<p>Intelligent manufacturing has raised the bar for product quality in industrial production, where the impact of faults on product quality directly determines maintenance costs. As such, quality-related fault detection is critical for the efficient and stable operation of industrial processes. However, in large-scale industrial systems, the influence of faults is typically confined to a few specific variables. Incorporating all variables into a single model may dilute or obscure critical fault information. This issue is addressed by introducing a sparsity block division with Hilbert–Schmidt Independence Criterion (HSIC) to segregate variables into quality-related and quality-unrelated groups. For the quality-unrelated group, we developed a stacked autoencoder with global and local structural constraints (GLSAE). For the quality-related group, a hybrid GLSAE model, consisting of the knowledge-sharing nonlinear encodernets and the linear decodernet, was constructed. This framework facilitates the effective prediction of quality indicators and process variables through the shared knowledge mechanism of encodernets. Finally, we constructed the quality-related subspace and the residual subspace through linear decomposition and combined them with the information of the unrelated group to devise relevant statistical metrics. The superiority of the algorithm over current state-of-the-art methods is demonstrated through its successful application to both the Tennessee Eastman (TE) process and an industrial-scale fluid catalytic cracking (FCC) unit.</p>