In-situ catalytic reaction infrared spectrum characterization data intelligent analysis system
By establishing an in-situ infrared spectroscopy intelligent analysis system and utilizing preprocessing and machine learning models, the real-time and accuracy issues of infrared spectral data analysis in complex in-situ catalytic reactions were solved, achieving efficient data analysis and species identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DALIAN INSTITUTE OF CHEMICAL PHYSICS CHINESE ACADEMY OF SCIENCES
- Filing Date
- 2024-11-01
- Publication Date
- 2026-05-08
AI Technical Summary
Existing infrared spectroscopy data analysis methods struggle to achieve real-time, automated, and intelligent data analysis in complex in-situ catalytic reactions. This is especially true when faced with complex changes in adsorption states, spectral baseline drift, and increased noise, making data analysis challenging and requiring high timeliness.
An in-situ infrared spectroscopy intelligent analysis system was established, including a preprocessing module, a feature extraction module, and a machine learning recognition module. Preprocessing methods such as segmented difference, baseline correction, smoothing and noise reduction, and fitting peak division were used, combined with multiple machine learning models for feature recognition and analysis. A database was established for self-correction and reinforcement training.
It enables real-time automated data analysis of in-situ catalytic reactions, improves the reliability and accuracy of data preprocessing and feature identification, reduces the impact of unknown changes, and enhances the efficiency and accuracy of dynamic data analysis and species identification.
Smart Images
Figure CN121994735A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent analysis technology of infrared spectral data, and in particular to a method for preprocessing and feature recognition of infrared spectra of in-situ catalytic reactions. Background Technology
[0002] Infrared spectroscopy has wide applications in energy catalysis and biomedicine. Currently, infrared spectral data analysis mainly relies on expert experience and theoretical calculations. However, due to the large amount of information contained in spectral data, the presence of multiple peaks for specific functional groups, and the overlap of peaks for similar functional groups, data analysis is challenging and requires a high level of knowledge in physicochemical fields such as molecular vibrational theory. For example, carbon monoxide adsorption on metal surfaces exhibits complex patterns such as linear adsorption, bridged adsorption, and polyatomic adsorption. Especially during complex in-situ catalytic reaction detection, these adsorption states are affected by continuous temperature changes and chemical reactions, leading to peak position and shape shifts, spectral baseline drift, increased local noise, and signal intensity shifts. These unknown changes further increase the difficulty of data analysis. Simultaneously, the in-situ reaction process generates a large amount of spectral data in real time, placing high demands on the timeliness of data analysis. In the field of energy catalysis research, there is an urgent need for real-time, automated, and intelligent data analysis of dynamically changing in-situ infrared spectra. Summary of the Invention
[0003] This invention addresses the real-time in-situ infrared spectroscopy data analysis in energy catalysis research, proposing an intelligent in-situ infrared spectroscopy analysis system with strong adaptive and error-correcting capabilities. An in-situ infrared spectroscopy database is established, and each infrared spectrum in the database is analyzed individually. The in-situ infrared spectral data is preprocessed using a segmentation and difference feedback model. Subsequently, spectral data features are extracted and corrected using existing knowledge and experience. The preprocessed and analyzed spectral data is merged and stored with the original data to form a new database used for training the spectral recognition model. Furthermore, enhanced training is conducted for weak spectral signals in the in-situ reaction process to further improve the accuracy of data analysis and feature recognition.
[0004] The technical solution adopted by this invention to achieve the above objectives is: an intelligent analysis system for in-situ catalytic reaction infrared spectral characterization data, comprising:
[0005] The preprocessing module is used to perform real-time dynamic preprocessing on in-situ infrared spectral data.
[0006] The feature extraction module is used to extract spectral feature parameters from the preprocessed spectral data.
[0007] The machine learning recognition module is used to identify spectral feature parameters through a machine learning model to obtain the chemical structure information of the catalytic reaction intermediates and products contained therein.
[0008] The data analysis module is used to obtain the dynamic change trend of in-situ reaction products based on the changes in chemical structure with reaction conditions.
[0009] The preprocessing module includes:
[0010] 1) Spectral data difference preprocessing module, used to make a standard curve of the infrared spectrum of the sample in the initial state before the catalytic reaction, use the difference between the spectral data during the reaction process and the standard curve, and divide the spectral range into segments based on the data difference subtraction result;
[0011] 2) Baseline correction module, used to select different correction models for each spectral data according to the segmentation results to perform baseline removal processing and form corrected spectral data;
[0012] 3) Smoothing and noise reduction module, used to smooth and reduce noise for each corrected spectral data;
[0013] 4) Valley finding and segmentation module, which is used to identify the location of valley points based on the smoothed and denoised data by calculating the first and second derivatives of the spectral data, and automatically segment the spectrum according to the location of the valley points;
[0014] 5) Peak fitting module, used to perform peak fitting on each segment of spectral data after segmentation and output peak characteristic parameters;
[0015] 6) The cyclic feedback evaluation module is used to compare the characteristic differences between the fitted curve and the original infrared spectrum, including peak position, peak area and peak characteristic parameters, to obtain the ratio of peak area and peak height of the fitted curve and the original infrared spectrum, which are used as the determination coefficient and peak height weighting factor, respectively. After the current i-th iteration of fitting the peak, if the determination coefficient and weighting factor are still lower than the threshold, return to step 4) to adjust the segment position in the valley segmentation module.
[0016] If the weight factor obtained in the (i+1)th iteration is still lower than the threshold, return to step 3) to adjust the filter window size in the smoothing and noise reduction module;
[0017] If the weight factor obtained in the (i+2)th iteration is still lower than the threshold, return to step 2) to adjust the parameters of different calibration models until the weight factor is not lower than the threshold.
[0018] The baseline correction module includes multiple correction models:
[0019] A polynomial fitting model is used to fit data using the least squares method to obtain polynomial-fitted data.
[0020] The segmented baseline fitting model is used to divide the data into several segments and use different fitting methods for each segment to obtain the data after segmented baseline fitting.
[0021] The moving window smoothing model is used to slide a window of a set length over a data sequence. The statistical value of each data point within the window is used as the smoothing value of the window's center point, thus obtaining the smoothed data.
[0022] The penalized least squares model is used to minimize the squared error between the predicted and actual values in a regression model, thus obtaining the regressed data.
[0023] The fitting peak division module performs the following steps:
[0024] 1) Based on the symmetry and distribution of spectral features, Gaussian and Lorentz fitting operations are performed in parallel to evaluate the applicability of Gaussian and Lorentz fitting based on the minimum deviation of the fitting:
[0025] The formula for Gaussian fitting is:
[0026] The formula for Lorentz fitting is:
[0027] Where a is the peak height, b is the peak position, and c is the peak width; y(x) represents the fitted value, and x represents the input spectral wavenumber;
[0028] 2) Evaluate the fitting effect using two evaluation criteria: For each fitting method, either Gaussian or Lorentz, obtain the coefficient of determination R. 2 The ratio of peak height to peak height, H;
[0029] Coefficient of determination R 2 Used to measure how well the fitted curve fits the actual data:
[0030]
[0031] Among them, y i It is the actual data, that is, the spectral data after wave trough segmentation, f i It is fitted data. It is the average of the actual data;
[0032] Ratio of peak height values:
[0033]
[0034] Among them, h fitted It is the height of the fitted peak, h originalIt is the height of the peak in the spectral data after valley segmentation;
[0035] 3) When the coefficient of determination R 2 When the peak height ratio H reaches the set range, the fitting effect is considered good, and the peak feature parameters, including peak height a, peak position b, and peak width c, are output; otherwise, return to step 1) to re-identify and segment the shoulder peaks until the coefficient of determination R is reached. 2 Until the peak height ratio H reaches the set range;
[0036] 4) Select the coefficient of determination R 2 The fitting method with a smaller peak height ratio H outputs peak characteristic parameters.
[0037] The feature extraction module obtains the peak area based on the peak feature parameters; the peak height, peak position, peak width, and peak area together constitute the spectral feature parameters, which are used for machine learning model recognition.
[0038] The machine learning recognition module includes:
[0039] Machine learning models, including support vector machines, random forests, backpropagation neural networks, and CNNs, are used to identify the chemical structure information of catalytic reaction intermediates and products contained in each model based on the input spectral feature parameters. The classification results of each model are used to obtain the final classification result, i.e., the chemical structure information of catalytic reaction intermediates and products contained in the model, through a voting method.
[0040] The model training module is used to pre-train each machine learning model; the input of the model is labeled in-situ infrared spectra, the labels including sample information, temperature information and time information; the output of the model is the chemical structure information of the included catalytic reaction intermediates and products.
[0041] The model evaluation and optimization module is used to evaluate and optimize machine learning models, and update machine learning models.
[0042] The data analysis module performs the following steps:
[0043] 1) Save the in-situ infrared spectra of the catalytic reaction line by line to the database and sort them according to the tag information;
[0044] 2) After analyzing each infrared spectrum, determine the characteristics of organic functional groups under each reaction temperature condition;
[0045] 3) Based on temperature and time changes, the changes in the content of organic functional groups in the catalytic reaction process are automatically generated, i.e., the dynamic change trend of the in-situ reaction products.
[0046] A method for intelligent analysis of infrared spectral characterization data of in-situ catalytic reactions includes the following steps:
[0047] For in-situ infrared spectral data, real-time dynamic preprocessing is performed through a preprocessing module;
[0048] After preprocessing, spectral data are used to extract spectral feature parameters.
[0049] The spectral feature parameters are identified using a machine learning model to obtain the chemical structure information of the catalytic reaction intermediates and products contained therein.
[0050] The dynamic change trend of in-situ reaction products can be obtained by measuring the chemical structure under different reaction conditions.
[0051] The in-situ infrared spectral data undergoes real-time dynamic preprocessing via a preprocessing module, including the following steps:
[0052] 1) The spectral data difference preprocessing module generates a standard curve from the infrared spectrum of the sample in its initial state before the catalytic reaction. It then calculates the difference between the spectral data during the reaction process and the standard curve, and segments the spectral range based on the data difference subtraction result.
[0053] 2) The baseline correction module selects different correction models for each spectral data point based on the segmentation results to perform baseline removal processing, thereby generating corrected spectral data;
[0054] 3) The smoothing and noise reduction module performs smoothing and noise reduction on each corrected spectral data;
[0055] 4) The valley segmentation module identifies the location of valley points based on the smoothed and denoised data by calculating the first and second derivatives of the spectral data, and automatically segments the spectrum according to the location of the valley points.
[0056] 5) The peak fitting module performs peak fitting on each segment of the spectral data and outputs peak characteristic parameters;
[0057] 6) The cyclic feedback evaluation module compares the characteristic differences between the fitted curve and the original infrared spectrum, including peak position, peak area, and peak characteristic parameters, and obtains the ratio of peak area and peak height of the fitted curve to the original infrared spectrum, which are used as the determination coefficient and peak height weighting factor, respectively. After the current i-th iteration of fitting the peak, if the determination coefficient and weighting factor are still lower than the threshold, return to step 4) to adjust the segment position in the valley finding segmentation module.
[0058] If the weight factor obtained in the (i+1)th iteration is still lower than the threshold, return to step 3) to adjust the filter window size in the smoothing and noise reduction module;
[0059] If the weight factor obtained in the (i+2)th iteration is still lower than the threshold, return to step 2) to adjust the parameters of different calibration models until the weight factor is not lower than the threshold.
[0060] The step of identifying spectral feature parameters using a machine learning model to obtain the chemical structure information of the included catalytic reaction intermediates and products includes the following steps:
[0061] Machine learning models include support vector machines, random forests, backpropagation neural networks, and CNNs; for each model, based on the input spectral feature parameters, the chemical structure information of the catalytic reaction intermediates and products is identified; the classification results of each model are used to obtain the final classification result, i.e., the chemical structure information of the catalytic reaction intermediates and products, through a voting method.
[0062] The model training module pre-trains each machine learning model; the input to the model is a labeled in-situ infrared spectrum, the labels including sample information, temperature information and time information; the output of the model is the chemical structure information of the included catalytic reaction intermediates and products.
[0063] The present invention has the following beneficial effects and advantages: it performs real-time automatic data preprocessing and feature recognition for in-situ catalytic reaction processes, improves the reliability and scientific nature of data preprocessing and feature recognition by utilizing a self-correcting loop function, and realizes intelligent analysis of in-situ infrared spectral data by combining database query and deep learning technology.
[0064] 1. The intelligent preprocessing unit focuses on the in-situ catalytic reaction process. It uses temperature and time to label spectral data, adding sequence information to each infrared spectrum. Based on this sequence information, it performs spectral subtraction on the in-situ infrared spectra to reduce interference caused by sample differences. Furthermore, to address the reliability and accuracy of the spectral preprocessing, it employs limiting conditions such as the coefficient of determination and achieves self-correction by progressively returning to the upper-level self-check, thus overcoming overfitting issues.
[0065] 2. The modules in the preprocessing unit work together to handle complex in-situ catalytic reaction detection processes, such as peak position and shape shifts, spectral baseline drift, increased local noise, and signal strength shifts. They can promptly process these changes in different ways to reduce the impact of these unknown changes on the spectral data and ensure the real-time performance and robustness of the data processing.
[0066] 3. In the analysis of in-situ infrared spectra, not only are the characteristics of the spectral data themselves analyzed, but also the organic functional groups corresponding to the spectral characteristics are intelligently analyzed by establishing a standard database and using deep learning methods. The analysis results are then used to plot the variation patterns of catalytic reaction intermediates and products.
[0067] 4. This system establishes and continuously enriches the in-situ infrared spectral database for catalytic reactions, improving the efficiency and accuracy of dynamic data analysis and species identification during the reaction process. In addition, new experimental results are placed into the deep learning training set, and through unsupervised training, the accuracy of intelligent spectral recognition is further improved, forming a closed loop in which the dataset and model improve simultaneously. Attached Figure Description
[0068] Figure 1 This is a schematic diagram showing the connections of each module in the in-situ reaction infrared spectroscopy identification system;
[0069] Figure 2 This is a flowchart of the workflow of each unit in the preprocessing module;
[0070] Figure 3 This is a diagram of the internal structure of a CNN (Convolutional Neural Network).
[0071] Figure 4 Here is an example of CO in-situ reaction analysis: (a) raw infrared spectrum; (b) baseline correction; (c) fitting and peak separation; (d) feature identification of species variation with temperature. Detailed Implementation
[0072] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.
[0073] like Figure 1 As shown, the present invention is achieved through the following process:
[0074] 1. Establish a database
[0075] This invention constructs an in-situ infrared spectroscopy database based on MySQL version 3.8.37. For each catalytic reaction's in-situ infrared spectrum, a corresponding data table and functional group characteristic spectrum are created in the database. Each data table stores all sampling data for the entire in-situ reaction, which can be used for subsequent data analysis. In addition, the database stores the characteristic values corresponding to each functional group after processing the in-situ standard data, for rapid identification in the subsequent prediction process.
[0076] 2. Intelligent preprocessing module
[0077] like Figure 2As shown, the intelligent preprocessing module standardizes the raw spectral data, including six units: standard spectral difference, segmented baseline correction, smoothing and noise reduction, valley finding and segmentation, fitting peak division, and cyclic feedback evaluation. Based on knowledge and experience, the functional group characteristics corresponding to the wavenumber range of infrared spectra have strong scientific basis. The 4000–1000 cm⁻¹ range is divided into three segments: 1000–1800 cm⁻¹, 1800–2200 cm⁻¹, and 2650–4000 cm⁻¹. For different wavenumber ranges, the system of this invention will automatically perform standard spectral difference based on the data characteristics. This system tags the real-time acquired in-situ infrared spectra, including sample information, temperature information, and time information. According to the temperature change law of the chemical reaction process, it performs differential spectral processing on the high-temperature infrared spectra. By automatically judging the validity of the differential spectral results based on the preset segment range, if the data after difference subtraction within a segment is negative, the segment is automatically deleted and regarded as a saturated signal, and no analysis is performed to reduce interference (when the differential spectral result is negative, it is regarded as an invalid interval). The system uses intelligent decision-making feature analysis to analyze the interval.
[0078] Based on the difference spectrum analysis results, the system intelligently selects the baseline correction method, including iterative polynomial fitting, piecewise baseline fitting, moving window smoothing, and penalized least squares method. The system simultaneously uses the above baseline correction methods in parallel, evaluates the reliability of the baseline correction by peak position feature restoration, and selects the baseline correction method.
[0079] 1) Polynomial fitting
[0080] Polynomial curve fitting is a technique for fitting data by using polynomial functions to approximate data points, thereby helping to reveal trends and patterns in the data. Polynomial fitting can handle various forms of relationships and is suitable for situations where data trends need to be described. The goal of polynomial fitting is to find a polynomial function that approximates a given set of data points as closely as possible. This is achieved by minimizing the error between the data points and the fitted curve. A commonly used error metric is the sum of squared errors, also known as least squares. Suppose we have a set of data points (x... i y i (where i = 1, 2, 3, ..., n), where n represents the sampling length (the number of data points on a spectral curve). If we want to find a P(x) to fit these points, the general form of the polynomial is: P(x) = a0 + a1x + a2x 2 +a m x m , where m represents the order or degree of the polynomial, x represents the intensity of the sampled data, i.e., the spectrum, y is the wavenumber on the horizontal axis of the spectrum, and y is the intensity on the vertical axis.
[0081] 2) Piecewise baseline fitting
[0082] Piecewise baseline fitting is a technique for identifying and adjusting baselines in data, particularly useful for processing data with distinct interval characteristics, such as in spectral data. This method more accurately captures the baseline trend of the data by dividing it into segments and then applying a different fitting method to each segment. The basic idea of piecewise baseline fitting is to divide the data into several intervals, applying a different baseline fitting method or model to each interval. This can handle nonlinear trends and abrupt changes that may exist in the data, thus better describing the baseline of the data.
[0083] 3) Smooth window movement
[0084] Moving window smoothing is a technique used to reduce noise in sequential data, helping to reveal trends and patterns. The basic idea is to smooth the data by calculating the average of data points within a window surrounding each data point. The fundamental principle of moving window smoothing is to slide a fixed-length window across the data sequence, using the average (or other statistic) of each data point within the window as the smoothing value at the window's center. This smooths out short-term fluctuations in the data sequence, thus highlighting long-term trends. A simple formula for moving window smoothing is... Where k is the window size, x i This represents the i-th sampling point of the infrared sampling data. SMA t This is the smoothed data.
[0085] 4) Penalized Least Squares Method
[0086] Penalized least squares is an extension of regression analysis that incorporates a penalty term to control model complexity, thereby improving generalization ability and reducing overfitting. The most commonly used penalized least squares method is ridge regression. The goal of traditional least squares is to find a regression model that minimizes the squared error between the model's predictions and actual values. The specific objective function is... in, The value of y is the spectrum after smoothing through the model. i Here, represents the actual value of the spectrum, n represents the number of data points, i is the index of the sampling point (i.e., the i-th point), and j is the total number of sampling points. Penalized least squares adds a penalty term to this objective function to control the complexity of the model or the selection of features. This avoids overly complex models and reduces the risk of overfitting. A common objective function for ridge regression is... in, The value of y is the spectrum after smoothing through the model. i The actual value of the spectrum, i is the sampling point index, i.e., the i-th point, j is the total number of sampling points, σ is the penalty coefficient, and β is the actual value of the spectrum. jThese are the regression coefficients of the model. Ridge regression controls the complexity of the model by penalizing the sum of squares of the coefficients, and is particularly suitable for situations where there is multicollinearity among features.
[0087] After baseline correction, the spectral data undergoes smoothing and denoising. A Savitzky-Golay filter is applied for smoothing to reduce data noise fluctuations. A custom local window size is used, and the mean square error (MSE) is employed as the smoothing evaluation metric. The main task of the smoothing and denoising unit is to improve signal quality for data with water noise. First, a Savitzky-Golay filter is used to smooth the spectral data. The Savitzky-Golay filter performs a polynomial fit on the data within a local window and replaces the original data points with the fitted result, thus effectively removing noise while preserving the main characteristics of the signal. The specific formula is... Among them, C i These are the filter coefficients, where m represents half of the window and i represents the offset from the center point. They are obtained by performing polynomial fitting on the data within the local window. To evaluate the smoothing effect, the smoothing and denoising unit introduces mean square error (MSE) as an evaluation metric. Specifically, the MSE value between the smoothed spectral data and the original spectral data is calculated. Where n represents the number of sampling points, and i is the index of the sampling point, i.e., the i-th point. original (x i ) represents the sampled value of the original data, y smoothed (x i The value at that point represents the sampled value after smoothing. If the MSE is less than 0.01, it indicates that the smoothing effect is excellent and the data can be included in the fitting peak division unit; otherwise, reset the window size of the Savitzky-Golay filter and perform smoothing until the MSE meets the requirements.
[0088] Due to the severe overlap of organic functional group characteristics during the catalytic reaction, peak fitting is significantly interfered with. This system utilizes prior knowledge to establish a trough-finding segmentation principle, avoiding the splitting of characteristic peaks into two segments that would affect the fitting results. The location of troughs is identified by calculating the first and second derivatives of the spectral data. The first derivative f′(x) represents the change in the slope of the spectral curve, and the second derivative f′′(x) represents the rate of change of the slope. When the spectral curve changes from a downward to an upward trend, the first derivative changes from negative to positive, and the second derivative is positive at this point. The specific formula is as follows, where the definition of the first derivative is... The definition of the second derivative is The process of identifying trough points involves traversing the spectral data and calculating the first and second derivatives for each point. This is done by determining whether the first derivative changes from negative to positive, i.e., f′(x i-1 )<0 and f′(x)i )>0, and whether the second derivative at the corresponding point is positive, i.e., f″(x) i If the value is greater than 0, the location of the trough points can be determined. After identifying all the trough points, the spectrum is automatically segmented according to the location of the trough points.
[0089] After automatic segmentation, peak fitting is performed on the spectrum to output spectral features, preparing for the feature recognition unit. In the peak fitting unit, the spectral data is precisely fitted and peak-segmented to extract the characteristic peaks in the spectrum. First, based on the symmetry and distribution of the characteristic peaks, the applicability of Gaussian and Lorentz fitting is evaluated using the minimum deviation method. The specific formulas are as follows: The formula for Gaussian fitting is: The formula for Lorentz fitting is: Where a is the peak height, b is the peak position, and c is the peak width. After fitting, the fitting effect is evaluated using two criteria: the coefficient of determination R0. 2 The ratio of peak height to peak height (H). Coefficient of determination (R). 2 The formula used to measure the degree of fit between a fitted curve and actual data is: Among them, y i It is actual data, f i It is fitted data. This is the average of the actual data. The formula for calculating the ratio of peak height values is: Among them, h fitted It is the height of the fitted peak, h original It is the height of the peak in the original signal. For example, when R 2 If the value is greater than 0.98 and H > 0.98, the fitting effect is considered good, and the peak feature parameters are output; otherwise, the shoulder peaks are re-identified and processed. During the refitting process for the shoulder peaks, Gaussian or Lorentz fitting is continued, and R² is recalculated. 2 And H until the optimal values of R2 and H are reached (close to 1).
[0090] The cyclic feedback evaluation unit enables the preprocessing module to self-correct. It intelligently judges the reliability of the preprocessing results by using weighted factors such as the coefficient of determination of peaks. When the weighted factor is lower than the standard, the intelligent preprocessing module will sequentially backtrack to the previous standard processing unit for readjustment. For example, when the coefficient of determination of peaks is always lower than 0.8, the system will readjust the valley finding segmentation process and then perform peak fitting again. If the coefficient of determination still does not meet the requirements, the smoothing and noise reduction process will be readjusted. This cyclic feedback is performed until the fitted peak result meets the weighted factor requirements.
[0091] After intelligent preprocessing and peak fitting, important parameters such as peak position, peak height, half-maximum width, and peak area of each characteristic peak are extracted. Accurate extraction and quantification of these characteristic parameters provide high-quality input data for machine learning species identification, ensuring the accuracy and reliability of species identification and quantitative analysis. Weak signal recognition and reinforcement training employs quadratic difference spectral processing. By subtracting the spectra at different temperatures from the reference spectrum, subtle changes in the spectrum under different temperature conditions are determined.
[0092] 3. Machine Learning Recognition Module
[0093] The machine learning recognition module is a key component of this invention. Its main function is to utilize spectral feature parameters obtained from the feature extraction module to perform classification and recognition through multiple classification models, such as chemical formulas or chemical chains, achieving high-precision spectral recognition and analysis of product variation trends. The feature parameters extracted from the feature extraction module include peak position, peak height, full width at half maximum (FWHM), and peak area. In this invention, to mitigate the limitations of a single model and smooth out its errors, four classification models—Support Vector Machine, Random Forest, Backpropagation Neural Network, and CNN—are trained separately on the training set. By adjusting the parameters of each individual model, each model achieves optimal performance on the training set. Figure 3 As shown.
[0094] During the model training phase, the preprocessed spectral data was divided into training and testing sets in an 8:2 ratio, and each model was trained separately for each set. After each round of training, a 5-fold cross-validation method was used to evaluate the training results, and the accuracy on the validation set was calculated.
[0095] During the model evaluation phase, model performance is assessed using accuracy, calculated using the following formula: Where TP represents the number of samples predicted as positive, TN represents the number of samples predicted as negative, FP represents the number of samples predicted as positive, and FN represents the number of samples predicted as negative. Evaluation results show an accuracy of 94.0% on the random forest model, 99.5% on the backpropagation (BP) neural network, and 98.2% on the CNN model.
[0096] Finally, the classification results from the four models are combined using a voting method to obtain the final classification result. This multi-model result fusion approach improves the robustness of the model in data prediction. Furthermore, the database stores the feature information corresponding to each functional group of the standard data. Therefore, before inputting the data into the model for prediction, the features extracted from the data to be predicted can be matched with the information in the database to achieve rapid identification of the data. If the match with the standard database is low, the data is input into the classification model again for identification. In summary, in the classification and identification module, the data to be predicted is first matched with the database, and then, based on the matching results, classification and identification are performed using multiple models, greatly improving the accuracy of data classification.
[0097] like Figure 4 As shown in Figure 4(a), this is a set of in-situ infrared reaction data of the catalyst. After preprocessing and segmentation, the data from 1900 to 2100 cm⁻¹ were analyzed. -1 Range characteristics ( Figure 4 (b) shows that the positions of the three peaks are obtained by peak fitting. Figure 4 (c) shows that the values are 2047cm. -1 2042cm -1 and 1996cm -1 The intelligent model identifies the products; the positions of these three peaks all belong to the characteristics of linear CO adsorption (Table 1). Based on the characteristics of the in-situ reaction data, information on catalytic reaction intermediates and products is plotted as follows: Figure 4 As shown in (d).
[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions for some or all of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. An intelligent analysis system for in-situ catalytic reaction infrared spectral characterization data, characterized in that, include: The preprocessing module is used to perform real-time dynamic preprocessing on in-situ infrared spectral data. The feature extraction module is used to extract spectral feature parameters from the preprocessed spectral data. The machine learning recognition module is used to identify spectral feature parameters through a machine learning model to obtain the chemical structure information of the catalytic reaction intermediates and products contained therein. The data analysis module is used to obtain the dynamic change trend of in-situ reaction products based on the changes in chemical structure with reaction conditions.
2. The intelligent analysis system for in-situ catalytic reaction infrared spectral characterization data according to claim 1, characterized in that, The preprocessing module includes: 1) Spectral data difference preprocessing module, used to make a standard curve of the infrared spectrum of the sample in the initial state before the catalytic reaction, use the difference between the spectral data during the reaction process and the standard curve, and divide the spectral range into segments based on the data difference subtraction result; 2) Baseline correction module, used to select different correction models for each spectral data according to the segmentation results to perform baseline removal processing and form corrected spectral data; 3) Smoothing and noise reduction module, used to smooth and reduce noise for each corrected spectral data; 4) Valley finding and segmentation module, which is used to identify the location of valley points based on the smoothed and denoised data by calculating the first and second derivatives of the spectral data, and automatically segment the spectrum according to the location of the valley points; 5) Peak fitting module, used to perform peak fitting on each segment of spectral data after segmentation and output peak characteristic parameters; 6) The cyclic feedback evaluation module is used to compare the characteristic differences between the fitted curve and the original infrared spectrum, including peak position, peak area and peak characteristic parameters, to obtain the ratio of peak area and peak height of the fitted curve and the original infrared spectrum, which are used as the determination coefficient and peak height weighting factor, respectively. After the current i-th iteration of fitting the peak, if the determination coefficient and weighting factor are still lower than the threshold, return to step 4) to adjust the segment position in the valley segmentation module. If the weight factor obtained in the (i+1)th iteration is still lower than the threshold, return to step 3) to adjust the filter window size in the smoothing and noise reduction module; If the weight factor obtained in the (i+2)th iteration is still lower than the threshold, return to step 2) to adjust the parameters of different calibration models until the weight factor is not lower than the threshold.
3. The intelligent analysis system for in-situ catalytic reaction infrared spectral characterization data according to claim 1, characterized in that, The baseline correction module includes multiple correction models: A polynomial fitting model is used to fit data using the least squares method to obtain polynomial-fitted data. The segmented baseline fitting model is used to divide the data into several segments and use different fitting methods for each segment to obtain the data after segmented baseline fitting. The moving window smoothing model is used to slide a window of a set length over a data sequence. The statistical value of each data point within the window is used as the smoothing value of the window's center point, thus obtaining the smoothed data. The penalized least squares model is used to minimize the squared error between the predicted and actual values in a regression model, thus obtaining the regressed data.
4. The intelligent analysis system for in-situ catalytic reaction infrared spectral characterization data according to claim 1, characterized in that, The fitting peak division module performs the following steps: 1) Based on the symmetry and distribution of spectral features, Gaussian and Lorentz fitting operations are performed in parallel to evaluate the applicability of Gaussian and Lorentz fitting based on the minimum deviation of the fitting: The formula for Gaussian fitting is: The formula for Lorentz fitting is: Where a is the peak height, b is the peak position, and c is the peak width; y(x) represents the fitted value, and x represents the input spectral wavenumber; 2) Evaluate the fitting effect using two evaluation criteria: For each fitting method, either Gaussian or Lorentz, obtain the coefficient of determination R. 2 The ratio of peak height to peak height, H; Coefficient of determination R 2 Used to measure how well a fitted curve fits the actual data: Among them, y i It is the actual data, that is, the spectral data after wave trough segmentation, g i It is fitted data. It is the average of the actual data; Ratio of peak height values: Among them, h fitted It is the height of the fitted peak, h original It is the height of the peak in the spectral data after valley segmentation; 3) When the coefficient of determination R 2 When the peak height ratio H reaches the set range, the fitting effect is considered good, and the peak feature parameters, including peak height a, peak position b, and peak width c, are output; otherwise, return to step 1) to re-identify and segment the shoulder peaks until the coefficient of determination R is reached. 2 Until the peak height ratio H reaches the set range; 4) Select the coefficient of determination R 2 The fitting method with a smaller peak height ratio H outputs peak characteristic parameters.
5. The intelligent analysis system for in-situ catalytic reaction infrared spectral characterization data according to claim 1, characterized in that, The feature extraction module obtains the peak area based on the peak feature parameters; the peak height, peak position, peak width, and peak area together constitute the spectral feature parameters, which are used for machine learning model recognition.
6. The intelligent analysis system for in-situ catalytic reaction infrared spectral characterization data according to claim 1, characterized in that, The machine learning recognition module includes: Machine learning models, including support vector machines, random forests, backpropagation neural networks, and CNNs, are used to identify the chemical structure information of catalytic reaction intermediates and products contained in each model based on the input spectral feature parameters. The classification results of each model are used to obtain the final classification result, i.e., the chemical structure information of catalytic reaction intermediates and products contained in the model, through a voting method. The model training module is used to pre-train each machine learning model; the input of the model is labeled in-situ infrared spectra, the labels including sample information, temperature information and time information; the output of the model is the chemical structure information of the included catalytic reaction intermediates and products. The model evaluation and optimization module is used to evaluate and optimize machine learning models, and update machine learning models.
7. The intelligent analysis system for in-situ catalytic reaction infrared spectral characterization data according to claim 1, characterized in that, The data analysis module performs the following steps: 1) Save the in-situ infrared spectra of the catalytic reaction line by line to the database and sort them according to the tag information; 2) After analyzing each infrared spectrum, determine the characteristics of organic functional groups under each reaction temperature condition; 3) Based on temperature and time changes, the changes in the content of organic functional groups in the catalytic reaction process are automatically generated, i.e., the dynamic change trend of the in-situ reaction products.
8. A method for intelligent analysis of infrared spectral characterization data of in-situ catalytic reactions, characterized in that, Includes the following steps: For in-situ infrared spectral data, real-time dynamic preprocessing is performed through a preprocessing module; After preprocessing, spectral data are used to extract spectral feature parameters. The spectral feature parameters are identified using a machine learning model to obtain the chemical structure information of the catalytic reaction intermediates and products contained therein. The dynamic change trend of in-situ reaction products can be obtained by measuring the chemical structure under different reaction conditions.
9. The intelligent analysis method for in-situ catalytic reaction infrared spectral characterization data according to claim 8, characterized in that, The in-situ infrared spectral data undergoes real-time dynamic preprocessing via a preprocessing module, including the following steps: 1) The spectral data difference preprocessing module generates a standard curve from the infrared spectrum of the sample in its initial state before the catalytic reaction. It then calculates the difference between the spectral data during the reaction process and the standard curve, and segments the spectral range based on the data difference subtraction result. 2) The baseline correction module selects different correction models for each spectral data point based on the segmentation results to perform baseline removal processing, thereby generating corrected spectral data; 3) The smoothing and noise reduction module performs smoothing and noise reduction on each corrected spectral data; 4) The valley segmentation module identifies the location of valley points based on the smoothed and denoised data by calculating the first and second derivatives of the spectral data, and automatically segments the spectrum according to the location of the valley points. 5) The peak fitting module performs peak fitting on each segment of the spectral data and outputs peak characteristic parameters; 6) The cyclic feedback evaluation module compares the characteristic differences between the fitted curve and the original infrared spectrum, including peak position, peak area, and peak characteristic parameters, and obtains the ratio of peak area and peak height of the fitted curve to the original infrared spectrum, which are used as the determination coefficient and peak height weighting factor, respectively. After the current i-th iteration of fitting the peak, if the determination coefficient and weighting factor are still lower than the threshold, return to step 4) to adjust the segment position in the valley finding segmentation module. If the weight factor obtained in the (i+1)th iteration is still lower than the threshold, return to step 3) to adjust the filter window size in the smoothing and noise reduction module; If the weight factor obtained in the (i+2)th iteration is still lower than the threshold, return to step 2) to adjust the parameters of different calibration models until the weight factor is not lower than the threshold.
10. The intelligent analysis method for in-situ catalytic reaction infrared spectral characterization data according to claim 8, characterized in that, The step of identifying spectral feature parameters using a machine learning model to obtain the chemical structure information of the included catalytic reaction intermediates and products includes the following steps: Machine learning models include support vector machines, random forests, backpropagation neural networks, and CNNs. For each model, the chemical structure information of the catalytic reaction intermediates and products is identified based on the input spectral feature parameters. The classification results of each model are used to obtain the final classification result, i.e., the chemical structure information of the catalytic reaction intermediates and products, through a voting method. The model training module pre-trains each machine learning model; the input to the model is a labeled in-situ infrared spectrum, the labels including sample information, temperature information and time information; the output of the model is the chemical structure information of the included catalytic reaction intermediates and products.