Recurrent Neural Network‐Driven Interpretation of Anomalous Signals in Compositional Data Using MEWMA Control Charts

Compositional data (CoDa) monitoring in production processes often relies on multivariate control charts (CCs) with isometric log‐ratio () transformation to distinguish in‐control and out‐of‐control (OOC) conditions. Traditional CCs effectively detect shifts in the mean vector but struggle with anomalies affecting specific variables. A recurrent neural network (RNN) with backpropagation through time is introduced to interpret OOC signals in the multivariate exponentially weighted moving average CoDa control chart (MEWMA‐CoDa CC) to address this challenge. The proposed model pinpoints anomalous variables responsible for OOC events, enhancing process control strategies and improving production efficiency. The model's effectiveness is demonstrated through two CoDa case studies. A P = 5‐part CoDa dataset on oil production across five U.S. states (Alaska, Arkansas, Arizona, Colorado, California) is analyzed, with the RNN achieving 93.5% accuracy in identifying the shifted variable. Another case examines agricultural production across four Chinese cities using a P = 4‐part CoDa framework, comparing MEWMA‐CoDa CC to Hotelling's ‐CoDa CC. Results indicate superior accuracy for MEWMA‐CoDa CC in interpreting OOC signals. A novel approach for CoDa process control is presented, offering significant advancements in anomaly detection and production process optimization.