Deep learning is gradually mainstreaming into data‐driven methods, relying on the advantages of extracting complicated nonlinear features. However, the black‐box property makes its decision rules non‐transparent, resulting in difficulty in attribution tasks, which aim to backtrack the contribution of network inputs to the outputs. Fault isolation and localization are techniques for diagnosing the root cause of system failures, which have a consistent objective with attribution for a deep learning‐based fault observer or classifier. Unfortunately, most fault isolation methods are based on shallow learning methods. Also, many attribution algorithms are linear without considering the influence of nonlinear activation functions. The related concerns motivate us to propose a new approach, namely layer‐wise contribution‐filtered propagation (LCP), for deep learning‐based fault isolation. In LCP, reasonable contributions are defined based on the influence of each layer input on maximizing the absolute output activation. A symbolic function is designed to identify neurons with negative contributions, which are then filtered and forbidden to backpropagate to the previous layer. By guiding correct attribution, LCP is available for any nonlinear activation functions and their combinations. It also provides a solution for fault isolation with stacked sample inputs, in which one single variable has several attributions associated with different times. Finally, two chemical simulations verify the effectiveness of the proposed method.