The independent monitoring sources and single-type fault features fail to comprehensively reflect the health status of rotating machinery, leading to insufficient diagnostic capabilities. To address this issue, a fault diagnosis method based on temporal-spatial vibration feature fusion extraction is proposed. First, an embedded parallel attention convolutional neural network (PACNN) is utilized to fuse vibration signals from different sensors into early features. Then, the improved Resnet18 and BiLSTM are used to extract the deep spatial features and temporal features from the early fusion features respectively based on the parallel extraction framework, which avoids the problem of mutual interference when getting the combined features, and the suitable attention modules for different models are applied to enhance the diagnosis performance. Additionally, global sequential features such as kurtosis and RMS are extracted to increase the feature diversity. All the features are then concatenated and input into a trainable fully connected network to identify the health status. The proposed method is applied to a fault dataset, and its effectiveness is demonstrated.
1 October 2024 in Advanced Engineering Informatics