阅读说明：本技术 一种用于实时识别物体形状的声学方法 (Acoustic method for identifying shape of object in real time ) 是由 梁彬 翁经锴 丁玉江 胡成博 于 2019-09-16 设计创作，主要内容包括：本发明公开了一种用于实时识别物体形状的声学方法,包括以下步骤：1得出神经网络的训练集和测试集；2将训练集输入声学超神经网络,经过多层声学超表面后,在探测面形成一定的声压分布；3探测面被划分为N个区域,取得这N个区域的总声能量值；4计算取得的N个值与物体的标签之间的误差,并运用误差来计算每一个超神经元所施加的相位调制的梯度,以此来更新梯度,从而更新相位,直到得到稳定的输出,并可以正确地识别目标对象；5通过更新后的相位值,确定超表面上每一个单元的相位偏移值,根据相位偏移值制作声学超表面；6将声学超表面放置合适的位置,声波经过声学超表面后,位于声学超表面后的探头即可识别出目标对象。(The invention discloses an acoustic method for identifying the shape of an object in real time, which comprises the following steps: 1, obtaining a training set and a testing set of the neural network; 2, inputting the training set into an acoustic super-neural network, and forming certain sound pressure distribution on a detection surface after passing through a multilayer acoustic super-surface; 3 dividing the detection surface into N areas, and acquiring the total acoustic energy value of the N areas; 4 calculating the error between the obtained N values and the label of the object, and calculating the gradient of the phase modulation applied by each super neuron by using the error so as to update the gradient, thereby updating the phase until a stable output is obtained and the target object can be correctly identified; 5, determining a phase deviation value of each unit on the super surface through the updated phase value, and manufacturing the acoustic super surface according to the phase deviation value; and 6, placing the acoustic super-surface at a proper position, and identifying the target object by the probe positioned behind the acoustic super-surface after the acoustic wave passes through the acoustic super-surface.)

1. An acoustic method for identifying the shape of an object in real time, comprising the steps of:

step 1, after sound waves pass through an N-type target object, sound pressure distribution on a cross section which is vertical to a propagation direction after a formed scattering sound field is propagated for a certain distance is used as a data set, a label of the data set is consistent with that of the target object, and partial data in the data set are randomly extracted and respectively used as a training set and a test set;

step 2, inputting the training set into an acoustic super-neural network, and forming certain sound pressure distribution on a detection surface after passing through a multilayer acoustic super-surface;

step 3, dividing the detection surface into N areas, wherein N is the number of the objects to be distinguished, and acquiring the total acoustic energy value of the N areas;

step 4, calculating the error between the obtained N values and the label of the object, and calculating the gradient of the phase modulation applied by each super neuron by using the error so as to update the gradient, thereby updating the phase until stable output is obtained and the target object can be correctly identified;

step 5, determining a phase deviation value of each unit on the super surface through the updated phase value, and manufacturing the acoustic super surface according to the phase deviation value;

and 6, placing the acoustic super surface at a proper position, and identifying the target object by the probe positioned behind the acoustic super surface after the acoustic waves pass through the acoustic super surface.

2. An acoustic method for real-time recognition of the shape of an object according to claim 1, characterized in that said step 1 comprises the following steps:

step 1.1, incident plane waves on an object to generate a fringe field map of the object is described

Step 1.2, constructing a label corresponding to the object, wherein the label consists of N-1 0 s and 1 s, the position of the number 1 in the label represents the shape of the object, and the label is marked as

Step 1.3, constructing a test set according to steps 1.1 and 1.2

3. The acoustic method for recognizing the shape of the object in real time according to claim 1, wherein in the step 2, the acoustic device in the acoustic hyper-neural network is composed of M layers of acoustic hyper-surfaces, each layer of the hyper-surfaces is composed of n x M hyper-neurons, and a training set is used as an input; the forward propagation function model of the acoustic super-neural network is

4. The acoustic method for recognizing the shape of the object in real time according to claim 1, wherein in the step 3, the detection surface is divided into N regions, each region corresponds to the shape of an object, and by comparing the sizes of the N regions, the shape of the object corresponding to the region with the largest acoustic energy value is the shape of the object predicted by the acoustic super neural network.

5. An acoustic method for recognizing a shape of an object in real time according to claim 3, in the step 4, M × n × M phase initial values are generated by using a random function, the scattered field of the target object in the training set is input to the super surface, the distribution of the sound pressure of the scattered field on the super surface is calculated by using the relation between layers, n values are obtained by calculating the sum value of the acoustic energy of N pre-divided areas on the detection plane, an error function between the N values and the label of the object is calculated, calculating the variation gradient of the phase value on the super surface of the M layers by using the error function and a gradient descent method, and updating the phase of the whole super surface, and continuously updating the phase value on the super surface by continuously inputting the training set until stable output is obtained and the target object can be correctly identified.

6. An acoustic method for real-time shape recognition of an object according to claim 5, wherein in step 4, the N values are processed by a Softmax function, the probability that the neural network predicts the different shapes to which it corresponds is calculated, the error between the probability and the label is calculated, the error is calculated by a cross entropy error function, then the gradient of the phase modulation applied by each super neuron is found by a gradient descent method, the phase value is updated by subtracting the phase gradient from the phase value, and the Softmax function is:

7. An acoustic method for real-time recognition of the shape of an object according to claim 3, characterized in that: the distance between the adjacent acoustic super surfaces is 0.5-5 incident sound wave wavelengths.

Technical Field

The invention relates to an acoustic method for identifying the shape of an object in real time, and belongs to the field of acoustics.

Background

Recognizing the shape of an object is a commonly used technique in acoustics, such as ultrasonic non-destructive testing and medical ultrasound techniques. Conventional ultrasound imaging techniques rely on arrays of a large number of transducers, and the size of the individual transducers is often larger than the wavelength, which limits further improvements in imaging resolution, and is disadvantageous for accurate diagnosis. In addition, the conventional method often requires complex post-processing of the received signal, relies on a powerful processor, consumes a high amount of energy, and cannot automatically identify an object.

Disclosure of Invention

The purpose of the invention is as follows: in order to overcome the defects in the prior art, the invention provides the acoustic method for identifying the shape of the object in real time, only a small number of transducers are needed, the neural network can utilize scattered acoustic energy to analyze in real time, a data post-processing device is not needed, the overall structure size is the wavelength magnitude, and the acoustic method is simple in structure and easy to implement.

The technical scheme is as follows: in order to solve the above technical problem, an acoustic method for identifying the shape of an object in real time according to the present invention includes the following steps:

step 1, after sound waves pass through an N-type target object, sound pressure distribution on a cross section which is vertical to a propagation direction after a formed scattering sound field is propagated for a certain distance is used as a data set, a label of the data set is consistent with that of the target object, and partial data in the data set are randomly extracted and respectively used as a training set and a test set;

step 2, inputting the training set into an acoustic super-neural network, and forming certain sound pressure distribution on a detection surface after passing through a multilayer acoustic super-surface;

step 3, dividing the detection surface into N areas, wherein N is the number of the objects to be distinguished, and acquiring the total acoustic energy value of the N areas;

step 4, calculating the error between the obtained N values and the label of the object, and calculating the gradient of the phase modulation applied by each super neuron by using the error so as to update the gradient;

and 5, inputting the objects in the test set into the trained acoustic hyper-neural network, and predicting the shapes of the objects.

Preferably, the step 1 comprises the following steps:

step 1.1, incident plane waves on an object to generate a fringe field map of the object is described

Step 1.2, constructing a label corresponding to the object, wherein the label consists of N-1 0 s and 1 s, the position of the number 1 in the label represents the shape of the object, such as a recognition sphere, a cube and a regular tetrahedron, the first position represents the sphere, the second position represents the cube, the third position represents the regular tetrahedron, and if the object is the cube, the label is (0,1, 0). Inscription of the label as

Step 1.3, constructing a test set according to steps 1.1 and 1.2

Preferably, in step 2, an acoustic device in the acoustic super-neural network is composed of M layers of acoustic super-surfaces, each layer of super-surface is composed of n × M super-neurons, a training set is used as an input, a scattered field is incident on the acoustic super-surface, each super-neuron on each layer adds a phase modulation to an input sound pressure, an output sound wave propagates for a distance, the super-neuron on the layer also adds a phase modulation to the second layer of acoustic super-surface, the subsequent transmission process is similar, and finally a certain sound field distribution is formed on a detection surface; the forward propagation function model of the acoustic super-neural network is

P ^lIs the input wave, G, on the superneural unit of layer I ^lIs a matrix of wave propagation equations that,

is the phase modulation applied by the super neurons of layer i,

representing a dot product.

Preferably, in step 3, the detection surface is divided into N regions, each region corresponds to a shape of an object, and by comparing sizes of the N regions, the shape of the object corresponding to the region with the largest acoustic energy value is the shape of the object predicted by the acoustic super neural network.

Preferably, in step 4, M × N × M initial phase values are generated by using a random function, the fringe field of the target object in the training set is input to the super-surface, the distribution of the sound pressure of the fringe field on the super-surface is calculated by using the relationship between layers, N values are obtained by calculating the total value of the sound energy of N pre-divided regions on the detection plane, an error function between the N values and the label of the object itself is calculated, the change gradient of the phase values on the super-surface of M layers is calculated by using the gradient descent method through the error function, the phase of the whole super-surface is updated, and the phase values on the super-surface are continuously updated by continuously inputting the training set until stable output is obtained and the target object can be correctly identified.

Preferably, in step 4, the N values are processed by a Softmax function, the probability that the neural network predicts the different shapes corresponding to the N values is calculated, the error between the probability and the label is calculated, the error is calculated by a cross entropy error function, then the gradient of the phase modulation applied to each super neuron is obtained by a gradient descent method, and the phase value is updated by subtracting the phase value from the phase gradient, where the Softmax function is:

prob _qprobability of input scattered field due to q-th shape of scattered field, I _qIs the total value of the sound energy of the qth area; the cross entropy error function is:

g _qis the qth tag.

Preferably, the distance between the adjacent acoustic super surfaces is 0.5-5 incident sound wave wavelengths.

In order to make the traditional acoustic super surface have the capability of recognizing the shape of an object, the traditional acoustic super surface and deep learning are combined together, and the traditional acoustic super surface is called an acoustic super neural network. The following describes how the super-surface can be given the ability to identify objects.

Firstly, a scattered sound field of sound waves after being subjected to digital processing is used as a training set, namely an input layer;

second, the meta-surface phase modulation factor is variable, and is analogous to the connection weight of neurons, so that meta-surface phase units can be analogous to neurons;

third, the action of the multilayer acoustic super surface on sound waves is similar to that of a multilayer hidden layer;

fourthly, the sound waves are converged in different areas in the plane after passing through the multilayer hidden layers, and the values of the different areas detected by the probe are output layers.

This is the particular training process that recognizes the ten digits 0-9. Firstly, a training set and a test set of the acoustic super-neural network are generated, because the sound wave is generally a scattered field for detecting an object, the scattered field is formed by irradiating plane waves on the object, and the MNIST training set used by the traditional neural network is not the MNIST training set itself.

Initializing a phase plane, and generating M784 initial phase values by using a random function, wherein M is the number of layers of the acoustic super-neural network. A part of digital scattered field in the training set is input into a phase plane, and the distribution of sound pressure of the scattered field on the super surface can be calculated through a wave propagation equation. After the sound wave passes through each super neuron, a phase modulation is added, and the phase modulation is a parameter to be optimized. After the phase modulation, the back sound pressure phase distribution of the phase surface in the super-neural network is completely different from the incident sound pressure phase distribution, the back sound pressure phase distribution is propagated to the next super-surface layer again to obtain the phase modulation again, and after the modulation of the multi-layer super-surface layer, 10 values are obtained by calculating the sum of sound energy of ten areas on the detection plane. And calculating an error function between the ten values and the label of the number, and calculating the gradient of the phase value on the super surface by using the error function and a gradient descent method so as to update the phase of the whole super surface. By continuously inputting the training set, the phase values on the super-surface are continuously updated, eventually reaching a stable and recognizable number.

The phase required for the ultrasound surface can be obtained by training, and it is sub-essential to what acoustic surface element is used for such phase, in our case with a structure in which four helmholtz resonators are placed at the sides of the acoustic waveguide. Specifically, simulation software is used to calculate that the phase of the ultrasonic surface unit for incident and emergent sound waves changes from 0 to 2pi along with the change of the structure geometric parameters and the change of the height h of the sound wave guide. We then need only to fill in building blocks with such super surface units at the desired phase obtained by training to achieve such an acoustic super neural network.

In the invention, after an object to be recognized is determined, the recognized object is taken as a target object, sound pressure distribution formed after sound waves pass through the target object is taken as a training set, a final phase surface is obtained through simulation by utilizing the existing neural network theory, in the acoustic super surface, after the phase is determined, the acoustic super surface meeting conditions can be designed, the acoustic super surface is placed according to the design requirements, after incident sound waves pass through the trained acoustic super surface, the target object can be recognized through a probe behind the acoustic super surface, and the recognition task can be completed only through the trained acoustic super surface without calculation of a processor.

Has the advantages that: the acoustic method for identifying the shape of the object in real time only needs a small number of transducers, the neural network can utilize scattered acoustic energy to analyze in real time, a data post-processing device is not needed, the overall structure size is the wavelength order, and the acoustic method is simple in structure and easy to implement.

Drawings

FIG. 1(a) is an acoustic super-neural network operation diagram, and (b) and (c) are super-neural network and traditional neural network diagrams.

Fig. 2(a) shows a graph in which the accuracy of classification increases with the number of layers, and (b) shows a graph in which the accuracy of classification of a two-layer super neural network increases with the number of training algebras, and the total error of a test set decreases with the number of training algebras.

FIGS. 3(a) and (b) fusion matrix and energy distribution of two-layer super neural networks, (d) is the number 0 and its acoustic energy distribution at the output face, (c) and (e) the energy distribution in the simulation and experiment of the 20 numbers selected in the experiment.

Detailed Description

The present invention will be further described with reference to the accompanying drawings.

By way of example with the number '8', it is illustrated how to make an acoustic super-surface corresponding to the identification number, the super-surface being in two layers, each layer of super-surface comprising 28 × 28 cells, comprising the following steps:

step 1, after sound waves pass through a digital '8', sound pressure distribution on a cross section which is vertical to a propagation direction after a formed scattering sound field is propagated for a certain distance is used as a data set, a label of the data set is consistent with a target object, and partial data in the data set are randomly extracted and respectively used as a training set and a test set;

step 2, inputting the training set into an acoustic super-neural network, wherein the acoustic super-neural network is obtained by software simulation and calculation of the existing formula by using MATLAB (matrix laboratory), and a certain sound pressure distribution is formed on a detection surface after two layers of acoustic super-surfaces are processed;

step 3, dividing the detection surface into 10 areas, and acquiring the total acoustic energy value of the 10 areas;

step 4, calculating the error between the obtained 10 values and the label of the object, and calculating the gradient of the phase modulation applied by each super neuron by using the error so as to update the gradient, thereby updating the phase until stable output is obtained and the target object can be correctly identified;

and 5, determining the phase deviation value of each unit on the super surface through the updated phase value, manufacturing the acoustic super surface according to the phase deviation value, and filling the corresponding super surface units in a building block manner.

As shown in fig. 1, we constructed two layers (M ═ 2) of hypernerves with 28 units per layer (n ═ 28, M ═ 28) using handwritten numbers in MNIST datasets as real-world acoustic scatterers. Taking the figure '8' as an example, to illustrate the training process, when a plane wave is incident on an object of the figure '8', the sound wave is scattered, a certain sound pressure distribution is formed on the first super surface, and the sound wave on each super neuron is added with a phase modulation after passing through the super neuron. This will result in a redistribution of the acoustic field, with new acoustic field re-radiating onto the second super-surface, which continues to phase modulate the acoustic field, eventually resulting in a distribution of the acoustic field over the detection plane. The detection surface is divided into 10 (N is 10) regions, the 10 regions correspond to 10 numbers to be classified, and the region with the maximum acoustic energy on the detection surface is the shape of the object predicted by the sonosuper-neural network. For example, if the acoustic energy in the region '6' is the largest, then the system predicts that the plane wave is impinging on the number '6', which is not the actual case (number '8'), indicating a recognition error; accordingly, if the acoustic energy of the region '8' on the detection plane is the largest, the system predicts that the plane wave is irradiated on the number '8', which coincides with the actual situation (number '8') indicating correct recognition. During the training process, the 10 values are taken out, the size of the error between the 10 values and the label of the input number per se is evaluated by applying a cross entropy loss function, and the phase change gradient corresponding to the phase modulation of each super neuron on the super surface is calculated, namely, the error between the output label and the label of the number per se can be reduced by changing the phase value on the super surface to which direction. By continuously inputting the numbers, the error between the actual label corresponding to the numbers in the test set and the label predicted by the neural network is continuously reduced, and finally the numbers can be accurately identified.

FIG. 1(a) is an acoustic super-neural network operation diagram, and (b) and (c) are super-neural network and traditional neural network diagrams.

Fig. 2(a) shows that the accuracy of classification increases with the number of layers, (b) shows that the classification accuracy of a two-layer hyperneural network increases with the number of training algebras, and the total error of the test set decreases with the number of training algebras. To verify the utility of the acoustic device, we constructed a two-layer acoustic hyper-neural network for experiments. We randomly chosen 20 samples (two samples per number) for the experiment. In fig. 3 a is the confusion matrix for the identification accuracy in the simulation and b is the total energy distribution map. c and e are the energy distribution plots of the 20 numbers under the experiment in the simulation and experiment, and we can see that the energy is focused to the correct region. d is the energy profile of the number 0.

The above description is only of the preferred embodiments of the present invention, and it should be noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the invention and these are intended to be within the scope of the invention.