Application of a Novel Data-Driven Framework in Anomaly Detection of Industrial Data

Abstract

In recent years, deep learning algorithms represented by the Transformer model have gained widespread recognition in the application of time series anomaly detection. Based on the Boosting algorithm, this paper proposes a novel data-driven improved Saxformer framework for multivariate time series anomaly detection. This framework enhances the Transformer model’s ability to capture local feature information and improves both the robustness to noise and the flexibility in practical applications. A method using Boosting classifiers for fine control is designed to improve the Transformer output layer. By outputting text letter by letter, the method aims to balance the model’s creativity and conservatism, suppressing significant fluctuations in predictions and thereby enhancing detection accuracy. Experimental results show that the proposed model outperforms existing mainstream algorithms on multiple evaluation metrics. The F1 score, Precision, and Recall reach 0.882, 0.946, and 0.825, respectively. Compared with the best-performing baseline model, the F1 score is enhanced by 1.4%, demonstrating high accuracy in time series anomaly detection tasks.

Keywords

• Anomaly detection
• boosting
• 1-d convolution
• Transformer
• time series data