Abstract: In large-scale distributed systems, such as the Internet of Vehicles (IoV), the conflict between high-frequency data sampling and limited communication bandwidth presents a critical challenge for remote state estimation. Traditional quantization methods often suffer from saturation failures when the system state diverges, as observations exceed the dynamic range of the quantizer. To address this issue, this paper proposes an Innovation-Based Encoding-Decoding Quantization (IEQ) method and a Reconstruction-Error-Based Feedback Quantization (REFQ) algorithm. Firstly, the IEQ method constructs a novel type of measurement innovation. We theoretically prove that this innovation is statistically independent of the system state and follows a stable distribution. This characteristic fundamentally eliminates quantizer saturation caused by large-amplitude observations. Secondly, to solve the problem of error accumulation in differential encoding, the REFQ algorithm introduces a feedback mechanism. The sender simulates the decoding process to calculate the reconstruction error and corrects the quantization input in real-time. Theoretical analysis demonstrates that REFQ strictly constrains the reconstruction error within a single quantization step size. Finally, an optimization-based strategy is developed to allocate quantization bits adaptively based on system matrices. Crucially, it is further established that under this bit allocation scheme, the error covariance of the remote Kalman filter remains uniformly bounded, thereby providing a rigorous theoretical guarantee for the stability of the overall estimation system. Experiments are conducted on a self-built real-world vehicle platform to emulate an IoV remote state estimation scenario. The results show that the proposed method reduces bandwidth consumption by 4 times compared to raw data transmission while maintaining estimation accuracy. Moreover, the REFQ algorithm effectively suppresses error drift, keeping reconstruction errors within a stable bound regardless of time steps, whereas traditional methods show divergent errors.