UWB non-line-of-sight signal identification method based on lightweight neural network

Abstract

The invention discloses a lightweight ultra-wideband non-line-of-sight signal identification neural network, and belongs to the field of signal identification by neural networks. The method is characterized in that a portable lightweight neural network for ultra-wideband non-line-of-sight signals is designed, a DWM1000 module is adopted to achieve signal sending and receiving, and a data set is collected in a complex real environment; in the aspect of neural network design, input data is converted into two-dimensional image data from a long sequence, a Dropout layer is added into a convolutional network to reduce overfitting, a gating attention mechanism is introduced to improve accuracy, and compression is performed by using an X-CUBE-AI import model and resources occupied during transplantation are calculated.

Description

Technical Field

[0001] The present invention relates to the field of ultra-wideband non-line-of-sight signal recognition using neural networks, and in particular to signal recognition under resource-constrained conditions. Background Art

[0002] With the rapid development of positioning and navigation technology, people's demand for precise indoor positioning is gradually increasing. Ultra-wideband signals are widely used in indoor positioning due to their wide spectrum, low power consumption and short pulse characteristics. However, due to the complex indoor environment, the direct path is often blocked, resulting in large errors in the final positioning results. Therefore, identifying non-line-of-sight signals is particularly important for achieving high-precision positioning.

[0003] Traditional recognition methods often rely on manual feature extraction, with low utilization of big data and therefore poor generalization ability. The use of neural network methods can make up for this shortcoming, but neural networks often require a large amount of sample data for training and occupy a lot of computing resources. Therefore, it is necessary to design a lightweight neural network to ensure the accuracy of recognition while requiring a simpler model and fewer parameters. This is more suitable for low-cost ultra-wideband positioning systems, and can improve recognition speed, which is conducive to real-time positioning and navigation research.

[0004] Based on the ultra-wideband positioning platform, the present invention designs a lightweight neural network for identifying ultra-wideband non-line-of-sight signals, while taking into account the recognition accuracy and the total number of parameters after training. It proposes to transform the one-dimensional long sequence channel impulse response feature input into a two-dimensional image-like input, introduces a gated attention mechanism into the convolutional neural network, and optimizes the network structure to ensure the recognition accuracy while being deployable. Summary of the invention

[0005] Most of the existing ultra-wideband non-line-of-sight signal recognition are relatively complex and not suitable for resource-constrained conditions. To achieve the above purpose, the present invention designs an ultra-wideband non-line-of-sight recognition lightweight neural network based on channel impulse response (CIR) data.

[0006] The technical solution adopted by the present invention is as follows:

[0007] Step 1: By using the DWM1000 module as the transmitting and receiving module of the ultra-wideband signal, design a program to realize the communication between the base station and the tag, and collect the CIR data of the received signal, and mark it according to the actual collection situation.

[0008] Step 2: The collected 1016-bit one-dimensional long sequence data is expanded to 1024 bits using the one-dimensional linear interpolation method, and then converted into 32*32 image-like data. The public dataset of the eWINE project of the European Horizon 2020 program is used to expand the dataset.

[0009] Step 3: Build a neural network model consisting of a convolutional layer, a batch normalization layer, a dropout layer, a fully connected layer, and an attention mechanism.

[0010] Step 4: Use the transformed data set in step 2 as input to train the neural network, optimize the network structure and parameters, determine the use of the gated attention mechanism, and obtain the trained model.

[0011] Step 5: Use X-CUBE-AI in STM32CubeMX to further optimize and compress the network model trained in step 4, and calculate the final size of Flash and Ram occupied by the model.

[0012] Furthermore, in step 2, if we want to insert the value f(x) at position x, we can calculate the one-dimensional linear interpolation formula based on the values ​​of f(x1) and f(x0) before and after it:

[0013]

[0014] Furthermore, the gated attention mechanism in step 4 realizes the attention to the important features in each step through two dense layer calculations. The formula of the gated attention mechanism is:

[0015] A t =σ(W q Q t +W k K t +W v V t +b)C

[0016] Where At represents the attention weight at time step t. σ is a nonlinear activation function. Wq, Wk, Wv, and b are learnable parameters. Qt, Kt, and Vt are the dense representations of query, key, and value, respectively. C is the context vector output by the encoder.

[0017] The beneficial effects of the present invention are as follows:

[0018] 1. The present invention collects data sets in a real environment to improve the practicability, accuracy and reliability of the model.

[0019] 2. The present invention converts a one-dimensional long sequence into image-like data input, thereby increasing the training and recognition speed of the model and improving the accuracy of the model.

[0020] 3. The present invention introduces a gated attention mechanism into the network, so that the model can dynamically adjust its attention to the data, improve the accuracy, and add a Dropout layer to reduce overfitting.

[0021] 4. The present invention uses X-CUBE-AI to load and compress the trained model, reduce the total number of model parameters, and calculate the occupied resources, proving the feasibility of deployment under limited resource conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0023] Figure 1 It is the signal acquisition hardware and program block diagram.

[0024] Figure 2 It is an analysis diagram of line-of-sight and non-line-of-sight signal collection environment.

[0025] Figure 3 It is to design a lightweight neural network structure diagram.

[0026] Figure 4 This is the final network resource usage when using X-CUBE-AI. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0029] like Figure 1 As shown in the figure, (a) is the base station used for data set collection, which mainly realizes the receiving function of ultra-wideband signal, and determines whether the received signal is complete after receiving the signal. If the received signal is good, the CIR data of the received signal is read and uploaded to the host computer through UDP. The flowchart of the base station function is shown on the left side of (c). (b) is a tag, which is powered by a lithium battery for easy mobility and realizes the sending function of ultra-wideband signal. The flowchart of the tag function is shown on the right side of (c). Since the CIR data is stored in the accumulator of DWM1000, it is 4064 bytes, and each four bytes is one bit, which are the lower eight bits of the real part, the upper eight bits of the real part, the lower eight bits of the imaginary part, and the upper eight bits of the imaginary part. Since the program crashes when reading once, it is read multiple times in a loop. Due to the internal memory access delay when reading the accumulator, the first byte output is a virtual byte that should be discarded, and finally communication is realized and the complete data set is collected. The number of CIR data samples is 1016, which is expanded to 1024 by a one-dimensional linear interpolation method to convert it into 32*32 image-like data.

[0030] The open source dataset is a public dataset of the European Horizon 2020 eWINE project. The dataset also uses the DWM1000 RF transceiver to collect indoor UWB signals. The dataset collects a total of 42,000 data, including 21,000 LOS data and 21,000 NLOS data. The dataset collected in the real environment is relatively small, about 5,000, and the dataset will continue to be collected in the future to make it suitable for the real environment.

[0031] like Figure 2 As shown, the signal transmission environment in a complex environment is mainly divided into three types, (a) is the line-of-sight situation, there is a direct path between the base station and the tag, and the direct path is not blocked. In this case, the collected channel impulse response is often an ideal single peak, but due to the complex surrounding environment, there are often multiple propagation paths, resulting in a multipath effect, which is reflected in the channel impulse response. There will be multiple peaks with peak lags and lower peaks, which has little effect on positioning accuracy. (b) is the non-line-of-sight multipath condition. In this case, the direct path is blocked by obstacles, but there is a clear reflection path. Both the direct path and the reflection path will cause peak lags and energy loss. (c) is the non-line-of-sight condition, the direct path is blocked and there is no direct reflection path. Non-line-of-sight multipath and non-line-of-sight both have a great impact on the positioning results, so the present invention mainly distinguishes between non-line-of-sight multipath and signals collected under line-of-sight conditions.

[0032] like Figure 3 As shown in the figure, it is a schematic diagram of the lightweight neural network model designed by the present invention. In the entire neural network, the input is a 32*32 two-dimensional matrix, which passes through three groups of convolutional layers, batch normalization layers and Dropout layers. After global average pooling, a gated attention mechanism layer is introduced for flattening, and finally the classification is output through a fully connected layer. The CNN convolutional neural network is combined with the gated attention mechanism. The gating mechanism is used to regulate the transmission and retention of information to prevent gradient problems, thereby more effectively capturing information in the sequence. The attention mechanism can automatically focus on more important information in the sequence, so that the model can more accurately understand and process important information in the input data.

[0033] Import open source data sets and collected data sets, and compare the training and testing of one-dimensional neural networks and two-dimensional neural networks. After training, the total number of parameters of the one-dimensional neural network is 125.77K, but the accuracy is low at about 83.72%. Although the total number of parameters of the two-dimensional neural network has increased to 366.52K, the accuracy has increased by about 4%. At the same time, the two-dimensional convolutional network also has a greater advantage in training and recognition speed. Compare neural networks with different numbers of layers and structures, and optimize the training parameters. Finally, the three-layer convolutional neural network plus the gated attention mechanism has the best effect. The accuracy and F1 evaluation parameters are close to 90%, and the total number of parameters of the uncompressed model is 495.52K. The specific comparison results are shown in Table 1.

[0034] Table 1 Comparison of network experimental results

[0035]

[0036] like Figure 4 As shown in the figure, the memory usage of the neural network model is 36KiB at the convolution layer. Finally, the trained model is imported and compressed using X-CUBE-AI. The resource usage is calculated as shown in Table 2. The network itself occupies 36.00KiB of Ram and 380.13KiB of Flash respectively. At the same time, its computational complexity is 2485750MACC, indicating that 2485750 multiplication and accumulation operations are required. At the same time, the library file also occupies certain resources. Finally, the lightweight model occupies a total of 397.56KiB of Flash, 40.14KiB of RAM, and a computational complexity of 2485750MACC. This lightweight neural network is feasible for deployment in embedded systems.

[0037] Table 2 Model resource usage

[0038]

[0039] Example 1

[0040] The method of the present invention has been verified by experiments to have good effects in ultra-wideband non-line-of-sight signal recognition. The experiment uses 42,000 open source data sets of the European Horizon 2020 eWINE project and real scene acquisition data sets. By converting the input into image-like data, all data sets are divided into training sets, validation sets and test sets in a ratio of 7:2:1 and input into the neural network for training. In the experiment, the learning rate is set to 0.01, the Dropout layer coefficient is set to 0.3, the data batch size is set to 32, and the number of training times is set to 65. After multiple experiments, the recognition accuracy and the average value of the F1 evaluation parameters of the model are both above 88%. The trained model is compressed, and the final total parameter amount is reduced from 495.52K to 397.56K.

Claims

1. A UWB non-line-of-sight signal recognition method based on a lightweight neural network, characterized in that: By adding data sets collected under real conditions for training, it is suitable for ultra-wideband non-line-of-sight signal recognition in complex indoor environments. The model recognition accuracy and speed are improved by changing the input format of the neural network training data set, adding a Dropout layer, and introducing a gated attention mechanism. By using X-CUBE-AI, a tool that can quickly implement AI applications on STM32 devices, the model is compressed and optimized, and the calculation resources are reduced.

2. The data set collection in real scenarios according to claim 1, characterized in that: Base stations are placed in real complex indoor positioning scenarios, and mobile tags are used to collect data sets. The data sets are annotated according to real scenarios as a supplement to the neural network training data set. At the same time, experimental equipment is used to collect data for verification.

3. The method for changing the format of a data set according to claim 1, characterized in that: The collected 1016-bit one-dimensional long sequence data is expanded to 1024 bits using the one-dimensional linear interpolation method. The calculation formula is: The data is then converted into 32*32 image-like data as input for neural network training.

4. The Dropout layer and gated attention mechanism according to claim 1, characterized in that: Dropout reduces overfitting during training by randomly blocking some neurons. The gated attention mechanism can focus on important information in each step. The calculation formula is: A t =σ(W q Q t +W k K t +W v V t +b)C Where At represents the attention weight at time step t. σ is a nonlinear activation function, and Wq, Wk, Wv, and b are learnable parameters. Qt, Kt, and Vt are dense representations of query, key, and value, respectively. C is the context vector output by the encoder.

5. The method for optimizing and compressing a model according to claim 1, characterized in that: The X-CUBE-AI 8.1.0 version was used to import the final trained model into the STM32CubeMX project, and the model was optimized and compressed to prove its feasibility of deployment in STM32.

Images