Method for establishing recognition model of picosecond ultrasonic measurement signal feature and recognition method

Abstract

The application discloses a picosecond ultrasonic measurement signal feature recognition model establishing method and a recognition method, belongs to the picosecond ultrasonic measurement field, and comprises the following steps: obtaining multiple groups of one-dimensional time sequence picosecond ultrasonic measurement signals measured from different types of micro-nano films, marking time domain intervals and types of each feature signal, and obtaining multiple samples; a hybrid neural network model comprising at least one 1D convolutional neural network layer and at least one recurrent neural network layer is built, and is used for predicting time domain distribution probability curves of various feature signals and non-feature signals in a target one-dimensional time sequence picosecond ultrasonic measurement signal; all samples are divided into a training set, a verification set and a test set; after the hybrid neural network model is trained, verified and tested, a picosecond ultrasonic measurement signal feature recognition model is obtained. The application can simultaneously capture the spatial characteristics and the time characteristics of the picosecond ultrasonic measurement signal, so that the accuracy of picosecond ultrasonic measurement signal feature recognition is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of picosecond ultrasonic measurement, and more particularly relates to a method for establishing a recognition model of picosecond ultrasonic measurement signal features and a recognition method. BACKGROUND

[0002] Photoacoustic measurement is a physical property characterization technology based on the photoacoustic effect. The basic principle is to use a periodically intensity-modulated light source to excite sound waves in non-transparent materials through the photoacoustic effect, and then to deduce the relevant physical properties of the sample under test according to the propagation characteristics of the sound waves in the sample under test. Picosecond ultrasonic measurement is a branch of photoacoustic measurement. Specifically, it uses ultrashort pulse laser to excite ultrasonic waves with a frequency as high as tens of GHz or even THz, achieving nanoscale high-resolution characterization. With the advantages of high resolution, speed, non-destructive, etc., picosecond ultrasonic measurement has become one of the key characterization technologies in the integrated circuit manufacturing process.

[0003] A picosecond ultrasonic measurement system is usually built based on a pump-probe structure. The basic principle of this structure is to divide the ultrashort pulse laser into two beams. The beam with higher energy is called pump light, which is used to excite high-frequency ultrasonic waves in the sample under test. The beam with lower energy is called probe light, which is used to detect the generation and propagation of sound waves in the sample. During the generation and propagation of sound waves, the intensity, phase, deflection angle, etc. of the probe light will change accordingly. This change is received by the detector and recorded in time sequence to form a one-dimensional time sequence signal, i.e. a picosecond ultrasonic measurement signal. This signal contains some characteristic signals, such as: a zero signal formed by a sharp change in the amplitude of the measurement signal at the moment of sound wave generation; and a echo signal formed by a short-term amplitude fluctuation of the measurement signal when the longitudinally propagating sound wave is reflected back to the sample surface. By recognizing, extracting and analyzing the characteristic signals in the picosecond ultrasonic measurement signal, the physical properties of the sample under test such as film thickness, density, interface adhesion, etc. can be obtained.

[0004] The existing feature recognition method is mostly to locate the zero point and echo in the measurement signal by using peak seeking algorithm, however, such method is greatly affected by noise, and in actual measurement environment, the problems of poor robustness and low accuracy often occur. In the patent document with the application publication number CN113607068A, a photoacoustic measurement signal feature recognition model establishment method and extraction method are disclosed, which modifies all convolution layers, maximum pooling layers and up-sampling in the U-Net model into 1D form to establish a one-dimensional U-Net neural network model, and constructs a corresponding data set to train, verify and test the model, and obtains a model for photoacoustic measurement signal feature recognition. Since the one-dimensional U-Net neural network model is a deep learning model, it can learn the inherent law of data in the training process, and can accurately recognize the feature signal in the one-dimensional time sequence photoacoustic measurement signal obtained in the complex noise environment, and has high accuracy and robustness.

[0005] In the above patent document, the one-dimensional U-Net neural network model belongs to a convolutional neural network model, and the data law it captures is mainly the spatial characteristics of the signal. For signals with large amplitude, good results can be achieved in actual application. However, the amplitude of picosecond ultrasonic measurement signal is usually in the range of 10 -4 ~ 10 -8 , which belongs to weak signal, and its spatial characteristics are easily covered by noise. In actual engineering environment, the noise level is high, and the model established based on this method usually has low accuracy in identifying the characteristics of picosecond ultrasonic measurement signal. SUMMARY

[0006] In view of the defects and improvement needs of the prior art, the present application provides a picosecond ultrasonic measurement signal feature recognition model establishment method and recognition method, which aims to establish a new picosecond ultrasonic measurement signal feature recognition model, so as to capture the spatial characteristics and time characteristics of picosecond ultrasonic measurement signal at the same time, thereby improving the accuracy of picosecond ultrasonic measurement signal feature recognition.

[0007] To achieve the above purpose, according to one aspect of the present application, a picosecond ultrasonic measurement signal feature recognition model establishment method is provided, comprising:

[0008] Different kinds of micro-nano thin films are used as samples respectively, at least one group of one-dimensional time sequence picosecond ultrasonic measurement signals of each sample is obtained, and the time domain interval of each feature signal in each group of one-dimensional time sequence picosecond ultrasonic measurement signals and the corresponding type are labeled; the number of film layers of each sample is N, N is a positive integer; the feature signals include zero point signals and echo signals returned by each film layer of the sample;

[0009] A hybrid neural network model comprising at least one 1D convolutional neural network layer and at least one recurrent neural network layer is built to predict the time-domain distribution probability curves of various feature signals and non-feature signals in a target one-dimensional time-series picosecond ultrasonic measurement signal; the target one-dimensional time-series picosecond ultrasonic measurement signal is generated by a micro-nano film with N film layers in a picosecond ultrasonic measurement;

[0010] Each group of one-dimensional time-series picosecond ultrasonic measurement signals and its corresponding labeled results are taken as a sample, and after standardization, all samples are divided into a training set, a validation set and a test set, and after training, validation and testing of the hybrid neural network model using the training set, the validation set and the test set, a picosecond ultrasonic measurement signal feature recognition model is obtained.

[0011] Further, the convolution kernels in the 1D convolutional neural network layer are all greater than or equal to 9x1.

[0012] Further, the recurrent kernels in the recurrent neural network layer are all LSTM recurrent kernels.

[0013] Further, the number of memories in the recurrent kernel of the recurrent neural network layer is all greater than or equal to 100.

[0014] Further, the obtained one-dimensional time-series picosecond ultrasonic measurement signals are all obtained by picosecond ultrasonic measurement in an actual engineering environment.

[0015] Further, for each sample, multiple groups of one-dimensional time-series picosecond ultrasonic measurement signals are measured, and the multiple groups of one-dimensional time-series picosecond ultrasonic measurement signals are measured at different measurement points on the sample.

[0016] Further, before labeling the one-dimensional time-series picosecond ultrasonic measurement signals, the one-dimensional time-series picosecond ultrasonic measurement signals are subjected to data enhancement.

[0017] The data enhancement method includes adding noise.

[0018] Further, when training the hybrid neural network model, weighted cross-entropy is used as the loss function.

[0019] According to another aspect of the present application, a picosecond ultrasonic measurement signal feature recognition method is provided, comprising:

[0020] For a micro-nano film to be measured with N film layers, its one-dimensional time-series picosecond ultrasonic measurement signal is obtained and input into the picosecond ultrasonic measurement signal feature recognition model to predict the time-domain distribution probability curves of various feature signals and non-feature signals in the one-dimensional time-series picosecond ultrasonic measurement signal;

[0021] Corresponding to the maximum value of the probability value of each time domain distribution probability curve at the same time point, the category is taken as the category of the signal at the time point.

[0022] The recognition model of the picosecond ultrasonic measurement signal feature is established by the method for establishing the recognition model of the picosecond ultrasonic measurement signal feature provided in the application.

[0023] According to another aspect of the application, a computer readable storage medium is provided, comprising a stored computer program; when the computer program is read by a processor, the computer readable storage medium controls the device where the computer readable storage medium is located to execute the method for establishing the recognition model of the picosecond ultrasonic measurement signal feature provided in the application, and / or the recognition method of the picosecond ultrasonic measurement signal feature provided in the application.

[0024] Overall, the above technical solutions conceived by the application can achieve the following beneficial effects:

[0025] (1) The recognition model of the picosecond ultrasonic measurement signal feature established by the application comprises at least one 1D convolutional neural network layer and at least one recurrent neural network layer, wherein the 1D convolutional neural network layer is used to capture the spatial characteristics of the signal, and the recurrent neural network layer is used to capture the temporal characteristics of the signal. Since the temporal characteristics of the picosecond ultrasonic measurement signal are less affected by noise, the model can accurately capture the temporal characteristics of the signal even in the case of high noise level, and cooperate with the spatial characteristics to effectively improve the recognition accuracy of the picosecond ultrasonic measurement signal feature.

[0026] (2) Each feature signal in the picosecond ultrasonic measurement signal comprises at least two feature points (extreme points / inflection points). Based on this, in the preferred scheme of the application, the recognition model of the picosecond ultrasonic measurement signal feature established by the application has a larger convolution kernel (all greater than or equal to 9x1) in the 1D convolutional neural network layer, which can fully capture the spatial characteristics of the picosecond ultrasonic measurement signal.

[0027] (3) In picosecond ultrasonic measurement, at least two feature signals in the picosecond ultrasonic measurement signal need to be extracted to complete subsequent measurement and calculation, so the sequence length of the one-dimensional time sequence picosecond ultrasonic measurement signal is often long, and there may be a gradient disappearance problem in the model training process. Based on this, in the preferred scheme of the application, the recognition model of the picosecond ultrasonic measurement signal feature established by the application has LSTM recurrent kernels in the recurrent neural network layer, which can effectively avoid the problem of gradient disappearance that may occur in the model training process. In a further preferred scheme, the number of memory cells in the recurrent kernel of the recurrent neural network layer is large (greater than or equal to 100), which can ensure sufficient capture of the temporal characteristics of the one-dimensional time sequence picosecond ultrasonic measurement signal.

[0028] (4) In the preferred scheme of the present application, the construction of the data set of the model is optimized, specifically including: obtaining a one-dimensional time sequence picosecond ultrasonic measurement signal in an actual engineering environment, measuring a one-dimensional time sequence picosecond ultrasonic measurement signal at multiple different measurement points for the same sample, and adding noise. These means can make the noise level of the samples in the data set as close as possible to the noise level of the one-dimensional time sequence picosecond ultrasonic measurement signal actually measured, thereby improving the generalization performance of the trained model, and being beneficial to improving the feature recognition accuracy of the picosecond ultrasonic measurement signal in actual application.

[0029] (5) In the one-dimensional time sequence picosecond ultrasonic measurement signal, the proportion of various types of feature signals and non-feature signals is usually unbalanced, which can cause the trained model to be biased towards the category with a higher proportion when performing feature recognition; based on this, in the preferred scheme of the present application, the weighted cross-entropy is used as the loss function, which can effectively avoid the influence of class imbalance on the performance of the model. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The flow chart of the picosecond ultrasonic measurement signal feature recognition model establishment method provided for the embodiments of the present application;

[0031] Figure 2 The one-dimensional time sequence picosecond ultrasonic measurement signal and the labeling result after classification labeling by Labelme provided for the embodiments of the present application;

[0032] Figure 3 The mixed neural network model schematic diagram provided for the embodiments of the present application;

[0033] Figure 4 The model prediction result of the time domain distribution probability curve of various types of feature signals and non-feature signals in the target one-dimensional time sequence picosecond ultrasonic measurement signal provided for the embodiments of the present application;

[0034] Figure 5 The model recognition result schematic diagram of various types of feature signals in the target one-dimensional time sequence picosecond ultrasonic measurement signal provided for the embodiments of the present application;

[0035] Figure 6 The feature recognition result of the target one-dimensional time sequence picosecond ultrasonic measurement signal with a signal-to-noise ratio of about 40 dB by the recognition model established according to the present application;

[0036] Figure 7 The feature recognition result of the target one-dimensional time sequence picosecond ultrasonic measurement signal with a signal-to-noise ratio of about 40 dB by the recognition model established according to the Chinese invention patent CN113607068A. DETAILED DESCRIPTION

[0037] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0038] In the present application, the terms "first", "second", etc. (if any) in the present application and the drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0039] In order to solve the problem of low accuracy of existing photoacoustic measurement signal feature recognition model based on one-dimensional U-Net model in identifying picosecond ultrasonic measurement signal under the noise level of actual engineering environment, the present application provides a picosecond ultrasonic measurement signal feature recognition model establishment method and a recognition method, the overall idea of which is: in-depth analysis of the reason why the existing photoacoustic measurement signal feature recognition model has low recognition accuracy when used to identify the characteristic signal of picosecond ultrasonic measurement signal, i.e. one-dimensional U-Net model can only effectively capture the spatial characteristics of the signal, and the amplitude of picosecond ultrasonic measurement signal is low, so under the noise level of actual measurement environment, the spatial characteristics of the signal are easy to be covered by noise; based on this, a new picosecond ultrasonic measurement signal feature recognition model is established, which can capture the spatial characteristics and time characteristics of picosecond ultrasonic measurement signal at the same time, thereby improving the accuracy of picosecond ultrasonic measurement signal feature recognition.

[0040] The following is an embodiment.

[0041] Embodiment 1

[0042] A picosecond ultrasonic measurement signal feature recognition model establishment method, as shown in Figure 1 , includes the following steps (S1) to (S7):

[0043] (S1) Obtain at least one group of one-dimensional time sequence picosecond ultrasonic measurement signals of each sample by taking different types of micro-nano films as samples.

[0044] In this embodiment, the number of film layers of different types of micro-nano films is the same, which is N=1, and the micro-nano film is a metal film; the types of micro-nano films are different, specifically at least one of the attributes of the film layer material, the substrate material and the film layer thickness of the micro-nano film is different;

[0045] The one-dimensional time sequence picosecond ultrasonic measurement signal of the sample can be obtained by using an existing picosecond ultrasonic measurement system or by theoretical modeling simulation. In this embodiment, the one-dimensional time sequence picosecond ultrasonic measurement signal of the sample is obtained by using an existing picosecond ultrasonic measurement system to perform picosecond ultrasonic measurement on the sample. The specific process of the picosecond ultrasonic measurement is as follows: the output of the ultrashort pulse laser is divided into two beams with an energy ratio of about 9:1; the beam with higher energy is irradiated to the position of the measurement point on the surface of the micro-nano film sample, used to excite high-frequency ultrasonic waves at the near-surface of the sample, and this beam is called pump light; the beam with lower energy is irradiated to the same position on the surface of the sample after passing through an adjustable time delay line, used to detect the generation and propagation process of the acoustic wave in the sample, and this beam is called probe light; the probe light reflected by the surface of the sample is received by a photodetector and converted into an electrical signal, which is then output to a computer for processing; the time delay line is adjusted to make the photodetector output electrical signals corresponding to different time points in the generation and propagation process of the acoustic wave; the electrical signals corresponding to different time points are recorded in chronological order by the computer, thereby obtaining the one-dimensional time sequence picosecond ultrasonic measurement signal of the sample. It should be noted that the description of the picosecond ultrasonic measurement is only an optional embodiment of the present application and should not be understood as the only limitation of the present application. In some other embodiments of the present application, other measurement systems or methods can be used to obtain the one-dimensional time sequence picosecond ultrasonic measurement signal of the sample.

[0046] In order to make the noise level of the training data of the model as close as possible to the noise level of the one-dimensional time sequence picosecond ultrasonic measurement signal obtained by actual measurement, thereby improving the generalization performance of the trained model, and facilitating the improvement of the feature recognition accuracy of the picosecond ultrasonic measurement signal in actual application, the above-mentioned picosecond ultrasonic measurement is performed in an actual engineering environment, and for the same sample, a plurality of one-dimensional time sequence picosecond ultrasonic measurement signals of the sample are obtained by changing the measurement points. By obtaining the one-dimensional time sequence picosecond ultrasonic measurement signal at different measurement points on the same sample, the sample noise caused by the surface roughness, internal defects and interface defects of the micro-nano film can be included in the training samples as much as possible, further improving the generalization ability of the trained model.

[0047] In order to provide sufficient training data for subsequent neural network training and ensure the model training effect, the number of one-dimensional time sequence picosecond ultrasonic measurement signals obtained in this embodiment is n≥30.

[0048] (S2) Data augmentation is performed on the basis of the obtained one-dimensional time sequence picosecond ultrasonic measurement signal.

[0049] From the time cost of measurement, the number of one-dimensional time sequence picosecond ultrasonic measurement signals obtained in step S1 is often limited, and it is difficult to cover all scenarios, so it is necessary to expand the sample size of the measurement signal through data enhancement, thereby improving the generalization ability of the neural network model trained subsequently;

[0050] Optionally, in the embodiment, data enhancement is realized by adding white noise to each obtained one-dimensional time sequence picosecond ultrasonic measurement signal, specifically, four different intensity white noises are added to each one-dimensional time sequence picosecond ultrasonic measurement signal, and the signal-to-noise ratios corresponding to the four intensities are 60 dB, 50 dB, 40 dB and 30 dB respectively.

[0051] Through data enhancement, the training data is expanded from n groups of one-dimensional time sequence picosecond ultrasonic measurement signals to n x 4 groups of one-dimensional time sequence picosecond ultrasonic measurement signals; at the same time, the data enhancement method of adding noise can make the final recognition model more robust under noise conditions.

[0052] (S3) Label the time domain interval and the corresponding type of each feature signal in each one-dimensional time sequence picosecond ultrasonic measurement signal.

[0053] For a metal film sample, the feature signals in the one-dimensional time sequence picosecond ultrasonic measurement signal include zero signals and echo signals returned by each film layer. When high-frequency ultrasonic waves are generated near the surface of the sample, they will cause a dramatic change in the optical properties of the sample surface, resulting in a dramatic change in the intensity and other properties of the reflected probe light, which appears as a high-amplitude, steep left and relatively flat right mutation signal in the one-dimensional time sequence picosecond ultrasonic measurement signal, which is called a zero signal. During the longitudinal propagation of ultrasonic waves in the sample, partial reflection occurs at the sample film interface, causing changes in the optical properties of the sample surface and changes in the intensity and other properties of the probe light, which appears as a low-amplitude fluctuation signal in the one-dimensional time sequence picosecond ultrasonic measurement signal, which is called an echo signal. When identifying the one-dimensional time sequence picosecond ultrasonic measurement signal, the echo signals reflected from different film layer interfaces belong to different categories of feature signals. For a sample with N film layers, the number of feature signal categories in the one-dimensional time sequence picosecond ultrasonic measurement signal is N+1, where the number of zero signal categories is 1 and the number of echo signal categories is N. In the embodiment, the number of films of the sample is N = 1, so the one-dimensional time sequence picosecond ultrasonic measurement signal contains two types of feature signals, namely zero signals and echo signals reflected from the film layer and the substrate interface.

[0054] For a known micro-nano film sample, according to the picosecond ultrasonic correlation theory, the spatial morphology and time domain interval of each characteristic signal in the one-dimensional time sequence picosecond ultrasonic measurement signal can be known. Based on this, under the condition that the annotated time domain interval contains the complete morphology characteristics of the characteristic signal and the time domain interval of each characteristic signal does not overlap with that of other characteristic signals, the time domain interval and corresponding type of each characteristic signal in the one-dimensional time sequence picosecond ultrasonic measurement signal can be annotated;

[0055] Optionally, in the embodiment, the above annotation work is completed with the help of the deep learning image annotation tool Labelme software: after converting the data file for saving the one-dimensional time sequence picosecond ultrasonic measurement signal into an image file, the image file is opened in Labelme, and the polygon tool is used to perform quadrilateral frame selection on the regions where the above two types of characteristic signals are located in the image, Figure 2 As shown in the figure, a group of one-dimensional time sequence picosecond ultrasonic measurement signals and the annotation result schematic diagram after classification labeling by Labelme are shown, wherein the category label of the zero signal is zero, and the category label of the echo signal is echo.

[0056] In the embodiment, after the annotation is completed, the obtained annotation result (i.e., the category of each labeled frame and the horizontal and vertical coordinate information of each vertex thereof) is saved as a.json file, and n×4.json files corresponding to n×4 data files of one-dimensional time sequence picosecond ultrasonic measurement signals are obtained.

[0057] (S4) Each group of one-dimensional time sequence picosecond ultrasonic measurement signals and the corresponding annotation result is taken as a sample and is subjected to standardization processing.

[0058] Optionally, in the embodiment, the data file of the one-dimensional time sequence picosecond ultrasonic measurement signal and the.json file containing the annotation result thereof are renamed with the same file name, so that the corresponding group of measurement signals and the annotation result jointly constitute a sample, and n×4 samples are obtained.

[0059] The sample is subjected to standardization processing, including: the one-dimensional time sequence picosecond ultrasonic measurement signal is subjected to standardization processing, and the annotation result thereof is subjected to standardization processing;

[0060] In the embodiment, the one-dimensional time sequence picosecond ultrasonic measurement signal is subjected to standardization processing, specifically: a one-dimensional array in the data file for saving the one-dimensional time sequence picosecond ultrasonic measurement signal is extracted and is interpolated to a specified length L; optionally, in the embodiment, the interpolation method is linear interpolation, and the specified length L is 1000; the one-dimensional array after interpolation is subjected to amplitude standardization according to x i_stan = (x i - μ) / σ, wherein x irepresents the i-th data in the array, μ represents the mean of all data in the array, σ represents the variance of all data in the array, and x i_stan represents the i-th data in the array after amplitude normalization; the one-dimensional array after interpolation and amplitude normalization is saved as a.npy file, named after the original data file, and stored as a signal data file in the specified path, thereby completing the standardization processing of the one-dimensional time sequence picosecond ultrasonic measurement signal;

[0061] In this embodiment, the labeling results of the one-dimensional time sequence picosecond ultrasonic measurement signal are standardized, specifically: the vertex horizontal coordinate information (i.e. the time domain information of the characteristic signal) of each marking box in the.json file used to save the labeling results is extracted; the time domain information of the characteristic signal of the same category is parsed into a one-dimensional array of length L, and the time domain information of the area without marking box (i.e. non-characteristic signal) is also parsed into a one-dimensional array of length L. For a sample with N layers of membranes, the number of categories of characteristic signals is N+1, corresponding to N+1 one-dimensional arrays of length L, plus one one-dimensional array of length L corresponding to non-characteristic signals, a total of N+2 one-dimensional arrays of length L; the above N+2 one-dimensional arrays of length L are combined into a two-dimensional array of size L×(N+2) along the vertical length direction, and saved as a.npy file, named after the original.json file, and stored as a label data file in the specified path, thereby completing the standardization processing of the labeling results; in this embodiment, the number of layers of the sample is N=1, and the length of the one-dimensional array is L=1000, therefore, the single label data file contains a two-dimensional array of size 1000×3, which is composed of three one-dimensional arrays of length 1000, and the three one-dimensional arrays correspond to three signal categories: zero signal (zero), echo signal (echo) and non-characteristic signal (background);

[0062] In this embodiment, after the standardization processing of the sample is completed, n×4 signal data files and n×4 label data files corresponding thereto are obtained, all the signal data files are stored under the signal path, and all the label data files are stored under the label path.

[0063] (S5) Divide all samples into a training set, a validation set and a test set.

[0064] Optionally, in this embodiment, the data proportion of the training set, the validation set and the test set is 7:2:1, and accordingly, the n x 4 signal data files and the n x 4 label data files are randomly sampled and stored in the signal path and the label path in the training set, the validation set and the test set according to the proportion;

[0065] It should be noted that in actual application, the division proportion of the training set, the validation set and the test set can be adjusted according to actual needs.

[0066] (S6) A hybrid neural network model containing at least one 1D convolutional neural network layer and at least one recurrent neural network layer is built for predicting the time domain distribution probability curve of each type of feature signal and non-feature signal in the target one-dimensional time sequence picosecond ultrasonic measurement signal. The target one-dimensional time sequence picosecond ultrasonic measurement signal is generated by a micro-nano film with N layers in picosecond ultrasonic measurement.

[0067] The hybrid neural network model established in this embodiment, in which the 1D convolutional neural network layer is used to capture the spatial characteristics of the signal, and the recurrent neural network layer is used to capture the time characteristics of the signal. Since the time characteristics of the picosecond ultrasonic measurement signal are less affected by noise, this model can accurately capture the time characteristics of the signal even in the case of high noise level, and cooperate with the spatial characteristics to effectively improve the recognition accuracy of the picosecond ultrasonic measurement signal features.

[0068] Considering that each feature signal in the picosecond ultrasonic measurement signal contains at least two feature points (extreme points / inflection points), in order to enable the model to fully capture the spatial characteristics of the picosecond ultrasonic measurement signal, in this embodiment, the convolution kernel size of the 1D convolutional neural network layer is set to be greater than or equal to 9 x 1. At the same time, considering that at least two feature signals in the picosecond ultrasonic measurement signal need to be extracted in order to complete the subsequent measurement and calculation in the picosecond ultrasonic measurement, the sequence length of the one-dimensional time sequence picosecond ultrasonic measurement signal is often long, and there may be a problem of gradient disappearance in the model training process. In order to avoid the possible problem of gradient disappearance in the model training, in this embodiment, the recurrent neural network layer adopts an LSTM recurrent kernel; further, the number of memories in the recurrent kernel of the recurrent neural network layer is set to be greater than or equal to 100, so as to ensure that the model fully captures the time characteristics of the measurement signal.

[0069] It should be noted that in actual application, the number of 1D convolutional neural network layers in the model needs to be determined according to the specific morphology of the characteristic signal in the picosecond ultrasonic measurement signal, so as to ensure that the regularity of the spatial characteristics of the characteristic signal can be fully learned; the number of recurrent neural network layers in the model needs to be determined according to the richness of the data, so as to ensure that the regularity of the time characteristics of the characteristic signal can be fully learned; and the relative position relationship between the 1D convolutional neural network layer and the recurrent neural network layer needs to be adjusted according to the final training result.

[0070] Finally, the hybrid neural network model established in this embodiment has a straight line structure as shown in Figure 3 The first layer is a 1D convolutional neural network layer containing 8 convolutional kernels of 9x1 size, the second layer is a 1D convolutional neural network layer containing 4 convolutional kernels of 9x1 size, the third layer is an LSTM recurrent neural network layer containing 100 memory cells, and the fourth layer is a time-distributed fully connected layer with a second dimension of N+2 for controlling the output of the model.

[0071] In this embodiment, the input of the hybrid neural network model is a one-dimensional array of length 1000 extracted from the signal data file, and the output of the model is a two-dimensional array of 1000x3, which can be decomposed into three one-dimensional arrays of length 1000, corresponding to the prediction results of the time-domain distribution probability curves of the three signals (zero, echo and background) respectively.

[0072] (S7) After training, verifying and testing the hybrid neural network model using the training set, the verification set and the test set respectively, the recognition model of the picosecond ultrasonic measurement signal characteristics is obtained.

[0073] After the above neural network model is constructed, the data in the signal data file in the training set is used as the input of the model, the output of the model is compared with the data in the corresponding label data file, the loss is calculated, and the learning parameters such as model weight are updated by backward propagation of parameters, so as to realize the training of the model;

[0074] The trained model is verified using the verification set, and the following adjustments are made according to the training and verification results to optimize the training effect of the model: adjusting the model hyperparameters such as the size of the convolutional kernel, the number of convolutional kernels, the number of memory cells in the recurrent kernel, adjusting the model structure characteristics such as the number of network layers, the order, the arrangement mode, and adjusting the training parameters such as epoch and learning rate;

[0075] The trained and verified model is tested by using the test set to determine whether the accuracy of signal feature recognition and classification of the model meets the standard. If not, the model hyperparameters, structure and training parameters are adjusted, and the training, verification and testing are performed again until the test result of the model meets the standard, so as to obtain the picosecond ultrasonic measurement signal feature recognition model.

[0076] Considering that the proportion of various feature signals and non-feature signals in the one-dimensional time sequence picosecond ultrasonic measurement signal is usually unbalanced, which may cause the model trained to be biased towards the category with a higher proportion during feature recognition, in the embodiment, a weighted cross-entropy is used as a loss function in the above training process to avoid the influence of class imbalance on the performance of the model.

[0077] Optionally, in the above training, verification and testing process, an intersection over union (IOU) is used as an evaluation index for the pros and cons of the model. Optionally, after obtaining the picosecond ultrasonic measurement signal feature recognition model, the model is saved as a.h5 file for calling.

[0078] In the embodiment, the final picosecond ultrasonic measurement signal feature recognition model can be used to predict the time domain distribution probability curve of various feature signals in the one-dimensional time sequence picosecond ultrasonic measurement signal of the micro-nano film with 1 film layer. It should be noted that if the number of film layers N of the micro-nano film sample in the S1 step changes, the number of feature signal categories N+1 and the total number of signal categories N+2 will also change. Only the operations related to the three numbers in the subsequent S2-S7 steps need to be adjusted accordingly. The final picosecond ultrasonic measurement signal feature recognition model can be used to predict the time domain distribution probability curve of various feature signals in the one-dimensional time sequence picosecond ultrasonic measurement signal of the micro-nano film with N film layers.

[0079] A group of target one-dimensional time sequence picosecond ultrasonic measurement signals are obtained by performing picosecond ultrasonic measurement on a micro-nano metal single-layer film, and the target one-dimensional time sequence picosecond ultrasonic measurement signals are standardized according to the standardization method of the one-dimensional time sequence picosecond ultrasonic measurement signal described in the S4 step of the above embodiment. The feature recognition of the standardized target one-dimensional time sequence picosecond ultrasonic measurement signals is performed by using the picosecond ultrasonic measurement signal feature recognition model obtained in the above embodiment. Specifically, the model is obtained by calling the corresponding.h5 file, and the one-dimensional array in the signal data file corresponding to the target one-dimensional time sequence picosecond ultrasonic measurement signal is extracted and input into the model to obtain a two-dimensional array with a size of 1000x3 output by the model. The two-dimensional array is decomposed into three one-dimensional arrays with a length of 1000, and the three arrays are plotted in the same coordinate system as the input of the model, so that the probability curve of the time domain distribution of various feature signals in the one-dimensional time sequence picosecond ultrasonic measurement signal of the micro-nano film with 1 film layer is obtained as shown in FIG. 6.Figure 4 The model prediction result of the time-domain distribution probability curve of each type of feature signal and non-feature signal in the target one-dimensional time sequence picosecond ultrasonic measurement signal is shown.

[0080] Embodiment 2:

[0081] A picosecond ultrasonic measurement signal feature recognition method comprises:

[0082] For a micro-nano metal film to be measured with N film layers, a one-dimensional time sequence picosecond ultrasonic measurement signal thereof is obtained and input into a picosecond ultrasonic measurement signal feature recognition model to predict the time-domain distribution probability curve of each type of feature signal and non-feature signal in the one-dimensional time sequence picosecond ultrasonic measurement signal. It is easy to understand that the sum of the probability values of each time-domain distribution probability curve at the same time point is 1.

[0083] By comparing the probability values of each time-domain distribution probability curve at the same time point, the type corresponding to the maximum value is taken as the type of the signal at the time point.

[0084] The picosecond ultrasonic measurement signal feature recognition model is established by the picosecond ultrasonic measurement signal feature recognition model establishment method provided in Embodiment 1.

[0085] Take Figure 4 for example. The sum of the three probability curve values at the same horizontal coordinate is 1. The three vertical coordinate values are the probability values of the signal point at the horizontal coordinate belonging to the three signal types of zero, echo and background. The type corresponding to the maximum value in the three probability values is the model prediction result of the signal type of the signal point. Accordingly, the model recognition result of each type of feature signal in the target one-dimensional time sequence picosecond ultrasonic measurement signal can be further obtained from the time-domain distribution probability curve in Figure 4 , as shown in Figure 5 . The vertical coordinate value of 0 indicates that it does not belong to this type, and the vertical coordinate value of 1 indicates that it belongs to this type.

[0086] In actual application, after the signal types of each signal point in the one-dimensional time sequence picosecond ultrasonic measurement signal are recognized, the feature signals of each type can be easily extracted.

[0087] Embodiment 3:

[0088] A kind of picosecond ultrasonic measurement signal feature identification model establishment method, this embodiment is similar to the above-mentioned embodiment, the difference is that, in this embodiment, the sample is transparent film of non-transparent material substrate, the characteristic signal in one-dimensional time sequence picosecond ultrasonic measurement signal obtained by the sample through picosecond ultrasonic measurement, in addition to zero signal and each film layer returned echo signal, it also includes Brillouin oscillation signal, the characteristic signal is in the form of periodic oscillation, and oscillation period is determined by the sound velocity of transparent film, refractive index and the wavelength of probe light, incident angle, correspondingly, the class number of characteristic signal in one-dimensional time sequence picosecond ultrasonic measurement signal is N+2, wherein the class number of zero signal is 1, the class number of echo signal is N, and the class number of Brillouin oscillation signal is 1;In the mixed neural network model, the fourth layer is a time distribution type full connection layer with the second dimension of N+3 for controlling the output of the model.

[0089] Embodiment 4:

[0090] A picosecond ultrasonic measurement signal feature identification method, comprising:

[0091] For the micro-nano transparent film to be measured with N film layers and non-transparent material substrate, obtain its one-dimensional time sequence picosecond ultrasonic measurement signal and input it into the picosecond ultrasonic measurement signal feature identification model to predict the time domain distribution probability curve of each type of characteristic signal and non-characteristic signal in the one-dimensional time sequence picosecond ultrasonic measurement signal.

[0092] Compare the probability values of each time domain distribution probability curve at the same time point, and take the class corresponding to the maximum value as the class of the signal at the time point.

[0093] The picosecond ultrasonic measurement signal feature identification model is established by the picosecond ultrasonic measurement signal feature identification model establishment method provided in the above-mentioned embodiment 3.

[0094] Embodiment 5:

[0095] A computer-readable storage medium includes a stored computer program;When the computer program is read by the processor, the computer-readable storage medium controls the device where the computer-readable storage medium executes the picosecond ultrasonic measurement signal feature identification model establishment method provided in the above-mentioned embodiment 1 or 3, and / or the picosecond ultrasonic measurement signal feature identification method provided in the above-mentioned embodiment 2 or 4.

[0096] The beneficial effects that can be achieved by the present application are further explained and described below in connection with the feature recognition results of picosecond ultrasonic measurement signals by different models. Specifically, a group of target one-dimensional time sequence picosecond ultrasonic measurement signals with a signal-to-noise ratio of about 40 dB are subjected to feature recognition by using the models established according to the feature recognition model establishment method disclosed in the present application and the feature recognition model establishment method disclosed in CN113607068A, respectively.

[0097] Figure 6 The characteristic recognition result of the target one-dimensional time sequence picosecond ultrasonic measurement signal with a signal-to-noise ratio of about 40 dB according to the model established by the present application is shown, and the coincidence degree with the true result is high, and the corresponding intersection-over-union ratio is as high as 94.87; Figure 7 The characteristic recognition result of the target one-dimensional time sequence picosecond ultrasonic measurement signal according to the model established by the Chinese invention patent CN113607068A is shown, and the coincidence degree with the true result is low, and the corresponding intersection-over-union ratio is only 69.43. It can be seen that, compared with the existing method, the present application can realize higher accuracy of picosecond ultrasonic measurement signal characteristic recognition.

[0098] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the present application and is not intended to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A picosecond ultrasonic measurement signal feature recognition model establishment method, characterized in that, The method comprises the following steps: Each sample has N layers, and N is a positive integer; The characteristic signals include zero signals and echo signals returned by each layer of the sample; A hybrid neural network model comprising at least one 1D convolutional neural network layer and at least one recurrent neural network layer is built to predict the time-domain distribution probability curves of each type of characteristic signal and non-characteristic signal in the target one-dimensional time-series picosecond ultrasonic measurement signal generated by the micro-nano film with N layers in the picosecond ultrasonic measurement; the convolution kernel in the 1D convolutional neural network layer is greater than or equal to 9x1; Each one-dimensional time-series picosecond ultrasonic measurement signal and its corresponding annotation result are standardized and divided into a training set, a validation set and a test set, and the hybrid neural network model is trained, validated and tested using the training set, the validation set and the test set, respectively, to obtain a picosecond ultrasonic measurement signal feature recognition model. The recurrent kernel in the recurrent neural network layer is an LSTM recurrent kernel.

2. The picosecond ultrasonic measurement signal feature recognition model establishment method of claim 1, wherein, The number of memory cells in the recurrent kernel of the recurrent neural network layer is greater than or equal to 100.

3. The picosecond ultrasonic measurement signal feature recognition model establishment method of claim 2, wherein, The obtained one-dimensional time-series picosecond ultrasonic measurement signals are obtained by picosecond ultrasonic measurement in an actual engineering environment.

4. The method of claim 1 to 3, wherein For each sample, multiple one-dimensional time-series picosecond ultrasonic measurement signals are measured, and the multiple one-dimensional time-series picosecond ultrasonic measurement signals are measured at different measurement points on the sample.

5. The method of claim 1 to 3, wherein Before annotating the one-dimensional time-series picosecond ultrasonic measurement signal, the one-dimensional time-series picosecond ultrasonic measurement signal is subjected to data enhancement.

6. The method of claim 1 to 3, wherein The data enhancement method includes adding noise. When training the hybrid neural network model, weighted cross-entropy is used as the loss function.

7. The method of claim 1 to 3, wherein The method comprises the following steps:

8. A method of identifying features of a picosecond ultrasound measurement signal, characterized by, For the micro-nano film with N layers to be measured, its one-dimensional time-series picosecond ultrasonic measurement signal is obtained and input into the picosecond ultrasonic measurement signal feature recognition model to predict the time-domain distribution probability curves of each type of characteristic signal and non-characteristic signal in the one-dimensional time-series picosecond ultrasonic measurement signal; The probability values of the time-domain distribution probability curves at the same time point are compared, and the class corresponding to the maximum value is taken as the class of the signal at the time point. The picosecond ultrasonic measurement signal feature recognition model is established by the picosecond ultrasonic measurement signal feature recognition model establishment method of any one of claims 1-7. The computer program is read by the processor to control the device where the computer readable storage medium executes the picosecond ultrasonic measurement signal feature recognition model establishment method of any one of claims 1-7 and / or the picosecond ultrasonic measurement signal feature recognition method of claim 8.

9. A computer-readable storage medium, characterized in that, ​

Images