Estudo de viabilidade de um método baseado em inteligência artificial para classificação de anomalias em máquina embaladora usando dados coletados em produção, para auxiliar na manutenção

Abstract

This work evaluates the prediction of using an Autoencoder based on Long Short-Term Memory (LSTM) recurrent neural networks for anomaly detection in operational signals of a machine packaged with Vega Shrink-Wrapper. The model is trained exclusively with data considered normal, learning the typical temporal pattern of machine operation, and anomaly identification is performed through the occurrence error of data windows, calculated by the Occurrence Error (MSE). A publicly available real dataset containing time series from industrial sensors available over six months of operation is used. The methodology involves data preparation and windowing, training the LSTM Autoencoder, and evaluating the model's performance in two data splitting scenarios (70/30 and 90/10), using metrics such as accuracy, precision, recall, F1-score, and Precision-Recall curves. As a result, performance analyses were calculated that allowed a direct comparison between the two data splitting scenarios.