Adaptive Mobile Location Estimator with NLOS Mitigation using Fuzzy Inference Scheme

 This paper proposes a fuzzy line of sight (LOS)/non-line of sight (NLOS) smoother based on an adaptive Kalman filter, which can be used for mobile location estimation with the time of arrival (TOA) measurement data in cellular networks to meet the Federal Communications Commission (FCC) requirement for phase . A fuzzy inference scheme is used by the proposed location estimator to detect LOS condition, NLOS condition or LOS/NLOS transition condition and to estimate the noise covariance via fuzzy interpolation. With an accurate estimation of noise covariance, an adaptive Kalman filter is proposed for the range estimation between the base station (BS) and mobile station (MS). Therefore, the proposed mobile location estimator can efficiently mitigate the NLOS effects of the simulated measurement range error even changing condition between LOS and NLOS. Simulation results demonstrate that the performance of the proposed fuzzy LOS/NLOS smoother is improved significantly over the FCC target in both fixed LOS/NLOS and LOS/NLOS transition condition, and outplays other location estimators employing the Kalman filter and NLOS mitigation techniques. . INTRODUCTION Mobile position location has attracted considerable interest in recent research. The mobile location service was driven by FCC, which mandated the wireless networks to provide the location for emergency calls. The Enhanced 911 demands the location accuracy requirement for phase as 100 m for 67% of time and 300 m for 95% of time for network-based location systems [1]. The most accurate and popular position location system is the Global Positioning System (GPS), but it is currently not a viable option for solving the Enhanced 911 location problem due to the expense and technical challenge of replacement of all existing cellular handsets with GPS receivers. Therefore, we limit discussion in this study to those techniques which can be used with existing handsets. Recently, several methods have been employed to estimate the mobile location, signal strength [2], time difference of arrival (TDOA) [3], time of arrival (TOA) [4], and angle of arrival (AOA) in cellular networks. Hearability and NLOS problems are the two biggest challenges for accurate mobile location estimation. When a mobile station is near the serving BS, it must reduce its power to avoid causing interference to other users. However, a too weak transmitted power may not be received by three or more nearby BSs to detect and estimate the mobile location. To solve this problem, the Power Up Function (PUF) was recommended for the current cellular network system IS-95B [4]. Furthermore, in an NLOS condition, the range measurement data was analyzed in [5], and the mean and the standard deviation of range measurement error were in the order of 513 m and 436 m, respectively. NLOS mitigation techniques [6, 7] are necessary for canceling the effects of NLOS measurement error. The Kalman filter was employed in [7] for mobile range measurement smoothing and NLOS effects mitigation. In a realistic condition, the communication channel between the corresponding BS and MS is often changed when the mobile moves randomly with different velocities and accelerations. The LOS/NLOS transition condition will cause a serious measurement error for the range estimation, because the covariance matrices of the measurement noise employed by the corresponding Kalman filter are not adaptively adjusted to match the true covariance variation in the LOS and NLOS cases. In this paper, we propose a fuzzy LOS/NLOS smoother-based adaptive Kalman filter, which can be used for mobile location estimation with the TOA measurement data in cellular networks to meet the FCC requirement for phase A moving estimation window is employed to calculate the real covariance of the simulated measurement data for real-time identification of LOS/NLOS conditions. A fuzzy logic system is introduced to effectively describe the possible changes among LOS, NLOS and transition condition in cellular networks. Therefore, the fuzzy inference system can be used to mitigate the NLOS effects by a more accurate covariance estimate. With the accurate total noise covariance estimated by the proposed fuzzy inference scheme through fuzzy interpolation, we can employ an adaptive Kalman filter with total noise covariance to smoothly estimate the range distance between the corresponding BS and MS. Simulation results demonstrate that performance of the proposed fuzzy LOS/NLOS smoother is further over the FCC target of both fixed LOS/NLOS and LOS/NLOS transition conditions. . SYSTEM MODEL AND THE PROPOSED MOBILE LOCATION ESTIMATOR ARCHITECTURE The architecture of the proposed mobile location estimator is illustrated in Figure 1, the raw TOA measurement data at consecutive time samples obtained from different BSs are first calculated to identify any NLOS condition by a moving estimation window. The proposed adaptive identification of LOS/NLOS condition techniques instead of repeatedly checking in [7] is expected to get a better accuracy of location estimation in the complicated mobile environment. According to [6], when the NLOS condition exists, the measurement error covariance is significantly greater than the case of LOS. In the proposed location estimator, a fuzzy inference scheme is employed to mitigate the NLOS effects by a more accurate covariance estimation through fuzzy interpolation method. From the accurately estimated covariance by fuzzy inference scheme, we can apply an adaptive Kalman filter with the total noise covariance update every iteration to smoothly estimate the range distance between the corresponding BS and MS. Finally, these estimated range distances between the three BSs and MS can be calculated to obtain the MS location estimation. A. System Model Assume that there are K BSs to detect the range signal from the MS. The simulated range measurement corresponding to TOA data between BSk and MS at time tn can be modeled as [6, 7] K k t NLOS t n t d t r n k n k n k n k L , 1 ), ( ) ( ) ( ) ( = + + = (1) where ) ( n k t d is the true range between the corresponding BSk and MS, nk(tn) is the measurement noise modeled as a zero mean white Gaussian noise N(0,σm), and NLOSk(tn) is the NLOS measurement error of the kth BS at time sample tn, the NLOS measurement error can be modeled as a positive mean white Gaussian noise N(mNLOS,σNLOS) [5, 6]. In an LOS case, the BS range measurement is only corrupted by the system measurement noise nk(tn). In this situation, there is no NLOS measurement error and NLOSk(tn) can be set to zero. Otherwise, in an NLOS case, the range measurement is corrupted by two sources of errors, nk(tn) and NLOSk(tn). Measurements taken by Nokia [5] confirm that the NLOS problem dominates the measurement error in the range estimation for the mobile location. We can further transfer the range measurement model in (1) to a corresponding discrete time sequence as K k n w n b m n d n r k k NLOS k k k L , 1 ), ( ) ( ) ( ) ( _ = + + = (2) where mk_NLOS is the NLOS range offset between the corresponding BSk and MS, wk(n) is the normalized zero mean white Gaussian noise N(0,1), and bk(n) is the standard deviation of the total range measurement error for both LOS and NLOS effects between the corresponding BSk and MS. So the mk_NLOS and bk(n) can be defined respectively as :   = condition NLOS if m condition LOS if m NLOS NLOS k , , 0 _   + = condition NLOS if condition LOS if n b NLOS m m k , , ) ( 2 2 σ σ σ (3) B. Moving Window for LOS/NLOS Identification For the purpose of LOS/NLOS identification, we employ a moving estimation window with size M to make a real-time calculation for the standard deviation k m _ σ̂ of the range measurement error for the BSk. This means that the last M samples of rk(n) are used to calculate its rough covariance of the range measurement error. The window size is chosen empirically to make a better statistical smoothing performance. The rough standard deviation of the range measurement error can be calculated by K k for n r j r M n n M n j k k k m L , 1 , )) ( ) ( ( 1 ) ( ˆ 1 2 _ = − = ∑ + − = σ (4) where ) (n rk is the window average mean of ) ( j rk from j = n – M +1 to j = n and can be expressed as