Multidomain neural process model based on source attention for industrial robot anomaly detection

Abstract

Industrial robots are vital intelligent equipment in modern industries. Periodic maintenance, which is costly and cannot prevent unexpected failures, is necessary to reduce the probability of failure and extend their service life. Therefore, this study pioneers the application of neural processes in industrial robot anomaly detection. On the basis of the attentive neural process framework, a multidomain fusion neural process (MNP) model based on source attention (SA), which introduces a multidomain path that improves the ability of the model to decouple latent distributions of observed data in industrial environments, is proposed. The multidomain path consists of the following parts: First, a time–frequency domain feature extraction module (TFDFEM) is proposed to extract rich time–frequency domain features from raw signals. Second, a dual-purpose SA module is designed to calibrate the temporal and spectral features within the signal, enabling the model to prioritize relevant features. Last, an SA-based multidomain fusion strategy (MDFS) is developed to fuse and complement features from different domains. Numerous experiments based on robots in an automotive welding and bolt fastening lines show that the MNP achieves an average accuracy of 90.8%, outperforming existing models by at least 6.2%. The average F1 is 94.7%, which outperforms existing models by 4.2%. Therefore, our model provides a promising tool for the state-based maintenance of industrial robots. The code for this project is available at https://github.com/hyh7323/Multi-domain-Neural-Process.