A Non-calibration-based Meter Reading Method for Substation Using Multi-Level Visual Model Aggregation

Existing methods for substation meter reading rely on manual calibration to obtain the meter's type, pointer position, and range information. Utilizing the prior information and neural network methods, meter values can be calculated. However, this approach relies on manual calibration data heavily, which limits its generalizability. Furthermore, the time-consuming manual calibration and configuration processes result in higher debugging expenses, which impends large-scale deployment. To overcome the challenge, this study proposes a non-calibration-based meter reading method using multi-level visual model aggregation, which is applicable to the complex environment of substations. First, a meter detection model is designed, spatial and channel attention mechanisms. This model is capable of extracting meter positions and identifying meter types from inspection images captured in substations. Subsequently, an instance segmentation model is adopted to extract meter range area information, which combines prototype mask generation and mask coefficient prediction to predict parallel sub-branches. Finally, a key point-based alignment correction model is proposed to solve the camera deviation caused by mechanical errors or environmental disturbances. It can achieve camera alignment correction and re-captures the target meter even under displacement scenarios of up to 30%. The effectiveness of the proposed method has been validated through practical applications in substations, demonstrating its robust performance in real-world scenarios.