Sensitivity Analysis of Text Vectorization Techniques for Failure Analysis: A Latent Dirichlet Allocation and Generalized Variational Autoencoder Approach

Abstract

Abstract Failure analysis has grown in importance as a means of ensuring high quality in the production of electronic components. The findings of a failure analysis can be used to pinpoint weaknesses in a component and get a deeper understanding of the mechanisms and causes of failure, enabling the adoption of corrective actions to raise the quality and reliability of the final products. A failure reporting, analysis, and corrective action system (FRACAS) is a method for organizations to record, categorize, and assess failures as well as plan corrective actions. Any reports of failure, together with a history of failure and any related corrective activities, should be formally documented in order to achieve standards. These text feature datasets must first be preprocessed by pipeline techniques and converted to digital by the vectorization method to be ready to begin extracting information and building a predictive model to predict the topics of failure conclusions from failure description features. Text data is an important data type that directly reflects semantic information. However, the optimal choice of text vectorization method is an important concept in natural language processing tasks. In fact, text data cannot be directly used for model parameter training, it is necessary to vectorize the original text data of failure analysis and make it numerical, and then the feature extraction operation can be carried out or creating predictive models suitable for failure analysis We are initially focused on studying sensitivity analysis in relation to the implementation of various vectorization techniques for textual data in the context of failure analysis. To do this, we propose a new methodology based on the combination of latent Dirichlet allocation (LDA) topic model which discovers underlying topics in a collection of failure conclusion and infers word probabilities in topics, and Generalized Variational Autoencoder which is an unsupervised neural network model with objective of reconstructing its input of vectorized data of description failure analysis by compressing it into a latent space using an encoder-decoder network. The comparison of text vectorization methods is possible by checking the accuracy of supervised classification. Experiments of our proposed methodology on textual datasets of failure analysis demonstrate the effectiveness of the Wored2Vec technique, which allows better discrimination of textual classes compared to the use of Doc2Vec or Term Frequency-Inverse Document Frequency (TFIDF).