Strong ground motion recordings are crucial for earthquake engineering, serving as the foundation for seismic design, performance assessment of structures, and the development of building codes. However, validating these recordings presents significant challenges due to the traditionally low correlation observed among their three components. This study introduces a novel approach to assess the intrinsic compatibility of strong motion recording components using deep learning techniques. We hypothesize that despite their low correlation, components from genuine strong ground motion recordings should exhibit distinct characteristics that set them apart from non-earthquake-induced motions, and thus should be intrinsically compatible. Our research develops a convolutional neural network model trained on a comprehensive dataset of horizontal components from strong motion recordings, primarily sourced from the Japan KiK-net database. The model achieves 98% accuracy in recognizing component compatibility on the test dataset. We demonstrate that component compatibility is a fundamental feature of all genuine strong ground motion recordings and can be effectively quantified using our deep learning approach. Applying our model to existing strong motion databases, including the NGA-West2 project, we identify several problematic recordings. These findings highlight potential issues in data quality that traditional assessment methods may have overlooked. Notable discoveries include recordings with misaligned components and anomalous high-frequency content in borehole sensors, which could significantly impact structural response analyses and ground motion prediction equations. This study provides engineers with a powerful new tool for assessing the quality and reliability of strong ground motion recordings. This methodology opens new avenues for applying machine learning techniques in earthquake engineering, particularly in the crucial area of strong motion data quality assurance.