Analyzing abnormal pattern of hotelling T2 control chart for compositional data using artificial neural networks

• Published in: Computers & industrial engineering 2023-06, Vol.180, Article 109254
• Main Authors: Zaidi, Fatima Sehar, Dai, Hong-Liang, Imran, Muhammad, Tran, Kim Phuc
• Format: Article
• Language: English
• Subjects: Artificial neural networks; CoDa; Hotelling T2; Multilayer feed-forward; Pattern recognition
• ISSN: 0360-8352; 1879-0550
• DOI: 10.1016/j.cie.2023.109254

Compositional data (CoDa) has been monitored in statistical process monitoring, where multivariate control charts (CCs) such as Hotelling TC2, MEWMA-CoDa, and MCUSUM-CoDa are commonly used to determine if a process is in-control. However, these charts can encounter problems when there is an out-of-control (OOC) process due to various factors such as shifts in variables, outliers, or trends. To address this issue, a pattern recognition (PR) tool using multilayer feed-forward neural networks (MLFFNN) is proposed to accurately recognize CoDa patterns. In the simulation study, six different models in simplex sample space are used to induce trends and shifts in CoDa, and sufficient samples are generated to evaluate the proposed PR model’s performance. The isometric log-ratio (ilr) transformation is applied to CoDa to convert the data from simplex sample space to real space. The Hotelling TC2 statistic is obtained from the generated values after applying the ilr transformation. TC2 statistic is then standardized for MLFFNN, and a back-propagation learning algorithm is used to accurately fit the PR model. Results show the proposed model accurately identifies the CCs pattern, even during OOC processes. A time budget CoDa is analyzed to demonstrate the proposed model’s effectiveness in recognizing patterns. •CoDa is generated using the different models in simplex sample space to induce patterns.•Hotelling TC2 CC are applied on generated CoDa using ilr transformation.•The trend in CoDa using the BP technique of MLFFNN.•An ANN PR model is established with different assumptions to detect the data patterns.•The performance is studied using different types of shifts and trends.