Non-line-of-sight classification and error compensation method and system based on double convolution neural network

Abstract

The application relates to the technical field of positioning device signal processing and positioning precision, and particularly discloses a non-line-of-sight classification and error compensation method and system based on a double convolution neural network, the method comprising the following steps: acquiring channel impulse response data of an ultra-wideband signal, and pre-processing the channel impulse response data to obtain an amplitude sequence in real number form; constructing a double convolution neural network classification model, the double convolution neural network classification model comprising parallel large-kernel convolution blocks and small-kernel convolution blocks, which are respectively used for extracting macroscopic waveform features and microscopic fluctuation features from the amplitude sequence; and performing classification training after fusing the macroscopic waveform features and the microscopic fluctuation features; and based on the trained classification model, the regression compensation of ranging errors of non-line-of-sight signals is performed, macroscopic waveform contour features and microscopic local fluctuation features can be simultaneously extracted from data, and a complementary feature extraction mechanism can more comprehensively describe the differences between LOS / NLOS signals compared with a single-structure network.

Description

Technical Field

[0001] This invention relates to the field of signal processing and positioning accuracy technology for positioning devices, specifically to a non-line-of-sight classification and error compensation method and system based on dual convolutional neural networks. Background Technology

[0002] Indoor positioning technology primarily relies on wireless signals. Whether the signal is obstructed is a key factor affecting positioning accuracy. In unobstructed environments, signal propagation follows a line-of-sight (LOS) path, resulting in relatively stable signal attenuation and allowing positioning algorithms to accurately calculate location. For example, in open shopping mall lobbies or office corridors, Wi-Fi or UWB signals propagate smoothly, enabling positioning systems to achieve high accuracy. However, once obstructions (such as walls, furniture, or people) are present, signal propagation shifts to a non-line-of-sight (NLOS) path, leading to signal reflection, refraction, diffraction, and multipath effects, causing abnormal signal attenuation or fluctuations.

[0003] Traditional classification methods employ relatively simple channel features, typically with fewer than ten digits of feature values. Methods used include k-means, LS, SVM, LSTM, RF, MLP, and CNN. Compared to directly using Channel Impulse Response (CIR) values, traditional methods utilize only known channel features such as signal strength, first, second, and third path amplitudes, and delays. While computationally less, they suffer from lower classification and ranging accuracy. This is especially problematic when signal obstruction causes significant signal fluctuations, leading to larger ranging accuracy deviations and positioning errors. To address the challenges posed by signal obstruction, this paper proposes a novel neural network architecture. This architecture uses continuous signal amplitudes from the positioning device to analyze and determine whether the current location is in a signal obstruction state (i.e., a non-line-of-sight scenario), thereby improving positioning accuracy. Comparison with machine learning methods demonstrates that the proposed neural network model exhibits excellent classification accuracy in line-of-sight scenarios. Finally, based on the previous classification results, the proposed structure is used to further improve ranging accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a non-line-of-sight classification and error compensation method and system based on dual convolutional neural networks to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A non-line-of-sight classification and error compensation method based on dual convolutional neural networks, the method comprising:

[0007] The channel impulse response (CIR) data of the ultra-wideband (UWB) signal is acquired and preprocessed to obtain a real-valued amplitude sequence.

[0008] A dual convolutional neural network classification model is constructed, which includes parallel large convolutional kernel convolutional blocks and small convolutional kernel convolutional blocks, respectively used to extract macroscopic waveform features and microscopic fluctuation features from the amplitude sequence;

[0009] The macroscopic waveform features and microscopic fluctuation features are fused together for classification training;

[0010] Based on the trained classification model, regression compensation for ranging errors is performed on non-line-of-sight signals, and the corrected distance value is output.

[0011] As a further embodiment of the present invention, the large convolutional kernel convolutional block includes two layers of 7x1 large convolutional kernel convolutional layers, which are downsampled based on the max pooling layer;

[0012] The small convolutional kernel convolutional block includes a two-layer 3x1 small convolutional kernel convolutional layer, which is downsampled based on an average pooling layer.

[0013] As a further embodiment of the present invention, the large convolutional kernel block specifically uses 15x1 and 11x1 convolutional kernels connected in series; the small convolutional kernel block specifically uses 7x1 and 5x1 convolutional kernels connected in series.

[0014] As a further aspect of the present invention, a channel attention module is also provided in the process of fusing macroscopic waveform features and microscopic fluctuation features to perform weighted processing on the feature channels.

[0015] As a further embodiment of the present invention, the error compensation step specifically comprises:

[0016] Freeze the feature extraction parameters in the trained classification model;

[0017] A regression subnetwork consisting of fully connected layers is connected after the feature fusion layer;

[0018] The regression subnetwork is trained using a combined loss function, which is a weighted sum of the classification loss function and the regression loss function.

[0019] As a further embodiment of the present invention, the combined loss function is:

[0020] ;

[0021] in, For cross-entropy loss, For root mean square error loss, aThe adjustable weighting coefficient is between 0 and 1.

[0022] This invention also provides a non-line-of-sight classification and error compensation system based on a dual convolutional neural network, used to implement the aforementioned non-line-of-sight classification and error compensation method based on a dual convolutional neural network, the system comprising:

[0023] The data acquisition module is used to acquire the channel impulse response (CIR) data of the ultra-wideband (UWB) signal and preprocess it to obtain the amplitude sequence in real number form.

[0024] The model building module is used to build a dual convolutional neural network classification model, which includes parallel large convolutional kernel convolutional blocks and small convolutional kernel convolutional blocks, respectively used to extract macroscopic waveform features and microscopic fluctuation features from the amplitude sequence;

[0025] The training module is used to perform classification training by fusing macroscopic waveform features and microscopic fluctuation features;

[0026] The compensation module is used to perform regression compensation for ranging errors of non-line-of-sight signals based on the trained classification model, and output the corrected distance value.

[0027] Compared with the prior art, the beneficial effects of the present invention are: it can simultaneously extract macroscopic waveform profile features and microscopic local fluctuation features from CIR data. This complementary feature extraction mechanism can more comprehensively describe the differences between LOS / NLOS signals compared with a single-structure network.

[0028] The introduced channel attention mechanism further enhances the contribution of important feature channels and suppresses redundant information.

[0029] High-precision classification is taken as a prerequisite for error correction. Different correction strategies are adopted for different propagation paths. In particular, for the NLOS path, a deep learning model is used to directly learn the complex mapping relationship of error from CIR features, rather than using a simple fixed value or linear model, so as to more accurately estimate and compensate for NLOS error. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0031] Figure 1 This is a flowchart illustrating the non-line-of-sight classification and error compensation method based on a dual convolutional neural network provided in an embodiment of the present invention.

[0032] Figure 2A typical CIR waveform diagram under the NLOS environment provided for embodiments of the present invention.

[0033] Figure 3 A typical CIR waveform diagram under the LOS environment provided for embodiments of the present invention.

[0034] Figure 4 This is a schematic diagram of the structure of the dual convolutional neural network classification model provided in an embodiment of the present invention.

[0035] Figure 5 The image shows the classification results based on a dual convolutional network structure provided in this embodiment of the invention.

[0036] Figure 6 The image shows the result of distance measurement error correction based on classification results, as provided in an embodiment of the present invention. Detailed Implementation

[0037] To make the technical problems, solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0038] Figure 1 This is a flowchart of a non-line-of-sight classification and error compensation method based on a dual convolutional neural network. In this embodiment of the invention, the non-line-of-sight classification and error compensation method based on a dual convolutional neural network includes:

[0039] The channel impulse response (CIR) data of the ultra-wideband (UWB) signal is acquired and preprocessed to obtain a real-valued amplitude sequence.

[0040] A dual convolutional neural network classification model is constructed, which includes parallel large convolutional kernel convolutional blocks and small convolutional kernel convolutional blocks, respectively used to extract macroscopic waveform features and microscopic fluctuation features from the amplitude sequence;

[0041] The macroscopic waveform features and microscopic fluctuation features are fused together for classification training;

[0042] Based on the trained classification model, regression compensation for ranging errors is performed on non-line-of-sight signals, and the corrected distance value is output.

[0043] In this embodiment, the data consists of multiple sources, including data collected by our own UWB device and raw channel impulse response (CIR) data obtained from open-source datasets. After acquiring the data, it is processed. First, the names and quantities of each feature value are checked to determine if the data acquisition is accurate. Then, invalid rows at the beginning and end of each data set are deleted. Data with a large number of missing values ​​are deleted, and data with a small number of missing values ​​are padded with zeros to ensure data integrity.

[0044] The data format is rectified. For CIR data in complex form, it is processed using the absolute square root formula to convert it into real values ​​for unified calculation during training.

[0045] ;

[0046] Convert the magnitude to real number form, where I It is the result of multiplying the signal by the in-phase component of the local oscillator, representing the in-phase component of the signal. Q It is the result of multiplying the signal by the orthogonal (90° phase shift) component of the local oscillator, representing the orthogonal component of the signal.

[0047] The NLOS values ​​for whether the data is from a line-of-sight environment or not are padded to indicate whether it is from a line-of-sight environment or not.

[0048] Figure 2 The display shows a set of data waveforms under LOS and NLOS environments. First, the characteristics of the CIR waveform are analyzed, and the segment containing the most significant discriminative information is extracted. The 700th to 900th sampling points, a total of 200 points, are used as the input of the model to focus on key features and reduce computational complexity.

[0049] Dataset partitioning and standardization: The processed data is divided into training, validation, and test sets in a 6:2:2 ratio. The input data is standardized to have a mean of 0 and a standard deviation of 1.

[0050] As a preferred embodiment of the present invention, the large convolutional kernel convolutional block includes a two-layer 7x1 large convolutional kernel convolutional layer, which is downsampled based on the max pooling layer;

[0051] The small convolutional kernel convolutional block includes a two-layer 3x1 small convolutional kernel convolutional layer, which is downsampled based on an average pooling layer.

[0052] The large convolutional kernel block specifically uses 15x1 and 11x1 convolutional kernels connected in series; the small convolutional kernel block specifically uses 7x1 and 5x1 convolutional kernels connected in series.

[0053] In the process of fusing macroscopic waveform features and microscopic fluctuation features, a channel attention module is also set up to perform weighted processing on the feature channels.

[0054] In this embodiment, the dual convolutional neural network classification model structure comprises two parallel and structurally distinct convolutional blocks used to extract complementary features from the CIR data. Channel attention is then used to focus on the CIR features after the subsequent parallel dual convolutions, followed by concatenation for feature fusion and classification.

[0055] The dual convolutional structure uses a large convolutional kernel, followed by a 15x1 convolutional layer and then an 11x1 convolutional layer. The large convolutional kernel has a large receptive field, used to capture broad macroscopic shape features such as peaks and troughs in the CIR waveform. Each convolutional layer is followed by a ReLU activation function, and finally, a max-pooling layer is used for downsampling to highlight the salient features of extreme values ​​in the data.

[0056] ;

[0057] ;

[0058] Where Largeconv1 is the result of the first convolution layer, Largeconv2 is the result of the second convolution layer, inch is the input, outch is the output, ksize is the kernel size, and st is the convolution stride.

[0059] In the dual convolutional structure, the small convolutional kernel convolutional block first uses a 7x1 small convolutional kernel convolutional layer, followed by a 5x1 convolutional layer. The small convolutional kernel can capture subtle fluctuations and local ripples in the CIR waveform. Each convolutional layer is followed by a ReLU activation function, and finally an average pooling layer is used for downsampling to preserve the smoothness of the overall distribution.

[0060] ;

[0061] ;

[0062] Conv represents performing a convolution operation, smallconv1 is the result of the first convolution layer, smallconv2 is the result of the second convolution layer, inch is the input, outch is the output, ksize is the kernel size, and st is the convolution stride.

[0063] In the feature fusion part after convolution, a channel attention is added to each of the two convolutional layers to focus on the more subtle features of the CIR value during the convolution process, thereby improving feature extraction capability. The subsequent pooling part is also divided into max pooling and average pooling, inheriting and preserving the different feature inputs from the preceding convolutional parts.

[0064] ;

[0065] ;

[0066] Among them, X large and X small The outputs are two layers, where chatt is the channel attention convolution, inch is the number of input channels, max uses max pooling, and avg uses average pooling.

[0067] Feature fusion and classification are achieved through concatenation. This part of the structure first concatenates the feature maps output from two convolutional blocks along the channel dimension, forming a comprehensive feature vector that integrates macroscopic and microscopic information. This feature vector is then flattened and input into a fully connected layer to enhance important feature channels. Next, it passes through a fully connected network with a Dropout layer to reduce overfitting, followed by an output layer to output the final classification result (LOS or NLOS).

[0068] ;

[0069] ;

[0070] The `combined` option concatenates the two inputs. X large It is the output of the large convolutional layer, X small This is the output of a small convolutional layer. `reshape` is an operation that changes the shape of the tensor, rearranging its dimensions while keeping the total number of elements constant. `B` is the batch size, `C` is the number of channels, and `N` is the number of spatial locations.

[0071] In a preferred embodiment of the present invention, the error compensation step specifically comprises:

[0072] Freeze the feature extraction parameters in the trained classification model;

[0073] A regression subnetwork consisting of fully connected layers is connected after the feature fusion layer;

[0074] The regression subnetwork is trained using a combined loss function, which is a weighted sum of the classification loss function and the regression loss function.

[0075] The combined loss function is:

[0076] ;

[0077] in, For cross-entropy loss, For root mean square error loss, a The adjustable weighting coefficient is between 0 and 1.

[0078] In this embodiment, multiple training sets are used for training during the training phase, and the classification results and training parameters of each set are saved. The results are also printed, and the optimal results are achieved by adjusting the training batch size and learning rate. The preprocessed CIR data is input into a dual convolutional neural network, sequentially passing through a large convolutional kernel block, a small convolutional kernel block, a feature fusion layer, a channel attention module, and a fully connected classifier to obtain the prediction results. The error between the model's predicted output and the true label is calculated using the cross-entropy loss function. The gradient of the loss function with respect to the parameters of each layer is calculated using the chain rule. The Adam optimizer updates the network weights based on the gradient information, and the parameter update direction and step size are adjusted by combining momentum and adaptive learning rate.

[0079] The specific formula used is as follows. The activation function used is the ReLU function, and the formula is:

[0080] ;

[0081] Its function is to keep the gradient constant at 1 in the positive interval, thus avoiding the vanishing gradient problem. Setting all negative values ​​to 0 while keeping positive values ​​unchanged helps the neural network learn data features.

[0082] The pooling layer uses a combination of max pooling and average pooling, employing a sliding window to extract the maximum value within each window, thereby reducing feature dimensionality while preserving salient features. The specific formula is as follows:

[0083] ;

[0084] ;

[0085] in, C is the number of channels, L is the length, k is the kernel size, s is the stride, and j is the index within the window.

[0086] The loss function used is the cross-entropy loss function, which measures the difference between the probability distribution predicted by the model and the true label in classification tasks. It is a combination of Softmax and negative log-likelihood loss (NLLLoss), and the formula is:

[0087] ;

[0088] z i Let C be the original output score of the i-th category, and C be the total number of categories. In this application, C=2, corresponding to LOS and NLOS. Sum the index scores for all categories, ensuring normalization.

[0089] ;

[0090] Among multiple batches, the loss function is its average value, and the formula is:

[0091] ;

[0092] N is the batch size, and i is the index of the sample in the batch.

[0093] After each training round, model performance is evaluated on the validation set, calculating metrics such as accuracy, precision, recall, and F1 score. The performance differences between the training and validation sets are analyzed to determine if the model is overfitting or underfitting. Hyperparameters, including learning rate, batch size, and dropout ratio, are adjusted based on the validation results. Loss and accuracy curves during training are visualized to aid in analyzing model convergence. Training is repeated using multiple sets of data from different sources, and the parameter results are saved.

[0094] The loss for error correction is referenced to the line-of-sight classification, and the cross-entropy loss of the classification and the root mean square loss of the ranging part are weighted together:

[0095] ;

[0096] Among them cls loss For classification loss, reg loss It is the regression loss, where 'a' is an adjustable constant that can be adjusted within the range of zero to one.

[0097] The classification loss function and the regression loss function used are as follows:

[0098] ;

[0099] ;

[0100] Where N is the batch size, i is the index of the sample in the batch, and Z is the index of the sample in the batch. i This represents the result value for each batch.

[0101] During classification training, model performance is evaluated using metrics such as accuracy, precision, recall, and F1 score. Furthermore, the performance differences between the training and validation sets are analyzed, and hyperparameters, including learning rate, batch size, and dropout ratio, are adjusted based on the validation results. The calculation formulas used are as follows:

[0102] ;

[0103] ;

[0104] ;

[0105] ;

[0106] TP represents a true positive instance, meaning the true value is positive and the predicted value is positive; TN represents a true negative instance, meaning the true value is negative and the predicted value is negative; FP represents a false positive instance, meaning the true value is negative but the predicted value is positive; and FN represents a false negative instance, meaning the true value is positive but the predicted value is negative.

[0107] Furthermore, the Roc curve is used to demonstrate the model's classification ability. It is a tool for evaluating the performance of binary classification models, calculated by weighing the true positive rate (TPR) and false positive rate (FPR).

[0108] ;

[0109] ;

[0110] The evaluation methods used for ranging error correction include MSE, RMSE, and R², and their calculation methods are as follows:

[0111] ;

[0112] ;

[0113] ;

[0114] Where N is the batch number and i is the sequential index. For the true value, For predicted values, This is the average value.

[0115] Both the classification and regression environments used were configured with Windows system, Python 3.9 language, Intel i7 14650 processor, RTX 4060 GPU, and CUDA 11.3.

[0116] The accuracy of the classification model used in this invention is shown below:

[0117]

[0118] This invention also provides a non-line-of-sight classification and error compensation system based on a dual convolutional neural network, used to implement the aforementioned non-line-of-sight classification and error compensation method based on a dual convolutional neural network, the system comprising:

[0119] The data acquisition module is used to acquire the channel impulse response (CIR) data of the ultra-wideband (UWB) signal and preprocess it to obtain the amplitude sequence in real number form.

[0120] The model building module is used to build a dual convolutional neural network classification model, which includes parallel large convolutional kernel convolutional blocks and small convolutional kernel convolutional blocks, respectively used to extract macroscopic waveform features and microscopic fluctuation features from the amplitude sequence;

[0121] The training module is used to perform classification training by fusing macroscopic waveform features and microscopic fluctuation features;

[0122] The compensation module is used to perform regression compensation for ranging errors of non-line-of-sight signals based on the trained classification model, and output the corrected distance value.

[0123] The proposed dual-convolutional block structure can simultaneously extract macroscopic waveform contour features and microscopic local fluctuation features from CIR data. This complementary feature extraction mechanism can more comprehensively describe the differences between LOS / NLOS signals compared to single-structure networks. The introduced channel attention mechanism further enhances the contribution of important feature channels and suppresses redundant information. Experimental results show that the proposed model outperforms traditional machine learning methods and basic neural network models in terms of classification accuracy, precision, recall, and F1 score on multiple public and privately collected datasets.

[0124] This invention takes high-precision classification as a prerequisite for error correction. Different correction strategies are adopted for different propagation paths. In particular, for the NLOS path, a deep learning model is used to directly learn the complex mapping relationship of error from CIR features, rather than using a simple fixed value or linear model, thereby enabling more accurate estimation and compensation of NLOS error.

[0125] This invention integrates multi-source data from different environments and devices during the model training phase, enhancing the model's adaptability to unknown environments and devices and ensuring the robustness of the method in actual deployment.

[0126] This invention provides a complete solution from raw CIR data processing to final high-precision distance estimation, with a clear process that is easy to integrate into existing UWB positioning systems.

[0127] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-line-of-sight classification and error compensation method based on dual convolutional neural networks, characterized in that, The method includes: The channel impulse response (CIR) data of the ultra-wideband (UWB) signal is acquired and preprocessed to obtain a real-valued amplitude sequence. A dual convolutional neural network classification model is constructed, which includes parallel large convolutional kernel convolutional blocks and small convolutional kernel convolutional blocks, respectively used to extract macroscopic waveform features and microscopic fluctuation features from the amplitude sequence; The macroscopic waveform features and microscopic fluctuation features are fused together for classification training; Based on the trained classification model, regression compensation for ranging error is performed on non-line-of-sight signals, and the corrected distance value is output. The error compensation step is specifically as follows: Freeze the feature extraction parameters in the trained classification model; A regression subnetwork consisting of fully connected layers is connected after the feature fusion layer; The regression subnetwork is trained using a combined loss function, which is a weighted sum of the classification loss function and the regression loss function.

2. The non-line-of-sight classification and error compensation method based on dual convolutional neural networks according to claim 1, characterized in that, The large convolutional kernel convolutional block includes two layers of 7x1 large convolutional kernel convolutional layers, which are downsampled based on max pooling layers; The small convolutional kernel convolutional block includes a two-layer 3x1 small convolutional kernel convolutional layer, which is downsampled based on an average pooling layer.

3. The non-line-of-sight classification and error compensation method based on dual convolutional neural networks according to claim 1, characterized in that, The large convolutional kernel block specifically uses 15x1 and 11x1 convolutional kernels connected in series; the small convolutional kernel block specifically uses 7x1 and 5x1 convolutional kernels connected in series.

4. The non-line-of-sight classification and error compensation method based on dual convolutional neural networks according to claim 3, characterized in that, In the process of fusing macroscopic waveform features and microscopic fluctuation features, a channel attention module is also set up to perform weighted processing on the feature channels.

5. The non-line-of-sight classification and error compensation method based on dual convolutional neural networks according to claim 1, characterized in that, The combined loss function is: ； in, For cross-entropy loss, denoted as root mean square error loss, where a is an adjustable weighting coefficient between 0 and 1.

6. A non-line-of-sight classification and error compensation system based on a dual convolutional neural network, used to implement the non-line-of-sight classification and error compensation method based on a dual convolutional neural network as described in any one of claims 1-5, characterized in that, The system includes: The data acquisition module is used to acquire the channel impulse response (CIR) data of the ultra-wideband (UWB) signal and preprocess it to obtain the amplitude sequence in real number form. The model building module is used to build a dual convolutional neural network classification model, which includes parallel large convolutional kernel convolutional blocks and small convolutional kernel convolutional blocks, respectively used to extract macroscopic waveform features and microscopic fluctuation features from the amplitude sequence; The training module is used to perform classification training by fusing macroscopic waveform features and microscopic fluctuation features; The compensation module is used to perform regression compensation for ranging errors of non-line-of-sight signals based on the trained classification model, and output the corrected distance value.

Images