Spectral data enhancement method for regression prediction of continuous variable-oriented generative adversarial network

Abstract

The application discloses a continuous variable regression prediction-oriented generative adversarial network spectrum data enhancement method, carries out spectrum data acquisition and pretreatment, constructs pre-training networks T1 and T2 of a continuous conditional generative adversarial network and the continuous conditional generative adversarial network, constructs a discriminant network model D by using generated spectrum and experimental true spectrum, trains a generated network model G by sharing the training parameters of the D network, generates new synthetic spectrum according to the new generated network G obtained through training, makes the synthetic spectrum data gradually approach the true experimental spectrum data, judges the similarity of the synthetic spectrum and the true experimental spectrum, inputs preset label values into a label input network to generate corresponding synthetic spectrum when the similarity meets the requirements, and forms a training set of a machine learning prediction model by the synthetic spectrum data and corresponding preset labels and the true experimental spectrum data, so that the machine learning prediction model is trained and tested. The application improves the precision of the machine learning prediction model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of spectral detection, and relates to a generative adversarial network spectral data enhancement method for continuous variable regression prediction. BACKGROUND

[0002] Spectral technology is widely used in many fields such as chemistry, physics, biology, environmental science and material science. Some common spectral technologies include ultraviolet-visible spectroscopy, infrared spectroscopy, Raman spectroscopy, mass spectrometry and nuclear magnetic resonance spectroscopy. Spectral technology generates a large amount of data, and traditional methods are difficult to comprehensively analyze and process these data. In recent years, big data and deep learning methods have been increasingly applied to the analysis of spectral data, making the analysis and processing of data more accurate, fast and comprehensive.

[0003] It is well known that machine learning algorithms are data-driven and require a large amount of sufficient spectral data to obtain stable and accurate predictions. However, in the application of spectral technology, it may be costly or difficult to obtain high-quality spectral data. The reasons include the following aspects: (1) collecting spectral data from various sources or fields may be time-consuming and costly; (2) certain spectral data may be restricted by privacy regulations, making it difficult to obtain; (3) certain spectral detection tasks may require professional knowledge or skills to accurately label or annotate data; (4) ensuring the high quality of certain spectral data is challenging, especially if the spectral data may contain errors, biases or noise, which may affect the performance of the prediction model. Therefore, it is a proper method to use a generative model to generate a certain amount of synthetic spectral data for data augmentation based on a small amount of real spectral data.

[0004] Currently, generative adversarial networks (GAN, Generative Adversarial Networks) are used to generate data in the field of spectral technology for machine learning classification tasks (Document 1: A spectral database expansion method based on generative adversarial networks, CN 109508647A). Regression tasks with continuous variables have wide applications in the field of spectral detection, such as chemistry, biology and material science. At this time, spectral data are used to predict continuous output variables. However, in these continuous variable regression predictions based on spectral detection, the general GAN network is used to generate synthetic spectral data with pre-set corresponding labels, which faces problems such as inability to pre-set labels, low generation quality and efficiency, and is not suitable for continuous variable regression prediction. SUMMARY

[0005] In order to solve the problem that it is difficult to obtain enough spectral data for existing continuous variable regression tasks, the application provides a generative adversarial network spectral data enhancement method for continuous variable regression prediction, which improves the accuracy of the machine learning prediction model; the method is widely applicable to the generation of near-infrared spectrum, Raman spectrum, laser-induced breakdown spectrum, fluorescence spectrum and terahertz synthetic spectrum data, and can generate synthetic spectrum data with preset labels, which is high in quality and efficient, and is completely applicable to continuous variable regression prediction.

[0006] The technical scheme of the application is as follows:

[0007] A generative adversarial network spectral data enhancement method for continuous variable regression prediction, comprising the following steps:

[0008] Step 1. Perform spectral data acquisition, the process is:

[0009] Build a spectral acquisition experimental device to perform spectral data acquisition of the sample to be measured;

[0010] Step 2. Spectral data set preprocessing, the process is:

[0011] Divide the one-dimensional real spectrum data into several parts in the wavelength dimension, and then convert it into a two-dimensional spectrum image to adapt to the input of the constructed generative adversarial network, and normalize all real labels;

[0012] Step 3. Construct a pre-training network (T1+T2) of a continuous conditional generative adversarial network,

[0013] Step 4. Construct a continuous conditional generative adversarial network, including a generative network G, a discriminative network D and a label embedding network T3;

[0014] Step 5. Construct a discriminative network model D using generated spectra and experimental real spectra;

[0015] Step 6. Train the generative network model G by sharing the training parameters of the D network;

[0016] Step 7. Generate new synthetic spectra according to the new generative network G obtained by training;

[0017] Step 8. Repeat steps 5-7 to gradually approach the real experimental spectral data for the synthetic spectral data;

[0018] Step 9. Perform similarity judgment on the synthetic spectrum and the real experimental spectrum;

[0019] Step 10. After the similarity of step 9 meets the requirements, input the preset label value into the label input network, and generate the corresponding synthetic spectrum through the trained continuous generative adversarial network;

[0020] Step 11. The synthetic spectral data and the corresponding preset labels generated in step 10 above and the real experimental spectral data constitute a training set for a machine learning prediction model, and the machine learning prediction model is trained and tested.

[0021] Further, the process of step 3 is:

[0022] The pre-trained network is a fully connected neural network that realizes the output of the corresponding prediction label The pre-trained network can be divided into two sub-networks, (T1 and T2), sub-network 1 (T1) learns to realize the mapping from the input spectral image sample (x) to the output high-dimensional feature (h) through the training process, and sub-network 2 (T2) learns to realize the mapping from the input high-dimensional feature (h) to the output prediction label Through the use of the trained sub-network 2 (T2), the label embedding network (T3) can be guided to map the real label (y) to the high-dimensional feature (h). The mean square error (MSE) of the prediction label and the real label is used as the loss function to train the pre-trained network.

[0023] Further, the process of step 4 is:

[0024] The label embedding network T3 is constructed using a fully connected neural network model, the input is the real label or the target label, and the output is the high-dimensional feature h described in step 3. The high-dimensional feature h will act as a continuous condition on the generation network G and the discriminant network D, and then generate and train; the training of the label embedding network T3 is completed with the trained sub-network T2; select a real label (y), add a noise that meets the normal distribution, and use it as the input of the label embedding network T3. The output of the label embedding network T3 is the high-dimensional feature (h). Then input the output high-dimensional feature (h) into the trained sub-network T2, and after mapping by the sub-network T2, the output of the sub-network T2 is the prediction label of the high-dimensional feature (h); compare the prediction label with the noisy real label input into the label embedding network T3, and calculate the MSE of the two as the loss, aiming to minimize the optimization training label embedding network (T3); construct the generation network G, input a random noise that meets the normal distribution, and generate a synthetic spectral image sample; construct the discriminant network D, and use the synthetic spectral image generated by the generation network G and the real experimental spectral image as the two inputs of the discriminant network D for discriminant analysis, thereby guiding the generation network G to generate high-quality data.

[0025] Further, the process of step 5 is:

[0026] The discriminant network model D adopts a binary classification deep convolutional neural network model, the discriminant network D accepts the real experimental spectrum image and the synthetic spectrum image generated by the generation network G as two inputs, and the high-dimensional feature (h) output by the label embedding network T3 is used as the condition input, and the output is the probability of discriminating the synthetic spectrum image as real, and the loss function of the discriminant network D during training is defined as follows:

[0027]

[0028] Wherein, N r and N g are the number of real spectrum image samples and synthetic spectrum image samples respectively, ε r and ε g are two noise variables conforming to the standard normal distribution, and are the ith real spectrum image sample and synthetic spectrum image sample respectively, and are the corresponding labels, and The calculation method of the hyperparameter σ is Wherein, is the standard deviation of the normalized label in the training set, and the calculation method of the hyperparameter κ is Wherein is the smallest normalized different real label of the lth, is the number of different labels in the training set, and the target labels of the real samples and the generated samples satisfy and This means that the labels y provided to the generation network and the discriminant network have a certain deviation from and The formula and represent the weight assigned to the real or synthetic spectrum sample within a small range of the target label, and the closer the distance, the greater the weight assigned.

[0029] The process of step 6 is:

[0030] A deep convolutional neural network is used to construct the generation network G, which accepts random noise and high-dimensional features mapped by T3 as input, and outputs corresponding synthetic spectrum images, and the loss function used by the generator G during training is as follows:

[0031]

[0032] Wherein, z i is a random noise conforming to the standard normal distribution.

[0033] The process of step 9 is:

[0034] The structural similarity method is used to judge the similarity of the synthetic spectrum and the real experiment spectrum.

[0035] Further, in the step 11, the DNN or CNN deep regression network model is used for continuous variable regression prediction of the spectrum data.

[0036] The present application has the advantages that: the machine learning model is widely used in the field of spectrum detection for continuous variable regression prediction. The machine learning model usually needs a large amount of effective data driving, but it is difficult to obtain a large amount of effective data in practice. Data enhancement technology is a relatively effective solution. However, the previous spectrum data generation technology based on ordinary GAN network cannot preset the generated label, and has the problems of low generation quality and efficiency, which is difficult to apply to continuous variable regression prediction in the field of spectrum detection. The present application is aimed at continuous variable regression prediction in the field of spectrum detection, based on a continuous conditional generative adversarial network, special label embedding technology and improved loss function, and combined with the characteristics of spectrum data, the control granularity of generated spectrum data is improved, and the preset label synthetic spectrum data sample can be generated efficiently and with high quality, which is suitable for continuous variable regression prediction. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 It is a method flowchart of the present application;

[0038] Figure 2 It is a block diagram of a full spectrum data acquisition system based on microbubble cavity temperature sensing measurement;

[0039] Figure 3 It is a continuous conditional generative adversarial network structure;

[0040] Figure 4 It is a pre-training network and label embedding network structure;

[0041] Figure 5 It is a discriminator network sub-block structure;

[0042] Figure 6 It is a generated network sub-block structure;

[0043] Figure 7 It is a comparison of the similarity between the real experiment spectrum image sample and the synthetic spectrum image sample;

[0044] Figure 8 It is a box plot of the mean square error (MSE) of the DNN model prediction;

[0045] Figure 9To enhance the comparison between each predicted label value and the true predicted value before;

[0046] Figure 10 To enhance the comparison between each predicted label value and the true predicted value after. DETAILED DESCRIPTION

[0047] The present application will be further described below with reference to the accompanying drawings and examples.

[0048] Reference Figures 1-10 A generative adversarial network spectrum data enhancement method for continuous variable regression prediction is provided, and a microbubble cavity temperature sensing measurement is taken as an example to illustrate the application of the spectrum data enhancement method based on the continuous condition generative adversarial network in the regression scene. The measurement is based on a microbubble cavity temperature sensor to collect multimode spectrum data at a specific temperature, and a mapping relationship between the temperature and the multimode spectrum is established through a machine learning model to perform regression prediction on the continuous variable temperature. Due to the limited precision of the physical device, the number of spectrum data obtained is limited, and therefore the temperature sensing data is enhanced by using the method provided in the present application, and the method comprises the following steps:

[0049] Step 1, collecting spectrum

[0050] An experimental device based on microbubble cavity (Microbubble resonator, MBR) temperature sensing measurement is built as shown in Figure 2 The experimental device based on microbubble cavity (Microbubble resonator, MBR) temperature sensing measurement is built as shown in FIG. 1, wherein 1 is an arbitrary signal generator, 2 is a tunable laser, 3 is a fiber polarization controller, 4 is a microbubble resonator, 5 is a semiconductor refrigeration piece thermistor, 6 is a semiconductor refrigeration piece controller, 7 is a data acquisition card, 8 is a photodetector, 9 is an oscilloscope, and 10 is an upper computer.

[0051] A 1550nm tunable laser was used to excite the resonant mode of the MBR via a tapered optical fiber. The laser wavelength was precisely scanned using a triangular wave generated by an arbitrary function generator at a scanning speed of 15MHz / s. The MBR was placed inside a copper box with dimensions of 3cm × 1.5cm × 1cm, and the MBR temperature was indirectly controlled by directly controlling the temperature of the copper box. This temperature measurement experimental setup employed a thermoelectric cooler (TEC) and a temperature controller to achieve high-precision temperature control. During temperature control, the heat-conducting side of the TEC was placed in close contact with the surface of the copper box, and a thermistor was used to detect the temperature of the copper box. The temperature controller obtained the temperature of the measured object in real time through the thermistor, compared it with the actual set temperature, and then output appropriate power to the cooler to bring the copper box temperature to the set temperature. A fiber optic polarization controller was used to adjust the polarization state of the input light to ensure optimal coupling efficiency. The light was coupled into the MBR via a tapered optical fiber. The transmission spectrum was recorded in real time by a low-noise photodetector and then split into two paths, connected to an oscilloscope and a data acquisition card, respectively. The former was used for real-time monitoring of the sensor spectrum, and the latter for automatic acquisition of the transmission spectrum. When the temperature changes, the computer program controls the laser, temperature control equipment, and data acquisition card to work together to automatically collect multiple sets of transmission spectra at different temperatures. The temperature change step is controlled at 0.005℃, and a total of 40 sets of spectral data are collected. The temperature corresponding to each spectral sample is regarded as its label, ranging from 35℃ to 35.195℃.

[0052] Step 2: Preprocessing of Spectral Dataset

[0053] One-dimensional spectral data is converted into two-dimensional spectral image data suitable for input to deep learning models. Specifically, the 149,999 sampling points of each spectral data point are sliced ​​into 61 parts, and then these 61 parts are stitched together to form a 61×2459 two-dimensional image. Finally, this two-dimensional spectral image is resized to 3×64×64 and normalized to improve the robustness and generalization performance of the CcGAN model. Simultaneously, the label (temperature) is normalized.

[0054] Step 3: Construct the pre-trained network T1+T2 of the continuous conditional generative adversarial network.

[0055] like Figure 3 As shown, the continuous conditional generative adversarial network structure mainly consists of four parts: generator network, discriminator network, pre-trained network, and label embedding network.

[0056] like Figure 4 As shown in the left figure, the pre-trained network is a fully connected neural network that outputs corresponding predicted labels based on the input experimental spectral image samples (x). The pre-training network can be divided into two sub-networks T1 and T2. Sub-network T1 learns to realize the mapping from the input spectral image sample x to the output high-dimensional feature h through the training process. Sub-network T2 learns to realize the mapping from the input high-dimensional feature h to the output predicted label Through the use of the trained sub-network T2, the label embedding network T3 can be guided to map the real label y to the high-dimensional feature h. The mean squared error (MSE) of the predicted label and the real label is used as the loss function to train the pre-training network. The number of neurons in the input layer is set to 12288. The number of neurons in the two hidden layers is set to 128, which is equal to the dimension of the high-dimensional feature (h). After the pre-training network is trained, sub-network T2 can map the high-dimensional feature h to the output predicted label Based on the real experimental spectral image dataset, the pre-training network is trained using the Adam optimizer with a learning rate of 0.001. The batch_size is set to 40, and after 20000 epochs of training, the loss of the pre-training network decreases significantly, which represents that the pre-training network is trained.

[0057] Step 4, construct a continuous conditional generative adversarial network (including a generative network G, a discriminative network D, and a label embedding network T3).

[0058] As Figure 4 As shown in the right graph, a fully connected neural network model is used to construct the label embedding network T3, the input is the real label or the target label (used in the training of G and D), and the output is the high-dimensional feature (h) described in step 3, which will be used as a continuous condition for the generative network G and the discriminative network D, and then for generation and training. The training of the label embedding network T3 is completed with the help of the trained sub-network T2. Select a real label y, add a noise conforming to the normal distribution as the input of the label embedding network T3, and the output of the label embedding network T3 is the high-dimensional feature h. Then input the output high-dimensional feature h into the trained sub-network T2, and after mapping through the sub-network T2, the output of the sub-network T2 is the predicted label of the high-dimensional feature h. The predicted label The input of the label embedding network T3 is compared with the real label with noise, and the MSE of the two is calculated as the loss to optimize the training label embedding network T3. The number of nodes of the input layer of T3 is 1, which matches the dimension of the label (temperature). Both the hidden layer and the output layer consist of 128 nodes, which match the dimension of the high-dimensional feature h. T3 uses stochastic gradient descent with a learning rate of 0.0001 as the optimizer, ReLU as the activation function, and Groupnorm to stabilize the training of the model. The batch_size is set to 1, and after 5000 epochs of training, the training loss no longer decreases, and the training process stops.

[0059] Step 5, construct a discriminant network model D using the generated spectrum and the experimental real spectrum.

[0060] The discriminator network consists of 2 linear layers and 5 discriminant network sub-blocks, as shown in Figure 5 The sub-blocks have a unique residual structure, and through the skip connection in them, the problem of gradient disappearance can be effectively alleviated. Each sub-block contains a spectral normalization layer (SN) that can stabilize the training of the model; after the convolutional layer, an average pooling layer is connected to gradually reduce the spatial dimensions of the feature maps at the end of the main path and the skip connection. The discriminant network D accepts the real experimental spectrum image and the synthetic spectrum image generated by the generator network G as two inputs, the former is generated by the generator network, and the latter is obtained by converting the real spectrum data into a two-dimensional spectrum image. At the same time, the high-dimensional feature h output by the label embedding network T3 is input as a condition, and the output is the probability that the synthetic spectrum image is real. The loss function of the discriminant network D during training is defined as follows:

[0061]

[0062] where N r and N g are the number of real spectrum image samples and synthetic spectrum image samples, respectively, ε r and ε g are two noise variables conforming to the standard normal distribution, and are the i-th real spectrum image sample and synthetic spectrum image sample, and are and corresponding labels. The calculation method of the hyperparameter σ is where is the standard deviation of the normalized labels in the training set. The calculation method of the hyperparameter κ is where is the l-th smallest normalized different real label, is the number of different labels in the training set. The target labels of real samples and generated samples satisfy and This means that the labels y provided to the generator and discriminator are biased from the true labels and Equation and represent the weights assigned to real or synthetic spectral samples within a small range of the target label, and the closer the distance, the greater the weight assigned. The training of D uses the Adam optimizer to optimize the learning rate of 0.0001, the batch size is 40, and the iteration number is 15000.

[0063] Step 6, share the training parameters of the D network, and train the generator model G.

[0064] The generator consists of 1 linear layer and 4 generator sub-blocks, as shown in Figure 6 The sub-blocks of the generator also introduce a residual structure, as do the discriminator sub-blocks. Each sub-block contains some identical operations, such as conditional batch normalization (cond. BN), activation function ReLU, and convolutional layer, followed by a upsampling layer to enable the generator to learn complex spectral representations. The input to the generator is a 256-dimensional random noise conforming to the normal distribution and the high-dimensional feature h mapped by the T3 network, which will be treated as a continuous condition and input to the network through cond. BN, with an output size of 3x64x64 synthetic spectral images. The loss function used when training the generator G is as follows:

[0065]

[0066] where z i is a random noise conforming to the standard normal distribution. The training parameters of the generator and the optimizer used are the same as those of the discriminator.

[0067] Step 7, generate new synthetic spectra based on the new generator G obtained from training.

[0068] Step 8, repeat steps 5-7 to gradually approach the real experimental spectral data.

[0069] Step 9, perform similarity judgment on the synthetic spectra and real experimental spectra.

[0070] Set several real labels as the label input CcGAN of the synthetic spectrum, and generate several corresponding synthetic spectral images. The results are shown in Figure 7As shown, it can be seen that the real spectral images of the same label are very similar to the synthetic spectral images, indicating that the training of CcGAN is effective. In addition to direct observation, an image similarity index (SSIM) can be introduced to quantitatively analyze the similarity of real spectra and synthetic spectra.

[0071] Step 10, after the similarity of the above step 9 meets the requirements, the preset label value is input into the label input network, and the corresponding synthetic spectrum is generated by the trained continuous generative adversarial network.

[0072] The value of the input label is set to m / (M-1), where M is the total number of synthetic spectra, and m is an integer (m=0, 1,..., M-1). Taking the generation of 2000 groups of synthetic spectral images as an example, if 2000 synthetic spectra are needed to be generated, 2000 continuous labels are input into the trained label embedding network T3, and M is assigned a value of 2000, m=0, 1,..., 1999. In addition, 2000 groups of 256-dimensional random noise z are input into the generator. In this way, 2000 groups of synthetic spectral images can be obtained.

[0073] Step 11, the synthetic spectral data and the corresponding preset label generated in the above step 10 are used to form a training set of a DNN (Deep neural network, DNN) regression model, and the model is trained and tested.

[0074] The DNN regression model adopts the following structure: the input layer is composed of 12288 nodes, which matches the length of the flattened spectral image. There are two hidden layers, each including 128 nodes; the output layer is a single node representing the predicted result (temperature). After each hidden layer, BN is used, followed by a softmax activation function layer. The training process is started by using the Adam optimizer, with a learning rate of 0.001. The MSE of the real label value and the predicted label value is used as an index to evaluate the prediction ability of the model.

[0075] Figure 8 The box plot of the MSE predicted by the DNN model is shown as the number of synthetic spectra increases, and each box contains 40 MSEs, which are obtained by averaging 20 measurements of a single test sample. According to the experimental results, when 4000 synthetic samples are added, the MSE predicted by the DNN model is the smallest. Without data augmentation, the height of the box is 1.796x10 -5 , representing the numerical range. The horizontal line in the middle of the box represents the median, which is equal to 1.46x10 -5, which represents the midpoint of these values. Compared with the box without the CcGAN enhancement stage, the range and median of the box with 4000 enhanced samples are reduced by 59.35% and 81.2%, respectively, which indicates that the prediction accuracy of the model has been significantly improved. With the increase of the number of enhanced samples, the prediction performance of the DNN will gradually improve until it reaches a threshold that has a significant negative impact on performance (more than 4000 samples). This is because there is an acceptable deviation between the synthetic spectrum generated by CcGAN and the real spectrum. When the number of synthetic spectra is too large, this deviation will exceed the acceptable threshold, thereby affecting the improvement of the prediction performance of the model.

[0076] Figure 9 The comparison between the predicted value and the actual value at each temperature and the MSE without data enhancement are shown. It can be seen that the mean value predicted by the DNN model is 1.69E-05. Figure 10 The comparison between the predicted value and the actual value at each temperature and the MSE after adding 4000 synthetic spectral images are shown. The mean value predicted by the DNN model is 4.05E-06, which also proves that CcGAN has a good enhancement effect on spectral data.

[0077] The above describes the spectral data enhancement method of the regression prediction generation adversarial network for continuous variable in detail. The above examples are only used to help understand the method and core idea of the present application, but not to limit it. Any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application are equivalent replacement methods and are included in the protection scope of the present application.

Claims

1. A generative adversarial network (GAN) method for spectral data augmentation in continuous variable regression prediction, characterized in that, The method includes the following steps: Step 1: Set up the spectral acquisition experimental setup and collect spectral data; Step 2: Preprocessing the spectral dataset; Step 3: Construct the pre-trained network T1 + T2 of the continuous conditional generative adversarial network; Step 4: Construct a continuous conditional generative adversarial network, including a generator network G, a discriminator network D, and a label embedding network T3; the process is as follows: A fully connected neural network model is used to construct a label embedding network T3. The input is either the real label or the target label, and the output is the high-dimensional feature h described in step 3. This high-dimensional feature h will be used as a continuous condition to act on the generator network G and the discriminator network D, thereby performing generation and training. The training of the label embedding network T3 is completed with the help of the trained sub-network T2. A real label is selected. y After adding normally distributed noise, it is used as the input to the label embedding network T3, and the output of the label embedding network T3 is a high-dimensional feature h. Then, the output high-dimensional feature h is input into the trained sub-network T2. After being mapped by the sub-network T2, the output of the sub-network T2 is the predicted label of the high-dimensional feature h. The prediction label The input of the label embedding network T3 is compared with the noisy real label, and the MSE of the two is calculated as the loss. The training of the label embedding network T3 is optimized with the goal of minimizing the loss. A generative network G is constructed, which takes a normally distributed random noise as input and generates synthetic spectral image samples. A discriminative network D is constructed, which takes the synthetic spectral image generated by the generative network G and the real experimental spectral image as two inputs to the discriminative network D for discriminative analysis, thereby guiding the generative network G to generate high-quality data. Step 5: Construct a discriminant network model D using the generated spectrum and the experimental real spectrum; Step 6: Share the training parameters of the D network and train the generative network model G; Step 7: Generate new synthetic spectra based on the newly generated network G obtained from training; Step 8: Repeat steps 5-7 to gradually bring the synthesized spectral data closer to the actual experimental spectral data. Step 9: Determine the similarity between the synthesized spectrum and the actual experimental spectrum; Step 10: After the similarity in step 9 above meets the requirements, the preset label value is input into the label input network, and the trained continuous generative adversarial network generates the corresponding synthetic spectrum. Step 11: Combine the synthetic spectral data and corresponding preset labels generated in Step 10 with the real experimental spectral data to form the training set for the machine learning prediction model, and train and test the machine learning prediction model.

2. The generative adversarial network spectral data augmentation method for continuous variable regression prediction according to claim 1, characterized in that: The process of step 3 is as follows: The pre-trained network is a fully connected neural network that implements training based on input experimental spectral image samples. x Output the corresponding predicted labels The pre-trained network can be divided into two sub-networks, T1 and T2. Sub-network T1 learns from the input spectral image samples through the training process. x The subnetwork T2 learns the mapping from the input high-dimensional feature h to the output predicted label through the training process. The mapping, by using the trained subnetwork T2, can guide the label embedding network T3 to embed the real labels. y The mapping is to a high-dimensional feature h, and the mean squared error (MSE) between the predicted label and the true label is used as the loss function to train the pre-trained network.

3. The generative adversarial network spectral data augmentation method for continuous variable regression prediction according to claim 1 or 2, characterized in that: The process of step 5 is as follows: The discriminant network model D employs a binary classification deep convolutional neural network model. Discriminant network D accepts two inputs: a real experimental spectral image and a synthetic spectral image generated by the generator network G. Simultaneously, the high-dimensional feature h output by the label embedding network T3 serves as its conditional input. The output is the probability of classifying the synthetic spectral image as real. The loss function for training discriminant network D is defined as follows: ； in, and These represent the number of real spectral image samples and the number of synthetic spectral image samples, respectively. and These are two noise variables that conform to a standard normal distribution. and The first i One real spectral image sample and one synthetic spectral image sample, and They are respectively and Corresponding tags, hyperparameters The calculation method is as follows ,in, To train the standard deviation of the normalized labels, hyperparameters The calculation method is as follows ,in For the first l The smallest normalized different real label, The number of different labels in the training set, and the target labels of the real samples and generated samples respectively satisfy... and This means that labels are provided to the generator network and the discriminator network. y and and There is some deviation, the formula and This represents the weight assigned to real or synthetic spectral samples within a small range of the target label; the closer the sample, the greater the weight.

4. The generative adversarial network spectral data augmentation method for continuous variable regression prediction according to claim 3, characterized in that: The process of step 6 is as follows: A generator network G is constructed using a deep convolutional neural network. G takes random noise and high-dimensional features obtained from T3 mapping as input and outputs the corresponding synthetic spectral image. The loss function used during the training of generator G is as follows: ； in, It is a random noise that conforms to a standard normal distribution.

5. The generative adversarial network spectral data augmentation method for continuous variable regression prediction according to claim 1 or 2, characterized in that: In step 9, the structural similarity method is used to judge the similarity between the synthesized spectrum and the real experimental spectrum. If the requirements are not met, steps 5-7 are repeated until the requirements are met.

6. The generative adversarial network spectral data augmentation method for continuous variable regression prediction according to claim 1 or 2, characterized in that: In step 11, a DNN or CNN deep regression network model is used to perform continuous variable regression prediction on the spectral data, and the improvement in the prediction ability of the prediction model after data augmentation is compared.

Images