Hyper-redundant manipulators have a wide range of applications in narrow environments. However, due to the difficulty of obtaining the joint state with multiple degrees of freedom (DoFs), the motion of the manipulator is generally controlled by the open-loop strategy, which is susceptible to interference. To provide precise joint-level closed-loop control, this work presents an optical fiber sensor for monitoring the joint bending state. The framework of the sensor is a beam with a circular cross section, and the sensing unit comprises three identical fiber Bragg gratings (FBGs). In addition, a multifidelity bending state measurement model is established. The model takes a physics-informed neural network as the basic structure. During the modeling process, the calibration data are constrained by the sensing mechanism using the transfer learning technique. The experimental results demonstrate average measurement errors of 1.22% for the joint bending angle and 1.09% for the bending direction, confirming the effectiveness of the proposed FBG sensor and the multifidelity measurement model. However, the training results with different sample sizes demonstrate that the proposed measurement model consistently achieves the highest measurement accuracy across all sample sizes, with a smaller variation in error compared to other models. This indicates that the proposed multifidelity model exhibits superior robustness.