摘要
Deep learning has boosted process soft sensing, but field performance often falls short because measurements are imperfect. Outliers, noise, and missing data are common in plants, and they distort the learned mapping between inputs and targets. This review outlines a practical agenda for deep learning-assisted soft sensors under imperfect measurement. First, a measurement-centered view is adopted to scope three major imperfection classes in practice. Second, methods are organized along three fronts that align with deployment needs: resisting outliers, reducing noise, and learning with incomplete data. The main ideas are synthesized for each front, and their strengths, limits, and suitability to plant constraints are summarized. The review goes beyond single imperfections. Interactions among outliers, noise, and missing values are analyzed, and their joint impact on training, validation, and online use is explained. A deployment pathway is then given, including stress testing under mixed imperfection, calibration of uncertainty, and rules for safe action when confidence is low. Compared with prior surveys focusing on outlier detection, robust modeling, or data cleaning in isolation, this work provides a unified and deployment-focused view specific to soft sensing, with clear links from measurement defects to reliable practice in plants.