Absorption spectrum inversion method based on neural network and correlation method

Abstract

The invention discloses an absorption spectrum inversion method based on a neural network and a correlation method, and mainly solves the problems of low accuracy, poor real-time performance and strict requirements on absorption spectrum data standards of the existing absorption spectrum temperature inversion. According to the implementation scheme, the method comprises the following steps: acquiring original absorption spectrum data, and preprocessing the original absorption spectrum data; constructing a neural network model comprising a full convolutional network FCN and a long short-term memory network LSTM; training a neural network model by using the preprocessed absorption spectrum data set through a back propagation method to obtain a trained neural network model; inputting to-be-measured absorption spectrum data into the neural network model, and outputting a temperature range corresponding to an absorption spectrum; and post-processing the temperature range output by the neural network based on a correlation method to obtain a temperature value. The method effectively improves the precision and real-time performance of the absorption spectrum temperature inversion, reduces the requirements on the absorption spectrum data standard, and can be used for gas component analysis, environmental monitoring, drug analysis and food safety detection in the fields of chemistry, physics and biology.

Description

Technical Field

[0001] The present invention belongs to the field of absorption optical technology, and specifically relates to an absorption spectrum inversion method, which can be used for gas composition analysis, environmental monitoring, drug analysis, biomolecule detection, and food safety detection in the fields of chemistry, physics and biology. Background Art

[0002] Absorption spectroscopy is a method used to study the interaction between matter and light. By analyzing the degree of absorption of light of different wavelengths by a substance, information about the composition, temperature, concentration, and reaction kinetics of the substance can be obtained. It has a wide range of applications in chemistry, physics, biology, and other fields.

[0003] Existing absorption spectrum inversion methods include spectral temperature method, blackbody radiation method, color temperature measurement method, etc. These methods generally have some disadvantages. First, external environmental factors such as light and background radiation may interfere with the measurement results, especially in spectral temperature measurement. Second, some methods are limited by the resolution and sensitivity of the instrument and may lose accuracy when measuring extremely high or low temperatures. In addition, the blackbody radiation method needs to consider the emissivity of the surface of the object, which is difficult to measure accurately. These methods usually rely on physical assumptions, such as the blackbody radiation law, which may not hold true under certain conditions, resulting in measurement errors. It is also necessary to calibrate the instrument regularly, otherwise it may lead to uncertainty in temperature measurement. When choosing an appropriate method, it is necessary to comprehensively consider the application requirements, environmental conditions and characteristics of the method to achieve a better temperature measurement effect.

[0004] The patent document with application number CN20220391067.7 discloses a method and system for extracting tissue optical parameters based on long short-term memory network, which introduces long short-term memory network into the tissue spatial resolution diffuse reflectance spectrum generation model, fully explores the correlation in the spectral data points, and increases the model prediction ability while reducing the model parameters; when training the tissue spatial resolution diffuse reflectance spectrum generation model, the fiber probe structure is not considered, and all spatial resolution diffuse reflectance spectra emitted from the tissue surface are included in the pre-data set, and the data dimension is larger. Although this method can effectively improve the accuracy of the tissue optical parameter and diffuse reflectance spectrum mapping model. However, due to the insufficient ability of the long short-term memory network to extract local features, the optical parameter extraction accuracy is not high.

[0005] The patent document with application number CN202010585647.0 discloses a "Deep Learning Model for Small Sample Hyperspectral Image Classification", which is based on an encoder-decoder. The original data set is randomly sampled twice to generate a set of random training sample sequences. The global convolutional long short-term memory module is used to extract the spectral dependency features in the long and short ranges and the spatial relationship of pixels in the local range. The global spectrum and spatial attention mechanism are used to extract detailed spectral dependency features and spatial details respectively, so that when the training samples are limited, the most discriminative features can still be extracted to ensure the accuracy of classification. However, the structure of this model is relatively complex, the computing resource requirements are high, and it is sensitive to parameter tuning, which makes it difficult to meet the current needs in the industrial field.

[0006] In summary, the existing absorption spectrum temperature inversion methods generally have the defects of low inversion accuracy, high computational complexity, poor real-time performance, and strict requirements on absorption spectrum data standards. Summary of the invention

[0007] The purpose of the present invention is to propose an absorption spectrum inversion method based on neural network and correlation method to address the shortcomings of existing absorption spectrum inversion methods in terms of computational complexity, inversion accuracy and real-time performance. By combining the deep feature extraction capabilities of the fully convolutional network (FCN) and the long short-term memory network (LSTM) and the precise matching characteristics of the correlation method, the training efficiency and inversion accuracy of the neural network model can be effectively improved, and the computational complexity can be effectively reduced.

[0008] To achieve the above object, the technical solution of the present invention comprises the following steps:

[0009] 1. An absorption spectrum inversion method based on neural network and correlation method, characterized by comprising the following:

[0010] (1) Obtaining raw absorption spectrum data and preprocessing it;

[0011] (2) Construct a neural network model including a fully convolutional network (FCN) and a long short-term memory (LSTM) network;

[0012] (3) using the preprocessed absorption spectrum data set to train the neural network model through a back propagation method to obtain a trained neural network model;

[0013] (4) inputting the absorption spectrum data to be measured into the neural network model, and outputting the temperature range corresponding to the absorption spectrum;

[0014] (5) Post-process the temperature range output by the neural network based on the correlation method to obtain the temperature value.

[0015] Furthermore, the construction includes a neural network model of a fully convolutional network FCN and a long short-term memory network LSTM, and its implementation includes the following:

[0016] Construct a fully convolutional network, which contains three convolutional units cascaded in sequence. Each convolutional unit contains a batch normalization layer, a Relu activation function, and a global pooling operation. The convolution kernel size of the first convolutional unit is 8*8, and the number of convolution kernels is 128; the convolution kernel size of the second convolutional unit is 5*5, and the number of convolution kernels is 256; the convolution kernel size of the third convolutional unit is 3*3, and the number of convolution kernels is 128; each convolution kernel is initialized using Kaiming to improve the convergence speed;

[0017] A long short-term memory network is constructed, which includes three long short-term memory network layers cascaded in sequence, each long short-term memory network layer includes memory neural units, the number of memory neural units in the first long short-term memory network layer is 32; the number of memory neural units in the second long short-term memory network layer is 64; the number of memory neural units in the third long short-term memory network layer is 128;

[0018] Construct a feature fusion module, which includes a convolution layer and a feature aggregation layer. The convolution layer is used to fuse the output features of the full convolution network module and the long short-term memory network module, and the convolution kernel size is 3*3; the feature aggregation layer is used to average the weighted pixels on all channels of the convolution layer output features at each spatial position to achieve feature aggregation;

[0019] The fully convolutional network (FCN) and the long short-term memory (LSTM) network are combined in parallel to form a dual-branch processing architecture for absorption spectral data. The features extracted from the two paths are integrated through a feature fusion module. The fusion process includes a 3×3 convolutional layer for joint feature modeling and a feature aggregation layer for global weighted feature processing. Finally, the integrated feature representation is mapped to the output space through a fully connected layer to form a complete neural network model.

[0020] Compared with the prior art, the present invention has the following advantages:

[0021] Firstly, the present invention realizes efficient inference of the temperature range corresponding to the absorption spectrum curve through the constructed FCN-LSTM neural network, which is significantly better than the traditional method that requires a complex process of comparing temperature curves one by one. The neural network model can eliminate the dependence on experimental conditions and effectively solve the problem of difficulty in temperature measurement caused by the high similarity of the absorption spectrum curves. That is, the temperature range can be quickly inverted with only one spectrum extraction.

[0022] Secondly, the present invention organically integrates the features extracted by the fully convolutional network FCN and the long short-term memory network LSTM, making full use of the advantages of both, namely, utilizing the advantage of the long short-term memory network LSTM in accurately controlling the flow of information through the input gate, forget gate and output gate, and efficiently extracting long sequence features; utilizing the fully convolutional network FCN in effectively extracting local features in sequence data and retaining the temporal structure, better capturing the advantages of contextual information, and achieving high-precision and high-efficiency temperature inversion.

[0023] Thirdly, the present invention utilizes the powerful feature extraction capability of the fully convolutional network (FCN) and does not set a limit on the length of the spectral curve sequence. It can effectively realize temperature inversion regardless of the length of the sequence, thus overcoming the limitations of traditional methods that have strict requirements on absorption spectrum data standards and making it suitable for a wider range of application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is an implementation flow chart of the present invention;

[0025] Figure 2 is a diagram of the neural network structure constructed in the present invention;

[0026] Figure 3 is an absorption spectrum data graph obtained in the present invention;

[0027] Figure 4 It is a graph of the absorption spectrum data and the processed data in the present invention;

[0028] Figure 5 This is a result diagram of temperature inversion of the selected absorption spectrum data in the present invention. DETAILED DESCRIPTION

[0029] The embodiments and effects of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0030] Reference Figure 1 , the implementation steps of this embodiment are as follows:

[0031] Step 1: Obtain raw absorption spectrum data and preprocess it.

[0032] In this embodiment, the input raw absorption spectrum data is as follows: Figure 3 As shown, Z-Score standardization is performed to eliminate the dimension effect and improve data consistency. The standardized data is as follows Figure 4 As shown;

[0033] Subsequently, the principal component analysis (PCA) method was used to reduce the dimensionality of the standardized data: the principal components were extracted by calculating the covariance matrix and performing eigenvalue decomposition, sorted by eigenvalue size, and the top k principal components with the largest contribution were selected to reduce the number of data features and retain the main information, completing the dimensionality reduction process and obtaining the preprocessed absorption spectrum data set.

[0034] Step 2: Build a neural network model including a fully convolutional network (FCN) and a long short-term memory network (LSTM).

[0035] Reference Figure 2 , the implementation of this step is as follows:

[0036] 2.1) The neural network model constructed in this embodiment includes a fully convolutional network, a long short-term memory network, and a feature fusion module. Among them:

[0037] The fully convolutional network part includes three convolution units cascaded in sequence, each of which includes a batch normalization layer, a Relu activation function, and a global pooling operation. The convolution kernel size of the first convolution unit is 8*8, and the number of convolution kernels is 128; the convolution kernel size of the second convolution unit is 5*5, and the number of convolution kernels is 256; the convolution kernel size of the third convolution unit is 3*3, and the number of convolution kernels is 128; each convolution kernel is initialized using Kaiming to improve the convergence speed;

[0038] The long short-term memory network part comprises three long short-term memory network layers cascaded in sequence, each long short-term memory network layer comprises memory neural units, the number of memory neural units in the first long short-term memory network layer is 32; the number of memory neural units in the second long short-term memory network layer is 64; the number of memory neural units in the third long short-term memory network layer is 128;

[0039] The feature fusion module consists of a convolution layer and a feature aggregation layer. The convolution layer is used to fuse the features extracted by the full convolution network and the long short-term memory network, and its convolution kernel size is 3×3; the feature aggregation layer achieves efficient feature integration by averaging the pixel values ​​of the channel dimension at each spatial position.

[0040] 2.2) The fully convolutional network FCN and the long short-term memory network LSTM are combined in parallel to form a dual-branch processing architecture for absorption spectrum data; the dual-branch architecture is used to extract the features of the absorption spectrum data; the extracted features are integrated through a feature fusion module, that is, joint feature modeling is performed through a 3×3 convolutional layer in the fusion module, and the feature aggregation layer performs global weighted processing on the features; the integrated feature representation is then mapped to the temperature range corresponding to the absorption spectrum curve through an external fully connected layer to form a complete neural network model;

[0041] Step 3: Define the loss function of the neural network model.

[0042] Since the absorption spectrum curve data has a "long tail" effect, that is, the number of sample categories is unbalanced. In order to improve the accuracy of model training, this embodiment uses the hinge cross entropy loss function training as the loss function Loss of the neural network, which is expressed as follows:

[0043]

[0044] Among them, y represents the prediction result of the neural network model, margin threshold represents the boundary threshold, and y i represents the true label, Represents the predicted probability of the neural network model for the i-th sample.

[0045] Step 4: Use an adaptive stochastic gradient descent algorithm to train the neural network model to obtain a trained neural network model.

[0046] 4.1) Input the preprocessed absorption spectrum data set into the constructed neural network model to obtain the temperature range corresponding to the absorption spectrum data;

[0047] 4.2) Using the loss function, calculate the regression loss between the output temperature range and the corresponding original temperature range;

[0048] 4.3) Using an adaptive stochastic gradient descent algorithm, the neural network model based on the fully convolutional network FCN and the long short-term memory network LSTM is trained for multiple rounds until the regression loss converges to obtain a trained neural network model.

[0049] Step 5: Use the trained neural network model to obtain the temperature range corresponding to the test set.

[0050] 5.1) Input the data to be tested into the trained neural network model, perform multi-layer convolution operations through the fully convolutional network FCN to extract the local features of the spectral data, extract the global sequence features of the spectral data through the long short-term memory network LSTM, and prevent overfitting through the Dropout mechanism;

[0051] 5.2) The fusion module performs multi-scale fusion of the features extracted by the FCN branch and the LSTM branch, and then further processes the fused features through an external fully connected layer to output the temperature range corresponding to the measured spectrum.

[0052] Step 6: Post-process the temperature range output by the neural network based on the correlation method to obtain a determined temperature value.

[0053] The correlation method is an algorithm for calculating the similarity between two absorption spectrum curves.

[0054] In this embodiment, the specific implementation of post-processing the absorption spectrum data to be measured by using the correlation method is as follows:

[0055] 6.1) Calculate the correlation coefficient r between the measured data and each curve in the reference curve set:

[0056]

[0057] Among them, the correlation coefficient r is an important indicator to measure the matching degree between the measured data and the reference curve, i Represents the i-th predicted value output by the neural network, y i represents the i-th value of the true target parameter, and are the means of the predicted values ​​and the true values ​​respectively, and N is the total number of data samples;

[0058] 6.2) Select the reference curve temperature corresponding to the maximum value of r as the final temperature value output by the neural network. The result is as follows: Figure 5 shown.

[0059] The above content is a preferred embodiment of the present invention and does not constitute any limitation to the present invention. Obviously, a person skilled in the art in the technical field to which the present invention belongs can make some simple deductions or substitutions without departing from the concept of the present invention, but these should be regarded as the protection scope of the present invention.

Claims

1. An absorption spectrum inversion method based on neural network and correlation method, characterized in that: These include: (1) Obtaining raw absorption spectrum data and preprocessing it; (2) Construct a neural network model including a fully convolutional network (FCN) and a long short-term memory (LSTM) network; (3) using the preprocessed absorption spectrum data set to train the neural network model through a back propagation method to obtain a trained neural network model; (4) inputting the absorption spectrum data to be measured into the neural network model, and outputting the temperature range corresponding to the absorption spectrum; (5) Post-process the temperature range output by the neural network based on the correlation method to obtain the temperature value.

2. The method according to claim 1, characterized in that In step (1), the original absorption spectrum data is preprocessed, and the implementation steps include the following: (1a) performing Z-Score standardization processing on the original absorption spectrum data; (1b) The principal component analysis (PCA) method is used to reduce the dimension of the standardized data to reduce the number of data features and retain the main information.

3. The method according to claim 1, characterized in that: In step (2), a neural network model including a fully convolutional network FCN and a long short-term memory network LSTM is constructed, and the implementation steps include the following: (2a) Construct a fully convolutional network, which includes three convolutional units cascaded in sequence. Each convolutional unit includes a batch normalization layer, a Relu activation function, and a global pooling operation. The convolution kernel size of the first convolutional unit is 8*8, and the number of convolution kernels is 128; the convolution kernel size of the second convolutional unit is 5*5, and the number of convolution kernels is 256; the convolution kernel size of the third convolutional unit is 3*3, and the number of convolution kernels is 128; each convolution kernel is initialized using Kaiming to improve the convergence speed; (2b) constructing a long short-term memory network, which includes three long short-term memory network layers cascaded in sequence, each long short-term memory network layer includes memory neural units, the number of memory neural units in the first long short-term memory network layer is 32; the number of memory neural units in the second long short-term memory network layer is 64; the number of memory neural units in the third long short-term memory network layer is 128; (2c) constructing a feature fusion module, which includes a convolution layer and a feature aggregation layer. The convolution layer is used to fuse the output features of the full convolution network module and the long short-term memory network module, and the convolution kernel size is 3*3; the feature aggregation layer is used to average the weighted pixels on all channels of the output features of the convolution layer at each spatial position to achieve feature aggregation; (2d) The fully convolutional network (FCN) and the long short-term memory (LSTM) network are combined in parallel to form a dual-branch processing architecture for absorption spectrum data. The dual-branch architecture is used to extract the features of the absorption spectrum data. The extracted features are integrated through a feature fusion module, that is, joint feature modeling is performed through a 3×3 convolutional layer in the fusion module, and the feature aggregation layer performs global weighted processing on the features. The integrated feature representation is then mapped to the temperature range corresponding to the absorption spectrum curve through an external fully connected layer to form a complete neural network model.

4. The method according to claim 1, characterized in that: In step (3), the preprocessed absorption spectrum data set is used to train the neural network model through the back propagation method, and the implementation steps include the following: (3a) inputting the preprocessed absorption spectrum data set into the constructed neural network model to obtain the temperature range corresponding to the absorption spectrum data; (3b) Combining the characteristics of cross entropy loss and hinge loss, the loss function Loss of the neural network model is defined as: Among them, y represents the prediction result of the model, margin threshold represents the boundary threshold, and y i represents the true label, Represents the model's predicted probability for the i-th sample; (3c) using the loss function, calculating the regression loss between the output temperature range and the corresponding original temperature range; (3d) Using an adaptive stochastic gradient descent algorithm, the neural network model based on the fully convolutional network (FCN) and the long short-term memory network (LSTM) is trained for multiple rounds until the regression loss converges, thereby obtaining a trained neural network model.

5. The method according to claim 1, characterized in that Step (4) inputs the absorption spectrum data to be measured into the neural network model, and outputs the temperature range corresponding to the absorption spectrum. The implementation steps include the following: (4a) The absorption spectrum data to be measured is copied and input into the fully convolutional network FCN and the long short-term memory network LSTM respectively; (4b) Fully convolutional network (FCN) extracts local features and multi-scale information of spectral data through multi-layer convolution operations; (4c) The long short-term memory network LSTM is first dimensionally adjusted to adapt to the sequence modeling requirements, and then the LSTM layer is used to model the temporal dependency of the spectral sequence data, and the Dropout mechanism is used to prevent overfitting; (4d) The features extracted by the FCN branch and the LSTM branch are fused at multiple scales, and then the fused features are further processed by the fully connected layer to output the temperature range corresponding to the measured spectrum.

6. The method according to claim 1, characterized in that In step (5), the temperature range output by the neural network is post-processed based on the correlation method, and the implementation steps include the following: (5a) obtaining corresponding absorption spectrum data as a reference curve set according to the temperature range predicted by the neural network; (5b) calculating the correlation coefficient r between the absorption spectrum data to be measured and each curve in the reference curve set; (5c) All correlation coefficients are sorted and the reference curve temperature corresponding to the maximum value is selected as the temperature value output by the final neural network.

7. The method according to claim 6, characterized in that In step (5b), the correlation coefficient r is calculated as follows: Among them, x i Represents the i-th predicted value output by the neural network, y i represents the i-th value of the true target parameter, and are the means of the predicted values ​​and the true values, respectively, and N is the total number of data samples.

Images