2024 | Fadhila Lachekhab, Messouada Benzaoui, Sid Ahmed Tadjer, Abdelkrim Bensmaïne, Hichem Hamma
This paper presents an LSTM-autoencoder deep learning model for anomaly detection in electrical motors. The primary goal is to develop a solution that can detect anomalies in the variation vibrations of an electrical motor along three axes: axial (X), radial (Y), and tangential (Z). The model combines LSTM layers with an autoencoder to leverage the LSTM's ability to handle large amounts of temporal data. The effectiveness of the LSTM-autoencoder is compared to a regular autoencoder using Python and TensorFlow. The comparison is based on three main metrics: training time, loss function, and mean squared error (MSE) anomalies. The results show that the LSTM-autoencoder achieves significantly smaller loss values and MSE anomalies compared to the regular autoencoder, despite being slower due to the added LSTM layers. The study highlights the advantages of using LSTM-based architectures for handling sequential and time-dependent data, making it a promising approach for predictive maintenance in industrial settings.This paper presents an LSTM-autoencoder deep learning model for anomaly detection in electrical motors. The primary goal is to develop a solution that can detect anomalies in the variation vibrations of an electrical motor along three axes: axial (X), radial (Y), and tangential (Z). The model combines LSTM layers with an autoencoder to leverage the LSTM's ability to handle large amounts of temporal data. The effectiveness of the LSTM-autoencoder is compared to a regular autoencoder using Python and TensorFlow. The comparison is based on three main metrics: training time, loss function, and mean squared error (MSE) anomalies. The results show that the LSTM-autoencoder achieves significantly smaller loss values and MSE anomalies compared to the regular autoencoder, despite being slower due to the added LSTM layers. The study highlights the advantages of using LSTM-based architectures for handling sequential and time-dependent data, making it a promising approach for predictive maintenance in industrial settings.