Abstract
The energy consumption data of beneficiation and metallurgy enterprises has the characteristics of multi-dimensional and timing. An energy consumption data anomaly detection model based on transfer learning is proposed to help enterprises monitor the abnormal energy use. The combination of transfer learning and DTW algorithm is used to eliminate the adverse impact of timing on the model, and the ensemble learning is used to realize data detection. Finally, an efficient and accurate energy consumption data anomaly detection model is formed. The detection accuracy of the model is 91.6%, which can effectively meet the requirements of enterprise energy management.
Introduction
Beneficiation and metallurgy enterprises are among the largest industrial energy consumers. Historically, these enterprises have relied on "extensive" management styles, which lack the precision of real-time, data-driven monitoring. As the industry moves toward "digital energy management," the need for automated anomaly detection becomes critical. However, industrial energy data presents unique challenges: it is highly multi-dimensional, sensitive to temporal cycles (seasonal and shift-based), and suffers from a "small sample" problem. Unlike fields like computer vision, individual factories often lack the massive labeled datasets required to train deep learning models from scratch. This paper proposes a hybrid model that leverages transfer learning and ensemble methods to overcome these barriers.
Methodology
The researchers developed a four-stage pipeline to process and classify energy consumption data effectively.
Dimensionality Reduction and Alignment
To handle the multi-dimensional nature of the data—where factors range from ore mining volume to local weather—Principal Component Analysis (PCA) was employed. This step identifies the most influential variables, such as product output and unit power consumption, while stripping away noise. To address temporal sensitivity, Dynamic Time Warping (DTW) was used to align time-series sequences. DTW ensures that the model compares similar production cycles even if they vary in speed or time-shifts.
Transfer and Ensemble Learning
The core innovation lies in the use of Transfer Learning via a Fully Convolutional Network (FCN). Instead of training from zero, the model utilizes weights from a "source domain" (data from similar subsidiaries) and fine-tunes them for the "target" enterprise. To further enhance robustness and prevent overfitting on small datasets, a Stacking framework was implemented. This framework integrates multiple base learners, such as Random Forest (RF), AdaBoost (AB), and Support Vector Machines (SVM), to produce a final classification.
Experimental Setup
The study utilized data from a copper mine company (2018–2019) as the target dataset, including production metrics and environmental factors like temperature and wind. The source dataset consisted of energy data from other subsidiaries within the same industrial group. Anomalies were manually labeled based on internal management logs to provide a ground truth for training and validation.
Key Results
The model's performance was evaluated primarily using the F1-score to ensure a balance between precision and recall. The proposed hybrid model (referred to as SNN in the results) achieved an F1-score of 0.916 and an overall accuracy of 92%.
As shown in the table above, the hybrid approach significantly outperformed standalone traditional algorithms. For instance, Random Forest and AdaBoost achieved F1-scores of 0.846 and 0.875, respectively, while K-Nearest Neighbors lagged at 0.761. These results demonstrate that combining transfer learning with ensemble stacking provides a significant performance boost that individual models cannot achieve on limited industrial datasets.
Conclusion
This research demonstrates that transfer learning is highly viable in the metallurgy sector due to the similarity in production modes across different factories. By utilizing a relatively simple FCN structure supplemented by ensemble methods, enterprises can gain the benefits of deep learning without the need for massive "Big Data" volumes. Future work will focus on optimizing the selection of base learners and addressing the challenges of real-time deployment in live industrial environments.
Summary updated