the data matrix is decomposed using a mixed-integer quadratic optimization method to obtain sparse canonical variates. This enhances intuitive understanding of potential relationships among variables and helps identify key fault variables. Second, sparse canonical variates obtained during the normal phase are used to construct canonical vectors, residual vectors, and canonical variate dissimilarity. Furthermore, the difference in the probability distribution between the normal sample and the faulty sample is considered, and the Wasserstein distance is introduced to construct probabilistic fault detection indicators to improve the detection performance of incipient faults. Additionally, the kernel density estimation is used to determine control limits for statistical indicators under non-Gaussian distributions. Finally, the experimental results from the Tennessee-Eastman(TE) process and the continuous stirred tank reactor(CSTR) system show that compared to canonical variate dissimilarity analysis(CVDA) and probability CVDA, the proposed method achieves detection rate gains of 27.53 % and 10.68 % in the TE process, and raises fault detection by 106 and 60 samples in the CSTR system, respectively.