ELM-Assisted Particle Filter for INS/UWB-Integrated Quadrotor Positioning

At present, with the wide application of the quadrotor, accurate positioning has become an increasingly important problem and needs to be considered. In this study, the inertial navigation system (INS) and ultra-wideband (UWB) technology are used together to collect the ight data of the quadrotor, and the particle lter (PF) algorithm will be employed to fuse the data information of the two sensors for reducing the error and obtaining the movement path of the quadrotor in the three dimensional (3D) space. Meanwhile, for making PF work more accurate, the extreme learning machine (ELM) is adopted to map the equation of state in the ltering process to make position information more reliable. In addition, ELM can also establish a new signal through mapping when UWB’s signal is interrupted, so that the whole system can work normally. To verify the eectiveness of the proposed method, a real experiment was carried out. ­e experimental results show that the ELM-assisted PF strategy has a good eect on INS/UWB-integrated navigation quadrotor positioning. When UWB signals are normal, compared with the single PF, the ELM-assisted PF is able to improve positioning accuracy by about 18.44%. When UWB’s signal is interrupted, compared with the least square support vector machine(LS-SVM-) assisted PF, the ELM-assisted PF could improve positioning accuracy by about 1.15m. On the whole, the proposed design algorithm not only improves the positioning accuracy but also can predict UWB’s signal when it is interrupted and thus make the lter work normally.


Introduction
Quadrotor aircraft can adjust its attitude in the space by changing the speed. Because of the development of modern control technology, quadrotor aircraft has been more and more widely concerned. e stable ight of the quadrotor can be used in aerial photography, rescue and search, security, re-ghting, pesticide spraying, construction, and other high-risk operating environments [1,2]. erefore, how to control the moving path and space position of the quadrotor, that is, how to achieve accurate positioning of the quadrotor, is a much more important subject that needs to be considered [3][4][5][6][7][8][9]. e inertial navigation system measures the acceleration of the target in the reference frame and obtains velocity, heading angle, and position information by integrating it twice over time. e inertial navigation system is based on Newtonian mechanics, which belongs to the way of calculation navigation. It has good independence and continuity, does not depend on external information, nor does it radiate energy outward, and has low requirements for the working environment. However, due to the lack of reference, the solution position of the inertial navigation system a ects each other, and the error will gradually increase with time accumulation; therefore, INS is not suitable for a long period, long distance navigation, and positioning. e UWB technology obtains the position information by collecting distance information between the reference node and the motion carrier, and the positions are independent from each other, so there is no time error accumulation problem.
However, the signal transmission of the reference node is affected by the external environment, so there is the possibility of occlusion, deviation, or interruption. erefore, it is difficult for a single navigation technology to meet the requirements of accurate positioning. For avoiding that, INS and UWB-integrated technology was proposed. Such multisensor fusion manner can give full play to their respective advantages, eliminate the error accumulation problem of INS, and ensure the continuity of signal data [10][11][12][13][14].
Measurement data from the sensor are subjects to various kinds of noise, such as measurement environment, measuring technology, or hardware itself. Using the appropriate filtering algorithm can suppress the influence of bad noise, improve the signal quality, and make the positioning more accurate and reliable, so as to improve the performance of the whole navigation system. e particle filter algorithm is an effective sampling method, and it approximates the density function by calculating and comparing a random sample. In this method, the sample mean replaced the integral operation.
e Monte Carlo method is its central idea, and probability is expressed by the particle set, which is widely used in practice [15,16]. In the positioning process, in order to obtain more stable and superior effects, we can also use some learning strategies to match with filtering algorithms, such as the BP neural network, ELM, and LS-SVM. e extreme learning machine was proposed with the development of computer network technology. It was presented aiming to improve the backward propagation (BP) algorithm to improve low learning efficiency and simplify the setting of learning parameters. ELM is based on the feedforward neural network and improves the fuzzy neural network and thence has good adaptability in the supervised learning problem. In some current views, ELM has advantages in speed and generalization [17][18][19][20].
In this study, the extreme learning machine will be adopted to assist the particle filter algorithm to calibrate the INS/UWB-integrated navigation system mounted on a quadrotor. In this process, the extreme learning machine plays two roles. ELM establishes the relationship between each state value in the filtering process, so as to make the whole model more stable. By establishing the mapping relationship between INS and UWB, even UWB's signal is interrupted, and the necessary data needed for filtering can be provided to make the system work normally and improve the robustness of the whole system. In order to prove the availability of the proposed strategy, the experiment has been carried out in a real environment, and the experimental results prove the effectiveness of the proposed strategy.
e main contributions of this study are as follows: (i) A new PF algorithm based on the extreme learning machine is proposed to implement the fusion of INS and UWB's data, and the mapping relationship of state values in the filtering process is built by training of the extreme learning machine, which improves the reliability of the positioning system. (ii) e mapping between INS and UWB can be built by training of the extreme learning machine, and the signal compensation can be carried out when UWB's signal is interrupted, so that the whole system can work normally.
e remaining construction of this study is outlined as follows. e ELM-assisted PF strategy is explained in "INS/ UWB-Integrated Navigation Quadrotor Positioning Scheme," the PF filtering strategy for INS/UWB positioning and the ELM-assisted strategy are given in the "Particle Filtering Algorithm" and " e Scheme of ELM," and the real test and error analysis are completed in "Test." Finally, "Conclusion" describes the effectiveness of the proposed method for INS/UWB-integrated navigation quadrotor positioning.

INS/UWB-Integrated Navigation Quadrotor
Positioning Scheme e INS/UWB-integrated navigation quadrotor positioning scheme will be discussed for the following two situations. In the filtering process, ELM establishes the mapping of the state deviation Δx t , t ∈ [1, ∞) generated at each moment of PF and correct Δx by the more accurate estimate Δx generated by ELM to reduce the state deviation. Meanwhile, the state vector x t|t−1 , t ∈ [1, +∞) output from PF and δPo are taken as input and targets of ELM, and ELM keeps training to establish the mapping relationship between them. e whole system works normally, and ELM plays two roles: (1) to constantly correct Δx t along with PF, and ELM outputs deviation information; (2) through training, the mapping relationship between the Δx t output by PF and δPo is constantly established. At this time, ELM works offline and does not output observation information.
e positioning strategy is shown in Figure 1.

Particle Filtering Algorithm
On the basis of INS/UWB integrated navigation, we use the particle filter algorithm to fuse sensor information. e state equation of particle filter is as follows: 2 Mathematical Problems in Engineering In (1), t is the time index, ϕ t , δV n t , and δP n t denote the errors of the INS's attitude, velocity, and position, (∇ b t , ε b t ) represent the accelerometers bias and gyroscope drift employ. ΔT is the sample time, and ω t ∼ N(0, Q t ) represent the noise of the system. In (2), (f n Ut , f n Et , f n Nt ) is the acceleration up, east, and north of the coordinate system. In (3), C n b is the rotation matrix, and θ, c, ψ represent the rotation angle around x, y, and z axis. e observation equation of the particle filter is as follows.

Mathematical Problems in Engineering
In the equation above, Y t is the observation vector of PF, δ(d (i) t ) 2 is the square of the difference between the distances collected by INS and UWB, (δx (i)

The Scheme of ELM
In the navigation positioning process, we use the PF algorithm to fuse the positioning information of the two sensors and reduce the bad noise and error. ELM is used to assist the PF algorithm. ELM has two functions here. One is to establish the mapping relationship between various state values in the PF filtering process to optimize the accuracy. e other is when UWB's measurement being interrupted; the mapping relationship between the state value and the observation value is established to compensate the signal, so that the filter can work normally and enhance the stability of the system.

Scheme 1: Correction of State Deviation in the Filtering
Process. In the particle filter state update process, the state value x state describes the state information of the motion quadrotor, while the state update of the filter x update represents the predicted value of the filter, and the particle filter outputs the filtering results by comparing the weight relation. e state value x state and state update x update are two different descriptions of the target state, and there are deviations between them, given by where t is the time index. Since the filtering effect will always be affected by the state value, so before the prediction stage, the state deviation Δx should be optimized. According to equation (1) and the real motion model, it can be known that Δx is affected by the previous moment at all moments, that is to say, the present Δx t is affected by Δx at all previous moments. e (5) is expressed as where f represents the relational function between them. It should be emphasized that f is unknown, so we use ELM to build this function. In the training model, the quantity of Δx keeps increasing with the change of time, and the training speed of ELM will decline rapidly. We usually set a training scale n to limit the number to the right of (6), for ensuring that the number of states affected by the previous state is the same. So, (6) can be written as When t > � n + 2, the training model is shown in Figure 3.
In the filtering process, when time index t > � n + 2, ELM collects training data (Δx t−1−n , . . . . . . , Δx t−2 ), (Δx t−n , . . . . . . , Δx t−1 ) that are generated by PF as input and output to establish the mapping relationship among them. en, (Δx t−n , . . . . . . , Δx t−1 ) as the input of ELM and predict (Δx t−n+1 , . . . . . . , Δx t ) through mapping established in the previous stage. e output state deviation array contains Δx t needed by PF state update at the next moment. Δx t is corrected by ELM. After entering PF, x update is calculated by (5) to obtain more accurate estimation information. e process is then repeated until filtering is complete. us, the state value of PF is constantly updated by ELM, and the current state value is jointly determined by the previous multiple state values, so the influence of noise is reduced and the positioning accuracy is improved.

Scheme 2: Replenish Measurement Information
When UWB Interrupted. Complicated or hostile indoor environment may lead to the interruption of UWB's signal, and the Training data observation vector δPo of PF will not be available. Before signal interruption, UWB provides observation data, and ELM is used to establish the mapping between δPo and state value x t|t−1 , t ∈ [1, +∞) output by PF. As in the previous section, this functional relationship is difficult to describe with exact mathematical equations. We use x t|t−1 as the input training data of ELM and δPo as the training target, ELM through a certain amount of data training to establish the mapping relationship between them. In this way, in the absence of UWB's signal, ELM can predict δPo through the x t|t−1 and mapping relationship, so that the filter can work normally and ensure the stability of the system. It should be noted here that among the training data used, the x t|t−1 output by the filter is obtained through the assistance of ELM, so it is more accurate than the simple PF.
ELM provides the observation data-assisted PF process when UWB measurement is interrupted as shown in Figure 4.
During the training stage, when UWB measurement Po (U) is available, PF provides x t|t−1 . e observation vector t is the training target and ELM establishes offline mapping through training. In the prediction stage, Po (U) of UWB is not available at this time, ELM predicts δPo by input x t|t−1 and mapping relationship instead of Y t , and PF works normally by using δPo provided by ELM.

Test
In order to verify the effectiveness of the proposed strategy, a real experiment has been designed. e availability of the proposed strategy is proved by results of the simulation and error analysis.

Experimental Settings.
e experimental site was selected in an empty space, and (x, y, z) represents the actual directions east, north, and up. e quadrotor is equipped with the inertial measurement unit (IMU) and UWB blind node (UWB BN) to provide reference for distance information. e quadrotor used in the experiment is shown in Figure 5. In order to better position the quadrotor, six UWB    Mathematical Problems in Engineering reference nodes (UWB RNs) were placed at known coordinate points. ree were placed high and three were placed low (height 2.7 m and 0.1 m, horizontal distances between each UWB RN were 2.3 m and 4.6 m, as shown in yellow, i � 1 − 6). e UWB RN form a 3D space in which the target quadrotor moves along a planned path. e upper computer collects INS and UWB's data at an interval of ΔT � 0.02S. e actual test scene is shown in Figure 6.

Localization Errors.
In this part, the effectiveness of the proposed strategy is verified by simulation calculation: Scheme 1: ELM-assisted PF can effectively reduce positioning error Scheme 2: when UWB's signal is interrupted, ELM compensates the signal through the mapping relationship to make the filter work normally

Error Analysis of Scheme 1: Correction of State
Deviation. For scheme 1, the position information Po (I) of INS and Po (U) of UWB can be calculated, respectively, through the data collected by the upper computer, and the flight trajectory in 3D space can be fitted out. e ELMassisted PF algorithm mentioned in scheme 1 is used to process the data to reduce the influence of bad noise and improve the positioning accuracy. Here, an individual PF algorithm is also used to prove the superiority of the proposed strategy by comparison. e trajectory comparison diagram is shown in Figure 7. e error reduction of the ELM-assisted PF model with that of individual PF was compared. e root mean square error (RMSE) has been used in the x, y, and z directions at all time indexes, which was calculated as   After filtering, RMSE of UWB, PF, and ELM-assisted PF were calculated, and the results are shown in Figure 8 and Table 1.
In Table 1, mean represents the average value of the errors in the three directions. It can be seen that comparing with UWB, the PF algorithm error is reduced by 10.88% and the PF + ELM algorithm error is reduced by 27.31%. Compared with the PF algorithm, PF + ELM algorithm error is reduced by 18.44%. It should be emphasized that UWB positioning and PF are quite exact.
Otherwise, the cumulative distribution function (CDF) is also used to describe the achievement rate of each algorithm. e comparison of CDF is shown in Figure 9. As shown in the figure, the PF + ELM strategy achieves the ideal state faster than PF alone, indicating that for the INS/UWB combined navigation model, scheme 1 has a better effect on reducing noise and optimizing accuracy.

Error Analysis of Scheme 2: Replenish Measurement
Information. For scheme 2, after the position information Po (I) of INS and Po (U) of UWB are fitted, four interrupt regions (#1, #2, #3, #4) were set to simulate the interruption of UWB's signal, as shown in Figure 10.
To demonstrate the performance of the proposed strategy, LS-SVM was also used. In the outage area, ELM and LS-SVM perform signal compensation, respectively, as shown in Figure 11. Meanwhile, RMSEs of four areas are calculated, respectively, according to (8), as shown in Figure 12 and Table 2.
As shown in Figure 11, the red line represents the outage areas of the reference trajectory, the green line represents the Obviously, the difference between PF + ELM and the reference value is smaller, proving that compared with LS-SVM, the proposed strategy is more effective. According to Table 2, the error of PF + ELM is reduced by 1.15 m compared with LS-SVM.

Conclusion
In this study, ELM-assisted PF for INS/UWB combined navigation quadrotor positioning strategy is studied. In this scheme, INS and UWB acquire the position information of the motion quadrotor, respectively, PF performs data fusion and filtering, and ELM is used to assist the filter. ELM establishes the mapping between the difference deviation of each state in the filtering process to correct state deviation. Meanwhile, during the normal operation of PF, ELM keeps training and establishing the relationship between the state vector output by PF and the observation vector. When UWB's signal is interrupted, the observation vector fails; under this condition, ELM makes signal prediction and provides observation signal to make the filter work normally. e experimental results show that the proposed ELM-

Data Availability
e data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare that they have no conflicts of interest.