To enhance the demodulation accuracy of vernier spectral signals in Fabry⁃Pérot (F⁃P) sensors, this study proposes a direct deep learning⁃based demodulation method for spectral data. The method involves preprocessing spectral data to convert complex vernier spectral information into formats compatible with Convolutional Neural Network (CNN), followed by training and testing deep learning models on the processed full⁃spectrum data. The CNN architecture was employed for feature extraction and classification of spectral data, enabling accurate demodulation of target signals. Experimental validation was conducted utilizing spectral data collected from a dual⁃cavity F⁃P sensor with 112.5 nm / MPa sensitivity. The results demonstrate that the CNN model achieved an average accuracy of 92.49% with 10⁃fold cross⁃validation, accompanied by a Root Mean Square Error (RMSE) of 0.039 2 MPa and a mean relative error of 3.31%. The hybrid Convolutional Neural Network⁃Long Short Term Memory (CNN⁃LSTM) model exhibited superior performance with an average accuracy of 96.98%, an RMSE of 0.039 0 MPa, and a mean relative error of 3.28%. Notably, the CNN⁃LSTM approach attained high precision using only 256 sampled data points, demonstrating remarkable efficiency. This method provides an effective technical pathway for advancing spectral signal demodulation technology, offering significant reference value for developing intelligent optical sensing systems.