Ruthenium-zinc benzene hydrogenation catalyst activity analysis and performance prediction method and system based on deep learning

Abstract

The invention relates to a deep learning-based activity analysis and performance prediction method and system for a ruthenium-zinc benzene hydrogenation catalyst. The method comprises the following steps: automatically preprocessing an HRTEM image and an XRD spectrum of a catalyst sample; deep morphological characteristics of ruthenium nanoparticles in the image are extracted through an encoder, a probability distribution diagram is generated, and chemical characteristics representing a zinc phase in a spectrum are extracted through a one-dimensional convolutional neural network; performing deep fusion on the image features and the spectral features by using a cross attention mechanism to generate unified comprehensive feature representation; based on the comprehensive characteristics, through parallel regression prediction branches, analysis results of the average size and the dispersity of the ruthenium nanoparticles and performance prediction results of the benzene conversion rate and the cyclohexene selectivity are output at the same time. The method overcomes the defects that a traditional method depends on manual analysis, and multi-modal data and performance cannot be associated, and end-to-end intelligent association and prediction of the microstructure and the macroscopic performance of the catalyst are achieved.

Description

Technical Field

[0001] This invention relates to the field of industrial catalysis and artificial intelligence, specifically to a method and system for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst based on deep learning. Background Technology

[0002] The selective hydrogenation of benzene to cyclohexene is an important chemical process, and the target product, cyclohexene, is a key intermediate in the production of chemicals such as nylon. The core of this reaction lies in developing high-performance catalysts. Among these, supported catalyst systems using ruthenium as the active component and zinc as a promoter have become a research hotspot due to their high selectivity for cyclohexene. The performance of the catalyst, especially its activity and selectivity, fundamentally depends on the microstructure (such as size and dispersion state) of the active phase ruthenium nanoparticles and the chemical state of the promoter zinc. Currently, the performance analysis and structural determination of such catalysts mainly rely on physical characterization techniques such as high-resolution transmission electron microscopy (HRTEM) and X-ray diffraction (XRD) to obtain information on their microstructure and crystal structure, which is then combined with experimental testing to evaluate their macroscopic catalytic performance.

[0003] However, existing techniques have significant limitations. On the one hand, the identification, size measurement, and dispersion assessment of ruthenium nanoparticles in HRTEM images heavily rely on researchers' experience for manual interpretation or simple image processing based on fixed rules. This results in low processing efficiency, poor adaptability to irregular particles, and highly subjective and reproducible results. On the other hand, the key chemical information provided by XRD spectroscopy regarding zinc phases and lattice strain is disconnected from the geometric morphology information provided by HRTEM images. Traditional analytical methods struggle to organically combine these two heterogeneous data sets, thus failing to quantitatively reveal the intrinsic relationship between "ruthenium particle structure, zinc chemical state, and macroscopic catalytic performance." This keeps catalyst development in a "trial and error" mode, requiring extensive repeated experiments to screen formulations—a process that is time-consuming, costly, and hinders a deep understanding of the mechanism of action of promoters. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide a method and system for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst based on deep learning, which can automatically, accurately, and multidimensionally analyze the catalyst structure and directly and quantitatively predict its catalytic performance.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst based on deep learning, comprising the following steps:

[0007] S1: Perform image enhancement, noise reduction, and spectral normalization alignment on the HRTEM image data of the catalyst samples obtained by high-resolution transmission electron microscopy and the XRD spectral data of the catalyst samples obtained by X-ray diffraction to generate preprocessed image data and standard spectral data.

[0008] S2: Call the deep learning encoder to extract the deep semantic features of ruthenium nanoparticles in the preprocessed image data, and call the decoder to reconstruct the probability distribution map representing the outline of ruthenium nanoparticles, generating an image feature sequence containing particle morphology and spatial distribution information;

[0009] S3: Based on the convolutional neural network, capture the local peak shape and global spectral line features in the standard spectral data, and map the local peak shape and global spectral line features into low-dimensional dense vectors that characterize the crystal structure and zinc phase information to generate spectral feature vectors;

[0010] S4: Based on the attention mechanism, the image features in the image feature sequence are dynamically focused on the spatial region most relevant to the current spectral information in the spectral feature vector, and the image feature sequence and spectral feature vector are integrated into a unified feature space to generate a comprehensive feature representation;

[0011] S5: Parallel regression prediction is performed based on the comprehensive feature representation. One branch regression layer is called to map the comprehensive feature representation to predict the average size and dispersion of ruthenium nanoparticles. Another branch regression layer is used to map the comprehensive feature representation to predict the conversion rate and cyclohexene selectivity of the benzene hydrogenation reaction, generating comprehensive prediction results. The comprehensive prediction results are used to indicate the microstructure state and macroscopic catalytic performance of the catalyst sample.

[0012] In one embodiment, S1 of the deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst provided by the present invention specifically includes the following steps:

[0013] S11: Perform contrast-limited adaptive histogram equalization on the HRTEM image data of the catalyst sample obtained by high-resolution transmission electron microscopy to enhance the gray-level difference between ruthenium nanoparticles and the support background, and generate a contrast-enhanced image.

[0014] S12: Perform nonlocal mean denoising on the contrast-enhanced image to suppress Gaussian noise and speckle noise in the image while preserving grain edge details, and generate preprocessed image data.

[0015] S13: Perform minimum-maximum normalization on the XRD spectral data of the catalyst sample obtained by X-ray diffractometer to scale the original intensity values ​​to a uniform numerical range and generate normalized spectra.

[0016] S14: Perform peak alignment based on dynamic time warping on the normalized spectrum, match and calibrate the diffraction angle positions of the normalized spectrum and the standard substance spectrum, and generate standard spectral data.

[0017] In one embodiment, S2 of the deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst provided by the present invention specifically includes the following steps:

[0018] S21: Based on the encoder-decoder network architecture, multi-level convolution and pooling downsampling are performed on the preprocessed image data. The deep residual network is called to extract shallow geometric features containing basic edges and textures and high-level abstract features containing the overall shape and semantic category of particles from the image, generating a deep feature map containing multi-level semantic information.

[0019] S22: Perform progressive upsampling and feature stitching on the deep feature map. Through skip connections, the shallow detail features from the early layers of the encoder path, which contain particle edges and textures, are fused and the resolution is restored with the high-level semantic features from the deep feature map, which contain the overall shape and category information of the particles, to generate a high-resolution feature map.

[0020] S23: Perform pixel-by-pixel classification processing on the high-resolution feature map, calculate and output the probability value of each pixel belonging to ruthenium nanoparticles through convolutional layers and nonlinear activation functions, and generate a binary probability distribution map characterizing the contour of ruthenium nanoparticles.

[0021] S24: Based on high-resolution feature maps and probability distribution maps, feature selection and serialization are performed. By using the probability distribution map as a spatial weight to apply attention weight to the high-resolution feature map, and flattening the weighted features along the spatial dimension, an image feature sequence focused on the foreground particle region is generated.

[0022] In one embodiment, S3 of the deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst provided by the present invention specifically includes the following steps:

[0023] S31: Perform multi-scale one-dimensional convolution filtering on the standard spectral data, using three one-dimensional convolution kernels with widths of 3, 5, and 7 to capture the sharp diffraction peak features, medium-width diffraction peak features, and broad and gentle diffraction envelope features in the spectrum, respectively, to generate multi-scale spectral feature maps.

[0024] S32: Perform feature fusion and pooling on the multi-scale spectral feature map, and stitch sharp diffraction peak features, medium-width diffraction peak features and wide and gentle diffraction envelope features by channel dimension, and perform global average pooling on the stitched feature map to aggregate global spectral information and generate fused spectral feature vector;

[0025] S33: The fused spectral feature vector is subjected to nonlinear transformation and dimensionality reduction. The fused spectral feature vector is mapped to a latent space with significantly reduced dimensionality through a fully connected layer and the ReLU activation function, generating a low-dimensional dense vector characterizing the crystal structure and zinc phase information as the spectral feature vector.

[0026] In one embodiment, S4 of the deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst provided by the present invention specifically includes the following steps:

[0027] S41: Perform linear projection processing on the spectral feature vector. Based on the linear transformation layer in the cross-attention mechanism, the spectral feature vector is mapped to the key feature space used to calculate similarity and the value feature space used to transmit information, respectively, to generate key vector and value vector.

[0028] S42: Perform similarity calculation on the image feature sequence and key vector, measure the correlation between the feature and key vector at each spatial location in the image feature sequence through dot product operation, and generate attention weight distribution;

[0029] S43: Based on the attention weight distribution, the value vector is weighted and aggregated, and the spectral information carried by the value vector is injected into the corresponding image feature space location according to the intensity of the attention weight distribution, generating an enhanced image feature sequence after modal interaction;

[0030] S44: Perform global aggregation processing on the enhanced image feature sequence. Average pooling is used to average the features at all locations along the spatial dimension to compress the spatial dimension of the feature sequence and generate a comprehensive feature representation.

[0031] In one embodiment, the calculation formula for the attention weight distribution of the deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst provided by the present invention is as follows:

[0032]

[0033] in, For attention weight distribution, Let K be the sequence of image features used as the query, and K be the key matrix formed by copying and expanding the key vectors. Let T be the feature dimension of the key vector, T denote the matrix transpose, and softmax be the normalization exponential function.

[0034] In one embodiment, S5 of the deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst provided by the present invention specifically includes the following steps:

[0035] S51: The training dataset, consisting of multiple catalyst samples with known real active phase parameters and benzene hydrogenation performance, is used for model training. Based on the HRTEM images, XRD spectral data, and corresponding real measurements of ruthenium nanoparticle average size, dispersion, benzene conversion rate, and cyclohexene selectivity in the training dataset, the parameters of the preset multi-channel deep learning model are optimized by the gradient descent algorithm to establish a nonlinear mapping relationship from multimodal data to target labels and generate a pre-trained multi-channel deep learning model.

[0036] S52: Perform first branch regression processing on the comprehensive feature representation, input the comprehensive feature representation into the first fully connected regression layer in the pre-trained multi-channel deep learning model corresponding to the active phase parameter prediction task, and map the comprehensive feature representation to the predicted average size and dispersion score of ruthenium nanoparticles through the weights and bias parameters learned by the first fully connected regression layer, thereby generating the active phase analysis results;

[0037] S53: Perform second-branch regression processing on the comprehensive feature representation, input the comprehensive feature representation into the second fully connected regression layer corresponding to the catalytic performance prediction task in the pre-trained multi-channel deep learning model, and map the comprehensive feature representation into the predicted benzene conversion percentage and cyclohexene selectivity percentage through the weights and bias parameters learned by the second fully connected regression layer, generating the benzene hydrogenation performance prediction result.

[0038] Secondly, this invention provides a deep learning-based system for analyzing the activity and predicting the performance of ruthenium zinc benzene hydrogenation catalysts, which includes the following modules:

[0039] The spectrum preprocessing module is used to perform image enhancement, noise reduction, and spectral normalization and alignment on the HRTEM image data of catalyst samples obtained by high-resolution transmission electron microscopy and the XRD spectral data of catalyst samples obtained by X-ray diffractometer, and to generate preprocessed image data and standard spectral data.

[0040] The image feature extraction module is used to call the deep learning encoder to extract the deep semantic features of ruthenium nanoparticles in the preprocessed image data, and call the decoder to reconstruct the probability distribution map representing the outline of ruthenium nanoparticles, generating an image feature sequence containing particle morphology and spatial distribution information.

[0041] The spectral feature mapping module is used to capture local peak shapes and global spectral line features in standard spectral data based on convolutional neural networks, and map the local peak shapes and global spectral line features into low-dimensional dense vectors that characterize crystal structure and zinc phase information, thereby generating spectral feature vectors.

[0042] The feature fusion coding module is used to dynamically focus the image features in the image feature sequence on the spatial region most relevant to the current spectral information in the spectral feature vector based on the attention mechanism, and integrate the image feature sequence and the spectral feature vector into a unified feature space to generate a comprehensive feature representation;

[0043] The catalytic performance prediction module is used to perform parallel regression prediction based on the comprehensive feature representation. It calls one branch regression layer to map the comprehensive feature representation to predict the average size and dispersion of ruthenium nanoparticles, and another branch regression layer to map the comprehensive feature representation to predict the conversion rate and cyclohexene selectivity of the benzene hydrogenation reaction, generating a comprehensive prediction result. The comprehensive prediction result is used to indicate the microstructure state and macroscopic catalytic performance of the catalyst sample.

[0044] Thirdly, this application provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned deep learning-based methods for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalysts.

[0045] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-mentioned deep learning-based methods for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalysts.

[0046] In summary, the deep learning-based method for activity analysis and performance prediction of ruthenium-zinc benzene hydrogenation catalysts provided in this application efficiently and objectively obtains the precise morphological distribution of ruthenium nanoparticles and the chemical state information of zinc species in the catalyst through automated preprocessing and feature extraction of HRTEM images and XRD spectra, overcoming the shortcomings of traditional manual analysis, which is characterized by strong subjectivity and low efficiency. By introducing a cross-attention mechanism for deep fusion of multimodal features, the intrinsic correlation between the geometric structure of ruthenium particles and the chemical state of zinc promoters can be effectively established, thereby enabling the analysis of the promoter's action mechanism at the microscale. Based on the fused comprehensive features, end-to-end parallel regression prediction can synchronously and quantitatively output the active phase parameters (average size and dispersion) and performance indicators (conversion rate and selectivity) of the catalyst, thus directly establishing the complex mapping relationship between "structure and performance". This method integrates the previously fragmented characterization analysis and performance evaluation into a unified, automated intelligent process, which not only significantly improves analysis efficiency and repeatability, but also provides a reliable data model foundation for performance-oriented reverse design of catalysts. It is expected to fundamentally change the traditional R&D model that relies on trial and error and accelerate the development of high-performance catalysts.

[0047] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0048] Figure 1 A flowchart illustrating a deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalysts, provided for embodiments of this application;

[0049] Figure 2 This is a schematic diagram of a deep learning-based system for analyzing the activity and predicting the performance of a ruthenium zinc benzene hydrogenation catalyst, provided as another embodiment of this application. Detailed Implementation

[0050] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0051] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0052] In one embodiment, such as Figure 1 As shown, a deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalysts is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0053] S1: Perform image enhancement, noise reduction, and spectral normalization alignment on the HRTEM image data of the catalyst samples obtained by high-resolution transmission electron microscopy and the XRD spectral data of the catalyst samples obtained by X-ray diffraction to generate preprocessed image data and standard spectral data.

[0054] Specifically, the system receives HRTEM image data of catalyst samples acquired through high-resolution transmission electron microscopy and XRD spectral data of catalyst samples acquired through X-ray diffraction. The system enhances the HRTEM image data by dividing the image into several grid blocks and independently performing histogram equalization on each grid block, limiting contrast variations in local areas and adjusting grayscale differences between different regions of the image to highlight the distinction between ruthenium nanoparticles and the background. After enhancement, the system performs noise reduction on the image data. First, it suppresses high-frequency noise by weighted averaging of the neighborhood of each pixel in the image. Then, it sorts and statistically analyzes the neighborhood of each pixel, replacing the original pixel value with the median value of the sorted neighborhood to remove isolated noise points and maintain the integrity of the ruthenium nanoparticle outline in the image.

[0055] Furthermore, the system performs baseline correction on the XRD spectral data. By fitting the spectral background curve, the influence of background interference factors on the spectrum is eliminated, ensuring that the intensity of the diffraction peaks in the spectrum accurately reflects the phase information of the catalyst. After baseline correction, the system performs a normalization operation on the spectral data, calculates the total peak area of ​​the spectrum, and calculates the ratio of all diffraction intensity values ​​in the spectrum to the total peak area, ensuring that the intensity of spectral data from different sources is within a uniform range. Using the characteristic diffraction peaks of the standard sample as a reference, the system performs peak alignment on the spectral data, adjusting the position of the diffraction peaks of corresponding phases in the spectrum to maintain consistency. Through the above processing, the system generates preprocessed image data and standard spectral data. The preprocessed image data can present the morphology and position of ruthenium nanoparticles, while the standard spectral data can reflect the crystal structure of the catalyst and related information about the zinc phase.

[0056] S2: Call the deep learning encoder to extract the deep semantic features of ruthenium nanoparticles in the preprocessed image data, and call the decoder to reconstruct the probability distribution map representing the outline of ruthenium nanoparticles, generating an image feature sequence containing particle morphology and spatial distribution information.

[0057] Specifically, the system invokes a deep learning encoder to process the preprocessed image data and extract deep semantic features of ruthenium nanoparticles. The deep learning encoder is built on a specific pre-trained network architecture, containing multiple convolutional blocks and pooling layers. The system extracts features from the preprocessed image data through convolutional operations within the convolutional blocks, progressively capturing semantic features from shallow edges and texture information to deeper features such as particle shape and spatial distribution. The pooling layers downsample the feature maps through specific operations, reducing data dimensionality while preserving key features and improving processing efficiency. After multiple rounds of convolution and pooling operations, the encoder outputs a feature map containing the core information of the ruthenium nanoparticles.

[0058] Furthermore, the system calls the decoder to process the feature map output by the encoder, reconstructing a probability distribution map representing the contour of ruthenium nanoparticles. The decoder employs a structure combining upsampling and skip connections. Upsampling operations gradually restore the spatial resolution of the feature map, while skip connections fuse feature maps from different levels of the encoder with corresponding levels of the decoder, compensating for details lost during upsampling. Through multiple rounds of upsampling and feature fusion operations, the decoder finally outputs a feature map with the same size as the original preprocessed image data. The system then activates this feature map to generate a binary classification probability distribution map. Each pixel in the probability distribution map corresponds to the probability that its location in the original image belongs to either a ruthenium nanoparticle or the background. Finally, the system flattens the feature map output by the encoder, transforming it into an ordered vector sequence. Each vector in this sequence contains information about the morphology and spatial distribution of ruthenium nanoparticles in the corresponding spatial region, forming an image feature sequence.

[0059] S3: Based on convolutional neural networks, capture local peak shapes and global spectral features in standard spectral data, and map local peak shapes and global spectral features into low-dimensional dense vectors that characterize crystal structure and zinc phase information, generating spectral feature vectors.

[0060] Specifically, the system processes standard spectral data based on a convolutional neural network (CNN) to capture local peak shapes and overall spectral features. The CNN consists of multiple one-dimensional convolutional layers, pooling layers, and fully connected layers. After the system inputs the standard spectral data into the CNN, it first performs calculations through the first convolutional layer. The kernel of this layer, designed with a specific size, captures local features corresponding to narrow diffraction peaks in the spectrum; these features are related to the crystallization state of a specific phase. The system then performs calculations through the second convolutional layer, whose kernel size differs from the first layer. This second layer captures local features corresponding to broad diffraction peaks in the spectrum; these features are related to lattice strain, amorphous phases, or the phase of a specific additive. The third convolutional layer fuses the features output from the first two layers, integrating local peak shape features at different scales.

[0061] Furthermore, the system processes the feature map output from the convolutional layer using a global average pooling layer to eliminate spatial dimensional redundancy and retain the overall spectral line characteristics. These characteristics include the distribution pattern of diffraction peaks and peak intensity ratios, reflecting the overall crystal structure and phase composition ratio of the catalyst. The feature vector output from the global average pooling layer is input to a fully connected layer. The first fully connected layer filters and transforms the feature vector, retaining key features that affect catalytic performance. The second fully connected layer further compresses the feature dimension, mapping high-dimensional features into a low-dimensional dense vector. This low-dimensional dense vector integrates local peak shapes and overall spectral line characteristics from standard spectral data, comprehensively characterizing the catalyst's crystal structure and zinc phase information, forming a spectral feature vector.

[0062] S4: Based on the attention mechanism, the image features in the image feature sequence are dynamically focused on the spatial region most relevant to the current spectral information in the spectral feature vector, and the image feature sequence and spectral feature vector are integrated into a unified feature space to generate a comprehensive feature representation.

[0063] Specifically, the system processes image feature sequences and spectral feature vectors based on an attention mechanism. It performs format conversion on both, setting the image feature sequences as a query vector set and mapping the spectral feature vectors to a set of key vectors and a set of value vectors with the same dimensions as the query vector set through linear transformation. Further, the system calculates the correlation strength between the query vector set and the key vector set, obtaining an attention weight matrix through specific mathematical operations. Each element in this matrix reflects the correlation between the corresponding vector in the image feature sequence and the corresponding information in the spectral feature vector. The system then performs a weighted summation of the value vector set based on the attention weight matrix, dynamically focusing each vector in the image feature sequence on the information with the highest current correlation to the spectral feature vector, highlighting the feature combinations that play a crucial role in catalytic performance.

[0064] Through this process, the system injects chemical state information from the spectral feature vector into the image feature sequence, achieving dynamic correlation between the two types of heterogeneous features. The system then further processes the fused features, compressing them into fixed-dimensional vectors through linear transformation. This integrates the image feature sequence and the spectral feature vector into a unified feature space, forming a comprehensive feature representation. This comprehensive feature representation simultaneously contains both the catalyst's microstructure and chemical state information, providing comprehensive feature support for subsequent prediction tasks.

[0065] S5: Parallel regression prediction is performed based on the comprehensive feature representation. One branch regression layer is called to map the comprehensive feature representation to predict the average size and dispersion of ruthenium nanoparticles. Another branch regression layer is used to map the comprehensive feature representation to predict the conversion rate and cyclohexene selectivity of the benzene hydrogenation reaction, generating comprehensive prediction results. The comprehensive prediction results are used to indicate the microstructure state and macroscopic catalytic performance of the catalyst sample.

[0066] Specifically, the system inputs the comprehensive feature representation into two independent regression paths, each containing continuous linear transformations and activation operations. One regression path progressively maps the comprehensive feature representation, establishing a correspondence between the comprehensive feature representation and relevant parameters of ruthenium nanoparticles, ultimately outputting the predicted results related to the average size and dispersion of ruthenium nanoparticles. The other regression path similarly progressively maps the comprehensive feature representation, establishing a correspondence between the comprehensive feature representation and relevant performance parameters of the benzene hydrogenation reaction, ultimately outputting the predicted results related to the conversion rate of the benzene hydrogenation reaction and the selectivity of cyclohexene. During the regression prediction process, the system uses a specific loss calculation method to measure the deviation between the predicted results and the actual situation, adjusting the operational parameters in the regression path based on the loss value to continuously bring the predicted results closer to the actual situation. The system uses optimization algorithms to regulate the entire regression process, ensuring the effectiveness and stability of parameter adjustments and improving the reliability of the prediction results. The generated comprehensive prediction results include the average size and dispersion of ruthenium nanoparticles, as well as the conversion rate and cyclohexene selectivity of the benzene hydrogenation reaction. These results collectively reflect the microstructure and macroscopic catalytic performance of the catalyst sample, providing data support for related research and applications of catalysts. Key information about the catalyst can be obtained without additional manual analysis and experimental testing.

[0067] In summary, the deep learning-based method for activity analysis and performance prediction of ruthenium-zinc benzene hydrogenation catalysts provided in this application efficiently and objectively obtains the precise morphological distribution of ruthenium nanoparticles and the chemical state information of zinc species in the catalyst through automated preprocessing and feature extraction of HRTEM images and XRD spectra, overcoming the shortcomings of traditional manual analysis, which is characterized by strong subjectivity and low efficiency. By introducing a cross-attention mechanism for deep fusion of multimodal features, the intrinsic correlation between the geometric structure of ruthenium particles and the chemical state of zinc promoters can be effectively established, thereby enabling the analysis of the promoter's action mechanism at the microscale. Based on the fused comprehensive features, end-to-end parallel regression prediction can synchronously and quantitatively output the active phase parameters (average size and dispersion) and performance indicators (conversion rate and selectivity) of the catalyst, thus directly establishing the complex mapping relationship between "structure and performance". This method integrates the previously fragmented characterization analysis and performance evaluation into a unified, automated intelligent process, which not only significantly improves analysis efficiency and repeatability, but also provides a reliable data model foundation for performance-oriented reverse design of catalysts. It is expected to fundamentally change the traditional R&D model that relies on trial and error and accelerate the development of high-performance catalysts.

[0068] In one embodiment, S1 of the deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst provided by the present invention specifically includes the following steps:

[0069] S11: Perform contrast-limited adaptive histogram equalization on the HRTEM image data of the catalyst sample obtained by high-resolution transmission electron microscopy to enhance the grayscale difference between ruthenium nanoparticles and the support background, and generate a contrast-enhanced image.

[0070] Specifically, the system performs a full-area scan of the input HRTEM image data. Based on the overall image size and grayscale distribution characteristics, it automatically divides the image into multiple non-overlapping sub-regions. This division process ensures that each sub-region can independently reflect local grayscale variation patterns, and that the number and size of the sub-regions are adapted to the grayscale detail distribution of the image. For each sub-region, the system statistically analyzes the grayscale values ​​of all pixels within it, generating a corresponding grayscale histogram to clearly define the proportion of pixels with different grayscale values ​​within that region. To avoid noise amplification caused by excessive local grayscale stretching, the system pre-sets contrast limiting logic. When the proportion of pixels in a certain grayscale interval in the grayscale histogram exceeds the set limit, the system proportionally distributes the excess pixels to other grayscale intervals, readjusting the dynamic range of the grayscale distribution in that sub-region, thus reasonably amplifying local grayscale differences.

[0071] After adjusting the histogram of a single sub-region, the system performs transition processing on the gray-level boundaries of adjacent sub-regions. By adjusting the gradient of the gray-level values ​​of pixels in the boundary regions, it eliminates the abrupt gray-level changes caused by sub-region division, ensuring the gray-level continuity and visual consistency of the entire image. Through this series of sequential operations, the system effectively enhances the gray-level distinction between the region containing ruthenium nanoparticles and the carrier background region. The structural information such as particle edges and contours, which were originally obscured by the background or had indistinct gray-level differences, are clearly presented, generating a contrast-enhanced image that meets the requirements of subsequent processing.

[0072] S12: Perform nonlocal mean denoising on the contrast-enhanced image to suppress Gaussian noise and speckle noise while preserving grain edge details, generating preprocessed image data.

[0073] Specifically, the system locates each pixel in the contrast-enhanced image one by one according to a preset scanning order, and takes the currently located pixel as the target pixel. Centered on the target pixel, the system automatically defines a fixed-range neighborhood window that completely includes the target pixel and all pixels within a certain surrounding range, thus comprehensively capturing the local grayscale environment information of the target pixel. Subsequently, within the effective range of the entire contrast-enhanced image, the system searches for all candidate neighborhood windows with similar grayscale distribution characteristics to the target pixel's neighborhood window. The search process covers all areas of the image, ensuring that no possible similar windows are missed. After finding candidate neighborhood windows, the system establishes a similarity evaluation criterion by calculating the grayscale difference between the target pixel's neighborhood window and each candidate neighborhood window. The smaller the grayscale difference of the candidate neighborhood window, the higher its similarity to the target pixel's neighborhood, and the higher its corresponding weight allocation.

[0074] Furthermore, the system performs a weighted average of the grayscale values ​​of corresponding pixels in all candidate neighborhood windows according to their respective similarity weights. The result is the grayscale value of the target pixel after noise reduction. Throughout the processing, the system effectively filters Gaussian noise and speckle noise through preset neighborhood window specifications and similarity evaluation algorithms. Because the calculation of similarity weights relies heavily on local grayscale distribution characteristics, the edge details of the ruthenium nanoparticles are not destroyed during noise reduction, thus preserving the original morphology and structural features of the particles. After the system completes noise reduction processing on all pixels in the image, it outputs preprocessed image data with noise suppression and intact edge details.

[0075] S13: Perform minimum-maximum normalization on the XRD spectral data of the catalyst sample obtained by X-ray diffractometer to scale the original intensity values ​​to a uniform numerical range and generate normalized spectra.

[0076] Specifically, the system comprehensively traverses the input XRD spectral data, collecting all raw intensity values. Through data statistics during the traversal, the maximum and minimum values ​​of the raw intensity values ​​in the XRD spectral data are determined. These two extreme values ​​will serve as the core benchmark parameters for subsequent normalization processing. Further, following a preset minimum-maximum normalization calculation logic, the system performs a linear transformation operation on each raw intensity value based on the determined maximum and minimum intensity values. The core logic of the linear transformation operation is to scale each raw intensity value proportionally according to the range corresponding to the extreme value, ensuring that all transformed intensity values ​​fall within a uniform numerical range. This effectively eliminates fluctuations in the raw spectral intensity values ​​caused by differences in the amount of catalyst used during preparation, differences in crystallinity, and minor fluctuations in instrument parameters during detection, thus avoiding interference from inconsistencies in intensity magnitude between different samples in subsequent spectral feature extraction and analysis.

[0077] During the normalization process, the system strictly adheres to linear transformation rules, ensuring that the relative intensity relationships between diffraction peaks of different intensity levels remain unchanged. Whether it's a high-intensity strong diffraction peak or a low-intensity weak diffraction peak, their relative proportion and distribution characteristics in the spectrum are fully preserved, without distortion or loss of accuracy due to normalization. Through this standardized processing procedure, the system generates normalized spectra, providing a basis for direct comparison and analysis of the XRD spectra of different catalyst samples in the intensity dimension.

[0078] S14: Perform peak alignment based on dynamic time warping on the normalized spectrum, match and calibrate the diffraction angle positions of the normalized spectrum and the standard substance spectrum, and generate standard spectral data.

[0079] Specifically, the system loads standard substance spectra from a pre-set database. These standard substance spectra contain the position information of characteristic diffraction peaks of known phases, serving as the core reference benchmark for peak alignment processing. The positions of these characteristic diffraction peaks have been verified multiple times, ensuring high accuracy and authority. Further, the system performs data analysis on the normalized spectrum and the standard substance spectrum, extracting their corresponding diffraction angle and intensity sequences to construct an original data mapping framework between the normalized spectrum and the standard substance spectrum. Based on this mapping framework, the system initiates a dynamic time warping algorithm to elastically adjust the diffraction angle sequence of the normalized spectrum. During the adjustment process, the relative intensity relationship of the diffraction peaks within the normalized spectrum remains unchanged; only the position of the diffraction angles is adaptively adjusted.

[0080] Preferably, the system calculates the deviation between the position of each characteristic diffraction peak in the normalized spectrum and the corresponding position of the characteristic diffraction peak in the standard substance spectrum. Based on the magnitude and distribution of the deviation, the system gradually adjusts the diffraction angle axis of the normalized spectrum. The adjustment process employs an iterative optimization method, recalculating the deviation after each iteration until all characteristic diffraction peaks in the normalized spectrum and their corresponding positions in the standard substance spectrum achieve optimal matching in diffraction angle. This process effectively eliminates peak position shifts caused by differences in instrument precision, minor deviations during sample placement, and fluctuations in the detection environment, ensuring consistency in the diffraction angle dimension of XRD spectra for different catalyst samples. After completing the peak alignment process, the system generates standard spectral data, enabling subsequent spectral feature extraction to be conducted based on a unified peak position benchmark, effectively guaranteeing the reliability of feature extraction results and the comparability of features between different samples.

[0081] In one embodiment, S2 of the deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst provided by the present invention specifically includes the following steps:

[0082] S21: Based on the encoder-decoder network architecture, multi-level convolution and pooling downsampling are performed on the preprocessed image data. The deep residual network is called to extract shallow geometric features containing basic edges and textures and high-level abstract features containing the overall shape and semantic category of particles from the image, generating a deep feature map containing multi-level semantic information.

[0083] Specifically, the system first performs an initial convolution operation on the preprocessed image data. The convolution kernel traverses the image along a preset path, aggregating and calculating the pixel information of each local region, and transforming the correlation information of adjacent pixels into an initial feature response. After the initial convolution, the system performs a pooling downsampling operation to compress the spatial dimension of the feature map generated by the convolution, retaining key intensity change information in the feature map and reducing the amount of data in subsequent calculations. The system repeats the convolution and pooling downsampling operations according to the above process, forming a multi-level processing chain. Each level of processing is based on the feature map output from the previous level, and the feature representation capability is improved by increasing the number of convolution kernels.

[0084] During multi-level processing, the system invokes a deep residual network to participate in the computation. By constructing identity mapping paths, the deep residual network enables image data to be directly transmitted across some convolutional layers while being processed by the convolutional layers, avoiding the attenuation of feature information in deep networks. Through multi-level processing by the deep residual network, the system extracts features from the early layers of the network. These features originate from the basic grayscale changes of the image, corresponding to the edge contours and surface texture information of ruthenium nanoparticles. The system extracts another type of feature from the later layers of the network. These features, after multiple convolutions and aggregations, have decoupled from the direct association with the original pixels, corresponding to the overall morphology and semantic category information of ruthenium nanoparticles. The system fuses the shallow geometric features and high-level abstract features extracted from different layers along the channel dimension, integrating the complementary information of the two types of features to form a deep feature map containing multi-level semantic information. This feature map fully preserves features across all dimensions, from basic details to abstract semantics.

[0085] S22: Progressive upsampling and feature stitching are performed on the deep feature map. Through skip connections, shallow detail features containing particle edges and textures from the early levels of the encoder path are fused and resolution restored with high-level semantic features containing the overall shape and category information of the particles from the deep feature map, generating a high-resolution feature map.

[0086] Specifically, the system upsamples deep feature maps, expanding their spatial resolution through interpolation to gradually bring their size closer to the preprocessed image data. After each upsampling, the system performs convolution on the expanded feature map, adjusting the number of channels to ensure its dimensionality matches that of subsequently fused features, while also optimizing feature representation accuracy. The system repeats the upsampling and convolution optimization operations at a preset number of levels, forming a progressive resolution recovery chain. During progressive upsampling, the system extracts feature maps from corresponding levels in the encoder path using a skip connection mechanism. These feature maps from earlier encoder levels have not undergone multiple downsampling compressions, thus fully preserving the edge details and texture information of the ruthenium nanoparticles.

[0087] Furthermore, the system concatenates the extracted early encoder feature maps with the upsampled deep feature maps, enabling the shallow detail features and the high-level semantic features in the deep feature maps to complement each other in the same dimension. After concatenation, the system performs a convolutional integration operation on the fused feature maps to eliminate redundant information between features from different sources and enhance the consistent representation of features. The above-mentioned processes of upsampling, skip connections, feature concatenation, and convolutional integration are advanced step by step, with each level achieving the fusion of corresponding shallow details and high-level semantics, until the resolution of the feature maps is restored to the preset standard, ultimately generating a high-resolution feature map that contains both the overall shape and category information of the ruthenium nanoparticles and clearly presents the detailed features of edges and textures.

[0088] S23: Perform pixel-by-pixel classification processing on the high-resolution feature map, calculate and output the probability value of each pixel belonging to ruthenium nanoparticles through convolutional layers and nonlinear activation functions, and generate a binary probability distribution map characterizing the contour of ruthenium nanoparticles.

[0089] Specifically, the system performs convolution processing on the high-resolution feature map. The kernel size of this convolutional layer is set according to the feature integration requirements. By traversing the high-resolution feature map, the system extracts and aggregates the local features around each pixel, converting the high-dimensional feature vector corresponding to each pixel into a low-dimensional feature representation, simplifying subsequent classification and judgment operations. After convolution processing, the system calls a non-linear activation function to transform the low-dimensional features. The activation function maps the low-dimensional features to a fixed numerical range through specific operational logic, giving the feature information comparable probabilistic attributes. Based on the transformed feature information, the system calculates the probability of each pixel belonging to a ruthenium nanoparticle in the high-resolution feature map. The calculation process determines the probability that each pixel belongs to a ruthenium nanoparticle based on the degree of matching between the feature and the feature pattern of the ruthenium nanoparticle. The system records the probability calculation results of all pixels across the entire image space, forming probability distribution data covering the entire image space. The system generates a binary probability distribution map based on the probability distribution data. The value of each spatial location in the distribution map directly corresponds to the probability of the pixel at that location. Through this distribution form, the boundary between the area where the ruthenium nanoparticles are located and the background area in the image is clearly delineated, and the contour morphology of the ruthenium nanoparticles is fully characterized.

[0090] S24: Based on high-resolution feature maps and probability distribution maps, feature selection and serialization are performed. By using the probability distribution map as a spatial weight to apply attention weight to the high-resolution feature map, and flattening the weighted features along the spatial dimension, an image feature sequence focused on the foreground particle region is generated.

[0091] Specifically, the system first uses a binary probability distribution map as a spatial weight template to perform attention-weighted operations on the high-resolution feature map. It traverses each feature point in the high-resolution feature map in a preset order, extracting the corresponding feature value. Simultaneously, it extracts the probability value of the same spatial location in the binary probability distribution map. The feature value and probability value are multiplied, thus strengthening feature points in areas with higher probability values ​​and weakening feature points in areas with lower probability values. This achieves the filtering of feature information in the high-resolution feature map, focusing on the foreground region where the ruthenium nanoparticles are located. After the weighting process, the system performs spatial dimension flattening on the weighted feature map. Each feature point in the two-dimensional feature map is arranged sequentially from left to right and top to bottom according to a preset order, converting the feature information in two-dimensional space into a one-dimensional continuous feature sequence.

[0092] During the flattening process, the system retains the relative spatial correlation information of each feature point, ensuring that the feature elements in the sequence can reflect the spatial distribution relationship of ruthenium nanoparticles in the original image. Through the above processing, the system integrates the feature information related to ruthenium nanoparticles scattered in two-dimensional space into a unified one-dimensional feature sequence. This sequence contains key features such as the morphology, outline, and spatial distribution of ruthenium nanoparticles, generating an image feature sequence focused on the foreground particle region, providing accurate image feature input for subsequent multimodal fusion with spectral features.

[0093] In one embodiment, S3 of the deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst provided by the present invention specifically includes the following steps:

[0094] S31: Perform multi-scale one-dimensional convolution filtering on the standard spectral data, using three types of one-dimensional convolution kernels with widths of 3, 5, and 7 to capture the sharp diffraction peak features, medium-width diffraction peak features, and broad and gentle diffraction envelope features in the spectrum, respectively, to generate multi-scale spectral feature maps.

[0095] Specifically, the system first formats the standard spectral data, organizing the correspondence between diffraction angles and intensities into a one-dimensional sequence. This ensures the data arrangement order aligns with the diffraction angle variation trend, adapting to the subsequent sliding traversal requirements of the convolution kernel. The system then calls three different widths of one-dimensional convolution kernels to perform parallel convolution operations on the formatted spectral data. The core of the convolution operation follows the formula:

[0096]

[0097] in, This represents the feature response value at the i-th position after the convolution operation, used to characterize the aggregation result of the spectral intensity at that position and in the neighborhood; This represents the k-th weight parameter of the convolution kernel, used to adjust the contribution of the original spectral data at the corresponding position; Indicating the first in standard spectral data The original intensity values ​​at each location are the basis for the input of the convolution operation; K represents the width of the convolution kernel, which determines the spectral neighborhood covered by the convolution operation. The three types of convolution kernels correspond to different K values, and this formula enables the accurate capture of peak features at different scales.

[0098] Preferably, the first type of convolution kernel, with a small K value, aggregates intensity changes within a narrow local range, extracting sharp diffraction peaks with small wavelength spans. These features correspond to the lattice diffraction signals of specific zinc phases in the catalyst. The second type of convolution kernel, with a moderate K value, aggregates intensity changes with a medium span, extracting diffraction peaks with medium widths between sharp and gentle peaks. These features are related to the lattice strain of the crystal structure or the diffraction signals of some alloy phases in the catalyst. The third type of convolution kernel, with a large K value, aggregates intensity change trends within a larger wavelength span, extracting broadened peak shape information, i.e., gentle diffraction envelope features. These features correspond to scattering signals related to amorphous phases or supports in the catalyst. After each type of convolution kernel completes traversal operations in order from the start to the end of the spectrum, it generates its corresponding single-scale feature map. The system integrates the three types of single-scale feature maps according to the channel dimension to form a multi-scale spectral feature map.

[0099] S32: Perform feature fusion and pooling on the multi-scale spectral feature map, and stitch together sharp diffraction peak features, medium-width diffraction peak features and wide and gentle diffraction envelope features through the channel dimension. Then, perform global average pooling on the stitched feature map to aggregate global spectral information and generate a fused spectral feature vector.

[0100] Specifically, the system performs channel-dimensional stitching on three types of single-scale feature maps in the multi-scale spectral feature map. Sharp diffraction peak feature maps, medium-width diffraction peak feature maps, and broad diffraction envelope feature maps are superimposed and integrated along the channel dimension in a preset order. This ensures that the original information of each type of peak shape feature is fully preserved and complementary, so that the integrated feature map simultaneously covers spectral peak shape information at different scales. After stitching, the system performs a global average pooling operation on the fused feature map to aggregate the global feature information of the entire spectral sequence. This operation follows the formula:

[0101]

[0102] in, represents the pooling output value of the c-th channel, used to characterize the average response of the scale feature corresponding to that channel across the entire spectral range; H represents the spatial dimension length of the fused feature map, corresponding to the length of the diffraction angle sequence of the spectral data; This represents the eigenvalue at position i in channel c of the fused feature map, which is the feature response at the i-th diffraction angle position at the corresponding scale for that channel. The system uses this formula to traverse the entire spatial dimension of the fused feature map, summing and averaging all eigenvalues ​​for each channel, thus converting the high-dimensional spatial features corresponding to each channel into a single eigenvalue. This operation effectively eliminates the influence of local intensity fluctuations on feature representation, enhances the stability of the overall spectral features, and significantly reduces the dimensionality of the feature data, simplifying the computational complexity of subsequent processing. Through the coherent processing of channel splicing and global average pooling, the system generates a fused spectral feature vector. This vector integrates global information on multi-scale peak shapes and can initially reflect the crystal structure type of the catalyst and the overall state of the zinc phase.

[0103] S33: The fused spectral feature vector is subjected to nonlinear transformation and dimensionality reduction. The fused spectral feature vector is mapped to a latent space with significantly reduced dimensionality through a fully connected layer and the ReLU activation function, generating a low-dimensional dense vector characterizing the crystal structure and zinc phase information as the spectral feature vector.

[0104] Specifically, the system inputs the fused spectral feature vector into the fully connected layer, and performs a spatial mapping on the fused feature vector through a linear transformation of the fully connected layer. The linear transformation follows the formula:

[0105]

[0106] Where h represents the feature vector after linear transformation, which is the intermediate feature expression of the fused feature vector after spatial mapping; W represents the weight matrix of the fully connected layer, used to weight and integrate the features of each dimension of the fused feature vector, thereby strengthening useful features and weakening redundant features; z represents the input fused spectral feature vector, which is the original data basis of the linear transformation; b represents the bias vector of the fully connected layer, used to adjust the overall offset of the linear transformation result and improve the accuracy of feature mapping. Through matrix multiplication and vector superposition operations of this formula, the fused feature vector is mapped to a new feature space, while simultaneously completing the enhanced screening of feature components related to crystal structure and zinc phase.

[0107] After the linear transformation is completed, the system calls the ReLU activation function to perform nonlinear processing on the transformed feature vector to introduce complex correlations between features. The ReLU activation function follows the formula:

[0108]

[0109] in, This represents the j-th feature value after activation. This represents the j-th element of the eigenvector after linear transformation. This function sets eigenvalues ​​less than 0 to 0 and retains eigenvalues ​​greater than 0, resulting in a non-linear distribution of the eigenvector, thus more accurately characterizing the intrinsic relationship between the crystal structure and the zinc phase. After the non-linear transformation, the system inputs the eigenvector into a subsequent fully connected layer for further dimensionality reduction. Through continuous weight matrix operations, the dimensionality of the eigenvector is gradually compressed, mapping the high-dimensional eigenvector to a latent space with significantly reduced dimensionality. During dimensionality reduction, the system retains the core feature information related to the crystal structure and zinc phase, generating a low-dimensional dense vector as the spectral feature vector. This vector can compactly and accurately characterize the crystal structure information of the catalyst and the existence state of the zinc phase.

[0110] In one embodiment, S4 of the deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst provided by the present invention specifically includes the following steps:

[0111] S41: Perform linear projection processing on the spectral feature vectors. Based on the linear transformation layer in the cross-attention mechanism, the spectral feature vectors are mapped to the key feature space used to calculate similarity and the value feature space used to transmit information, respectively, to generate key vectors and value vectors.

[0112] Specifically, the system invokes the linear transformation layer built into the cross-attention mechanism to perform a linear transformation operation on the spectral feature vector. The linear transformation follows the formula:

[0113]

[0114] in, This represents the output feature vector after linear projection. This represents the weight matrix of the linear transformation layer, used to weight and integrate the features of each dimension of the spectral feature vector to achieve spatial mapping. This represents the input spectral feature vector. The bias vector, representing the linear transformation layer, is used to adjust the overall distribution of the linear transformation results to improve mapping accuracy. The system performs two linear projections using two independent sets of weight matrices and bias vectors. The first projection maps the spectral feature vectors to the key feature space used for similarity calculation, generating a key vector. The second projection maps the spectral feature vectors to the value feature space used for information transmission, generating a value vector. Both projection processes strictly adhere to the aforementioned linear transformation formula, ensuring that the output key and value vectors satisfy the feature requirements for similarity calculation and information transmission, respectively.

[0115] S42: Perform similarity calculation on the image feature sequence and key vector, measure the correlation between the feature and key vector at each spatial location in the image feature sequence through dot product operation, and generate attention weight distribution.

[0116] Specifically, the system first performs a copying and expansion process on the key vector to generate a key matrix with the same length as the image feature sequence, ensuring that the features at each spatial location in the image feature sequence can be correlated with the key vector one by one. The image feature sequence serves as the query vector. The expanded key matrix serves as the key vector. The system and A dot product operation is performed, and an initial correlation matrix is ​​obtained by multiplying element-wise and then summing the results. This operation is used to quantify the correlation between image features and spectral features. Preferably, the system performs normalization by substituting the values ​​into the attention weight distribution calculation formula. The attention weight distribution calculation formula is as follows:

[0117]

[0118] in, For attention weight distribution, Let K be the sequence of image features used as the query, and K be the key matrix formed by copying and expanding the key vectors. Let T represent the feature dimension of the key vector, T denote the matrix transpose, and softmax be the normalization exponential function. The system sequentially performs matrix transpose, division, and normalization operations according to this formula, converting the correlation result obtained from the dot product operation into an attention weight distribution that conforms to the characteristics of a probability distribution. This distribution can accurately characterize the correlation strength between spatial location features and spectral features in the image feature sequence.

[0119] S43: The value vector is weighted and aggregated based on the attention weight distribution. The spectral information carried by the value vector is injected into the corresponding image feature space location according to the intensity of the attention weight distribution, generating an enhanced image feature sequence after modal interaction.

[0120] Specifically, after obtaining the attention weight distribution and value vector, the system initiates a weighted aggregation process. The system first performs a copying and expansion process on the value vector to ensure its dimension perfectly matches the attention weight distribution, guaranteeing that each attention weight accurately corresponds to a specific element in the value vector. Preferably, the weighted aggregation process follows the formula:

[0121]

[0122] in, This represents the enhanced image feature sequence after modal interaction. This represents the expanded value vector, which carries spectral information related to the catalyst's crystal structure and zinc phase. The system uses this formula to perform matrix multiplication between the attention weight distribution and the expanded value vector, achieving weighted aggregation. Value vector elements with larger values ​​in the attention weight distribution are assigned higher contributions, while those with smaller values ​​have reduced contributions. This process precisely injects the spectral information carried by the value vector into the corresponding spatial locations in the image feature sequence according to the intensity of the attention weight distribution. This enhances the spectral information in regions of the image feature sequence that are strongly correlated with spectral features, generating an enhanced image feature sequence that simultaneously contains information on the microstructure of ruthenium nanoparticles and the catalyst crystal and zinc phase information.

[0123] S44: Perform global aggregation processing on the enhanced image feature sequence. Average pooling is used to average the features at all locations along the spatial dimension to compress the spatial dimension of the feature sequence and generate a comprehensive feature representation.

[0124] Specifically, the system performs average pooling on the enhanced image feature sequence, following the formula:

[0125]

[0126] in, This is the final generated comprehensive feature representation. This represents the spatial dimension length of the enhanced image feature sequence, i.e., the number of spatial locations contained in the feature sequence. This represents the feature vector at the i-th spatial location in the enhanced image feature sequence. The system traverses the feature vectors at all locations along the spatial dimensions of the feature sequence, performing a summation and averaging operation on all elements in each feature dimension. This operation eliminates local detail differences between different spatial locations, preserving the core information of the fused dual-modality model. After completing the average pooling process, the system generates a comprehensive feature representation that integrates the microstructure information of ruthenium nanoparticles with the crystal and zinc phase information of the catalyst, providing a unified feature input for the subsequent parallel regression prediction module.

[0127] In one embodiment, S5 of the deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst provided by the present invention specifically includes the following steps:

[0128] S51: The training dataset, consisting of multiple catalyst samples with known real active phase parameters and benzene hydrogenation performance, is used for model training. Based on the HRTEM images, XRD spectral data, and corresponding real measurements of ruthenium nanoparticle average size, dispersion, benzene conversion rate, and cyclohexene selectivity in the training dataset, the parameters of the pre-set multi-channel deep learning model are optimized using the gradient descent algorithm to establish a nonlinear mapping relationship from multimodal data to target labels, thereby generating a pre-trained multi-channel deep learning model.

[0129] Specifically, the system first loads a training dataset. The dataset contains HRTEM image data, which characterizes the microstructure, spatial distribution, and other chemical structural information of ruthenium nanoparticles in the catalyst; XRD spectral data, which reflects the crystal structure, zinc phase composition, and ruthenium-zinc interaction of the catalyst; and corresponding target labels, including the average size and dispersion of ruthenium nanoparticles (two core parameters of the active phase), and benzene conversion and cyclohexene selectivity (two key catalytic performance indicators of the benzene hydrogenation reaction). The system uses HRTEM image data and XRD spectral data to form multimodal input features, and four types of target labels as chemical characteristic benchmarks, all input into a pre-defined multi-channel deep learning model. The training process aims to minimize the deviation between the model's predicted values ​​and the chemical characteristic benchmarks, employing a gradient descent algorithm to optimize the model parameters. Parameter updates follow the formula:

[0130]

[0131] in, This represents the updated model parameters. This represents the model parameters in the current iteration round. The learning rate is used to adjust the step size of parameter updates to ensure training stability. Represents the loss function For the current parameter The gradient loss function specifically measures the prediction bias of active phase parameters and the prediction bias of catalytic performance. The active phase parameter bias is related to the prediction accuracy of the ruthenium nanoparticle chemical structure, while the catalytic performance bias is related to the prediction accuracy of the benzene hydrogenation reaction chemical process. The system continuously updates the model's weights, biases, and other parameters through multiple iterations, gradually optimizing the nonlinear mapping relationship from multimodal characterization data to catalyst chemical property labels. Training terminates when the loss function value converges to a preset stable range, indicating that the model's prediction accuracy of the catalyst's chemical properties meets the requirements, generating a pre-trained multi-channel deep learning model.

[0132] S52: Perform first-branch regression processing on the comprehensive feature representation, input the comprehensive feature representation into the first fully connected regression layer in the pre-trained multi-channel deep learning model corresponding to the active phase parameter prediction task, and map the comprehensive feature representation into the predicted average size and dispersion score of ruthenium nanoparticles through the weights and bias parameters learned by the first fully connected regression layer, thereby generating the active phase analysis results.

[0133] Specifically, the system inputs the comprehensive feature representation generated by S44 into a pre-trained multi-channel deep learning model. Through the model's internal directional transfer mechanism, the comprehensive features are imported into the first fully connected regression layer corresponding to the active phase parameter prediction task. The comprehensive features integrate chemical correlation information between the ruthenium nanoparticle microstructure from HRTEM and the crystal structure from XRD. The weight matrix and bias vector of the first fully connected regression layer have been learned during the training phase and are specifically used to extract and map chemical features related to the active phase. The first fully connected regression layer performs a linear transformation operation on the comprehensive feature representation, following the formula:

[0134]

[0135] in, The output represents the average size and dispersion score of ruthenium nanoparticles. These two parameters directly determine the chemical activity of the active phase of the catalyst. The average size affects the specific surface area of ​​the active sites, and the dispersion affects the uniformity of the distribution of the active sites. This represents the weight matrix of the first fully connected regression layer. Its parameter distribution is adapted to the extraction requirements of the chemical characteristics of ruthenium nanoparticles, and can accurately capture the core information related to activity in the comprehensive characteristics. This represents the bias vector of the first fully connected regression layer, used to calibrate the linear transformation results to improve the prediction accuracy of the active phase parameters; This represents the comprehensive feature representation of the input. Through the linear transformation described above, the system accurately maps the comprehensive feature representation into prediction results directly related to the chemical properties of the active phase, generating active phase analysis results that include the average size and dispersion of ruthenium nanoparticles. These results directly reflect the chemical state of the catalyst's active centers.

[0136] S53: Perform second-branch regression processing on the comprehensive feature representation, input the comprehensive feature representation into the second fully connected regression layer corresponding to the catalytic performance prediction task in the pre-trained multi-channel deep learning model, and map the comprehensive feature representation into the predicted benzene conversion percentage and cyclohexene selectivity percentage through the weights and bias parameters learned by the second fully connected regression layer, generating the benzene hydrogenation performance prediction result.

[0137] Specifically, the system inputs the comprehensive feature representation into a pre-trained multi-channel deep learning model, and imports it into a second fully connected regression layer corresponding to the catalytic performance prediction task through a directed transfer mechanism. The second fully connected regression layer uses a fixed weight matrix and bias vector trained independently of the first fully connected regression layer. Its parameters have been trained to adapt to the chemical characteristics associated with the benzene hydrogenation reaction, enabling it to accurately extract the core chemical information affecting the catalytic reaction from the comprehensive features. The second fully connected regression layer performs a linear transformation operation on the comprehensive feature representation, following the formula:

[0138]

[0139] in, The predicted output represents the percentage of benzene conversion and the percentage of cyclohexene selectivity. Benzene conversion reflects the catalytic activity of the catalyst for the hydrogenation of benzene, while cyclohexene selectivity reflects the directional catalytic ability of the catalyst for the target product cyclohexene. Both are core chemical indicators for evaluating the performance of the catalyst in the hydrogenation of benzene. This represents the weight matrix of the second fully connected regression layer, which is specifically used to capture chemical features related to catalytic activity and selectivity in the comprehensive features, such as the strength of ruthenium-zinc interaction and crystal structure integrity. This represents the bias vector of the second fully connected regression layer, which is used to correct linear transformation bias and improve performance prediction accuracy. This represents the comprehensive feature representation of the input. The system uses this linear transformation to map the comprehensive feature representation into a catalytic performance prediction result directly related to the benzene hydrogenation reaction, generating a benzene hydrogenation performance prediction result that includes the percentage of benzene conversion and the percentage of cyclohexene selectivity.

[0140] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0141] Based on the same inventive concept, this application also provides a deep learning-based system for analyzing and predicting the activity and performance of ruthenium-zinc-benzene hydrogenation catalysts, used to implement the aforementioned deep learning-based method for analyzing and predicting the activity and performance of ruthenium-zinc-benzene hydrogenation catalysts. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more deep learning-based systems for analyzing and predicting the activity and performance of ruthenium-zinc-benzene hydrogenation catalysts provided below can be found in the limitations of the deep learning-based method for analyzing and predicting the activity and performance of ruthenium-zinc-benzene hydrogenation catalysts described above, and will not be repeated here.

[0142] Preferably, such as Figure 2 As shown, this invention provides a deep learning-based system 600 for analyzing the activity and predicting the performance of ruthenium zinc benzene hydrogenation catalysts. This system is configured with the following modules:

[0143] The spectrum preprocessing module 610 is used to perform image enhancement, noise reduction, and spectral normalization and alignment on the HRTEM image data of the catalyst sample obtained by high-resolution transmission electron microscopy and the XRD spectral data of the catalyst sample obtained by X-ray diffractometer, and to generate preprocessed image data and standard spectral data.

[0144] The image feature extraction module 620 is used to call the deep learning encoder to extract the deep semantic features of ruthenium nanoparticles in the preprocessed image data, and call the decoder to reconstruct the probability distribution map representing the outline of ruthenium nanoparticles, generating an image feature sequence containing particle morphology and spatial distribution information.

[0145] The spectral feature mapping module 630 is used to capture local peak shape and overall spectral line features in standard spectral data based on a convolutional neural network, and map the local peak shape and overall spectral line features into low-dimensional dense vectors that characterize crystal structure and zinc phase information to generate spectral feature vectors.

[0146] The feature fusion coding module 640 is used to dynamically focus the image features in the image feature sequence on the spatial region most relevant to the current spectral information in the spectral feature vector based on the attention mechanism, and integrate the image feature sequence and the spectral feature vector into a unified feature space to generate a comprehensive feature representation;

[0147] The catalytic performance prediction module 650 is used for parallel regression prediction based on a comprehensive feature representation. It calls one branch regression layer to map the comprehensive feature representation to predict the average size and dispersion of ruthenium nanoparticles, and another branch regression layer to map the comprehensive feature representation to predict the conversion rate and cyclohexene selectivity of the benzene hydrogenation reaction, generating a comprehensive prediction result. This comprehensive prediction result indicates the microstructural state and macroscopic catalytic performance of the catalyst sample.

[0148] Preferably, the spectrum preprocessing module 610 provided in this application is configured with the following units:

[0149] The image contrast enhancement unit is used to perform contrast-limited adaptive histogram equalization processing on the HRTEM image data of the catalyst sample obtained by high-resolution transmission electron microscopy, enhance the gray-level difference between ruthenium nanoparticles and the support background, and generate a contrast-enhanced image.

[0150] The image noise reduction processing unit is used to perform non-local mean noise reduction on the contrast-enhanced image, suppressing Gaussian noise and speckle noise in the image while preserving grain edge details, and generating preprocessed image data.

[0151] The spectral normalization processing unit is used to perform minimum-maximum normalization processing on the XRD spectral data of the catalyst sample obtained by the X-ray diffractometer, scaling the original intensity values ​​to a uniform numerical range and generating normalized spectra.

[0152] The spectral peak alignment unit is used to perform dynamic time warping-based peak alignment on the normalized spectrum, matching and calibrating the diffraction angle positions of the normalized spectrum with those of the standard substance spectrum to generate standard spectral data.

[0153] Preferably, the image feature extraction module 620 provided in this application is configured with the following units:

[0154] The deep feature extraction unit is used to perform multi-level convolution and pooling downsampling on preprocessed image data based on the encoder-decoder network architecture. It calls the deep residual network to extract shallow geometric features containing basic edges and textures and high-level abstract features containing the overall shape and semantic category of particles from the image, generating a deep feature map containing multi-level semantic information.

[0155] The feature fusion upsampling unit is used to progressively upsample and stitch features into deep feature maps. Through skip connections, it gradually fuses and restores the resolution of shallow detail features containing particle edges and textures from early levels of the encoder path with high-level semantic features containing overall particle shape and category information from deep feature maps, generating high-resolution feature maps.

[0156] The particle probability map generation unit is used to perform pixel-by-pixel classification processing on the high-resolution feature map. It calculates and outputs the probability value of each pixel belonging to ruthenium nanoparticles through convolutional layers and nonlinear activation functions, and generates a binary probability distribution map that characterizes the contour of ruthenium nanoparticles.

[0157] The image feature sequence generation unit is used for feature selection and serialization based on high-resolution feature maps and probability distribution maps. It generates an image feature sequence focused on the foreground particle region by using the probability distribution map as a spatial weight to apply attention weight to the high-resolution feature map and flattening the weighted features along the spatial dimension.

[0158] Preferably, the spectral feature mapping module 630 provided in this application is configured with the following units:

[0159] The multi-scale spectral convolution unit is used to perform multi-scale one-dimensional convolution filtering on standard spectral data. It uses three one-dimensional convolution kernels with widths of 3, 5, and 7 to capture the sharp diffraction peak features, medium-width diffraction peak features, and broad and gentle diffraction envelope features in the spectrum, respectively, to generate multi-scale spectral feature maps.

[0160] The spectral fusion pooling unit is used to perform feature fusion and pooling on multi-scale spectral feature maps. It splices sharp diffraction peak features, medium-width diffraction peak features, and broad and gentle diffraction envelope features through the channel dimension, and performs global average pooling on the spliced ​​feature map to aggregate global spectral information and generate a fused spectral feature vector.

[0161] The spectral feature dimensionality reduction unit is used to perform nonlinear transformation and dimensionality reduction on the fused spectral feature vector. Through a fully connected layer and the ReLU activation function, the fused spectral feature vector is mapped to a latent space with significantly reduced dimensionality, generating a low-dimensional dense vector that characterizes the crystal structure and zinc phase information as the spectral feature vector.

[0162] Preferably, the feature fusion coding module 640 provided in this application is configured with the following units:

[0163] The spectral feature projection unit is used to perform linear projection processing on the spectral feature vector. Based on the linear transformation layer in the cross-attention mechanism, the spectral feature vector is mapped to the key feature space used to calculate similarity and the value feature space used to transmit information, respectively, to generate key vectors and value vectors.

[0164] The attention weight calculation unit is used to perform similarity calculation on the image feature sequence and the key vector. It measures the correlation between the feature and the key vector at each spatial location in the image feature sequence through dot product operation and generates the attention weight distribution.

[0165] The modal feature enhancement unit is used to perform weighted aggregation processing on the value vector based on the attention weight distribution, and inject the spectral information carried by the value vector into the corresponding image feature space location according to the intensity of the attention weight distribution, thereby generating an enhanced image feature sequence after modal interaction.

[0166] The comprehensive feature aggregation unit is used to perform global aggregation processing on the enhanced image feature sequence. It averages the features at all locations along the spatial dimension through average pooling, compresses the spatial dimension of the feature sequence, and generates a comprehensive feature representation.

[0167] Preferably, the catalytic performance prediction module 650 provided in this application is configured with the following units:

[0168] The model parameter training unit is used to train the model on a training dataset consisting of multiple catalyst samples whose true active phase parameters and benzene hydrogenation performance are known. Based on the HRTEM images, XRD spectral data and the corresponding true measured values ​​of ruthenium nanoparticle average size, dispersion, benzene conversion rate and cyclohexene selectivity in the training dataset, the parameters of the preset multi-channel deep learning model are optimized by the gradient descent algorithm to establish a nonlinear mapping relationship from multimodal data to target labels and generate a pre-trained multi-channel deep learning model.

[0169] The active phase parameter prediction unit is used to perform first branch regression processing on the comprehensive feature representation. The comprehensive feature representation is input into the first fully connected regression layer in the pre-trained multi-channel deep learning model corresponding to the active phase parameter prediction task. The comprehensive feature representation is mapped to the predicted average size and dispersion score of ruthenium nanoparticles through the weights and bias parameters learned by the first fully connected regression layer, and the active phase analysis results are generated.

[0170] The catalytic performance prediction unit is used to perform second-branch regression processing on the comprehensive feature representation. The comprehensive feature representation is input into the second fully connected regression layer in the pre-trained multi-channel deep learning model corresponding to the catalytic performance prediction task. The comprehensive feature representation is mapped to the predicted percentage of benzene conversion and the percentage of cyclohexene selectivity through the weights and bias parameters learned by the second fully connected regression layer, generating the benzene hydrogenation performance prediction result.

[0171] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst.

[0172] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described deep learning-based method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst.

[0173] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0174] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0175] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for activity analysis and performance prediction of ruthenium zinc benzene hydrogenation catalyst based on deep learning, characterized in that, Includes the following steps: S1: Perform image enhancement, noise reduction, and spectral normalization alignment on the HRTEM image data of the catalyst samples obtained by high-resolution transmission electron microscopy and the XRD spectral data of the catalyst samples obtained by X-ray diffraction to generate preprocessed image data and standard spectral data. S2: Call the deep learning encoder to extract the deep semantic features of ruthenium nanoparticles in the preprocessed image data, and call the decoder to reconstruct the probability distribution map representing the outline of ruthenium nanoparticles, generating an image feature sequence containing particle morphology and spatial distribution information; S3: Based on a convolutional neural network, capture the local peak shape and overall spectral line features in the standard spectral data, and map the local peak shape and overall spectral line features into a low-dimensional dense vector characterizing the crystal structure and zinc phase information to generate a spectral feature vector; S4: Based on the attention mechanism, the image features in the image feature sequence are dynamically focused on the spatial region most relevant to the current spectral information in the spectral feature vector, and the image feature sequence and the spectral feature vector are integrated into a unified feature space to generate a comprehensive feature representation; S5: Based on the comprehensive feature representation, perform parallel regression prediction, call a branch regression layer to map the comprehensive feature representation to predict the average size and dispersion of ruthenium nanoparticles, and use another branch regression layer to map the comprehensive feature representation to predict the conversion rate and cyclohexene selectivity of the benzene hydrogenation reaction, and generate comprehensive prediction results. The comprehensive prediction results are used to indicate the microstructure state and macroscopic catalytic performance of the catalyst sample.

2. The method according to claim 1, characterized in that, S1 includes: S11: Perform contrast-limited adaptive histogram equalization on the HRTEM image data of the catalyst sample obtained by high-resolution transmission electron microscopy to enhance the gray-level difference between ruthenium nanoparticles and the support background, and generate a contrast-enhanced image. S12: Perform nonlocal mean denoising on the contrast-enhanced image to suppress Gaussian noise and speckle noise in the image while preserving particle edge details, and generate the preprocessed image data. S13: Perform minimum-maximum normalization on the XRD spectral data of the catalyst sample obtained by X-ray diffractometer to scale the original intensity values ​​to a uniform numerical range and generate normalized spectra. S14: Perform peak alignment based on dynamic time warping on the normalized spectrum, match and calibrate the diffraction angle positions of the normalized spectrum and the standard substance spectrum, and generate standard spectral data.

3. The method according to claim 1, characterized in that, S2 includes: S21: Based on the encoder-decoder network architecture, the preprocessed image data is subjected to multi-level convolution and pooling downsampling processing. The deep residual network is called to extract shallow geometric features containing basic edges and textures and high-level abstract features containing the overall shape and semantic category of particles from the image, generating a deep feature map containing multi-level semantic information. S22: The deep feature map is progressively upsampled and feature stitched. Through skip connections, shallow detail features from the early levels of the encoder path, which contain particle edges and textures, are fused and their resolution restored with high-level semantic features from the deep feature map, which contain overall particle shape and category information, to generate a high-resolution feature map. S23: Perform pixel-by-pixel classification processing on the high-resolution feature map, calculate and output the probability value of each pixel belonging to ruthenium nanoparticles through convolutional layers and nonlinear activation functions, and generate a binary probability distribution map characterizing the contour of ruthenium nanoparticles. S24: Based on the high-resolution feature map and the probability distribution map, feature filtering and serialization processing are performed. The high-resolution feature map is then attention-weighted by using the probability distribution map as a spatial weight, and the weighted features are flattened along the spatial dimension to generate an image feature sequence focused on the foreground particle region.

4. The method according to claim 1, characterized in that, S3 includes: S31: Perform multi-scale one-dimensional convolution filtering on the standard spectral data, using three types of one-dimensional convolution kernels with widths of 3, 5, and 7 to capture the sharp diffraction peak features, medium-width diffraction peak features, and broad and gentle diffraction envelope features in the spectrum, respectively, to generate a multi-scale spectral feature map. S32: Perform feature fusion and pooling processing on the multi-scale spectral feature map, and stitch the sharp diffraction peak features, the medium-width diffraction peak features, and the wide and gentle diffraction envelope features through the channel dimension. Then, perform global average pooling on the stitched feature map to aggregate global spectral information and generate a fused spectral feature vector. S33: The fused spectral feature vector is subjected to nonlinear transformation and dimensionality reduction processing. The fused spectral feature vector is mapped to a latent space with significantly reduced dimensionality through a fully connected layer and the ReLU activation function, generating a low-dimensional dense vector characterizing the crystal structure and zinc phase information as the spectral feature vector.

5. The method according to claim 1, characterized in that, S4 includes: S41: Perform linear projection processing on the spectral feature vector, and map the spectral feature vector to the key feature space for calculating similarity and the value feature space for transmitting information based on the linear transformation layer in the cross attention mechanism, respectively, to generate key vector and value vector; S42: Perform similarity calculation on the image feature sequence and the key vector, measure the correlation between the features at each spatial location in the image feature sequence and the key vector through dot product operation, and generate attention weight distribution; S43: Based on the attention weight distribution, the value vector is weighted and aggregated, and the spectral information carried by the value vector is injected 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; S44: Perform global aggregation processing on the enhanced image feature sequence, and average the features at all locations along the spatial dimension by average pooling to compress the spatial dimension of the feature sequence and generate a comprehensive feature representation.

6. The method according to claim 5, characterized in that, The formula for calculating the attention weight distribution is: in, For attention weight distribution, Let K be the sequence of image features used as the query, and K be the key matrix formed by copying and expanding the key vectors. Let T be the feature dimension of the key vector, T denote the matrix transpose, and softmax be the normalization exponential function.

7. The method according to any one of claims 1-6, characterized in that, S5 includes: S51: The training dataset consisting of multiple catalyst samples with known real active phase parameters and benzene hydrogenation performance is subjected to model training processing. Based on the HRTEM images, XRD spectral data and the corresponding real measured values ​​of ruthenium nanoparticle average size, dispersion, benzene conversion rate and cyclohexene selectivity in the training dataset, the parameters of the preset multi-channel deep learning model are optimized by gradient descent algorithm to establish a nonlinear mapping relationship from multimodal data to target labels and generate a pre-trained multi-channel deep learning model. S52: Perform first branch regression processing on the comprehensive feature representation, input the comprehensive feature representation into the first fully connected regression layer corresponding to the active phase parameter prediction task in the pre-trained multi-channel deep learning model, and map the comprehensive feature representation into the predicted average size and dispersion score of ruthenium nanoparticles through the weights and bias parameters learned by the first fully connected regression layer, thereby generating active phase analysis results; S53: Perform second branch regression processing on the comprehensive feature representation, input the comprehensive feature representation into the second fully connected regression layer corresponding to the catalytic performance prediction task in the pre-trained multi-channel deep learning model, and map the comprehensive feature representation into the predicted benzene conversion percentage and cyclohexene selectivity percentage through the weights and bias parameters learned by the second fully connected regression layer, thereby generating the benzene hydrogenation performance prediction result.

8. A deep learning-based system for analyzing the activity and predicting the performance of ruthenium-zinc benzene hydrogenation catalysts, characterized in that, The system includes: The spectrum preprocessing module is used to perform image enhancement, noise reduction, and spectral normalization and alignment on the HRTEM image data of catalyst samples obtained by high-resolution transmission electron microscopy and the XRD spectral data of catalyst samples obtained by X-ray diffractometer, and to generate preprocessed image data and standard spectral data. The image feature extraction module is used to call the deep learning encoder to extract the deep semantic features of ruthenium nanoparticles in the preprocessed image data, and call the decoder to reconstruct the probability distribution map representing the outline of ruthenium nanoparticles, generating an image feature sequence containing particle morphology and spatial distribution information. The spectral feature mapping module is used to capture local peak shapes and overall spectral line features in the standard spectral data based on a convolutional neural network, and map the local peak shapes and overall spectral line features into low-dimensional dense vectors that characterize crystal structure and zinc phase information, thereby generating spectral feature vectors. The feature fusion encoding module is used to dynamically focus the image features in the image feature sequence on the spatial region most relevant to the current spectral information in the spectral feature vector based on the attention mechanism, and integrate the image feature sequence and the spectral feature vector into a unified feature space to generate a comprehensive feature representation; The catalytic performance prediction module is used to perform parallel regression prediction based on the comprehensive feature representation. It calls a branch regression layer to map the comprehensive feature representation to predict the average size and dispersion of ruthenium nanoparticles, and uses another branch regression layer to map the comprehensive feature representation to predict the conversion rate and cyclohexene selectivity of the benzene hydrogenation reaction, generating a comprehensive prediction result. The comprehensive prediction result is used to indicate the microstructure state and macroscopic catalytic performance of the catalyst sample.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images