An Improved LSTM Text Classification Model for Factory Report

Effectively used maintenance data can greatly benefit the manufacturing process. However, managing massive quantities of data in a variety of formats without performing the necessary analysis might result in unplanned, expensive, and long-running downtime. Text preprocessing is a crucial stage in text classification under the manufacturing domain. Using text preprocessing techniques, several word forms are condensed into a single form. Moreover, text preprocessing methods are given a great deal of significance and are extensively researched in machine learning. Preprocessing features, extracting significant characteristics, and comparing them to features in a database are the fundamental steps in text classification. Most studies considered three main steps; tokenization, stemming, stop words removal. However, stemmed words can sometimes be out of context and affect the accuracy of the training model. This paper investigates the effect of the preprocessing tasks on English manufacturing report for text classification using LSTM network due to its ability to classify sequential data with memory cells. Phase I shows that there is a significant increment on average 3.12–3.13% of accuracy on every preprocessing task added, while in Phase II, the accuracy difference between lemmatization and stemming demonstrates for 5.21%. Lemmatization outperforms stemming due to the importance of context in the factory report dataset.