Technical Summary: Temporal-Spectral Alignment with Frequency Adaptation for Source-Free Time-Series Adaptation

Problem Definition

The paper addresses the challenge of Source-Free Domain Adaptation (SFDA) for time-series data. In standard Unsupervised Domain Adaptation (UDA), models align distributions using both labeled source data and unlabeled target data. However, in many real-world scenarios (e.g., sensitive industrial or medical applications), source data cannot be accessed due to privacy regulations, security concerns, or transmission costs. Consequently, the goal is to adapt a pre-trained source model to an unlabeled target domain using only the source model weights and the target data.

Existing SFDA methods for time-series primarily focus on feature-space alignment via self-supervised mechanisms (e.g., pseudo-labeling, entropy minimization). The authors argue that these approaches largely overlook spectral distribution shifts. Time-series signals often exhibit intrinsic variations in frequency bands and amplitude distributions due to factors like sensor sensitivity changes, environmental conditions, or user behavior (e.g., stride length). Adapting solely in the time or feature domain may fail to capture these spectral discrepancies, leading to misaligned features and unreliable pseudo-labels.

Methodology: SAFA Framework

The authors propose SAFA (Temporal-Spectral Alignment with Frequency Adaptation), a framework that operates in two distinct stages. Unlike traditional SFDA methods that fine-tune the source model's feature extractor or classifier, SAFA freezes the entire source model and introduces a lightweight, trainable module to adapt the input signals in the frequency domain.

Stage 1: Source Pre-training

Before adaptation, the source model is pre-trained on labeled source data ($D_s$) using a dual-objective approach to ensure robust temporal representation:

1. Supervised Classification: Standard cross-entropy loss ($L_{ce}$) is minimized to learn discriminative features.
2. Temporal Imputation (Self-Supervised): To capture temporal dependencies, a masking strategy is applied where random time segments are masked. An auxiliary Imputer Network ($j_s$), typically an LSTM, is trained to reconstruct the original feature representation from the masked features. This minimizes a Mean Squared Error (MSE) loss ($L_{imp}$).
   The total source objective is $L_{source} = L_{ce} + L_{imp}$. The resulting feature extractor ($g_s$), classifier ($h_s$), and imputer ($j_s$) are obtained.

Stage 2: Source-Free Frequency Adaptation

In the adaptation stage, the weights of $g_s$, $h_s$, and $j_s$ are completely frozen. Adaptation is driven solely by a trainable Frequency Adaptation Layer (FAL) inserted before the feature extractor. The process involves:

1. Spectral Decomposition: Target signals ($x_t$) are transformed into the frequency domain via Fast Fourier Transform (FFT), separating them into Amplitude ($A$) and Phase ($\Phi$) components.
2. Learnable Perturbation: Two lightweight Multi-Layer Perceptrons (MLPs), initialized with zero weights, predict adaptive perturbations ($\Delta A$ and $\Delta \Phi$) for the amplitude and phase, respectively. These are scaled by hyperparameters $\alpha$ and $\beta$ and applied multiplicatively:
   • $\hat{A} = A \odot (1 + \Delta A)$
   • $\hat{\Phi} = \Phi \odot (1 + \Delta \Phi)$
3. Signal Reconstruction: The modified spectrum is transformed back to the time domain via Inverse FFT (iFFT) to produce the adapted signal $\hat{x}_t$.

Optimization Objectives

Since target labels are unavailable, the FAL is optimized using two self-supervised losses:

1. Information Maximization (IM) Loss: Composed of Entropy Minimization (encouraging confident predictions) and Diversity Maximization (preventing collapse to a single class).
2. Temporal Output Verification (TOV) Loss: This leverages the frozen imputer network $j_s$ as a verifier. The adapted features $z_t = g_s(\hat{x}_t)$ are passed through $j_s$, and the reconstruction error is minimized. This ensures the adapted signals maintain the temporal logic learned during source pre-training.

The final objective is: $L_{total} = L_{ent} - \lambda_{div} L_{div} + \lambda_{tov} L_{TOV}$.

Key Contributions

The paper highlights three primary contributions:

1. Signal-Level Adaptation Paradigm: A novel SFDA approach for time-series where adaptation occurs directly at the input signal level via frequency modulation, rather than through feature-space fine-tuning. This avoids catastrophic forgetting and optimization instability associated with updating deep networks without source data.
2. Lightweight Frequency Adaptation Layer (FAL): A parameter-efficient module that explicitly aligns target distributions with the source by modulating amplitude and phase components. It preserves the pre-trained temporal representation capabilities while correcting spectral shifts.
3. Robustness and Generalization: The framework demonstrates superior performance across diverse benchmarks, effectively handling domain shifts caused by sensor variations and user differences.

Experimental Results

The authors evaluated SAFA on three heterogeneous benchmarks: WISDM (Human Activity Recognition), MFD (Machinery Fault Diagnosis), and Boiler (Industrial Fault Detection).

• Performance: SAFA consistently achieved state-of-the-art results in Macro F1-score (MF1) across all datasets.
  • MFD: Achieved an average MF1 of 86.12%, outperforming the second-best method (CE-SFDA, 83.02%) by 3.1%. Notably, it showed high robustness in difficult transfer scenarios (e.g., 0→1 and 3→0) where frequency shifts are pronounced.
  • WISDM: Achieved an average MF1 of 65.40%, surpassing the runner-up (CE-SFDA, 60.91%) by nearly 4.5%. The method was particularly effective in tasks involving significant inter-subject variability.
  • Boiler: Achieved an average MF1 of 64.67%, outperforming the best baseline (TemSR, 57.38%) by over 7%, demonstrating scalability to complex multi-sensor industrial data.
• Ablation Studies:
  • Removing the FAL caused a drastic performance drop (average MF1 fell from 86.12% to 65.05%), confirming that time-domain alignment alone is insufficient for spectral shifts.
  • Removing the TOV module reduced performance to 78.06%, highlighting its role in stabilizing training against noisy pseudo-labels.
  • Unfreezing the feature extractor degraded performance to 79.18%, indicating that freezing the backbone prevents catastrophic forgetting.
• Parameter Sensitivity: The method showed robustness to variations in amplitude ($\alpha$) and phase ($\beta$) scaling factors, maintaining stable performance across a wide range of hyperparameters.

Significance and Claims

The paper claims that SAFA addresses a critical gap in time-series domain adaptation: the neglect of spectral distribution shifts. By explicitly modeling and adapting the frequency domain (amplitude and phase), the method effectively bridges the gap between source and target distributions without requiring access to source data.

The authors assert that their approach offers a more practical and stable solution for privacy-sensitive applications. By freezing the source model and only adapting the input signal via a lightweight layer, the framework mitigates the risks of overfitting and catastrophic forgetting inherent in methods that attempt to update deep neural networks using only unlabeled target data. The results suggest that for time-series data, where domain shifts often manifest as changes in frequency characteristics (e.g., due to sensor placement or speed), frequency-domain adaptation is a crucial mechanism for robust knowledge transfer.