To address the issues of nonlinearity and high dimensionality in industrial process data, a fault detection method combining Canonical Variate Dissimilarity Analysis and Locally Linear Embedding is proposed. The dissimilarity matrix constructed by the CVDA algorithm can effectively monitor faults, but it relies on linear projections and is only sensitive to changes in linear features of the data structure. The LLE algorithm is used to map high-dimensional data to a low-dimensional space by preserving local relationships between samples, further extracting features and uncovering nonlinear characteristics and local neighborhood information. Finally, an isolation forest model is established in the low-dimensional manifold space to obtain anomaly scores of sample points as the fault detection evaluation criterion. Through a set of nonlinear numerical examples and the Tennessee Eastman chemical process data, the proposed method is compared and analyzed with traditional KPCA、PPA and CVDA to verify its effectiveness and superiority.