Indoor positioning method combining base station screening strategy and neural network

Abstract

The invention discloses an indoor positioning method combining a base station screening strategy and a neural network, and the method comprises the steps: judging whether there is an NLOS base station of which the signal propagation is blocked due to the obstruction of an obstacle or not through a rapid detection algorithm link; in the base station screening algorithm link, the base stations are divided into a plurality of subsets, the estimated positions of the rss subsets of the base stations without NLOS tend to form dense clusters, the estimated positions of the RSS subsets are grouped according to the clusters, the dense clusters are detected, and the centroid positions of the dense clusters are returned to determine reliable positioning results; the NLoS base station is identified by calculating the frequencies of all the base stations appearing in the reliable result; in the step of optimizing the BP neural network distance measurement by using the crown porcupine algorithm, a CPO-BP neural network is established to construct a distance measurement model by using the advantage that the neural network can effectively capture a nonlinear relationship, so that the relationship between the RSSI and the distance in a complex environment is accurately captured, and the positioning precision is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of indoor positioning technology, and in particular to an indoor positioning method for improving positioning accuracy by constructing a neural network ranging model through base station screening and group intelligent optimization algorithm. Background Art

[0002] In indoor environments, there is a lack of line-of-sight transmission channels between satellites and indoor receivers, and existing outdoor positioning systems are difficult to provide accurate positioning services in indoor environments. Therefore, people are committed to the development of indoor positioning systems to achieve reliable and accurate indoor positioning and navigation. In fact, the WiFi indoor positioning algorithm based on RSSI has the advantages of simple implementation, low cost, and wide coverage, and has been studied by many scholars. WiFi indoor positioning methods are mainly classified into two categories: methods based on signal propagation models and methods based on fingerprint recognition. However, the construction of fingerprint databases is cumbersome and the maintenance cost is high, which limits the use of fingerprint recognition. Therefore, how to achieve RSSI ranging-based positioning based on signal models has become a hot issue at present.

[0003] The non-line-of-sight error between the base station and the test point has always hindered the development of indoor positioning technology based on RSSI ranging. To address this problem, there are currently different improvement directions, including:

[0004] 1. A WiFi RTT positioning method based on LOS compensation and credible NLOS (Non Line-Of-Sight) and NLOS identification uses the compensated LOS distance and credible NLOS distance to estimate the position.

[0005] 2. A method to measure multiple RSSI values ​​of WiFi signals to solve the problem of NLOS identification and mitigation. Using several statistical features of RSSI time series, two machine learning-based algorithms and a third hypothesis testing-based algorithm are developed and compared to identify Line-Of-Sight (LOS) and NLOS.

[0006] 3. The method based on Morlet wave transform and convolutional neural network can identify LOS / NLOS in the time-frequency domain.

[0007] 4. Use the Backward Propagation (BP) neural network to build a signal path loss model to achieve coarse positioning of the node to be located.

[0008] However, the positioning accuracy of these algorithms is not particularly high, and the ability to jump out of the local optimal solution is not obvious. In addition, they have shortcomings such as slow convergence speed and poor local search ability. In addition, they do not consider the impact of selecting different subsets of WiFi base stations on the positioning results, resulting in the presence of signals transmitted by unreliable base stations in the data set used to build the neural network, which leads to a decrease in positioning accuracy. Summary of the invention

[0009] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a positioning method that combines a base station screening strategy and a crested porcupine algorithm (CPO) to optimize a BP neural network strategy, thereby solving the problems of low positioning accuracy and long positioning time of the existing indoor positioning technology.

[0010] The present invention is implemented by the following technical solutions: an indoor positioning method combining a base station screening strategy and a neural network, comprising three steps:

[0011] S1. Rapid detection algorithm link. When there are a large number of base stations that can be used for positioning in the indoor positioning algorithm, the use of cluster-based subset sampling will make the indoor positioning cost too high. The present invention uses a fast detection algorithm to determine whether there is an NLOS base station whose signal propagation is blocked by obstacles, so as to improve the operating efficiency of the positioning algorithm.

[0012] S2. In the base station screening algorithm, if there are unreliable NLOS base stations, the base stations are first divided into several subsets. The estimated positions of the RSS subsets without NLOS base stations tend to form a dense cluster. On this basis, the present invention groups the estimated positions of the RSS subsets into clusters, detects dense clusters and returns their centroid positions to determine reliable positioning results, and identifies NLoS base stations by calculating the frequencies of all base stations that appear in the reliable results;

[0013] S3. The crested porcupine algorithm optimizes the BP neural network ranging link. The traditional indoor positioning algorithm based on RSSI ranging usually relies on empirical parameters to construct a path loss model, which is easily affected by environmental noise, resulting in large errors and inaccurate positioning. This paper takes advantage of the neural network's ability to effectively capture nonlinear relationships and establishes a CPO-BP neural network to construct a ranging model to accurately capture the relationship between RSSI and distance in complex environments, effectively improving the positioning accuracy.

[0014] S1 specifically includes:

[0015] S11. Divide the Wifi set subset into several subsets consisting of three base stations, including a subset with high-precision positioning composed entirely of LOS base stations and a subset with low positioning accuracy where NLOS base stations exist;

[0016] S12. After the RSS subsets are generated, the collected RSSI values ​​are converted into distance values ​​for each RSS subset through the path loss model, and then the subset is positioned through the trilateral positioning method;

[0017] S13. After the positioning result is obtained, the RSS subset may contain NLOS base stations, which may cause the estimated position of the same test point to be dispersed. The positioning results may be of the following two types:

[0018] a. If all positioning results are close to the actual position, then all base stations used for positioning are LOS base stations;

[0019] b. If the mutual Euclidean distance of the positioning results is greater than a certain threshold, it is considered that there is an NLOS base station and the base station screening step is performed.

[0020] S2 specifically includes:

[0021] S21. First, cluster the positioning results to find the most dense category, calculate the discreteness of the clusters through the mutual Euclidean distance between the subset positioning results of each cluster, and select the cluster with the smallest discreteness (smallest average Euclidean distance) as the reliable positioning subset;

[0022] S22. In general, most reliable results are obtained based on LoS base stations, while there are few results involving NLoS base stations. Therefore, the present invention identifies NLoS base stations by calculating the frequencies of all base stations that appear in reliable results; specifically, all base stations are divided into two clusters according to frequency: one is to select the cluster with high frequency as LoS base stations, and the other is to select the cluster as NLoS base station category, that is, the higher the frequency of a certain base station, the greater the probability of the LoS base station, and finally the NLOS base stations that are not conducive to indoor positioning are screened out based on frequency statistics.

[0023] S3 specifically includes:

[0024] S31. Database establishment stage: First, collect RSSI data sent by different anchor nodes received by the reference point, then perform Gaussian filtering on the collected RSSI data, then calculate the coordinates of the anchor node and the reference point, calculate the distance between the anchor node and the reference point by the Euclidean distance formula, and finally build a database based on the RSSI value after filtering of the reference point and its corresponding distance value;

[0025] S32. Network training phase: Set the relevant parameters of the BP neural network model and CPO, such as the number of nodes H in the hidden layer, the number of nodes N in the input layer, the number of nodes S in the output layer, the population size p of crested porcupines, etc.; randomly generate the initial positions of p crested porcupines in the range of [-0.5, 0.5], and the initial positions of crested porcupines are the initial weights and thresholds of the BP neural network; use the prediction error function of the BP neural network as the fitness function of the CPO optimization algorithm, set the initial weights and thresholds of the BP neural network, and use the RSSI value and its corresponding distance value in the database established in the offline phase as the input and output data of the BP neural network to train the BP neural network, and judge whether the current number of iterations is greater than iter m , if it is greater than iter m The iteration is terminated, and the position of the best individual in the current crested porcupine population is the optimal initial weight and threshold of the BP neural network;

[0026] S33. Network calling stage: Gaussian filtering is performed on the RSSI value received by the node to be located, and the filtered data is used as the input value of the BP neural network. The input value is input into the CPO-BP neural network that has been trained in the network training stage to obtain the predicted distance value. According to the distance value predicted by the CPO-BP neural network, the MLE method is used to calculate the estimated coordinates of the node to be located to realize the positioning function.

[0027] Different from conventional positioning methods, the indoor positioning method combining base station screening strategy and neural network proposed in the present invention improves the computing efficiency in complex positioning environments and reduces the computing cost; the established RSSI ranging model based on BP neural network suppresses the fluctuation of RSSI signal within a certain range and avoids the dependence on environmental parameters, so that the indoor positioning accuracy is higher. At the same time, the introduction of crown porcupine optimized BP neural network makes the positioning accuracy and stability of the positioning algorithm proposed in the present invention higher. Experiments show that the average positioning errors of GA-BP, BP and traditional landmarc positioning methods are 1.85m, 2.95m and 5.4m, and the average positioning error of this algorithm is 0.94m, which is better than others. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is the principle diagram of three-side positioning;

[0029] Figure 2 is the frequency of occurrence of different base stations in the reliable base station set;

[0030] Figure 3 It is a flow chart of indoor positioning;

[0031] Figure 4 are the parameters related to BP neural network;

[0032] Figure 5 is the indoor environment simulation parameter;

[0033] Figure 6 is the positioning error of different algorithms. DETAILED DESCRIPTION

[0034] The present invention will be further described below in conjunction with the embodiments and drawings.

[0035] An indoor positioning method combining a base station screening strategy and a neural network comprises the following steps:

[0036] Step 1: First, determine whether there is non-line-of-sight error in the environment to avoid meaningless positioning efficiency reduction. Therefore, a fast detection algorithm is run to determine whether there is an unreliable NLOS base station;

[0037] Specifically include:

[0038] S11: The base station set used for positioning is represented as M, including the NLoS base station set (represented as N1) and the LoS base station set (represented as N2), |N1|+|N2|=|M|, there are two phenomena:

[0039] a. If the three selected base stations are all from N2, the ranging results are accurate and the difference is not large. Selecting any three base stations in N2 can achieve good positioning results.

[0040] b. Since NLOS base stations are affected by obstacles, the random ranging interference of the subset of NLOS base stations is large, and the positioning result is inaccurate. Therefore, M is divided into multiple subsets to obtain positioning results;

[0041] S12: After obtaining the subset consisting of three base stations, the three-side positioning method is used to implement positioning research. The principle diagram of the three-side positioning is as follows: Figure 1 As shown in the figure, AN is the base station, and d is the distance from the base station to the positioning point. First, the signal strength RSSI value of the base station is converted into a distance value, and then the positioning is performed using the radius of the three base stations to finally obtain the positioning result;

[0042] S13: After obtaining the positioning results of different base station subsets, it is determined whether there is an NLOS base station. Since the existence of a subset of NLOS base stations will cause the estimated position of the test point to be dispersed, there may be two possibilities for the positioning results: a. If all positioning results are close to the actual position, then all base stations used for positioning are LOS base stations. Specifically, calculate all positioning results lu i Helu j The mutual Euclidean distance dis between (i,j) , as shown in formula (1):

[0043] dis(i,j) =||lu i -lu j ||,i,j=1,2,…,m (1)

[0044] S14. Then, take the average of all mutual Euclidean distances. If the average mutual distance is less than the threshold value γ (γ=6m is selected in the present invention), it is considered that there is no NLOS base station, and no subsequent base station screening is required;

[0045] S15. If the mutual Euclidean distance dis of the positioning results (i,j) If it is greater than the threshold value γ=6m, it is considered that there is an NLOS base station and the base station screening step is performed.

[0046] Step 2: Divide the base stations into several subsets. The estimated positions of the RSS subsets without NLOS base stations will tend to form a dense cluster. On this basis, the present invention groups the estimated positions of the RSS subsets into clusters, detects dense clusters and returns their centroid positions to determine reliable positioning results, and identifies NLoS base stations by calculating the frequencies of all base stations in the current reliable results;

[0047] Specifically include:

[0048] S21. First, you need to calculate lu again i Helu j The spatial distance between (i,j) It means, that is, formula (1);

[0049] S22. Then, calculate the local density of each point. i The local density of i , d c is the location of the clustering point, ρ i As shown in formula (2):

[0050]

[0051] in,

[0052] S23. Similarly, ρ of all subset positions can be calculated. Based on the local density, this algorithm calculates the relative minimum distance δ of each estimation result, lu i The relative minimum distance is denoted as δ i , defined as formula (4):

[0053]

[0054] Where, d ij is the distance from the location point with x=i and y=j to the cluster center;

[0055] S24. Therefore, select the point with high ρ value and δ value as the cluster center, and the point with small ρ value and high δ value as the outlier. Finally, divide the other points into corresponding clusters according to the distance from the point to the cluster center, and obtain the required positioning result of the clustering algorithm;

[0056] S25. After obtaining the positioning results, without losing generality, cluster the positioning results of the subsets using a clustering algorithm, calculate the discreteness of each cluster, and select the cluster with the smallest discreteness as the reliable positioning subset;

[0057] S26. In general, most reliable positioning results are obtained based on LoS base stations, while results involving NLoS base stations are rare. Therefore, the present invention identifies NLoS base stations by calculating the frequencies of all base stations that appear in reliable results;

[0058] S27. All base stations are divided into two clusters according to frequency: one is to select the cluster with high frequency as LoS base stations, and the other is to select the cluster with high frequency as NLoS base stations. That is, the higher the frequency of a base station, the greater the probability of the LoS base station. In this example, the set consists of 10 base stations. The frequencies of different base stations appearing are as follows: Figure 2 As shown;

[0059] S28. At this point, the NLoS base station is identified as a reliable base station for subsequent positioning through a method based on frequency statistics;

[0060] Step 3: Use the RSSI value received by the base station selected in step 2 as input, and use the good nonlinear fitting ability of the neural network to better construct the mapping relationship between the RSSI value and the distance value to achieve higher precision positioning. The flow chart is as follows: Figure 3 As shown;

[0061] Specifically include:

[0062] S31. Parameter initialization, including the crested porcupine population size P, the convergence rate alpha used to adjust the convergence speed of the algorithm, the trade-off ratio Tf between the third and fourth defense mechanisms, and the number of iterations of the algorithm iter m ;

[0063] S32. randomly generate the initial positions of p crested porcupine individuals in the range of [-0.5, 0.5], the initial positions of the crested porcupine are the initial weights and thresholds of the BP neural network, and the prediction error function of the BP neural network is used as the fitness function of the CPO optimization algorithm;

[0064] S33. Entering the exploration phase, when the predator is far away, the crested porcupine has two defense strategies, namely the visual strategy and the sound strategy. These strategies refer to investigating different areas and devoting themselves to global exploration. The formula is shown in (5):

[0065]

[0066] In the formula, For the best CP position, is the vector generated between the current crested porcupine individual and the random crested porcupine individual, and are the positions of crested porcupines before and after the update, τ1 and τ2 are random values ​​in the interval [0,1];

[0067] S34. According to the defensive behavior of crested porcupines, when predators approach, crested porcupines have two defensive strategies, namely, odor attack strategy and physical attack strategy. These strategies refer to the development of potential areas and the focus on (local) development search. In the odor attack stage, crested porcupines will secrete a foul odor that spreads in the surrounding area to prevent predators from approaching it. The formula is shown in (6):

[0068]

[0069] In the formula, is a parameter that controls the search direction, γ t is the defense factor, S i ′ is the odor diffusion factor, τ3 is a random value in the interval [0,1],

[0070] and These are the positions of crested porcupines before and after the update. and is the position of the crested porcupine with the second highest fitness in the population, Trends for crested porcupine location updates;

[0071] S35. In the physical attack phase, when the predator is very close to it, the crested porcupine will attack it with its short and thick feathers, as shown in formula (7):

[0072]

[0073] In the formula, α is the convergence speed factor, and are the positions of the crested porcupine before and after the update, τ4 and τ5 are random values ​​in the interval [0,1]. is the best solution obtained, indicating the position of the i-th individual at iteration t, is the average force of the crested porcupine affecting the i-th predator, γ t It is the defined defense factor;

[0074] S36. The defense strategies of porcupines are imitated by CPO algorithm, including visual intimidation, sound intimidation, odor attack and physical attack, so as to find the optimal parameters. The present invention uses CPO algorithm to find the weights and thresholds of the optimal neural network for indoor positioning, and then constructs a suitable ranging model;

[0075] S37. Collect RSSI data sent by different anchor nodes received by reliable LOS base stations, and then perform Gaussian filtering on the collected RSSI data. Given the coordinates of the anchor node and the reference point, calculate the distance between the anchor node and the reference point using the Euclidean distance formula, and build a database based on the RSSI value of the reference point after filtering and its corresponding distance value;

[0076] S38. Set the relevant parameters of the BP neural network model, such as the number of nodes H in the hidden layer, the number of nodes N in the input layer, the number of nodes S in the output layer, etc. The parameter setting table is as follows: Figure 4 As shown;

[0077] S39. Use the prediction error function of the BP neural network as the fitness function of the CPO optimization algorithm. After setting the initial weights and thresholds of the BP neural network, use the RSSI value and its corresponding distance value in the database established in the offline stage as the input and output data of the BP neural network to train the BP neural network. Calculate the fitness value of each individual in the crested porcupine population based on the prediction error function of the BP neural network. The smaller the fitness value, the better the result of the optimization algorithm. Select the best individual in the crested porcupine population based on the fitness value and output its position;

[0078] S40. In the example of the present invention, the effectiveness of the positioning method in an indoor NLOS environment is tested. In the simulated indoor environment, the length and width of the positioning area are set to 20m. In order to enable each node to receive RSSI smoothly, four anchor nodes are set at the four corners of the positioning area, with specific coordinates of (0,0), (0,20), (20,0), (20,20), and their parameters are set as follows: Figure 5 At the same time, a reference point is selected every 0.1m in the positioning area from (0,0) to (20,20), and then the RSSI data sent by the anchor node received by the reference point is collected and preprocessed, and the processed RSSI data is used as the training data of the BP neural network;

[0079] S41. In order to verify that the present invention can better achieve high-precision positioning in NLOS environment compared with other positioning algorithms, the present invention compares the fitting effect of the proposed algorithm with GA-BP, BP and traditional landmarc positioning algorithms, as shown in the comparison diagram Figure 6As shown in the figure, the experiment shows that the average positioning error of the present invention is 0.94m, which is lower than 1.85m, 2.95m and 5.4m of BP, GA-BP and traditional landmarc positioning algorithms, indicating that the positioning effect is better.

Claims

1. An indoor positioning method combining a base station screening strategy and a neural network, characterized in that It includes three parts: S1. Rapidly detect the algorithm link to determine whether there are NLOS base stations whose signal propagation is blocked by obstacles, so as to improve the operating efficiency of the positioning algorithm; S2. In the base station screening algorithm, the base stations are first divided into several subsets. The estimated positions of the RSS subset without NLOS base stations tend to form a dense cluster. The estimated positions of the RSS subset are grouped into clusters. The dense clusters are detected and their centroid positions are returned to determine the reliable positioning results. The NLoS base stations are identified by calculating the frequencies of all base stations that appear in the reliable results. S3. The crested porcupine algorithm optimizes the BP neural network ranging link, takes advantage of the neural network's ability to effectively capture nonlinear relationships, and establishes a CPO-BP neural network to construct a ranging model to accurately capture the relationship between RSSI and distance in complex environments, effectively improving positioning accuracy.

2. The indoor positioning method combining base station screening strategy and neural network according to claim 1 is characterized in that S1 specifically includes: S11. Divide the Wifi set subset into several subsets consisting of three base stations, including a subset with high-precision positioning composed entirely of LOS base stations and a subset with low positioning accuracy where NLOS base stations exist; S12. After the RSS subsets are generated, the collected RSSI values ​​are converted into distance values ​​for each RSS subset through the path loss model, and then the subset is positioned through the trilateral positioning method; S13. After the positioning result is obtained, the RSS subset may contain NLOS base stations, which may cause the estimated position of the same test point to be dispersed. The positioning results may be of the following two types: a. If all positioning results are close to the actual position, then all base stations used for positioning are LOS base stations; b. If the mutual Euclidean distance of the positioning results is greater than a certain threshold, it is considered that there is an NLOS base station and the base station screening step is performed.

3. The indoor positioning method combining base station screening strategy and neural network according to claim 1 is characterized in that S2 specifically includes: S21. First, cluster the positioning results to find the most dense category, calculate the cluster dispersion by the mutual Euclidean distance between the subset positioning results of each cluster, and select the cluster with the smallest dispersion as the reliable positioning subset; S22. Identify NLoS base stations by calculating the frequencies of all base stations appearing in reliable results; specifically, all base stations are divided into two clusters according to the frequencies: one is to select the cluster with high frequency as LoS base stations, and the other is to select the cluster as NLoS base station category, that is, the higher the frequency of a base station, the greater the probability of the LoS base station. Finally, NLOS base stations that are not conducive to indoor positioning are screened out based on frequency statistics.

4. The indoor positioning method combining base station screening strategy and neural network according to claim 1 is characterized in that S3 specifically includes: S31. Database establishment stage: First, collect RSSI data sent by different anchor nodes received by the reference point, then perform Gaussian filtering on the collected RSSI data, then calculate the coordinates of the anchor node and the reference point, calculate the distance between the anchor node and the reference point by the Euclidean distance formula, and finally build a database based on the RSSI value after filtering of the reference point and its corresponding distance value; S32. Network training phase: Set the relevant parameters of the BP neural network model and CPO, randomly generate the initial positions of p crested porcupine individuals in the range of [-0.5, 0.5], and the initial positions of the crested porcupines are the initial weights and thresholds of the BP neural network; use the prediction error function of the BP neural network as the fitness function of the CPO optimization algorithm, set the initial weights and thresholds of the BP neural network, and use the RSSI value and its corresponding distance value in the database established in the offline phase as the input and output data of the BP neural network to train the BP neural network, and judge whether the current number of iterations is greater than iter m , if it is greater than iter m The iteration is terminated, and the position of the best individual in the current crested porcupine population is the optimal initial weight and threshold of the BP neural network; S33. Network calling stage: Gaussian filtering is performed on the RSSI value received by the node to be located, and the filtered data is used as the input value of the BP neural network. The input value is input into the CPO-BP neural network that has been trained in the network training stage to obtain the predicted distance value. According to the distance value predicted by the CPO-BP neural network, the MLE method is used to calculate the estimated coordinates of the node to be located to realize the positioning function.

Images