Gas photoacoustic spectrum enhanced recognition device and method and application

Abstract

The invention provides a gas photoacoustic spectrum enhancement recognition device and method and application, and belongs to the technical field of photoacoustic spectrum data processing, and the method comprises the following steps: obtaining a photoacoustic spectrum data set and a noise data set; extracting sound features of each photoacoustic spectrum and fusing the sound features to obtain a multi-dimensional feature vector of each photoacoustic spectrum; taking the enhanced noise data, the multi-dimensional feature vector and the target feature gas concentration as a negative sample, a positive sample and a label, and constructing a sample data set; inputting the sample data set into a dilated convolutional neural network for training to obtain a trained dilated convolutional neural network; and inputting the to-be-identified photoacoustic spectrum into the trained dilated convolutional neural network, and identifying the concentration of the target characteristic gas in the to-be-identified photoacoustic spectrum. The method can be used in photoacoustic spectrum recognition in different scenes, different acousto-optic pools do not need to be frequently replaced, the dilated convolutional neural network is used for photoacoustic spectrum recognition for the first time, and the recognition accuracy can be effectively enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of photoacoustic spectroscopy data processing, and in particular to a gas photoacoustic spectroscopy enhanced identification device and method and application thereof. Background Art

[0002] Gas photoacoustic spectroscopy is a new gas analysis method that measures absorption by detecting the absorption of photon energy by gas molecules to quantitatively analyze the concentration of gas. Compared with the traditional detection method that directly measures the energy of light radiation, gas photoacoustic spectroscopy converts heat energy into sound signals, has higher detection sensitivity, and can shorten the detection time.

[0003] In photoacoustic spectroscopy recognition, due to the complex measurement environment, in order to improve the measurement accuracy, it is often necessary to replace different photoacoustic cells to improve the signal-to-noise ratio of the photoacoustic signal to adapt to different sound environments. In the prior art, photoacoustic spectroscopy data and noise data are collected, and the authenticity of the data is improved by using a generative adversarial network, and then the sample data is input into the recurrent neural network RNN ​​to identify the concentration of the target characteristic gas. However, the recurrent neural network RNN ​​has certain limitations when identifying gas concentrations. For example, as the number of convolutional layers continues to increase, it will face the problem of gradient disappearance and inability to continue training. Therefore, a new gas photoacoustic spectroscopy enhanced recognition method is urgently needed to solve the above problems. Summary of the invention

[0004] The purpose of the present invention is to provide a gas photoacoustic spectroscopy enhanced identification device, method and application, which can be used in photoacoustic spectroscopy identification in different scenarios without the need to frequently replace different acousto-optic cells. It is the first time that a void convolutional neural network is used for photoacoustic spectroscopy identification, which can effectively enhance the accuracy of identification.

[0005] To achieve the above object, the present invention provides a method for enhancing gas photoacoustic spectroscopy identification, comprising the following steps:

[0006] S1. Obtain a photoacoustic spectroscopy dataset and a noise dataset, mix the photoacoustic spectroscopy dataset and the noise dataset to obtain mixed data, and pass the mixed data through a generative adversarial network based on a U-Net structure to obtain an enhanced noise dataset;

[0007] Among them, the construction method of the generative adversarial network based on the U-Net structure is as follows:

[0008] S11. Set up the U-Net network: including downsampling and upsampling operations;

[0009] The mixed data is input into the downsampling path, two 1×3 convolution kernels are connected, the convolution operation is repeated, and then the activation function ReLU is used for correction. A 1×2 convolution kernel with a stride of 2 is used for the first downsampling operation;

[0010] After each downsampling operation is completed, the number of channels is doubled, and the output of the previous downsampling operation is used as the input of the next downsampling operation. The downsampling operation is repeated four times;

[0011] The output of the fifth downsampling is flattened and sent to the first fully connected layer. The number of neurons in the first fully connected layer is 32, and the activation function is LeakyReLU. The output of the first fully connected layer is sent to the second fully connected layer. The number of neurons in the second fully connected layer is 2, and the activation function is Softmax.

[0012] The output of the fifth downsampling is used as the input of the first upsampling. Each upsampling operation includes a deconvolution operation with a convolution kernel size of 1×3, and the convolution output of the same scale in the downsampling path is concatenated. The concatenated result is repeated twice with a 1×3 convolution operation. In the first convolution, the number of channels is halved, and then the activation function ReLU is used for correction. The output of the previous upsampling is used as the input of the next upsampling, and the upsampling operation is repeated four times. After the fifth upsampling operation, a sparse layer is added and mapped to a one-dimensional vector A with the same number of categories, which is the output of the network.

[0013] S12, setting up a generative adversarial network;

[0014] S13, Generative Adversarial Network based on U-Net structure;

[0015] S2, extracting the sound features of each photoacoustic spectrum, and fusing all the sound features to obtain a multidimensional feature vector of each photoacoustic spectrum;

[0016] S4, using the enhanced noise data, multidimensional feature vectors and target characteristic gas concentration as negative samples, positive samples and labels to construct a sample data set;

[0017] S5. Input the sample data set into the dilated convolutional neural network for training to obtain a trained dilated convolutional neural network, including the following sub-steps:

[0018] S51. Building a dilated convolutional neural network

[0019] The atrous convolutional neural network includes multiple convolutional layers, multiple atrous convolutional layers, multiple deconvolutional layers, a normalization operation, an activation function and an output layer;

[0020] Multiple convolutional layers are connected in series, multiple atrous convolutional layers are connected in series after multiple convolutional layers, and multiple deconvolutional layers are connected in series after the atrous convolutional layers.

[0021] S52. Training a dilated convolutional neural network

[0022] Use sample data as input to train the dilated convolutional neural network;

[0023] S6. Input the photoacoustic spectrum to be identified into the trained atrous convolutional neural network to identify the concentration of the target characteristic gas in the photoacoustic spectrum to be identified.

[0024] Preferably, the sound features include Mel-spectrogram cepstral coefficients, frequency, amplitude and phase.

[0025] Preferably, the normalization operation is a BatchNormalization operation.

[0026] Preferably, the activation function is a Leaky ReLU activation function.

[0027] Preferably, the output layer is a Sigmoid output layer, and the loss function is adopted, and the formula is as follows:

[0028]

[0029] Where N represents the number of samples; l (i) represents the true label of the i-th sample, Represents the predicted label of the i-th sample.

[0030] Preferably, the photoacoustic spectroscopy dataset and the noise dataset are mixed, including random matching, pitch shifting, time stretching, and cyclic shifting.

[0031] Preferably, the generative adversarial network based on the U-Net structure uses the U-Net network as a generator.

[0032] The present invention also provides a device for implementing the above-mentioned gas photoacoustic spectroscopy identification enhancement method, comprising a data acquisition module, an extraction module, a sample data construction module, a hole convolutional neural network module and a recognition module;

[0033] The data acquisition module includes a photoacoustic spectroscopy data acquisition module, a noise data acquisition module, a mixing module and a boosting module;

[0034] The photoacoustic spectrum data acquisition module is used to acquire photoacoustic spectrum data;

[0035] The noise data acquisition module is used to acquire noise data;

[0036] The mixing module is used to mix the photoacoustic spectroscopy data and the noise data;

[0037] The enhancement module is a generative adversarial network based on the U-Net structure;

[0038] The extraction module is used to extract the sound features of each photoacoustic spectrum and fuse all the sound features to obtain a multi-dimensional feature vector of each photoacoustic spectrum;

[0039] The sample data construction module is used to construct a sample data set by taking the enhanced noise data, multi-dimensional feature vectors and target characteristic gas concentration as negative samples, positive samples and labels;

[0040] The dilated convolutional neural network module is used to build and train dilated convolutional neural networks;

[0041] The identification module is used to identify the concentration of the gas to be measured.

[0042] The present invention also provides an application of a gas photoacoustic spectrum identification enhancement device in identifying the gas concentration in transformer oil.

[0043] Therefore, the present invention adopts the above-mentioned gas photoacoustic spectroscopy enhanced identification device and method and application, and the technical effects are as follows:

[0044] (1) Constructing a sample dataset can reduce the dependence on the specific structure of the photoacoustic pool, and through the generative adversarial network based on the U-Net structure, the authenticity of the noise reconstruction samples is improved, and the robustness of the model is enhanced;

[0045] (2) By using a dilated convolutional neural network, the recognition accuracy is higher than that of a recurrent neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flow chart of a method for enhanced identification of gas photoacoustic spectroscopy;

[0047] Figure 2 A flowchart for building a generative adversarial network based on the U-Net structure;

[0048] Figure 3 This is a schematic diagram of the structure of a gas photoacoustic spectroscopy enhanced identification device. DETAILED DESCRIPTION

[0049] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0050] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0051] Embodiment 1

[0052] like Figure 1 FIG. 1 is a flow chart of a method for enhancing gas photoacoustic spectroscopy identification according to the present invention, comprising the following steps:

[0053] S1. Obtain a photoacoustic spectroscopy dataset and a noise dataset, mix the photoacoustic spectroscopy dataset and the noise dataset to obtain mixed data, and pass the mixed data through a generative adversarial network based on a U-Net structure to obtain an enhanced noise dataset;

[0054] Mixing of photoacoustic spectroscopy datasets and noise datasets, including random matching, pitch shifting, time stretching, and cyclic shifting.

[0055] Among them, the construction method of the generative adversarial network based on the U-Net structure is as follows (such as Figure 2 shown):

[0056] S11. Set up the U-Net network: including downsampling and upsampling operations;

[0057] The mixed data is input into the downsampling path, two 1×3 convolution kernels are connected, the convolution operation is repeated, and then the activation function ReLU is used for correction. A 1×2 convolution kernel with a stride of 2 is used for the first downsampling operation;

[0058] After each downsampling operation is completed, the number of channels is doubled, and the output of the previous downsampling operation is used as the input of the next downsampling operation. The downsampling operation is repeated four times;

[0059] The output of the fifth downsampling is flattened and sent to the first fully connected layer. The number of neurons in the first fully connected layer is 32, and the activation function is LeakyReLU. The output of the first fully connected layer is sent to the second fully connected layer. The number of neurons in the second fully connected layer is 2, and the activation function is Softmax.

[0060] The output of the fifth downsampling is used as the input of the first upsampling. Each upsampling operation includes a deconvolution operation with a convolution kernel size of 1×3, and the convolution output of the same scale in the downsampling path is concatenated. The concatenated result is repeated twice with a 1×3 convolution operation. In the first convolution, the number of channels is halved, and then the activation function ReLU is used for correction. The output of the previous upsampling is used as the input of the next upsampling, and the upsampling operation is repeated four times. After the fifth upsampling operation, a sparse layer is added and mapped to a one-dimensional vector A with the same number of categories, which is the output of the network.

[0061] S12, setting up a generative adversarial network;

[0062] Adversarial networks are an existing technology, including generators and discriminators.

[0063] The objective function of the adversarial network is:

[0064]

[0065] Where D represents the discriminator and G represents the generator. D(x) represents the probability that x is a real data sample, and G(z) represents the probability distribution of the generated sample z.

[0066] S13, Generative Adversarial Network based on U-Net structure;

[0067] The generative adversarial network based on the U-Net structure uses the U-Net network as the generator. The output of the U-Net network and the mixed data are simultaneously input into the discriminator for optimization.

[0068] S2, extracting the sound features of each photoacoustic spectrum, and fusing all the sound features to obtain a multidimensional feature vector of each photoacoustic spectrum;

[0069] The sound features include Mel-spectrogram cepstral coefficients, frequency, amplitude and phase.

[0070] S4, using the enhanced noise data, multidimensional feature vectors and target characteristic gas concentration as negative samples, positive samples and labels to construct a sample data set;

[0071] S5. Input the sample data set into the dilated convolutional neural network for training to obtain a trained dilated convolutional neural network, including the following sub-steps:

[0072] S51. Building a dilated convolutional neural network

[0073] The atrous convolutional neural network includes multiple convolutional layers, multiple atrous convolutional layers, multiple deconvolutional layers, a normalization operation, an activation function and an output layer;

[0074] Multiple convolutional layers are connected in series, multiple atrous convolutional layers are connected in series after multiple convolutional layers, and multiple deconvolutional layers are connected in series after the atrous convolutional layers.

[0075] The normalization operation is the BatchNormalization operation.

[0076] The activation function is the Leaky ReLU activation function.

[0077] The output layer is a Sigmoid output layer, using a loss function, the formula is as follows:

[0078]

[0079] Where N represents the number of samples; l (i) represents the true label of the i-th sample, Represents the predicted label of the i-th sample.

[0080] S52. Training a dilated convolutional neural network

[0081] Use sample data as input to train the dilated convolutional neural network;

[0082] S6. Input the photoacoustic spectrum to be identified into the trained atrous convolutional neural network to identify the concentration of the target characteristic gas in the photoacoustic spectrum to be identified.

[0083] like Figure 3 As shown, the present invention also provides a device for implementing the above-mentioned gas photoacoustic spectroscopy identification enhancement method, including a data acquisition module, an extraction module, a sample data construction module, a hole convolutional neural network module and a recognition module;

[0084] The data acquisition module includes a photoacoustic spectroscopy data acquisition module, a noise data acquisition module, a mixing module and a boosting module;

[0085] The photoacoustic spectrum data acquisition module is used to acquire photoacoustic spectrum data;

[0086] The noise data acquisition module is used to acquire noise data;

[0087] The mixing module is used to mix the photoacoustic spectroscopy data and the noise data;

[0088] The enhancement module is a generative adversarial network based on the U-Net structure;

[0089] The extraction module is used to extract the sound features of each photoacoustic spectrum and fuse all the sound features to obtain a multi-dimensional feature vector of each photoacoustic spectrum;

[0090] The sample data construction module is used to construct a sample data set by taking the enhanced noise data, multi-dimensional feature vectors and target characteristic gas concentration as negative samples, positive samples and labels;

[0091] The dilated convolutional neural network module is used to build and train dilated convolutional neural networks;

[0092] The identification module is used to identify the concentration of the gas to be measured.

[0093] Therefore, the present invention adopts the above-mentioned gas photoacoustic spectroscopy enhanced identification device, method and application, which can be used in photoacoustic spectroscopy identification in different scenarios without the need to frequently replace different acousto-optic cells. It also uses the hole convolutional neural network for photoacoustic spectroscopy identification for the first time, which can effectively enhance the recognition accuracy.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A gas photoacoustic spectroscopy enhanced identification method, characterized in that: The steps include: S1. Obtain a photoacoustic spectroscopy dataset and a noise dataset, mix the photoacoustic spectroscopy dataset and the noise dataset to obtain mixed data, and pass the mixed data through a generative adversarial network based on a U-Net structure to obtain an enhanced noise dataset; Among them, the construction method of the generative adversarial network based on the U-Net structure is as follows: S11. Set up the U-Net network: including downsampling and upsampling operations; Input the data into the downsampling path, connect two 1×3 convolution kernels, repeat the convolution operation, and then use the activation function ReLU correction, use a 1×2 convolution kernel and a maximum pooling operation with a stride of 2 for the first downsampling operation; After each downsampling operation is completed, the number of channels is doubled, and the output of the previous downsampling operation is used as the input of the next downsampling operation. The downsampling operation is repeated four times; The output of the fifth downsampling is flattened and sent to the first fully connected layer. The number of neurons in the first fully connected layer is 32, and the activation function is LeakyReLU. The output of the first fully connected layer is sent to the second fully connected layer. The number of neurons in the second fully connected layer is 2, and the activation function is Softmax. The output of the fifth downsampling is used as the input of the first upsampling. Each upsampling operation includes a deconvolution operation with a convolution kernel size of 1×3, and the convolution output of the same scale in the downsampling path is concatenated. The concatenated result is repeated twice with a 1×3 convolution operation. In the first convolution, the number of channels is halved, and then the activation function ReLU is used for correction. The output of the previous upsampling is used as the input of the next upsampling, and the upsampling operation is repeated four times. After the fifth upsampling operation, a sparse layer is added and mapped to a one-dimensional vector A with the same number of categories, which is the output of the network. S12, setting up a generative adversarial network; S13, Generative Adversarial Network based on U-Net structure; S2, extracting the sound features of each photoacoustic spectrum, and fusing all the sound features to obtain a multidimensional feature vector of each photoacoustic spectrum; S3, using the enhanced noise data, multidimensional feature vectors and target characteristic gas concentration as negative samples, positive samples and labels to construct a sample data set; S4. Inputting the sample data set into the dilated convolutional neural network for training to obtain a trained dilated convolutional neural network, including the following sub-steps: S41. Building a dilated convolutional neural network The atrous convolutional neural network includes multiple convolutional layers, multiple atrous convolutional layers, multiple deconvolutional layers, a normalization operation, an activation function and an output layer; Multiple convolutional layers are connected in series, multiple atrous convolutional layers are connected in series after multiple convolutional layers, and multiple deconvolutional layers are connected in series after the atrous convolutional layers. S42. Training a dilated convolutional neural network Use sample data as input to train the dilated convolutional neural network; S5. Input the photoacoustic spectrum to be identified into the trained dilated convolutional neural network to identify the concentration of the target characteristic gas in the photoacoustic spectrum to be identified.

2. A gas photoacoustic spectroscopy enhanced identification method according to claim 1, characterized in that: The sound features include Mel-spectrogram cepstral coefficients, frequency, amplitude and phase.

3. A gas photoacoustic spectroscopy enhanced identification method according to claim 2, characterized in that: The normalization operation is the BatchNormalization operation.

4. A gas photoacoustic spectroscopy enhanced identification method according to claim 3, characterized in that: The activation function is the Leaky ReLU activation function.

5. A gas photoacoustic spectroscopy enhanced identification method according to claim 4, characterized in that: The output layer is a Sigmoid output layer, using a loss function, the formula is as follows: Where N represents the number of samples; l (i) represents the true label of the i-th sample, Represents the predicted label of the i-th sample.

6. A gas photoacoustic spectroscopy enhanced identification method according to claim 5, characterized in that: Mix photoacoustic spectroscopy datasets with noise datasets, including random matching, pitch shifting, time stretching, and cyclic shifting.

7. A gas photoacoustic spectroscopy enhanced identification method according to claim 6, characterized in that: The generative adversarial network based on the U-Net structure uses the U-Net network as the generator.

8. A device for implementing the gas photoacoustic spectroscopy identification enhancement method as described in any one of claims 1 to 7.

9. Application of the gas photoacoustic spectroscopy identification enhancement device as claimed in claim 8 in identifying the gas concentration in transformer oil.

Images