High-precision anti-interference distance measurement method based on UWB

Abstract

The invention discloses a UWB-based high-precision anti-interference distance measurement method, which belongs to the technical field of distance measurement, is used for anti-interference distance measurement, and comprises the following steps: obtaining a distance observation value between a tag and a base station by using a bilateral bidirectional distance measurement method, reading a channel pulse response signal in ultra-wideband distance measurement, obtaining an optimal variational mode decomposition parameter by using particle swarm optimization, and obtaining a distance measurement result by using the optimal variational mode decomposition parameter. An intrinsic mode function is reconstructed, decomposed and reconstructed, and dimension rising is carried out through the GASF and the GASF; and calculating a statistical parameter sequence of the CIR, inputting the statistical parameter sequence into the improved MobileNetV3 in combination with the GASF to obtain a non-line-of-sight noise factor, and performing improved Kalman filtering to obtain a distance measurement result. According to the invention, the average pooling retains the average information of the features, and draws the overall contour; the maximum pooling highlights prominent features and captures key details; the recognition rate reaches 78.0%, the calculated distance root mean square error RMSE is within 7cm, and the precision requirement of practical application is met.

Description

Technical Field

[0001] The invention discloses a high-precision anti-interference ranging method based on UWB, belonging to the technical field of ranging. Background Art

[0002] Ultra-wideband (UWB) wireless communication is a technology that uses nanosecond short pulses for wireless communication. It uses the time-of-flight (TOF) algorithm to calculate the distance between the base station and the tag, which can achieve high-precision and low-latency positioning. Compared with satellite positioning, ultra-wideband is mostly used in the field of indoor positioning, such as indoor positioning and navigation of large urban complexes, personnel positioning in smart factories, UWB anti-lost devices, etc. Multipath effect and non-line-of-sight (NLOS) propagation are the main factors affecting the ranging accuracy. In the prior art, Kalman filtering and its variants are used to solve the multipath effect and non-line-of-sight propagation errors, but the noise of the measured data does not belong to Gaussian noise. Although the variance of the data can be reduced after Kalman filtering, due to non-line-of-sight propagation, the data still has an error with a positive mean, and its value is related to the environment. In the prior art, neural networks are also used to first identify the non-line-of-sight environment and then eliminate the error. How to improve the accuracy of neural network recognition and reduce the number of parameters to ensure real-time performance is also a difficult problem in this field. In addition, there is also the problem of untimely correction of NLOS errors due to the different time resolutions of NLOS recognition and ranging algorithms. Summary of the invention

[0003] The purpose of the present invention is to provide a high-precision anti-interference ranging method based on UWB to solve the problem of untimely correction of NLOS error during ranging in the prior art.

[0004] A high-precision anti-interference ranging method based on UWB, including using a bilateral two-way ranging method DS-TWR to obtain a distance observation value between a tag and a base station, reading a channel impulse response signal CIR in ultra-wideband UWB ranging, using particle swarm optimization to obtain optimal variational mode decomposition VMD parameters, reconstructing decomposition and reconstruction of intrinsic mode functions DR-IMFs, and upgrading the dimension through Gram angle and field GASF; calculating a statistical parameter sequence of CIR, combining GASF input to improve MobileNetV3, obtaining a non-line-of-sight NLOS noise factor, and performing an improved Kalman filter to obtain a ranging result.

[0005] The method of using particle swarm optimization to obtain the best variational mode decomposition VMD parameters includes initializing the particle swarm, calculating individual fitness, obtaining the optimal value of the group, and judging whether the permutation entropy is the minimum. If not, the particles are updated and the individual fitness is recalculated. If so, the best variational mode decomposition VMD parameters are obtained. The best variational mode decomposition VMD parameters include a penalty factor and the modal decomposition number K.

[0006] The reconstruction decomposition and reconstruction of the intrinsic mode function DR-IMFs include judging the value of K. If K is not greater than 3, then set K=3 to obtain the Gram angle and field GASF. If K is greater than 3, perform permutation entropy and center frequency clustering, perform k-means clustering of the intrinsic mode function IMFs sequence, and retain the clustering method with a small sum of dynamic time warping DTW.

[0007] The improved Kalman filter includes initialization Status at all times , Moment covariance matrix , Noise Variance and 、NLOS environmental noise error , combined with the NLOS noise factor prediction a priori NLOS noise error estimate ,predict Prior state estimation at time and Time prior error cosquare matrix , calculate the Kalman factor and attenuation factor , update the posterior NLOS noise error estimate to obtain , combined with the distance observation between the tag and the base station, update the posterior state estimate to obtain the updated time and The posterior covariance matrix at time , update the residual mean , output As the ranging result.

[0008] When K is greater than 3, the intrinsic mode functions IMFs after VMD are clustered and reconstructed to synthesize three DR-IMFs; The entropy of the array and center frequency clustering As the basis for cluster analysis, k-means clustering of IMFs series was performed; The k-means clustering includes dividing the number of clusters into 3, calculating the distance between each data point and the center of each cluster in the space and updating and iterating until the cluster center no longer changes, and clustering all IMFs again based on the distance between the IMFs sequences. The sum of the DTW from the IMFs in each cluster to the central IMFs of the two k-means clusterings is calculated, and the clustering method with the smallest DTW sum is retained. The IMFs in each cluster are added together to obtain the final DR-IMFs.

[0009] Permutation Entropy Greater than threshold The modal function is used for discrete wavelet denoising, the soft threshold is selected for wavelet denoising, and the Mallat algorithm is used to accelerate the wavelet transform; DR-IMFs and CIR are upgraded to 300*300 grayscale images through GASF. First, all sequences are normalized and scaled to Convert the normalized sequence to polar coordinates: ; In the formula, is the normalized values, It is The polar coordinates of the values; ; In the formula, is the GASF matrix, It is The polar coordinates of the value.

[0010] The statistical parameter sequence of CIR includes the kurtosis, maximum amplitude, average excess delay, root mean square delay spread, average energy, strong path energy, variance and mean of CIR. The first input of the improved MobileNetV3 is 4 300*300 grayscale images, and the second input is a 1*8 one-dimensional sequence. The 4 300*300 grayscale images correspond to the grayscale images of 3 DR-IMFs and CIR after dimensionality upgrade through GASF. The 1*8 one-dimensional sequence corresponds to the statistical parameter sequence of CIR. The first input enters 11 serial MobileNet convolution blocks to generate 4 feature maps. The 4 feature maps are fused with the first input to form an overall feature map. The overall feature map output generates three branches. The first branch performs minimum and maximum normalization. The second branch performs a softmax activation function and then combines with the third branch to enter 11 serial MobileNet convolution blocks. The result is then fused with the result of the first branch. The result is divided into two branches, one for maximum pooling and the other for average pooling. The second input is linearly encoded and fused with the two branches to obtain the non-line-of-sight NLOS noise factor.

[0011] The MobileNet convolution block includes input, 1×1 convolution + h-swish activation function block, 3×3 depth-separable convolution + h-swish activation function block, feature fusion layer, 1×1 convolution, feature splicing layer and output. There is a shortcut connection directly to the feature splicing layer after the input. The 3×3 depth-separable convolution + h-swish activation function block has branches, which are connected to the pooling layer, fully connected + Relu activation function + fully connected + sigma activation function layer in sequence, and finally connected to the feature fusion layer.

[0012] The line-of-sight LOS noise factor is introduced, and then an improved Kalman filter is performed, and the distance observation value is used as the observation value , update the posterior state estimate posterior covariance matrix, update the residual mean, and finally output the ranging result.

[0013] In the improved Kalman filter: ; ; ; In the formula, , is a known matrix, for Moment control matrix; ; ; In the formula, is the non-line-of-sight NLOS noise factor, when the current signal propagation path is NLOS ,at this time = , Pick a random number between 15 and 30, in cm; ; In the formula, is the observation matrix; ; ; In the formula, and Respectively Moment and time; ; ; ; ; In the formula, is through The calculated value.

[0014] Compared with the prior art, the present invention has the following beneficial effects: the present invention accurately and efficiently extracts individual features, and its lightweight characteristics effectively reduce the computational complexity, further mining the common features contained in the data; the normalization operation eliminates the differences in different feature dimensions and value ranges, making the features at the same scale, providing a fair benchmark for subsequent processing; average pooling retains the average information of the features and outlines the overall contour; maximum pooling highlights prominent features and captures key details; the recognition rate of the present invention reaches 78.0%, and the calculated distance root mean square error (RMSE) is within 7 cm, meeting the accuracy requirements of practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is the technical flow chart of the first part of the present invention; Figure 2 This is the technical flow chart of the second part of the present invention; Figure 3 It is the CIR waveform of the line-of-sight channel LOS; Figure 4 It is the CIR waveform of the non-line-of-sight channel NLOS; Figure 5 It is the GASF-encoded map of high-frequency DR-IMFs; Figure 6 It is the GASF-encoded image of intermediate frequency DR-IMFs; Figure 7 It is the GASF-encoded image of low-frequency DR-IMFs; Figure 8 It is the original CIR signal encoded by GASF; Fig. 9 To improve the MobileNetV3 structure diagram; Fig.10 This is the MobileNet convolution block structure diagram. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] A high-precision anti-interference ranging method based on UWB, including using a bilateral two-way ranging method DS-TWR to obtain a distance observation value between a tag and a base station, reading a channel impulse response signal CIR in ultra-wideband UWB ranging, using particle swarm optimization to obtain optimal variational mode decomposition VMD parameters, reconstructing decomposition and reconstruction of intrinsic mode functions DR-IMFs, and upgrading the dimension through Gram angle and field GASF; calculating a statistical parameter sequence of CIR, combining GASF input to improve MobileNetV3, obtaining a non-line-of-sight NLOS noise factor, and performing an improved Kalman filter to obtain a ranging result.

[0018] The method of using particle swarm optimization to obtain the best variational mode decomposition VMD parameters includes initializing the particle swarm, calculating individual fitness, obtaining the optimal value of the group, and judging whether the permutation entropy is the minimum. If not, the particles are updated and the individual fitness is recalculated. If so, the best variational mode decomposition VMD parameters are obtained. The best variational mode decomposition VMD parameters include a penalty factor and the modal decomposition number K.

[0019] The reconstruction decomposition and reconstruction of the intrinsic mode function DR-IMFs include judging the value of K. If K is not greater than 3, then set K=3 to obtain the Gram angle and field GASF. If K is greater than 3, perform permutation entropy and center frequency clustering, perform k-means clustering of the intrinsic mode function IMFs sequence, and retain the clustering method with a small sum of dynamic time warping DTW.

[0020] The improved Kalman filter includes initialization Status at all times , Moment covariance matrix , Noise Variance and 、NLOS environmental noise error , combined with the NLOS noise factor prediction a priori NLOS noise error estimate ,predict Prior state estimation at time and Time prior error cosquare matrix , calculate the Kalman factor and attenuation factor , update the posterior NLOS noise error estimate to obtain , combined with the distance observation between the tag and the base station, update the posterior state estimate to obtain the updated time and The posterior covariance matrix at time , update the residual mean , output As the ranging result.

[0021] When K is greater than 3, the intrinsic mode functions IMFs after VMD are clustered and reconstructed to synthesize three DR-IMFs; The entropy of the array and center frequency clustering As the basis for cluster analysis, k-means clustering of IMFs series was performed; The k-means clustering includes dividing the number of clusters into 3, calculating the distance between each data point and the center of each cluster in the space and updating and iterating until the cluster center no longer changes, and clustering all IMFs again based on the distance between the IMFs sequences. The sum of the DTW from the IMFs in each cluster to the central IMFs of the two k-means clusterings is calculated, and the clustering method with the smallest DTW sum is retained. The IMFs in each cluster are added together to obtain the final DR-IMFs.

[0022] Permutation Entropy Greater than threshold The modal function is used for discrete wavelet denoising, the soft threshold is selected for wavelet denoising, and the Mallat algorithm is used to accelerate the wavelet transform; DR-IMFs and CIR are upgraded to 300*300 grayscale images through GASF. First, all sequences are normalized and scaled to Convert the normalized sequence to polar coordinates: ; In the formula, is the normalized values, It is The polar coordinates of the values; ; In the formula, is the GASF matrix, It is The polar coordinates of the value.

[0023] The statistical parameter sequence of CIR includes the kurtosis, maximum amplitude, average excess delay, root mean square delay spread, average energy, strong path energy, variance and mean of CIR. The first input of the improved MobileNetV3 is 4 300*300 grayscale images, and the second input is a 1*8 one-dimensional sequence. The 4 300*300 grayscale images correspond to the grayscale images of 3 DR-IMFs and CIR after dimensionality upgrade through GASF. The 1*8 one-dimensional sequence corresponds to the statistical parameter sequence of CIR. The first input enters 11 serial MobileNet convolution blocks to generate 4 feature maps. The 4 feature maps are fused with the first input to form an overall feature map. The overall feature map output generates three branches. The first branch performs minimum and maximum normalization. The second branch performs a softmax activation function and then combines with the third branch to enter 11 serial MobileNet convolution blocks. The result is then fused with the result of the first branch. The result is divided into two branches, one for maximum pooling and the other for average pooling. The second input is linearly encoded and fused with the two branches to obtain the non-line-of-sight NLOS noise factor.

[0024] The MobileNet convolution block includes input, 1×1 convolution + h-swish activation function block, 3×3 depth-separable convolution + h-swish activation function block, feature fusion layer, 1×1 convolution, feature splicing layer and output. There is a shortcut connection directly to the feature splicing layer after the input. The 3×3 depth-separable convolution + h-swish activation function block has branches, which are connected to the pooling layer, fully connected + Relu activation function + fully connected + sigma activation function layer in sequence, and finally connected to the feature fusion layer.

[0025] The line-of-sight LOS noise factor is introduced, and then an improved Kalman filter is performed, and the distance observation value is used as the observation value , update the posterior state estimate posterior covariance matrix, update the residual mean, and finally output the ranging result.

[0026] In the improved Kalman filter: ; ; ; In the formula, , is a known matrix, for Moment control matrix; ; ; In the formula, is the non-line-of-sight NLOS noise factor, when the current signal propagation path is NLOS ,at this time = , Pick a random number between 15 and 30, in cm; ; In the formula, is the observation matrix; ; ; In the formula, and Respectively Moment and time; ; ; ; ; In the formula, is through The calculated value.

[0027] The technical process of the present invention consists of two parts, respectively as follows: Figure 1 and Figure 2 As shown, it includes using the bilateral two-way ranging method DS-TWR to obtain the distance observation value between the tag and the base station, reading the channel impulse response signal CIR in the ultra-wideband UWB ranging, using particle swarm optimization to obtain the best variational mode decomposition VMD parameters, reconstructing the decomposition and reconstruction of the intrinsic mode function DR-IMFs, and upgrading the dimension through the Gram angle and field GASF; calculating the statistical parameter sequence of CIR, combining GASF input to improve MobileNetV3, obtaining the non-line-of-sight NLOS noise factor, and performing improved Kalman filtering to obtain the ranging result. The method of using particle swarm optimization to obtain the best variational mode decomposition VMD parameters includes initializing the particle swarm, calculating the individual fitness, obtaining the optimal value of the group, and judging whether the permutation entropy is the smallest. If not, the particles are updated and the individual fitness is recalculated. If so, the best variational mode decomposition VMD parameters are obtained. The best variational mode decomposition VMD parameters include a penalty factor and the modal decomposition number K. Reconstructing the intrinsic mode function DR-IMFs includes determining the value of K. If K is not greater than 3, set K=3 to obtain the Gram angle and field GASF. If K is greater than 3, perform permutation entropy and center frequency clustering, perform k-means clustering of the IMFs sequence, and retain the clustering method with a small DTW sum. The improved Kalman filter includes initializing Status at all times , Moment covariance matrix , Noise Variance and 、NLOS environmental noise error , combined with the NLOS noise factor prediction a priori NLOS noise error estimate ,predict Prior state estimation at time and Time prior error cosquare matrix , calculate the Kalman factor and attenuation factor , update the posterior NLOS noise error estimate to obtain , combined with the distance observation between the tag and the base station, update the posterior state estimate to obtain the updated time and The posterior covariance matrix at time , update the residual mean , output As the ranging result.

[0028] The CIR invalid data segment is regarded as environmental noise. In the present invention, the 650th to 950th data are intercepted as valid data segments. The CIR valid data segment contains characteristic information for distinguishing NLOS (non-line-of-sight) and LOS (line-of-sight) channels. The CIR waveform of the line-of-sight channel LOS is as follows: Figure 3 As shown in the figure, the CIR signal under the LOS channel is characterized by the peak amplitude of the first path; the CIR waveform of the non-line-of-sight channel NLOS is as follows Figure 4 As shown in the figure, the CIR in the NLOS channel is characterized by multiple peaks, and the first path is not the point where the peak is the largest. The statistical characteristics of the CIR signal can also be used as a basis for distinguishing NLOS from LOS, such as kurtosis, maximum amplitude, average excess delay, root mean square delay spread, average energy, strong path energy, variance, and mean. In addition, the CIR signal also contains abstract classification information, which requires the use of a neural network for classification.

[0029] The present invention calculates 8 statistical parameters of 42,000 CIR signals in the eWINE open source data set and standardizes the statistical parameters of each CIR signal. The GASF matrix images of the three DR-IMFs and the original CIR signal are drawn. The high-frequency DR-IMFs are GASF-encoded as shown in the figure. Figure 5 As shown in the figure, the intermediate frequency DR-IMFs are encoded by GASF. Figure 6 As shown in the figure, the low-frequency DR-IMFs are encoded by GASF. Figure 7 As shown in the figure, the original CIR signal is encoded by GASF. Figure 8 shown.

[0030] Improve MobileNetV3 such as Fig. 9As shown, the first input of the improved MobileNetV3 is 4 300*300 grayscale images, and the second input is a 1*8 one-dimensional sequence. The 4 300*300 grayscale images correspond to the grayscale images of 3 DR-IMFs and CIR after dimensionality upgrade through GASF, and the 1*8 one-dimensional sequence corresponds to the statistical parameter sequence of CIR; The first input enters 11 serial MobileNet convolution blocks to generate 4 feature maps. The 4 feature maps are fused with the first input to form an overall feature map. The overall feature map output generates three branches. The first branch performs minimum and maximum normalization. The second branch performs a softmax activation function and then combines with the third branch to enter 11 serial MobileNet convolution blocks. The result is then fused with the result of the first branch. The result is divided into two branches, one for maximum pooling and the other for average pooling. The second input is linearly encoded and fused with the two branches to obtain the non-line-of-sight NLOS noise factor.

[0031] The MobileNet convolutional block structure is as follows Fig.10 As shown in the figure, the MobileNet convolution block includes input, 1×1 convolution + h-swish activation function block, 3×3 depth-separable convolution + h-swish activation function block, feature fusion layer, 1×1 convolution, feature concatenation layer and output. After the input, there is a shortcut connection directly to the feature concatenation layer. The 3×3 depth-separable convolution + h-swish activation function block is equipped with branches, which are connected to the pooling layer, fully connected + Relu activation function + fully connected + sigma activation function layer in sequence, and finally connected to the feature fusion layer.

[0032] The indoor ranging system experiment is set up in the laboratory corridor. The experimental system consists of two development boards equipped with DW1000 and STM32 chips. The two development boards serve as a base station and a tag for ranging, respectively, and are 1.5m above the ground. In addition, the experimental system also includes a personal computer. The development board as a tag communicates with the serial port of the personal computer through WiFi transparent transmission. The training and testing of the NLOS recognition model mentioned in the present invention are based on an open source data set, which contains 7 different indoor scenes and a total of 42,000 sets of CIR data. The amount of data is sufficient to illustrate various situations in the actual application of UWB. Cross-validation is used to divide the training set and the test set into 4:1. The recognition accuracy of the model on this data set reaches 78.0%. At the same time, the recognition accuracy and model parameter quantity of other classic models under the same data set are also tested, as shown in Table 1.

[0033] Table 1. Comparison of model accuracy and parameter quantity ; Compared with MobileNetV3, the recognition accuracy of the model of the present invention is improved by 3.9%. Compared with ResNet-34, the number of parameters is reduced by 50.46% while the recognition rate is almost the same. This means that the model of the present invention has lower computing power requirements and is easy to deploy on edge computing devices.

[0034] In order to test the ranging effect of the system, multiple comparative tests were set up. The human body was used as an occlusion to simulate the NLOS environment. The tag and the base station were fixed on a 1.5m high bracket. The floor tiles (600mm*600mm) were used as a reference. The bracket was placed on the top of the tile and kept still. The tag and the base station calculated the observation value through the DS-TWR algorithm. , and the observations are labeled The CIR is sent to the host computer for processing. Since the ranging error is small in the LOS environment, this experiment only uses the data under the NLOS condition to illustrate the experimental effect. The placement distance between the base station and the tag is changed to 180cm, 240cm, and 300cm, with a total of 501 data in each group, and the average value, root mean square error, and variance of the algorithm of the present invention and the Kalman filter output are compared. The results of 180cm are shown in Table 2, the results of 240cm are shown in Table 3, and the results of 300cm are shown in Table 4.

[0035] Table 2. Comparison of distance measurement effect at 180cm ; Table 3. Comparison of distance measurement effect at 240cm ; Table 4. Comparison of distance measurement effect at 300cm .

[0036] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features may be replaced by equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-precision anti-interference ranging method based on UWB, characterized in that: It includes using the bilateral two-way ranging method DS-TWR to obtain the distance observation value between the tag and the base station, reading the channel impulse response signal CIR in the ultra-wideband UWB ranging, using particle swarm optimization to obtain the optimal variational mode decomposition VMD parameters, reconstructing the decomposition and reconstruction of the intrinsic mode function DR-IMFs, and upgrading the dimension through the Gram angle and field GASF; calculating the statistical parameter sequence of CIR, combining GASF input to improve MobileNetV3, obtaining the non-line-of-sight NLOS noise factor, and performing improved Kalman filtering to obtain the ranging result.

2. The high-precision anti-interference ranging method based on UWB according to claim 1, characterized in that: The method of using particle swarm optimization to obtain the best variational mode decomposition VMD parameters includes initializing the particle swarm, calculating individual fitness, obtaining the optimal value of the group, and judging whether the permutation entropy is the minimum. If not, the particles are updated and the individual fitness is recalculated. If so, the best variational mode decomposition VMD parameters are obtained. The best variational mode decomposition VMD parameters include a penalty factor and the modal decomposition number K.

3. The high-precision anti-interference ranging method based on UWB according to claim 2 is characterized in that: The reconstruction decomposition and reconstruction of the intrinsic mode function DR-IMFs include judging the value of K. If K is not greater than 3, then set K=3 to obtain the Gram angle and field GASF. If K is greater than 3, perform permutation entropy and center frequency clustering, perform k-means clustering of the intrinsic mode function IMFs sequence, and retain the clustering method with a small sum of dynamic time warping DTW.

4. The high-precision anti-interference ranging method based on UWB according to claim 3 is characterized in that: The improved Kalman filter includes initialization Status at all times , Moment covariance matrix , Noise Variance and 、NLOS environmental noise error , combined with the NLOS noise factor prediction a priori NLOS noise error estimate ,predict Prior state estimation at time and Time prior error cosquare matrix , calculate the Kalman factor and attenuation factor , update the posterior NLOS noise error estimate to obtain , combined with the distance observation between the tag and the base station, update the posterior state estimate to obtain the updated time and The posterior covariance matrix at time , update the residual mean , output As the ranging result.

5. The high-precision anti-interference ranging method based on UWB according to claim 4, characterized in that: When K is greater than 3, the intrinsic mode functions IMFs after VMD are clustered and reconstructed to synthesize three DR-IMFs; The entropy of the array and center frequency clustering As the basis for cluster analysis, k-means clustering of IMFs series was performed; The k-means clustering includes dividing the number of clusters into 3, calculating the distance between each data point and the center of each cluster in the space and updating and iterating until the cluster center no longer changes, and clustering all IMFs again based on the distance between the IMFs sequences. The sum of the DTW from the IMFs in each cluster to the central IMFs of the two k-means clusterings is calculated, and the clustering method with the smallest DTW sum is retained. The IMFs in each cluster are added together to obtain the final DR-IMFs.

6. The high-precision anti-interference ranging method based on UWB according to claim 5, characterized in that: Permutation Entropy Greater than threshold The modal function is used for discrete wavelet denoising, the soft threshold is selected for wavelet denoising, and the Mallat algorithm is used to accelerate the wavelet transform; DR-IMFs and CIR are upgraded to 300*300 grayscale images through GASF. First, all sequences are normalized and scaled to Convert the normalized sequence to polar coordinates: ； In the formula, is the normalized values, It is The polar coordinates of the values; ； In the formula, is the GASF matrix, It is The polar coordinates of the value.

7. The high-precision anti-interference ranging method based on UWB according to claim 6, characterized in that: The statistical parameter sequence of CIR includes the kurtosis, maximum amplitude, average excess delay, root mean square delay spread, average energy, strong path energy, variance and mean of CIR. The first input of the improved MobileNetV3 is 4 300*300 grayscale images, and the second input is a 1*8 one-dimensional sequence. The 4 300*300 grayscale images correspond to the grayscale images of 3 DR-IMFs and CIR after dimensionality upgrade through GASF. The 1*8 one-dimensional sequence corresponds to the statistical parameter sequence of CIR. The first input enters 11 serial MobileNet convolution blocks to generate 4 feature maps. The 4 feature maps are fused with the first input to form an overall feature map. The overall feature map output generates three branches. The first branch performs minimum and maximum normalization. The second branch performs a softmax activation function and then combines with the third branch to enter 11 serial MobileNet convolution blocks. The result is then fused with the result of the first branch. The result is divided into two branches, one for maximum pooling and the other for average pooling. The second input is linearly encoded and fused with the two branches to obtain the non-line-of-sight NLOS noise factor.

8. The high-precision anti-interference ranging method based on UWB according to claim 7, characterized in that: The MobileNet convolution block includes input, 1×1 convolution + h-swish activation function block, 3×3 depth-separable convolution + h-swish activation function block, feature fusion layer, 1×1 convolution, feature splicing layer and output. There is a shortcut connection directly to the feature splicing layer after the input. The 3×3 depth-separable convolution + h-swish activation function block has branches, which are connected to the pooling layer, fully connected + Relu activation function + fully connected + sigma activation function layer in sequence, and finally connected to the feature fusion layer.

9. The high-precision anti-interference ranging method based on UWB according to claim 8, characterized in that: The line-of-sight LOS noise factor is introduced, and then an improved Kalman filter is performed, and the distance observation value is used as the observation value , update the posterior state estimate posterior covariance matrix, update the residual mean, and finally output the ranging result.

10. The high-precision anti-interference ranging method based on UWB according to claim 9, characterized in that: In the improved Kalman filter: ； ； ； In the formula, , is a known matrix, for Moment control matrix; ； ； In the formula, is the non-line-of-sight NLOS noise factor, when the current signal propagation path is NLOS ,at this time = , Pick a random number between 15 and 30, in cm; ； In the formula, is the observation matrix; ； ； In the formula, and Respectively Moment and time; ； ； ； ； In the formula, is through The calculated value.

Images