A Hybrid CNN-LSTM Based Network for Rolling Bearing Fault Detection

Anomaly or fault detection on machines is relevant in many industries, like wind power generation. Bearing failures, a common cause of breakdowns, pose significant challenges, leading to increased downtime and maintenance costs. Recent machine learning techniques, such as convolutional neural networks (CNNs) and long short-term memory (LSTM), have been exploited to extract spatial and temporal relationships from data to detect failures. Many literature works use the Paderborn bearing vibration data in their studies, which mainly rely on either CNNs or LSTM models to accomplish the detection task. In this paper, we use a hybrid CNN-LSTM network architecture to achieve high accuracy for bearing fault detection using the Paderborn dataset. In addition, a two-sliding window technique is used to prepare the data to pass through both types of layers. Our study emphasizes the importance of selecting appropriate architectures and considering bearing characteristics and operation conditions for effective fault detection. Our results achieved up to 99.58% accuracy. However, when tested on different machine operating conditions, the model’s performance degraded, indicating the necessity of other methods to address this issue. In addition, we verified that an improper model tunning can degrade its performance, even if a complex hybrid method is considered, highlighting the importance of hyperparameters tunning in fault detection problems.