基于遗传算法优化ELM神经网络的室内可见光定位系统

Indoor Visible Light Localization System Based on Genetic Algorithm-Optimized Extreme Learning Machine Neural Network

针对室内定位精度不高、稳定性差的问题,本文提出了一种基于遗传算法优化极限学习机(GA-ELM)的多LED灯室内可见光定位系统。该系统先通过遗传算法(GA)优化ELM神经网络的初始权值和阈值,使优化后的ELM网络的拓扑结构更加稳定,不易陷于局部最优;然后利用各个参考点处LED接收的光功率以及接收机的位置坐标作为指纹数据,构建指纹数据库,将指纹数据库引入到优化好的ELM网络模型中进行训练,建立定位模型。实验结果表明,在4 m×4 m×3 m的测试空间内,本文算法的平均定位误差为0.9214 cm,最大定位误差为3.9192 cm,定位精度可达毫米级。

Objective Recently, with the advancement of science and technology, public desire for indoor positioning services has grown significantly. Thus, most researchers have turned their attention to indoor positioning technology. Visible light is harmless to the human body, does not interfere with other electronic devices, and has a low effect by multipath reflection;it can be employed as the information carrier of the indoor positioning system. Thus, visible light positioning is one of the most promising indoor positioning technologies. LED lights are frequently employed in indoor positioning studies because of their high cost performance, high broadband, and long service life. With the maturation of machine learning technology, the application of machine learning algorithms to indoor visible light localization has become the focus of several scholars’ research, and good localization results have been obtained. To further enhance the accuracy and stability of indoor visible light localization, this study proposes an indoor visible light localization system based on a genetic algorithm-optimized extreme learning machine(ELM) neural network. The genetic algorithm can efficiently enhance the stability of the ELM neural network, which in turn improves the global localization accuracy.Methods First, the fingerprint database was constructed. In this research, 441 sets of data were selected as the training set and 225 sets of data were employed as the test set. Second, the training set was employed as the ELM neural network input, which was fed into the neural network for training. Since the ELM neural network was prone to local optimum and instability, its weights and thresholds were sought out using a genetic algorithm during training, and the optimal weights and thresholds were found and assigned to the ELM neural network after selection, crossover, and variation operations. Then, the test set was sent into the trained neural network, which predicted coordinate points. Finally, the error between the actual and predicted locations was computed to examine the localization performance of the system.Results and Discussions The average localization errors of the receiver were 1.39, 2.23, 3.75, and 6.64 cm at the test heights of 0.2, 0.4, 0.6, and 0.8 m, respectively, and the maximum localization errors were 6.86, 11.04, 16.41, and 24.11 cm, respectively. As the height of the receiver increases, the channel gain decreases because of the increasing emission angle of the LEDs as well as the reception angle of the receiver, which causes the optical signal to fade and decreases the optical power received by the receiver, thereby decreasing localization accuracy. In the experimental situation, the indoor localization system based on a genetic algorithm-optimized ELM neural network(GA-ELM) achieved an average localization error of 0.9214 cm and a maximum localization error of 3.9192 cm. Compared with the findings obtained from the indoor positioning system based on the ELM neural network, which show an improvement of 86.1% on the average localization error and 70.16% on the maximum localization error, the average localization error of the GA-ELM positioning algorithm reached the millimeter level. Furthermore, this research compares the algorithm with the BP neural network, support vector machine(SVM), and GA-BP algorithms. The maximum and average localization errors of the proposed algorithm are significantly smaller than those of the other three algorithms. Finally, the average localization time of the GA-ELM algorithm is compared with the other three algorithms to illustrate the timeliness of the GA-ELM algorithm, and the average localization time needed by the GA-ELM algorithm is 0.04235 s. Compared with the SVM, BP, and GA-BP algorithms, the localization time of the proposed algorithm is shortened dramatically, indicating that the proposed algorithm is better in terms of timeliness. Conclusions In this research, a multi-LED light localization system with a genetic algorithm-optimized ELM neural network is employed to obtain indoor high-precision localization. The received optical power value of PD is measured through the direct line-of-sight link to establish the fingerprint database in the offline stage. In the online stage, the fingerprint data are introduced into the optimized ELM neural network to complete the localization. In this study, the performance simulation and experiments of the localization algorithm are performed in a space of 4 m×4 m×3 m. The simulation findings reveal that the average localization error of the localization algorithm employed in this study increases as the receiver height increases, and the experimental findings reveal that the average localization error of the localization algorithm employed in this research is 0.9214 cm and the maximum localization error is 3.9192 cm, which is comparable to ELM, GA-BP, SVM, and BP, improves the localization accuracy, and has broader range of application than ELM, GA-BP, SVM, and BP. From the domestic and international investigation studies over the past 1 or 2 years, there are substantial enhancements in localization accuracy using various localization algorithms in various environments, which are highly applicable. The investigation in this paper is at a high level in terms of localization accuracy, and a significant improvement is observed in the stability of the system.