<scp>LSTM</scp> ‐ <scp>iRealNVP</scp> : Enhanced industrial process monitoring via improved <scp>RealNVP</scp> flow models and fault‐free samples

Abstract Industrial process monitoring is the cornerstone of safe, high‐quality, and profitable production. Over the past decade, deep autoencoders have become a popular tool for extracting latent representations from high‐dimensional sensor streams. Yet every autoencoder ultimately learns a mere ‘numerical shadow’ of the data, scattering latent variables in unconstrained spaces. The resulting representations are irregular, discontinuous, and highly sensitive to noise. To address these challenges, we propose LSTM‐iRealNVP–an innovative normalizing flow framework integrating temporal feature extraction module. Our solution introduces three key advancements: (1) A mathematically rigorous invertible block, iRealNVP, replaces traditional autoencoders. By enforcing an explicit Gaussian prior through bijective transformations, it ensures constraints on latent variables. (2) A hybrid encoder merges LSTM dynamics with iRealNVP layers, simultaneously capturing non‐linear temporal dependencies and regularizing the latent space into a smooth, unimodal distribution. (3) An anomaly score derived directly from the exact log‐likelihood employs an adaptive threshold and remains robust across varying operating modes. Extensive validation on two industrial benchmarks confirms the transformative impact of these contributions. On the Tennessee Eastman process, LSTM‐iRealNVP pushes the detection rate to 82.41% while suppressing false alarms to 3.06%. On a full‐scale wastewater treatment plant, the method attains a near‐perfect detection rate of 99.75% at a false‐alarm rate of merely 0.8%, outperforming state‐of‐the‐art monitors. These results demonstrate that principled, likelihood‐driven representations can simultaneously elevate detection sensitivity and operational stability, setting a new performance frontier for industrial anomaly detection.