Selective Reporting and Evaluation Protocol Shape the Evidence on Deep Learning Fault Detection in the Tennessee Eastman Process: A Systematic Review and Meta-Analysis

The reported gains of deep learning for fault detection on the Tennessee Eastman Process (TEP) have not been synthesised. In this systematic review, registered retrospectively in OSF Registries, we searched Scopus and the Web of Science Core Collection for journal articles from 2017 to 2026 reporting scenario-level results of deep learning methods. Results were pooled per fault scenario with random-effects models, separately for monitoring schemes, which report the fault detection rate, and supervised classifiers, which report per-class recall. Of 266 records, 101 studies were included. Most scenarios were detected almost completely, but monitoring studies detected only 15.8%, 11.2% and 18.0% of IDV 3, 9 and 15, whereas classifiers reported per-class recall of 94.8%, 87.2% and 81.9%. The difficult scenarios were the least often reported, with an odds ratio of 1.26 per 10 percentage points of detection, 95% CI 1.21 to 1.32. Under paradigm-specific imputation, omitting them raised the published averages of monitoring studies by about 6.8 percentage points, an estimate that rests on assumptions. Against the strongest comparator in the same experiment, the proposed method was inferior or tied in 32% of scenario pairs away from the detection ceiling. Most scenarios leave little room for gains, and progress on the difficult faults requires complete and comparable reporting.