CN115201750A - A NLOS identification method for ultra-wideband positioning system - Google Patents

Abstract

The invention discloses an NLOS (non line of sight) identification method of an ultra-wideband positioning system, which comprises the steps of acquiring channel impulse response time sequence data of an anchor point or a mobile tag in the UWB system; performing wavelet transformation on the acquired channel impulse response data, and calculating to acquire a time-frequency information graph of the channel impulse response data; constructing a deep neural network, and generating a time-frequency information graph by using the acquired LOS and NLOS channel impulse response data to train the time-frequency information graph to obtain an NLOS identification model; and transplanting the constructed model to a mobile tag hardware platform of the UWB system, and performing NLOS identification and positioning performance verification in an indoor environment. According to the invention, the time-frequency information graph of the channel impulse response time sequence is processed by using the deep neural network, time and frequency information is effectively extracted, and the constructed deep neural network can effectively perform NLOS identification in the UWB indoor positioning system.

Description

A NLOS identification method for ultra-wideband positioning system

technical field

The invention belongs to the technical field of indoor positioning, and in particular relates to an NLOS identification method of an ultra-wideband positioning system.

Background technique

In order to meet the growing demand for high-precision positioning of people and assets, radio positioning technology has developed rapidly in various indoor positioning scenarios and has broad market application prospects. With the increase of human indoor activities, various wireless positioning indoor services such as Bluetooth, WIFI, ZigBee, Near Field Communication (NFC) and UWB are emerging. Among these indoor positioning technologies, UWB is considered as one of the most promising wireless collaborative indoor positioning solutions. At present, UWB is widely used in indoor positioning systems due to its low power consumption, high transmission rate and strong penetration. Compared with other RF technologies, UWB has a bandwidth greater than 500MHz, extremely short transmit pulses, high temporal and spatial resolution, and considerable error suppression of multipath effects. It differs from existing narrowband communications because the UWB spectrum consists of multiple subbands, which enables wirelessly transmitted data to travel through obstacles at relatively high data rates. In addition, by utilizing the characteristics of UWB narrow pulse and high time resolution, the ranging accuracy can be effectively improved, thereby improving the positioning accuracy.

Currently, UWB positioning technology has been demonstrated to achieve centimeter-level accuracy under line-of-sight (LOS) propagation. However, various obstacles are encountered when using UWB in practical indoor environments. These obstacles, such as human bodies, concrete walls, glass windows, metal panels and wooden doors, may block and reflect UWB wireless signals, introduce multipath interference and generate NLOS signals. Compared with the LOS condition, the transmitted signal arrives at the receiver with different degrees of delay under the NLOS condition. Furthermore, when the delay spread is relatively large, the received NLOS signal becomes attenuated and distorted. Therefore, in UWB positioning applications, it is very important to reduce and eliminate the error caused by NLOS propagation.

Current LOS/NLOS identification methods can be classified into two different types based on channel impulse response (CIR) and other signal metrics. For example, the statistical parameter difference of estimated distance information can be used for NLOS and LOS classification. The distance noise under the LOS condition obeys a Gaussian distribution with zero mean, while the distance noise under NLOS condition can be modeled as a Gaussian distribution with a non-zero mean. Therefore, hypothesis testing can be performed based on distance variance and mean difference. Borras J. also took into account the variance of the distance measure and used a binary hypothesis test to determine the NLOS condition. Jo et al. combined ray tracing algorithms with statistical or map-based methods for NLOS recognition. However, these techniques may not be applicable when the tag is moving in the actual positioning system, as it requires multiple measurements of a pair of positions. Additionally, detection thresholds may vary across sites and environments.

The main idea of another class of CIR-based NLOS identification methods is that the energy of the first path is significantly larger than that of the delayed path. In previous studies, NLOS identification was performed based on traditional statistical features of the CIR of the received signal, such as kurtosis, peak lead delay, average excess delay, and root mean square (RMS), which appeared differently under NLOS and LOS conditions value of . But the discrimination threshold of NLOS varies in different locations and environments. Therefore, it is difficult to identify NLOS in various scenarios using standard statistical parameters. To address this problem, some nonparametric machine learning methods are used for NLOS and LOS classification. For example, using input vector machines (IVM) for NLOS recognition has the advantage of having few input vectors and minimal sparsity. However, in a UWB localization system, its efficiency may be severely degraded with the increase of anchor and label observation data. In addition, Support Vector Machine (SVM) and Correlation Vector Machine (RVM) are applied to NLOS recognition, by using SVM and correlation vector respectively to establish decision boundary according to the input features, so as to obtain more accurate and robust NLOS recognition. However, manually selected vector features generated from UWB signal propagation path loss models may not be sufficient for identification requirements in various localization scenarios. With the development of machine learning, convolutional neural network (CNN) and long short-term memory (LSTM) have been shown to show superior performance in ultra-wideband LOS and NLOS time series data classification. A CNN is used to automatically detect and extract the CIR features of the received UWB signal, and then the CNN output is fed into an LSTM for classification. However, this method does not consider the difference in spectral information between the received LOS and NLOS signals, and the serial structure design of CNN and LTSM is relatively complicated. Simone et al. achieved effective ranging error mitigation at the edge through high-semantic feature extraction of CIR signals, correction under LOS or NLOS conditions, and deep learning and graph optimization techniques. However, this method only uses the time domain information of the received signal, and the NLOS positioning accuracy only reaches the decimeter level.

SUMMARY OF THE INVENTION

Aiming at the problem of NLOS recognition of UWB system by different obstacles such as human body, concrete wall, glass window, metal plate and wooden door that may exist in the room, the present invention proposes a LOS and NLOS signal based on CIR time-frequency diagram and deep neural network model. recognition methods. First, the neural network model takes as input the time-frequency information map of the raw channel impulse response CIR of the base stations and tags of the UWB positioning system. At the same time, the original CIR signal input by continuous wavelet transform (CWT) is processed to generate a time-frequency information map, which can be regarded as an A×A pixel image. Then, a deep learning neural network is designed, which is trained with the time-frequency information graph, and the trained model can be applied to the recognition of NLOS and LOS signals received on the UWB system mobile tags.

An ultra-wideband positioning system NLOS identification method disclosed by the present invention comprises the following steps:

Obtain the CIR time series data of anchor points or mobile tags in the UWB system;

The obtained CIR data is subjected to wavelet transform, and the time-frequency information map of the CIR is obtained by calculation;

Build a deep neural network and train it with the collected LOS and NLOS data to obtain a NLOS recognition model;

The constructed model is transplanted to the mobile tag hardware platform of UWB system for NLOS recognition in indoor environment.

Further, each anchor or mobile tag transmits a symbol modeled as:

where P is the transmit amplitude of a pulse, s(τ) is a single Gaussian pulse waveform, and N pulses with a period of t _s constitute a specific frame;

The transmission channel pulse is as follows:

where γ _m and β _m represent the fading coefficient and time delay of the mth path, respectively. Therefore, the received signal is the accumulation of fading and delayed transmission signals in several different transmission paths, and is described as:

加性高斯白噪音。where ψ(t) represents the variance as

Additive white Gaussian noise.

Further, the time-frequency information map of CIR is generated using continuous wavelet transform, and continuous wavelet transform is defined as:

where p(τ) is the received impulse response sequence signal, κ _b,μ (τ) is the scaling and shift of the basic wavelet κ(τ), expressed as:

where b is the scale factor, θ is the delay factor, and the non-analytic Mollet wavelet is used as the basic wavelet, and its frequency domain mathematical expression is defined as:

The default value of ω ₀ in the formula is 6.

Finally, a time-frequency information graph is drawn according to the values of the continuous wavelet transform coefficient matrix of the input CIR signal.

Further, the deep neural network includes 4 convolutional layers, 4 max pooling layers and 1 flattening layer;

Firstly, the time-frequency information map of the input CIR is processed into a two-dimensional matrix through pixel mathematical transformation;

With random kernel initialization, the convolution layer convolves the input time-frequency information map to extract its edge features;

In order to reduce the computational complexity, the maximum pooling layer is used to optimize the matrix size of the output of the convolutional layer;

After extracting and optimizing the time-frequency information map features of CIR, the obtained feature matrix is put into the flat layer and transformed as a one-dimensional vector;

To balance the information of the time-frequency information map and CIR, the output of the flattening layer is sent to the hidden layer to reduce the size of subsequent combinations.

Further, the connection between hidden layers is achieved in the following ways:

r _i =q _i *e _j +u _i

Among them, qi and _{ui represent} the weight coefficient and bias coefficient between the hidden layers, respectively, e _j is the neuron vector, and the neuron vector of the hidden layer is defined by the active function of the rectified linear unit ReLU as:

Further, for the final NLOS and LOS binary classification, the loss function is calculated by the cross-entropy function as:

where D is the number of samples, f is a label with a value of 0 or 1, and p(f) represents the probability of prediction; in backpropagation, the weight coefficients and bias coefficients of each layer are automatically optimized to ensure that the prediction gradually approaches the target; The weight coefficients and bias coefficients are updated in the following ways:

where x is the learning rate, u ^new and u ^old are the new and old bias coefficients, respectively, and q ^new and q ^old are the new and old weight coefficients, respectively.

The beneficial effects of the present invention are as follows:

A novel deep neural network is proposed to process the time-frequency information map of CIR time series for NLOS/LOS identification of UWB-based localization systems, which is effective compared to existing CIR-only NLOS classification methods The time and frequency information are extracted efficiently, and the problem of signal type identification is efficiently transformed into an image classification problem. The constructed deep neural network has more advantages in image feature extraction and classification, and can effectively perform NLOS identification in UWB indoor positioning system.

The indoor UWB positioning test was carried out using the comparison results of different NLOS identification methods, and the ranging curve was fitted from the identified measurement value and the low positioning average error. Compared with the real indoor travel route, it further confirmed that the invention can complete complex positioning. The NLOS identification task in the environment can effectively improve the accuracy of the UWB positioning system.

Description of drawings

Fig. 1 overall flow chart of the present invention;

FIG. 2 is a partial architecture diagram of the identification method of the present invention;

Fig. 3 another part of the structure diagram of the identification method of the present invention;

Figure 4 Difference between CIR of LOS and NLOS;

Figure 5. CIR time-frequency information diagram in the case of LOS;

Figure 6. CIR time-frequency information diagram in the case of NLOS;

Figure 7. UWB positioning system results based on least squares LS positioning and weighted least squares WLS positioning.

Detailed ways

The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way, and any transformation or replacement based on the teachings of the present invention belongs to the protection scope of the present invention.

The NLOS and LOS identification workflow of the present invention based on the CIR time-frequency diagram and the deep neural network structure is shown in Figures 1-3 and consists of the following steps:

Step 1: Receive CIR time series data from anchors or mobile tags in the UWB system. A UWB-based positioning system, where each anchor or tag transmits a symbol can be modeled as:

where P is the transmit amplitude of a pulse, s(τ) is a single Gaussian pulse waveform, and N pulses with a period of t _s constitute a specific frame. A transmission channel pulse can look like this:

where γ _m and β _m represent the fading coefficient and time delay of the mth path, respectively. Therefore, the received signal is an accumulation of attenuated and delayed transmission signals in several different transmission paths, which can be described as:

加性高斯白噪音(AWGN)。where ψ(t) represents the variance as

Additive White Gaussian Noise (AWGN).

For UWB-based positioning systems, the location of the tag is unknown, while the locations of at least three base stations are known in advance. To measure the distance between the tag and the base station in the positioning work, the time-of-flight (TOF) method has emerged, which has the advantage that there is no error caused by clock synchronization deviation in the UWB-based positioning system. The ranging error is determined by the detection of the first occurrence of the received CIR signal. CIR signals are often affected by external interference and physical phenomena such as obstacle fading and multipath fading. Figure 2 shows the difference between the CIRs of LOS and NLOS. By comparing the first and peak path positions in the accumulator, the first and peak paths of LOS are much closer than in the NLOS case. Furthermore, the normalized magnitude of LOS is larger than that of NLOS. Therefore, due to the different transmission paths, LOS and NLOSCIR can be considered as one-dimensional time series with their own characteristics.

Step 2: Construct the time-frequency map of the CIR signal for NLOS and LOS identification. Due to the different properties of CIR signals under LOS and NLOS conditions, it can be predicted that the signal of the first path and the signal of the peak path should exhibit different frequencies in the frequency domain. In order to highlight the frequency domain features of the CIR signal and retain its time-varying information, the present invention constructs a time-frequency diagram of the CIR signal for NLOS and LOS identification.

Continuous wavelet transform (CWT) can detect certain features in CIR signals and images, so continuous wavelet transform (CWT) is chosen to generate the time-frequency information map of CIR (refer to the non-patent paper C. Peng, et al., A Noise-Robust ModulationSignal Classification Method Based on Continuous Wavelet Transform, 5th IEEE Information Technology and Mechatronics Engineering Conference, ITOEC 2020, June 12, 2020-June 14, 2020, Institute of Electrical and Electronics Engineers Inc., Chongqing, China, 2020, pp.745-750.). Continuous wavelet transform is the basis of wavelet analysis theory and can be defined as:

where p(τ) is the received impulse response sequence signal and κ _b,μ (τ) is the scaling and shift of the basic wavelet κ(τ). It can be expressed as:

where b is the scale factor, θ is the delay factor, and the non-analytic Mollet wavelet is used as the basic wavelet, and its frequency domain mathematical expression is defined as:

The default value of ω ₀ in the formula is 6. Finally, according to the value of the CWT coefficient matrix of the input CIR signal, its time-frequency information is plotted.

The CIR time-frequency information graphs of LOS and NLOS are shown in Figure 4 and Figure 5. Obviously, the time-frequency infographics of NLOS and LOS are different in terms of the energy propagation in the frequency and time of the same wavelet. Furthermore, the LOS energy is much more concentrated than the NLOS energy in the time-frequency information map due to CIR signal delay and attenuation of multipath effects. Unlike features in CIR, features in its time-frequency infomap are two-dimensional. The present invention combines the time and frequency features of the time-frequency infographic to increase the information provided for LOS and NLOS identification.

Step 3: In order to ensure high feature extraction efficiency, the time-frequency information two-dimensional image feature extraction module of CIR includes 4 convolutional layers, 4 max-pooling layers and 1 flattening layer. Therefore, the input time-frequency information map is processed into a two-dimensional matrix by pixel mathematical transformation. With random kernel initialization, the convolutional layer convolves the input time-frequency information map to extract its edge features. To reduce the computational complexity, a max-pooling layer is used to optimize the matrix size of the output of the convolutional layer. After extracting and optimizing the time-frequency information map features of CIR, the obtained feature matrix is put into a flat layer and transformed as a one-dimensional vector. To balance the information of the time-frequency information map and CIR, the output of the flattening layer is sent to the hidden layer to reduce the size of subsequent combinations.

The connections between hidden layers are achieved in the following ways:

r _i =q _i *e _j +u _i (7)

Among them, qi and _{ui represent} the weight coefficient and bias coefficient between the hidden layers, respectively. e _j is the neuron vector, and the neuron vector of the hidden layer is defined by the active function of the rectified linear unit (ReLU) as:

In the output layer, the sigmoid activation function is defined as:

For the final NLOS and LOS binary classification, the loss function can be calculated by the cross-entropy function as:

where D is the number of samples, f is a label with a value of 0 or 1, and p(f) represents the likelihood of the prediction. In backpropagation, the weight coefficients and bias coefficients of each layer can be automatically optimized to ensure that the predictions gradually approach the target. The weight factor and bias factor can be updated in the following ways:

where x is the learning rate. At each training stage, these equations are used to update the hyperparameters and reduce the loss function until its value becomes a steady state. Finally, the proposed model-driven prediction is closer to the intended target.

In the present invention, a 4-layer convolutional CNN is designed to construct a network model for NLOS identification. Figure 3 depicts the structure of the deep neural network model. The detailed parameters of the neural network are as follows:

Table 1. Deep Neural Network Model Structure

It is worth noting that for the 4 convolutional layers, the padding used is "same", and the size of the input matrix can be kept to simplify the operation. The number of convolution kernels is 32, 64, 128 and 128, and the size of the convolution kernel is 3×3. To address the vanishing gradient problem and short model training time, ReLU is used as the activation function for each convolutional layer. A Maxpooling layer is set after each convolutional layer, and the sliding window size is set to 2×2. Also, the Dropout regularization coefficient for each layer is set to 0.2 to deal with overfitting. The output of the fourth pooling layer is sent to the flattening layer below, which is used to convert multidimensional vectors to 1D vector data, which is then fully connected with a 64-unit hidden layer to reduce size. The 64 units output by the hidden layer are combined together, and then the output layer of 1 unit is connected to classify the NLOS and LOS signals.

In the present invention, 70,000 NLOS and 70,000 LOS measurements from different locations in typical indoor obstacle environments (including wooden doors, concrete walls, metal panels, human bodies, and glass windows, etc.) are used to construct the overall dataset. 105,000 samples are randomly selected from the dataset for training and testing of the NLOS recognition model of the deep neural network to reduce the possibility of overfitting of the recognition model in some special positions. The selected test data and training data samples are randomly mixed, and the numbers of training and testing data sets are 87500 and 17500. The designed network is trained with 20 epochs. The final NLOS overall recognition rate is 87.45%, which proves that the present invention can effectively discriminate NLOS and LOS on the received CIR time series in the UWB positioning system in the indoor environment with obstacles.

Indoor positioning test results and analysis

In order to verify that the invention reduces the positioning error of the UWB system through NLOS identification, indoor positioning experiments are carried out, and the use of least squares (LS) and weighted least squares (WLS) to solve the UWB system label moving position is compared. In the indoor experiment, six base stations are arranged, and one mobile tag moves in a one-room, one-hall suite of about 50 square meters. The location coordinates of the 6 anchors are: A1(0, 4.93), A2(1.42, 1.62), A3(4.57, 6.22), A4(6.02, 2.72), A5(4.57, 0) and A6(1.70, 0.62). The labels move along a predefined path, as shown in Figure 6 (black straight trajectory). In the UWB-based system, the two-way ranging method is used to calculate the distance between the anchor and the tag in the UWB system. The received CIRs of these 6 anchors are fed into a trained neural network model to classify as NLOS or LOS cases. If there are z base stations classified as NLOS propagation, w ₁ represents the weight corresponding to the label of NLOS propagation and the minimum distance j _{NLOS_min} in the base station, then the remaining z-1 label weights can be determined as:

where j _i represents the distance between the i-th base station and the tag. Similarly, in r LOS-propagated base stations, v ₁ represents the weight corresponding to the LOS-propagated label and the maximum distance j _{LOS_max} in the base station, then the remaining r-1 label weights can be determined as:

where j _t represents the distance between the t-th base station and the tag. In the present invention, z+r=6, w ₁ =0.1, and v ₁ =1 are obtained according to the number of base stations used and the environment setting empirical values. Therefore, the weight matrix can be obtained:

W=diag(w ₁ ,w ₂ ,...,w _n ,v ₁ ,v ₂ ,...,v _m ) (15)

Figure 6 shows the CIR position estimation data collected on the moving trajectory, in which the point set is estimated using WLS, the dotted line trajectory is the curve after fitting using WLS, and the black solid line trajectory is the real moving path. The comparison found that using WLS has higher overall location estimation accuracy because its weights are used for NLOS measurements. The 9th-order polynomial fitting curve of the location dataset obtained using WLS estimation is smooth, which is almost consistent with the actual path, proving that the proposed deep learning neural network can effectively reduce the impact of NLOS on the precise positioning of the UWB system.

The beneficial effects of the present invention are as follows:

A novel deep neural network is proposed to process the time-frequency information map of CIR time series for NLOS/LOS identification of UWB-based localization systems, which is effective compared to existing CIR-only NLOS classification methods The time and frequency information are extracted efficiently, and the problem of signal type identification is efficiently transformed into an image classification problem. The constructed deep neural network has more advantages in image feature extraction and classification, and can effectively perform NLOS identification in UWB indoor positioning system.

The indoor UWB positioning test was carried out using the comparison results of different NLOS identification methods, and the ranging curve was fitted from the identified measurement value and the low positioning average error. Compared with the real indoor travel route, it further confirmed that the invention can complete complex positioning. The NLOS identification task in the environment can effectively improve the accuracy of the UWB positioning system.

As used herein, the word "preferred" means serving as an example, instance, or illustration. Any aspect or design described herein as "preferred" is not necessarily to be construed as advantageous over other aspects or designs. Rather, use of the word "preferred" is intended to present concepts in a specific manner. The term "or" as used in this application is intended to mean an inclusive "or" rather than an exclusive "or." That is, unless specified otherwise or clear from context, "X employs A or B" is meant to naturally include either of the permutations. That is, "X uses A or B" is satisfied in any of the preceding examples if X uses A; X uses B; or X uses both A and B.

Furthermore, although the present disclosure has been shown and described with respect to one implementation or implementation, equivalent variations and modifications will occur to those skilled in the art based on a reading and understanding of this specification and drawings. The present disclosure includes all such modifications and variations and is limited only by the scope of the appended claims. In particular with respect to the various functions performed by the above-described components (eg, elements, etc.), the terms used to describe such components are intended to correspond to any component that performs the specified function of the component (eg, which is functionally equivalent) (unless otherwise indicated), even if not structurally equivalent to the disclosed structures that perform the functions of the exemplary implementations of the present disclosure shown herein. Furthermore, although a particular feature of the present disclosure has been disclosed with respect to only one of several implementations, such feature may be combined with one or other features of other implementations as may be desired and advantageous for a given or particular application combination. Also, to the extent that the terms "including," "having," "containing," or variations thereof, are used in the detailed description or the claims, such terms are intended to include in a manner similar to the term "comprising."

Each functional unit in this embodiment of the present invention may be integrated into one processing module, or each unit may exist physically alone, or multiple or more units may be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware, and can also be implemented in the form of software function modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may also be stored in a computer-readable storage medium. The above-mentioned storage medium may be a read-only memory, a magnetic disk or an optical disk, and the like. The above-mentioned apparatuses or systems may execute the storage methods in the corresponding method embodiments.

To sum up, the above-mentioned embodiment is an embodiment of the present invention, but the embodiment of the present invention is not limited by the embodiment, and any other changes that deviate from the spirit and principle of the present invention, Modifications, substitutions, combinations, and simplifications should all be equivalent substitutions, which are all included within the protection scope of the present invention.

Claims (6)

1. An identification method of an ultra-wideband positioning system NLOS is characterized by comprising the following steps:

acquiring channel impulse response time sequence data of an anchor point or a mobile tag in a UWB system;

performing wavelet transformation on the acquired channel impulse response data, and calculating to acquire a time-frequency information graph of a channel impulse response sequence;

constructing a deep neural network, and training the deep neural network by using the acquired LOS and NLOS time-frequency information graph data to obtain an NLOS identification model;

and transplanting the constructed model to a mobile tag hardware platform of the UWB system to perform NLOS identification in an indoor environment.

2. The ultra-wideband positioning system NLOS identification method of claim 1, wherein each anchor or mobile tag transmission symbol is modeled as:

wherein P is oneThe amplitude of emission of each pulse, s (τ) being a single Gaussian pulse shape with t _s The N pulses of the cycle form a specific frame;

the transmission channel pulses are as follows:

wherein gamma is _m And beta _m Respectively representing the fading coefficient and the time delay of the mth path, M being the total number of paths, and therefore the received signal is the accumulation of the attenuated and delayed transmission signals in several different transmission paths, described as:

where ψ (t) represents a variance of

Additive white gaussian noise.

3. The method of claim 1, wherein the graph of the acquired impulse shock response time-frequency information is generated using a continuous wavelet transform, the continuous wavelet transform being defined as:

where p (τ) is the received impulse response sequence signal, κ _b,μ (τ) is the scaling and shifting of the basic wavelet κ (τ) expressed as:

wherein b is a scale factor, mu is a shift factor, theta is a time delay factor, a non-analytic Moret wavelet is used as a basic wavelet, and a frequency domain mathematical expression is defined as:

in the formula of omega ₀ The default value is a value of 6 which,

and finally, drawing a time-frequency information graph according to the value of the continuous wavelet transform coefficient matrix of the input CIR signal.

4. The ultra-wideband positioning system NLOS identification method of claim 1, wherein the deep neural network comprises 4 convolutional layers, 4 max-pool layers, and 1 flat layer;

firstly, processing an input time-frequency information graph into a two-dimensional matrix through pixel mathematical transformation;

initializing through a random core, and performing convolution processing on an input time-frequency information graph by a convolution layer to extract edge characteristics of the time-frequency information graph;

in order to reduce the computational complexity, the size of a matrix output by the convolutional layer is optimized by using the maximum pool layer;

after extracting and optimizing the characteristics of the time-frequency information graph, putting the obtained characteristic matrix into a flat layer to be used as a one-dimensional vector for transformation;

to balance the information of the time-frequency information map and the CIR, the output of the flat layer is sent to the hidden layer to reduce the size of the subsequent combination.

5. The method for identifying the NLOS of the ultra-wideband positioning system according to claim 1, wherein the connection between the hidden layers is implemented by:

r _i ＝q _i *e _j +u _i

wherein q is _i And u _i Respectively representing the weight coefficient and the bias coefficient between hidden layers, e _j Is a neuron vector, the neuron vector of the hidden layer is defined by an active function of the rectifying linear unit ReLU as follows:

6. the ultra-wideband positioning system (NLOS) identification method of claim 5, wherein for the final NLOS and LOS binary classification, the cross entropy function using the LOSs function is calculated as:

where D is the number of samples, f is a label with a value of 0 or 1, and p (f) represents the predicted likelihood; in back propagation, automatically optimizing the weight coefficient and the deviation coefficient of each layer to ensure that the prediction is gradually close to a target; the weight coefficient and the deviation coefficient are updated by:

where x is the learning rate, u ^new And u ^old New and old deviation coefficients, q, respectively ^new And q is ^old Respectively a new weighting factor and an old weighting factor.

Images