Adopting Machine Learning to GNSS Positioning on MediaTek P60 Platform

Abstract

In this article, we are going to discuss how machine learning can be adopted to MediaTek P60 platform in order to increase the capability to against the harsh enviroment, such as dense urban and urban canyon. With the mechanism of machine learning, the receiver will have better chance to identify the abnormality of measurement. These additional information can be provided to the traditional Kalman Filter (KF) based algorithm a better opportunity to select better measurement and mitigate the effect of abnormal measurement. Furthermore, the training result of machine learning might not be exactly the same due to the difference between devices. We are also exploring for the possibility to combine with AI (Artificial Intelligence) computation unit on mobile device in order to achieve the customized learning along with difference devices design. In the most common traditional design, the Kalman Filter (KF) has been widely applied to GNSS positioning receiver in the past decades. The filter plays the rule of smoothing, estimating and predicting in a sense of least-square methodology in order to minimize the sum of squared residuals ( a residuals being: the difference between a measured value and the fitted valued provided by a model). With the assumption of Linear Time-Invariant (LTI) system with Gaussian noises, the result obtaining from KF can be considered as an optimized one. In practical, the KF structure is simple and robust nature of recursive algorithm which is best suited for estimation on the go, instead of implementing a batch-filter when you are processing each measurement whenever it is sampled. This makes the KF one of the best methods to perform estimation on the fly. Also, even though it is rare that the optimal model exist in real-life applications, the filter often works well and appealing for resolving GNSS navigation and data fusion with sensor based measurements in automotive-control problems. As a matter of fact, even though the KF smoothing is also used extensively to correct statistical models of substantial uncertain variables when receivers operating under harsh working environments and tracking the received signal despite of being affected by time-variant signal fading or blockage. The major pursue of GNSS positioning/tracking algorithms is to provide reliable position fixes with optimal design of improved algorithm with special emphasis in their application on mobile terminals, wearable devices and the moving vehicles. The measurements used for fix are often referred to as high-sensitivity GNSS receivers, and the fact of keeping track of the received signal, as signal tracking which are typically sudden fading events with more than tens dB of signal attenuation in urban canyons and with blockage of the line-of-sight signals in moving portable devices. Moreover, in favor of incorporating with manufactures of mass-market GNSS capable devices, simple design on RF components and smaller or more crowded assembly on PCB are becoming important in wearable or capable devices. More wireless interference and in device jamming sources are aggravated when hardware implementation constraints need to be satisfied. In state of the art, the use of extended filter mitigates the measurement noise and allows the tracking operation under these harsh conditions. However, when people need more accurate and robust positioning results, lots of fine-tuned parameters and factor-functional design should be considered. The aim of this work will propose a synthetic machine learning design of a Kalman filter based architecture and evaluate the values for robust and high-sensitivity tracking in a GNSS capable smartphone. The expected results will make the algorithm designers easier to optimize the parameters for platform dependent varieties and make the navigation fix more accurate.