<p>Accurate and non-invasive diagnostics are crucial for the safety and reliability of lithium-ion batteries. Lithium iron phosphate (LFP) cells, though stable and cost-effective, pose challenges for conventional monitoring due to complex degradation mechanisms. This work proposes a machine learning approach that integrates electrochemical impedance spectroscopy (EIS) with a variational autoencoder (VAE) to capture degradation signatures directly from spectral data. Commercial 6.5-Ah LFP pouch cells were characterized at multiple temperatures and aging levels. Equivalent circuit modeling and distribution of relaxation times (DRT) analysis were employed to generate a synthetic training dataset. The VAE was trained to map impedance spectra into a latent representation reflecting relaxation dynamics, enabling accurate spectrum reconstruction without unstable inversion procedures. Results show precise impedance reconstruction with low error and detection of degradation through DRT peak shifts toward lower frequencies at low temperatures, and conversely toward higher frequencies at high temperatures. The proposed framework combines interpretability with computational efficiency, supporting real-time diagnostics and integration into battery management systems.</p>