A Dynamic-Data-Driven Method for Improving the Performance of Receiver Autonomous Integrity Monitoring

Abstract

In this article, the problem of receiver autonomous integrity monitoring (RAIM) is transformed into a modeling problem using dynamic data and an artificial neural network. A new RAIM method based on a probabilistic neural network (P-RAIM) is presented to improve integrity monitoring performance. Compared with existing RAIM methods, P-RAIM has a greater ability to meet the monitoring requirements for localizer performance with vertical guidance down to altitudes of 250 feet (LPV-250) in a single global navigation satellite system. First, by projecting the pseudorange error model from the measurement domain into the positioning domain through multiconvolution, patterns including a satellite fault pattern and a faultfree pattern are obtained based on variance inflation theory. Second, the P-RAIM model is proposed as a modified dynamic-data-driven probabilistic neural network with five layers; moreover, unique methods for training sample collection and integrity support are presented. Then, particle swarm optimization is applied to optimize a fitness function based on the false alarm probability and missed detection probability thereby improving the ability of P-RAIM to meet the LPV-250 requirements, including the false alarm probability, missed detection probability, vertical alarm limit and alarm time. Finally, utilizing real satellite data from a receiver located in Beijing to verify the effectiveness and universality of P-RAIM, evaluation experiments show that both the false alarm probability and missed detection probability can be effectively reduced to meet the LPV-250 requirements when the positioning bias is no less than 40 m. Compared with least-squaresresiduals RAIM, P-RAIM can more easily detect potential faulty satellites in a single constellation.INDEX TERMS Receiver autonomous integrity monitoring, LPV-250, global navigation satellite system, multi-layer neural network, alarm systems.

Nonparametric estimation-based five-layer neural network RAIM with improved availability

“…In outlier detection, one of the most commonly utilized models is the mean shift model (MS), where the expectation of a random variable is shifted from zero to ξ bias , and another is the variance inflation model (VI). According to the variance inflation theory [20][21][22], a variable with faulty bias, denoted by µ f , can be modeled as a random variable with the same expectation but greater variance than the fault-free one. The MS and VI models are illustrated in figure 1 and formulated as follows:…”

“…Under the fault-free condition H 0 in accordance with (26), supposed that n s (⩾ 30) samples are collected from the positioning system, i.e. sampled from the position error distribution according to (5), to form a sample set B, the sample standard deviation S of this sample set B then obeys the modified normal distribution [21]:…”

“…The smoothing factor λ in NE-RAIM is determined in this section according to the false alarm probability that is estimated based on the stringent navigation requirements of LPV-250 [21]. The LPV-250 requirements include the integrity requirement I req , the continuity requirement C req , the AL, and the time-to-alert (TTA), as summarized in table 1.…”

“…These methods require considerable computing resources or introduce excessive hyperparameters to be determined in prior models. Zheng et al [21] utilized a probabilistic neural network to avoid complex integral computation of Bayesian posterior probability and then optimized the activation function to satisfy both the false alarm probability and missed detection probability. Sun [16] extracted feature vectors from separate solutions and then adopted the recurrent neural network to boost the detection performance and shorten the alert delay in tiny fault scenarios.…”

“…fault detection, fault exclusion, and a protection level algorithm. First, a classification variable based on multi-epoch information and the variance inflation hypothesis [20][21][22] is constructed, transforming the challenge of satellite fault detection into one of recognizing the classification variable's statistical models. Following that, dynamic sampling methods are proposed for fault-free and faulty patterns.…”