Ultra-wideband (UWB) positioning technology has been widely used in indoor positioning systems since its inception because of its signal stability and low power consumption. It provides good technical support for intelligent life interconnection, Internet of Things, industrial automation, medical tracking, supermarket assistance and other applications. However, in indoor positioning, due to the influence of obstacles such as walls, human bodies, furniture, etc., the signal propagation between the target and the base station often forms a non-line-of-sight (NLOS) situation with occlusion and a multipath (MP) situation with reflection and refraction, which causes a great distance estimation deviation and further affects the accuracy of target positioning.

In the learning process of previous research results, it is found that most scholars 'research on UWB signal recognition is often limited to the binary classification of LOS and NLOS, and MP signals are often ignored or mixed with NLOS signals. Another notable problem is that the existing methods of signal recognition using machine learning often use empirical methods to manually test and determine the parameters needed by the classifier, which consumes a lot of time, and the parameters determined in this way are often not the optimal solutions of the parameters. To solve these problems, a UWB anomaly signal recognition method based on genetic algorithm optimized random forest (GA-RF) is proposed in this paper. MP signal is regarded as an independent signal category to participate in signal recognition and genetic algorithm is introduced to optimize the parameters needed by RF classifier, thus realizing high accuracy classification of LOS, NLOS and MP signals.

In the aspect of feature selection of the training data set, considering that it is difficult for low-cost devices to obtain channel impulse response information, two types of features, i.e., measurement distance and received signal strength (RSS), are selected for classification model training. GA-RF method and machine learning classical classification model support vector machine (SVM) are compared and analyzed for UWB outlier signal recognition performance by testing and training scenes. The experimental results show that the recognition accuracy of GA-RF classifier can reach 92.52% and 79.83% respectively when the test environment is the same as the training environment and when the training environment is different. Compared with SVM, the recognition accuracy is improved by 22.2% and 8.7% respectively. Therefore, it can be considered that GA-RF method can better identify NLOS signals and multipath signals.