This paper proposes a deep learning-based demodulation method for all-fiber Mach-Zehnder interferometer (MZI). By leveraging convolutional neural network (CNNs) to establish a nonlinear mapping model between transmission spectral curvature and applied curvature, breaking through the dynamic range limitation caused by spectral saturation in traditional demodulation methods. Two intensity-modulated fiber sensors (two-path type and in-line type) are fabricated, demonstrating original curvature demodulation ranges of 0.05747?0.10449 m^-1 and 0.02612?0.49106 m?1, respectively. To achieve extended dynamic range demodulation, a Gramian angular field (GAF) encoding technique is introduced to transform one-dimensional spectral signals into two-dimensional images. The CNN regression architecture implements a neural network structure with progressively decreasing neuron counts in fully connected layers, replacing conventional classification output layers to establish nonlinear spectral-curvature mapping. Experimental results demonstrate that under constant maximum sensitivity conditions, both sensor types achieve expanded curvature measurement range of 0?1.5672 m?1, has been increased to 33 times and 3 times the original level, respectively. Validation across four network architectures (ResNet, GoogleNet, etc.) confirm the universality of this proposed method in overcoming traditional spectral limitations in fiber sensing, establishing a novel methodological framework for deep learning-enhanced fiber optic sensing technology.