This study proposes a state of health (SOH) estimation method for the lithium-ion batteries based on a Large Language Model (LLM) and BiLSTM model, aiming to capture the complex dynamic characteristics during the performance degradation process of lithium batteries through deep learning models. First, the charge-discharge data of lithium batteries is pr-processed, including data cleaning and normalization, to ensure the effectiveness of model inputs. Next, cross-attention is employed to reconstruct the charge-discharge data into text format, aligning data modalities with text modalities. Text prompts for this task are then input into the LLM model, with tokenization and encoding performed to obtain input encodings. Subsequently, these input encodings and the cross-attention reconstructed features are jointly fed into the LLM model for feature extraction. Finally, the features extracted by the LLM model, along with the charge-discharge data of the lithium batteries, are input into the BiLSTM model for SOH prediction. To validate the model's effectiveness, various experiments were conducted using a historical dataset of cyclic aging from commercial 21700 lithium-ion batteries (LG M50T).