Magnet resonance imaging (MRI) using a permanent magnet array for polarization and encoding drastically simplifies the system hardware, enabling portability. The image reconstruction is model based, where a pre-measured magnetic field map is required and its accuracy is critical for imaging. However, the magnets inevitably decay and thus requires impractical regular measurements of the magnetic field map. This study propose usage of physics-guided neural network (PgNN) to selfcorrect the encoding magnetic field decay, eliminating the need for regular measurements of field maps and making model-based MRI imaging approach practical. PgNN is trained using reconstructed images and acquired signals of phantom scans with known ground truth images. The image reconstruction signal equation is integrated into PgNN as physics concepts to guide the training to generate a correction term for the decaying magnetic field. The loss function, derived from comparing the reconstructed image to the phantom ground truth, improve PgNN's performance. The corrected encoding magnetic field improves the reconstructed image's structural similarity index measure and its perspective from the human visual system. Owing to the physics guidance, the training time is below 20 minutes and only one set of training data is required. PgNN rapidly corrects the encoding field and improves the image quality without additional cost, replacing regular and repeated measurement of the encoding magnetic fields.