Deep learning NLOS identification method suitable for multi-scene complex environment

Abstract

The invention discloses a deep learning NLOS identification method suitable for a multi-scene complex environment, and the method comprises the steps: firstly converting collected multi-scene UWB channel pulse response data to obtain discrete multipath pulse response data, and obtaining multipath delay power data and multipath similarity data through calculation; constructing a new sample data set, and performing data preprocessing; and then constructing a dual-input feature fusion deep learning NLOS identification model based on multipath information to complete NLOS classification in the UWB positioning system. According to the deep learning NLOS identification method suitable for the multi-scene complex environment, the problems of low multi-scene identification capability, poor generalization precision and poor complex scene identification effect can be solved on the premise of a small amount of data, so that relatively high NLOS identification precision is realized, and the UWB positioning precision of the multi-scene complex environment can be effectively improved.

Description

Technical Field

[0001] The present invention relates to a method for identifying wireless signals propagated in non-line-of-sight (NLOS), in particular to a deep learning NLOS identification method applicable to complex environments with multiple scenarios, and belongs to the technical field of ultra-wideband. Background Art

[0002] Positioning technology is widely used in modern life, from navigation to asset tracking, to smart manufacturing and other fields. Although there are various positioning methods, in indoor environments, ultra-wideband (UWB) positioning technology has become a popular choice due to its advantages such as high precision, low power consumption and resistance to multipath interference. UWB positioning mainly relies on ranging technology. In complex environments with multiple scenarios, due to factors such as building structures and obstacles, wireless signals are prone to non-line of sight (NLOS) propagation, and NLOS propagation will cause ranging errors, which in turn leads to inaccurate positioning. Therefore, in UWB positioning systems, accurate identification of NLOS signals is crucial. Correct NLOS identification can not only significantly improve positioning accuracy, but also enhance the robustness of the system, thereby adapting to various complex multi-scenario environments.

[0003] At present, NLOS identification methods are mainly divided into methods based on channel data and methods based on channel features. The method based on channel data directly uses the original channel data for identification, but its identification performance depends on a huge amount of data. In the massive channel data, a considerable part is actually useless data or noise. These noises not only increase the data processing cost, but also may lead to the risk of overfitting of the model. The method based on channel features extracts features from the original channel data, and then uses machine learning or sets multi-type threshold methods for classification. However, the feature extraction process is complicated and may lose key information. The threshold-based method is difficult to effectively distinguish between Line Of Sight (LOS) propagation and NLOS propagation in complex environments with multiple scenarios. Summary of the invention

[0004] In view of the problems existing in the above-mentioned prior art, the present invention provides a deep learning NLOS recognition method suitable for multi-scenario complex environments, which can effectively avoid the shortcomings of traditional channel data-based methods and channel feature-based methods, and realize accurate recognition of NLOS signals. It is particularly suitable for the recognition of NLOS signals during UWB system positioning in multi-scenario complex environments.

[0005] To achieve the above purpose, the deep learning NLOS recognition method applicable to multi-scenario complex environment specifically includes the following steps:

[0006] Step 1: Based on the multi-scenario public data set collected by the UWB positioning system, the channel pulse response data between the anchor point and the tag in the UWB positioning system is collected, and the LOS data label and NLOS data label are marked;

[0007] Step 2: Obtain CIR’s multipath discrete data h based on FPI m (n);

[0008] Step 3: According to the multipath discrete data h m (n) obtaining multipath signal delay power data TDP(x);

[0009] Step 4: First, the multipath discrete data h m (n) Obtain multipath continuous data x(t) by linear interpolation method, and then reconstruct the multipath continuous data x(t) into denoised multipath data by VMD Then calculate the multipath continuous data x(t) and the denoised multipath data The similarity characteristics of are used as multipath similarity characteristic data;

[0010] Step 5, for the delay power data TDP(x) of the multipath discrete data and the multipath similarity feature data, respectively construct a multipath data feature extraction module and a multipath similarity data extraction module to extract the features of the multipath information, fuse the multipath data features and the multipath similarity features to form a dual-channel multipath feature fusion NLOS classification module, and the multipath data feature extraction module, the multipath similarity data extraction module and the multipath feature fusion NLOS classification module together constitute a dual-input feature fusion deep learning NLOS recognition model based on multipath information;

[0011] Step 6, transform all channel impulse response data of the multi-scenario public data set in Step 1, obtain multipath delay power data TDP(x) and multipath similarity feature data, construct a new sample data set, and then perform data preprocessing of setting training set, validation set and test set based on the new sample data set;

[0012] Step 7, input the preprocessed training data into the dual-input feature fusion deep learning NLOS recognition model based on multipath information to train and verify the model. When the model converges, stop training and complete the NLOS classification in the UWB positioning system.

[0013] Furthermore, Step 2 is as follows:

[0014] For the collected channel impulse response data, the PRF of the collection device is set, and the channel model of CIR is expressed as:

[0015]

[0016] Where: t represents the timestamp of each value in CIR; S is the number of multipath components; a s is the amplitude of the Sth multipath component; τ s is the time delay of the Sth multipath component; δ(·) is the Dirac function; n(t) represents the additive white Gaussian noise in the channel;

[0017] The acquisition device represents the collected CIR as a complex-valued IQ sample of 1 ns, and then obtains the corresponding discrete CIR. The discrete CIR is expressed as:

[0018]

[0019] Where: I n ,Q n They represent the real and imaginary parts of the nth complex value of the channel impulse response data collected in units of 1 ns; P is the number of sampling points for each CIR data;

[0020] Take 150 channel impulse response data including FPI and the first path signal as multipath data, and the multipath channel impulse response can be expressed as:

[0021] h m (n)=h(n),n0≤t<n0+150

[0022] Where: h m (n) represents the multipath discrete data of the channel impulse response; n0 is the first path index time value, unit is ns; h(n) represents the discrete CIR.

[0023] Furthermore, Step 3 is as follows:

[0024] The delay power data TDP(x) of multipath discrete CIR data is specifically defined as follows:

[0025]

[0026] Where: L is the sampling factor; x is the data interval based on the sampling factor; n0 corresponds to h m (n) the initial moment.

[0027] Furthermore, the sampling factor L=5 is taken to obtain the delayed power data TDP(x) of the multipath discrete CIR data.

[0028] Furthermore, Step 4 is as follows:

[0029] Step 4-1, multipath discrete data h m (n) obtaining multipath continuous data x(t) by linear interpolation method;

[0030] The multipath continuous data x(t) is expressed as:

[0031]

[0032] Where: is an interval indicator function, when t∈[t n ,t n+1 ] takes the value 1, otherwise it takes the value 0; t n and t n+1 is the discrete multipath CIR time point; t is the interpolation time point; h m (n) and h m (n+1) is the discrete CIR signal value adjacent to the corresponding discrete time point; h m (n) is the multipath discrete data; N is the length of the discrete multipath data;

[0033] Step 4-2, introduce VMD to reconstruct and reduce the noise of multipath continuous data x(t) to obtain denoised multipath data

[0034] VMD solves each mode u by the following optimization method: k (t) and the corresponding center frequency w k :

[0035]

[0036] Where: K is the number of modes, that is, the number of decomposed sub-signals; u k (t) is the kth modal signal; w k is the center frequency of the kth mode signal;

[0037] De-noised multipath data obtained by denoising and reconstruction It is expressed as:

[0038]

[0039] Where: K is the number of modes, that is, the number of decomposed sub-signals; u k (t) is the kth modal signal;

[0040] Step 4-3, calculate the multipath continuous data x(t) and the denoised multipath data The similarity characteristics of are used as multipath similarity characteristic data;

[0041] Calculate the multipath continuous data x(t) and the denoised multipath data The similarity index of CS, SES, PCC, SCC, Manhattan Distance, and Chebyshev Distance are introduced. The similarity characteristics of multipath data are expressed as:

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

[0048] Where: A and B represent the multipath continuous data x(t) and the denoised multipath data respectively. and are the ranks of A and B respectively.

[0049] Furthermore, Step 5 is as follows:

[0050] The multipath data feature extraction module is based on the Attentive CNN model. For the input multipath delay power data, the multipath data feature extraction module uses the convolution layer to extract local features and enhances the perception of key features by introducing the attention mechanism.

[0051] The multipath similarity feature extraction module is based on the MLP model. Aiming at the similarity relationship between multipath data, the multipath similarity feature extraction module uses a fully connected structure to extract similarity features;

[0052] The multipath feature fusion NLOS classification module is based on the DNN model. After extracting the multipath data features and multipath similarity features, the multipath feature fusion NLOS classification module adopts a feature fusion strategy to integrate the two features in a dual-channel manner. The fused features are sent as input to the DNN model for NLOS classification.

[0053] Furthermore, in Step 6, when performing data preprocessing based on the new sample data set, the ratio of the training set, validation set, and test set is set to 7:1:2.

[0054] Furthermore, in Step 7, when the preprocessed training data is input into the dual-input feature fusion deep learning NLOS recognition model based on multipath information to train and verify the model, the dual-input feature fusion deep learning NLOS recognition model based on multipath information sets the loss function to adopt cross entropy, the optimizer to be Adam, and the classification layer activation function to be Softmax. After the sample data of the training set is input into the model, the features are first extracted and the feature fusion is completed through forward propagation, and then the loss between the predicted value and the true value is calculated, and the gradient is calculated and the parameters are updated through back propagation; when the accuracy of the verification set has not been significantly improved within multiple rounds of training cycles, the training is stopped, and the model with the best performance in the verification set is saved for testing to complete the model training.

[0055] Compared with the prior art, the present deep learning NLOS recognition method applicable to complex environments of multiple scenarios first converts the collected multi-scenario UWB channel impulse response data to obtain its discrete multipath impulse response data, obtains its multipath delay power data and multipath similarity data by calculation, constructs a new sample data set, and performs data preprocessing; then constructs a multipath data feature extraction module, a multipath similarity feature extraction module, and a multipath feature fusion classification module, and the above three modules constitute a dual-input feature fusion deep learning NLOS recognition model based on multipath information; then the preprocessed training data is input into the dual-input feature fusion deep learning NLOS recognition model based on multipath information to train and verify the model, and the training is stopped when the model converges to complete the NLOS classification in the UWB positioning system. In terms of accurate recognition of NLOS signals in the positioning process of UWB systems in complex environments of multiple scenarios, through the above experimental simulation verification, the present deep learning NLOS recognition method applicable to complex environments of multiple scenarios can solve the problems of low multi-scenario recognition ability, poor generalization accuracy, and poor complex scene recognition effect under the premise of a small amount of data, thereby achieving a higher NLOS recognition accuracy, thereby effectively improving the UWB positioning accuracy in complex environments of multiple scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the present invention;

[0057] Figure 2 This is a network architecture diagram of the dual-input feature fusion deep learning NLOS recognition model based on multipath information of the present invention;

[0058] Figure 3 It is a graph showing the comparison results of binary classification confusion matrices of the present invention and a support vector machine method based on radial basis function, a support vector machine method based on linear kernel, and a convolutional bidirectional long short-term memory network model method based on original channel impulse response. DETAILED DESCRIPTION

[0059] This deep learning NLOS identification method, which is suitable for complex environments in multiple scenarios, achieves NLOS identification by preventing key features from being submerged by noise, enhancing the effective fusion of features, and making full use of the targeted features of multipath data.

[0060] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0061] like Figure 1 As shown, the deep learning NLOS recognition method applicable to multi-scenario complex environments specifically includes the following steps:

[0062] Step 1, based on the multi-scenario public dataset collected by the UWB positioning system, collect the channel pulse response data between the anchor point and the tag in the UWB positioning system, and mark the data label to indicate whether the data is a LOS signal or a NLOS signal.

[0063] In this embodiment, a multi-scenario public data set collected by a collection device based on the DW 1000UWB chip is used to evaluate the performance of the NLOS identification method. The channel impulse response data of NLOS and LOS between the anchor point and the tag in the UWB system of the data set are collected from the following 7 different indoor environments: Office I, Office II, Small Apartment, Small Studio, Kitchen with Living Room, Bedroom and Boiler Room. 3000 NLOS channel impulse response measurements and 3000 LOS channel impulse response measurements were collected in each environment, totaling 42,000 sets of sample data.

[0064] Step 2: Obtain the multipath discrete data h of the channel impulse response (CIR) m (n).

[0065] The pulse repetition frequency (PRF) of the DW 1000UWB acquisition device is set to 64MHz. Each CIR raw data has 1016 sampling points. The channel model of CIR is expressed as:

[0066]

[0067] Where: t represents the timestamp of each value in CIR; S is the number of multipath components; a s is the amplitude of the Sth multipath component; τ s is the time delay of the Sth multipath component; δ(·) is the Dirac function; n(t) represents the additive white Gaussian noise (AWGN) present in the channel.

[0068] The DW 1000UWB acquisition device represents the collected CIR as a complex-valued IQ sample of 1ns. The IQ sample includes an I (real part) sample and a Q (imaginary part) sample. Then, the corresponding discrete CIR value is obtained, with a total of 1016 data points. The discrete CIR is expressed as:

[0069]

[0070] Where: I n ,Q n They represent the I (real part) sample and Q (imaginary part) sample of the nth complex value of the channel impulse response data collected in units of 1ns.

[0071] For the CIR received between the anchor point and the tag in the LOS environment and the NLOS environment collected in the UWB system, the first path index (First Path Index, FPI) is the position of the first path determined by the DW 1000UWB chip, the data before FPI is the noise stage, and the data after FPI reflects most of the useful information related to the propagation characteristics. Since the selection length of the data sample is related to the environment, a long selection length will introduce useless information and noise, so usually 150 samples are selected, and the 150 channel impulse response data including the first path signal after FPI are taken as multipath data. The multipath channel impulse response can be expressed as:

[0072] h m (n)=h(n),n0≤t<n0+150

[0073] Where: h m (n) represents the multipath discrete data of the channel impulse response; n0 is the first path index time value, unit is ns; h(n) represents the discrete CIR.

[0074] Step 3: According to the multipath discrete data h m (n) Obtain multipath signal delay power data (Time-Delay Power, TDP) TDP(x).

[0075] Although 150 samples are usually required to fully describe the multipath CIR information, the multipath discrete data can be optimized to reduce the number of these samples without sacrificing performance, thereby speeding up the network training process. The feasibility of this optimization is based on two reasons: first, there is a high correlation between adjacent CIR samples; second, the amplitude and overall shape of the CIR are key to channel-based NLOS identification. Optimizing the multipath discrete data can introduce TDP data, which provides a function of the average power of the multipath channel impulse response as a function of delay.

[0076] The multipath signal delay power data TDP(x) can be obtained by calculating the square of the modulus of the multipath impulse response (i.e. |h m (n) 2 ) is used to obtain. In order to effectively utilize the multipath impulse response data and reduce the input dimension of the neural network, |h m (n) 2 The downsampled version of h is calculated with a specific downsampling factor L. The proposed training features are based on |h m (n) 2 The average value in equal intervals, the delay power data TDP(x) of multipath discrete CIR data is specifically defined as follows:

[0077]

[0078] Where: L is the sampling factor; x is the data interval based on the sampling factor; n0 corresponds to h m (n) the initial moment.

[0079] In this embodiment, the sampling factor L=5 is used to obtain the delayed power data TDP(x) of the multipath discrete CIR data.

[0080] Step 4: Calculate and obtain multipath similarity feature data to effectively utilize the multipath discrete data h obtained in Step 2 m (n).

[0081] The multipath discrete data h m (n) The multipath continuous data x(t) is obtained by linear interpolation method. The multipath continuous data x(t) contains environmental noise. The multipath continuous data x(t) can be de-noised and reconstructed by introducing variational mode decomposition (VMD) to obtain de-noised multipath data. Calculate the multipath continuous data x(t) and the denoised multipath data The similarity characteristics of are used as multipath similarity characteristic data. The details are as follows:

[0082] Step 4-1, multipath discrete data h m (n) Obtain multipath continuous data x(t) through linear interpolation method.

[0083] The multipath continuous data x(t) can be expressed as:

[0084]

[0085] Where: is an interval indicator function, when t∈[t n ,t n+1 ] takes the value 1, otherwise it takes the value 0; t nand t n+1 is the discrete multipath CIR time point; t is the interpolation time point; h m (n) and h m (n+1) is the discrete CIR signal value adjacent to the corresponding discrete time point; h m (n) is the multipath discrete data; N is the length of the discrete multipath data.

[0086] Step 4-2, introduce VMD to reconstruct and reduce the noise of multipath continuous data x(t) to obtain denoised multipath data

[0087] VMD decomposes the continuous signal x(t) into several modes u k (t), each mode represents a signal component in a different frequency band. The goal of VMD is to minimize the bandwidth of each mode signal. In order to obtain the variational mode decomposition multipath reconstruction signal, VMD solves each mode u by the following optimization method: k (t) and the corresponding center frequency w k :

[0088]

[0089] Where: K is the number of modes, that is, the number of decomposed sub-signals; u k (t) is the kth modal signal; w k is the center frequency of the kth mode signal.

[0090] In this embodiment, the number of modes K=5 is used to obtain the modal signal and the center frequency of the modal signal.

[0091] De-noised multipath data obtained by denoising and reconstruction It is expressed as:

[0092]

[0093] Where: K is the number of modes, that is, the number of decomposed sub-signals; u k (t) is the kth modal signal.

[0094] Step 4-3, calculate the multipath continuous data x(t) and the denoised multipath data The similarity features of are used as multipath similarity feature data.

[0095] Calculate the multipath continuous data x(t) and the denoised multipath data Similarity indexes are introduced, including cosine similarity (CS), Euclidean similarity (SES), Pearson Correlation Coefficient (PCC), Spearman Correlation Coefficient (SCC), Manhattan Distance, and Chebyshev Distance. The similarity characteristics of multipath data are expressed as follows:

[0096]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102] Where: A and B represent the multipath continuous data x(t) and the denoised multipath data respectively. and are the ranks of A and B respectively.

[0103] Step 5, build a dual-input feature fusion deep learning NLOS recognition model based on multipath information.

[0104] The network architecture of the deep learning NLOS recognition model based on dual input feature fusion of multipath information used in this deep learning NLOS recognition method suitable for multi-scenario complex environments is as follows: Figure 2 As shown. For the delay power data TDP(x) and multipath similarity feature data of multipath discrete data, a multipath data feature extraction module and a multipath similarity data extraction module are respectively constructed to extract the features of multipath information, and the multipath data features and multipath similarity features are fused to form a dual-channel multipath feature fusion NLOS classification module. The multipath data feature extraction module, the multipath similarity data extraction module and the multipath feature fusion NLOS classification module together constitute a dual-input feature fusion deep learning NLOS recognition model based on multipath information.

[0105] The multipath data feature extraction module of the embodiment is based on a convolutional neural network model (Attentive CNN) combined with an attention mechanism, and its parameter settings are shown in the following Table 1. For the input multipath delay power data, the multipath data feature extraction module uses a convolutional layer to extract local features and enhances the perception of key features by introducing an attention mechanism.

[0106] Table 1 Parameter settings of multipath data feature extraction module

[0107]

[0108] The multipath similarity feature extraction module of the embodiment is based on a multilayer perceptron (MLP) model, and its parameter settings are shown in the following Table 2. In view of the similarity relationship between multipath data, the multipath similarity feature extraction module uses a fully connected structure to extract similarity features.

[0109] Table 2 Parameter settings of multipath similarity feature extraction module

[0110]

[0111] The multipath feature fusion NLOS classification module of the embodiment is based on a deep neural network model (Deep Neural Networks, DNN), and its parameter settings are shown in the following Table 3. After extracting features from the above two types of multipath information, the multipath feature fusion NLOS classification module adopts a feature fusion strategy to perform dual-channel integration of the two features, and the fused features are sent as input to the DNN model for NLOS classification. The strong expression ability of DNN is used to classify the fused features, thereby achieving high-precision classification of LOS and NLOS.

[0112] Table 3 Parameter settings of multipath feature fusion NLOS classification module

[0113]

[0114]

[0115] The above-mentioned data processing and algorithm implementation of the embodiment are carried out on a computer with an operating system of win10 (x64), which is configured with an i5-12600KF CPU (3.70GHz), 32.00GBRAM, and NVIDIA RTX 4070SUPER (12GB), and the deep learning and machine learning models of the embodiment are implemented based on the open source machine learning library Pytorch.

[0116] Step 6, transform all channel impulse response data of the multi-scenario public data set in Step 1, obtain the multipath delay power data TDP(x) and multipath similarity feature data including CS, SES, PCC, SCC, HD, Manhattan Distance, and Chebyshev Distance, build a new sample data set, and then preprocess the data to set the training set, validation set, and test set based on the new sample data set.

[0117] The embodiment updates 42,000 sets of sample data to obtain 42,000 sets of new sample data to form a new sample data set. Based on the new sample data set, the ratio of the training set, the validation set, and the test set is set to 7:1:2. 29,400 sets of samples are randomly selected as the training set, 4,200 sets of samples are used as the validation set, and 8,400 sets of samples are used as the test set through a random seed. Randomly selecting these samples helps prevent the model from overfitting.

[0118] Step 7, input the preprocessed training data into the dual-input feature fusion deep learning NLOS recognition model based on multipath information, train and verify the model, and stop training when the model converges to complete the NLOS classification in the UWB positioning system.

[0119] The double-input feature fusion deep learning NLOS recognition model based on multipath information in the embodiment sets the loss function to adopt cross entropy, the optimizer is Adam, the classification layer activation function is Softmax, the discarding probability of the Dropout layer is set to 0.5, and the training Batchsize is set to 200 and the Epoch is set to 50. After the 29,400 groups of sample data of the training set are input into the model, the features are first extracted and the feature fusion is completed through forward propagation, and then the loss between the predicted value and the true value is calculated, and the gradient is calculated through back propagation and the parameters are updated. An early stopping mechanism is set during the training process. When the accuracy of the verification set has not been significantly improved within multiple rounds of training cycles, the training is stopped to prevent overfitting, and the model with the best performance in the verification set is saved for testing to complete the training of the NLOS classification model.

[0120] In order to verify the recognition accuracy of the deep learning NLOS recognition method applicable to multi-scenario complex environment, the classification performance of the deep learning NLOS recognition method applicable to multi-scenario complex environment is compared with the support vector machine (SVM) method based on radial basis function (RBF), the support vector machine (Linear Support Vector Machine, Linear SVM) method based on linear kernel, and the convolutional bidirectional long short-term memory network model (CNN-BiLSTM) method based on original channel impulse response. The comparison results are shown in Table 4 and Figure 3 shown.

[0121] Table 4 Comparison of classification performance of different classification methods

[0122]

[0123] The accuracy of the deep learning NLOS recognition method for complex multi-scenario environments is 88.26%, the precision is 90.32%, the recall rate is 85.61%, and the F1 score is 87.90%. Accuracy is the key to providing accurate positioning information sources for UWB positioning systems. As can be seen from Table 4, the accuracy of the deep learning NLOS recognition method for complex multi-scenario environments is 8.67 percentage points, 9.53 percentage points, and 4.51 percentage points higher than the radial basis function-based support vector machine (RBF SVM) method, the linear kernel-based support vector machine (Linear SVM) method, and the convolutional bidirectional long short-term memory network model (CNN-BiLSTM) based on the original channel impulse response method. The deep learning NLOS recognition method for multi-scenario complex environments has also significantly improved in other indicators, including an accuracy of 88.26%, and very close precision and recall, 90.32% and 85.61% respectively, indicating that the deep learning NLOS recognition method for multi-scenario complex environments can accurately identify NLOS and LOS to the greatest extent, and will not miss NLSO samples excessively; the F1 score reaches 87.90%, indicating that the deep learning NLOS recognition method for multi-scenario complex environments can achieve a good balance between precision and recall. Overall, the deep learning NLOS recognition method for multi-scenario complex environments has the best performance in NLOS recognition in UWB systems.

[0124] Regarding the accurate identification of NLOS signals during the UWB system positioning process in multi-scenario complex environments, the above-mentioned experimental simulation has verified that the deep learning NLOS identification method suitable for multi-scenario complex environments can solve the problems of low multi-scenario recognition ability, poor generalization accuracy and poor complex scene recognition effect under the premise of a small amount of data, thereby achieving higher NLOS recognition accuracy, thereby effectively improving the UWB positioning accuracy in multi-scenario complex environments.

Claims

1. A deep learning NLOS recognition method suitable for multi-scenario complex environments, characterized in that: The specific steps include: Step 1: Based on the multi-scenario public data set collected by the UWB positioning system, the channel pulse response data between the anchor point and the tag in the UWB positioning system is collected, and the LOS data label and NLOS data label are marked; Step 2: Obtain CIR’s multipath discrete data h based on FPI m (n); Step 3: According to the multipath discrete data h m (n) obtaining multipath signal delay power data TDP(x); Step 4: First, the multipath discrete data h m (n) Obtain multipath continuous data x(t) by linear interpolation method, and then reconstruct the multipath continuous data x(t) into denoised multipath data by VMD Then calculate the multipath continuous data x(t) and the denoised multipath data The similarity characteristics of are used as multipath similarity characteristic data; Step 5, for the delay power data TDP(x) of the multipath discrete data and the multipath similarity feature data, respectively construct a multipath data feature extraction module and a multipath similarity data extraction module to extract the features of the multipath information, fuse the multipath data features and the multipath similarity features to form a dual-channel multipath feature fusion NLOS classification module, and the multipath data feature extraction module, the multipath similarity data extraction module and the multipath feature fusion NLOS classification module together constitute a dual-input feature fusion deep learning NLOS recognition model based on multipath information; Step 6, transform all channel impulse response data of the multi-scenario public data set in Step 1, obtain multipath delay power data TDP(x) and multipath similarity feature data, construct a new sample data set, and then perform data preprocessing of setting training set, validation set and test set based on the new sample data set; Step 7, input the preprocessed training data into the dual-input feature fusion deep learning NLOS recognition model based on multipath information to train and verify the model. When the model converges, stop training and complete the NLOS classification in the UWB positioning system.

2. The deep learning NLOS recognition method applicable to multi-scenario complex environments according to claim 1 is characterized in that: Step 2 is as follows: For the collected channel impulse response data, the PRF of the collection device is set, and the channel model of CIR is expressed as: Where: t represents the timestamp of each value in CIR; S is the number of multipath components; a s is the amplitude of the Sth multipath component; τ s is the time delay of the Sth multipath component; δ(·) is the Dirac function; n(t) represents the additive white Gaussian noise in the channel; The acquisition device represents the collected CIR as a complex-valued IQ sample of 1 ns, and then obtains the corresponding discrete CIR. The discrete CIR is expressed as: Where: I n ,Q n They represent the real and imaginary parts of the nth complex value of the channel impulse response data collected in units of 1 ns; P is the number of sampling points for each CIR data; Take 150 channel impulse response data including FPI and the first path signal as multipath data, and the multipath channel impulse response can be expressed as: h m (n)=h(n),n0≤t<n0+150 Where: h m (n) represents the multipath discrete data of the channel impulse response; n0 is the first path index time value, unit is ns; h(n) represents the discrete CIR.

3. The deep learning NLOS recognition method applicable to multi-scenario complex environments according to claim 1 is characterized in that: Step 3 is as follows: The delay power data TDP(x) of multipath discrete CIR data is specifically defined as follows: Where: L is the sampling factor; x is the data interval based on the sampling factor; n0 corresponds to h m (n) the initial moment.

4. The deep learning NLOS recognition method applicable to multi-scenario complex environments according to claim 3 is characterized in that: The sampling factor L=5 is taken to obtain the delayed power data TDP(x) of the multipath discrete CIR data.

5. The deep learning NLOS recognition method applicable to multi-scenario complex environment according to claim 1 is characterized in that: Step 4 is as follows: Step 4-1, multipath discrete data h m (n) obtaining multipath continuous data x(t) by linear interpolation method; The multipath continuous data x(t) is expressed as: Where: is an interval indicator function, when t∈[t n ,t n+1 ] takes the value 1, otherwise it takes the value 0; t n and t n+1 is the discrete multipath CIR time point; t is the interpolation time point; h m (n) and h m (n+1) is the discrete CIR signal value adjacent to the corresponding discrete time point; h m (n) is the multipath discrete data; N is the length of the discrete multipath data; Step 4-2, introduce VMD to reconstruct and reduce the noise of multipath continuous data x(t) to obtain denoised multipath data VMD solves each mode u by the following optimization method: k (t) and the corresponding center frequency w k : Where: K is the number of modes, that is, the number of decomposed sub-signals; u k (t) is the kth modal signal; w k is the center frequency of the kth mode signal; De-noised multipath data obtained by denoising and reconstruction It is expressed as: Where: K is the number of modes, that is, the number of decomposed sub-signals; u k (t) is the kth modal signal; Step 4-3, calculate the multipath continuous data x(t) and the denoised multipath data The similarity characteristics of are used as multipath similarity characteristic data; Calculate the multipath continuous data x(t) and the denoised multipath data The similarity index of CS, SES, PCC, SCC, Manhattan Distance, and Chebyshev Distance are introduced. The similarity characteristics of multipath data are expressed as: Where: A and B represent the multipath continuous data x(t) and the denoised multipath data respectively. and are the ranks of A and B respectively.

6. The deep learning NLOS recognition method applicable to multi-scenario complex environment according to claim 5 is characterized in that: In Step 4-2, solve each mode u k (t) and the corresponding center frequency w k When , the number of modes K is taken as 5.

7. The deep learning NLOS recognition method applicable to multi-scenario complex environment according to claim 1 is characterized in that: Step 5 is as follows: The multipath data feature extraction module is based on the Attentive CNN model. For the input multipath delay power data, the multipath data feature extraction module uses the convolution layer to extract local features and enhances the perception of key features by introducing the attention mechanism. The multipath similarity feature extraction module is based on the MLP model. Aiming at the similarity relationship between multipath data, the multipath similarity feature extraction module uses a fully connected structure to extract similarity features; The multipath feature fusion NLOS classification module is based on the DNN model. After extracting the multipath data features and multipath similarity features, the multipath feature fusion NLOS classification module adopts a feature fusion strategy to integrate the two features in a dual-channel manner. The fused features are sent as input to the DNN model for NLOS classification.

8. The deep learning NLOS recognition method applicable to multi-scenario complex environment according to claim 1 is characterized in that: In Step 6, when performing data preprocessing based on the new sample data set, the ratio of the training set, validation set, and test set is set to 7:1:

2.

9. The deep learning NLOS recognition method applicable to multi-scenario complex environment according to claim 1 is characterized in that: In Step 7, when the preprocessed training data is input into the dual-input feature fusion deep learning NLOS recognition model based on multipath information to train and verify the model, the dual-input feature fusion deep learning NLOS recognition model based on multipath information sets the loss function to adopt cross entropy, the optimizer to be Adam, and the classification layer activation function to be Softmax. After the sample data of the training set is input into the model, the features are first extracted and the feature fusion is completed through forward propagation, and then the loss between the predicted value and the true value is calculated, and the gradient is calculated and the parameters are updated through back propagation; when the accuracy of the verification set has not been significantly improved within multiple rounds of training cycles, the training is stopped, and the model with the best performance in the verification set is saved for testing to complete the model training.

Images