A chemical reaction pattern recognition method based on infrared spectroscopy and reaction calorimetry

Abstract

This invention discloses a chemical reaction pattern recognition method based on the combined use of infrared spectroscopy and reaction calorimetry. This invention establishes kinetic models for different reaction types based on reaction kinetics principles, improving pattern recognition accuracy and model interpretability while generating massive amounts of infrared spectral and reaction calorimetric data. This avoids experimental costs and time expenditure, providing effective data support for the study of reaction kinetics principles. This invention sets up independent encoders to extract key features for each modality, and further employs a Transformer architecture to achieve cross-modal collaborative perception. Based on deep feature fusion, modal task weight vectors and modal input confidence vectors are generated simultaneously. From the perspectives of the global task and the current sample, the contribution of each modality and the real-time reliability of the input data are evaluated, achieving accurate identification and classification of chemical reaction patterns.

Description

Technical Field

[0001] This invention belongs to the field of kinetic analysis, specifically relating to a chemical reaction pattern recognition method based on infrared spectroscopy and reaction calorimetry. Background Technology

[0002] Currently, the most widely used online reaction process analysis techniques for kinetic analysis of different types of chemical reactions include online infrared spectroscopy and automated reaction calorimetry.

[0003] The principle of online infrared spectroscopy is based on the selective absorption of infrared light by molecules at specific wavelengths, and the infrared absorption spectrum of a substance is obtained by the change in transmitted light intensity. Its limitations are: (1) transient intermediates or rapid side reactions may not be captured; (2) it cannot capture key safety parameters such as reaction exothermic rate and adiabatic temperature rise, resulting in the thermal runaway risk of highly exothermic reactions being ignored; (3) isomers or those with different spatial conformations may have similar responses in infrared spectra, but their thermal stability may be vastly different.

[0004] Reaction calorimeters can calculate heat by measuring the heat exchange rate between the reaction system and the environment. Its limitations are: (1) the same exothermic curve may correspond to different reaction paths; (2) calorimetric data cannot verify the transformation of functional groups; (3) the generation of low-concentration highly toxic substances may be accompanied by weak thermal signals, which are easily masked by system noise.

[0005] Meanwhile, existing methods require a large amount of infrared spectral data and reaction calorimetric data of different reaction types when training multimodal neural network models, which is difficult to achieve easily in terms of time, cost and data completeness. In addition, multiple variables need to be considered during the experiment.

[0006] The significance of multimodal neural networks lies in their ability to fuse and process information from different perceptual modalities, thereby improving the model's understanding and analytical accuracy. However, there is a strong correlation between the molecular bond vibration characteristics in infrared spectroscopy and the energy release dynamics in reaction calorimetry data, making it difficult for a single-modal model to fully capture the reaction mechanism. Summary of the Invention

[0007] To overcome the problems of large bias from a single data source, incomplete dynamic process modeling, low distinguishability of similar reaction patterns, and poor interpretability of training models in traditional reaction monitoring methods, this invention provides a chemical reaction pattern recognition method based on infrared spectroscopy and reaction calorimetry.

[0008] This invention performs kinetic analysis on different types of chemical reactions, simulates and generates infrared spectral databases and reaction calorimetric databases; constructs a single-modal classification model and a multi-modal fusion model, and synchronously generates modal task weight vectors and modal input confidence vectors based on fusion features, completing the meta-fusion of the first and second decisions, and realizing the automatic identification and classification of chemical reaction modes.

[0009] To achieve the above objectives, the present invention provides a chemical reaction pattern recognition method based on infrared spectroscopy and reaction calorimetry, comprising the following steps:

[0010] Step 1: Establish kinetic models for different reaction types, simulate and generate infrared spectral data and reaction calorimetric data, and load and preprocess them;

[0011] Step 2: Divide the preprocessed data and perform standardization;

[0012] Step 3: Establish a single-modal classification model and obtain single-modal prediction results; at the same time, establish a multimodal fusion model and obtain multimodal deep fusion features;

[0013] Step 4: Based on the fusion features, simultaneously generate the modal task weight vector and the modal input confidence vector;

[0014] Step 5: Perform a first weighted fusion on the single-modal prediction results using the task weight vector to obtain a first decision; perform a second weighted fusion using the input confidence vector to obtain a second decision;

[0015] Step 6: Perform meta-fusion on the first decision and the second decision, and output the final classification result.

[0016] Based on the above technical content, compared with the prior art, the present invention has the following beneficial effects:

[0017] This invention establishes kinetic models for different reaction types based on reaction kinetics principles. While improving pattern recognition accuracy and model interpretability, it generates a large volume of infrared spectral and calorimetric data, avoiding experimental costs and time expenditure, and providing effective data support for the study of reaction kinetics principles. Simultaneously, appropriate random noise is added to the simulation data to mimic data obtained from real experiments.

[0018] This invention employs independent encoders to extract key features from each modality, and further utilizes a Transformer architecture to achieve cross-modal collaborative perception. Based on deep fusion features, modal task weight vectors and modal input confidence vectors are generated simultaneously. From the perspectives of the global task and the current sample, the contribution of each modality and the real-time reliability of each modal input data are evaluated, thereby achieving accurate identification and classification of chemical reaction patterns.

[0019] Compared to traditional single-data models, this invention enables the combined use of multimodal data and splits the traditional fuzzy fusion weights into two parts: task weights and input confidence. This allows for a direct observation of the contribution of sample data to model training and data quality, effectively improving prediction accuracy and making the decision-making process transparent and interpretable. Attached Figure Description

[0020] Figure 1 This is a flowchart of a chemical reaction pattern recognition method based on infrared spectroscopy and reaction calorimetry according to an embodiment of this application;

[0021] Figure 2 This is an example of an infrared spectrum generated through simulation in an embodiment of this application;

[0022] Figure 3 This is an example of a reaction calorimetry curve generated by simulation in an embodiment of this application;

[0023] Figure 4 This is a core structure diagram of the neural network model in an embodiment of this application;

[0024] Figure 5 This is a flowchart of step S11 in an embodiment of this application. Detailed Implementation

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] like Figure 1 As shown in the figure, this application provides a chemical reaction pattern recognition method based on infrared spectroscopy and reaction calorimetry, which specifically includes the following steps:

[0027] S1. Based on the principles of reaction kinetics, establish kinetic models for n different reaction types, and simulate and generate infrared spectral data and reaction calorimetric data for each of the n reaction types. Specifically:

[0028] Based on the Lorentz equation, the following settings were applied: the number of random absorption peaks ranged from 3 to 10, the maximum absorption value ranged from 0.0001 to 0.8, and the wavenumber corresponding to the maximum absorption value ranged from 600 to 1800 cm⁻¹. -1 The full width at half maximum (FWHM) of the absorption peak ranges from 1 to 100 cm⁻¹. -1 .

[0029] Based on the established reaction kinetic model, the following settings were used: the initial reactant concentration ranged from 0.5 to 1.5 mol / L, the reaction rate constant ranged from 0.0001 to 0.01, the reaction order ranged from 1 to 2, and the enthalpy ranged from -10 to -200 kJ / mol. The initial reactant concentration, enthalpy, and rate constant k were all taken from a triangular distribution, while other parameters were taken from a uniform distribution.

[0030] Simulate and generate 4×10 5 The results obtained from the infrared spectral data and reaction calorimetric data are as follows: Figures 2-3 As shown in the figure. Among them, the infrared spectral data is a t×λ dimensional matrix, with row vectors corresponding to the time series and column vectors corresponding to the wavelength dimension; the reaction calorimetric data is a 1×t dimensional matrix, which characterizes the dynamic changes of heat flux during the reaction process.

[0031] S2. Based on the generated data, add random noise to the data. Specifically:

[0032] Based on the generated multimodal data, random Gaussian noise is added. The standard deviation of the noise error is 1% of the maximum value of the spectral data and 0.1% of the maximum value of the calorimetric data. The mean error of both is 0, so as to simulate the spectral and calorimetric data obtained in the experiment under real conditions.

[0033] S3. Load and preprocess the two types of data, including:

[0034] The two types of data are loaded and preprocessed. The category directory is traversed, and the spectral data directory and calorimetric data directory are accessed respectively. The two types of modal data are paired to achieve modality alignment. The spectral file path, calorimetric file path and category label are combined into a tuple and added to the metadata list to ensure that the spectral data and calorimetric data in each training sample come from the same observation object.

[0035] To address the issue of asynchronous experimental data, this application's embodiments compare the data dimension with the model's defined dimension: when the data dimension is smaller than the model's defined dimension, the multimodal data dimension is supplemented by linear interpolation; when the data dimension is larger than the model's defined dimension, the multimodal data dimension is aligned by sampling at equal intervals.

[0036] S4. Divide the preprocessed data into training, validation, and test sets, and then perform standardization. Specifically:

[0037] The preprocessed data samples were divided into training, validation, and test sets in an 8:1:1 ratio, using stratified sampling to ensure that the sample proportions for each category were consistent with the original dataset. The mean and standard deviation of the training set were calculated, and the results were used to standardize the validation and test set data.

[0038] S5, such as Figure 4 As shown, a single-modal classification model is established to obtain single-modal prediction results; at the same time, a multi-modal fusion model is established to obtain multi-modal deep fusion features.

[0039] Furthermore, the single-modal classifier model employs a residual neural network to process spectral data and a multilayer perceptron to process calorimetric data to obtain single-modal prediction results. These prediction results enable mid-term predictions and are also used for task weight fusion and data confidence fusion to achieve first and second decisions. The specific working method of this model is as follows:

[0040] For infrared spectral data, a large convolutional kernel is first used to capture a wide range of spatial features, and max pooling is used for downsampling to increase the receptive field. Secondly, multiple residual blocks are used to extract key features. Finally, global average pooling and fully connected layers are used to realize feature projection and result output. Regularization is used to prevent overfitting and increase generalization ability.

[0041] For reaction calorimetry data, multiple fully connected layers are used to reduce the data dimensionality, and regularization is also introduced to prevent overfitting.

[0042] Furthermore, the multimodal fusion model works as follows:

[0043] For infrared spectral data, an improved residual network is used for preliminary feature extraction; position encoding is superimposed, and a single-layer Transformer encoder is used to achieve intra-modal interaction. For reaction calorimetry data, due to its low complexity, a fully connected network is used to extract high-dimensional feature representations while preserving the time series, and a single-layer Transformer encoder is also used for intra-modal interaction.

[0044] Next, multimodal feature fusion is performed: first, features are concatenated along the last dimension of the tensor, preserving the influence of the time dimension; second, features are projected using several fully connected layers, and after adding positional encoding, they are input into the Transformer model for deep fusion to obtain deep fused features.

[0045] The Transformer model employs a multi-head attention mechanism for relevance weighting, and its structure uses only the encoder part, enabling classification tasks. In this embodiment, the number of Transformer heads is set to 8, dropout is set to 0.1, and the number of encoder layers is 6.

[0046] Furthermore, because traditional residual network architectures have fixed convolutional kernel sizes, they cannot achieve the desired results when processing specific data. This application's embodiments employ a spectral encoder architecture, which improves upon the traditional residual architecture to maintain the continuity of the time series when processing infrared spectral data. Its specific workflow is as follows:

[0047] In the first stage, feature compression is performed in the spectral dimension by using convolutional layers for downsampling and using larger convolutional kernels to capture features over a wider spectral range. As the level of feature abstraction increases, the size of the convolutional kernel is gradually reduced, while the size of the convolutional kernel in the residual block is modified to 1×3 to maintain the stability of the feature dimension and preserve the continuity of the time series.

[0048] In the second stage, feature extraction is performed simultaneously in the time and spectral dimensions. The convolution kernel size of the residual blocks is set to 3×3, and complex temporal variations are learned by connecting 3 residual blocks.

[0049] In the third stage, the output dimensions are adjusted to facilitate feature fusion.

[0050] It is worth noting that the GELU activation function is also used in the spectral encoder in this embodiment of the application, and a regularization operation is added after each convolutional layer, which has better gradient smoothing than ReLU and prevents model overfitting.

[0051] S6. The two-factor generation module, based on the fusion features, simultaneously generates a modal task weight vector and a modal input confidence vector. Specifically:

[0052] The fused features are input into two parallel fully connected layers. The first fully connected layer is followed by a Softmax activation function, and the output modality task weight vector is W=[w1,w2], where The second fully connected layer is followed by a Sigmoid activation function, which outputs a modality input confidence vector C=[c1,c2], where... .

[0053] S7. The single-modal prediction results are weighted and fused using the task weight vector to obtain a first decision. Specifically:

[0054] Obtain the single-mode prediction results, use each element in the modal task weight vector W as the weighting coefficient of the single-mode prediction results of the corresponding mode, and perform a weighted linear combination of the single-mode prediction results of all modes to generate the first decision.

[0055] The weighted linear combination is performed according to the following formula:

[0056]

[0057] Where N is the number of modes, wi is the weight value of the i-th mode in the modal task weight vector, Pi is the single-mode prediction result of the i-th mode, and P weight It is the first decision.

[0058] S8. The single-modal prediction results are weighted and fused using the input confidence vector to obtain a second decision.

[0059] Obtain the single-mode prediction results, input the mode confidence vector C as the confidence coefficient of the single-mode prediction results of the corresponding mode, and perform weighted normalization fusion on all single-mode prediction results based on the confidence coefficients of each mode to generate the second decision.

[0060] The weighted normalization fusion is performed according to the following formula:

[0061]

[0062] Where N is the number of modes, ci is the confidence coefficient of the i-th mode in the modal input confidence vector, Pi is the single-mode prediction result of the i-th mode, and P confidence As a second decision, It is a very small positive constant that prevents division by zero.

[0063] S9. Perform meta-fusion on the first decision and the second decision, and output the final classification result.

[0064] Set up a context extractor to obtain a context vector from the deep fusion feature (the context vector is selected from at least one of the deep fusion feature, modal task weight vector and modal input confidence vector), input the context vector to the meta-fusion weight predictor, and output the meta-fusion weight coefficient β, where 0≤β≤1.

[0065] Based on the aforementioned meta-fusion weight coefficient β, the first decision P weight Second decision P confidence Weighting is performed to generate the final classification result P. final ,in:

[0066]

[0067] S10. Train the model and obtain the optimal model parameters. Specifically:

[0068] Before training begins, the hyperparameters for the model training process are set, and the weights are initialized. Multimodal data is input into the network, and the predicted output is obtained through calculation. The predicted results are compared with the true labels, and the error is calculated using a loss function.

[0069] The error is propagated back from the output layer to the input layer, the gradients of each parameter are calculated, and the weight parameters are updated using an optimization algorithm. An early stopping mechanism is added to stop training prematurely when performance no longer improves.

[0070] Finally, determine if the model has reached the upper limit of the number of iterations. If it has, end the training normally and save the optimal parameters of the model.

[0071] S11, such as Figure 5 As shown, the model file is read, the optimal parameters of the model are restored, and the test set data is input. According to steps S3-S9, chemical reaction pattern recognition based on infrared spectroscopy and reaction calorimetry is realized.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A chemical reaction pattern recognition method based on infrared spectroscopy and reaction calorimetry, characterized in that, Includes the following steps: Step 1: Establish kinetic models for different reaction types, simulate and generate infrared spectral data and reaction calorimetric data, and load and preprocess them; Step 2: Divide the preprocessed data and perform standardization; Step 3: Establish a single-modal classification model and obtain single-modal prediction results; at the same time, establish a multimodal fusion model and obtain multimodal deep fusion features; Step 4: Based on the fusion features, simultaneously generate the modal task weight vector and the modal input confidence vector; Step 5: Perform a first weighted fusion on the single-modal prediction results using the task weight vector to obtain a first decision; perform a second weighted fusion using the input confidence vector to obtain a second decision; Step 6: Perform meta-fusion on the first decision and the second decision, and output the final classification result; The working method of the single-modal classification model includes: For spectral data, convolutional kernels are used to capture large-scale spatial features, and max pooling is used for downsampling to increase the receptive field. Multiple residual blocks are used to extract key features. Global average pooling and fully connected layers are used to realize feature projection and result output. Regularization is used to prevent overfitting and increase generalization ability. For calorimetric data, multiple fully connected layers are used to reduce data dimensionality, and regularization is used to prevent overfitting. The working method of the multimodal fusion model includes: For spectral data, an improved spectral encoder is used for preliminary feature extraction; for overlay position encoding, a single-layer Transformer encoder is used to achieve intramodal interaction. For calorimetric data, a fully connected network is used to extract high-dimensional feature representations while preserving the time series, and a single-layer Transformer encoder is used for intramodal interaction. Features are concatenated along the last dimension of the tensor, preserving the influence of the time dimension; features are projected using a fully connected layer and positional encoding is added; the input is fed into the Transformer model, and a multi-head attention mechanism is used for relevance weighting.

2. The method according to claim 1, characterized in that, Step 1 includes: Step 1.1: Establish kinetic models for different reaction types based on the principles of reaction kinetics, and simulate and generate infrared spectral data and reaction calorimetric data corresponding to each reaction type; Step 1.2: Based on the generated data, add random noise to the data to simulate the data obtained in the experiment under real conditions; Step 1.3: Load and preprocess the processed data.

3. The method according to claim 2, characterized in that, Step 1.3 includes: The infrared spectral data and reaction calorimetric data are loaded, the category directory is traversed, the spectral data directory and the calorimetric data directory are accessed respectively, and the two types of modal data are paired to achieve modal alignment. The spectral file path, calorimetric file path, and category label are grouped into a tuple and added to the metadata list to ensure that the spectral and calorimetric data in each training sample come from the same observation object.

4. The method according to claim 3, characterized in that, Step 1.3 also includes: To address the issue of asynchronous experimental data, the data dimensions are compared with the dimensions defined in the model for assessment. When the data dimension is smaller than the model-defined dimension, the multimodal data dimension is completed by linear interpolation; when the data dimension is larger than the model-defined dimension, the multimodal data dimension is aligned by equal-interval sampling.

5. The method according to claim 1, characterized in that, The improved spectral encoder operates as follows: In the first stage, feature compression is performed in the spectral dimension by using convolutional layers for downsampling to capture spectral range features. As the level of feature abstraction increases, the size of the convolutional kernel is gradually reduced, while the size of the convolutional kernel of the residual block is modified to 1×3 to maintain the stability of the feature dimension and the continuity of the temporal sequence. In the second stage, feature extraction is performed simultaneously in the time and spectral dimensions. The convolution kernel size of the residual block is modified to 3×3, and three residual blocks are connected to learn complex temporal changes. In the third stage, the output dimensions are adjusted to facilitate feature fusion.

6. The method according to claim 1, characterized in that, Step 5 specifically involves: Obtain the single-modal prediction results, use each element in the modal task weight vector as the weighting coefficient of the corresponding modal prediction results, and perform a weighted linear combination of the single-modal prediction results of all modalities to generate the first decision; The modal input confidence vector is used as the confidence coefficient of the corresponding modal prediction result. The single-modal prediction results of all modalities are weighted, normalized, and fused to generate the second decision.

7. The method according to claim 6, characterized in that, Step 6 specifically involves: Set up a context extractor, obtain the context vector and input it into the meta-fusion weight predictor, and output the meta-fusion weight coefficients; Based on the aforementioned meta-fusion weight coefficients, the first decision and the second decision are weighted to generate the final classification result.

8. The method according to claim 7, characterized in that, The context vector is selected from at least one of the fusion features, modal task weight vector, and modal input confidence vector.

Images