Laser Powder Bed Fusion (LPBF) is a widely used additive manufacturing process that offers high precision and design flexibility but suffers from quality inconsistencies due to variations in layer thickness. Ensuring uniform layer thickness is critical, as deviations can lead to defects such as porosity and geometric distortion. Existing inspection methods rely on optical or thermographic imaging techniques that limit spatial and temporal resolution and require supervised machine learning techniques. This study introduces a novel self-supervised machine learning approach leveraging on-axis pyrometry data to infer local layer thickness variations during LPBF. A Temporal Convolutional Network (TCN) is trained using a unique data randomization technique to handle variable-length time-series signals. The model is designed to learn representations without requiring labelled data, addressing a key challenge in real-time process monitoring. Experimental validation was conducted using a controlled LPBF setup with varying layer thicknesses. The trained model successfully classified different thickness regimes and demonstrated the ability to capture process anomalies such as short-feeding or warping. Analysis using t-distributed stochastic neighbour embedding (t-SNE) revealed well-separated clusters for distinct layer thicknesses, validating the model’s effectiveness. However, the sensor’s resolution limited discrimination below 20 µm, highlighting the need for sensor fusion strategies. Future work will focus on integrating additional data sources, such as acoustic emissions and photonic sensing, to improve resolution and extend the model’s applicability to complex geometries and scan patterns. The proposed method provides a foundation for real-time LPBF quality control, enabling adaptive process optimization and defect prevention, paving the way for industrial-scale adoption of in-situ monitoring solutions.