Objective: In modern chemical production, process monitoring is important for ensuring operational safety, improving product quality, and enhancing economic benefits. In recent years, the deep integration and widespread application of distributed control systems, the Internet of Things, and large-scale data storage infrastructure have significantly improved the capability to acquire industrial process data in real time. Within this context, data-driven process monitoring systems have gradually become a major branch of research in the chemical production field because they can operate without relying on mechanistic models and have strong capabilities for learning from historical data. However, practical industrial processes generally exhibit significant non-stationary characteristics. Unlike stationary processes, the statistical characteristics of non-stationary process variables change over time, making it difficult to effectively distinguish between actual process faults and normal non-stationary variations using traditional multivariate statistical process monitoring methods, such as principal component analysis (PCA). Consequently, the rate of false alarms increases, the fault detection sensitivity declines, and the reliability and accuracy of monitoring systems are severely affected. Furthermore, complex physicochemical reactions and production equipment result in nonlinear relationships among process variables, which further complicates process monitoring. Conventional linear methods are inadequate for capturing the complex nonlinear interactions between variables, result ing in unsatisfactory performance of models and monitoring systems. Methods: To address these challenges, an integrated process monitoring method that combines stationary subspace analysis (SSA) with kernel principal component analysis (KPCA) is proposed in this study. First, SSA is used to process the original data, decomposing it into stationary and non-stationary subspaces. By extracting and retaining the stationary components, the approach effectively eliminates interference caused by non-stationary trends and supplies data with stable statistical characteristics for subsequent analysis. The processed stationary data are then input into the KPCA model. Using a kernel function, the data are implicitly mapped into a high-dimensional feature space, where linear PCA is performed, substantially enhancing the ability to capture complex nonlinear relationships. This monitoring strategy effectively overcomes the limitations of conventional methods in handling both non-stationarity and nonlinearity. The effectiveness of the proposed method was validated by application in an industrial case study involving continuous catalytic reforming. Results: SSA successfully separated the stationary source signals, providing an ideal input for KPCA, fully leveraging its advantages in nonlinear feature extraction. The proposed method achieved effective fault detection while maintaining a low false-alarm rate. Comparative experiments with traditional methods, such as PCA and cointegration analysis, further highlight the superiority of the proposed approach. Conclusions: Conventional methods are ineffective for handling the combined effects of non-stationarity and nonlinearity and thus exhibit limited fault identification capability and high false-alarm rates. In contrast, the proposed method maintains an extremely low false-alarm rate under normal operating conditions while enabling rapid and accurate alarms, significantly improving the precision and reliability of process monitoring, demonstrating superior overall monitoring performance. Such improvements in practical industrial applications can greatly reduce unnecessary production interventions and shutdowns caused by false alarms, avoiding substantial economic losses and providing reliable technical support for achieving safe, efficient, and stable production.