Abstract
Predicting the Remaining Useful Life (RUL) of industrial machinery is critical for proactive maintenance. However, domain shifts caused by varying operating conditions often degrade the performance of data-driven models. This paper presents Contrastive Adversarial Domain Adaptation (CADA), a framework that integrates adversarial training with contrastive learning. By leveraging an InfoNCE loss alongside a domain discriminator, CADA aligns feature distributions across domains while preserving task-relevant local structures. Evaluations on the C-MAPSS benchmark demonstrate that CADA significantly outperforms state-of-the-art methods, achieving a 21% improvement in RMSE and a 38% improvement in scoring metrics.
Introduction
In industrial Prognostics and Health Management (PHM), data-driven models are highly effective at predicting failure times when training and testing data share the same distribution. However, machines in the field operate under diverse loads, speeds, and environments—a phenomenon known as domain shift. A model trained on a controlled test cell often fails when deployed in varying real-world conditions.
While standard Domain Adaptation (DA) techniques attempt to align these distributions to achieve domain invariance, they often discard unique, target-specific sensor patterns vital for accurate RUL forecasting. CADA addresses this challenge by combining adversarial alignment with a contrastive objective to preserve the intrinsic structure of the target data.
Methodology
The CADA framework follows a structured pipeline from data preparation to cross-domain testing, as illustrated in Figure 1.
Core Architecture
The model utilizes a deep feature extractor consisting of a 5-layer Bidirectional LSTM (BiLSTM). This depth allows for the automatic extraction of representative features from raw sensory data without manual engineering. The training process involves three key stages:
- Supervised Pre-training: The source feature extractor ($E_S$) is trained on labeled source data to learn fundamental degradation patterns.
- Adversarial Domain Adaptation: A discriminator network ($D$) is employed to align the source and target feature distributions. The feature extractor learns to produce features that "fool" the discriminator, ensuring domain invariance.
- Contrastive Learning: To prevent the loss of task-specific information during alignment, the authors integrate an InfoNCE (Information Noise-Contrastive Estimation) module. This module maximizes the mutual information between raw input signals and extracted features by contrasting matching time steps against negative samples.
As shown in Figure 4, the final objective function jointly optimizes the adversarial loss and the contrastive loss, ensuring that the aligned features remain grounded in the original signal's structure.
Experimental Results
The CADA model was evaluated using the C-MAPSS benchmark dataset across 12 cross-domain scenarios involving varying operating conditions and fault modes.
Performance Metrics
CADA significantly outperformed five state-of-the-art domain adaptation methods, including CORAL and ADDA. Key results include:
- Accuracy: An average improvement of over 21% in Root Mean Square Error (RMSE).
- Reliability: A 38% improvement in the Score metric, which heavily penalizes late RUL predictions—a critical factor for industrial safety.
- Ablation Study: Removing the InfoNCE module led to a consistent drop in performance, confirming that preserving target-specific information is essential for effective adaptation.
Conclusion
CADA provides a robust solution for RUL prediction under varying operating conditions. By anchoring domain-invariant features to the original signal structure via contrastive learning, the framework achieves superior accuracy and reliability. Future work will explore domain generalization to handle entirely unseen operating environments without explicit adaptation phases.
Summary updated