Training sample amplification method and device, recognition model generation method and spectrum recognition method

Abstract

The invention provides a training sample amplification method which can be applied to the technical field of laser-induced breakdown spectroscopy data processing and application. The method comprises the following steps: acquiring a real training sample which comprises a laser-induced breakdown spectrum and a corresponding sample label; converting the laser-induced breakdown spectroscopy into laser-induced breakdown spectroscopy images to obtain a sample image group, and dividing the sample image group into a training set and a test set; taking the random signal and the sample label in the training set as input, taking the laser-induced breakdown spectroscopy image in the training set as output, and training the deep generative adversarial neural network model to obtain the trained deep generative adversarial neural network model; and amplifying the training set by using the trained deep generative adversarial neural network model to obtain an amplified training set. By utilizing spectral image conversion, the accuracy of spectral feature extraction is improved, and the problem of insufficient samples is fundamentally solved by utilizing data amplification.

Description

Technical Field

[0001] The present invention relates to the field of laser induced breakdown spectroscopy data processing and application technology, and more specifically to a training sample amplification method and device, recognition model generation and spectrum recognition method. Background Art

[0002] Laser-induced breakdown spectroscopy (LIBS), an emerging rapid elemental analysis technique, has been hailed as the "future star" of chemical analysis due to its multiple advantages, including rapid response, in-situ detection, minimally destructive analysis, no need for complex sample preparation, simultaneous analysis of all elements, and remote detection. Currently, spectral data processing methods have become a hot topic in this research field, as they can significantly enhance LIBS detection performance without increasing hardware costs.

[0003] The traditional LIBS spectral analysis process consists of three key steps: first, spectral signal preprocessing, which involves a series of operations such as noise removal and background subtraction; second, feature extraction of spectral elemental information, which is the basis for identification and analysis; and finally, training a machine learning model based on the extracted features to achieve accurate element identification. However, high-performance machine learning models often rely on large amounts of sample data, a requirement that limits the application of LIBS technology to small sample areas such as biomedicine, space exploration, and cultural relic identification.

[0004] To overcome this limitation, it is necessary to build a new data analysis chain, from spectral generation to identification, starting from the data source. The core of this chain lies in fundamentally solving the problem of insufficient samples. To achieve this, spectral data must be amplified to generate new and diverse spectral data to assist in the training of deep learning models. This approach can improve the stability, accuracy, and automation of the identification model, enabling LIBS technology to better serve the small sample field and demonstrate its broader application potential and value. Summary of the Invention

[0005] (1) Technical issues to be resolved

[0006] To solve at least one of the above problems, the present disclosure provides a training sample amplification method and device, a recognition model generation and a spectral recognition method, which utilizes spectral image conversion to improve the accuracy of spectral feature extraction, utilizes data amplification to fundamentally solve the problem of insufficient samples, and uses the generated spectral images to assist in deep learning model training, thereby improving the stability, accuracy and automation of the recognition model.

[0007] (2) Technical solution

[0008] In response to the above technical problems, the embodiments of the present disclosure provide a training sample amplification method and device, a recognition model generation method, and a spectrum recognition method.

[0009] According to a first aspect of the present disclosure, a training sample amplification method is provided, comprising: obtaining a real training sample, including a laser-induced breakdown spectrum and a corresponding sample label, the sample label being used to represent a material classification corresponding to the laser-induced breakdown spectrum; converting the laser-induced breakdown spectrum into a laser-induced breakdown spectrum image to obtain a sample image group, and dividing the sample image group into a training set and a test set; training a deep generative adversarial neural network model using a random signal and sample labels in a training set as input and the laser-induced breakdown spectrum image in the training set as output to obtain a trained deep generative adversarial neural network model; and amplifying the training set using the trained deep generative adversarial neural network model to obtain an amplified training set.

[0010] In some exemplary embodiments, converting the laser induced breakdown spectrum into a laser induced breakdown spectrum image includes: determining the size of the image to be generated based on the number of bands of the preprocessed laser induced breakdown spectrum; extracting the spectral peaks corresponding to all wavelengths of the preprocessed laser induced breakdown spectrum to obtain the wavelength and intensity of the spectral peak position; based on the wavelength of all spectral peak positions, calculating the position coordinates of the spectral peak on the image to be generated; based on the intensity of the spectral peak position, calculating the pixel value of the spectral peak on the image to be generated; and generating a laser induced breakdown spectrum image based on the position coordinates and pixel value of each spectral peak on the image to be generated.

[0011] In some exemplary embodiments, a training set is augmented using a trained deep generative adversarial neural network model, including: obtaining sample labels in the training set and the number of laser-induced breakdown spectroscopy images corresponding to each sample label; and using the trained deep generative adversarial neural network model to augment the number of laser-induced breakdown spectroscopy images corresponding to each sample label to a preset number.

[0012] In some exemplary embodiments, expanding the number of laser-induced breakdown spectroscopy images corresponding to each sample label to a preset number includes: obtaining a maximum number of laser-induced breakdown spectroscopy images corresponding to the sample label in a training set;

[0013] Using the trained deep generative adversarial neural network model, the number of laser-induced breakdown spectroscopy images corresponding to each sample label was amplified to the maximum value.

[0014] The laser-induced breakdown spectroscopy images that have been amplified to the maximum value are amplified multiple times using a pre-trained deep generative adversarial neural network model.

[0015] In some exemplary embodiments, the convolution layer of the deep generative adversarial neural network model is a two-dimensional convolution kernel, which is generated using the Gauss-Lorentz formula.

[0016] In some exemplary embodiments, before converting the laser induced breakdown spectrum into a laser induced breakdown spectrum image, the method further includes: performing outlier elimination and normalization processing on the laser induced breakdown spectrum to obtain a preprocessed laser induced breakdown spectrum.

[0017] In some exemplary embodiments, the method for eliminating outliers includes a median filtering method; and the normalization method includes a mean-variance normalization method.

[0018] According to a second aspect of the present disclosure, a recognition model generation method is provided, including: setting training weights for the generated graphic group and the real image group in the expanded training set respectively; constructing a loss function based on the training weights; and using the loss function to train a deep convolutional neural network model so that the value of the loss function meets preset requirements, thereby obtaining a recognition model.

[0019] According to a third aspect of the present disclosure, a laser-induced breakdown spectrum recognition method is provided, comprising obtaining a laser-induced breakdown spectrum to be recognized; converting the laser-induced breakdown spectrum to be recognized into a laser-induced breakdown spectrum image; and using a recognition model to recognize the laser-induced breakdown spectrum image to obtain a sample label of the laser-induced breakdown spectrum to be recognized.

[0020] According to a fourth aspect of the present disclosure, a training sample amplification device is provided, comprising: an acquisition module for acquiring real training samples, including a laser-induced breakdown spectrum and a corresponding sample label, wherein the sample label is used to represent the material classification corresponding to the laser-induced breakdown spectrum; an image conversion module for converting the laser-induced breakdown spectrum into a laser-induced breakdown spectrum image to obtain a sample image group, and dividing the sample image group into a training set and a test set; a training module for taking a random signal and a sample label in a training set as input and the laser-induced breakdown spectrum image in the training set as output, to train a deep generative adversarial neural network model to obtain a trained deep generative adversarial neural network model; and an amplification module for amplifying the training set using the trained deep generative adversarial neural network model to obtain an amplified training set.

[0021] (3) Beneficial effects

[0022] It can be seen from the above technical solutions that the training sample amplification method and device, recognition model generation and spectrum recognition method provided by the embodiments of the present disclosure have at least the following beneficial effects:

[0023] (1) The one-dimensional spectrum is converted into a two-dimensional image, which is then processed using two-dimensional convolution combined with the Gauss-Lorentz formula. On this basis, a deep generative adversarial neural network model is used to amplify the image. The amplified image is used to align the data volume of each category, making the data distribution more balanced and improving training efficiency.

[0024] (2) A deep neural network was used to establish a LIBS spectrum recognition model. Compared with traditional machine learning models, the deep neural network can fully learn the differences in element content and the ratio of different elements between categories to fully identify the differences in different types of LIBS spectra, thereby improving the recognition accuracy and stability of the deep neural network. Ultimately, high-precision classification and recognition of small sample spectra can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The above contents and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

[0026] Figure 1 A flowchart of a training sample augmentation method according to an embodiment of the present invention is schematically shown;

[0027] Figure 2 Schematically shows a structural diagram of a deep generative adversarial neural network model according to an embodiment of the present disclosure;

[0028] Figure 3 A flowchart of a method for generating a recognition model according to an embodiment of the present invention is schematically shown;

[0029] Figure 4 Schematically shows a structural diagram of a deep convolutional neural network model according to an embodiment of the present disclosure;

[0030] Figure 5 The following schematically shows a flow chart of a laser induced breakdown spectroscopy identification method according to an embodiment of the present disclosure;

[0031] Figure 6 Schematically shows a schematic diagram of an original spectrum line of a LIBS spectrum according to an embodiment of the present disclosure;

[0032] Figure 7 Schematically shows an image diagram of LIBS spectrum conversion according to an embodiment of the present disclosure;

[0033] Figure 8 A training sample augmentation device according to an embodiment of the present invention is schematically shown; and

[0034] Figure 9 The block diagram schematically shows an electronic device for the training sample augmentation method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concept of the present invention.

[0036] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise," "include," etc. used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0037] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0038] When expressions such as "at least one of A, B, and C, etc." are used, they should generally be interpreted in accordance with the meaning commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0039] Figure 1 The flowchart of a training sample augmentation method according to an embodiment of the present invention is schematically shown.

[0040] like Figure 1 As shown, a training sample augmentation method according to an embodiment of the present invention includes steps S110 to S140.

[0041] In step S110 , a real training sample is obtained, including a laser-induced breakdown spectrum and a corresponding sample label, where the sample label is used to indicate a material classification corresponding to the laser-induced breakdown spectrum.

[0042] In step S120, the laser induced breakdown spectrum is converted into a laser induced breakdown spectrum image to obtain a sample image group, and the sample image group is divided into a training set and a test set. LIBS image group, randomly divide the sample image group into training sets (The sample size is ) and the test set (The sample size is ).

[0043] In some exemplary embodiments, step S120 includes steps S121 - S125 .

[0044] In step S121 , the size of the image to be generated is determined according to the number of wavelength bands of the pre-processed laser induced breakdown spectrum.

[0045] For example, the image size is determined according to the number of bands of the LIBS spectrum. The image size is ,in , is the number of bands, and ceil is the rounding up function.

[0046] In step S122, the peaks corresponding to all wavelengths of the pre-processed laser-induced breakdown spectrum are extracted to obtain the wavelength and intensity of the peak positions.

[0047] In step S123 , based on the wavelengths of all spectrum peak positions, the position coordinates of the spectrum peaks on the image to be generated are calculated.

[0048] For example, the wavelengths of all peak positions Corresponding to the image, for example, the wavelength corresponding to a peak The formula for converting to image position is as follows:

[0049] (1)

[0050] in is the initial wavelength of the entire spectrum, is the last wavelength of the entire spectrum. and are the horizontal and vertical coordinates of the image, respectively. is the floor function, is the remainder function.

[0051] In step S124 , based on the intensity of the spectrum peak position, the pixel value of the spectrum peak on the image to be generated is calculated.

[0052] For example, the spectral intensity Convert to graphics , where the calculation formula for each pixel value is:

[0053] (2)

[0054] in is the maximum value of the spectrum line, and Represent the horizontal and vertical points respectively, and each LIBS spectrum corresponds to an image.

[0055] In step S125 , a laser induced breakdown spectroscopy image is generated based on the position coordinates and pixel values ​​of each spectrum peak on the image to be generated.

[0056] In some exemplary embodiments, before step S120, the process further includes removing outliers and normalizing the laser-induced breakdown spectrum to be identified. The outlier removal method may be a median filter, and the normalization method may be a mean-variance normalization method. Image preprocessing techniques can effectively remove random noise from the spectrum and retain key features.

[0057] For example, for any wavelength range , The LIBS spectra of each band are firstly processed to exclude outliers and normalize them. The outliers are excluded by median filtering. The calculation formula is as follows:

[0058] (3)

[0059] in is the spectrum curve, x is the wavelength, The value of is the intensity value corresponding to the wavelength, where is the filter window size, is the median function, and the normalization is done by mean-variance normalization. The calculation formula is as follows:

[0060] (4)

[0061] in is the mean of the current spectrum, is the variance of the current spectrum.

[0062] In step S130, the random signal and the sample labels in the training set are used as input, and the laser induced breakdown spectrum image in the training set is used as output to train the deep generative adversarial neural network model to obtain a trained deep generative adversarial neural network model.

[0063] Figure 2 The schematic diagram shows the structure of a deep generative adversarial neural network model according to an embodiment of the present disclosure.

[0064] like Figure 2As shown, the deep generative adversarial neural network model according to the embodiment of the present disclosure includes a connection layer, a combination of multiple convolution layers, a normalization layer and an activation layer, and an output layer. Through the setting of the combination of multiple convolution, normalization and activation layers, it is possible to gradually extract more advanced and abstract features and learn a hierarchical representation of the input data. The generative adversarial network model has significant advantages in feature extraction and data representation through the structural design of the connection layer, the combination of multiple convolution-normalization-activation layers and the output layer. Among them, the multiple convolution layers gradually expand the receptive field and gradually extract high-level features (such as shape and semantics) from low-level features (such as edges and textures), forming a hierarchical representation from local to global; the normalization layer (such as BatchNorm) stabilizes the training process and accelerates convergence; the activation layer (such as ReLU) introduces nonlinearity, enabling the model to fit complex distributions. The combination of these two enhances the robustness and expressiveness of features. The generator learns the distribution of real data through a convolution-normalization-activation combination, generating realistic samples. The discriminator learns to distinguish between real and fake data through a similar structure. The two compete to improve performance, forming a positive feedback loop for adversarial training. Connection layers (such as skip connections) combine features from different levels, preserving both details and global information and preventing information loss. The hierarchical representation enables the model to handle input data of varying scales and complexities. Normalization layers mitigate vanishing or exploding gradients, and activation layers avoid linear dependencies, making it easier for the model to converge to the global optimal solution. By learning a hierarchical representation of images, the model can generate high-resolution and diverse images.

[0065] In this embodiment of the present invention, the convolutional layer of the deep generative adversarial neural network model uses a two-dimensional convolution kernel generated using the Gauss-Lorentz formula to improve the accuracy of feature extraction. The convolutional layer takes as input a random signal and sample labels, and outputs an image after passing through the deep generative adversarial neural network.

[0066] In step S140, the training set is amplified using the trained deep generative adversarial neural network model to obtain an amplified training set.

[0067] In some exemplary embodiments, step S140 includes steps S141 - S142 .

[0068] In step S141, sample labels in a training set and the number of laser induced breakdown spectroscopy images corresponding to each sample label are obtained.

[0069] In step S142, the trained deep generative adversarial neural network model is used to expand the number of laser-induced breakdown spectroscopy images corresponding to each sample label to a preset number.

[0070] In some exemplary embodiments, step S142 includes steps S1421 - S1423 .

[0071] In step S1421, the maximum number of laser-induced breakdown spectroscopy images corresponding to the sample labels in the training set is obtained.

[0072] In step S1422, the trained deep generative adversarial neural network model is used to expand the number of laser-induced breakdown spectrum images corresponding to each sample label to a maximum value.

[0073] In step S1423, the pre-trained deep generative adversarial neural network model is used to multiply the laser-induced breakdown spectrum images that have been amplified to the maximum value.

[0074] For example, for training samples, where the number of sample categories is , the number of images of each type of sample is .

[0075] First, align the number of samples and take the largest number of samples As a benchmark, we use the trained deep generative adversarial neural network model Generate separately image samples. The number of samples in the final training set is indivual.

[0076] After aligning the number of samples, the samples were amplified and the amplification factor was , so the number of images in the final training set is , using generated images to align the data volume of each category, making the data distribution more balanced and improving training efficiency.

[0077] After obtaining the expanded training samples based on the above method, combined with Figure 3-Figure 4 A method for obtaining a pre-trained deep convolutional neural network model according to an embodiment of the present disclosure is introduced.

[0078] Figure 3 The flowchart of the recognition model generation method according to an embodiment of the present invention is schematically shown.

[0079] like Figure 3 As shown, the recognition model generation method according to an embodiment of the present invention includes steps S210-S230.

[0080] In step S210 , training weights are set for the generated graphics group and the real image group in the expanded training set respectively.

[0081] For example, after completing sample generation, taking the first type of samples as an example, the number of real samples is , the number of samples generated is . For this, we mark the secondary label, which is used to distinguish the real sample from the generated sample. The training weight is , for generating samples The weight is . And apply it to the training of recognition models.

[0082] In step S220, a loss function is constructed based on the training weights.

[0083] For example, in order to avoid overfitting due to too many generated samples, the designed loss function is as follows:

[0084] (5)

[0085] in is the training set sample, is its true probability distribution, is the predicted probability distribution.

[0086] In step S230, the deep convolutional neural network model is trained using the loss function so that the value of the loss function meets the preset requirements, thereby obtaining a recognition model.

[0087] Figure 4 A schematic diagram of the structure of a deep convolutional neural network model according to an embodiment of the present disclosure is shown schematically.

[0088] like Figure 4 As shown, a deep convolutional neural network model input layer, a first convolutional layer, a first pooling layer, a normalization layer, a first activation layer, a second convolutional layer, a second pooling layer, a second activation layer, a first fully connected layer, a second activation layer, a second fully connected layer, a normalized index layer and an output layer according to an embodiment of the present disclosure can automatically learn and extract deep-level features of input data by stacking multiple convolutional layers, pooling layers, activation layers and fully connected layers. The model achieves efficient feature extraction, parameter sharing, training stability and powerful classification capabilities through a stacking design of convolution-pooling-activation-fully connected, and has broad application prospects in the fields of images, text, speech and so on. Its hierarchical representation capability enables the model to capture the complex distribution of data and generate high-quality prediction results while maintaining the stability and generalization of training.

[0089] Use pre-trained deep convolutional neural network model to test the set Image data is classified and recognized to obtain test set labels, which are used to judge the accuracy of the pre-trained deep convolutional neural network model.

[0090] After obtaining the recognition model based on the above method, combined with Figure 5 A laser induced breakdown spectroscopy identification method according to an embodiment of the present disclosure is introduced.

[0091] Figure 5The flowchart of a laser induced breakdown spectroscopy identification method according to an embodiment of the present disclosure is schematically shown.

[0092] like Figure 5 As shown, a laser induced breakdown spectroscopy identification method according to an embodiment of the present disclosure includes steps S310 to S330.

[0093] In step S310, a laser induced breakdown spectrum to be identified is obtained.

[0094] In step S320 , the laser induced breakdown spectrum to be identified is converted into a laser induced breakdown spectrum image.

[0095] LIBS spectra are one-dimensional wavelength-intensity data. When converted into an image (such as a two-dimensional matrix or pseudo-color image), they can intuitively display the spatial distribution of different elemental features. Features such as texture and edges in the image can directly correspond to physical parameters such as peak position and peak width in the spectrum, facilitating algorithm recognition. LIBS images not only contain spectral features but also implicitly contain spatial distribution information on the sample surface. For example, in material defect detection, images can simultaneously reflect elemental composition (spectrum) and defect location (space). Compared to traditional spectral analysis, they can provide a more comprehensive sample characterization. Image data can be fused with other modalities (such as X-ray images and SEM images) to extract joint features through deep learning models (such as convolutional neural networks) to further improve classification or regression performance. Optionally, the method for converting the laser-induced breakdown spectrum to be identified into a laser-induced breakdown spectrum image in step S320 is the same as the conversion method in step S120 and is not further described here.

[0096] In step S330 , the laser induced breakdown spectrum image is identified using the recognition model to obtain a sample label of the laser induced breakdown spectrum to be identified.

[0097] Example 1:

[0098] Figure 6 A schematic diagram of an original spectral line of a LIBS spectrum according to an embodiment of the present disclosure is shown schematically.

[0099] like Figure 6 As shown in FIG, according to an embodiment of the present disclosure, a LIBS spectrum original spectrum has a wavelength range of 200-1000 nm. For the original spectrum including 12567 bands of LIBS spectrum, the LIBS spectrum is first preprocessed to exclude outliers and normalize the spectrum, including using median filtering to exclude outliers. The calculation formula is shown in formula (3), where the filter window size is is 11; the mean variance is used for normalization, and the calculation formula is shown in formula (4).

[0100] After spectral normalization, the spectrum is converted into an image. The conversion rules are as follows:

[0101] Step 1: Determine the image size based on the number of bands in the LIBS spectrum. The image size is , which is equivalent to the aforementioned step S121.

[0102] Step 2: Extract the spectral peak of the LIBS spectrum and obtain the wavelength and intensity information of the peak position, which is equivalent to the aforementioned step S122.

[0103] Step 3: Set the wavelengths of multiple peak positions of the spectrum Convert into image position, =200, =1000 is substituted into formula (1), which is equivalent to the aforementioned step S123.

[0104] Step 4: Using formula (2), the intensity of the peak position Convert to graphics , obtaining the LIBS image of each spectrum, which is equivalent to the aforementioned steps S124 and S125.

[0105] Figure 7 The figure schematically shows an image diagram of LIBS spectrum conversion according to an embodiment of the present disclosure.

[0106] like Figure 7 As shown in the figure, the image size is 112×112, where the spectral peak feature of LIBS is converted into the pixel value of the image center, and the wavelength is converted into the position information of the image, which is more conducive to two-dimensional convolution for feature extraction and more accurate image generation.

[0107] For a LIBS image group with a sample size of 900, the LIBS spectra are randomly divided into a training set (sample size is 630) and test set (The sample size is 270).

[0108] Deep generative adversarial neural network model designed based on training set image data , where the model structure of G is as follows Figure 2 Its input is a random signal and a sample label, and it outputs an image after passing through a deep generative adversarial neural network.

[0109] After obtaining the deep generative adversarial neural network model, we performed sample amplification and sample number alignment. For 630 training samples, there are 6 sample categories, and the number of images in each category is [70, 105, 140, 35, 175, 105].

[0110] First, align the number of samples and take the largest number of samples As a benchmark, this case has 175 samples, using the trained deep generative adversarial neural network model Generate [105, 70, 35, 140, 0, 70] image samples respectively. The number of samples in the final training set is 1050.

[0111] After aligning the number of samples, the samples were amplified and the amplification factor was , this case is , so the final number of images in the training set is 2100.

[0112] After the sample generation is completed, taking the first type of sample as an example, the number of real samples is 70 and the number of generated samples is 280. For this, the secondary label is marked and the real sample is set The training weight of the real sample is In this case =0.35, for generating samples The weight is In this case =0.65. And it is used in the training of recognition model.

[0113] The training set model is trained using a deep convolutional neural network. The structure of the deep convolutional neural network model is as follows: Figure 4 The loss function is shown in formula (5).

[0114] Use pre-trained deep convolutional neural network model to test the set The image data is classified and recognized to obtain the test set labels. The accuracy of 100 repeated tests is shown in Table 1:

[0115] Table 1 Comparison of the test set between the method according to the embodiment of the present disclosure and the direct recognition method Comparison of accuracy of classification and recognition of image data

[0116] As can be seen from Table 1, the method according to the embodiment of the present disclosure is effective for the test set The accuracy of 5 out of the 6 categories of data in the image prediction results is higher than that of the direct recognition method.

[0117] Figure 8 A training sample augmentation device according to an embodiment of the present invention is schematically shown.

[0118] like Figure 8 As shown, the training sample amplification device 800 of this embodiment includes an acquisition module 810 , an image conversion module 820 , a training module 830 and an amplification module 840 .

[0119] The acquisition module 810 is used to acquire real training samples, including laser-induced breakdown spectra and corresponding sample labels, where the sample labels are used to indicate the material classification corresponding to the laser-induced breakdown spectra.

[0120] The image conversion module 820 is used to convert the laser-induced breakdown spectrum into a laser-induced breakdown spectrum image, obtain a sample image group, and divide the sample image group into a training set and a test set.

[0121] The training module 830 is used to take the random signal and the sample labels in the training set as input and the laser-induced breakdown spectrum image in the training set as output to train the deep generative adversarial neural network model to obtain a trained deep generative adversarial neural network model.

[0122] The amplification module 840 is used to amplify the training set using the trained deep generative adversarial neural network model to obtain an amplified training set.

[0123] According to embodiments of the present invention, any multiple modules among the acquisition module 810, image conversion module 820, training module 830, and augmentation module 840 may be combined into a single module, or any one of these modules may be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules may be combined with at least part of the functionality of other modules and implemented in a single module. According to embodiments of the present invention, at least one of the acquisition module 810, image conversion module 820, training module 830, and augmentation module 840 may be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or may be implemented in hardware or firmware through any other reasonable means of circuit integration or packaging, or implemented in any one of software, hardware, and firmware, or any suitable combination of these. Alternatively, at least one of the acquisition module 810, image conversion module 820, training module 830, and augmentation module 840 may be at least partially implemented as a computer program module that, when executed, performs the corresponding functionality.

[0124] Figure 9 The block diagram schematically shows an electronic device for the training sample augmentation method according to an embodiment of the present invention.

[0125] like Figure 9As shown, an electronic device 900 according to an embodiment of the present invention includes a processor 901, which can perform various appropriate actions and processes based on programs stored in a read-only memory (ROM) 902 or programs loaded from a storage unit 908 into a random access memory (RAM) 903. The processor 901 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or related chipsets and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 901 may also include onboard memory for caching purposes. The processor 901 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0126] The RAM 903 stores various programs and data required for the operation of the electronic device 900. The processor 901, ROM 902, and RAM 903 are connected to each other via a bus 904. The processor 901 executes the programs in the ROM 902 and / or RAM 903 to perform various operations according to the method flow of the embodiment of the present invention. It should be noted that the programs may also be stored in one or more memories other than the ROM 902 and RAM 903. The processor 901 may also execute the programs stored in one or more memories to perform various operations according to the method flow of the embodiment of the present invention.

[0127] According to an embodiment of the present invention, electronic device 900 may further include an input / output (I / O) interface 905, which is also connected to bus 904. Electronic device 900 may also include one or more of the following components connected to I / O interface 905: an input section 906 including a keyboard, mouse, etc.; an output section 907 including devices such as a cathode ray tube (CRT), liquid crystal display (LCD), and speakers; a storage section 908 including a hard disk; and a communication section 909 including a network interface card such as a LAN card or modem. Communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to I / O interface 905 as needed. Removable media 911, such as a magnetic disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed in drive 910 as needed, so that computer programs read from the removable media can be installed into storage section 908 as needed.

[0128] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments, or may exist independently and not incorporated into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present invention.

[0129] According to an embodiment of the present invention, a computer-readable storage medium may be a non-volatile computer-readable storage medium, and may include, for example, but not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to an embodiment of the present invention, a computer-readable storage medium may include the ROM 902 and / or RAM 903 described above, and / or one or more memories other than ROM 902 and RAM 903.

[0130] It will be understood by those skilled in the art that the features described in the various embodiments of the present invention may be combined and / or coupled in various ways, even if such combinations or couplings are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention may be combined and / or coupled in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or couplings fall within the scope of the present invention.

Claims

1. A training sample amplification method, characterized in that: include: Acquire real training samples, including laser-induced breakdown spectra and corresponding sample labels, where the sample labels are used to indicate the material classification corresponding to the laser-induced breakdown spectra; Converting the laser-induced breakdown spectrum into a laser-induced breakdown spectrum image to obtain a sample image group, and dividing the sample image group into a training set and a test set; Using a random signal and a sample label in the training set as input and a laser-induced breakdown spectroscopy image in the training set as output, training a deep generative adversarial neural network model to obtain a trained deep generative adversarial neural network model; The training set is amplified using the trained deep generative adversarial neural network model to obtain an amplified training set.

2. The training sample amplification method according to claim 1, characterized in that: The converting the laser induced breakdown spectrum into a laser induced breakdown spectrum image comprises: determining the size of the image to be generated according to the number of bands of the preprocessed laser-induced breakdown spectrum; Extracting the peaks corresponding to all wavelengths of the preprocessed laser-induced breakdown spectrum to obtain the wavelengths and intensities of the peak positions; Calculating the position coordinates of the spectral peaks on the image to be generated based on the wavelengths of all the spectral peak positions; Calculating a pixel value of the spectral peak on the image to be generated based on the intensity of the spectral peak position; A laser-induced breakdown spectroscopy image is generated based on the position coordinates and pixel values ​​of each spectrum peak on the image to be generated.

3. The training sample amplification method according to claim 1, characterized in that: The method of amplifying the training set by using the trained deep generative adversarial neural network model includes: Obtaining sample labels in the training set and the number of laser-induced breakdown spectroscopy images corresponding to each sample label; Using the trained deep generative adversarial neural network model, the number of laser-induced breakdown spectroscopy images corresponding to each sample label is expanded to a preset number.

4. The training sample amplification method according to claim 3, characterized in that: The step of expanding the number of laser-induced breakdown spectroscopy images corresponding to each sample label to a preset number includes: Obtaining the maximum value of the number of laser-induced breakdown spectroscopy images corresponding to the sample labels in the training set; Using the trained deep generative adversarial neural network model, the number of the laser-induced breakdown spectroscopy images corresponding to each of the sample labels is amplified to the maximum value; The laser-induced breakdown spectroscopy image, whose quantity has been amplified to the maximum value, is amplified multiple times using a pre-trained deep generative adversarial neural network model.

5. The training sample amplification method according to claim 1, characterized in that: The convolution layer of the deep generative adversarial neural network model is a two-dimensional convolution kernel, which is generated using the Gauss-Lorentz formula.

6. The training sample amplification method according to claim 1, characterized in that: Before converting the laser induced breakdown spectrum into a laser induced breakdown spectrum image, the method further includes: The laser-induced breakdown spectrum is subjected to outlier elimination and normalization processing to obtain a preprocessed laser-induced breakdown spectrum.

7. The training sample augmentation method according to claim 6, characterized in that: The method for eliminating abnormal values ​​includes median filtering; The normalization processing method includes a mean-variance normalization processing method.

8. A recognition model generation method, characterized in that: include: Set training weights for the generated graphics group and the real image group in the expanded training set respectively; Constructing a loss function based on the training weights; The deep convolutional neural network model is trained using the loss function so that the value of the loss function meets the preset requirements and the recognition model is obtained.

9. A laser induced breakdown spectroscopy identification method, characterized in that: include: obtaining a laser-induced breakdown spectrum to be identified; converting the laser induced breakdown spectrum to be identified into a laser induced breakdown spectrum image; The laser induced breakdown spectrum image is identified by using a recognition model to obtain a sample label of the laser induced breakdown spectrum to be identified.

10. A training sample amplification device, characterized in that: include: An acquisition module is used to acquire real training samples, including laser-induced breakdown spectra and corresponding sample labels, wherein the sample labels are used to indicate the material classification corresponding to the laser-induced breakdown spectra; an image conversion module, configured to convert the laser-induced breakdown spectrum into a laser-induced breakdown spectrum image to obtain a sample image group, and divide the sample image group into a training set and a test set; A training module is configured to take a random signal and a sample label in the training set as input, and a laser-induced breakdown spectroscopy image in the training set as output, to train a deep generative adversarial neural network model, thereby obtaining a trained deep generative adversarial neural network model; An amplification module is used to amplify the training set using the trained deep generative adversarial neural network model to obtain an amplified training set.

Images