Industrial-scale component maintenance is shifting towards novel Predictive Maintenance (PdM) strategies supported by Big Data Analytics (BDA). This has resulted in an increased effort to implement Artificial Intelligence (AI) decision making into new maintenance paradigms. The transition of AI into industry faces significant challenges due to the inherent complexities of industrial operations, such as variability in components due to manufacturing, integration, dynamic operating environments and variable loading conditions. Therefore, AI in critical industrial systems requires more advanced capabilities such as robustness, scalability and verifiability. This paper presents the first Deep Learning (DL) based strategy for the classification of the State-Of-Health (SOH) of Electromagnetic Relays (EMR). The DL strategy scales with high-volumes of multivariate time-series data whilst automating labour intensive feature extraction requirements. The method proposed in our paper, combines a Convolutional-Auto-Encoder (CAE) with a Temporal Convolutional Neural Network (TCN), referred to as EMR-SOH CAE-TCN pipeline. Model uncertainty and SOH confidence bounds are approximated by Monte-Carlo dropout. Our pipeline is trained and evaluated on data generated from EMR life-cycle tests. We report a high classification accuracy and discriminatory power of the EMR-SOH classifier. The findings from our paper demonstrate the potential of AI pipelines for maintenance decision making of components in critical applications, providing a transferable AI based PdM solution that scales with large data quantities.