Anomaly detection of power consumption in yarn spinning using transfer learning

Anomaly detection of spinning power consumption is crucial for energy saving in yarn manufacturing. Data based methods are widely adopted for anomaly detection in industry due to historical power consumption data can be easily obtained today. However, data cannot be collected sufficiently and representatively in a short time when we study a newly-built yarn spinning workshop. Transfer learning has become an effective approach because the data and knowledge of old spinning workshops with rich power consumption records can be utilized. However, the abnormal patterns in a new yarn spinning workshop may be not exactly the same as in an old one, in most cases, there are less anomaly patterns in a new one. This pattern mismatch results in underutilization of knowledge of the data-rich workshop, and makes a transfer learning model less effective. In this paper, we propose a Cluster-based Deep Adaptation Network (CDAN) model to improve the efficiency of transfer learning for spinning power consumption anomaly detection. A cluster-based adaptation layer is inserted between the feature layers of source and target networks. It is designed specially to reduce the mismatch of transfer learning. The proposed CDAN model was applied in a real case study: a yarn spinning workshop in Xinjiang, China. With effective consideration of the mismatch in transfer learning, experimental results show that the proposed method can detect the anomaly of spinning power consumption compared with higher accuracy than state-of-the-art methods.

Automated summary

Anomaly detection of spinning power consumption is crucial for energy saving in yarn manufacturing. Data based methods are widely adopted for anomaly detection in industry due to historical power consumption data can be easily obtained today. Transfer learning has become an effective approach in this context, however, there may be a pattern mismatch when applying it to new yarn spinning workshops. The Cluster-based Deep Adaptation Network (CDAN) model is proposed in this research to address this challenge and improve the accuracy of anomaly detection for spinning power consumption.