SEA++: Multi-Graph-Based Higher-Order Sensor Alignment for Multivariate Time-Series Unsupervised Domain Adaptation

Yucheng Wang, Yuecong Xu, Jianfei Yang, Min Wu, Xiaoli Li, Lihua Xie, Zhenghua ChenIEEE TPAMI 2024 · DOI

Abstract

Unsupervised Domain Adaptation (UDA) for multivariate time-series (MTS) data faces unique challenges due to complex inter-sensor dependencies and temporal dynamics. Standard UDA methods often rely on global feature alignment, which overlooks local sensor-level discrepancies and high-order statistical shifts. This paper presents SEA++, an advanced framework that utilizes multi-graph encoders and high-order statistics to achieve precise sensor alignment. By modeling spatial-temporal relationships and aligning both individual sensor features and their correlations, SEA++ achieves state-of-the-art performance across predictive maintenance and human activity recognition benchmarks.

Introduction

In real-world deployments, machine learning models frequently encounter domain shift—a phenomenon where the data distribution in a target environment differs from the source environment used for training. For multivariate time-series (MTS) data, such as those generated by industrial sensors or wearable devices, this shift is particularly nuanced.

Traditional Unsupervised Domain Adaptation (UDA) techniques typically treat MTS data as a single global feature vector. However, this approach is often insufficient for sensor-based systems. Each sensor in an MTS array may exhibit unique distribution shifts, and the physical relationships between sensors (e.g., the correlation between pressure and temperature) can evolve over time or vary across domains. SEA++ addresses these limitations by shifting the focus from global alignment to granular, sensor-level alignment using high-order statistics and dynamic graph structures.

Methodology

The SEA++ framework is structured around three primary architectural pillars designed to capture the complexity of MTS data.

1. Multi-branch Graph-based Encoding

To move beyond flat vector representations, SEA++ employs a Multi-branch Self-attention-based Graph Construction (MSGC) module. This component models the spatial dependencies between sensors as a graph. These graphs are processed through Message Passing Neural Networks (GNNs) and Long Short-Term Memory (LSTM) networks to simultaneously capture spatial interactions and temporal dependencies.

2. Endo-feature Alignment (Local)

Endo-feature alignment focuses on the internal characteristics of the sensor array. It consists of two sub-components:

  • Sensor Feature Alignment (SFA): Aligns the marginal distributions of individual sensor features across domains.
  • Sensor Correlation Alignment (SCA): Aligns the joint distributions (correlations) between different sensors, ensuring the underlying physical logic of the system is preserved.

3. Exo-feature Alignment (Global)

While local alignment is critical, SEA++ also maintains an exo-feature alignment module to ensure that the final global representation—the high-level summary of the entire system—remains consistent across the source and target domains.

The "++" Enhancements: High-Order Statistics and Multi-Graph Weights

SEA++ improves upon its predecessor (SEA) by incorporating high-order statistics. Rather than relying solely on first-order moments (means), SEA++ utilizes second-order moments (covariance) to capture more complex distribution shifts. Furthermore, it introduces a multi-graph alignment technique that adaptively weights different segments of a time series. This allows the model to prioritize the alignment of temporal intervals where domain discrepancy is most pronounced.

Experimental Evaluation

The researchers evaluated SEA++ on two high-dimensional benchmarks representing distinct MTS challenges:

  • C-MAPSS: A turbofan engine degradation dataset used for Remaining Useful Life (RUL) prediction. This regression task involves 14 sensors across four distinct operating conditions.
  • Opportunity HAR: A Human Activity Recognition classification dataset featuring 113 sensors (accelerometers and gyroscopes) across four different human subjects.

Key Results

SEA++ consistently outperformed both general UDA baselines (e.g., DANN) and specialized MTS-UDA methods (e.g., CoDATs, AdvSKM):

  • Predictive Accuracy: On the C-MAPSS dataset, SEA++ improved the average Root Mean Square Error (RMSE) by 11.4% compared to the best conventional methods. In specific domain transfers (e.g., Domain 1 to Domain 3), the error reduction reached nearly 10% over the nearest competitor.
  • Classification Performance: On the Opportunity HAR dataset, SEA++ achieved an average accuracy of 88.31%, a nearly 5% improvement over the leading baseline.
  • Ablation Studies: The importance of multi-level alignment was confirmed via ablation; combining sensor feature alignment and sensor correlation alignment yielded a 24.1% improvement over models using only global alignment.

Discussion and Insights

The success of SEA++ highlights several critical insights for MTS adaptation:

  1. Granularity Matters: Aligning individual "trees" (sensors) is often more effective than aligning the "forest" (global features) when dealing with localized sensor noise or drift.
  2. Relational Consistency: Preserving the correlations between sensors is as vital as preserving their individual values, as these correlations often represent the invariant physical laws governing the system.
  3. Dynamic Adaptation: Not all temporal segments are equally difficult to adapt. Adaptive weighting allows the model to focus computational resources on the most discrepant data patches.

Limitations and Future Work

Despite its performance, SEA++ introduces higher computational complexity due to the use of multi-branch graphs and second-order statistics. Future research may focus on optimizing the trade-off between alignment precision and inference speed. Additionally, exploring the integration of known physical constraints into the graph construction process and handling variable-length sequences or asynchronous sampling rates remain promising avenues for further development.

Summary updated

BibTeX
@article{wang2024sea,
  author = {Yucheng Wang and Yuecong Xu and Jianfei Yang and Min Wu and Xiaoli Li and Lihua Xie and Zhenghua Chen},
  title = {{SEA++: Multi-Graph-Based Higher-Order Sensor Alignment for Multivariate Time-Series Unsupervised Domain Adaptation}},
  journal = {IEEE Trans. Pattern Anal. Mach. Intell.},
  year = {2024},
  doi = {10.1109/tpami.2024.3444904},
}

Loading related content...

Figure