Industrial online detection method and system based on near infrared spectrum and medium

Abstract

The invention discloses an industrial online detection method and system based on near infrared spectroscopy and a medium, and aims to construct a CNN-FSA prediction model which comprises a convolutional layer, a pooling layer, a flattening layer, a multilayer sensor, a linear transformation layer, a feature space attention mechanism and an output layer which are connected in sequence. And training the CNN-FSA prediction model by adopting the preprocessed near infrared detection data. Through a CNN-FSA convolutional neural network model fusing a one-dimensional convolutional neural network (1D CNN) and a feature space attention mechanism (FSA), the spectral local feature extraction capability is strong, a feature space compression layer retains key information while dimension reduction is performed, the feature space attention mechanism can dynamically focus on key chemical features, and the feature space attention mechanism can dynamically focus on the key chemical features. Therefore, on-line accurate detection of the concentration of an industrial key reagent with O-H and C-H bond interference C = O as an index is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of near-infrared online detection, and in particular to an industrial online detection method, system and medium based on near-infrared spectroscopy. BACKGROUND

[0002] The material composition evolution of the industrial production process constitutes a multivariate nonlinear dynamic system, and the application of traditional chemical analysis methods in this scenario has certain limitations. In the industrial online detection process based on near-infrared spectroscopy, traditional chemical titration method is time-consuming and requires sample destruction, while the detection technology based on near-infrared spectroscopy (NIRS) has the core competitiveness of fast response, non-destructive sample and simple operation. NIR technology can realize online monitoring and optimize production efficiency. NIR spectral data has high dimensionality, multicollinearity and complex nonlinear characteristics, which poses challenges to modeling methods. Traditional partial least squares (PLSR) effectively handles multicollinearity through latent variable projection, but its inherent linear nature cannot fully capture the nonlinear relationship between the bottom material (glucose, alcohol) and the NIR spectral characteristics during the experiment, and the generalization ability is limited.

[0003] Taking tartaric acid as an example, tartaric acid is a key ingredient of wine and lemon juice beverages, and its concentration directly affects the taste (acidity) and stability of the product. In winemaking, tartaric acid is manually added according to different acidity needs, and the traditional measurement method makes winemaking more complex and relatively high in cost. In the tartaric acid detection scenario, tartaric acid has a carbon-oxygen double bond (C=O) functional group, and near-infrared spectroscopy (NIRS) detection is based on the detection of the concentration of tartaric acid based on the carbon-oxygen double bond. However, the detection environment contains substances containing C-O, C-H and O-H bonds, which will interfere with the carbon-oxygen double bond in near-infrared spectroscopy. When there are problems such as multicollinearity in spectral variables, it is usually difficult to effectively obtain the potential relationship between variables using traditional modeling methods, resulting in a decline in model prediction effect and robustness.

[0004] Therefore, how to improve the accuracy of industrial online detection using near-infrared spectroscopy is a problem to be solved in the field. SUMMARY

[0005] To solve the technical problems in the background art, the present application proposes an industrial online detection method, system and medium based on near-infrared spectroscopy.

[0006] The industrial online detection method based on near-infrared spectroscopy proposed by the present application comprises the following steps: Pretreatment of near-infrared detection data; The CNN-FSA prediction model is constructed, and the CNN-FSA prediction model comprises a convolution layer, a pooling layer, a flattening layer, a multilayer perceptron, a linear transformation layer, a feature space attention mechanism and an output layer connected in sequence, and the CNN-FSA prediction model is trained by using the preprocessed near-infrared detection data. The near-infrared detection data to be detected is input into the trained CNN-FSA prediction model, and a detection result is obtained.

[0007] Preferably, the near-infrared detection data is preprocessed, and specifically comprises a standard normal variable transformation.

[0008] Preferably, the CNN-FSA prediction model comprises at least two levels of convolution layers-pooling layers connected in an alternating series. The first-level convolution layer-pooling layer is processed by M convolution kernels through a ReLU activation function, is connected to a batch normalization layer, and is reduced in dimension through a pooling layer. The second-level convolution layer-pooling layer is processed by 2M convolution kernels through a ReLU activation function, is connected to a batch normalization layer, and is reduced in dimension through a pooling layer. Wherein, M is a natural number.

[0009] Preferably, the CNN-FSA prediction model comprises N levels of convolution layers-pooling layers connected in an alternating series, and the number of convolution kernels of the N+1 level convolution layer is twice the number of convolution kernels of the N level convolution layer.

[0010] Preferably, the first-level convolution layer uses 16 convolution kernels with a size of 20, a step of 6 and a padding of 2, is processed through a ReLU activation function, is connected to a batch normalization layer, and is reduced in dimension through a pooling layer with a size of 2 and a step of 2. The second-level convolution layer uses 32 convolution kernels with a size of 10, a step of 3 and a padding of 2, is processed through a ReLU activation function, is connected to a batch normalization layer, and is reduced in dimension through a pooling layer with a size of 2 and a step of 2.

[0011] Preferably, the output layer comprises a dropout layer and a fully connected layer.

[0012] The application further provides an industrial online detection system based on near-infrared spectroscopy, comprising: A preprocessing module is configured to preprocess near-infrared detection data. A network training module is configured to construct a CNN-FSA prediction model, wherein the CNN-FSA prediction model comprises a convolution layer, a pooling layer, a flattening layer, a multilayer perceptron, a linear transformation layer, a feature space attention mechanism and an output layer connected in sequence, and the CNN-FSA prediction model is trained by using the preprocessed near-infrared detection data. The detection module is configured to input the near-infrared detection data to be detected into the trained CNN-FSA prediction model to obtain a detection result.

[0013] Preferably, the preprocessing module is configured to preprocess the near-infrared detection data by a standard normal variable transformation.

[0014] Preferably, the CNN-FSA prediction model comprises at least two levels of convolutional layers-pooling layers connected in an alternating series. The first-level convolutional layer-pooling layer is processed by M convolutional kernels through a ReLU activation function, and after being connected to a batch normalization layer, the first-level convolutional layer-pooling layer is reduced in dimension by a pooling layer. The second-level convolutional layer-pooling layer is processed by 2M convolutional kernels through a ReLU activation function, and after being connected to a batch normalization layer, the second-level convolutional layer-pooling layer is reduced in dimension by a pooling layer. wherein M is a natural number.

[0015] The application further provides a computer readable storage medium for storing a computer program, wherein the computer program is called by a processor to implement the steps of the above-mentioned industrial online detection method based on near-infrared spectroscopy.

[0016] In the application, the industrial online detection method and system based on near-infrared spectroscopy are proposed to construct a CNN-FSA prediction model, the CNN-FSA prediction model comprises a convolutional layer, a pooling layer, a flattening layer, a multilayer perceptron, a linear transformation layer, a feature space attention mechanism and an output layer connected in sequence, and the CNN-FSA prediction model is trained by using the preprocessed near-infrared detection data. The CNN-FSA convolutional neural network model fusing a one-dimensional convolutional neural network (1D CNN) and a feature space attention mechanism (FSA) has a strong local feature extraction capability for spectra, a feature space compression layer retains key information while reducing dimensions, and the feature space attention mechanism can dynamically focus on key chemical features, so as to realize online accurate detection of the concentration of an industrial key reagent which is disturbed by C-O bonds, C-H bonds and O-H bonds in the detection environment. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The figure is a CNN-FSA model architecture diagram in an embodiment of the industrial online detection method based on near-infrared spectroscopy proposed in the application.

[0018] Figure 2 The figure is a PLSR model fitting regression curve in an embodiment of the industrial online detection method based on near-infrared spectroscopy proposed in the application.

[0019] Figure 3 The figure is a CNN-FSA model fitting regression curve in an embodiment of the industrial online detection method based on near-infrared spectroscopy proposed in the application.

[0020] Figure 4 PLSR model prediction results in an embodiment of the industrial online detection method based on near-infrared spectroscopy proposed in the application.

[0021] Figure 5 CNN-FSA model prediction results in an embodiment of the industrial online detection method based on near-infrared spectroscopy proposed in the application.

[0022] Figure 6 PLSR model prediction sample absolute error in an embodiment of the industrial online detection method based on near-infrared spectroscopy proposed in the application.

[0023] Figure 7 CNN-FSA model prediction sample absolute error in an embodiment of the industrial online detection method based on near-infrared spectroscopy proposed in the application. DETAILED DESCRIPTION

[0024] REFERENCE Figure 1 The industrial online detection method based on near-infrared spectroscopy proposed in the application comprises the following steps: First, the near-infrared detection data is preprocessed. Specifically, the near-infrared detection data is preprocessed by standard normal variable transformation.

[0025] In near-infrared detection, in order to eliminate or weaken physical interference (such as light scattering, baseline drift) irrelevant to target chemical information, the detection data needs to be preprocessed. The standard normal variable transformation (SNV) is used to eliminate the interference introduced by particle size, surface scattering and optical path variation in near-infrared diffuse reflectance spectroscopy. In actual processing, the SNV preprocessing is implemented on the spectral data after the outlier is removed, and the mathematical essence is to convert each spectrum into a standard normal distribution space with a mean of 0 and a standard deviation of 1. The calculation process is as follows:

[0026] Wherein, represents the original spectral value of the i-th sample at the k-th wavelength; is the average value of the i-th sample spectrum; k is 1, 2, …, m, and m is the number of wavelength points.

[0027] For NIR spectral data, the SNV preprocessing corrects the scattering effect and improves the correlation between the spectral data and the target variable (tartaric acid concentration). The optimized data set output by the preprocessing process ensures that the subsequent modeling stage has robust data support.

[0028] Secondly, a CNN-FSA prediction model is constructed, the CNN-FSA prediction model comprising a convolution layer, a pooling layer, a flattening layer, a multilayer perceptron, a linear transformation layer, a feature space attention mechanism and an output layer connected in sequence, and the CNN-FSA prediction model is trained by using the preprocessed near-infrared detection data.

[0029] Specifically, the CNN-FSA prediction model comprises at least two levels of convolution layers-pooling layers connected in series alternately. The first-level convolution layer-pooling layer is processed by M convolution kernels through a ReLU activation function, accesses a batch normalization layer, and is reduced in dimension through a pooling layer; the second-level convolution layer-pooling layer is processed by 2M convolution kernels through a ReLU activation function, accesses a batch normalization layer, and is reduced in dimension through a pooling layer; wherein M is a natural number. By doubling the number of convolution layer channels, the information capacity is maintained while the network is enabled to learn more types of features as the spatial dimension (length) is reduced.

[0030] In one specific implementation of the embodiment, the first-level convolution layer uses 16 convolution kernels with a size of 20, a step of 6 and a padding of 2, is processed through a ReLU activation function, accesses a batch normalization layer, and is reduced in dimension through a pooling layer with a size of 2 and a step of 2; the second-level convolution layer uses 32 convolution kernels with a size of 10, a step of 3 and a padding of 2, is processed through a ReLU activation function, accesses a batch normalization layer, and is reduced in dimension through a pooling layer with a size of 2 and a step of 2.

[0031] The first level uses a larger convolution kernel (20) and a larger step (6) to capture macro features of the spectrum at a higher level (a wider wavelength range). The second level uses a smaller convolution kernel (10) and a smaller step (3) to further capture finer features on the basis of the macro features extracted by the first level.

[0032] In a specific padding mode, the first-level convolution kernel size is 20, and the conventional symmetric padding should be 10. However, in the embodiment, only a small padding (2) is used, which belongs to asymmetric and insufficient padding. The purpose of this processing is: on the one hand, by controlling the padding amount, the invalid boundary compensation is reduced, and excessive pseudo information is avoided in the edge region; on the other hand, the model can be more focused on the features of the main body region of the spectrum during convolution, thereby weakening the interference of the spectrum boundary noise (such as instrument edge effect). In actual calculation, by reasonably selecting the step and the padding amount, the output feature map length is ensured to be a positive integer, while the utilization of boundary information and noise suppression are taken into account.

[0033] In the second level of convolution, the convolution kernel (10) is smaller, and the input feature map has been processed by the first level of convolution and pooling, and the boundary noise is suppressed to a certain extent, so only a small padding (2) can meet the size requirement of the output feature map, and further enhance the attention to the effective center area.

[0034] Further, when the data features are more and more complex, the convolution layers can be increased. In a specific embodiment, the CNN-FSA prediction model includes N levels of convolution layers-pooling layers connected in series alternately, and the number of convolution kernels of the convolution layer at the N+1 level is twice the number of convolution kernels of the convolution layer at the N level.

[0035] In addition, the output layer of the embodiment includes a dropout layer and a fully connected layer.

[0036] Among them, the batch normalization layer accelerates the model convergence, and the pooling layer effectively eliminates the noise in the original spectrum, thereby enhancing the spectral feature information, and the module can realize hierarchical extraction of spectral features. The activation function is a nonlinear operation on the output of the convolution layer to extract more data feature information. The feature flattening layer converts the three-dimensional feature tensor output by the convolution module into a one-dimensional feature vector to provide input for the subsequent fully connected layer. The multilayer perceptron module has a single hidden layer structure, an input dimension, and an output dimension. The LeakyReLU activation function (negative slope 0.1) is used to enhance the nonlinear expression ability. The batch normalization stabilizes the training process. The attention mechanism is inspired by the human visual system and simulates the ability to selectively focus on important areas, allowing the model to assign higher weights to key features of the input data, thereby enhancing feature expression ability. The feature space attention module uses a multi-head self-attention mechanism (4 attention heads) to dynamically weight the features in an 8-dimensional feature space. This module can adaptively enhance the contribution of key chemical features while suppressing noise interference. The output layer includes Dropout regularization (ratio 0.05) to prevent overfitting, and a fully connected layer to map the 8-dimensional features to a 1-dimensional concentration prediction value.

[0037] Finally, the near-infrared detection data to be detected is input into the trained CNN-FSA prediction model to obtain a detection result.

[0038] The embodiment also proposes an industrial online detection system based on near-infrared spectroscopy, comprising: A preprocessing module for preprocessing near-infrared detection data; A network training module for constructing a CNN-FSA prediction model, the CNN-FSA prediction model comprising a convolution layer, a pooling layer, a flattening layer, a multilayer perceptron, a linear transformation layer, a feature space attention mechanism, and an output layer connected in sequence, and the CNN-FSA prediction model is trained using the preprocessed near-infrared detection data. The detection module is configured to input the near-infrared detection data to be detected into the trained CNN-FSA prediction model, and obtain a detection result.

[0039] In the embodiment, the proposed industrial online detection method and system based on near-infrared spectroscopy is used to construct a CNN-FSA prediction model, which includes a convolution layer, a pooling layer, a flattening layer, a multilayer perceptron, a linear transformation layer, a feature space attention mechanism and an output layer connected in sequence, and the CNN-FSA prediction model is trained by using the preprocessed near-infrared detection data. The CNN-FSA convolutional neural network model fusing a one-dimensional convolutional neural network (1D CNN) and a feature space attention mechanism (FSA) has strong local feature extraction capability for spectra, the feature space compression layer retains key information while reducing dimensions, and the feature space attention mechanism can dynamically focus on key chemical features, thereby realizing online accurate detection of the concentration of an industrial key reagent with C=O as an index, which is interfered by O-H and C-H bonds.

[0040] The embodiment fuses a multilayer perceptron (MLP), a feature projection linear transformation layer and a feature space attention mechanism on the basis of a traditional CNN model.

[0041] The MLP connected after the CNN effectively solves the problem that feature peaks (O-H 1450nm, C=O 1710nm) are easily overlapped with coexisting substances (glucose, alcohol and lactic acid), realizes global feature integration across a wavelength domain, and solves the peak overlap misjudgment caused by the local receptive field of a traditional CNN. The nonlinear activation (LeakyReLU) accurately fits the concentration-absorbance segmented response curve specific to tartaric acid, and improves the robustness of the model to complex background interference.

[0042] The feature projection linear transformation layer added at the end of the MLP solves the problem of mixed feature space caused by the coexistence of multiple components in an industrial field, and constructs a tartaric acid specific subspace: automatically learns the negative weight matrix of interfering substances (such as alcohol and glucose feature weights).

[0043] The four feature space attention mechanisms are used to dynamically focus on key wavelengths and self-learn wavelength importance weights.

[0044] The embodiment also proposes a computer readable storage medium for storing a computer program, wherein the computer program is called by a processor to implement the steps of the above-mentioned industrial online detection method based on near-infrared spectroscopy.

[0045] The embodiment also evaluates the prediction performance of the CNN-FSA prediction model by R 2 , RMSE and MAE with partial least squares regression (PLSR).

[0046] 1) Coefficient of determination (R2): R2 reflects the degree of fitting of the model to the data, and the calculation formula is:

[0047] 2) Root mean square error (RMSE): RMSE measures the deviation between the predicted value and the reference value, and is an important indicator commonly used to evaluate the prediction performance of the model. The formula is:

[0048] 3) Mean absolute error (MAE): MAE represents the average absolute deviation between the predicted value and the actual value, and the formula is:

[0049] where is the number of samples; is the actual value of the th sample; is the predicted value of the th sample; is the arithmetic mean of the actual values of the samples.

[0050] Among the model evaluation indicators, R² reflects the overall fitting ability of the model, and RMSE and MAE quantify the prediction error from different angles to help compare model performance.

[0051] Taking tartaric acid detection as an example, the near-infrared detection data set is divided into a training set and a test set, and then the training set and the test set are preprocessed respectively.

[0052] Table 1: Comparison of fitting effects of two models on training set and test set

[0053] According to the Figure 2 , 3 and Table 1, the comparison of model performance shows that the predicted values of the two models are highly correlated with the actual observed values, and the scatter points are closely distributed near the ideal fitting line. Whether it is a training set or a test set, the R² value is extremely high, and the RMSE and MAE values are extremely low, indicating that both models have extremely strong fitting ability for training data, and also show excellent prediction accuracy on the internal test set, and the performance of the two is close.

[0054] Table 2: Comparison of prediction effects of two models on prediction samples

[0055] The Figure 4 , 5The prediction performance of the partial least squares regression (PLSR) and the convolutional neural network model based on two rounds of independent prediction data is shown in Table 2. The true value and the predicted value of the two groups of experiments show a consistent upward trend with the experiment, with a value range of 0.000-0.025, reflecting the dynamic change of the tartaric acid concentration in the experiment. Table 2 clearly lists the evaluation indicators of the PLSR and CNN-FSA models for independent prediction data. The determination coefficient R² of the PLSR model is 0.968750, the root mean square error (RMSE) is 0.001214, the mean absolute error (MAE) is 0.001059, and the maximum error is 0.003594; the R² of the convolutional neural network model is 0.989558, the RMSE is 0.000702, the MAE is 0.000580, and the maximum error is 0.002363. The predicted values of the two models are highly correlated with the true values, and the scatter distribution follows the time sequence, indicating that the model has strong prediction ability for unseen data. However, the CNN-FSA model effectively captures the nonlinear features in the spectrum through the convolutional layer, and shows better prediction accuracy than PLSR (RMSE is reduced by 42.17%, and MAE is reduced by 45.23%).

[0056] attached Figure 6 , 7 The absolute error and the mean absolute error (MAE) of each sample point of the PLSR and CNN-FSA models on the independent prediction data set are shown. It can be seen that the data points with large absolute error of the two models are mainly concentrated in the initial stage of the experiment. This is because the tartaric acid concentration is low at this stage, resulting in weak spectral change signal caused by tartaric acid, and increasing the difficulty of detection. Overall, the MAE of the CNN-FSA model is 0.000580, and the MAE of the PLSR model is 0.001059. The error fluctuation of the CNN-FSA model is generally smaller than that of the PLSR model, indicating that the CNN-FSA model is superior to the traditional PLSR in the overall prediction accuracy of the data.

[0057] The results show that the CNN-FSA model has an advantage in capturing the complex nonlinear relationship between the NIR spectrum and the tartaric acid concentration. This advantage is mainly due to the strong extraction ability of the CNN-FSA to the local features of the spectrum, the key information retained by the feature space compression layer while reducing the dimension, and the dynamic focusing on the key chemical features by the feature space attention mechanism. Combined with SNV preprocessing, the adaptability of the model to high-dimensional NIR data is effectively improved.

[0058] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An industrial online detection method based on near infrared spectroscopy, characterized in that: The following steps are involved: Preprocessing of near infrared detection data; Constructing a CNN-FSA prediction model, the CNN-FSA prediction model includes a convolutional layer, a pooling layer, a flattening layer, a multi-layer perceptron, a linear transformation layer, a feature space attention mechanism, and an output layer connected in sequence, and using the preprocessed near-infrared detection data to train the CNN-FSA prediction model; The near-infrared detection data to be detected is input into the trained CNN-FSA prediction model to obtain the detection results.

2. The industrial online detection method based on near infrared spectroscopy according to claim 1, characterized in that: The preprocessing of the near-infrared detection data specifically includes: standard normal variable transformation.

3. The industrial online detection method based on near infrared spectroscopy according to claim 1, characterized in that: The CNN-FSA prediction model includes at least two levels of convolutional layers and pooling layers connected in series alternately; The first-level convolution layer-pooling layer is processed by M convolution kernels and the ReLU activation function, connected to the batch normalization layer, and then reduced in dimension by the pooling layer; The secondary convolutional layer-pooling layer is processed by 2M convolution kernels and the ReLU activation function, connected to the batch normalization layer, and then reduced in dimension by the pooling layer; Wherein, M is a natural number.

4. The industrial online detection method based on near infrared spectroscopy according to claim 3, characterized in that: The CNN-FSA prediction model includes N levels of convolutional layers and pooling layers that are alternately connected in series, and the number of convolution kernels in the N+1th level convolutional layer is twice the number of convolution kernels in the Nth level convolutional layer.

5. The industrial online detection method based on near infrared spectroscopy according to claim 3, characterized in that: The first convolutional layer uses 16 convolution kernels of size 20, a stride of 6, and a padding of 2. It is processed by the ReLU activation function and connected to the batch normalization layer. The dimensionality is reduced by the pooling layer of size 2 and stride 2; The second convolutional layer uses 32 convolution kernels of size 10, a stride of 3, and a padding of 2. It is processed by the ReLU activation function, connected to the batch normalization layer, and then reduced in dimension by a pooling layer of size 2 and stride 2.

6. The industrial online detection method based on near infrared spectroscopy according to claim 1, characterized in that: The output layer includes a dropout layer and a fully connected layer.

7. An industrial online detection system based on near infrared spectroscopy, characterized in that: include: A preprocessing module, used to preprocess near-infrared detection data; A network training module is used to build a CNN-FSA prediction model, which includes a convolutional layer, a pooling layer, a flattening layer, a multi-layer perceptron, a linear transformation layer, a feature space attention mechanism, and an output layer connected in sequence, and is trained using preprocessed near-infrared detection data; The detection module is used to input the near-infrared detection data to be detected into the trained CNN-FSA prediction model to obtain the detection results.

8. The industrial online detection system based on near infrared spectroscopy according to claim 7 is characterized in that: The preprocessing module is used to preprocess the near-infrared detection data through standard normal variable transformation.

9. The industrial online detection system based on near infrared spectroscopy according to claim 7, characterized in that: The CNN-FSA prediction model includes at least two levels of convolutional layers and pooling layers connected in series alternately; The first-level convolution layer-pooling layer is processed by M convolution kernels and the ReLU activation function, connected to the batch normalization layer, and then reduced in dimension by the pooling layer; The secondary convolutional layer-pooling layer is processed by 2M convolution kernels and the ReLU activation function, connected to the batch normalization layer, and then reduced in dimension by the pooling layer; Wherein, M is a natural number.

10. A computer-readable storage medium, characterized in that Used to store a computer program, which is called by a processor to implement the steps of the industrial online detection method based on near-infrared spectroscopy according to any one of claims 1 to 6.

Images