Ultra wide band NLOS identification method and device based on deep neural network
By adopting a lightweight deep-separable convolutional network and gated loop unit deep neural network model in resource-constrained devices, the spatial and temporal characteristics of ultra-wideband channel impulse response signals are extracted, and the problem of insufficient NLOS recognition performance is solved, and efficient and low-complexity NLOS signal recognition is achieved, which is suitable for high-precision positioning in IoT scenarios.
Patent Information
- Application Number
- CN202510287861.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-12
- Publication Date
- 2025-06-27
AI Technical Summary
The prior art has insufficient NLOS recognition performance in resource-constrained devices, and the traditional deep learning model has high computational complexity, making it not suitable for resource-constrained devices.
The lightweight deep-separable convolutional network and gating cyclic unit are adopted to build a deep neural network model, and the spatial characteristics of the ultra-wideband channel impulse response signal are extracted through the deep-separable convolutional network, and the gated cyclic unit extracts time characteristics to realize NLOS recognition.
While maintaining high recognition accuracy, the parameter quantity and calculation complexity of the deep neural network model are greatly reduced, and are suitable for devices with resource limitations, improving the high-precision positioning capability of wireless signals in IoT scenarios.
Smart Images
Figure CN120217094A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wireless signal positioning, and more specifically, to a method and device for ultra-wideband NLOS identification based on a deep neural network. Background Art
[0002] In recent years, with the rapid development of the Internet of Things technology, applications based on location services have emerged continuously, and location information has become key data indispensable in modern life. Among many indoor positioning technologies, Ultra Wide Band (UWB) has attracted much attention due to its high precision, strong anti-interference ability, and excellent spatio-temporal resolution, and has become one of the most promising solutions. The UWB system realizes data transmission by sending and receiving nanosecond-level extremely narrow pulses, and theoretically can achieve centimeter-level or even higher positioning accuracy. However, in practical applications, under non-line-of-sight (NLOS) conditions, due to signal blockage by obstacles, multipath effects and signal attenuation are caused, and the system positioning accuracy often drops significantly. To solve this problem, how to accurately identify and distinguish NLOS signals has become the key to improving the overall positioning performance. The NLOS identification method based on a deep neural network can effectively extract and distinguish the characteristics of line-of-sight (LOS) and NLOS signals by learning a large amount of signal data, provide a more reliable correction basis for subsequent positioning algorithms, and thus significantly improve the positioning effect in complex environments.
[0003] Currently, common NLOS identification methods mainly include classification methods based on channel impulse response (CIR) and other signal statistical metrics. These methods alleviate the NLOS identification problem to a certain extent, but there are still the following deficiencies: (1) Feature extraction is complex, and manual selection of vector features is required, which is difficult to adapt to diverse indoor scenarios; (2) The model complexity is high. Although traditional deep learning models (such as CNN + LSTM) improve the identification accuracy, they have a large number of model parameters and high computational complexity, and are not suitable for resource-constrained devices. Summary of the Invention
[0004] 1. Technical Problems to be Solved
[0005] Aiming at the deficiencies in the prior art, the present invention provides a method and device for ultra-wideband NLOS identification based on a deep neural network. By combining a depthwise separable convolutional network and a gated recurrent unit, it can effectively solve the NLOS identification problem in resource-constrained devices and provide reliable support for high-precision indoor positioning of wireless signals in Internet of Things scenarios.
[0006] 2. Technical Solutions
[0007] The object of the present invention is achieved by the following technical solutions.
[0008] A method for ultra-wideband NLOS identification based on a deep neural network, comprising the following steps:
[0009] Collect an ultra-wideband channel impulse response signal dataset, perform preprocessing, and divide the preprocessed dataset into a training set, a validation set, and a test set;
[0010] Construct a deep neural network model, the deep neural network model includes a depthwise separable convolutional network and a gated recurrent unit, the depthwise separable convolutional network extracts the spatial features of the ultra-wideband channel impulse response signal, and the gated recurrent unit extracts the temporal features of the ultra-wideband channel impulse response signal;
[0011] Train the deep neural network model based on the training set, evaluate the deep neural network model during the training process based on the validation set, and obtain a trained deep neural network model;
[0012] Input the test set into the trained deep neural network model for NLOS identification to obtain an identification result.
[0013] As a further improvement of the present invention, the depthwise separable convolutional network includes a DepthWise convolutional layer, a PointWise convolutional layer, a normalization layer, a ReLU activation function, and a max pooling layer connected in sequence.
[0014] As a further improvement of the present invention, in the DepthWise convolutional layer, each input channel has an independent convolutional kernel, and the calculation formula for the number of parameters of the DepthWise convolutional layer is:
[0015] M DW = K × K × C in
[0016] Wherein, M DW represents the number of parameters of the DepthWise convolutional layer, K represents the convolutional kernel size, and C in represents the number of input channels.
[0017] As a further improvement of the present invention, in the PointWise convolutional layer, a 1×1 convolution is performed on the output of the DepthWise convolutional layer, and the calculation formula for the number of parameters of the PointWise convolutional layer is:
[0018] M PW = l × l × C in × C out
[0019] Wherein, M PWDenote the number of parameters of the PointWise convolutional layer, C out Denote the number of output channels.
[0020] As a further improvement of the present invention, based on the number of parameters of the DepthWise convolutional layer and the number of parameters of the PointWise convolutional layer, the total number of parameters of the depthwise separable convolutional network is obtained, and the total number of parameters of the depthwise separable convolutional network is expressed as:
[0021] M = K×K×C in + 1×1×C in ×C out
[0022] where M represents the total number of parameters of the depthwise separable convolutional network.
[0023] As a further improvement of the present invention, the gated recurrent unit includes a reset gate and an update gate.
[0024] As a further improvement of the present invention, in the gated recurrent unit, relevant parameters about the gated recurrent unit are set, including the feature dimension of the input data, the dimension of the hidden state, the number of layers of the gated recurrent unit layer, the bias term, random dropout, whether it is bidirectional, and the shape of the input data.
[0025] An ultra-wideband NLOS recognition device based on a deep neural network, comprising:
[0026] A data processing module, which collects an ultra-wideband channel impulse response signal dataset and performs preprocessing, and divides the preprocessed dataset into a training set, a validation set, and a test set;
[0027] A model construction module, which constructs a deep neural network model. The deep neural network model includes a depthwise separable convolutional network and a gated recurrent unit. The depthwise separable convolutional network extracts the spatial features of the ultra-wideband channel impulse response signal, and the gated recurrent unit extracts the temporal features of the ultra-wideband channel impulse response signal;
[0028] A model training module, which trains the deep neural network model based on the training set, evaluates the deep neural network model during the training process based on the validation set, and obtains a trained deep neural network model;
[0029] A signal recognition module, which inputs the test set into the trained deep neural network model for NLOS recognition and obtains a recognition result.
[0030] A computer device, comprising a memory and a processor. A computer program is stored on the memory and can run on the processor. When the processor executes the computer program, the method described in any one of the above is implemented.
[0031] A computer-readable storage medium has a computer program stored thereon, and when the computer program is run by a processor, it executes the method described in any one of the above.
[0032] 3. Advantageous Effects
[0033] Compared with the prior art, the advantages of the present invention are as follows:
[0034] (1) For a method and device for ultra-wideband NLOS recognition based on a deep neural network according to the present invention, by adopting a lightweight depthwise separable convolutional network and a gated recurrent unit, while maintaining a high recognition accuracy, the number of parameters and the computational complexity of the deep neural network model are significantly reduced. Compared with existing neural network recognition methods, the present invention reduces the number of parameters by about 70%, making the method more suitable for resource-constrained devices, especially having important practical value in Internet of Things applications.
[0035] (2) For a method and device for ultra-wideband NLOS recognition based on a deep neural network according to the present invention, by combining the depthwise separable convolutional network to extract the spatial features of the ultra-wideband channel impulse response signal and the gated recurrent unit to capture the time features, not only the computational amount of the deep neural network model is effectively reduced, but also the process of signal feature extraction and time series modeling is further optimized. Thus, in a complex indoor environment, the classification accuracy of the present invention is better than that of traditional methods, and it can effectively improve the accuracy of the ultra-wideband positioning system, ensuring efficient and real-time NLOS signal recognition.
[0036] (3) For a method and device for ultra-wideband NLOS recognition based on a deep neural network according to the present invention, the constructed deep neural network model has characteristics such as a small number of parameters and low computational requirements, is suitable for deployment in an Internet of Things environment, can meet the real-time processing requirements of resource-constrained devices, not only improves the processing ability of the devices, but also ensures efficient NLOS signal classification under low power consumption and low latency, providing an effective solution for high-precision wireless signal positioning applications in resource-constrained device environments. Description of the Drawings
[0037] Figure 1 It is a flowchart of the ultra-wideband NLOS recognition method according to an embodiment of the present invention;
[0038] Figure 2 It is a time-frequency information diagram of the ultra-wideband channel impulse response signals of LOS and NLOS according to an embodiment of the present invention;
[0039] Figure 3 It is an architecture diagram of the deep neural network model according to an embodiment of the present invention;
[0040] Figure 4 It is a schematic diagram of the structure of the depthwise separable convolutional network according to an embodiment of the present invention. Specific Embodiment
[0041] The present invention will be described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments.
[0042] Embodiment
[0043] As Figure 1 shown, a method for ultra-wideband NLOS recognition based on a deep neural network provided in this embodiment includes the following steps: collecting an ultra-wideband channel impulse response signal dataset and performing preprocessing, and dividing the preprocessed dataset into a training set, a validation set, and a test set; constructing a deep neural network model, the deep neural network model includes a depthwise separable convolutional network and a gated recurrent unit, the depthwise separable convolutional network extracts the spatial features of the ultra-wideband channel impulse response signal, and the gated recurrent unit extracts the temporal features of the ultra-wideband channel impulse response signal; training the deep neural network model based on the training set, evaluating the deep neural network model during the training process based on the validation set, and obtaining a trained deep neural network model; inputting the test set into the trained deep neural network model for NLOS recognition to obtain a recognition result.
[0044] Specifically in this embodiment, data collection and processing. Non-line-of-sight (NLOS) and line-of-sight (LOS) measurement data from seven different indoor scenarios are used, including Office 1, Office 2, small apartment, small workshop, kitchen with living room, bedroom, and boiler room. 3000 NLOS samples and 3000 LOS samples are collected in each scenario, for a total of 42000 data. By collecting samples in different scenarios, the generalization ability of the network model is ensured, and the influence of a specific scenario on the network model is avoided.
[0045] The specific data collection and preprocessing steps are as follows:
[0046] Data collection, using existing DW1000 base stations and tags to collect an ultra-wideband channel impulse response (CIR) signal dataset in NLOS and LOS environments. The ultra-wideband channel impulse response signal dataset contains 1031 features.
[0047] Feature screening, using an existing Python program to intercept the features after the 15th column, that is, CIR0 - CIR1015. The dimension of each input data after interception is 1×1016 dimensions, and a total of 1016-dimensional features are included. Each row of data corresponds to a label value, where 1 represents NLOS and 0 represents LOS.
[0048] Data division: 42,000 pieces of data are divided into a training set (37,800 pieces) and a test set (4,200 pieces) in a 9:1 ratio. Then, the training set is further divided into a training set (34,020 pieces) and a validation set (3,780 pieces) in a 9:1 ratio. The training set, validation set, and test set all contain the same proportion of NLOS and LOS data to ensure the balance of the network model.
[0049] Data loading and processing: Use the existing DataLoader to batch load the data and convert it into PyTorch tensors. The training set data is randomly shuffled, while the validation set and test set data are not shuffled. Specifically, the training set, validation set, and test set are batch loaded through the DataLoader, which supports randomly shuffling the data order and setting the batch size. The data is processed in the input format of the network model, including data dimensionality increase, PyTorch tensor conversion, etc. Among them, when loading the training set data, it will be randomly shuffled, and the data less than one batch will be discarded. When loading the validation set data, the data will not be randomly shuffled, and the data less than one batch will be discarded. When loading the test set data, the data will not be randomly shuffled, and the data less than one batch will be discarded.
[0050] The ultra-wideband channel impulse response signal dataset processed through the above steps lays a foundation for the efficient training and accurate evaluation of the network model.
[0051] As Figure 2 shown, the abscissa in the figure represents time, and the ordinate represents energy. As Figure 2 can be seen, the amplitude of the LOS signal is significantly higher than that of the NLOS signal. The ultra-wideband channel impulse response signal can be regarded as a time series. Due to the differences in signal transmission paths, the characteristic curves of the ultra-wideband channel impulse response signals under NLOS and LOS conditions are significantly different. Especially under the NLOS condition, the influence of the received signal on the ultra-wideband channel impulse response signal curve is more obvious. This difference provides an important basis for classifying NLOS and LOS signals using deep learning methods based on ultra-wideband channel impulse response signal data.
[0052] Furthermore, a deep neural network model is constructed. As Figure 3 shown, the deep neural network model includes a depthwise separable convolutional network (DepthWise CNN) and a gated recurrent unit (GRU). In order to extract the spatial and temporal features of the ultra-wideband channel impulse response signal, and at the same time be able to reduce the number of parameters of the network model so that it can be deployed on resource-constrained devices. In this embodiment, a lightweight depthwise separable convolutional network is selected to extract spatial features to reduce the number of parameters and computational complexity, and a lightweight gated recurrent unit is used to extract temporal features to further optimize the modeling of temporal information. The advantages of the two networks are fused to construct a deep neural network model for efficiently classifying NLOS signals.
[0053] Specifically, as Figure 4 shown, the depthwise separable convolutional network includes a DepthWise convolutional layer, a PointWise convolutional layer, a normalization layer, a ReLU activation function, and a max pooling layer connected in sequence.
[0054] In this embodiment, an existing PyTorch is used to build the depthwise separable convolutional network. In the depthwise separable convolutional network, the DepthWise layer is first constructed, then the PointWise layer is constructed, then the BatchNorm layer is used for normalization, and then the ReLU activation function is used to increase the non-linear expression of the depthwise separable convolutional network. Finally, the max pooling layer is used to compress the spatial features extracted by the depthwise separable convolutional network to reduce the amount of computation.
[0055] In this embodiment, the depthwise separable convolutional network decomposes a complete convolution operation into two steps, including DepthWise convolution and PointWise convolution. Using the depthwise separable convolutional network can significantly reduce the number of parameters of the deep neural network model.
[0056] In the prior art, the calculation of the number of parameters of the traditional convolutional layer is:
[0057] M0 = K × K × C in × C out
[0058] where M0 represents the number of parameters of the traditional convolutional layer, K represents the convolutional kernel size, and C in represents the number of input channels, and C out represents the number of output channels.
[0059] In this embodiment, the depthwise separable convolutional network divides the convolution into a DepthWise convolutional layer and a PointWise convolutional layer. In the DepthWise convolutional layer, each input channel has an independent convolutional kernel (i.e., each input channel is processed separately). The calculation of the number of parameters of the DepthWise convolutional layer is:
[0060] M DW = K × K × C in
[0061] where M DW represents the number of parameters of the DepthWise convolutional layer, K represents the convolutional kernel size, and C in represents the number of input channels.
[0062] In the PointWise convolutional layer, a 1×1 convolution is performed on the output of the DepthWise convolutional layer. The calculation of the number of parameters of the PointWise convolutional layer is:
[0063] M PW = l×l×C in ×C out
[0064] Among them, M PW represents the parameter quantity of the PointWise convolutional layer, and C out represents the number of output channels.
[0065] In this embodiment, based on the parameter quantity of the DepthWise convolutional layer and the parameter quantity of the PointWise convolutional layer, the total parameter quantity of the depthwise separable convolutional network is obtained. The total parameter quantity of the depthwise separable convolutional network is expressed as:
[0066] M = K×K×C in + 1×1×C in ×C out
[0067] Among them, M represents the total parameter quantity of the depthwise separable convolutional network.
[0068] Through the above calculations, it can be seen that the total parameter quantity of the depthwise separable convolutional network is significantly reduced compared with the parameter quantity of the traditional neural network, and thus the computational amount and memory consumption are also reduced.
[0069] Build a gated recurrent unit to extract the time features in the ultra-wideband channel impulse response signal. The gated recurrent unit has a simpler internal structure compared with the LSTM module. The LSTM mainly uses an input gate, a forget gate, an output gate and a cell state to control the flow of information, while the gated recurrent unit includes a reset gate and an update gate and does not have an independent cell state. Therefore, the calculation is simpler and more efficient and has fewer parameter quantities.
[0070] In this embodiment, an existing PyTorch is used to build a gated recurrent unit, and relevant parameters about the gated recurrent unit are set in the nn.GRU method, including the feature dimension (input_size) of the input data, the dimension of the hidden state (hidden_size), the number of layers (num_layers) of the gated recurrent unit layer, the bias term (bias), random dropout (dropout), whether it is bidirectional (bidirectional), and the shape of the input data (batch_first).
[0071] Thus, in this embodiment, the ultra-wideband channel impulse response signal data is input into the deep neural network model. The data dimension is 64×1×1016 dimensions. Among them, 64 represents the batch_size size of the input data, 1 represents the number of input channels, that is, each piece of data has only one feature channel, and 1016 represents that the feature dimension of one piece of data is 1016 dimensions.
[0072] Furthermore, in order to extract the spatial features of the ultra-wideband channel impulse response signal data, a deep neural network model is constructed. In this embodiment, the deep neural network model includes two layers of depthwise separable convolutional networks. The depthwise separable convolutional network inputs the ultra-wideband channel impulse response signal data into the DepthWise convolutional layer, and each input channel in this convolutional layer has an independent convolutional kernel. At this time, the number of input channels is 1, the number of convolutional kernels is 10, the size of the convolutional kernel is 4, and the groups value is set to be the same as the in_channels value, so that each channel will have a separate convolutional kernel for processing. Since the size of the convolutional kernel is set to 4, the length of each feature map after convolution will be reduced to 1016 - 4 + 1 = 1013. Each input sample (64 samples) is processed by 10 convolutional kernels, generating 10 channels. Therefore, the final output dimension is 64 × 10 × 1013 dimensions.
[0073] Then, the output data of the previous layer is used as the input data to enter the PointWise convolutional layer, and a 1×1 convolution is performed on each channel to change the number of channels. At this time, the number of channels is increased from 10 to 20. This convolutional operation does not change the time length of the data. Therefore, the input dimension of this layer is 64 × 20 × 1013 dimensions. Furthermore, the output data is normalized through the BatchNorm layer, aiming to accelerate training and reduce overfitting, and then passes through the ReLU activation function layer to enhance the non-linear expression ability of the model.
[0074] In this embodiment, after being processed by the DepthWise convolutional layer and the PointWise convolutional layer, the dimension of the input ultra-wideband channel impulse response signal data becomes 64 × 20 × 1009 dimensions. The output dimension then undergoes a max pooling operation through the last layer of the depthwise separable convolutional network. The pooling window size is 2 and the stride is 2, reducing the temporal length of the data. At this time, the final data output dimension of the depthwise separable convolutional network is 64 × 20 × 504 dimensions.
[0075] Furthermore, the extracted spatial features are input into a gated recurrent unit to further extract the temporal features of the ultra-wideband channel impulse response signal data. In this embodiment, the gated recurrent unit includes two layers, and the internal bidirectional is set to False, that is, a unidirectional gated recurrent unit, thereby avoiding overfitting caused by the excessive complexity of the deep neural network model and reducing the parameters of the deep neural network model at the same time. When the spatial features enter the first layer of the gated recurrent unit, a hidden state will be generated, and then the hidden state of the first layer will be used as input data to enter the second layer of the gated recurrent unit for processing to generate the final output. The dimension of the input data is 64×20×504 dimensions, and the gated recurrent unit will process the data according to the timing information and output a new hidden state. After passing through two layers of gated recurrent units, the output dimension is 64×504×16 dimensions, where 504 represents the feature length at each moment, and 16 represents the number of hidden units of the gated recurrent unit.
[0076] In this embodiment, in the actual calculation, the processing of the subsequent layers of the gated recurrent unit is performed, and the output of the last moment of the gated recurrent unit is extracted, that is, from all the moment outputs of the gated recurrent unit, only the hidden state of the last moment is concerned. In a timing task, the last moment in the timing usually contains the key features of the entire sequence and can usually represent the sequence features well, especially in a classification task. Thus, the timing data is converted into a fixed-length representation. Finally, after the output processing of the gated recurrent unit, the data dimension is 64×16 dimensions.
[0077] A BatchNorm normalization layer is connected after the gated recurrent unit, and then a fully connected layer, which is used to flatten the input data. Then there is a ReLU activation function layer, which is used to further enhance the non-linear expression ability of the deep neural network model. Then there is a fully connected classification layer, and the output is 64×2 dimensions, which is used to output NLOS and LOS signals.
[0078] Furthermore, network model training and optimization. In this embodiment, the deep neural network model is trained based on the training set, and the deep neural network model during the training process is evaluated based on the validation set to obtain the trained deep neural network model. Specifically, the data set is loaded and the constructed deep neural network model is used for training on the server. During the training process of the deep neural network model, the performance evaluation results of the validation set are used to further optimize the performance of the deep neural network model. Monitor the classification accuracy and loss function value on the validation set, and adjust the hyperparameters of the deep neural network model according to their changes. For example, learning rate, regularization parameter, and network structure depth, etc. If overfitting or underfitting problems are found, corresponding measures are taken, such as adjusting the Dropout rate, increasing data augmentation, etc., to improve the generalization ability and robustness of the deep neural network model. Through this cyclic optimization, the optimal deep neural network model configuration is finally selected to ensure the efficiency and accuracy in practical applications.
[0079] In this embodiment, the training set data is used to train the deep neural network model. Cross-entropy is used as the loss function, and the Adam optimizer is used for gradient update. The initial learning rate is set to 0.001. During the training process, the Dropout regularization mechanism is added, and the Dropout rate is set to 0.3 to prevent overfitting.
[0080] Finally, network model test and recognition are performed. In this embodiment, the test set is input into the trained deep neural network model for NLOS recognition to obtain the recognition result. Specifically, after the training and optimization of the deep neural network model are completed, an independent test set is used to evaluate the classification result of the constructed deep neural network model. By inputting the data in the test set into the trained deep neural network model, metrics such as classification accuracy, precision, recall, and F1-score are calculated to comprehensively measure the performance of the deep neural network model on unknown data. The use of the test set ensures the generalization ability of the deep neural network model and can verify its stability and accuracy in practical applications. Finally, an evaluation is made on the effectiveness of the deep neural network model based on the test results, and it is determined whether the actual requirements are met. If the performance of the deep neural network model on the test set reaches the expected standard, and the number of parameters of the deep neural network model is small and the computational complexity is low to meet the stable and real-time recognition on resource-constrained devices, then the optimal deep neural network model is saved to ensure its efficiency and stability in practical applications. In addition, the saved deep neural network model can be used for subsequent inference tasks and actual deployment.
[0081] Therefore, a UWB NLOS identification method based on a deep neural network proposed in this embodiment constructs a deep neural network model, trains the deep neural network model using the obtained UWB channel impulse response signals, continuously optimizes the hyperparameters according to the evaluation metrics of the validation set, saves the best deep neural network model during the training process, and then uses it for classification testing of the test set. The deep neural network model file with the best performance is saved locally, and then signals causing positioning errors due to NLOS are identified in actual UWB positioning tasks.
[0082] A UWB NLOS identification method based on a deep neural network proposed in this embodiment combines a depthwise separable convolutional network and a gated recurrent unit to replace the traditional CNN and LSTM architectures, significantly reducing the number of network model parameters while achieving efficient and robust NLOS signal identification on resource-constrained devices. Compared with existing NLOS signal identification methods, the UWB NLOS identification method based on a deep neural network proposed in this embodiment further optimizes the process of signal feature extraction and time series data modeling, improving the accuracy and real-time performance of NLOS signal identification while significantly reducing the number of network model parameters, effectively solving the NLOS identification problem in resource-constrained devices, and providing reliable support for high-precision positioning of wireless signals in Internet of Things scenarios.
[0083] A UWB NLOS identification method based on a deep neural network proposed in this embodiment constructs a lightweight deep neural network model, which not only solves the problem of excessive computational resource requirements in NLOS signal identification but also improves the classification performance and real-time performance, having important practical application value and broad market prospects.
[0084] This embodiment also provides a UWB NLOS recognition device based on a deep neural network, including a data processing module, a model construction module, a model training module, and a signal recognition module. The data processing module is used to collect a UWB channel impulse response signal dataset and perform preprocessing, and divide the preprocessed dataset into a training set, a validation set, and a test set. The model construction module is used to construct a deep neural network, which includes a depthwise separable convolutional network and a gated recurrent unit. The depthwise separable convolutional network extracts spatial features in the UWB channel impulse response signal, and the gated recurrent unit extracts temporal features in the UWB channel impulse response signal. The model training module trains the deep neural network based on the training set and evaluates the deep neural network during the training process based on the validation set to obtain a trained deep neural network. The signal recognition module inputs the test set into the trained deep neural network for NLOS recognition to obtain a recognition result. The UWB NLOS recognition device based on a deep neural network provided in this embodiment can implement any of the methods of the UWB NLOS recognition method based on a deep neural network, and the specific working process of the UWB NLOS recognition device based on a deep neural network can refer to the corresponding process in the embodiment of the UWB NLOS recognition method based on a deep neural network. The methods and devices provided in this embodiment can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for example, the division of a certain module is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed connections or communication connections with each other can be indirect coupling or communication connections through some interfaces, devices or units, and can also be electrical, mechanical or other forms of connections.
[0085] This embodiment also provides a computer device. A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the UWB NLOS recognition method based on a deep neural network described above.
[0086] This embodiment also provides a computer-readable storage medium. A computer-readable storage medium stores a computer program thereon, and when the computer program is run by a processor, it executes the method for identifying ultra-wideband NLOS based on a deep neural network described in this embodiment. Among them, the computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, device, or component; the program code contained on the computer-readable medium can be transmitted by any suitable medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination of the above.
[0087] The present invention and its implementation manners have been schematically described above. This description is not restrictive. Without departing from the spirit or basic features of the present invention, the present invention can be implemented in other specific forms. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. Any reference numeral in the claims should not limit the claimed claim. Therefore, if those of ordinary skill in the art are inspired by it and, without departing from the purpose of this creation, design a structural manner and an embodiment similar to this technical solution without creative efforts, they shall fall within the protection scope of the present invention. In addition, the term "including" does not exclude other elements or steps, and the term "a" before an element does not exclude including "a plurality of" such elements. The plurality of elements stated in the product claims can also be implemented by one element through software or hardware. The terms such as first and second are used to represent names and do not represent any specific order.
Claims
1. A method for ultra-wideband NLOS recognition based on deep neural network, comprising the following steps: Collecting an ultra-wideband channel impulse response signal data set, performing preprocessing, and dividing the preprocessed data set into a training set, a validation set, and a test set; Constructing a deep neural network model, the deep neural network model includes a deep separable convolutional network and a gated recurrent unit, the deep separable convolutional network extracts the spatial characteristics of the ultra-wideband channel impulse response signal, and the gated recurrent unit extracts the temporal characteristics of the ultra-wideband channel impulse response signal; The deep neural network model is trained based on the training set, and the deep neural network model in the training process is evaluated based on the validation set to obtain a trained deep neural network model; The test set is input into the trained deep neural network model for NLOS recognition to obtain the recognition result.
2. The method for ultra-wideband NLOS recognition based on deep neural network according to claim 1, characterized in that: The depth-separable convolutional network includes a DepthWise convolutional layer, a PointWise convolutional layer, a normalization layer, a ReLU activation function and a maximum pooling layer connected in sequence.
3. The method for ultra-wideband NLOS recognition based on deep neural network according to claim 2, characterized in that: In the DepthWise convolution layer, each input channel has an independent convolution kernel, and the calculation formula of the DepthWise convolution layer parameter is: M DW =K×K×C in Among them, M DW DepthWise represents the number of convolutional layer parameters, K represents the convolution kernel size, C in Indicates the number of input channels.
4. The method for ultra-wideband NLOS recognition based on deep neural network according to claim 3, characterized in that: In the PointWise convolution layer, a 1×1 convolution is performed on the output of the DepthWise convolution layer. The calculation formula of the PointWise convolution layer parameter is: M PW =1×1×C in ×C out Among them, M PW Represents the number of PointWise convolutional layer parameters, C out Indicates the number of output channels.
5. The method for ultra-wideband NLOS recognition based on deep neural network according to claim 4, characterized in that: Based on the DepthWise convolution layer parameters and the PointWise convolution layer parameters, the total parameters of the depth-separable convolution network are obtained, and the total parameters of the depth-separable convolution network are expressed as: M=K×K×C in +1×1×C in ×C out Among them, M represents the total number of parameters of the depthwise separable convolutional network.
6. The method for ultra-wideband NLOS recognition based on deep neural network according to claim 1, characterized in that: The gated recurrent unit includes a reset gate and an update gate.
7. The method for ultra-wideband NLOS recognition based on deep neural network according to claim 6, characterized in that: In the gated recurrent unit, set the relevant parameters of the gated recurrent unit, including the feature dimension of the input data, the dimension of the hidden state, the number of layers of the gated recurrent unit layer, the bias term, random dropout, whether it is bidirectional, and the shape of the input data.
8. An ultra-wideband NLOS recognition device based on deep neural network, characterized in that: include: The data processing module collects the ultra-wideband channel impulse response signal data set, performs preprocessing, and divides the preprocessed data set into a training set, a validation set, and a test set; A model building module, constructing a deep neural network model, wherein the deep neural network model includes a deep separable convolutional network and a gated recurrent unit, wherein the deep separable convolutional network extracts spatial features of an ultra-wideband channel impulse response signal, and the gated recurrent unit extracts temporal features of an ultra-wideband channel impulse response signal; The model training module trains the deep neural network model based on the training set, and evaluates the deep neural network model during the training process based on the validation set to obtain a trained deep neural network model; The signal recognition module inputs the test set into the trained deep neural network model for NLOS recognition and obtains the recognition result.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is executed.