DOA estimation method based on convolutional neural network and received signal strength
阅读说明：本技术 一种基于卷积神经网络和接收信号强度的doa估计方法 (DOA estimation method based on convolutional neural network and received signal strength ) 是由 蒋伊琳 张昊平 李向 于 2019-10-09 设计创作，主要内容包括：本发明属于阵列信号处理技术领域,具体涉及降低了各阵元间对采样频率和信道同步依赖的一种基于卷积神经网络和接收信号强度的DOA估计方法。本方法包括以下步骤：1.定义DOA的范围,传感器的数量K和位置,角度空间的分辨率和类别的总数M；2.根据每个传感器接收的RSS值构建K×1的RSS向量；3.将RSS向量归一化至0到1之间；4.根据归一化的RSS向量构建K×K的RSS图像；5.定义网络中的层和参数,生成各类别的训练样本进行模型训练；6.根据训练好的模型进行实际的DOA预测。本发明的DOA方法降低了各阵元间对采样频率和信道同步的依赖,适用于实现相对简单,复杂度较低,定位精度不需要很高的环境。此外,该方法还具有应用于均匀线阵等较一般的DOA场景的潜力。(The invention belongs to the technical field of array signal processing, and particularly relates to a DOA estimation method based on a convolutional neural network and received signal strength, which reduces the synchronous dependence of each array element on sampling frequency and a channel. The method comprises the following steps: 1. defining the range of DOA, the number K and the positions of sensors, the resolution of an angle space and the total number M of categories; 2. constructing a Kx 1 RSS vector according to the RSS value received by each sensor; 3. normalizing the RSS vector to between 0 and 1; 4. constructing a KxK RSS image according to the normalized RSS vector; 5. defining layers and parameters in a network, and generating training samples of various types for model training; 6. and (4) performing actual DOA prediction according to the trained model. The DOA method reduces the dependence of each array element on the sampling frequency and the channel synchronization, and is suitable for the environment with relatively simple realization, lower complexity and no need of high positioning precision. In addition, the method has the potential of being applied to the uniform linear array and other more common DOA scenes.)
1. A DOA estimation method based on a convolutional neural network and received signal strength is characterized by comprising the following steps:
step 1: defining the range of DOA, the number K and the positions of sensors, the resolution of an angle space and the total number M of categories;
step 2: constructing a Kx 1 RSS vector according to the RSS value received by each sensor;
and step 3: normalizing the RSS vector to between 0 and 1;
and 4, step 4: constructing a KxK RSS image according to the normalized RSS vector;
and 5: defining layers and parameters in a network, and generating training samples of various types for model training;
step 6: and (4) performing actual DOA prediction according to the trained model.
2. A method of estimating DOA based on convolutional neural network and received signal strength as claimed in claim 1, wherein the range of DOA in step 1 is [ -30 °, +30 ° ]]49 sensors are uniformly distributed at [ -45 °, +45 ° ]]Defines a resolution of 1 deg., the entire DOA range is divided into 61 categories, specifically denoted as Θ ═ { θ ═ 611...θMIn which θiIndicating the DOA corresponding to the ith category.
3. A convolution-based nerve according to claim 1DOA estimation method of network and received signal strength, characterized in that, the RSS vector in step 2 is marked as RθExpressed as the following formula:
Rθ＝[Pr1…PrK]T+n
wherein n is [ n ]1…nK]TIs noise on reception, substituted by white noise, theta represents a certain DOA, PriRepresents the RSS value received by the ith sensor, expressed as:
where Pt is the transmission power of the signal source, Gt and Gr are the gains of the transmitting antenna and the receiving antenna, diWhich is the distance between the signal source and the ith sensor, lambda represents the wavelength, the transmission frequency of the signal is set to 50MHz,
wherein
4. A method of DOA estimation based on convolutional neural network and received signal strength as claimed in claim 1, wherein the RSS vector normalization in step 3 can be expressed as:
wherein R isMaxAnd RMinRepresenting the maximum and minimum values in the RSS vector, respectively.
5. A DOA estimation method based on convolutional neural network and received signal strength as claimed in claim 1, wherein the RSS vector imaging in step 4 specifically is:
in the normalized RSS vector, 0 represents the weakest energy, and 1 represents the strongest energy; dividing the energy into K levels, if the normalized RSS value of the ith sensor belongs to the jth energy level, setting the element of the jth row and the ith column of the RSS image to be 1, setting the rest elements to be 0, and generating a pair of [ KxK ] RSS images according to the level of each element.
6. The DOA estimation method based on the convolutional neural network and the received signal strength as claimed in claim 1, wherein the model in step 5 is specifically: the model comprises two hidden layers and two full-connection layers, wherein each hidden layer is divided into three parts: a convolutional layer, a ReLU (transformed linear units) activation function and a max pooling layer; the two convolutional layers contain 64 and 128 convolutional kernels, respectively, with a size [3 × 3], the step size is set to 1, the activation function of the first fully-connected layer is ReLU and the total number of output neurons is 2048; the second full-connection layer adopts a Softmax activation function to calculate the posterior probability of each category; the Adam gradient is used as an optimizer during training, and the cost function adopts cross-entropy (cross-entropy); the weight of the model adopts a truncated positive-space distribution during initialization, and the bias value is initialized to be constant 0.1.
7. The DOA estimation method based on the convolutional neural network and the received signal strength as claimed in claim 1, wherein the step 6 is specifically: and (3) converting the RSS value actually measured by the sensor into an RSS image, then putting the RSS image into the model trained in the step (5), and predicting the DOA according to the category with the highest probability in the output result of the CNN.
Technical Field
The invention belongs to the technical field of array signal processing, and particularly relates to a DOA estimation method based on a convolutional neural network and received signal strength, which reduces the synchronous dependence of each array element on sampling frequency and a channel.
Background
In recent years, the accuracy and super-resolution of DOA estimation are improved, and the adaptability to high-requirement scenes such as limited snapshots, low signal-to-noise ratios and the like is enhanced, so that the DOA estimation method is a main trend of DOA estimation research. Common DOA estimation methods such as MUSIC, ESPRIT, etc. all contribute significantly to high-precision positioning. However, in reviewing the features of these methods, we have readily discovered that there are some well-known limitations that remain. These phase-based methods require sufficient snapshots and accurate time synchronization, which will increase the burden on the hardware.
With the development of new theories and new technologies, the DOA estimation method presents more possibilities. DNN has now demonstrated its ability to perform array signal processing. After this, DNN also frequently occurs in the task of DOA estimation. Youssef Harkouss and Hassan Shraim et al (International Journal of RF and microwave computer-Aided Engineering, Vol.28, No. 6, 8, 2018, Direction of estimation for sample anti-purposes in multi-path environment using volumetric network) proposed a DOA estimation strategy based on CNN. The strategy is suitable for an intelligent antenna system in a multipath environment, the real part and the imaginary part of a covariance matrix are used for constructing input characteristics, and a bicubic interpolation method is adopted for converting the input characteristics into images. Liu Zhang-Meng and Zhang Chenwei et al (IEEE Transactions on Antennas and Propagation, volume 66, 12, 12.2018, Direction-of-Arrival Estimation based on deep Neural Networks with Robusting to Array improvements) propose a hierarchical framework based on DNN to solve the problem of lack of adaptability to various Array defects in DOA Estimation. But the input features of the network are still covariance matrices. Chakrabarty and E.A.P.Habets ("Broadband and DOA estimation using connected neural network networks with noise signals,"2017IEEE work kshopon Applications of Signal Processing to Audio and Acoustics (WASPAA), NewPaltz, NY,2017, pp.136-140.) propose a Broadband DOA estimation classification method based on CNN, and the phase component of the short-time Fourier transform (STFT) coefficient of the input Signal is directly used as the input characteristic of the neural network. The method has strong robustness to noise and small disturbance at the position of the microphone. In addition, Zhang Xueliang and Li Hao ("online direction of Arrival Estimation Based on Deep Learning,"2018ieee international Conference on Acoustics, Speech and Signal Processing (ICASSP), Calgary, AB,2018, pp.2616-2620) propose a method combining CNN and long-short term memory (LSTM) to solve the problem of online DOA Estimation in noisy and reverberant environments. A special generalized cross-correlation (GCC) is used as an input feature. In summary, a common feature of all DNN-based methods is that the extraction of input features still relies on sufficient snapshots and accurate phase synchronization, which is costly in hardware implementation. Especially when the sampling frequency is limited or multi-channel synchronization is required, such as an unmanned aerial vehicle platform, the extraction of the input features is greatly affected, and finally, the methods cannot easily achieve good DOA estimation effect.
Unlike the phase-based DOA estimation methods above, RSS-based methods utilize signal strength or amplitude information for estimation in a relatively simple and low hardware cost manner. The idea of DOA estimation by RSS is possible, since the antenna has a certain radiation pattern. This radiation model provides us with information on the variation of antenna directional gain with angle. Therefore, we can use this information to establish a non-linear mapping relationship with the DOA.
Disclosure of Invention
The invention aims to provide a DOA estimation method based on a convolutional neural network and received signal strength.
The purpose of the invention is realized as follows:
a DOA estimation method based on a convolutional neural network and received signal strength comprises the following steps:
step 1: defining the range of DOA, the number K and the positions of sensors, the resolution of an angle space and the total number M of categories;
step 2: constructing a Kx 1 RSS vector according to the RSS value received by each sensor;
and step 3: normalizing the RSS vector to between 0 and 1;
and 4, step 4: constructing a KxK RSS image according to the normalized RSS vector;
and 5: defining layers and parameters in a network, and generating training samples of various types for model training;
step 6: and (4) performing actual DOA prediction according to the trained model.
The DOA range stated in step 1 is [ -30 °, +30 ° ]]49 sensors are uniformly distributed at [ -45 °, +45 ° ]]Defines a resolution of 1 deg., the entire DOA range is divided into 61 categories, specifically denoted as Θ ═ { θ ═ 611…θMIn which θiIndicating the DOA corresponding to the ith category.
Step 2, the RSS vector is marked as RθExpressed as the following formula:
Rθ＝[Pr1…PrK]T+n
wherein n is [ n ]1…nK]TIs noise on reception, substituted by white noise, theta represents a certain DOA, PriRepresents the RSS value received by the ith sensor, expressed as:
where Pt is the transmission power of the signal source, Gt and Gr are the gains of the transmitting antenna and the receiving antenna, diWhich is the distance between the signal source and the ith sensor, lambda represents the wavelength, the transmission frequency of the signal is set to 50MHz,is a two-dimensional far-field radiation model, taking gaussian antenna as an example, the radiation model with 0 ° main beam is expressed as the following formula:
wherein
Step 3 the RSS vector normalization can be expressed as:
wherein R isMaxAnd RMinRepresenting the maximum and minimum values in the RSS vector, respectively.
Step 4, the RSS vector imaging specifically comprises the following steps:
in the normalized RSS vector, 0 represents the weakest energy, and 1 represents the strongest energy; dividing the energy into K levels, if the normalized RSS value of the ith sensor belongs to the jth energy level, setting the element of the jth row and the ith column of the RSS image to be 1, setting the rest elements to be 0, and generating a pair of [ KxK ] RSS images according to the level of each element.
The model in step 5 is specifically as follows: the model comprises two hidden layers and two full-connection layers, wherein each hidden layer is divided into three parts: a convolutional layer, a ReLU (transformed linear units) activation function and a max pooling layer; the two convolutional layers contain 64 and 128 convolutional kernels, respectively, with a size [3 × 3], the step size is set to 1, the activation function of the first fully-connected layer is ReLU and the total number of output neurons is 2048; the second full-connection layer adopts a Softmax activation function to calculate the posterior probability of each category; the Adam gradient is used as an optimizer during training, and the cost function adopts cross-entropy (cross-entropy); the weight of the model adopts a truncated positive-space distribution during initialization, and the bias value is initialized to be constant 0.1.
The step 6 specifically comprises the following steps: and (3) converting the RSS value actually measured by the sensor into an RSS image, then putting the RSS image into the model trained in the step (5), and predicting the DOA according to the category with the highest probability in the output result of the CNN.
The invention has the beneficial effects that: the DOA method reduces the dependence of each array element on the sampling frequency and the channel synchronization, and is suitable for the environment with relatively simple realization, lower complexity and no need of high positioning precision. In addition, the method has the potential of being applied to the uniform linear array and other more common DOA scenes.
Drawings
FIG. 1 is a modeled DOA scenario;
FIG. 2 is the complete process of RSS preprocessing;
FIG. 3 is a radiation pattern used in the present invention;
FIG. 4 is a network structure of a model;
FIG. 5 is a flow chart of the implementation and usage of the method of the present invention;
FIG. 6 is a representation before and after pretreatment;
FIG. 7 is a comparison of the present invention with the MUSIC method at low complexity;
FIG. 8 is the DOA estimation accuracy within a certain error range of the present invention.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
The invention discloses a DOA estimation method based on a convolutional neural network and received signal strength, which mainly solves the problem that when the existing DOA estimation method based on DNN (deep neural network) is used for extracting input characteristics, a receiving array element needs accurate time synchronization, higher sampling rate and other hardware costs are increased. The method comprises the following implementation steps: 1) converting the DOA problem into a CNN (volumetric neural network) -based multi-classification problem; 2) forming an initial RSS vector with a fixed dimension according to the RSS (received signal strength) acquired by the sensors; 3) carrying out normalization processing on the RSS vector; 4) converting the normalized RSS vector into an RSS image; 5) training a CNN model; 6) and selecting the class with the highest prediction probability to estimate the DOA according to the output result of the trained CNN. The DOA method reduces the dependence of each array element on the sampling frequency and the channel synchronization, and is suitable for the environment with relatively simple realization, lower complexity and no need of high positioning precision. In addition, the method has the potential of being applied to the uniform linear array and other more common DOA scenes.
The invention provides a DOA estimation method based on CNN and RSS, which is used for solving the problems that parameters adopted by the existing DOA method depend on sufficient snapshot number and the precise synchronization among array elements during extraction. The invention converts the DOA problem into the angle space classification problem based on the CNN, and only adopts a special RSS image as the input characteristic, thereby reducing the burden of hardware. The adaptive capacity of the DOA estimation method under low cost and low complexity is improved, and the high-probability DOA estimation within a certain error range is finally realized.
The technical scheme of the invention is a DOA estimation method based on CNN and RSS, and the invention comprises the following steps:
1) the range of DOAs, the number K and position of sensors, the resolution of the angular space and the total number M of classes are defined.
2) An RSS vector of K × 1 is constructed from the RSS values received by each sensor.
3) The RSS vector is normalized to between 0 and 1.
4) And constructing a K multiplied by K RSS image according to the normalized RSS vector.
5) And defining layers and parameters in the network, and generating training samples of various classes for model training.
6) And (4) performing actual DOA prediction according to the trained model.
The method comprises the following steps: the range of DOAs, the number K and position of sensors, the resolution of the angular space and the total number M of classes are defined.
FIG. 1 is a modeled DOA scenario. DOA is defined as being in the range of [ -30 °, +30 ° ]]The sensors with K equal to 49 are uniformly distributed at [ -45 °, +45 ° ]]On the circumference of (a). Defining a resolution of 1 deg., the entire DOA range is divided into M-61 classes, specifically denoted Θ - θ1…θM}. Wherein theta isiIndicating the DOA corresponding to the ith category.
When the antenna of the signal source rotates to a certain category, the RSS measured by the K sensors forms a group of samples, and each sample corresponds to a label. The tag is a 1 × M vector in which only one element is 1 and the rest are 0. The category corresponding to the position of 1 is the main beam direction of the antenna, i.e. the incoming wave direction. The DOA problem has thus far been transformed into a classification problem for supervised learning.
Step two: an RSS vector of K × 1 is constructed from the RSS values received by each sensor.
A two-dimensional far-field radiation model
WhereinIs the Half Power Beamwidth (HPBW) of the antenna,
In theory the RSS received by the ith sensor is PriThe concrete form is as follows.
Where Pt is 56dBm, Gt is 20dB, Gr is 15dB, and d is the gains of the transmitting and receiving antennasi1km means the distance between the signal source and the i-th sensor, λ represents the wavelength, and the transmission frequency of the signal is set to 50 MHz.
The Pr of the K sensors constitutes an RSS vector, denoted as Rθ。
Rθ＝[Pr1…PrK]T+n
Wherein n is [ n ]1…nK]TThe noise at the time of reception is replaced by white noise. Theta represents a certain DOA.
Step three: the RSS vector is normalized to between 0 and 1.
Extracting the maximum value and the minimum value in the RSS vector formed in the previous step, and recording as RMaxAnd RMinR is represented by the following formulaθ＝[R1…RK]TNormalized to between 0 and 1.
The normalized result is shown in fig. 2 (b).
Step four: and constructing a K multiplied by K RSS image according to the normalized RSS vector.
In the normalized RSS vector, 0 represents the weakest energy and 1 represents the strongest energy. The energy is divided into K levels, and a set of K × K RSS images is generated according to the level of each element. If the normalized RSS value of the ith sensor belongs to the jth energy level, the element of the jth row and ith column of the RSS image is 1, and the rest elements are set to be 0. The result of the imaging is shown in fig. 2 (c).
Step five: and defining layers and parameters in the network, and generating training samples of various classes for model training.
Fig. 3 is a four antenna radiation model used to generate a data set. The training set is enriched from three angles of signal-to-noise ratio, radiation model and angle category. The total number of samples was 73,200(61 angle classes x 4 radiation models x 5 SNRs x 60 trials per SNR), where 61,000 samples were used to generate the training set and 12,200 samples were used to generate the test set.
Fig. 4 is a structural diagram of a CNN network model. The model includes two hidden layers and two fully connected layers. Each hidden layer passes through three parts: a convolutional layer, a ReLU (transformed linear units) activation function, and a max pooling layer. The two convolutional layers contain 64 and 128 convolutional kernels, respectively, with a size of [3 × 3], with the step size set to 1. The activation function of the first fully-connected layer is ReLU and the total number of output neurons is 2048. The second fully-connected layer uses the Softmax activation function to calculate the posterior probability for each class. The Adam gradient serves as an optimizer during training, and the cost function employs cross-entropy (cross-entropy). The weight of the model adopts a truncated positive-space distribution during initialization, and the bias value is initialized to be constant 0.1.
Step six: and (4) performing actual DOA prediction according to the trained model.
And (4) converting the RSS value actually measured by the sensor into an RSS image through the second step and the third step, then putting the RSS image into the result of the fifth step, namely a trained model, and predicting the DOA according to the category with the highest probability in the output result of the CNN. The specific flow is shown in fig. 5.
The effect of the method will be further explained by combining with simulation experiments as follows:
the method provided by the invention is compared before and after pretreatment, and compared with the MUSIC algorithm under low complexity, and finally the precision of the method in a certain error range is verified.
Fig. 2 is a complete RSS vector preprocessing process that converts RSS vectors into RSS images while preserving sensor location information. In fig. 3, P1 and P2 are gaussian antennas with HPBW of 10 ° and 16 °, respectively, and P3 and P4 are antennas with side lobes formed by beams. FIG. 6 is a comparison before and after pretreatment. Pre-processing means that no imaging is performed, only the normalized RSS vector is used as the input feature. It can be seen that the trained model has relatively high accuracy in each SNR after the pre-processing proposed by the present invention. The invention proves that the method has obvious advantages in the training stage, and the treatment mode is more effective.
Fig. 7 uses P4 to simulate the radiation pattern of the signal source, and Root Mean Square Error (RMSE) is used as an evaluation index. The value of DOA is no longer a standard class but rather a more realistic random angle, e.g. 11.2 deg., -28.1 deg., etc. The method of the invention is compared with the MUSIC algorithm under 5 snapshots and 10 snapshots respectively. From the results, it can be seen that under the conditions of insufficient sampling and few snapshots, the method of the present invention exhibits relatively smaller DOA estimation error, and when the SNR is larger, the advantage is significant. The present method is still subject to gaps when the SNR is small.
Fig. 8 shows the accuracy under certain errors. The result of the DOA estimation can be considered accurate if the difference between the true angle and the predicted DOA is less than some determined error. It is clear that the results at + -1 deg. are better than those at + -0.5 deg.. Therefore, the present invention is more suitable for applications in the context where the positioning accuracy does not need to be too high. In summary, the present invention provides a new idea and a solution to mitigate the dependence on high sampling rate and accurate time synchronization for the field of DOA estimation.
- 上一篇：一种医用注射器针头装配设备
- 下一篇：一种角度与频率联合估计的增广DOA矩阵方法