AI-based Retrieval of Chemical Material Detection Information Database and Standard Spectrum Comparison System

Abstract

The present invention relates to the technical field of chemical material detection, and discloses a retrieval and standard spectrum comparison system for chemical material detection information database based on AI. The system includes data acquisition, spectral preprocessing, feature extraction, retrieval matching, and dynamic comparison modules. The data acquisition module uses a multispectral sensor to collect real-time spectral data covering ultraviolet, visible, and infrared bands, and denoises it through the spectral preprocessing module. The feature extraction module uses a deep residual convolutional network to extract spectral fingerprint feature vectors. The retrieval matching module inputs the feature vectors into a graph structure retrieval model to screen candidate standard spectra. The dynamic comparison module performs multi-dimensional comparison of the candidate standard spectra based on an improved quantum particle swarm optimization algorithm to obtain the optimal result. The system can accurately collect and process spectral data, quickly retrieve and accurately compare standard spectra, improve the efficiency and accuracy of chemical material detection, and has a wide application prospect in the fields of chemical material research, production quality control, etc.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical material detection, and specifically to an AI-based chemical material detection information database retrieval and standard spectrum comparison system. Background Art

[0002] In many aspects such as chemical material research, production, and quality control, accurately identifying the composition and structure of chemical materials is crucial. Traditional chemical material detection methods mainly rely on manual operations and simple instrumental analysis, and expose many drawbacks when facing complex and diverse chemical materials.

[0003] In the early stage, chemical analysis mainly used classical chemical methods such as titration and gravimetric analysis. Although these methods have a certain degree of accuracy, the operation process is cumbersome, and it requires professional technicians to spend a lot of time and energy. For example, when performing quantitative analysis of multiple components in a complex sample, multiple chemical reactions and separation operations are required, which is not only inefficient but also prone to introducing errors. With the development of science and technology, instrumental analysis methods have gradually emerged, such as spectral analysis and chromatographic analysis. These methods have improved the detection efficiency and accuracy to a certain extent, but there are still limitations.

[0004] Taking spectral analysis as an example, for common spectral detection instruments such as ultraviolet spectrophotometers and infrared spectrometers, when detecting chemical materials, the collected spectral data often contains a large amount of noise. The existence of noise not only interferes with the judgment of the true spectral characteristics of chemical materials but also may lead to misjudgment. Moreover, the data collected by different instruments vary in format, wavelength range, etc., making it difficult to integrate and analyze the data. When comparing the detected spectrum with the standard spectrum, due to the lack of effective algorithms and models, only simple qualitative comparisons can be carried out, and it is difficult to achieve precise quantitative analysis.

[0005] In terms of information management, with the continuous increase in the types of chemical materials, the related detection data and standard spectrum data have grown explosively. Traditional database management systems are difficult to efficiently store, retrieve, and utilize these massive data. When researchers search for the standard spectrum of a specific chemical material, they often need to spend a lot of time screening through numerous data, with extremely low efficiency. In addition, the existing retrieval and comparison systems lack intelligence and cannot be adaptively adjusted according to the actual detection situation, making it difficult to meet the requirements of modern chemical material detection for speed, accuracy, and intelligence.

[0006] In industrial production, the quality control of chemical materials is directly related to the quality and safety of products. For example, in industries such as pharmaceuticals, food, and electronics, extremely high requirements are placed on the purity, composition, etc. of chemical materials. If errors or omissions occur during the detection process, serious consequences may result. In the pharmaceutical industry, the presence of impurities in drugs may affect the efficacy of the drugs and even pose a hazard to the health of patients. Therefore, it is urgent to develop an efficient, accurate, and intelligent chemical material detection information library retrieval and standard spectrum comparison system. It can not only improve the efficiency and accuracy of chemical material detection but also provide strong technical support for the research, production, and quality control of chemical materials. Summary of the Invention

[0007] The purpose of the present invention is to provide an AI-based chemical material detection information library retrieval and standard spectrum comparison system to solve the problems raised in the above-mentioned background technology.

[0008] To achieve the above purpose, the present invention provides the following technical solution: an AI-based chemical material detection information library retrieval and standard spectrum comparison system, the system includes:

[0009] Data acquisition module: used to collect real-time spectral data of chemical materials through a multispectral sensor, and the multispectral sensor covers ultraviolet, visible, and infrared bands;

[0010] Spectral preprocessing module: based on the adaptive wavelet threshold algorithm, perform multi-scale decomposition and noise elimination on the real-time spectral data to generate a denoised spectral sequence;

[0011] Feature extraction module: use a deep residual convolutional network to perform multi-level feature extraction on the denoised spectral sequence to generate a spectral fingerprint feature vector;

[0012] Retrieval and matching module: input the spectral fingerprint feature vector into a pre-constructed graph structure retrieval model, and screen candidate standard spectra from the chemical material information library based on the node similarity measurement algorithm;

[0013] Dynamic comparison module: perform multi-dimensional comparison on the candidate standard spectra based on an improved quantum particle swarm optimization algorithm to generate an optimal matching result, and the improved quantum particle swarm optimization algorithm integrates dynamic inertia weight adjustment and crossover mutation strategies.

[0014] Preferably, the collection of real-time spectral data of chemical materials through the multispectral sensor includes:

[0015] The multispectral sensor includes an ultraviolet spectrophotometer, a Fourier transform infrared spectrometer, and a Raman spectroscopy probe;

[0016] Align the bands of the ultraviolet spectrophotometer data and the Fourier transform infrared spectroscopy data to construct a full-band spectral curve; perform baseline correction on the Raman spectroscopy probe data to eliminate the interference of the fluorescence background;

[0017] Construct a two-channel spectral fusion network. The first channel uses a one-dimensional convolutional kernel to extract the global features of the full-band spectral curve, and the second channel uses a gated recurrent unit to extract the local peak position features of the Raman spectroscopy data;

[0018] Fuse the global features and the local peak position features through a multi-scale attention mechanism to generate a joint spectral feature map; perform temporal modeling on the joint spectral feature map based on a bidirectional long short-term memory network, and output a denoised spectral sequence containing chemical bond vibration modes and functional group information.

[0019] Preferably, the graph structure retrieval model adopts a dynamic attribute graph neural network structure and realizes similarity measurement based on a node embedding algorithm; the dynamic attribute graph neural network structure includes:

[0020] Construct a chemical material-functional group association graph. The nodes in the graph include compound nodes, functional group nodes, and spectral feature nodes, and the node attributes include molecular weight, polarity index, and absorbance;

[0021] Adopt a two-stage attention mechanism. In the first stage, calculate the association weights between compound nodes and adjacent functional group nodes through a spatial graph convolutional layer, and in the second stage, perform importance screening on historical matching records through a temporal graph convolutional layer;

[0022] Iteratively update the node attributes based on a multi-head graph attention module. Each attention head fuses the chemical bond type and the solvent environment parameters; stabilize the training process through a skip connection and a layer normalization mechanism, and finally output a candidate standard spectrum containing the spectral peak offset tolerance.

[0023] Preferably, the improved quantum particle swarm optimization algorithm integrating dynamic inertia weight adjustment and crossover mutation strategy includes:

[0024] Model the comparison problem as a multi-objective optimization problem, and the decision variables include the spectral peak position offset, the peak intensity weight coefficient, and the baseline fitting parameters;

[0025] Initialize the quantum particle swarm and calculate the fitness function, which includes a spectral shape similarity term, a peak position alignment term, and a noise robustness term;

[0026] In the inertia weight adjustment stage, dynamically adjust the global search and local development capabilities of the particles according to the number of iterations; in the crossover mutation stage, introduce a chaotic map to generate the mutation direction;

[0027] The particle swarm is distributedly optimized using a parallel computing framework. The particle positions are updated in each iteration, and the global optimal solution is synchronized through the elitist retention strategy.

[0028] Preferably, the deep residual convolutional network adopts a dilated residual block structure to accelerate feature extraction, including:

[0029] The denoised spectral sequence is input into the pyramid convolutional layer, and multi-scale spectral features are extracted through dilated convolutional kernels with different dilation rates;

[0030] In the residual connection stage, an adaptive channel weighting mechanism is adopted to fuse shallow and deep features;

[0031] A spatial attention module is introduced to recalibrate the feature map and suppress the interference of irrelevant spectral bands.

[0032] Preferably, the adaptive wavelet threshold algorithm includes: decomposing the original spectral data into approximation coefficients and detail coefficients, and dynamically adjusting the threshold parameters according to the noise energy at each scale.

[0033] Preferably, the spatial graph convolutional layer adopts a relative chemical distance encoding mechanism, including:

[0034] Defining the relative chemical distance vector of compound nodes and functional group nodes, including electronegativity difference, bond length similarity, and steric hindrance coefficient;

[0035] Mapping the relative chemical distance vector to the bias parameter of the graph convolutional kernel through a non-linear transformation layer;

[0036] Superimposing the bias parameter on the standard graph convolutional operation to enhance the modeling ability of functional group correlation relationships.

[0037] Preferably, the chaotic mapping to generate the mutation direction includes:

[0038] Using the Logistic mapping to generate a chaotic sequence and mapping the chaotic variable to the particle position update direction;

[0039] Designing a dynamic scaling factor to adjust the mutation step size according to the particle distribution density;

[0040] When a decrease in population diversity is detected, a chaotic re-initialization operation is triggered.

[0041] Preferably, the spatial attention module is implemented through spectral domain pooling, including:

[0042] Dividing the feature map along the wavelength dimension into multiple sub-intervals, calculating the energy distribution entropy of each sub-interval, generating an attention weight matrix according to the energy distribution entropy, and enhancing the features of the high-entropy regions;

[0043] Adjusting the receptive field shape through a deformable convolutional kernel to adapt to the non-uniform spectral peak distribution.

[0044] Preferably, the adaptive wavelet threshold algorithm further includes:

[0045] Dividing the denoised spectral sequence into overlapping sub-windows, and calculating the signal-to-noise ratio and peak position stability of each sub-window;

[0046] Dynamically adjusting the wavelet decomposition level according to the sub-window verification results. When it is detected that the reconstruction error of consecutive sub-windows exceeds the standard, a threshold parameter adaptive reset mechanism is triggered.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] In the data acquisition and processing link, the multispectral sensor covers the ultraviolet, visible and infrared bands, and can comprehensively collect the real-time spectral data of chemical materials. Among them, the ultraviolet spectrophotometer, Fourier transform infrared spectrometer and Raman spectroscopy probe work together to obtain the characteristic information of different bands respectively. Preprocessing operations such as band alignment and baseline correction are performed on the data of each instrument, and a joint spectral feature map is generated through a dual-channel spectral fusion network and a multi-scale attention mechanism, and then temporal modeling is performed through a bidirectional long short-term memory network, effectively extracting the chemical bond vibration modes and functional group information in chemical materials. This process can obtain richer and more accurate chemical material information compared with the traditional single-spectral detection method, greatly improving the quality and usability of the data, and reducing misjudgments caused by incomplete or inaccurate data.

[0049] The spectral preprocessing module performs multi-scale decomposition and noise elimination on the real-time spectral data based on the adaptive wavelet threshold algorithm. This algorithm dynamically adjusts the threshold parameter according to the noise energy of each scale, can accurately remove noise, and at the same time retains the useful information of the spectrum to the greatest extent. Compared with the traditional fixed-threshold denoising method, the adaptive wavelet threshold algorithm can show better denoising effects in different noise environments, making the generated denoised spectral sequence clearer, and providing a reliable data basis for subsequent feature extraction and comparison analysis.

[0050] The feature extraction module adopts a deep residual convolutional network, especially the dilated residual block structure, which greatly accelerates the feature extraction process. The pyramid convolutional layer extracts multi-scale spectral features through dilated convolutional kernels with different dilation rates, the adaptive channel weighting mechanism fuses shallow and deep features, and the spatial attention module suppresses the interference of irrelevant spectral bands. This multi-level feature extraction method can more accurately capture the unique fingerprint features of the chemical material spectrum, and the generated spectral fingerprint feature vector contains rich chemical structure information, improving the accuracy and reliability of chemical material identification.

[0051] The retrieval matching module utilizes a pre - constructed graph - structured retrieval model, and adopts a dynamic attribute graph neural network structure and a node embedding algorithm to achieve similarity measurement. By constructing a chemical material - functional group association graph, combining a two - stage attention mechanism and a multi - head graph attention module to iteratively update node attributes, a candidate standard spectrum containing spectral peak offset tolerance is finally output. This retrieval model can fully explore the internal associations between chemical materials. Compared with traditional retrieval methods, it can more quickly and accurately screen out candidate standard spectra from the chemical material information database, greatly improving the retrieval efficiency and accuracy.

[0052] The dynamic comparison module is based on an improved quantum particle swarm optimization algorithm, which models the comparison problem as a multi - objective optimization problem. It integrates a dynamic inertia weight adjustment and a crossover - mutation strategy, dynamically adjusts the global search and local development capabilities of particles according to the number of iterations, introduces a chaos mapping to generate the mutation direction, and uses a parallel computing framework for distributed optimization. This algorithm can perform a more comprehensive and detailed comparison of candidate standard spectra in multiple dimensions, generate the optimal matching result, effectively improving the accuracy and reliability of the comparison between chemical materials and standard spectra. At the same time, the application of the parallel computing framework and the elite retention strategy greatly shortens the comparison time and improves the overall operation efficiency of the system. Brief Description of the Drawings

[0053] Figure 1 is the working principle diagram of the AI - based chemical material detection information database retrieval and standard spectrum comparison system described in the present invention;

[0054] Figure 2 is the working flowchart of multi - spectral sensor data acquisition and processing;

[0055] Figure 3 is the working flowchart of the improved quantum particle swarm optimization algorithm;

[0056] Figure 4 is the working flowchart of generating the mutation direction by chaos mapping. Detailed Embodiments

[0057] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0058] Please refer to Figures 1-4 , the present invention provides an AI - based chemical material detection information database retrieval and standard spectrum comparison system, and its overall implementation solution is as follows:

[0059] It mainly includes a data acquisition module, a spectral preprocessing module, a feature extraction module, a retrieval and matching module, and a dynamic comparison module. First, the data acquisition module is used to collect the real-time spectral data of chemical materials through a multispectral sensor that covers the ultraviolet, visible, and infrared bands. The collected real-time spectral data contains rich chemical material information, which serves as the basis for subsequent analysis.

[0060] The spectral preprocessing module performs multi-scale decomposition and noise elimination on the real-time spectral data based on the adaptive wavelet threshold algorithm to generate a denoised spectral sequence. Since the collected spectral data may be interfered by various noises, through this preprocessing step, the noise can be effectively removed, the quality of the spectral data can be improved, and more reliable data can be provided for subsequent analysis.

[0061] The feature extraction module uses a deep residual convolutional network to perform multi-level feature extraction on the denoised spectral sequence to generate a spectral fingerprint feature vector. The deep residual convolutional network can automatically learn the complex features in the spectral data and extract representative spectral fingerprint feature vectors, which can accurately reflect the spectral characteristics of chemical materials.

[0062] The retrieval and matching module inputs the spectral fingerprint feature vector into a pre-constructed graph structure retrieval model, and screens candidate standard spectra from the chemical material information database based on the node similarity measurement algorithm. The graph structure retrieval model can efficiently process a large-scale chemical material information database and quickly screen out candidate standard spectra that are similar to the spectral characteristics of the chemical material to be detected through the node similarity measurement algorithm.

[0063] The dynamic comparison module performs multi-dimensional comparison on the candidate standard spectra based on the improved quantum particle swarm optimization algorithm to generate an optimal matching result. The improved quantum particle swarm optimization algorithm integrates dynamic inertia weight adjustment and crossover mutation strategies. Through this algorithm, the candidate standard spectra can be carefully compared in multiple dimensions, various factors can be comprehensively considered, and finally the result that best matches the spectrum of the chemical material to be detected can be obtained.

[0064] The following further illustrates the implementation of the present invention in combination with Embodiments 1 to 5.

[0065] Embodiment 1:

[0066] In this embodiment, the specific process of the data acquisition module collecting the real-time spectral data of chemical materials is described in detail. The multispectral sensor consists of an ultraviolet spectrophotometer, a Fourier transform infrared spectrometer, and a Raman spectroscopy probe. The ultraviolet spectrophotometer is mainly used to measure the spectral information of chemical materials in the ultraviolet band, the Fourier transform infrared spectrometer focuses on obtaining the spectral data in the infrared band, and the Raman spectroscopy probe is used to collect Raman spectral information.

[0067] When collecting data, first align the bands of the ultraviolet spectrophotometer data and the Fourier transform infrared spectroscopy data to construct a full-band spectral curve. This step is to unify the spectral data obtained by different instruments in the wavelength dimension for subsequent analysis. Specifically, a specific algorithm is used to find the corresponding relationship between the two spectral data in terms of wavelength, and then calibration is performed to enable them to be seamlessly connected, constructing a spectral curve covering the entire ultraviolet to infrared band.

[0068] For the Raman spectroscopy probe data, baseline correction is required to eliminate the interference of the fluorescence background. Since Raman spectroscopy is prone to being affected by the fluorescence background during the acquisition process, resulting in inaccurate spectral signals, baseline correction can be used to remove this interference. The principle is to fit the baseline of the spectral data and then subtract the original spectral data from the baseline to obtain the Raman spectral data after removing the fluorescence background interference.

[0069] Next, construct a two-channel spectral fusion network. The first channel uses a one-dimensional convolutional kernel to extract the global features of the full-band spectral curve. The one-dimensional convolutional kernel slides on the full-band spectral curve and extracts the overall features of the spectrum through convolutional operations, such as the general shape of the spectrum and the positions of the main absorption peaks. These global features can reflect the overall spectral characteristics of chemical materials. The second channel uses a gated recurrent unit to extract the local peak position features of the Raman spectral data. The gated recurrent unit can effectively capture the local features in time series data. In Raman spectral data, it can accurately identify the positions and intensities of each peak.

[0070] The global features and local peak position features are fused through a multi-scale attention mechanism to generate a joint spectral feature map. The multi-scale attention mechanism can automatically assign weights according to the importance of features at different scales, effectively fusing the global features and local peak position features. Specifically, it calculates the attention weights of features at different scales, giving higher weights to important features, thereby highlighting the role of these features in the joint spectral feature map. Finally, based on a bidirectional long short-term memory network, temporal modeling is performed on the joint spectral feature map to output a denoised spectral sequence containing chemical bond vibration modes and functional group information. The bidirectional long short-term memory network can consider both the forward and backward information of the spectral data. By performing temporal modeling on the joint spectral feature map, it can discover the laws of spectral data changes over time (or wavelength), and then output a denoised spectral sequence containing chemical bond vibration modes and functional group information. These information are crucial for accurately analyzing the composition and structure of chemical materials in the subsequent process.

[0071] Example 2:

[0072] This embodiment focuses on the specific implementation of the adaptive wavelet threshold algorithm in the spectral preprocessing module. The core of the adaptive wavelet threshold algorithm is to decompose the original spectral data into approximation coefficients and detail coefficients, and dynamically adjust the threshold parameters according to the noise energy at each scale.

[0073] Let the original spectral data be , where represents time (which can be understood as variables such as wavelength in spectral data). Through wavelet transform, is decomposed into approximation coefficient and detail coefficient , where represents the scale of wavelet decomposition. Wavelet transform is a mathematical transformation method that can decompose a signal at different scales. The approximation coefficient reflects the low-frequency component of the signal, that is, the general trend and main features in the spectral data; the detail coefficient reflects the high-frequency component of the signal, including the detail information and noise in the spectral data.

[0074] When calculating the threshold parameter, it needs to be dynamically adjusted according to the noise energy at each scale. Assume that at scale , the estimated value of the noise energy is , then the threshold can be calculated by the formula , where is a regulation factor used to control the size of the threshold, which is set according to experimental experience or specific application scenarios; represents the length of the spectral data. The principle of this formula is based on the statistical characteristics of noise in the wavelet transform domain, and determines the appropriate threshold by estimating the noise energy. When the noise energy is large, the threshold also increases accordingly, so that noise can be more effectively removed; conversely, when the noise energy is small, the threshold decreases to avoid over-removing useful spectral detail information.

[0075] In actual operation, first perform wavelet decomposition on the original spectral data to obtain approximation coefficients and detail coefficients at different scales. Then, calculate the threshold at each scale according to the above formula. For the detail coefficients, compare their absolute values with the threshold. The detail coefficients less than the threshold are considered as noise and set to 0; the detail coefficients greater than the threshold are subjected to shrinkage processing, that is, subtract the threshold and then multiply by a scale factor (usually 1). Finally, reconstruct the processed approximation coefficients and detail coefficients through inverse wavelet transform to obtain the denoised spectral sequence. In this way, while effectively removing noise, the useful information in the spectral data can be retained to the greatest extent, providing a high-quality data basis for subsequent feature extraction and analysis.

[0076] Example 3:

[0077] This embodiment elaborates in detail the process of accelerating feature extraction by using the dilated residual block structure in the deep residual convolutional network. When processing the denoised spectral sequence, the deep residual convolutional network adopts the dilated residual block structure.

[0078] First, the denoised spectral sequence is input into the pyramid convolutional layer, and multi-scale spectral features are extracted through dilated convolutional kernels with different dilation rates. Let the denoised spectral sequence be , and the dilated convolutional kernel be , where represents the dilation rate. In the pyramid convolutional layer, multiple dilated convolutional kernels with different dilation rates work in parallel. For the dilated convolutional kernel with a dilation rate of , its convolutional operation can be expressed as , where represents the convolution operation, and is the feature map after convolution by this dilated convolutional kernel. Dilated convolutional kernels with different dilation rates can capture spectral features at different scales. Dilated convolutional kernels with larger dilation rates can obtain more extensive spectral information, corresponding to larger-scale features; dilated convolutional kernels with smaller dilation rates focus on extracting local and finer spectral features.

[0079] In the residual connection stage, an adaptive channel weighting mechanism is adopted to fuse shallow and deep features. Assume the shallow feature map is , the deep feature map is , and the adaptive channel weighting mechanism learns the channel weight vector . The fused feature map can be calculated by the formula . Here is a vector with the same number of channels as the feature map, and each of its elements represents the weight of the corresponding channel, with the value range between . By adaptively learning the channel weights, the network can dynamically adjust the proportion of shallow and deep features in the fusion process according to the importance of different channel features, so as to better retain and fuse spectral feature information at different levels.

[0080] Finally, a spatial attention module is introduced to recalibrate the feature map and suppress interference from irrelevant spectral segments. The spatial attention module is realized through spectral domain pooling. The feature map is divided into multiple sub-intervals along the wavelength dimension, the energy distribution entropy of each sub-interval is calculated, and an attention weight matrix is generated according to the energy distribution entropy to enhance the features in the high-entropy region. Let the feature map be , which is divided into sub-intervals, and the energy distribution entropy of the -th sub-interval is . The calculation formula for the energy distribution entropy is , where Indicates the proportion of the energy of the th pixel in the th sub - interval to the total energy of the sub - interval. is the number of pixels in the sub - interval. According to the calculated energy distribution entropy , an attention weight matrix is generated to recalibrate the feature map , and the enhanced feature map is obtained. In this way, the features of important regions in the spectrum can be highlighted, the interference of irrelevant spectral bands can be suppressed, and the accuracy and effectiveness of feature extraction can be further improved.

[0081] Example 4:

[0082] This example focuses on the specific construction and working process of the graph - structure retrieval model. The graph - structure retrieval model adopts a dynamic attribute graph neural network structure and realizes similarity measurement based on the node embedding algorithm.

[0083] First, a chemical material - functional group association graph is constructed. The nodes in the graph include compound nodes, functional group nodes, and spectral feature nodes, and the node attributes include molecular weight, polarity index, and absorbance. The compound nodes represent different chemical materials, the functional group nodes correspond to various functional groups in the chemical materials, and the spectral feature nodes are associated with specific features in the spectral data. Attributes such as molecular weight, polarity index, and absorbance can describe the physical and chemical properties and spectral characteristics of chemical materials and functional groups, providing rich information for subsequent analysis.

[0084] A two - stage attention mechanism is adopted. In the first stage, the association weights between compound nodes and adjacent functional group nodes are calculated through a spatial graph convolutional layer. The spatial graph convolutional layer adopts a relative chemical distance encoding mechanism, defining the relative chemical distance vector between compound nodes and functional group nodes, including electronegativity difference , bond - length similarity , and steric hindrance coefficient . The relative chemical distance vector is mapped to the bias parameter of the graph convolution kernel through a non - linear transformation layer, and its mapping relationship can be expressed as , where represents the non - linear transformation function. The bias parameter is superimposed on the standard graph convolution operation to enhance the modeling ability of the functional group association relationship. The formula for the standard graph convolution operation is , where is the node feature matrix of the th layer, is the adjacency matrix after adding self - loops, is 's degree matrix, is the The weight matrix of the layer is the activation function. The bias parameter obtained by introducing the relative chemical distance encoding mechanism can adjust the graph convolution operation, enabling the model to more accurately capture the correlation between compound nodes and functional group nodes.

[0085] In the second stage, the importance of historical matching records is screened through the temporal graph convolutional layer. The historical matching records contain information from previous retrieval processes. By using the temporal graph convolutional layer, the time series features in these records can be mined, and the information that is significant for the current retrieval can be screened out, improving the accuracy and efficiency of the retrieval.

[0086] Based on the multi-head graph attention module, the node attributes are iteratively updated. Each attention head fuses the chemical bond type and the solvent environment parameters. The training process is stabilized through skip connections and layer normalization mechanisms, and finally, a candidate standard spectrum containing the spectral peak shift tolerance is output. The multi-head graph attention module can analyze and update the node attributes from multiple perspectives, improving the expressive ability of the model. Skip connections can avoid the problem of gradient disappearance, enabling the network to better learn deep features; the layer normalization mechanism normalizes the input of each layer, accelerating the convergence speed of the model and ensuring that the model can be stably trained and output an accurate candidate standard spectrum.

[0087] Example 5:

[0088] This example details the specific implementation process of the improved quantum particle swarm optimization algorithm that integrates dynamic inertia weight adjustment and crossover mutation strategies. The improved quantum particle swarm optimization algorithm models the alignment problem as a multi-objective optimization problem. The decision variables include the spectral peak position offset , the peak intensity weight coefficient and the baseline fitting parameter .

[0089] First, initialize the quantum particle swarm and calculate the fitness function. The fitness function includes the spectral shape similarity term , the peak position alignment term and the noise robustness term . The fitness function can be expressed as , where , , are the weight coefficients used to adjust the importance of different terms in the fitness function, and their values are set according to specific application scenarios and requirements. The spectral shape similarity term is used to measure the shape similarity between the spectrum to be detected and the candidate standard spectrum, and can be obtained by calculating a certain distance metric between the two, such as Euclidean distance or cosine similarity, etc.; the peak position alignment term Used to evaluate the matching degree of spectral peak positions, for example, it can be measured by calculating the deviation of spectral peak positions; the noise robustness term reflects the performance of the algorithm in the presence of noise. For example, it can be determined by calculating the stability of the spectrum after adding a certain amount of noise.

[0090] In the stage of adjusting the inertia weight, according to the number of iterations dynamically adjust the global search and local development capabilities of the particles. The inertia weight is adjusted by the formula , where and are the maximum and minimum values of the inertia weight respectively, is the maximum number of iterations. At the beginning of the iteration, the inertia weight is larger, and the particles have stronger global search capabilities, enabling them to explore in a larger solution space; as the number of iterations increases, the inertia weight gradually decreases, and the local development capabilities of the particles are enhanced, enabling them to search the current nearby solution space more finely and improving the convergence accuracy of the algorithm.

[0091] In the crossover and mutation stage, a chaotic map is introduced to generate the mutation direction. A Logistic map is used to generate a chaotic sequence, and the Logistic map formula is , where is the chaotic variable at the th iteration, is the control parameter, usually taking values between . The chaotic variable is mapped to the particle position update direction, and a dynamic scaling factor is designed to adjust the mutation step size according to the particle distribution density. When a decrease in population diversity is detected, a chaotic re-initialization operation is triggered. By introducing a chaotic map, the randomness of particle search is increased, avoiding the algorithm from falling into a local optimal solution.

[0092] A parallel computing framework is used to perform distributed optimization on the particle swarm. In each iteration, the particle positions are updated and the global optimal solution is synchronized through the elitist retention strategy. Under the parallel computing framework, multiple particles can perform calculations and updates simultaneously, greatly improving the running efficiency of the algorithm. The elitist retention strategy ensures that in each iteration, the current optimal solution is saved to avoid being lost in subsequent iterations, thus ensuring that the algorithm can converge to the global optimal solution or a result close to the global optimal solution, and finally realizing the multi-dimensional comparison of candidate standard spectra to generate the optimal matching result.

[0093] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0094] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based chemical material detection information library retrieval and standard spectrum comparison system, characterized in that, Comprising: Data acquisition module: used to collect real-time spectral data of chemical materials through a multispectral sensor, and the multispectral sensor covers the ultraviolet, visible, and infrared bands; Spectral preprocessing module: perform multi-scale decomposition and noise elimination on the real-time spectral data based on the adaptive wavelet threshold algorithm to generate a denoised spectral sequence; Feature extraction module: use a deep residual convolutional network to perform multi-level feature extraction on the denoised spectral sequence to generate a spectral fingerprint feature vector; Retrieval and matching module: input the spectral fingerprint feature vector into a pre-constructed graph structure retrieval model, and screen candidate standard spectra from the chemical material information library based on the node similarity measurement algorithm; Dynamic comparison module: perform multi-dimensional comparison on the candidate standard spectra based on the improved quantum particle swarm optimization algorithm to generate an optimal matching result, and the improved quantum particle swarm optimization algorithm integrates dynamic inertia weight adjustment and crossover mutation strategies; The graph structure retrieval model adopts a dynamic attribute graph neural network structure and realizes similarity measurement based on the node embedding algorithm; the dynamic attribute graph neural network structure includes: Construct a chemical material - functional group association graph, where the nodes in the graph include compound nodes, functional group nodes, and spectral feature nodes, and the node attributes include molecular weight, polarity index, and absorbance; Adopt a two-stage attention mechanism. In the first stage, calculate the association weights between compound nodes and adjacent functional group nodes through a spatial graph convolutional layer, and in the second stage, perform importance screening on historical matching records through a temporal graph convolutional layer; Iteratively update the node attributes based on the multi-head graph attention module, and each attention head fuses the chemical bond type and solvent environment parameters; stabilize the training process through the skip connection and layer normalization mechanism, and finally output candidate standard spectra including spectral peak offset tolerances; The improved quantum particle swarm optimization algorithm integrating dynamic inertia weight adjustment and crossover mutation strategies includes: Model the comparison problem as a multi-objective optimization problem, and the decision variables include spectral peak position offset, peak intensity weight coefficient, and baseline fitting parameters; Initialize the quantum particle swarm and calculate the fitness function, and the fitness function includes spectral shape similarity term, peak position alignment term, and noise robustness term; In the inertia weight adjustment stage, dynamically adjust the global search and local development capabilities of the particles according to the number of iterations; in the crossover mutation stage, introduce a chaotic map to generate the mutation direction; Adopt a parallel computing framework to perform distributed optimization on the particle swarm, update the particle positions in each iteration, and synchronize the global optimal solution through the elite retention strategy.

2. The chemical material detection information database retrieval and standard spectrum comparison system according to claim 1, characterized in that The collection of real-time spectral data of chemical materials through the multispectral sensor includes: The multispectral sensor includes an ultraviolet spectrophotometer, a Fourier transform infrared spectrometer, and a Raman spectroscopy probe; Align the bands of the ultraviolet spectrophotometer data and the Fourier transform infrared spectral data to construct a full-band spectral curve; perform baseline correction on the Raman spectroscopy probe data to eliminate fluorescence background interference; Construct a two-channel spectral fusion network. The first channel uses a one-dimensional convolutional kernel to extract the global features of the full-band spectral curve, and the second channel uses a gated recurrent unit to extract the local peak position features of the Raman spectral data; Fusing the global features and local peak position features through a multi-scale attention mechanism to generate a joint spectral feature map; performing temporal modeling on the joint spectral feature map based on a bidirectional long short-term memory network to output a denoised spectral sequence containing chemical bond vibration modes and functional group information.

3. The chemical material detection information database retrieval and standard spectrum comparison system according to claim 1, characterized in that The deep residual convolutional network adopts a dilated residual block structure to accelerate feature extraction, including: Inputting the denoised spectral sequence into a pyramid convolutional layer, and extracting multi-scale spectral features through dilated convolutional kernels with different dilation rates; In the residual connection stage, an adaptive channel weighting mechanism is adopted to fuse shallow and deep features; Introducing a spatial attention module to recalibrate the feature map and suppress interference from irrelevant spectral segments.

4. The chemical material detection information library retrieval and standard spectrum comparison system according to claim 2, characterized in that, The adaptive wavelet threshold algorithm includes: decomposing the original spectral data into approximation coefficients and detail coefficients, and dynamically adjusting the threshold parameter according to the noise energy at each scale.

5. The chemical material detection information library retrieval and standard spectrum comparison system according to claim 1, characterized in that, The spatial graph convolutional layer adopts a relative chemical distance encoding mechanism, including: Defining a relative chemical distance vector for compound nodes and functional group nodes, including electronegativity difference, bond length similarity, and steric hindrance coefficient; Mapping the relative chemical distance vector to the bias parameter of the graph convolutional kernel through a non-linear transformation layer; Superimposing the bias parameter on the standard graph convolutional operation to enhance the modeling ability of functional group correlation relationships.

6. The chemical material detection information library retrieval and standard spectrum comparison system according to claim 1, characterized in that The chaotic mapping generates the mutation direction, including: Using a Logistic mapping to generate a chaotic sequence and mapping the chaotic variable to the particle position update direction; Designing a dynamic scaling factor to adjust the mutation step size according to the particle distribution density; When a decrease in population diversity is detected, triggering a chaotic re-initialization operation.

7. The chemical material detection information library retrieval and standard spectrum comparison system according to claim 3, characterized in that, The spatial attention module is implemented through spectral domain pooling, including: Dividing the feature map along the wavelength dimension into multiple sub-intervals, calculating the energy distribution entropy of each sub-interval, generating an attention weight matrix according to the energy distribution entropy, and enhancing the features in the high-entropy region; Adjusting the receptive field shape through a deformable convolutional kernel to adapt to the non-uniform spectral peak distribution.

8. The chemical material detection information library retrieval and standard spectrum comparison system according to claim 4, characterized in that, The adaptive wavelet threshold algorithm also includes: Dividing the denoised spectral sequence into overlapping sub-windows, and calculating the signal-to-noise ratio and peak position stability of each sub-window; Dynamically adjusting the wavelet decomposition layer number according to the sub-window verification result, and triggering a threshold parameter adaptive reset mechanism when the reconstruction error of consecutive sub-windows exceeds the standard.

Images