Interpretable Multi-Band Spectral Learning for In-Cabin Magnetic-Field Prediction and Equivalent-Source Attribution

Magnetic fields generated by electromagnetic radiation sources in electric vehicle cabins may affect passenger exposure and the reliable operation of onboard electronics. Physical measurement campaigns are time-consuming and spatially limited, motivating rapid field-prediction methods. Based on measured frequency-domain data, this study proposes the Multi-band Equivalent-Source Attribution Network (MESA-Net) that uses the S01 (Static background) spectrum to anchor a spatial background branch and learns condition- and frequency-dependent effective source strengths for powered and dynamic operating states. This static-background-conditioned decomposition is introduced because each cabin probe records an aggregate amplitude-only signal rather than an isolated source spectrum. Five predefined cuboid equivalent-source regions are coupled with normalized distance kernels and an optional diagnostic residual through incoherent power-domain superposition. The dataset comprises 12 passenger measurement points, two near-source auxiliary points, five operating conditions, and four independently acquired bands spanning 1 Hz–400 kHz. Under leakage-free nested spatial holdout, the model achieved a mean absolute amplitude error of 3.528 dB. Learning band-dependent attenuation improved balanced multi-band performance relative to fixed attenuation. MESA-Net therefore links broadband signal prediction with model-constrained candidate-source attribution, rather than unique physical source-power recovery, and provides an engineering reference for rapid cabin-field assessment and diagnostic prioritization.