A method for identifying NLOS based on the GMM-PSO-EM algorithm

Abstract

This invention discloses an NLOS identification method based on the GMM-PSO-EM algorithm. First, the time-delay data is preprocessed and denoised, and the distribution quantity is determined by combining AIC and BIC criteria. Then, the GMM parameters are iteratively updated using the PSO-SA-EM algorithm. After establishing the GMM model, the NLOS probability value within the scene is determined by the area of ​​the region determined by the PDF value. This GMM-PSO-EM algorithm exhibits excellent NLOS identification performance, demonstrating strong effectiveness and robustness in handling NLOS problems. It can solve the problems of high complexity in NLOS identification in complex indoor environments, the need for extensive data labeling, and poor performance with small data volumes. Furthermore, it is relatively simple to implement and has a faster computation speed, making it particularly suitable for identifying NLOS signals during UWB system positioning in complex indoor environments.

Description

Technical Field

[0001] This invention relates to a method for identifying non-line-of-sight (NLOS) propagating wireless signals, specifically an NLOS identification method based on the GMM-PSO-EM algorithm, belonging to the field of ultra-wideband technology. Background Technology

[0002] Positioning technology is widely used in modern life. Although there are various positioning methods, Ultra Wide Band (UWB) positioning technology has become a popular choice in indoor environments due to its advantages such as high accuracy, low power consumption, and resistance to multipath interference. UWB positioning mainly relies on ranging technology. However, in complex environments with multiple scenarios, factors such as building structures and obstacles can cause strong multipath effects during signal propagation. This means that the signal will be reflected, refracted, and scattered through multiple different paths before reaching the receiver, resulting in various problems such as signal delay spread, fading, and phase changes. At the same time, the non-line-of-sight (NLOS) propagation phenomenon between the positioning base station and the target to be located is extremely significant and a particularly prominent problem, seriously interfering with the positioning system's accurate determination of the target's location. Therefore, accurate identification of NLOS signals is crucial in UWB positioning systems. Correct NLOS identification can not only significantly improve positioning accuracy but also enhance the system's robustness, thereby adapting to various complex indoor multi-scenario environments.

[0003] In the UWB-based indoor NLOS identification problem, firstly, due to the differences in time delay and intensity between NLOS signals and LOS (Line of Sight) signals, NLOS signals are easily overwhelmed by background clutter and noise. Secondly, the time delay and intensity variations of NLOS signals are much smaller than those of LOS signals, and NLOS and LOS signals may have similar patterns and amplitudes, making it difficult to distinguish between NLOS and LOS signals using conventional time-domain or frequency-domain filtering methods. Therefore, how to effectively separate NLOS signals is a key issue of concern in the industry. Among existing NLOS identification methods, algorithms based on channel impulse response (CLS) require high computational complexity and need to be recalibrated or have their parameters adjusted in different environments; machine learning-based algorithms have high recognition rates, but require prior collection of sample data, making field applications complex; geometric feature-based methods identify NLOS conditions by analyzing the geometric characteristics of signal propagation, but are sensitive to environmental changes and may result in inaccurate identification in complex environments; deep learning and transfer learning-based algorithms require large amounts of training data and may need to be retrained in new environments to achieve better performance; statistical analysis methods identify NLOS conditions by statistically analyzing the range of results in complex indoor environments, but various physical effects may distort LOS signals along direct paths in different ways, making multi-class identification challenging. Summary of the Invention

[0004] To address the problems of the existing technologies, this invention provides an NLOS identification method based on the GMM-PSO-EM algorithm. This method can solve the problems of high complexity in NLOS identification in complex indoor environments, the need for a large amount of data labeling, and poor performance when the amount of data is small. It is also relatively simple to implement and has a faster calculation speed. It is particularly suitable for the identification of NLOS signals in the UWB system positioning process in complex indoor environments.

[0005] To achieve the above objectives, the NLOS identification method based on the GMM-PSO-EM algorithm specifically includes the following steps:

[0006] Step 1: Determine the number of Gaussian distributions in the GMM based on AIC and BIC;

[0007] Step 2: Use the PSO-SA-EM algorithm to determine the parameters of the distribution of each component of the GMM and obtain the GMM model;

[0008] Step 3: Collect latency data for LOS and NLOS scenarios;

[0009] Step 4: Fit the time delay GMM models for LOS and NLOS;

[0010] Step 5: Calculate the PDF value for each time delay data, use geometric methods to determine the probability of the current propagation being NLOS, and perform NLOS identification.

[0011] The method for calculating the NLOS probability value is as follows:

[0012]

[0013] Where: PDF NLOS and PDF LOS These are the probability density function values ​​for a given time delay under the NLOS and LOS models, respectively; Area NLOS and Area LOS These are the geometric regions obtained by squaring the marginal probabilities based on the NLOS and LOS models, respectively; TotalArea is the sum of the non-overlapping regions of NLOS and LOS; Probability NLOS It is the NLOS probability value;

[0014] If PDF NLOS and PDF LOS If the values ​​are equal, it is considered that there is no overlapping region, and the NLOS probability value is output as 50%; otherwise, it is considered that there is an overlapping region, and the TotalArea is calculated; the NLOS probability value is calculated by dividing the non-overlapping NLOS area by the TotalArea. NLOS If the probability value is greater than or equal to 50%, it is identified as NLOS; otherwise, it is identified as LOS.

[0015] Furthermore, Step 1 is detailed as follows:

[0016] Step 1-1: Introduce a histogram to visualize peak information and determine a general range of distribution quantities;

[0017] Step 1-2: Use AIC and BIC to make judgments, plot the curves of AIC and BIC within the distribution range, and smooth the curves.

[0018] Step 1-3: Select the value in the curve that minimizes the AIC or BIC value as the distribution quantity.

[0019] Furthermore, the PSO-SA-EM algorithm in Step 2 is as follows:

[0020] Step 2-1, Initialization:

[0021] ① Initialize the particle swarm size N, and randomly assign the position pBest(0) and velocity v to each particle. i (0);

[0022] ② Randomly initialize the individual optimal pBest(0) and the global optimal gBest(0) for each particle;

[0023] ③ Set the initial temperature T(0) and cooling rate α for the SA algorithm;

[0024] ④ Determine the convergence criteria and number of iterations for the EM algorithm;

[0025] Step 2-2, Determine the annular neighborhood: Assign a neighborhood to each particle i, and the leading particle x lead The index: (i+1) mod N, the lagging particle x lag The index is: (i-1+n) mod N;

[0026] Step 2-3: Determine the neighborhood optimum: Compare the individual optimal solutions of the leading particle and the lagging particle, and determine the neighborhood optimum solution nBest;

[0027] Steps 2-4, PSO-SA iterations, perform the following steps for each particle:

[0028] ① The position and velocity of the particle are updated using the PSO algorithm, as shown below:

[0029] v i (t+1)=w·v i (t)+c1·r1·(pBest i -x i (t))+c2·r2·(gBest-x i (t))

[0030] x i (t+1)=x i (t)+v i (t+1)

[0031] In the formula: v i (t) is the velocity of the i-th particle at time t; x i (t) is the position of the i-th particle at time t; w is the inertial weight used to balance the particle's exploration and development capabilities; c1 is the cognitive coefficient that affects the particle's movement toward the individual optimal position; c2 is the social coefficient that affects the particle's movement toward the global optimal position; r1 and r2 are random numbers between 0 and 1;

[0032] ② Neighborhood optimal guidance: Calculate the current particle position x i The difference Δx between (t+1) and the neighborhood best nBest is expressed as follows:

[0033] Δx=nBest-x i (t+1)

[0034] Update the particle position x based on the difference Δx and the step size λ. i (t+1)' represents the following:

[0035] x i (t+1)'=x i (t+1)+λΔx

[0036] In the formula: x i (t+1) represents the particle position before the update;

[0037] ③ Use the probabilistic acceptance criterion of the SA algorithm to decide whether to accept the new solution: Calculate the fitness of the new position and compare it with the individual best of the current particle. Then, according to the SA criterion, accept the new position with probability to decide whether to accept the new solution. The formula for calculating the SA criterion is as follows:

[0038]

[0039] In the formula: ΔE is the difference between the fitness of the new position and the current optimal fitness of the individual; T(t) is the current temperature;

[0040] ④ Update the individual optimal and global optimal;

[0041] ⑤ Update the temperature in the simulated annealing, as shown below:

[0042] T(t+1)=T(t) * α

[0043] In the formula: α is the cooling rate;

[0044] Steps 2-5: Initialize the EM algorithm by using the optimal particle position found by the PSO-SA algorithm as the initial parameters for the EM algorithm.

[0045] Steps 2-6: Execute the E and M steps of the EM algorithm until convergence.

[0046] Steps 2-7 output the final GMM parameters.

[0047] Furthermore, Step 4 is detailed below:

[0048] For a given data point x, the PDF calculation formula is as follows:

[0049]

[0050] In the formula: ω k It is the weight of the k-th component; It is a Gaussian distribution of the k-th component; μ k and σ k These are the mean and variance of each Gaussian distribution, respectively.

[0051] Based on the above formula, the number of distributions and parameters for NLOS and LOS scenarios are obtained, and the GMM model for NLOS scenario and the GMM model for LOS scenario are obtained.

[0052] Compared with existing technologies, this NLOS identification method based on the GMM-PSO-EM algorithm first preprocesses and denoises the time-delay data, and determines the distribution quantity by combining AIC and BIC criteria. Then, it iteratively updates the GMM parameters using the PSO-SA-EM algorithm, establishes the GMM model, and combines geometric methods to determine the NLOS probability value within the scene. Experimental results show that this NLOS identification method based on the GMM-PSO-EM algorithm has excellent NLOS identification performance, strong effectiveness and robustness in handling NLOS problems, and is relatively simple to implement and has a faster computation speed. It is particularly suitable for identifying NLOS signals in the UWB system positioning process in complex indoor environments. In addition, by combining the distribution models of NLOS and LOS using geometric methods, this method can output the current NLOS probability value, which can provide a new perspective and approach for improving positioning accuracy in practical applications. Attached Figure Description

[0053] Figure 1 This is a flowchart of the present invention;

[0054] Figure 2 This is a flowchart of the NLOS identification method of the present invention;

[0055] Figure 3 This is the EM algorithm fitting graph;

[0056] Figure 4 This is the fitting graph of the PSO-SA-EM algorithm of this invention;

[0057] Figure 5 This is a comparison chart of NLOS and LOS scene fitting in this invention. Detailed Implementation

[0058] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0059] like Figure 1 As shown, the NLOS identification method based on the GMM-PSO-EM algorithm specifically includes the following steps:

[0060] Step 1: Determine the number of Gaussian distributions in the Gaussian Mixture Module (GMM) based on the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC).

[0061] Step 1-1: First, in order to determine the range of Gaussian mixture distributions, a histogram is introduced to visualize the peak information and determine a general range of distributions.

[0062] Step 1-2: Next, use AIC and BIC to make a judgment. AIC and BIC are both criteria in the selection of statistical models. They are used to evaluate and compare the goodness of fit of different statistical models. Plot the curves of AIC and BIC within the range of the distribution, and then smooth the curves to reduce the impact of noise in the original data.

[0063] Steps 1-3: Finally, select the value in the curve that minimizes the AIC or BIC value as the distribution quantity.

[0064] Step 2: Improve the Expectation Maximization algorithm (EM) using Particle Swarm Optimization (PSO) and Simulated Annealing (SA) to determine the parameters of the distribution of each component of the GMM and obtain the GMM model.

[0065] Because the EM algorithm is prone to getting trapped in local optima, its convergence speed may be slow when the initial parameters are far from the optimal solution. Therefore, the PSO-SA algorithm is introduced to provide better initial parameters for the EM algorithm, thereby accelerating the convergence speed and improving the quality of the final solution. In addition, the probabilistic acceptance mechanism of SA allows the algorithm to accept suboptimal solutions with a certain probability during the search process, which helps to escape the local optimum trap, thereby improving the GMM model's ability to update parameters and its fitting effect.

[0066] The EM algorithm integrating PSO-SA is abbreviated as PSO-SA-EM, and its algorithm steps are described as follows:

[0067] Step 2-1, Initialization:

[0068] ① Initialize the particle swarm size N, and randomly assign the position pBest(0) and velocity v to each particle. i (0).

[0069] ② Randomly initialize the individual optimal pBest(0) and the global optimal gBest(0) for each particle.

[0070] ③ Set the initial temperature T(0) and cooling rate α for the SA algorithm.

[0071] ④ Determine the convergence criteria and number of iterations for the EM algorithm.

[0072] Step 2-2, Determine the annular neighborhood: Assign a neighborhood to each particle i, and the leading particle xlead The index: (i+1) mod N, the lagging particle x lag The index is (i-1+n)mod N, which determines that the particle swarm forms a closed loop, where the lagging particle of the first particle is the last particle, and the leading particle of the last particle is the first particle.

[0073] Step 2-3: Determine the neighborhood optimum: Compare the individual optimal solutions of the leading particle and the lagging particle, and determine the neighborhood optimum solution nBest.

[0074] Steps 2-4, PSO-SA iterations, perform the following steps for each particle:

[0075] ① The position and velocity of the particle are updated using the PSO algorithm, as shown below:

[0076] v i (t+1)=w·v i (t)+c1·r1·(pBest i -x i (t))+c2·r2·(gBest-x i (t))

[0077] x i (t+1)=x i (t)+v i (t+1)

[0078] In the formula: v i (t) is the velocity of the i-th particle at time t; x i (t) is the position of the i-th particle at time t; w is the inertia weight, used to balance the particle's exploration and development capabilities; c1 is the cognitive coefficient, which affects the particle's movement toward the individual optimal position; c2 is the social coefficient, which affects the particle's movement toward the global optimal position; r1 and r2 are random numbers between 0 and 1, used to simulate randomness.

[0079] ② Neighborhood optimal guidance: Calculate the current particle position x i The difference Δx between (t+1) and the neighborhood best nBest is expressed as follows:

[0080] Δx=nBest-x i (t+1)

[0081] Update the particle position x based on the difference Δx and the step size λ. i (t+1)' represents the following:

[0082] x i (t+1)'=x i (t+1)+λΔx

[0083] In the formula: x i (t+1) represents the particle position before the update.

[0084] ③ Use the probabilistic acceptance criterion of the SA algorithm to decide whether to accept the new solution: Calculate the fitness of the new position and compare it with the individual best of the current particle. Then, according to the SA criterion, accept the new position with probability to decide whether to accept the new solution. The formula for calculating the SA criterion is as follows:

[0085]

[0086] In the formula: ΔE is the difference between the fitness of the new position and the current optimal fitness of the individual; T(t) is the current temperature.

[0087] ④ Update the individual optimal and global optimal.

[0088] ⑤ Update the temperature in the simulated annealing, as shown below:

[0089] T(t+1)=T(t)*α

[0090] In the formula: α is the cooling rate.

[0091] Steps 2-5: Initialize the EM algorithm by using the optimal particle position found by the PSO-SA algorithm as the initial parameters for the EM algorithm.

[0092] Steps 2-6: Execute the E-step (calculate expectation) and M-step (maximize) of the EM algorithm until convergence.

[0093] Steps 2-7 output the final GMM parameters.

[0094] Step 3: Collect latency data for both line-of-sight and non-line-of-sight scenarios.

[0095] To collect indoor latency and other special environmental parameters, this embodiment adds a circuit board to the LD150-I UWB tag from Haoru Technology, which can complete sensor data acquisition and Wi-Fi communication. The overall data acquisition and communication integrated circuit board includes a power module, an MCU module, a DWM1000 and IMU positioning module, a sensor acquisition module, a Wi-Fi communication module, and an over-limit alarm module. The MCU used is an STM32F103C8T6 microcontroller, the gas sensor module is MQ-4, the carbon monoxide sensor module is MQ-7, the temperature sensor is DS18B20, and the Wi-Fi module is ESP8266. The function is to collect data on harmful gas concentration, temperature, latency, and location. The data storage format is TEMP, ppm_mq7, ppm_mq4, POSI_clock, POSI_distance(0, 1, 2...), POSI_accelX, POSI_accelY, POSI_accelZ, POSI_gyroX, POSI_gyroY, POSI_gyroZ, POSI_tof(0, 1, 2...), and POSI_position_(x, y, z), which respectively represent temperature, carbon monoxide concentration, methane concentration, clock, distance from each base station to the tag, three-axis velocity, three-axis acceleration, time delay from each base station to the tag, and the tag's calculated position coordinates.

[0096] Step 4: Fit the time-delay Gaussian mixture distribution model of LOS and NLOS.

[0097] The probability density function of a GMM can be expressed as a weighted sum of the probability density functions of each component. Therefore, for a given data point x, its probability density function (PDF) is calculated as follows:

[0098]

[0099] In the formula: ω k It is the weight of the k-th component; It is a Gaussian distribution of the k-th component; μ k and σ k These are the mean and variance of each Gaussian distribution, respectively.

[0100] This allows us to obtain the number of distributions and parameters for both NLOS and LOS scenarios, thus leading to the GMM model for the NLOS scenario and the GMM model for the LOS scenario.

[0101] Step 5: Calculate the PDF value for each time delay data point according to the GMM formula, determine the probability of the current propagation being NLOS using geometric methods, and perform NLOS identification, as follows:

[0102] Step 5-1: After determining the number of GMM distributions and updating the GMM parameters, construct GMM models for both NLOS and LOS scenarios. First, calculate the PDF value for each time delay data according to the GMM model formula.

[0103] Step 5-2: Estimate the probability that the current propagation is NLOS. Based on extensive experimental data, it is assumed that the probability of a random variable taking a value within a specific interval is proportional to the length of that interval. The core principle is to approximate the probability value using the area of ​​the region determined by the PDF value. The specific NLOS identification process is as follows: Figure 2 As shown.

[0104] The method for calculating the NLOS probability value is as follows:

[0105]

[0106] Where: PDF NLOS and PDF LOS These are the probability density function values ​​for a given time delay under the NLOS and LOS models, respectively; Area NLOS and Area LOS These are the geometric regions obtained by squaring the marginal probabilities based on the NLOS and LOS models, respectively; TotalArea is the sum of the non-overlapping regions of NLOS and LOS; Probability NLOS It is the NLOS probability value.

[0107] If PDF NLOS and PDF LOS If the values ​​are equal, it is considered that there is no overlapping region, and the NLOS probability value is output as 50%; otherwise, it is considered that there is an overlapping region, and the TotalArea is calculated, which is the sum of the areas of the non-overlapping regions of NLOS and LOS. The non-overlapping area of ​​NLOS divided by the total area is the calculated NLOS probability value. NLOS If the probability value is greater than or equal to 50%, it is identified as NLOS; otherwise, it is identified as LOS.

[0108] The fitting effect of the EM algorithm when the number of GMM components is 4 is as follows: Figure 3 As shown, the average residual of the PDF value for the fitting time delay is 0.01245. The fitting effect of the NLOS recognition method based on the GMM-PSO-EM algorithm is shown in the figure below. Figure 4As shown, the average residual of the fitted time delay PDF value is only 0.01056, which is about 15.16% better than the EM algorithm. Visualization images of GMM modeling for NLOS and LOS scenes are shown below. Figure 5 As shown, it is clear that the latency distribution curves are different in NLOS and LOS scenarios. In the LOS scenario, the path is direct and stable, and the latency distribution is relatively concentrated. Therefore, the probability density distribution curve of the latency is narrow. In contrast, the probability density distribution curve of the latency in the NLOS scenario is wider, indicating that the latency value is distributed over a larger range, reflecting the influence of multiple propagation paths.

[0109] To verify the accuracy of the NLOS recognition method based on the GMM-PSO-EM algorithm, the NLOS recognition method based on the GMM-PSO-EM algorithm was compared with the unmodified GMM algorithm, CNN algorithm and SVM algorithm in terms of recall, precision and F1 score. The comparison results are shown in Table 1 below.

[0110] Table 1 Performance Comparison of Different Methods

[0111]

[0112]

[0113] As shown in Table 1, the NLOS recognition method based on the GMM-PSO-EM algorithm has higher recall, accuracy, and F1 score than other algorithms. The average recall rate reaches 92.4%, the average accuracy reaches 93.7%, and the F1 score reaches 93.2%. Compared with the other three methods, the F1 score is improved by 3.78%, 2.3%, and 5.2%, respectively. This indicates that the NLOS recognition method based on the GMM-PSO-EM algorithm has better NLOS recognition accuracy because it better fits the LOS and NLOS features.

[0114] To address the issue of accurate identification of NLOS signals during UWB system positioning in complex indoor multi-scene environments, the above-mentioned experimental simulations verify that the proposed NLOS identification method based on the GMM-PSO-EM algorithm can solve the problems of high complexity in NLOS identification in complex indoor environments, the need for a large amount of data labeling, and poor performance when the amount of data is small. This results in higher NLOS identification accuracy and can effectively improve the UWB positioning accuracy in complex indoor multi-scene environments.

Claims

1. A method for identifying NLOS based on the GMM-PSO-EM algorithm, characterized in that, Specifically, the following steps are included: Step 1: Determine the number of Gaussian distributions in the GMM based on AIC and BIC; Step 2: Use the PSO-SA-EM algorithm to determine the parameters of the distribution of each component of the GMM and obtain the GMM model; Step 3: Collect latency data for LOS and NLOS scenarios; Step 4: Fit the time delay GMM models for LOS and NLOS; Step 5: Calculate the PDF value for each time delay data, use geometric methods to determine the probability of the current propagation being NLOS, and perform NLOS identification. The method for calculating the NLOS probability value is as follows: Where: PDF NLOS and PDF LOS These are the probability density function values ​​for a given time delay under the NLOS and LOS models, respectively; Area NLOS and Area LOS These are the geometric regions obtained by squaring the marginal probabilities based on the NLOS and LOS models, respectively; TotalArea is the sum of the non-overlapping regions of NLOS and LOS; Probability NLOS It is the NLOS probability value; If PDF NLOS and PDF LOS If the values ​​are equal, it is considered that there is no overlapping region, and the NLOS probability value is output as 50%; otherwise, it is considered that there is an overlapping region, and the TotalArea is calculated; the NLOS probability value is calculated by dividing the non-overlapping NLOS area by the TotalArea. NLOS If the probability value is greater than or equal to 50%, it is identified as NLOS; otherwise, it is identified as LOS.

2. The NLOS identification method based on the GMM-PSO-EM algorithm according to claim 1, characterized in that, Step 1 is as follows: Step 1-1: Introduce histograms to visualize peak information and determine a range of distribution quantities; Step 1-2: Use AIC and BIC to make judgments, plot the curves of AIC and BIC within the distribution range, and smooth the curves. Step 1-3: Select the value in the curve that minimizes the AIC or BIC value as the distribution quantity.

3. The NLOS identification method based on the GMM-PSO-EM algorithm according to claim 1, characterized in that, The PSO-SA-EM algorithm in Step 2 is as follows: Step 2-1, Initialization: ① Initialize the particle swarm size N, and randomly assign the position pBest(0) and velocity v to each particle. i (0); ② Randomly initialize the individual optimal pBest(0) and the global optimal gBest(0) for each particle; ③ Set the initial temperature T(0) and cooling rate α for the SA algorithm; ④ Determine the convergence criteria and number of iterations for the EM algorithm; Step 2-2, Determine the annular neighborhood: Assign a neighborhood to each particle i, and the leading particle x lead The index: (i+1) mod N, the lagging particle x lag The index is: (i-1+n) mod N; Step 2-3: Determine the neighborhood optimum: Compare the individual optimal solutions of the leading particle and the lagging particle, and determine the neighborhood optimum solution nBest; Steps 2-4, PSO-SA iterations, perform the following steps for each particle: ① The position and velocity of the particle are updated using the PSO algorithm, as shown below: v i (t+1)=w·v i (t)+c1·r1·(pBest i -x i (t))+c2·r2·(gBest-x i (t)) x i (t+1)=x i (t)+v i (t+1) In the formula: v i (t) is the velocity of the i-th particle at time t; x i (t) is the position of the i-th particle at time t; w is the inertial weight used to balance the particle's exploration and development capabilities; c1 is the cognitive coefficient that affects the particle's movement toward the individual optimal position; c2 is the social coefficient that affects the particle's movement toward the global optimal position; r1 and r2 are random numbers between 0 and 1; ② Neighborhood optimal guidance: Calculate the current particle position x i The difference Δx between (t+1) and the neighborhood best nBest is expressed as follows: Δx=nBest-x i (t+1) Update the particle position x based on the difference Δx and the step size λ. i (t+1)' represents the following: x i (t+1)'=x i (t+1)+λΔx In the formula: x i (t+1) represents the particle position before the update; ③ Use the probabilistic acceptance criterion of the SA algorithm to decide whether to accept the new solution: Calculate the fitness of the new position and compare it with the individual best of the current particle. Then, according to the SA criterion, accept the new position with probability to decide whether to accept the new solution. The formula for calculating the SA criterion is as follows: In the formula: ΔE is the difference between the fitness of the new position and the current optimal fitness of the individual; T(t) is the current temperature; ④ Update the individual optimal and global optimal; ⑤ Update the temperature in the simulated annealing, as shown below: T(t+1)=T(t)*α In the formula: α is the cooling rate; Steps 2-5: Initialize the EM algorithm by using the optimal particle position found by the PSO-SA algorithm as the initial parameters for the EM algorithm. Steps 2-6: Execute the E and M steps of the EM algorithm until convergence. Steps 2-7 output the final GMM parameters.

4. The NLOS identification method based on the GMM-PSO-EM algorithm according to claim 1, characterized in that, Step 4 is as follows: For a given data point x, the PDF calculation formula is as follows: In the formula: ω k It is the weight of the k-th component; It is a Gaussian distribution of the k-th component; μ k and σ k These are the mean and variance of each Gaussian distribution, respectively. Based on the above formula, the number of distributions and parameters for NLOS and LOS scenarios are obtained, and the GMM model for NLOS scenario and the GMM model for LOS scenario are obtained.

Images