Anomaly Detection of Copper Melting Process Data Based on Anomaly Transformer

In this paper, aiming at the copper melting process, the relevant process variable parameters are coupled with each other, complicated and change with the production cycle, which leads to the anomaly detection of copper melting data. low accuracy of copper melting data, we propose an Anomaly Transformer model based on the Transformer model, which replaces the original Self-Attention mechanism with Anomaly-Attention mechanism, to realize accurate and reliable copper melting process data anomaly detection. In order to realize accurate and reliable anomaly detection of copper melting process data. Firstly, the degree of anomaly of copper melting process data at each time point is measured by the correlation dispersion, and the KL dispersion distance value is utilized to quantify the degree of anomaly. Distance value of KL dispersion to quantify the change of the value of each neighboring time point, and secondly, the above quantized distance value is transform the its attention mechanism into the anomaly attention mechanism form., in order to present the anomaly deviation and weight of each time point, and distinguish normal points from anomaly points, and finally, the trained anomaly Transformer model is analyzed on the test set for anomaly detection. Finally, the trained anomaly Transformer model is analyzed on the test set for anomaly detection, and the results are analyzed and interpreted using F1 evaluation indexes. The experimental results show that in the production process dataset of copper melting. The application of correlation-based discretization brings a certain 3.7% improvement in the F1 score. The addition of the anomalous attention module achieved 43.4% performance improvement, and the final combination of the combination of the combined KL dispersion and the anomaly attention module yielded an F1 score of 71.31, achieving a better anomaly detection result.