A multi-class non-line-of-sight signal recognition system and method based on space-time features

Abstract

The application provides a multi-class non-line-of-sight signal recognition system and method based on space-time features, the method comprises collecting CIR data in multiple non-line-of-sight scenes, then the features extracted by the space-time feature extractor are selected and fused to make the fused features have effective representation ability for UWB line-of-sight / non-line-of-sight, finally, based on the difference of the fused features, the multi-class non-line-of-sight signal recognition of UWB signals in a complex indoor scene is realized, the method has strong generalization ability, stability and high recognition efficiency, and can be conveniently deployed on an edge platform, which provides a reliable basis for subsequent alleviation of indoor specific ranging or positioning errors, and helps to improve indoor positioning accuracy.

Description

Technical Field

[0001] The invention relates to the field of wireless positioning technology, specifically to a multi-category non-line-of-sight signal recognition system and method based on spatiotemporal characteristics. Background Technology

[0002] With the development of IoT technology, the demand for accurate indoor positioning has increased across various industries. UWB signals contain line-of-sight (LOS) information of the indoor environment, allowing indoor positioning systems to locate and track objects based on the acquired UWB signals. However, UWB-based indoor positioning systems are susceptible to non-line-of-sight (NLOS) conditions, leading to significant errors in positioning calculations. Therefore, accurately identifying and mitigating indoor NLOS signals is crucial.

[0003] In indoor NLOS scene recognition tasks, manually designed UWB signal features can accurately describe the characteristics of indoor NLOS scenes. For example, the "Line-of-Sight / Non-Line-of-Sight Path Recognition Method Based on Channel Impulse Response Energy Distribution" (CN109151724A) applied for by Central South University uses various machine learning methods to discriminate signal energy features, thereby determining the LOS / NLOS binary classification type of the signal. However, the design of these features often requires a large amount of prior knowledge, and the feature extraction process is quite cumbersome. Deep learning-based classification technology has gradually become a new hot topic in the field of NLOS recognition. Research on deep feature extraction and deep network structure optimization has promoted further improvement in the accuracy of NLOS scene recognition. For example, the "A Deep Learning-Based UWB NLOS Signal Recognition Method" (CN109151724A) applied for by Beijing University of Posts and Telecommunications builds a dual-stream deep neural network and trains the network using a public UWB measurement dataset to recognize LOS / NLOS signals collected in real-world environments. However, it requires the collection of a large amount of data for training, which is a cumbersome process. Furthermore, most current NLOS classification methods focus only on a simple binary classification of LOS / NLOS, which fails to adequately assess the severity of NLOS signals and thus cannot utilize the effective components within the NLOS signals to improve positioning accuracy. Therefore, improvements to existing technologies are urgently needed. Summary of the Invention

[0004] To address the positioning drift problem in UWB-based indoor positioning systems under NLOS occlusion conditions, this invention provides a multi-class NLOS signal identification method based on spatiotemporal features: spatiotemporal features are extracted from channel impulse response data using a designed model, and then fused with key features from a lightweight attention mechanism. Multi-class NLOS signal identification is then performed based on the fused feature pairs. The method employed by this invention to solve its technical problem is as follows:

[0005] Channel impulse response data were collected in an indoor environment with multiple NLOS scenarios. Training dataset, validation dataset and test dataset were constructed based on the channel impulse response data of the indoor environment.

[0006] A spatiotemporal feature model is constructed, which mainly includes four functional parts: a feature extractor, an attention mechanism, a feature fusion unit, and a classification and recognition unit. The feature extractor is used for deep extraction of spatial and temporal features of the channel impulse response data in UWB signals; the attention mechanism is used to adjust the weight of each spatiotemporal feature of the channel impulse response data; the feature fusion unit is used to fuse the original deep spatiotemporal features obtained from the feature extractor with the key features from the attention mechanism, outputting more detailed fused features; and the classification and recognition unit predicts the NLOS type of the UWB signal to be identified based on the fused features.

[0007] The CIR data from the collected UWB dataset is labeled and divided into training, validation, and test sets according to a preset ratio. The labels represent the non-line-of-sight (NLOS) category of the target in the CIR data. The CIR dataset contains several scenes, and each scene contains several target data. The CIR data from the training dataset is input into the initial network, and the initial network is iteratively trained until convergence using a preset loss function to obtain a converged network. During the iteration process, the network parameters in the feature extractor, attention mechanism, feature fusion unit, and classifier are continuously adjusted to ensure that the overall loss function meets the accuracy requirements. The validation dataset is input into the converged network to verify the NLOS recognition capability of the converged network. If the verification is successful, the trained spatiotemporal feature model is obtained. The CIR data from the test dataset is input into the trained spatiotemporal feature network to obtain the recognition results of the NLOS signal category output by the spatiotemporal feature model.

[0008] The feature extractor consists of a spatial feature extraction part based on a one-dimensional convolutional network and a temporal feature extraction part based on a long short-term memory network; it performs in-depth extraction of spatial and temporal features for the channel impulse response data of UWB signals, respectively.

[0009] The one-dimensional convolutional network used to extract the spatial features of the channel impulse response data in UWB signals consists of three sequential batch-normalized convolutional filters, namely the first convolutional filter, the second convolutional filter, and the third convolutional filter; the kernel sizes of the first convolutional filter to the third convolutional filter are K and K, respectively. h ×K w 2K h ×K w 4K h ×K w The definition of the above convolutional filter is as follows:

[0010]

[0011] In the formula, Z = [z1, z2, ..., z L [This is] the input channel impulse response signal, with a length L = 512. This represents the output feature map after the convolution operation; the feature map has spatial dimensions H×W and channels C; V=[v1, v2, …, v C [] is the set of learned filter kernels, where each filter v c Defined as

[0012] Adding normalization to the feature extractor provides overall regularization, defined as follows:

[0013]

[0014] In the formula, γ BN These are learnable parameters, μ and σ. 2 These are the batch mean and variance, where ε is a small constant added to the variance for numerical stability; after batch normalization, the output is... The feature map is spatially downsampled by passing through a linear rectified activation function and adding a pooling layer after the convolutional neural network module.

[0015] The Long Short-Term Memory (LSTM) network used to extract the temporal features of the channel impulse response signal consists of 512 hidden LSM units. The deep spatial features extracted by the one-dimensional convolutional network in the feature extractor are planarized through a flattening layer and then fed into the LSM network for temporal feature extraction to generate a spatiotemporal feature sequence. The calculation of one LSM network memory block is as follows:

[0016] f t =σ(w f ·[h t-1 ,x t ]+b f (3)

[0017] i t =σ(w i ·[h t-1 ,x t ]+b i (4)

[0018]

[0019]

[0020] O t =σ(w o·[h t-1 ,x t ]+b o (7)

[0021] h t =o f ·tanh(C t (8)

[0022] In the formula, f t Let C represent the activation vector of the forget gate, C represent the activation vector of the input / update gate, and C×t represent the cell input activation vector. t For the memory of the current cell, O t h represents the activation vector of the output gate. t Output for the current cell; the bias vectors and weight matrices of the input gate (i), output gate (o), forget gate (f), and memory unit (c) are represented by b, w, and h. t-1 It is the output of the previous unit, c t-1 It is the memory of the previous unit, σ represents the sigmoid function, and "·" represents the Hadamard product;

[0023] The attention mechanism used to adjust the weights of each spatiotemporal feature in the channel impulse response data mainly consists of an efficient channel attention mechanism; it generates global descriptors for the corresponding channel dimensions from the spatiotemporal features of the channel impulse response data, as shown in the following formula:

[0024]

[0025] In the formula, y c This represents the global descriptor for the c-th channel, where c = 1, 2, ..., C; x c (i, j) represents the value of the input feature X at position (i, j) in the c-th channel; then, the local dependencies between channels are modeled through one-dimensional convolution, with the kernel size k adaptively determined according to the number of channels C:

[0026]

[0027] In the formula, ψ(C) is a function based on the number of channels; γ and b are hyperparameters used to adjust the range of kernel size; |·| odd This means adjusting the result to the nearest odd number; next, the channel attention weights are generated, using the following formula:

[0028] σ c =f conv1d (y,k)=Conv1D(y,W k (11)

[0029] s c =sigmoid(σc (12)

[0030] In the formula, f conv1d This represents a one-dimensional convolution operation, with kernel weights of . y = [y1, y2, ..., y C ] represents the global descriptor vector; σ c This is an intermediate result after convolution; s c Let be the attention weight for the c-th channel, ranging from [0, 1]; sigmoid(·) is the activation function; finally, the weights generated above are weighted onto the original input feature H:

[0031] X c ′=s c ×X c (13)

[0032] In the formula, X c ′ represents the weighted output feature of the c-th channel; X c This represents the original input feature of the c-th channel;

[0033] The feature fusion unit processes the raw depth-space-time feature vector H obtained from the feature extractor. lstm The key feature vector H of the attention mechanism eca The fusion process is performed to generate a fused feature vector F. concat =[H lstm H eca Then, a flattening layer is used to flatten it; then two fully connected layers are used, with a ReLU layer and a dropout layer embedded between fully connected layer 1 and fully connected layer 2; furthermore, the number of neurons in fully connected layer 1 in the feature fusion unit is much greater than the number of neurons in fully connected layer 2; the second fully connected layer is defined as the output layer, consisting of five fully connected neurons;

[0034] The output of the classifier is based on the output of the feature fusion unit. The output of the softmax layer is defined as the output class probability. The NLOS scene recognition result of the UWB signal to be recognized is predicted by the following formula:

[0035]

[0036] in and These are the is-th and js-th input elements, respectively. Softmax(·) is the softmax activation function, and exp(·) is the exponential function. According to... The output size identifies the LOS / NLOS matching results in five scenarios.

[0037] This invention also designs a multi-classification recognition system for UWB NLOS signals based on a spatiotemporal feature model, comprising:

[0038] Data acquisition module: used to collect corresponding channel impulse response data in an indoor environment with multiple NLOS scenarios, and to build a classification training dataset based on the collected channel impulse response data in multiple NLOS scenarios;

[0039] The model building module is used to construct a spatiotemporal feature model, which includes a feature extractor, an attention mechanism, a feature fusion unit, and a classification and recognition unit. The feature extractor extracts temporal and spatial features from the channel impulse response data for NLOS multi-class recognition. The attention mechanism further extracts key features from the temporal and spatial features extracted by the feature extractor to improve the accuracy of NLOS multi-class recognition. The feature fusion unit fuses the original spatiotemporal features from the feature extractor and the key features from the attention mechanism. The classification and recognition unit determines the actual environment's channel impulse response based on the fused features output by the feature fusion unit. The model training module is used to train the spatiotemporal feature model for classification and recognition using the training dataset until the spatiotemporal feature model converges and the classifier achieves a certain accuracy in judging the NLOS scene type of unlabeled UWB signals. The classification and recognition module is used to obtain the NLOS type of the channel impulse response data to be identified. The channel impulse response data to be identified is input into the feature extractor, attention mechanism, feature fusion unit and classifier of the pre-trained spatiotemporal feature model. The classification and recognition module outputs the NLOS scene type of the channel impulse response data to be identified.

[0040] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the steps of the method described above by calling the computer program stored in the memory.

[0041] The beneficial effects of this invention are mainly reflected in:

[0042] (1) The spatiotemporal feature-based multi-class NLOS signal recognition method of the present invention is based on the channel impulse response data with high discriminativeness collected in multi-class NLOS scenarios. It uses a spatiotemporal feature model to extract deep spatiotemporal features for distinguishing similar NLOS scenarios, and further uses an attention mechanism to extract key NLOS representation features. This makes the fused features output by the feature fusion unit have sufficient multi-class NLOS signal discriminativeness, so that the classifier can determine the NLOS signal type of the channel impulse response data, so as to achieve accurate multi-class recognition of NLOS signals in UWB positioning networks in indoor scenarios.

[0043] (2) This method is based on a lightweight attention mechanism to assist in the extraction of channel impulse response features, thereby achieving a balance between high classification accuracy and low processing time, and can be rapidly deployed on embedded devices with limited computing resources. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the workflow of a deep learning-based multi-category NLOS signal recognition method according to a preferred embodiment of this application.

[0045] Figure 2 This is a schematic diagram of the sub-process of step S3.

[0046] Figure 3 This is a comparison chart of the NLOS recognition and classification performance of the spatiotemporal feature model.

[0047] Figure 4 This is a comparison chart of the recognition results of the spatiotemporal feature model and other mainstream classification models.

[0048] Figure 5 This is a diagram showing the recognition results of the NLOS multi-class recognition system.

[0049] Figure 6 This is a schematic diagram of the module structure of the NLOS multi-class recognition system. Detailed Implementation

[0050] The present invention will be further described below with reference to specific embodiments.

[0051] Example 1: A deep learning-based multi-class NLOS signal recognition method, the specific process of which is as follows: Figure 1 As shown, it includes:

[0052] Step S1: Collect channel impulse response data of UWB signals in indoor environments with multiple NLOS scenarios, and construct a training dataset based on the collected channel impulse response data.

[0053] Step S2: Construct a spatiotemporal feature model, which consists of four parts: a feature extractor, an attention mechanism, a feature fusion unit, and a classification and recognition unit. The feature extractor extracts temporal and spatial features from the channel impulse response data for NLOS multi-class recognition; the attention mechanism further extracts key features from the temporal and spatial features extracted by the feature extractor to improve the accuracy of NLOS multi-class recognition; the feature fusion unit fuses the original spatiotemporal features from the feature extractor and the key features from the attention mechanism; and the classification and recognition unit determines the NLOS scene type of the actual environment's channel impulse response data based on the fused features output by the feature fusion unit.

[0054] Step S3: Select channel impulse response datasets with different occlusion types from the training dataset for spatiotemporal feature model classification training until the spatiotemporal feature model is trained to convergence, obtain a converged network, and the classifier achieves a certain accuracy in recognizing NLOS scenes of unlabeled UWB signals, then the training ends.

[0055] Step S4: Input the channel impulse response data to be identified into the feature extractor, attention mechanism, feature fusion unit and classification unit of the pre-trained spatiotemporal feature model, and output the NLOS scene type of the channel impulse response data to be identified through the classification unit.

[0056] It is understood that the spatiotemporal feature-based multi-class NLOS signal recognition method in this embodiment, by collecting channel impulse response data with high discriminativeness under multiple NLOS scenarios, extracting deep spatiotemporal features for distinguishing similar NLOS scenarios using a spatiotemporal feature model, and giving sufficient attention to important features based on an attention mechanism, enables the fused features output by the feature fusion unit to have sufficient multi-class NLOS signal discriminativeness. Thus, the classifier can determine the NLOS signal type of the channel impulse response data, thereby achieving accurate multi-class recognition of NLOS signals in UWB positioning networks under changing scenarios, greatly improving the robustness of the model and the accuracy of multi-class recognition results.

[0057] Understandably, this method is based on deep learning and incorporates a lightweight and efficient channel attention mechanism to assist in the extraction of channel impulse response features. It eliminates the need for manual feature selection and extraction of UWB signals, which helps to identify deep spatiotemporal features with high distinguishing contribution. Thus, it achieves a balance between high classification accuracy and low processing time, enabling fast and effective differentiation of NLOS signals that are often confused by other methods on embedded devices with limited computing resources.

[0058] Specifically, in step S1, channel impulse response (CTR) data of UWB signals are collected in indoor environments with multiple NLOS scenarios. CTR data is a UWB signal feature used in the signal domain to identify LOS and NLOS signals. Multiple NLOS scenarios refer to scenarios that cause varying degrees of obstruction and multipath propagation effects on UWB signals, including but not limited to wall obstruction, human body obstruction, glass obstruction, metal obstacles, and other common indoor obstruction scenarios. Furthermore, a training dataset is constructed based on the collected indoor environment CTR data to facilitate subsequent classification and recognition training of the spatiotemporal feature model. Optionally, since there are significant differences in the dielectric constants among the various NLOS scenarios, the CTR data under different NLOS scenarios will exhibit certain differences, which is beneficial for improving the robustness of the model. It is understood that this invention does not require extensive UWB experimental data collection in indoor environments; only a small amount of CTR data under various NLOS environments is needed.

[0059] In step S2, a spatiotemporal feature model is constructed. This model primarily comprises four functional components: a feature extractor, an attention mechanism, a feature fusion unit, and a classification and recognition unit. The feature extractor is used for deep extraction of spatial and temporal features from the channel impulse response data in UWB signals. The attention mechanism is used to adjust the weights of each spatiotemporal feature in the channel impulse response data. The feature fusion unit fuses the original deep spatiotemporal features obtained from the feature extractor with the key features from the attention mechanism, outputting more detailed fused features. The classification and recognition unit predicts the NLOS type of the UWB signal to be identified based on the fused features, enabling multi-class recognition of NLOS signals and providing support for accurate and effective mitigation of ranging errors. The feature extractor employs a spatiotemporal feature extraction neural network, preferably including a one-dimensional convolutional neural network with multiple serial batch normalization and a long short-term memory network.

[0060] like Figure 2 As shown, in step S3, channel impulse response datasets with different occlusion types are selected from the training dataset for spatiotemporal feature model classification training until the spatiotemporal feature model training converges, resulting in a converged network. The classifier can then determine the NLOS scene type of unlabeled UWB signals, including the following:

[0061] Step S31: Select channel impulse response data from multiple categories of LOS / NLOS signals in the training dataset, and use a feature extractor to extract spatiotemporal features from them. During the iteration process, continuously adjust the network parameters of the feature extractor to obtain the spatiotemporal features of the channel impulse response data from multiple categories of LOS / NLOS signals in the training dataset and label and distinguish them to construct a spatiotemporal feature training dataset.

[0062] Step S32: Select channel impulse response data and corresponding spatiotemporal features of multi-class LOS / NLOS signals from the training dataset. Use an attention mechanism to adjust the weights of the spatiotemporal features of the channel impulse response data to obtain key features. Then, use a feature fusion device to fuse the original spatiotemporal features with the key features and add the corresponding NLOS signal type label to construct the fused feature training dataset.

[0063] Step S33: Set the overall loss function of the spatiotemporal feature model, input the constructed spatiotemporal feature training dataset and the fused feature training dataset into the classifier respectively, and continuously classify the network parameters of the classifier during the iteration process so that the overall loss function meets the accuracy requirements. Finally, the classifier determines the NLOS signal type of the fused features, and finally obtains the spatiotemporal feature model with attention mechanism and the spatiotemporal feature model without attention mechanism respectively.

[0064] It is understood that the attention mechanism is not unique, and efficient attention mechanisms are preferred. To fully compare the effectiveness of different attention mechanisms, comparisons can be made using other attention mechanisms against the attention mechanism selected in this invention. The comparison results are as follows: Figure 3 As shown, the NLOS signal type recognition accuracy achieved by the efficient channel attention mechanism is the highest, and it also performs well in terms of recognition resource consumption and recognition speed.

[0065] In the iterative training of the initial network, the global loss function uses an adjustment factor added to the cross-entropy loss to give more attention to these multi-class NLOS signals that are challenging to classify, as shown below:

[0066]

[0067] In the formula, α is a balancing factor used to balance the uneven proportion of positive and negative samples, and γ is a focusing parameter used to control the contribution of easily classified and difficult-to-classify samples to the loss. is i y The predicted probability of a class.

[0068] In step S4, the channel impulse response data to be identified is input into the feature extractor, attention mechanism, feature fusion unit, and classification unit of the trained spatiotemporal feature model. The classification unit then outputs the NLOS scene type of the channel impulse response data. Specifically, the channel impulse response data to be identified is sequentially input into the trained feature extractor and attention mechanism to extract the spatiotemporal features and key features of the data. These features are then fused by the feature fusion unit. The trained classification unit then performs multi-class NLOS identification based on the fused features and outputs the NLOS scene type of the channel impulse response data. This enables the detailed identification of NLOS occlusion between base stations and between base stations and tags in multi-base station indoor positioning systems, which is beneficial for accurately assessing the severity of NLOS and multipath effects.

[0069] The above steps S1 to S3 are offline training, and step S4 is online recognition. Of course, in other embodiments of the present invention, steps S1 to S3 can also be trained online.

[0070] It is understood that, in order to demonstrate the effectiveness of the UWB NLOS multi-class recognition method based on the spatiotemporal feature model of this invention, the inventors of this application conducted experimental comparisons with current mainstream NLOS recognition models on a validation dataset. The experimental results are as follows: Figure 4 As shown, the prediction rate and recall rate for the five types of NLOS signals are mostly over 95.5%, which demonstrates that the method proposed in this invention has a strong ability to identify and classify NLOS signals.

[0071] In addition, such as Figure 5 As shown, another embodiment of the present invention also provides a UWB NLOS multi-class recognition system based on spatiotemporal features, preferably employing the method described above. The system includes:

[0072] Data acquisition module: used to collect corresponding channel impulse response data in an indoor environment with multiple NLOS scenarios, and to build a classification training dataset based on the collected channel impulse response data in multiple NLOS scenarios;

[0073] The model building module is used to construct a spatiotemporal feature model, which includes a feature extractor, an attention mechanism, a feature fusion unit, and a classification and recognition unit. Specifically, the feature extractor extracts temporal and spatial features from the channel impulse response data for NLOS multi-class recognition; the attention mechanism further extracts key features from the temporal and spatial features extracted by the feature extractor to improve the accuracy of NLOS multi-class recognition; the feature fusion unit fuses the original spatiotemporal features from the feature extractor and the key features from the attention mechanism; and the classification and recognition unit determines the NLOS scene type of the actual environment's channel impulse response data based on the fused features output by the feature fusion unit.

[0074] Model training module: Used to train the spatiotemporal feature model for classification and recognition using the training dataset until the spatiotemporal feature model is trained to convergence, resulting in a converged network. The training ends when the classifier achieves a certain accuracy in recognizing NLOS scenes of unlabeled UWB signals.

[0075] Classification and Recognition Module: Used to obtain the NLOS type of the channel impulse response data to be identified. The channel impulse response data to be identified is input into the feature extractor, attention mechanism, feature fusion unit and classification and recognition unit of the trained spatiotemporal feature model. The classification and recognition unit outputs the NLOS scene type of the channel impulse response data to be identified.

[0076] It is understood that the UWB NLOS multi-class recognition system based on the spatiotemporal feature model in this embodiment, by collecting channel impulse response data with high discriminativeness under multiple NLOS scenarios, extracting deep spatiotemporal features for distinguishing similar NLOS scenarios using a deep learning network, and giving sufficient attention to important features based on a lightweight attention mechanism, enables the fused features output by the feature fusion unit to have sufficient discriminativeness for multi-class NLOS signals. Thus, the classifier can determine the NLOS signal type of the channel impulse response data, thereby achieving accurate multi-class recognition of NLOS signals in the UWB positioning network under changing scenarios, greatly improving the robustness of the model and the accuracy of the multi-class recognition results.

[0077] Finally, it should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for multi-class NLOS signal identification based on spatio-temporal features, the method comprising: Includes the following steps: CIR data were collected in an indoor environment with multiple NLOS scenarios. Training, validation, and test datasets were constructed based on the indoor CIR data. A spatiotemporal feature model was built, comprising four main functional parts: a feature extractor, an attention mechanism, a feature fusion unit, and a classifier. The feature extractor is used for depth extraction of spatial and temporal features from the CIR data in UWB signals. The attention mechanism is used to adjust the weights of each spatiotemporal feature in the CIR data. The feature fusion unit fuses the raw spatiotemporal features obtained from the feature extractor with the key features from the attention mechanism, outputting fused features. The classifier predicts the NLOS type of the UWB signal to be identified based on the fused features, achieving multi-class recognition of non-line-of-sight signals and providing support for accurately and effectively mitigating ranging errors in subsequent operations. The CIR data from the collected UWB dataset is labeled and divided into training, validation, and test sets according to a preset ratio. The labels represent the non-line-of-sight (NLOS) target category to which the CIR data belongs. The CIR dataset contains several scenes, each containing several target data points. The CIR data from the training dataset is input into an initial network, which is iteratively trained using a preset loss function until convergence, resulting in a converged network. During the iteration process, the network parameters in the feature extractor, attention mechanism, feature fusion unit, and classifier are continuously adjusted to ensure the overall loss function meets the accuracy requirements. The validation dataset is input into the converged network to verify its NLOS recognition capability. If the verification is successful, a trained spatiotemporal feature model is obtained. The CIR data from the test dataset is input into the trained spatiotemporal feature model to obtain the NLOS signal category recognition results output by the spatiotemporal feature model. The multi-category NLOS signal recognition method based on spatiotemporal features is characterized by comprising the following steps: Note that the mechanism adjusts the weights of each spatiotemporal feature of the CIR data. It first generates a global description of the spatiotemporal features of the CIR data corresponding to the channel dimensions, as shown in the following formula: where y c represents the global descriptor of the c-th channel, c = 1, 2, …, C; x c (i, j) represents the value of the input feature X at position (i, j) on the c-th channel; then, the local dependency relationship between channels is modeled by one-dimensional convolution, and the convolution kernel size k is adaptively determined according to the number of channels C: where ψ(C) is a function based on the number of channels; γ and b are hyperparameters to adjust the range of kernel size; denotes adjusting the result to the nearest odd number; next, the channel attention weight is generated by using a one-dimensional convolution operation, and the formula is as follows: In the formula, f conv1d represents a one-dimensional convolution operation, and the convolution kernel weight is ; y = [y1, y2, …, y C ] is a global descriptor vector; σ c is an intermediate result after convolution; s c is the attention weight of the cth channel, ranging between [0, 1]; sigmoid(·) is an activation function; finally, the above generated weight is weighted to the original input feature H: In the formula, represents the output feature of the cth channel after weighting; represents the original input feature of the cth channel.

2. The spatio-temporal feature based multi-class NLOS signal identification method according to claim 1, characterized in that, The process of inputting CIR data from the training dataset into the initial network, iteratively training the initial network using a preset loss function until convergence, and obtaining a converged network by continuously adjusting the network parameters in the feature extractor, attention mechanism, feature fusion unit, and classification unit during the iteration process to ensure that the overall loss function meets the accuracy requirements includes the following: Channel impulse response data from multiple categories of LOS / NLOS signals are selected from the training dataset, and spatiotemporal features are extracted from them using a feature extractor. The network parameters of the feature extractor are continuously adjusted during the iteration process to obtain the spatiotemporal features of the channel impulse response data from multiple categories of LOS / NLOS signals in the training dataset and to label and distinguish them in order to construct a spatiotemporal feature training sample set. Channel impulse response data and corresponding spatiotemporal features of multiple categories of LOS / NLOS signals are selected from the training dataset. The spatiotemporal features of the channel impulse response data are weighted and adjusted using an attention mechanism to obtain key features. The original spatiotemporal features and key features are then fused using a feature fusion ... The overall loss function of the spatiotemporal feature model is set, and the constructed spatiotemporal feature training sample set and the fused feature training sample set are respectively input into the classifier. During the iteration process, the network parameters of the classifier are continuously improved to ensure that the overall loss function meets the accuracy requirements. The classifier determines the NLOS signal type based on the fused features. Finally, spatiotemporal feature models with and without attention mechanisms are obtained respectively.

3. The spatio-temporal feature based multi-class NLOS signal identification method according to claim 1, characterized in that, The feature extractor is used for deep extraction of spatial and temporal features of CIR data in UWB signals, including the following: The feature extractor consists of a spatial feature extraction part and a temporal feature extraction part; the spatial feature extraction part and the temporal feature extraction part respectively perform depth extraction of spatial and temporal features from the CIR data of the UWB signal; The one-dimensional convolution network for extracting CIR data space features in the UWB signal is composed of three serial batch normalization convolution filters, and the three serial convolution filters are respectively a first convolution filter, a second convolution filter and a third convolution filter; the kernel sizes of the first convolution filter to the third convolution filter are respectively , , ; the definitions of the above convolution filters are as follows: wherein, is an input CIR signal, length L = 512, U = [u1, u2,..., u C ] e H×W×C represents the output feature map after convolution operation; the feature map has spatial dimensions H x W and channels C; V = [v1, v2,..., v C ] is a set of learned filter kernels, where each filter v c is defined as [v1 C, v2 C,..., vL C]; normalization is performed on the input to reduce the number of training rounds, which is defined as follows: In the formula, is a learnable parameter, and is the mean and variance of the batch, and ε is a parameter for maintaining numerical stability; after the batch normalization operation, the output is transmitted through a linear rectification activation function; a pooling layer is added after the convolutional neural network module to perform spatial downsampling of the feature map; the linear rectification activation function and the max pooling layer can introduce sparsity and enhance nonlinearity for the spatiotemporal feature model; The deep spatial features extracted by the one-dimensional convolutional network are planarized by the flattening layer and then fed into the long short-term memory network for temporal feature extraction to generate a spatiotemporal feature sequence.

4. The spatio-temporal feature based multi-class NLOS signal identification method according to claim 1, characterized in that, Includes the following steps: The feature fusion unit processes the raw deep spatiotemporal feature vectors obtained from the feature extractor. Key feature vectors of the attention mechanism The feature vector is fused to produce a fused feature vector. Then, a flattening layer is used to flatten it; further, two fully connected layers are used, with a linear rectified function layer and a dropout layer embedded between fully connected layer 1 and fully connected layer 2; further, the number of neurons in fully connected layer 1 in the feature fusion machine is much larger than the number of neurons in fully connected layer 2; the second fully connected layer is defined as the output layer, consisting of five fully connected neurons; the output of the classifier is based on the output of the feature fusion machine, and the output of the softmax layer is defined as the output class probability, predicting the NLOS scene recognition result of the UWB signal to be recognized, which is achieved by the following formula: wherein and are the first and the second input elements, Softmax(·) is a softmax activation function, is an exponential function; according to the output size of Softmax(x softmax is LOS / NLOS matching results in five scenarios are identified.

5. A UWB NLOS multi-classification identification system based on a space-time feature model, adopting the multi-class NLOS signal identification method based on deep learning according to any one of claims 1-4, characterized in that, Includes the following: Data acquisition module: used to collect corresponding CIR data in an indoor environment with multiple NLOS scenarios, and to build a classification training dataset based on the collected CIR data from multiple NLOS scenarios; The model building module is used to construct a spatiotemporal feature network model, which includes a feature extractor, an attention mechanism, a feature fusion unit, and a classifier. The feature extractor extracts temporal and spatial features from CIR data for NLOS multi-class recognition. The attention mechanism further extracts key features from the temporal and spatial features extracted by the feature extractor to improve the accuracy of NLOS multi-class recognition. The feature fusion unit fuses the original spatiotemporal features from the feature extractor and the key features from the attention mechanism. The classifier determines the non-line-of-sight scene type of the actual environment CIR data based on the fused features output by the feature fusion unit. Model training module: Used to train the spatiotemporal feature network model for classification and recognition using the training dataset until the spatiotemporal feature network model is trained to convergence and the classifier determines the NLOS scene type of the unlabeled UWB signal to a certain accuracy, then the training ends. Classification and Recognition Module: Used to obtain the non-line-of-sight type of the CIR data to be recognized. The CIR data to be recognized is input into the feature extractor, attention mechanism, feature fusion unit and classification and recognition unit of the trained spatiotemporal feature network model. The classification and recognition unit outputs the non-line-of-sight scene type of the CIR data to be recognized.

Images