A ResNet multi-target signal DOA estimation algorithm based on beamforming

Abstract

The application discloses a ResNet multi-target signal DOA estimation algorithm based on beam forming, and the method comprises the following steps: a detector array is adopted, an LCMV beam forming algorithm is applied to scan a space, phase adjustment is performed in the scanning process, signal enhancement is realized for signals from the scanning direction, and interference signals from other directions are suppressed, a residual network ResNet is used to convert a multi-target DOA estimation problem into a two-classification problem, an efficient channel attention mechanism is added to the residual network ResNet to enhance the extraction of signal characteristics of the model, and the accuracy and robustness of target DOA estimation are improved. The disclosed algorithm has the technical characteristics of high detection accuracy and precision under multiple interference signals and low signal-to-noise ratio.

Description

TECHNICAL FIELD

[0001] The application relates to an estimation algorithm, in particular to a ResNet multi-target signal DOA estimation algorithm based on beam forming, and belongs to the field of sound source positioning. BACKGROUND

[0002] DOA estimation is a sound source positioning technology, which refers to calculating the direction of a sound source in space by analyzing sound signals according to the time difference or signal intensity difference of sound waves reaching various sensors. This technology has wide application in the fields of speech recognition, audio signal processing, wireless communication, etc. However, when the number of sound sources increases, DOA estimation is more susceptible to interference in complex environments, and it will be difficult to accurately distinguish multiple sound sources in close directions, especially in the case of similar signal intensity or overlap, which greatly affects the accuracy of the estimation algorithm. The traditional DOA estimation method based on physical model may face the limitation of resolution capability in multi-target estimation, and it is difficult to cope with complex signal environment, and these methods have the problem of robustness under noise and non-ideal conditions. SUMMARY

[0003] In order to solve the above-mentioned prior art problems, the application provides a ResNet multi-target signal DOA estimation algorithm based on beam forming, which has the technical characteristics of being able to reduce the interference between targets and environmental noise interference, and effectively distinguishing the source angles of each target in the case of multi-target interference.

[0004] In order to achieve the above-mentioned purpose, the application is realized by the following technical scheme:

[0005] The ResNet multi-target signal DOA estimation algorithm based on beam forming comprises the following steps: adopting a detector array, applying an LCMV beam forming algorithm to scan the space, adjusting the phase during the scanning process, realizing signal enhancement for signals from the scanning direction and suppressing interference signals from other directions, and then using a residual network ResNet to convert the multi-target DOA estimation problem into a binary classification problem, and adding an efficient channel attention mechanism to the residual network ResNet to enhance the extraction of signal features by the model and improve the accuracy and robustness of target DOA estimation.

[0006] Preferably, the method comprises the following steps:

[0007] Step S1: a circular array receives a UAV signal, calculates the distance of a sound source to each array and samples accordingly;

[0008] Step S2: phase adjustment is performed on the sampled signal by using an LCLV beam forming algorithm;

[0009] Step S3: splicing the phase-adjusted array output and its weighted sum signal;

[0010] Step S4: using the ResNet-ECA network to extract features from the data, training, and obtaining a multi-target DOA estimation model;

[0011] Step S5: using the newly adjusted spliced data to infer the trained model;

[0012] Step S6: the network determines whether there is a target at each scanning angle according to the multi-target DOA estimation model.

[0013] Preferably, the step S1 is specifically using an array composed of omnidirectional sensors to receive ambient sound signals, and all detectors of the array sample the signals, and record the signals received by the circular array at time t.

[0014] Preferably, the steps S2 to S3 are specifically: for the signals received by each detector of the array at time t, using the LCMV beamforming algorithm to scan in all directions, for each scanning direction, first multiplying the signal received by each detector by a suitable weight vector to adjust the phase, then adding, and then adding the phase-adjusted signal to form a set of data for the determination of the presence of signals in the scanning direction.

[0015] Preferably, the step S4 is specifically: during the network training process, the signal data set in each direction after phase adjustment is made into a complete network training set, and three labels are added when naming, which are array scanning angle, target incident angle and prior determination result (whether there is a target), the above training set is used as the input of the ResNet-ECA signal recognition network and input into the ResNet network, and after training, a multi-target DOA estimation model is obtained.

[0016] Preferably, in the network inference process of step S5, the detector array is placed in the required environment, and steps S1 and S2 are repeatedly executed, the signals received by the array after LCMV beamforming algorithm processing are used as input data, and are input into the multi-target DOA estimation model in step S4, and then the model output is used to determine whether there is a target at each scanning angle.

[0017] Preferably, the input signal of the ResNet network (ResNet-ECA signal recognition network) includes the phase-adjusted signal of each detector and the sum of all adjusted signals.

[0018] Phase adjustment refers to multiplying the original signal by a weight vector according to the array scanning angle after the array receives the signal through the LCMV beamforming technology.

[0019] The sum of the phase-adjusted signals is adjusted so that the signals obtained after phase adjustment in the scanning angle are superimposed in phase, and the signals are enhanced, while the noise and interference direction signals in the non-scanning angle are not superimposed in phase, and are suppressed.

[0020] Preferably, the output of the ResNet-ECA signal recognition network is only to determine whether there is a target in each scanned direction, but not the arrangement of the non-scanning angle and the target incident angle. The network needs to record the number of targets detected by the scanning full direction.

[0021] Preferably, the input data set of the ResNet-ECA signal recognition network is a two-dimensional matrix, which is generated by the following steps:

[0022] The signal output by each detector is a one-dimensional signal, which is a sampling of the signal in a period of time. In a single scanning angle calculation, the signal after phase adjustment by the beam forming algorithm is still a one-dimensional signal, and the sum of the adjusted signals of the detectors is also a one-dimensional signal. These signals together form a two-dimensional data matrix.

[0023] Preferably, the input two-dimensional data matrix of the ResNet-ECA signal recognition network can be a two-dimensional data matrix obtained by a single scanning angle and / or a huge two-dimensional data matrix composed of all two-dimensional data matrices corresponding to all scanning angles.

[0024] When the input of the ResNet network is only a two-dimensional matrix obtained by a single scanning angle, the output of the ResNet network is a judgment of whether there is a target signal in the scanning angle. When the input of the ResNet network is a two-dimensional data matrix obtained by full-direction scanning angle, the output of the ResNet network (ResNet-ECA signal recognition network) is the number of targets existing in the full direction and the direction of each target.

[0025] Preferably, the phase-adjusted signal is obtained by multiplying the original signal collected by the detector by a weighting vector. This is a digital signal, and the weighting vector can be adjusted according to the accuracy requirement of the scanning angle or the specific environment without retraining the RecNet-ECA network. At the same time, because the weighting vector is a digital signal calculated in the system design process, the full-direction scanning can be completed in parallel without being substituted into the RecNet-ECA network one by one, saving calculation and judgment time.

[0026] Preferably, the beamforming-based ResNet network, the output of the ResNet network (ResNet-ECA signal recognition network) is 1 or 0, that is, only to judge whether there is a target in the scanning direction, when the ResNet-ECA signal recognition network input is the data matrix corresponding to the omnidirectional scanning angle, the ResNet-ECA signal recognition network can judge any multiple targets, rather than a specific number of targets according to the label of the training set.

[0027] Advantages:

[0028] (1) The traditional method relies on strong assumptions about the physical signal model, and the beamforming-based ResNet multi-target DOA estimation algorithm of the application can effectively distinguish the source angles of each target in the multi-target interference situation.

[0029] (2) The application beneficially combines the traditional method with the deep learning algorithm, and the traditional LCMV beamforming algorithm can enhance the signal of each scanning direction and suppress the interference of other directions by selecting a suitable weighting vector, greatly reducing the target interference and environmental noise interference.

[0030] (3) By combining the one-dimensional signal sampled by the detector with the signal weighted and synthesized using the LCMV beamforming algorithm, the ResNet-ECA signal recognition neural network can effectively distinguish whether the phase of each detector sampling signal is aligned with the phase of the synthesized signal. The more aligned the phase, the more likely there is a target in this direction.

[0031] (4) The weighting vector for phase adjustment can be adjusted appropriately according to the accuracy requirements of the scanning angle or the specific environment, without the need to retrain the RecNet-ECA network. At the same time, because the weighting vector is a digital signal calculated during system design, omnidirectional scanning can be completed in parallel without being entered into the network one by one, saving calculation and judgment time.

[0032] (5) In the present application, it is relatively easy to use a neural network to distinguish multiple target angles, only to determine whether there is a target in a certain direction. Compared with other neural networks that do not combine beamforming, the present application needs to determine the position of each target according to the classification probability of the target in each direction, and the computational complexity of the present application is lower than that of other existing target DOA estimation methods.

[0033] (6) The present application can be widely used in wireless communication, target detection and tracking, speech processing, autonomous driving cars, environmental monitoring and other fields, and also provides new ideas and methods for monitoring underwater targets in the ocean. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1This is a flow chart of the estimation algorithm of the present invention.

[0035] Figure 2 This is one of the visualizations of the dataset and the beamforming effect.

[0036] Figure 3 This is the second visualization of the data set and beamforming effect.

[0037] Figure 4 This is the third visualization of the data set and beamforming effect.

[0038] Figure 5 This is the fourth visualization of the data set and beamforming effect. DETAILED DESCRIPTION

[0039] The present invention will be further described below with reference to the accompanying drawings, but the present invention is not limited to the following embodiments.

[0040] like Figures 1-5 The figure shows a specific embodiment of a beamforming-based ResNet multi-target signal DOA estimation algorithm. First, a circular array receives sound source signals from multiple drones. The LCMV classic beamforming algorithm is used to perform 360-degree beam scanning. Signal phase is adjusted for each direction to enhance the signal and reduce noise and the interaction between multiple target signals. Subsequently, the phase-adjusted output data from each microphone and the beamforming output data are combined. This data is then trained using the ResNet-ECA target signal DOA estimation algorithm to generate a DOA estimation model. This model is used to classify and identify targets in the scanning direction, thereby determining the target's ultimate source direction.

[0041] This embodiment provides a ResNet multi-target signal DOA estimation algorithm based on beamforming, and the method specifically includes the following steps:

[0042] Step S1: Use an array of omnidirectional sensors to receive surrounding sound signals. All detectors in the array sample the signals and record the signals received by the circular array at time t.

[0043] In this embodiment, the microphone array is assumed to be an 8-element circular array with a radius of 0.15m. The coordinate system is established with the center of the circular array as the reference point. The array scans clockwise with the positive direction of the coordinate axis Y as 0°, and the range is Φ∈[0,2π). The number of sampling snapshots is 400, the sound speed is 340m / s, and the sampling frequency is f sThe sound signal is intercepted every 600 time points for one detection, which can ensure that all detectors can sample the signal completely. Each detector samples 400 time points, and 8 detectors simultaneously detect to obtain an 8*400 matrix data.

[0044] Step S2: The signals received by each detector of the array at time t are scanned in all directions using the LCMV beamforming algorithm. For each direction, the signals received by each detector are first multiplied by a suitable weighting vector (suitable meaning: data selected according to actual needs) for phase adjustment, and then added. The phase-adjusted signals and the added signals form a set of data for determining the presence of signals in the scanning direction.

[0045] In this embodiment, a suitable weighting vector is selected to enhance the signals received in the scanning angle and suppress signals in other directions. If the scanning direction is the same as the direction of the sound signal source of the unmanned aerial vehicle during the scanning process, the enhanced signal intensity in this direction is the largest, i.e., the propagation delay of each array element is basically filled, and the output signals of the array elements are basically in phase. The enhanced signal is obtained by weighting and summing the phase-adjusted output signals of the array elements, and the 9*400 matrix data is obtained by splicing the 1*400 signal data and the 8*400 matrix data of the phase-adjusted output data of the 8 array elements.

[0046] Step S3: During the ResNet network training process, a complete ResNet network training set is made from the phase-adjusted signal data set in each direction, and three labels are marked during naming, which are array scanning angle, target incident angle and priori determination result (whether there is a target). The above training set is used as the input of the ResNet-ECA signal recognition network and input into the ResNet network. After training, a multi-target DOA estimation model is obtained.

[0047] Figures 2-5 The output data of the detector and the synthesized signal of this embodiment are used to make a training set and a visualized graph of beamforming effect. Specifically Figure 3 The input network training set legend is shown. The first 8 signals show the phase-adjusted signals of the 8 array elements, and the 9th signal is the weighted and synthesized output data of the 8 microphones. It can be seen that at some time points marked in the box, the 8 microphone data phases are basically aligned and added in the 9th signal to form a relatively obvious wave peak. Figures 3-5 The signal processing process under different signal-to-noise ratios of 10dB, 0dB and -5dB is shown respectively, Figures 3-5 The sub Figure 1 ,2 The three figures respectively represent the first row signal of the noiseless data set, the first row signal of the data set with noise, and the ninth row signal of the data set, Figures 3-5 The sub Figure 2 It can be seen that as the signal-to-noise ratio continues to decrease, the signal gradually distorts. When the signal-to-noise ratio is -5dB, the signal is almost completely covered by noise. By comparison, it can be seen that Figures 3-5 The sub Figure 3 It can be clearly seen that after phase adjustment of the 8-segment signal, the noise and interference signal can be effectively removed.

[0048] In this embodiment, when constructing the training set, due to the symmetry of the circular array, the array scanning angle and the incident angle of the two unmanned aerial vehicle signal sources are set to be within the angle range of 0-90°, and three angles are freely combined every 10 degrees, a total of 1000 angle combinations. When the scanning angle coincides with any of the incident angles of the two signal sources, a "target" label is marked, and otherwise a "target" label is marked. Each combination is sampled 30 times, and a total of 30,000 data are established as the data set.

[0049] The sorted data set is input into the pre-built and initialized parameter ResNet-ECA neural network for training. Through iterative training of a large amount of data, the parameters are gradually updated, and finally a multi-target DOA estimation model is obtained.

[0050] Step S4: During the network inference process, the detector array is placed in the required environment, and steps S1 and S2 in the embodiment are repeatedly executed. The signal received by the array after being processed by the LCMV beamforming algorithm is input as input data into the multi-target DOA estimation model described in step S3, and the model outputs whether there is a target at each scanning angle.

[0051] In this embodiment, the accuracy and root mean square error are used to evaluate the method of the present application. When the signal-to-noise ratio is 15dB, the ResNet multi-target signal DOA estimation algorithm based on beamforming in the present application has a DOA estimation accuracy of 96.73% for two targets in the air, and a root mean square error of 2.53°. Compared with the multi-target DOA estimation results of ResNet, GCC-PHAT+ResNet and beamforming+MobileNet, it can be seen that the DOA estimation accuracy of ResNet for multi-target is 31.51%, and the root mean square error is 9.63°. The DOA estimation accuracy of GCC-PHAT+ResNet for multi-target is 85.12%, and the root mean square error is 5.81°. The DOA estimation accuracy of beamforming+MobileNet for multi-target is 82.48%, and the root mean square error is 7.39°. Therefore, the method of the present application effectively improves the DOA estimation accuracy compared with the existing method.

[0052] In addition, when there are three or more targets in the air, the method of the present application can still maintain a DOA estimation accuracy of 96% or more and a root mean square error of less than 5°, which further indicates that the present application has strong robustness and generalization ability in multi-target DOA estimation, while reducing the complexity of signal processing in multi-target estimation.

[0053] Finally, it should be noted that the present application is not limited to the above embodiments, but can have many variations. All variations that can be directly derived or inferred from the disclosed content by those of ordinary skill in the art should be considered within the scope of the present application.

Claims

1. A ResNet multi-target signal DOA estimation algorithm based on beamforming, characterized in that: The algorithm includes: using a detector array and applying the LCMV beamforming algorithm to scan the space, performing phase adjustment during the scanning process, achieving signal enhancement for signals from the scanning direction, and suppressing interference signals from other directions. Then, using the residual network ResNet to transform the multi-target DOA estimation problem into a binary classification problem, and adding an efficient channel attention mechanism to the residual network ResNet to enhance the model's extraction of signal features, thereby improving the accuracy and robustness of target DOA estimation. The algorithm includes the following steps: Step S1: The circular array receives the drone signal, calculates the distance from the sound source to each array and samples accordingly; Step S2: using the LCMV beamforming algorithm to adjust the phase of the sampled signal; Step S3: splicing the phase-adjusted array output with its weighted sum signal; Step S4: Use the ResNet-ECA network to extract features from the data and perform training to obtain a multi-target DOA estimation model; Step S5: Use the trained model to infer the newly adjusted and spliced ​​data; Step S6: The network determines whether there is a target at each scanning angle based on the multi-target DOA estimation model; Specifically, steps S2 to S3 are as follows: for the signal received by each detector in the array at time t, the LCMV beamforming algorithm is used to scan in all directions. For each azimuth scan, the signal received by each detector is first multiplied by a weighting vector, phase-adjusted, and then added. Each phase-adjusted signal and the added signal are combined into a set of data for determining the presence of the signal in the scanning direction.

2. The beamforming-based ResNet multi-target signal DOA estimation algorithm according to claim 1, characterized in that: The step S1 specifically includes: using an array composed of omnidirectional sensors to receive surrounding sound signals, all detectors in the array sample the signals, and recording the signals received by the circular array at time t.

3. The beamforming-based ResNet multi-target signal DOA estimation algorithm according to claim 1, characterized in that: Specifically, step S4 is as follows: during the network training process, the phase-adjusted signal data set in each direction is made into a complete network training set, and three labels are added when naming, namely, the array scanning angle, the target incidence angle, and the prior judgment result. The above training set is used as the input of the ResNet-ECA signal recognition network and input into the ResNet network. After the training is completed, a multi-target DOA estimation model is obtained.

4. The beamforming-based ResNet multi-target signal DOA estimation algorithm according to claim 1, characterized in that: In step S5, during the network inference process, the detector array is placed in the required environment, and steps S1 and S2 are repeated. The signal received by the array and processed by the LCMV beamforming algorithm is used as input data and input into the multi-target DOA estimation model in step S4. The model output is then used to determine whether there is a target at each scanning angle.

5. The beamforming-based ResNet multi-target signal DOA estimation algorithm according to claim 1, characterized in that: The input signal of the ResNet network includes the phase-adjusted signals of each detector and the sum of all adjusted signals; phase adjustment means that the signal received by the array is multiplied by a weighted vector according to the array scanning angle through the LCMV beamforming technology.

6. The beamforming-based ResNet multi-target signal DOA estimation algorithm according to claim 1, characterized in that: The output of the ResNet-ECA signal recognition network is to judge whether there is a target in each scanning direction, rather than the arrangement of the scanning angle and the target incident angle. The ResNet network needs to record the number of targets detected in all scanning directions.

7. The beamforming-based ResNet multi-target signal DOA estimation algorithm according to claim 1, characterized in that: The input dataset of the ResNet-ECA signal recognition network is a two-dimensional matrix, whose composition is generated by the following steps: The signal output by each detector is a one-dimensional signal, which is a sampling of the signal over a period of time. In the calculation of a single scanning angle, it is still a one-dimensional signal after phase adjustment by the beamforming algorithm. The sum of the adjusted signals of each detector is also a one-dimensional signal. These signals together constitute a two-dimensional data matrix.

8. The beamforming-based ResNet multi-target signal DOA estimation algorithm according to claim 3, characterized in that: The input 2D data matrix of the ResNet-ECA signal recognition network can be a 2D data matrix obtained from a single scanning angle and / or a giant 2D data matrix consisting of all 2D data matrices corresponding to all scanning angles; When the input of the ResNet network is only a two-dimensional matrix obtained from a scanning angle, the output of the ResNet network is the judgment of whether there is a target signal at this scanning angle; when the input of the ResNet network is a two-dimensional data matrix obtained from omnidirectional scanning angles, the output of the ResNet network is the number of targets judged by the ResNet network in all directions and the location of each target.

Images