UWB ranging error compensation method based on deep learning

Abstract

The invention provides a UWB ranging error compensation method based on deep learning. The method mainly comprises the steps that through data preprocessing, feature extraction is conducted on CIR based on a deep learning method, multi-scale convolution and an SE attention mechanism are combined, richer information is captured, global features can be paid attention to, local details can be carefully captured, an error compensation model is provided to predict errors generated by non-line-of-sight propagation, and therefore the accuracy of non-line-of-sight propagation is improved. And the distance measurement value is corrected, so that the positioning precision is improved, and errors caused by non-line-of-sight propagation are suppressed. Compared with a traditional UWB ranging error compensation method, the UWB ranging error compensation method has the advantages that the ranging value is corrected by predicting the error generated by non-line-of-sight propagation, the positioning precision is improved while the positioning cost of the UWB system is reduced, the number of base stations does not need to be increased, and the position of the mobile tag can be calculated all the time to compensate the error.

Description

Technical Field

[0001] The present invention is a method for compensating UWB ranging error based on deep learning. Background Art

[0002] With the continuous progress of society and the rapid development of science and technology, new technologies such as artificial intelligence and big data have been studied and applied more and more widely in the industrial field, and countries around the world are actively promoting the continuous innovation of industrial technologies. In the current era of rapid technological development, positioning and navigation technologies play a crucial role in many fields. Whether it is the precise driving of vehicles in intelligent transportation systems, the realization of autonomous driving technology, or the positioning and tracking of people and equipment in indoor scenarios such as large shopping malls and intelligent buildings, all highly rely on accurate and reliable positioning information.

[0003] Currently, the most widely used technologies in the field of vehicle navigation and positioning are the Global Navigation Satellite System (GNSS), the Inertial Navigation System (INS), and various integrated navigation technologies. However, once GNSS enters a denied environment and is affected by multipath and non-line-of-sight effects, such as the severe occlusion of satellite signals by tall buildings in urban canyons, the complete inability of satellite signals to penetrate inside tunnels, and the isolation of satellite signal reception in enclosed spaces such as indoor parking lots, GNSS will fail, resulting in a sharp decline in positioning accuracy and being unable to meet the actual application requirements. Ultra-wideband (UWB) technology has received extensive attention due to its low-power processing ability and wide bandwidth that provides high data rates. It can overcome the influence of complex environments to a certain extent and provide relatively accurate relative position information. However, as is well known, in wireless communication, due to various environmental limitations such as buildings, trees, and hills, the signal will undergo deep fading, so that the signal cannot reach the receiving end or may delay the received signal, resulting in problems such as multipath effects, noise interference, and non-line-of-sight (NLOS) propagation, leading to fluctuations or even a significant decline in positioning accuracy.

[0004] Based on the feature extraction of the CIR through deep learning, the present invention designs and constructs an error compensation model to correct the ranging value, thereby improving the positioning accuracy. This method does not require an increase in the number of base stations and can always calculate the position of the mobile tag. Therefore, the present invention adopts this method to suppress the error caused by non-line-of-sight propagation. Summary of the Invention

[0005] The following combines the accompanying drawings to specifically describe the technical solution of the present invention, and designs a method for compensating UWB ranging error based on deep learning. Figure 1 For the overall flowchart, it specifically includes the following steps:

[0006] Step 1: First, train and test on a public dataset;

[0007] Step 2: Preprocess the data and record the channel impulse response (CIR) data through precise instrument measurement;

[0008] Step 3: Based on the multi-scale convolution block and the one-dimensional SE block, the present invention proposes a UWB error suppression model;

[0009] Step 4: The present invention divides the data set collected by environment 1 under the public data set into a training data set and a verification data set in a ratio of 8:2, selects the mean square error function as the loss function of the experiment, and adopts the Adam optimization algorithm as the optimizer of the experiment;

[0010] Step 5: Evaluate the ranging error suppression effect. Input the UWB ranging data set to be identified (UWB ranging data set collected in other indoor environments) into the trained experimental model to evaluate the ranging error suppression effect;

[0011] Furthermore, in the process of step 1, a total of four experimental data in typical indoor scenes are included, including apartment, living room, factory and office. In order to facilitate research, the present invention selects the apartment data set as the training and verification data set, and uses the data set in the living room scene as the test data set.

[0012] Furthermore, in the process of step 2, the DW1000 UWB chip and Alt-TWR algorithm that comply with the IEEE 802.15.4 standard are used to implement UWB ranging, and the three-dimensional coordinates of the base station and the tag are obtained by a high-precision laser rangefinder with millimeter-level ranging accuracy.

[0013] A total of eight fixed base stations and one mobile tag were deployed in the experimental scenario. The movement trajectory of the mobile tag was pre-planned, and multiple sampling points were set on the trajectory according to the principle of equal spacing. At each sampling point, the tag performed 31 ranging measurements with the eight base stations, and accurately recorded the channel impulse response (CIR) data. It is worth mentioning that this dataset is not limited to a specific channel, but the above operations are repeated for all channels supported by DW1000, thus ensuring the wide coverage and depth of the dataset, which is helpful for the training and testing of complex models.

[0014] Further, in the process of the second step, in the context of channel complex number estimation in DW1000 technology, aiming to optimize model parameters and accelerate the model training process, the present invention adopts a data preprocessing strategy consistent with the literature "Transfer Learning for UWB Error Correction and (N)LOS Classification in Multiple Environments".

[0015] First, in order to obtain more intuitive time-domain information, the CIR estimator from the complex domain is converted into a time-domain representation. This conversion process is achieved by calculating the modulus (or absolute value) of the complex number, thereby effectively extracting the real CIR reflecting the signal strength from the complex CIR containing amplitude and phase information.

[0016] The above formula is the formula for conversion from the complex domain to the time domain, where I is the imaginary part value of the CIR, R is the real part value of the CIR, and RSSI is the channel gain strength in the time domain.

[0017] Further, according to the third step, the present invention takes measures to predict the error caused by non-line-of-sight propagation and correct the ranging value, thereby improving the positioning accuracy. This method does not require increasing the number of base stations and can always calculate the position of the mobile tag to suppress the error caused by non-line-of-sight propagation.

[0018] Further, according to the third step, the CIR data studied in the present invention is a one-dimensional sequence, so a multi-scale convolution block is proposed. By performing convolutions of different sizes on the input data for feature extraction in different dimensions, richer feature information can be obtained.

[0019] Through the padding operation, the scale of the output feature map is kept consistent, and then they are input together into the convolution layer with the same-sized convolution kernel in the next layer for feature fusion.

[0020] Then a one-dimensional squeeze-and-excitation block is designed, as Figure 2 shown, and the modeling of this non-linear relationship is realized by using a two-layer fully connected network.

[0021] Further, according to the third step, a model is established as Figure 3 , which is mainly composed of an encoder Encoder and a decoder Decoder. Among them, the Encoder is responsible for feature extraction of the CIR, and the Decoder is responsible for establishing a non-linear model from features to errors.

[0022] Further, in Step 4, the model is trained. In the present invention, the dataset collected in Environment 1 under the public dataset is divided into a training dataset and a validation dataset according to a ratio of 8:2.

[0023] Obtaining the ranging error value from the CIR training belongs to a regression problem. The mean squared error function is selected as the loss function for the experiment, and the Adam optimization algorithm is adopted as the optimizer for the experiment. The batchsize selected for the experiment is 32, the learning rate is 0.001, and the number of training epochs is set to 20. As Figure 4 is the result after training.

[0024] Evaluate the ranging error suppression effect. Input the UWB ranging dataset to be recognized (the UWB ranging dataset collected in other indoor environments) into the trained experimental model of this experiment to evaluate the ranging error suppression effect. As Figure 5 shown, it is the experimental ranging error map after calibration.

[0025] Compared with the prior art, the present invention provides a UWB ranging error compensation method based on deep learning. Through data preprocessing, feature extraction of the CIR is performed based on the deep learning method, and thus an error compensation model is proposed to predict the error generated by non-line-of-sight propagation, and the ranging value is corrected, thereby improving the positioning accuracy and suppressing the error caused by non-line-of-sight propagation. Compared with the traditional UWB ranging error compensation method, the solution of the present invention adopts multi-scale convolution and SE attention mechanism to improve the performance of target detection and recognition tasks, enabling the model to pay more attention to useful channel information, thereby improving the performance and accuracy of the model. By predicting the error generated by non-line-of-sight propagation and correcting the ranging value, it can effectively capture the influence of various factors in the environment on the ranging result, reduce the positioning cost of the UWB system while improving the positioning accuracy. This method does not require an increase in the number of base stations and can always solve the position of the mobile tag to compensate for the error. Its deep learning model can self-adjust according to different environmental conditions (such as multipath interference, object blocking, etc.) and has strong adaptability, achieving better effects and quality in actual measurement and positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is the flowchart of the present invention

[0027] Figure 2 is the SE block structure diagram

[0028] Figure 3 is the error compensation model block diagram

[0029] Figure 4 is the loss function descent diagram

[0030] Figure 5 is the ranging error CDF diagram in the scenario of Channel 3

[0031] Figure 6 It is the CDF graph of ranging error in the channel 7 scenario

[0032] Figure 7 It is the CDF graph of ranging error in the channel 3 scenario

[0033] Figure 8 It is the CDF graph of ranging error in the channel 7 scenario Detailed implementation manners

[0034] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the technical solutions of the present invention will be specifically described below in conjunction with the accompanying drawings and examples.

[0035] Step 1: First, training and testing are carried out under a certain public dataset. It includes experimental data in four typical indoor scenarios, including apartments, living rooms, factories and offices. For the convenience of research, the dataset of the apartment is selected as the training and validation dataset in this paper, and the dataset in the living room scenario is used as the test dataset.

[0036] Step 2: The DW1000 UWB chip compliant with the IEEE 802.15.4 standard and the Alt-TWR algorithm are adopted to realize UWB ranging, and the three-dimensional coordinates of the base station and the tag are obtained by a high-precision laser rangefinder with millimeter-level ranging accuracy.

[0037] In order to optimize the model parameters and improve the training efficiency, the key area near the first path is intercepted (including 50 sampling points before the first path and 100 sampling points after it, a total of 150 sampling points), and the CIR with a length of 1016 sampling points is shortened to 150 sampling points in length, effectively reducing the dimension and redundant information of the input data.

[0038] Subsequently, the maximum-minimum normalization method is adopted to standardize the CIR data to ensure the numerical stability and consistency of the data. The formula for maximum-minimum normalization is as follows:

[0039] Step 3: In terms of model construction, based on multi-scale convolutional blocks and one-dimensional SE blocks, a UWB error suppression model is proposed. The model mainly consists of an encoder Encoder and a decoder Decoder, where the Encoder is responsible for feature extraction of the CIR, and the Decoder is responsible for establishing a non-linear model from features to errors.

[0040] Process the CIR data with one-dimensional convolution operations.

[0041] Based on traditional convolution operations, a multi-scale convolution block is proposed. By performing convolutions of different sizes on the input data for feature extraction in different dimensions, richer feature information can be obtained. Through padding operations, the scales of the output feature maps are kept consistent, and then they are input together into the convolution layer with the same-sized convolution kernels in the next layer for feature fusion.

[0042] During the process of feature extraction from the input data through multiple convolution kernels, a set of feature outputs corresponding to the number of convolution kernels will be generated. Given that the feature vectors in different channels often carry different degrees of important information, these features are processed differentially.

[0043] The information aggregation is achieved through the adopted adaptive global average pooling.

[0044] The modeling of this non-linear relationship is realized using a two-layer fully connected network, as Figure 2 shown. Its specific operation expression is as follows: F ex (X) = σ(W2δ(W1X))

[0045] Among them, δ(x) represents the relu activation function, σ(x) represents the sigmoid activation function, and in order to meet the requirement of learning the relative importance among its channels with a certain complexity, this two-layer fully connected network is generally constructed in a way of first increasing the dimension and then decreasing the dimension.

[0046] The error compensation model block diagram model is established as Figure 3 shown. Based on the multi-scale convolution block and the one-dimensional SE block, a UWB error suppression model is proposed. This model is mainly composed of an encoder Encoder and a decoder Decoder. Among them, the Encoder is responsible for feature extraction of the CIR, while the Decoder is responsible for establishing a non-linear model from features to errors.

[0047] The Encoder consists of several Encoder layers. Using multiple Encoder layers can fully extract deeper feature information. Each Encoder layer can generally be composed of three parts, namely the multi-scale convolution feature extraction layer, the information fusion module, and the one-dimensional SE block.

[0048] Among them, the multi-scale convolution feature extraction layer extracts features in different dimensions from the CIR by using multiple convolution kernels of different sizes, and can obtain richer experimental results compared with the traditional method of using a single-sized convolution kernel.

[0049] The information fusion module first concatenates the outputs of the multi-scale convolutional layers, then fuses the features through multiple convolutional kernels of size 1*3, and normalizes them through batchnorm before outputting to the relu activation function, thus obtaining the fused feature map.

[0050] The final one-dimensional SE block selectively enhances the features of the fused feature map from the channel perspective, thereby enhancing the feature extraction ability of the model and facilitating subsequent learning.

[0051] The Decoder part consists of a fully connected network, which first performs a flattening operation on the feature map output by the Encoder.

[0052] Subsequently, the feature data will sequentially pass through two fully connected layers (FC), and each FC layer is followed by a ReLU activation function and a Dropout operation to aggregate and refine the features.

[0053] Then, the features will pass through another FC layer and a sigmoid function, which aims to further compress the feature information. Finally, through a fully connected layer (FC), the network will output a ranging error value Error.

[0054] Step 4: Train the model. In this paper, the dataset collected in Environment 1 under the public dataset is divided into a training dataset and a validation dataset according to a ratio of 8:2. Obtaining the ranging error value from CIR training belongs to a regression problem. In this paper, the mean squared error function is selected as the loss function for the experiment, and the Adam optimization algorithm is adopted as the optimizer for the experiment. The batchsize selected in the experiment in this paper is 32, the learning rate is 0.001, and the number of training epochs is set to 20.

[0055] First, the loss function MSE of the mean squared error is introduced, and its formula is as follows:

[0056] MSE can reflect the training effect of the model by squaring and then averaging the errors of the model, as Figure 4 shown.

[0057] The probability density cumulative graphs CDF of the original error and the corrected error on the test dataset are respectively plotted, and the results are as follows Figure 5 shown.

[0058] Step 5: Evaluate the ranging error suppression effect. The model obtained by training is tested with the data collected under different channels in Environment 2. The root mean square error is introduced as an index, and the cumulative distribution function is used to visualize the original ranging error and the corrected ranging error.

[0059] According to the final experimental results, it can be found that when DW1000 selects different channels, there will be significant differences in the MSE and RMSE of its original ranging values. However, under the action of the UWB error suppression model proposed in this paper, the ranging values after calibration have a significant decrease in both MSE and RMSE indicators, indicating that the model has an overall good error suppression effect on the test dataset.

[0060] To demonstrate the actual ranging error suppression effect of the model, the following takes Channel 3 (the channel corresponding to the largest RMSE after calibration) and Channel 7 (the channel corresponding to the smallest RMSE after calibration) as examples to show how the CDF of the ranging error will change after the ranging error is suppressed by the model, as Figure 5 , Figure 6 shown.

[0061] It is found through observation that although the ranging values have a large RMSE and MSE under Channel 3, 80% of the ranging errors of the ranging values are still lower than 20 cm, which is a 69% improvement compared to the original 65 cm. The model has greatly improved the ranging accuracy.

[0062] Under Channel 7, 80% of the ranging errors of the ranging values are lower than 10 cm, which is an 84% improvement compared to the original 65 cm, and 90% of the ranging value errors are lower than 20 cm, greatly suppressing the NLOS interference and achieving high-precision ranging.

[0063] Analyze the improvement of the ranging accuracy on the positioning accuracy. The following also takes Channel 3 and Channel 7 as examples for subsequent analysis.

[0064] Through Figure 7 it can be known that in the experimental scenario of Channel 3, for the experimental results of positioning using the calibrated ranging values, the average error is reduced from 0.29 m to 0.16 m. And by analyzing the error bars at each point, it is found that although the error increases slightly at some points, overall, the positioning accuracy is effectively improved, and the positioning variance at each point becomes smaller.

[0065] Through Figure 8 it can be known that in the experimental scenario of Channel 7, for the experimental results of positioning using the calibrated ranging values, the average error is reduced from 0.23 m to 0.09 m. And observing the positioning error at each point effectively decreases, and the positioning results are more stable. Through the above two test experiments, it is proved that the error suppression model proposed in this paper can effectively improve the positioning accuracy by reducing the ranging error.

[0066] The above embodiments merely illustrate the principles and effects of the present invention, rather than limiting the present invention. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes completed by those with ordinary knowledge in the technical field without departing from the spirit and technical idea disclosed by the present invention should still be covered by the claims of the present invention.

[0067] In summary, the present invention provides a UWB ranging error compensation method based on deep learning. Through data preprocessing and feature extraction of the channel impulse response (CIR), a model is developed to predict the error caused by non-line-of-sight propagation. The use of multi-scale convolution and SE attention mechanism improves the performance of target detection and recognition tasks, enabling the model to pay more attention to useful channel information, thereby improving the performance and accuracy of the model to correct the ranging value and improve the positioning accuracy. Compared with traditional UWB ranging error compensation methods, the present invention can effectively grasp the influence of environmental factors on the ranging result, reduce the positioning cost, and there is no need to add new base stations. It can calculate the position of the mobile tag in real time for error correction. Its deep learning model has strong adaptability and can self-adjust according to various environmental conditions (such as multipath interference and object blockage), so as to achieve better measurement and positioning effects.

Claims

1. A UWB error compensation model method based on deep learning, comprising the following steps: S1: This paper first conducts training and testing on a public dataset, which comes from the document Bregar, K. (2023) 'Indoor UWB positioning and position tracking data set', Scientific Data. It contains experimental data in four typical indoor scenes, including apartments, living rooms, factories, and offices. For the convenience of research, this paper selects the apartment dataset as the training and verification dataset, and the dataset in the living room scene as the test dataset. S2: Data preprocessing. The experiment used the DW1000 UWB chip and Alt-TWR algorithm that complies with the IEEE 802.15.4 standard to achieve UWB ranging, and the three-dimensional coordinates of the base station and the tag were obtained by a high-precision laser rangefinder with millimeter-level ranging accuracy. A total of eight fixed base stations and one mobile tag were deployed in the experimental scenario. The movement trajectory of the mobile tag was pre-planned, and multiple sampling points were set on the trajectory according to the principle of equal spacing. At each sampling point, the tag performed 31 ranging measurements with the eight base stations, and the channel impulse response (CIR) data was accurately recorded. It is worth mentioning that the data set is not limited to a specific channel, but the above operations are repeated for all channels supported by DW1000, thereby ensuring the wide coverage and depth of the data set, which is helpful for the training and testing of complex models. S3: In terms of model construction, this paper proposes a UWB error suppression model based on multi-scale convolutional blocks and one-dimensional SE blocks. The model is mainly composed of an encoder and a decoder. The encoder is responsible for feature extraction of CIR, while the decoder is responsible for building a nonlinear model from features to errors. S4: Training model. This paper divides the data set collected in environment 1 under the public data set into a training data set and a validation data set in a ratio of 8:

2. The ranging error value obtained from CIR training belongs to a regression problem. This paper selects the mean square error function as the loss function of the experiment and uses the Adam optimization algorithm as the optimizer of the experiment. The batch size selected in this experiment is 32, the learning rate is 0.001, and the number of training rounds is set to 20. S5: Evaluate the ranging error suppression effect. The UWB ranging data set to be identified (UWB ranging data set collected in other indoor environments) is input into the trained experimental model for testing, and the cumulative distribution function is used to visualize the original ranging error and the corrected ranging error to evaluate the ranging error suppression effect.

Images