By collecting the magnetic field information of each spatial point, we can build a magnetic field fingerprint map. When the user is positioning, the magnetic field measured by the sensor is matched with the magnetic field fingerprint map to identify the user’s location. However, since the magnetic field is easily affected by external magnetic fields and magnetic storms, which can lead to “local temporal-spatial variation”, it is difficult to construct a stable and accurate magnetic field fingerprint map for indoor positioning. This research proposes a new magnetic indoor positioning method, which combines a magnetic sensor array composed of three magnetic sensors and a recurrent probabilistic neural network (RPNN) to realize a high-precision indoor positioning system. The magnetic sensor array can detect subtle magnetic anomalies and spatial variations to improve the stability and accuracy of magnetic field fingerprint maps, and the RPNN model is built for recognizing magnetic field fingerprint. We implement an embedded magnetic sensor array positioning system, which is evaluated in an experimental environment. Our method can reduce the noise caused by the spatial-temporal variation of the magnetic field, thus greatly improving the indoor positioning accuracy, reaching an average positioning accuracy of 0.78 m.