CN-122024968-A - Deep learning-based ruthenium zinc benzene hydrogenation catalyst activity analysis and performance prediction method and system

CN122024968ACN 122024968 ACN122024968 ACN 122024968ACN-122024968-A

Abstract

The invention relates to a method and a system for predicting activity analysis and performance of a ruthenium-zinc-benzene hydrogenation catalyst based on deep learning, wherein the method comprises the steps of automatically preprocessing an HRTEM image and an XRD spectrum of a catalyst sample; the method comprises the steps of extracting deep morphology features of ruthenium nano particles in an image through an encoder, generating a probability distribution map, extracting chemical features representing zinc phases in a spectrum through a one-dimensional convolutional neural network, carrying out deep fusion on the image features and the spectrum features by utilizing a cross attention mechanism, generating unified comprehensive feature representation, and simultaneously outputting analysis results of average size and dispersity of the ruthenium nano particles and performance prediction results of benzene conversion rate and cyclohexene selectivity through parallel regression prediction branches based on the comprehensive features. The invention overcomes the defects that the traditional method depends on manual analysis and cannot correlate multi-mode data with performance, and realizes the end-to-end intelligent correlation and prediction of the microstructure and macroscopic performance of the catalyst.

Inventors

SUN HAIJIE
QIAO LEIYOU
WANG WEIMIN
ZHAO HAITAO
HU MENGQI
WANG SHIHAO

Assignees

郑州师范学院
江西心连心化学工业有限公司

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (10)

1. The method for analyzing the activity and predicting the performance of the ruthenium-zinc-benzene hydrogenation catalyst based on deep learning is characterized by comprising the following steps of: s1, performing image enhancement and noise reduction and spectrum normalization alignment on catalyst sample HRTEM image data obtained through a high-resolution transmission electron microscope and XRD spectrum data of a catalyst sample obtained through an X-ray diffractometer to generate preprocessed image data and standard spectrum data; S2, invoking a deep learning encoder to extract image deep semantic features of ruthenium nano particles in the preprocessed image data, and invoking a decoder to reconstruct a probability distribution map representing the outline of the ruthenium nano particles to generate an image feature sequence containing particle morphology and spatial distribution information; s3, capturing local peak shape and integral spectral line characteristics in the standard spectrum data based on a convolutional neural network, mapping the local peak shape and the integral spectral line characteristics into low-dimensional dense vectors representing crystal structures and zinc phase information, and generating a spectral characteristic vector; S4, dynamically focusing the image features in the image feature sequence on a spatial region most relevant to the current spectrum information in the spectrum feature vector based on an attention mechanism, and integrating the image feature sequence and the spectrum feature vector into a unified feature space to generate a comprehensive feature representation; s5, carrying out parallel regression prediction based on the comprehensive feature representation, calling a branch regression layer to map the comprehensive feature representation so as to predict the average size and the dispersity of ruthenium nano particles, and mapping the comprehensive feature representation through the other branch regression layer so as to predict the conversion rate of benzene hydrogenation reaction and the cyclohexene selectivity, so as to generate a comprehensive prediction result, wherein the comprehensive prediction result is used for indicating the microstructure state and the macroscopic catalytic performance of the catalyst sample.
2. The method according to claim 1, wherein S1 comprises: S11, performing contrast-limited self-adaptive histogram equalization processing on catalyst sample HRTEM image data obtained through a high-resolution transmission electron microscope, enhancing gray level difference between ruthenium nano particles and a carrier background, and generating a contrast enhanced image; s12, carrying out non-local mean value noise reduction treatment on the contrast enhanced image, inhibiting Gaussian noise and speckle noise in the image, and simultaneously keeping particle edge details to generate the preprocessed image data; S13, performing minimum-maximum normalization processing on XRD spectrum data of a catalyst sample obtained by an X-ray diffractometer, and scaling an original intensity value to a uniform numerical interval to generate a normalized spectrum; And S14, carrying out dynamic time-warping-based spectral peak alignment on the normalized spectrum, carrying out matching calibration on diffraction angle positions of the normalized spectrum and a standard substance spectrum, and generating standard spectrum data.
3. The method according to claim 1, wherein S2 comprises: S21, carrying out multistage convolution and pooling downsampling on the preprocessed image data based on an encoder-decoder network architecture, and calling a depth residual error network to extract shallow geometric features containing basic edges and textures and high-level abstract features containing particle integral shapes and semantic categories from the image to generate a deep feature map containing multi-level semantic information; s22, carrying out progressive up-sampling and feature stitching treatment on the deep feature map, and carrying out step-by-step fusion and resolution recovery on shallow detail features which are from an early stage of a coder path and contain particle edges and textures and high-level semantic features which are from the deep feature map and contain particle overall shape and category information through jump connection to generate a high-resolution feature map; S23, carrying out pixel-by-pixel classification processing on the high-resolution feature map, calculating and outputting probability values of each pixel belonging to ruthenium nano particles through a convolution layer and a nonlinear activation function, and generating a binary probability distribution map representing the outline of the ruthenium nano particles; And S24, carrying out feature screening and serialization processing based on the high-resolution feature map and the probability distribution map, carrying out attention weighting on the high-resolution feature map by taking the probability distribution map as a space weight, flattening the weighted features along the space dimension, and generating an image feature sequence focused on the foreground particle region.
4. The method according to claim 1, wherein S3 comprises: S31, carrying out multi-scale one-dimensional convolution filtering processing on the standard spectrum data, and respectively capturing sharp diffraction peak characteristics, medium-width diffraction peak characteristics and broad diffraction envelope characteristics in a spectrum by using three one-dimensional convolution kernels with widths of 3, 5 and 7 to generate a multi-scale spectrum characteristic diagram; S32, carrying out feature fusion and pooling treatment on the multi-scale spectrum feature map, splicing the sharp diffraction peak features, the medium-width diffraction peak features and the broad diffraction envelope features through channel dimensions, and carrying out global average pooling polymerization global spectrum information on the spliced feature map to generate a fusion spectrum feature vector; And S33, performing nonlinear transformation and dimension reduction processing on the fusion spectral feature vector, mapping the fusion spectral feature vector to a latent space with obviously reduced dimension through a full connection layer and a ReLU activation function, and generating a low-dimension dense vector representing crystal structure and zinc phase information as the spectral feature vector.
5. The method according to claim 1, wherein S4 comprises: S41, performing linear projection processing on the spectrum feature vectors, and respectively mapping the spectrum feature vectors to a key feature space for calculating similarity and a value feature space for transmitting information based on a linear transformation layer in a cross attention mechanism to generate key vectors and value vectors; S42, carrying out similarity calculation processing on the image feature sequence and the key vector, measuring the correlation between the feature of each spatial position in the image feature sequence and the key vector through dot product operation, and generating attention weight distribution; S43, carrying out weighted aggregation processing on the value vector based on the attention weight distribution, and injecting the spectrum information carried by the value vector into the corresponding image feature space position according to the intensity of the attention weight distribution to generate an enhanced image feature sequence after modal interaction; And S44, carrying out global aggregation processing on the enhanced image feature sequence, carrying out averaging and pooling on the features at all positions along the space dimension, compressing the space dimension of the feature sequence, and generating a comprehensive feature representation.
6. The method of claim 5, wherein the attention weight distribution is calculated by the formula: Wherein, the In order to pay attention to the weight distribution, For the image feature sequence as a query, K is a key matrix formed by a key vector copy extension, For the feature dimension of the key vector, T represents the matrix transpose and softmax is the normalized exponential function.
7. The method according to any one of claims 1-6, wherein S5 comprises: S51, performing model training treatment on a training data set formed by a plurality of catalyst samples with known real active phase parameters and benzene hydrogenation performance, optimizing parameters of a preset multichannel deep learning model through a gradient descent algorithm based on HRTEM images, XRD spectrum data and corresponding real measurement values of average size, dispersity, benzene conversion rate and cyclohexene selectivity of ruthenium nanoparticles in the training data set so as to establish a nonlinear mapping relation from the multi-mode data to a target label, and generating a pre-trained multichannel deep learning model; S52, carrying out first branch regression processing on the comprehensive feature representation, inputting the comprehensive feature representation into a first full-connection regression layer corresponding to an active phase parameter prediction task in the pre-trained multichannel deep learning model, mapping the comprehensive feature representation into a predicted ruthenium nanoparticle average size and dispersity score through the weight and bias parameters learned by the first full-connection regression layer, and generating an active phase analysis result; and S53, carrying out second branch regression processing on the comprehensive characteristic representation, inputting the comprehensive characteristic representation into a second full-connection regression layer corresponding to a catalytic performance prediction task in the pre-trained multichannel deep learning model, and mapping the comprehensive characteristic representation into a predicted benzene conversion percentage and cyclohexene selectivity percentage through the weight and bias parameters learned by the second full-connection regression layer to generate a benzene hydrogenation performance prediction result.
8. A deep learning-based ruthenium zinc benzene hydrogenation catalyst activity analysis and performance prediction system, the system comprising: The spectrogram preprocessing module is used for carrying out image enhancement noise reduction and spectrum normalization alignment on the HRTEM image data of the catalyst sample obtained by the high-resolution transmission electron microscope and the XRD spectrum data of the catalyst sample obtained by the X-ray diffractometer to generate preprocessed image data and standard spectrum data; The image feature extraction module is used for calling a deep learning encoder to extract image deep semantic features of ruthenium nano particles in the preprocessed image data, and calling a decoder to reconstruct a probability distribution map representing the outline of the ruthenium nano particles so as to generate an image feature sequence containing particle morphology and spatial distribution information; the spectrum characteristic mapping module is used for capturing local peak shape and integral spectral line characteristics in the standard spectrum data based on a convolutional neural network, mapping the local peak shape and the integral spectral line characteristics into low-dimensional dense vectors representing crystal structures and zinc phase information, and generating a spectrum characteristic vector; The feature fusion coding module is used for dynamically focusing the image features in the image feature sequence on a spatial region most relevant to the current spectrum information in the spectrum feature vector based on an attention mechanism, and integrating the image feature sequence and the spectrum feature vector into a unified feature space to generate a comprehensive feature representation; And the catalytic performance prediction module is used for carrying out parallel regression prediction based on the comprehensive characteristic representation, calling one branch regression layer to map the comprehensive characteristic representation so as to predict the average size and the dispersity of the ruthenium nano particles, and mapping the comprehensive characteristic representation through the other branch regression layer so as to predict the conversion rate of benzene hydrogenation reaction and the cyclohexene selectivity, so as to generate a comprehensive prediction result, wherein the comprehensive prediction result is used for indicating the microstructure state and the macroscopic catalytic performance of the catalyst sample.
9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the method of any one of claims 1 to 7 when executing the computer program.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the method of any of claims 1 to 7.

Description

Deep learning-based ruthenium zinc benzene hydrogenation catalyst activity analysis and performance prediction method and system Technical Field The invention relates to the technical field of industrial catalysis and artificial intelligence intersection, in particular to a method and a system for activity analysis and performance prediction of a ruthenium zinc benzene hydrogenation catalyst based on deep learning. Background The selective hydrogenation of benzene to prepare cyclohexene is an important chemical process, and the target product cyclohexene is a key intermediate for producing chemicals such as nylon and the like. The core of the reaction is to develop a high-performance catalyst, wherein a supported catalyst system with noble metal ruthenium as an active component and zinc as an auxiliary agent becomes a research hot spot due to higher selectivity to cyclohexene. The performance of the catalyst, in particular its activity and selectivity, is fundamentally dependent on the microstructure (e.g. size, dispersed state) of the active phase ruthenium nanoparticles and the chemical state of the promoter zinc. At present, performance analysis and structural analysis of the catalyst mainly depend on physical characterization technologies such as a high-resolution transmission electron microscope (HRTEM) and X-ray diffraction (XRD) to acquire microscopic morphology and crystal structure information, and then the macroscopic catalytic performance of the catalyst is evaluated by combining experimental tests. However, the prior art methods have significant limitations. On the one hand, identification, size measurement and dispersity evaluation of ruthenium nanoparticles in an HRTEM image are highly dependent on experience of researchers to perform manual interpretation or simple image processing based on fixed rules, so that the processing efficiency is low, the adaptability to irregular particles is poor, and the subjectivity and reproducibility of results are high. On the other hand, the key chemical information about zinc phases, lattice strains and the like provided by XRD spectra and the geometric shape information provided by HRTEM images are in a fracture state, and the traditional analysis method is difficult to organically combine the two types of heterogeneous data, so that the internal correlation between the ruthenium particle structure, the zinc chemical state and the macroscopic catalytic performance cannot be quantitatively disclosed. This allows the catalyst development to remain in the "trial and error" mode, i.e., the formulation is screened through a large number of repeated experiments, with long cycle times, high costs, and difficulty in understanding the mechanism of action of the adjuvant. Disclosure of Invention Based on the above, the invention aims to provide a method and a system for analyzing the activity and predicting the performance of a ruthenium-zinc-benzene hydrogenation catalyst based on deep learning, which can automatically, accurately and multi-dimensionally analyze the catalyst structure and can directly and quantitatively predict the catalytic performance of the catalyst. The invention adopts the following scheme: In a first aspect, the invention provides a method for analyzing the activity and predicting the performance of a ruthenium-zinc-benzene hydrogenation catalyst based on deep learning, which comprises the following steps: s1, performing image enhancement and noise reduction and spectrum normalization alignment on catalyst sample HRTEM image data obtained through a high-resolution transmission electron microscope and XRD spectrum data of a catalyst sample obtained through an X-ray diffractometer to generate preprocessed image data and standard spectrum data; s2, invoking a deep learning encoder to extract image deep semantic features of ruthenium nano particles in the preprocessed image data, and invoking a decoder to reconstruct a probability distribution map representing the outline of the ruthenium nano particles to generate an image feature sequence containing particle morphology and spatial distribution information; s3, capturing local peak shape and integral spectral line characteristics in standard spectral data based on a convolutional neural network, mapping the local peak shape and the integral spectral line characteristics into low-dimensional dense vectors representing crystal structures and zinc phase information, and generating spectral characteristic vectors; S4, dynamically focusing the image features in the image feature sequence on a spatial region most relevant to the current spectrum information in the spectrum feature vector based on an attention mechanism, and integrating the image feature sequence and the spectrum feature vector into a unified feature space to generate a comprehensive feature representation; And S5, carrying out parallel regression prediction based on the comprehensive feature representation, calling a branch regression