Dual-Domain Fusion for Semi-Supervised Learning

TL;DR

This work tackles the challenge of limited labeled data for time-series classification by introducing Dual-Domain Fusion (DDF), a model-agnostic semi-supervised framework that leverages both time-domain and time-frequency representations through a tri-model architecture. During training, separate time-domain and time-frequency classifiers collaborate with a fusion module to produce high-quality pseudo-labels, improving learning from unlabeled data, while inference remains efficient by using only the time-domain path. The approach is demonstrated on two bearing fault datasets (KAIST and SQV), where DDF achieves substantial accuracy gains (8–46%) over strong SSL baselines across varying amounts of unlabeled data and noise levels. The deployment strategy further enables cloud-based training with edge-friendly inference, making DDF particularly suitable for real-time fault diagnosis in resource-constrained environments. Overall, DDF provides a general, deployment-aware strategy to harness cross-domain information for SSL in time-series applications.

Abstract

Labeled time-series data is often expensive and difficult to obtain, making it challenging to train accurate machine learning models for real-world applications such as anomaly detection or fault diagnosis. The scarcity of labeled samples limits model generalization and leaves valuable unlabeled data underutilized. We propose Dual-Domain Fusion (DDF), a new model-agnostic semi-supervised learning (SSL) framework applicable to any time-series signal. DDF performs dual-domain training by combining the one-dimensional time-domain signals with their two-dimensional time-frequency representations and fusing them to maximize learning performance. Its tri-model architecture consists of time-domain, time-frequency, and fusion components, enabling the model to exploit complementary information across domains during training. To support practical deployment, DDF maintains the same inference cost as standard time-domain models by discarding the time-frequency and fusion branches at test time. Experimental results on two public fault diagnosis datasets demonstrate substantial accuracy improvements of 8-46% over widely used SSL methods FixMatch, MixMatch, Mean Teacher, Adversarial Training, and Self-training. These results show that DDF provides an effective and generalizable strategy for semi-supervised time-series classification.

Figures (5)

• Figure 1: The overall schematic of our proposed Dual-Domain Fusion method that combines the time and time-frequency models with a decision fusion model to generate more accurate and reliable predictions.
• Figure 2: The testing accuracies of the proposed Dual-Domain fusion and comparison techniques. Averaged results are depicted across the different SNR levels along with their corresponding 95% confidence intervals and relative gains in performance.
• Figure 3: The relationship between the number of training samples and the testing performance for our Dual-Domain Fusion and comparison methods. The results show the averaged testing accuracy $\pm$ standard deviation when utilizing the clean data in each case study.
• Figure 4: T-SNE results of the KAIST case study using clean data.
• Figure 5: T-SNE results of the SQV case study using clean data.