A convolutional autoencoder network based on multi-dimensional spatial feature reconstruction for anomaly detection of rotating machinery under complex operating conditions

Rotating machinery is critical equipment in industrial systems, and its operational status directly impacts production safety and equipment reliability. Due to complex operating conditions, continuous operation, and environmental disturbances, rotating machinery is prone to failures and abnormal states. However, in real-world engineering scenarios, abnormal samples are typically scarce and difficult to obtain, limiting the applicability of traditional label-dependent state detection methods. At the same time, existing unsupervised anomaly detection methods suffer from shortcomings in multi-scale feature extraction, channel feature utilization, and reconstruction error metrics. To address these issues, this paper proposes a Multi-Branch Parallel Attention Convolutional Autoencoder (MPACAE) for unsupervised anomaly detection in rotating machinery. This method constructs multi-scale parallel convolutional branches within the convolutional autoencoder framework to enhance the model’s ability to represent the complex features of vibration signals, and introduces a one-dimensional channel attention mechanism to adaptively reinforce key feature information. During the parameter optimization and training phase, to achieve effective reconstruction of multi-dimensional spatial features, this paper combines Euclidean distance with Pearson correlation distance to construct an EPCD loss function. This jointly constrains the model training process from the two dimensions of spatial differences and correlations, thereby improving the accuracy and stability of anomaly identification. Comparative experiments, ablation experiments, multi-operating-condition experiments, and noise-robustness experiments were conducted on multiple typical rotating machinery datasets. The results demonstrate that the proposed model can more accurately distinguish between normal and abnormal samples, exhibiting stronger anomaly detection capabilities; it also better fits the distribution of normal samples, thereby effectively achieving anomaly identification. Furthermore, under complex multi-operating-condition scenarios, this method still possesses excellent feature extraction and reconstruction capabilities and demonstrates good generalization performance, making it applicable to a wide range of rotating machinery anomaly detection tasks.