Abstract Deep learning-based process monitoring has achieved remarkable progress. Generally, a deep model is empirically selected before the data features are learned. In this study, the interpretability and suitability of stacked denoising autoencoder (SDAE) in process monitoring territory are theoretically analyzed and validated. Considering that the data will show different feature representations at different scales, such as overall outline, local information, and microscopic details, this study utilizes the concept of multiscale analysis to mine the feature information of raw data deeply in different scales. The multiscale analysis is performed on the basis of agglomerative hierarchical clustering and silhouette coefficient, which makes the analysis data characteristics-based and intelligently abandons the intervention of manual prior knowledge. Then, the SDAE models are established under each scale to learn the high-order and robust features from the data with noise and fluctuation, and all monitoring results of the different scales are integrated using the Bayesian inference. Finally, given the temporal information in sequence data, the state representation of previous events is embedded into the current decision through a sliding window. The numerical process, benchmark Tennessee Eastman and real steel plate process are used to analyze the superiority of the proposed method (MSDAE-TP) over other deep learning-based monitoring methods.