An on-line monitoring system for oil refining process based on near infrared spectrum

Abstract

The application provides an oil refining process online monitoring system based on near-infrared spectroscopy, acquires the original spectrum of the oil refining material to be measured in the oil refining process from an instrument, and performs baseline correction and smoothing denoising to obtain a pretreated spectrum; the pretreated spectrum is segmented, and then the segmented spectrum is subjected to nonlinear correction to obtain multiple corrected spectral segments; after splicing all the corrected spectral segments, reconstructed spectral data aligned with the main instrument spectral space are obtained; the reconstructed spectral data are subjected to segmented feature extraction, and the prediction values of various key quality indexes in the oil refining material to be measured are determined according to all the extracted local spectral features; and the control signals for adjusting the operating parameters of the corresponding oil refining device are generated according to the prediction values of each key quality index. The application can realize segmented nonlinear correction of near-infrared spectroscopy and accurate alignment of the spectral space to improve the prediction accuracy of the quality index, so as to complete intelligent closed-loop adjustment of the operating parameters of the oil refining device.

Description

Technical Field

[0001] This application relates to the field of industrial process control technology, and more specifically, to an online monitoring system for oil refining processes based on near-infrared spectroscopy. Background Technology

[0002] In the oil refining industry, real-time monitoring of key quality indicators such as octane number, density, and vapor pressure is crucial for ensuring the quality of finished oil products and optimizing equipment operating efficiency. Traditional offline laboratory testing has shortcomings such as detection lag and inability to adapt to real-time control in continuous production. Near-infrared spectroscopy technology, with its advantages of being fast, non-destructive, and capable of online detection, has become the mainstream technology for online quality monitoring in the oil refining process.

[0003] Existing near-infrared spectroscopy monitoring technologies for oil refining mostly employ a global linear approach for spectral correction and feature extraction, relying on a single model to predict quality indicators and implement simple deviation control. This approach cannot eliminate the nonlinear spectral differences between field instruments and the main laboratory instrument, and struggles to accurately extract local spectral features. It suffers from insufficient accuracy in predicting quality indicators and the inability to generate precise control signals under the coupling of multiple operating parameters. Therefore, achieving piecewise nonlinear correction and precise spectral spatial alignment of near-infrared spectroscopy to improve the prediction accuracy of quality indicators, thereby enabling intelligent closed-loop regulation of refining unit operating parameters, has become a significant challenge for the industry. Summary of the Invention

[0004] This application provides an online monitoring system for oil refining processes based on near-infrared spectroscopy, which can achieve segmented nonlinear correction and precise spatial alignment of near-infrared spectra to improve the prediction accuracy of quality indicators, thereby completing the intelligent closed-loop regulation of the operating parameters of oil refining units.

[0005] This application provides an online monitoring system for oil refining processes based on near-infrared spectroscopy, wherein the monitoring includes:

[0006] The spectral acquisition module acquires the raw spectrum of the oil refining material to be tested during the oil refining process using an online near-infrared spectrometer, and performs baseline correction and smoothing and noise reduction on the raw spectrum to obtain a preprocessed spectrum.

[0007] The model transfer module segments the preprocessed spectrum using a pre-trained segmented autoencoder spectral transfer network, then performs nonlinear correction on each segmented spectrum to obtain multiple corrected spectral segments, and finally stitches together all the corrected spectral segments to obtain reconstructed spectral data aligned with the spectral space of the main instrument.

[0008] The property prediction module performs segmented feature extraction on the reconstructed spectral data to obtain the local spectral features of each spectral segment, and determines the predicted values ​​of various key quality indicators in the refined oil material to be tested based on all the local spectral features.

[0009] The closed-loop control module compares the predicted value of each key quality indicator with the preset target quality indicator range. When the deviation exceeds the control threshold, it generates a control signal to adjust the operating parameters of the corresponding refining unit.

[0010] In this embodiment, the baseline correction and smoothing / denoising of the original spectrum from the instrument to obtain the preprocessed spectrum specifically includes:

[0011] The original spectrum from the instrument is subjected to mode coupling baseline correction processing to obtain a preliminary corrected spectrum;

[0012] By performing convolution smoothing filtering on the preliminary corrected spectrum, a smoothed corrected spectrum is obtained;

[0013] The preprocessed spectrum is obtained by performing a first-order derivative transformation based on the smoothed corrected spectrum.

[0014] In this embodiment, the preprocessed spectrum is segmented using a pre-trained segmented autoencoder spectral transfer network, and then nonlinear correction is performed on each segmented spectrum to obtain multiple corrected spectral segments, specifically including:

[0015] The preprocessed spectrum is uniformly divided into non-overlapping spectral sub-intervals according to a preset equal wavelength interval to obtain multiple original segmented spectra.

[0016] Each original segmented spectrum is input into a pre-trained segmented autoencoder spectral transfer network for compression mapping and nonlinear reconstruction dimensionality upscaling to obtain the preliminary corrected spectral segments for each spectrum.

[0017] Each original segmented spectrum is added point by point to the corresponding preliminary corrected spectrum segment to obtain multiple corrected spectrum segments.

[0018] In this embodiment, the process of stitching together all the corrected spectral bands to obtain reconstructed spectral data aligned with the spectral space of the main instrument specifically includes:

[0019] All the corrected spectral bands are arranged end to end according to the original acquisition wavelength to obtain the preliminary spliced ​​spectrum;

[0020] The initial spliced ​​spectrum and the preprocessed spectrum are subjected to wavelength-by-wavelength residual superposition processing to obtain a reconstructed spectrum with enhanced residuals;

[0021] By applying a weighted moving average filter to the reconstructed spectrum, reconstructed spectral data aligned with the spectral space of the main instrument is obtained.

[0022] In this embodiment, the instrument is a near-infrared spectrometer installed at the oil refining unit site for real-time online spectral acquisition.

[0023] In this embodiment, the main instrument is a near-infrared spectrometer used to establish a spectral data correction model.

[0024] In this embodiment, the segmented autoencoded spectral transfer network is a neural network specifically designed for nonlinear difference correction between near-infrared spectroscopy instruments in the oil refining process. It consists of multiple independent autoencoded blocks that correspond one-to-one with spectral segments and residual connection identity mapping branches.

[0025] In this embodiment, segmented feature extraction of the reconstructed spectral data to obtain local spectral features for each spectral band specifically includes:

[0026] The reconstructed spectral data is segmented into equal-length segments to obtain multiple reconstructed spectral segments;

[0027] Each reconstructed spectral segment is input into its corresponding independent fully connected hidden subnetwork for absorption response feature extraction, resulting in a local feature vector for each reconstructed spectral segment.

[0028] The local spectral features of each spectral band are determined based on all local feature vectors.

[0029] In this embodiment, determining the predicted values ​​of various key quality indicators in the refined oil material to be tested based on all local spectral characteristics specifically includes:

[0030] All local spectral features are spliced ​​together in segmented order to obtain a global feature representation of the complete band;

[0031] Based on the global features, the local spectral features of each spectrum segment are adaptively weighted and fused to obtain the fused features;

[0032] The fusion features are scaled to obtain predicted values ​​of various key quality indicators in the refined oil materials to be tested.

[0033] In this embodiment, the predicted value of each key quality indicator is compared with the preset target quality indicator range. When the deviation exceeds the control threshold, a control signal is generated to adjust the corresponding refining unit operating parameters. Specifically, this includes:

[0034] For each key quality indicator, the predicted value of the key quality indicator is compared with the preset target quality indicator range to obtain the quality deviation signal.

[0035] When the absolute value of the quality deviation signal exceeds the control threshold, the quality deviation signal is marked as a valid deviation signal, thereby obtaining the valid deviation signal for each key quality indicator.

[0036] Multivariate decoupling calculations are performed on all valid deviation signals to obtain the target correction amount for the operating parameters of each refining unit.

[0037] Based on all the target correction values, control signals are generated to adjust the operating parameters of the corresponding refining unit.

[0038] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0039] The raw spectra of the refined oil material to be tested are acquired by an online near-infrared spectrometer during the refining process. The raw spectra are then baseline-corrected and smoothed to obtain preprocessed spectra. The preprocessed spectra are segmented using a pre-trained segmented autoencoder spectral transfer network. Each segmented spectrum is then nonlinearly corrected to obtain multiple corrected spectral segments. All corrected spectral segments are stitched together to obtain reconstructed spectral data aligned with the spectral space of the master instrument. Segmented feature extraction is performed on the reconstructed spectral data to obtain the local spectral features of each spectral segment. Based on all the local spectral features, the predicted values ​​of various key quality indicators in the refined oil material to be tested are determined. The predicted value of each key quality indicator is compared with the preset target quality indicator range. When the deviation exceeds the control threshold, a control signal is generated to adjust the corresponding refining unit operating parameters.

[0040] Therefore, it can be seen that the control signal generated in this application to adjust the operating parameters of the corresponding refining unit firstly, by determining the calibration spectral segment, the spectral segment after correcting the nonlinear spectral response deviation component between the slave and master instruments can be obtained. The determination of the calibration spectral segment can accurately correct the instrument-specific nonlinear distortion and completely retain the true spectral component information of the refining material, avoiding spectral distortion caused by over-calibration. At the same time, the wide-band complex instrument differences are decomposed into local bands for independent calibration to reduce processing complexity and improve calibration efficiency, providing a precise and reliable data foundation for subsequent spectral spatial alignment and local spectral feature extraction, directly improving the prediction accuracy of key quality indicators and the accuracy of multi-parameter closed-loop control of the refining unit, effectively solving the industry pain points of insufficient calibration accuracy, large prediction deviation, and poor control adaptability of traditional technologies. Then, by determining the reconstructed spectral data, the final spectral data that retains the true spectral absorption characteristics of the refining material and completely corrects the nonlinear response deviation between instruments can be obtained. The determination of the reconstructed spectral data can achieve precise alignment with the spectral space of the master instrument and completely retain the true spectral characteristics of the refining material. By analyzing spectral absorption characteristics and thoroughly correcting nonlinear response deviations between instruments, spectral distortion and discontinuities are eliminated. This allows for the direct reuse of the high-precision quantitative calibration model of the main instrument, providing high-fidelity standardized data for subsequent local spectral feature extraction and accurate prediction of key quality indicators. This ensures the accuracy and stability of multi-parameter closed-loop control of refining units, fundamentally addressing the industry pain points of insufficient calibration accuracy, poor model adaptability, and lagging prediction and control in traditional technologies. Finally, the predicted values ​​provide real-time quality feedback values ​​with physical dimensions that match the actual refining process control parameters. The determination of predicted values ​​allows for the synchronous, accurate, and real-time acquisition of core indicators such as octane number, density, and vapor pressure, replacing lagging laboratory test data. This provides quantitative and immediate quality feedback for the closed-loop control module, effectively eliminating control deviations caused by detection lag. Simultaneously, relying on high-precision predicted values, precise decoupling and adjustment of multiple operating parameters are achieved, ensuring stable finished oil quality and efficient unit operation. This fundamentally solves the industry pain points of inaccurate prediction, data lag, and poor control adaptability in traditional technologies.

[0041] In summary, the technical solution adopted in this application can achieve piecewise nonlinear correction and precise spatial alignment of near-infrared spectra to improve the prediction accuracy of quality indicators, thereby completing the intelligent closed-loop adjustment of the operating parameters of the oil refining unit. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1This is a module structure diagram of the online monitoring system for oil refining processes based on near-infrared spectroscopy provided in this application;

[0044] Figure 2 This is an exemplary flowchart of determining reconstructed spectral data according to the present application;

[0045] Figure 3 This is a module interaction diagram of the online monitoring system for oil refining processes based on near-infrared spectroscopy provided in this application. Detailed Implementation

[0046] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0047] To better understand the above technical solutions, a detailed description of the technical solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. (Refer to...) Figure 1 As shown in the figure, this is a modular structure diagram of an online monitoring system for an oil refining process based on near-infrared spectroscopy according to this embodiment of the present application. The monitoring system includes: a spectral acquisition module 100, a model transfer module 200, a property prediction module 300, and a closed-loop control module 400, which are described below:

[0048] The spectral acquisition module 100 acquires the raw spectrum of the oil refining material to be tested during the oil refining process using an online near-infrared spectrometer, and performs baseline correction and smoothing and noise reduction on the raw spectrum of the instrument to obtain a preprocessed spectrum.

[0049] It should be noted that, in this application, the original spectrum from the instrument refers to the original absorption spectrum signal of the material to be tested, which is directly collected on-site by an online near-infrared spectrometer in the oil refining unit and has not yet undergone spectral difference correction between instruments.

[0050] In practice, firstly, a bypass sampling branch pipe can be drawn from the main pipeline of the refining unit to continuously draw the refined oil material to be tested from the main pipeline and send it to the sample pretreatment unit. A filter device with a heating jacket removes solid particulate impurities from the refined oil material while maintaining its temperature within a preset control range, resulting in a clean material stream. Secondly, the clean material stream is sent to a flow cell equipped with a sapphire optical window. The flow cell allows the clean material stream to form a uniformly thick liquid film within the near-infrared light transmission path region. Simultaneously, a temperature control device maintains the temperature fluctuation of the clean material stream within the flow cell within ±1 degree Celsius, resulting in a temperature-stable sample. Then, through an optical fiber transmission link composed of low-hydroxyl quartz optical fibers, the broadband near-infrared light generated by the near-infrared spectrometer is transmitted to the light incident end of the flow cell. The near-infrared beam passes perpendicularly through the temperature-stable sample in the flow cell and exits from the light exit end, where it is collected by the receiving optical fiber. Different molecular groups in the sample... The selective absorption of near-infrared light causes a characteristic attenuation of the intensity of transmitted light in the wavelength range of 780 nm to 2526 nm, resulting in a transmitted light signal carrying information about the material composition. Finally, the transmitted light signal is transmitted back to the detector module of the near-infrared spectrometer host via an optical fiber link. An indium gallium arsenide array detector converts the light intensity signals of different wavelengths into corresponding electrical signals, which are then converted into digital spectral data by an analog-to-digital converter. This digital spectral data is used as the raw spectrum from the instrument. The near-infrared spectrometer can be a Fourier transform type, a fixed grating diode array type, or an acousto-optic tunable filter type. The near-infrared spectrometer host and flow cell are installed in a positive-pressure explosion-proof analysis cabin or explosion-proof analysis cabinet that meets explosion-proof requirements, ensuring long-term safe and stable operation of the system in flammable and explosive hazardous areas of the oil refining plant. The filtration device used in the sample pretreatment unit can be a self-cleaning backwash filter or a multi-channel filter connected in parallel.

[0051] In this embodiment, the baseline correction and smoothing denoising of the original spectrum from the instrument to obtain the preprocessed spectrum can be achieved by the following steps:

[0052] The original spectrum from the instrument is subjected to mode coupling baseline correction processing to obtain a preliminary corrected spectrum;

[0053] By performing convolution smoothing filtering on the preliminary corrected spectrum, a smoothed corrected spectrum is obtained;

[0054] The preprocessed spectrum is obtained by performing a first-order derivative transformation based on the smoothed corrected spectrum.

[0055] It should be noted that, in this application, the preliminary calibration spectrum refers to the spectral data after separating and subtracting the low-frequency baseline drift signal. It can eliminate the slowly varying baseline shift in the spectrum while maintaining the integrity of the characteristic absorption peak shape, providing a pure spectral component for subsequent smoothing filtering. The smoothing calibration spectrum refers to the spectral data after suppressing high-frequency random noise. It can significantly improve the spectral signal-to-noise ratio while retaining the peak position and peak width characteristics of the absorption peaks, avoiding the amplification of noise in the subsequent derivative transformation and affecting the prediction accuracy. The preprocessed spectrum is the spectral data that shows the rate of change of absorbance with wavelength at each wavelength point. It can eliminate the additive interference caused by sample temperature fluctuations and residual baseline drift, while enhancing the distinction between overlapping absorption peaks, thereby improving the extraction efficiency and quantitative prediction accuracy of characteristic spectral information by the subsequent model transfer module and property prediction module.

[0056] In specific implementation, firstly, a mode-coupled hierarchical Gaussian prior model can be introduced to adaptively characterize the local coupling sparse structure between each wavelength point and its adjacent wavelength points in the instrument's original spectrum, resulting in a wavelength-coupled sparse structure model. Then, a sparse Bayesian learning framework is used to iteratively estimate the baseline drift in the wavelength-coupled sparse structure model point-by-point. During the iteration process, the mode-coupled hierarchical model is used to adaptively learn the block sparse structure of the spectral characteristic peaks, obtaining a baseline drift estimation sequence. Finally, the baseline drift estimation sequence is subtracted one by one from the original absorbance values ​​at each wavelength point in the instrument's original spectrum. Based on the baseline drift estimated at each wavelength point, the pure spectral signal is separated from the baseline signal to obtain the preliminary corrected spectrum after baseline subtraction. Then, the Savitzky-Gorye convolution smoothing filter parameters can be set according to the sampling interval and signal-to-noise ratio characteristics of the preliminary corrected spectrum. For example, a filter window width of 7 to 15 consecutive wavelength points can be selected, and the polynomial fitting order can be selected as 2 to 4. The sliding window is then placed sequentially on each wavelength point of the preliminary corrected spectrum, and a least-squares polynomial fit is performed on multiple adjacent wavelength points symmetrically distributed on both sides of the selected wavelength point within the window. The fitted data is then used to... The obtained polynomial curve is used to calculate the smoothed absorbance value at the center wavelength point of the sliding window, thus obtaining the smoothed value at the center wavelength point. The sliding window is then shifted point by point from the starting wavelength point to the ending wavelength point of the spectrum. For each wavelength point, the polynomial fitting and smoothing value calculation operations within the window are repeated to obtain a smoothed spectral curve covering the entire wavelength range, which serves as the smoothed correction spectrum. Finally, the difference step size parameter for performing first-order derivative differencing on the smoothed correction spectrum is determined based on the sampling interval of the smoothed correction spectrum. For example, the difference step size parameter is set to an interval of 1 to 3 sampling points. The smoothed absorbance values ​​at each wavelength point in the smoothed correction spectrum are then calculated. The difference operation is performed according to the difference step size parameter. That is, for each wavelength point in the smooth correction spectrum, the difference between the smooth absorbance value at the current wavelength point and the smooth absorbance value at the specified step size interval is calculated to obtain the difference sequence. The absorbance difference of each wavelength point in the difference sequence is divided by the corresponding wavelength interval value to obtain the first derivative value at each wavelength point. Thus, the additive baseline residue caused by sample temperature fluctuations or small changes in optical path in the smooth correction spectrum is eliminated by the first derivative transformation. Then, the first derivative spectrum formed by arranging the first derivative values ​​at each wavelength point in wavelength order is used as the preprocessed spectrum.

[0057] The model transfer module 200 segments the preprocessed spectrum using a pre-trained segmented autoencoder spectral transfer network, then performs nonlinear correction on each segmented spectrum to obtain multiple corrected spectral segments, and finally stitches together all the corrected spectral segments to obtain reconstructed spectral data aligned with the spectral space of the main instrument.

[0058] In this embodiment, the preprocessed spectrum is segmented using a pre-trained segmented autoencoder spectral transfer network, and then multiple corrected spectral segments are obtained by nonlinear correction of each segmented spectrum. This can be achieved through the following steps:

[0059] The preprocessed spectrum is uniformly divided into non-overlapping spectral sub-intervals according to a preset equal wavelength interval to obtain multiple original segmented spectra.

[0060] Each original segmented spectrum is input into a pre-trained segmented autoencoder spectral transfer network for compression mapping and nonlinear reconstruction dimensionality upscaling to obtain the preliminary corrected spectral segments for each spectrum.

[0061] Each original segmented spectrum is added point by point to the corresponding preliminary corrected spectrum segment to obtain multiple corrected spectrum segments.

[0062] It should be noted that in this application, the original segmented spectrum is an independent spectral segment obtained by dividing the preprocessed spectrum into preset equal wavelength intervals. It can be used to decompose the wide-band nonlinear instrument differences into multiple local band differences to reduce the complexity of a single calibration and reduce network redundancy parameters. The preliminary calibration spectral segment is an intermediate result output after segmented autoencoder block compression mapping and dimensional reconstruction. It can be used to fit the nonlinear spectral response deviation components between the slave instrument and the master instrument in the corresponding band. The calibration spectral segment refers to the spectral segment after correcting the nonlinear spectral response deviation components between the slave instrument and the master instrument. It can correct only the instrument-specific distortion while retaining the effective component information of the original spectrum, avoiding the introduction of additional spectral distortion due to overcalibration. Among them, the slave instrument is a near-infrared spectrometer installed on the site of the oil refining unit for real-time online acquisition of spectra. It is usually a compact, shock-resistant and explosion-proof online dedicated analyzer, such as a fixed grating diode array type near-infrared spectrometer or an acousto-optic tunable filter type near-infrared spectrometer. The master instrument is a near-infrared spectrometer used to establish a spectral data calibration model. It is usually a high-precision, high-stability large research-grade instrument placed in a central laboratory, such as a Fourier transform type near-infrared spectrometer.

[0063] In practice, the preprocessed spectrum is first uniformly divided according to a preset equal wavelength interval. For example, every twenty consecutive wavelength points are divided into a non-overlapping spectral sub-interval, thus decomposing the complete band spectrum into multiple independent spectral segments of equal length, resulting in multiple original segmented spectra. Then, each original segmented spectrum is individually input into its corresponding independent autoencoder block in a pre-trained segmented autoencoder spectral transfer network. The encoding layer of the independent autoencoder block uses a fully connected weight matrix to perform linear transformation and nonlinear activation processing on the original segmented spectrum, compressing and mapping the original segmented spectrum from the original high-dimensional wavelength space to a low-dimensional hidden layer feature space of a preset dimension. For example, a spectral segment containing twenty wavelength points is compressed into a thirty-two-dimensional feature vector, thereby extracting core feature information sensitive to instrument differences. The compressed feature vectors of each spectral segment are obtained. Then, the decoding layer of the independent autoencoder block performs a dimensionality-up reconstruction operation on the compressed feature vectors, symmetrical to that of the encoding layer. For example, the compressed feature vectors are restored layer by layer to the same number of wavelength points as the original input segmented spectrum through another set of fully connected weight matrices and nonlinear activation functions. This achieves an approximate fit to the instrument-specific nonlinear distortion components carried in the original segmented spectrum, resulting in the preliminary corrected spectral segments of each spectrum. Finally, for each original segmented spectrum and its corresponding preliminary corrected spectral segment, a point-by-point addition operation is performed at the corresponding wavelength point position. This allows the segmented autoencoder network to prioritize learning the instrument spectral differences rather than the identity mapping of the complete spectrum during training, ensuring the correction effect while avoiding the introduction of additional spectral distortion, thus obtaining multiple corrected spectral segments.

[0064] It should be noted that the segmented autoencoded spectral transfer network in this application is a neural network specifically designed for nonlinear difference correction between near-infrared spectroscopy instruments in the oil refining process. It consists of multiple independent autoencoded blocks corresponding one-to-one with spectral segments and residual connection identity mapping branches. Each independent autoencoded block contains an encoding layer responsible for spectral compression mapping and a decoding layer responsible for nonlinear reconstruction and dimensionality enhancement. This network can decompose a wide-band spectrum into independent local band corrections, accurately fit the nonlinear spectral response deviation between the field instrument and the main laboratory instrument, avoid spectral distortion caused by global linear correction, reduce network redundant parameters, improve correction efficiency, and, in conjunction with residual mapping and subsequent stitching smoothing processing, completely eliminate boundary jumps and artifacts in segmented correction, achieving accurate alignment of spectral data with the spectral space of the main instrument, thus laying a high-precision data foundation for subsequent spectral feature extraction and key quality indicator prediction.

[0065] Preferably, in this embodiment, reference Figure 2 As shown, this figure is an exemplary flowchart for determining reconstructed spectral data in an embodiment of this application. In this embodiment, the reconstructed spectral data aligned with the spectral space of the main instrument after stitching together all the corrected spectral segments can be achieved by the following steps:

[0066] In step S21, all the corrected spectral segments are arranged end to end according to the original acquisition wavelength order to obtain the preliminary spliced ​​spectrum;

[0067] In step S22, the preliminary spliced ​​spectrum and the preprocessed spectrum are subjected to wavelength-by-wavelength residual superposition processing to obtain a residual-enhanced reconstructed spectrum;

[0068] In step S23, the reconstructed spectrum is subjected to a weighted moving average filter to obtain reconstructed spectral data that is spatially aligned with the main instrument's spectral data.

[0069] It should be noted that, in this application, the preliminary spliced ​​spectrum refers to a continuous spectral curve that eliminates boundary intensity jumps caused by segmented independent corrections. It can provide a transitional spectrum with continuous spectral intensity and no abrupt changes for subsequent residual superposition processing. The reconstructed spectrum refers to spectral data that integrates the baseline profile features of the original spectrum with the nonlinear difference correction features between instruments. It can accurately correct the spectral response deviations between different instruments while preserving the true absorption profile of the sample. The reconstructed spectral data is the final spectral data that preserves the true spectral absorption features of the refined oil materials and thoroughly corrects the nonlinear response deviations between instruments. It can ensure that the final output spectral data is highly consistent with the main instrument spectrum in the feature space, so that the quantitative correction model established based on the main instrument spectrum can be directly applied to the reconstructed spectral data without additional correction.

[0070] In practice, firstly, all corrected spectral segments can be arranged end-to-end according to the original acquisition wavelength in ascending order. The difference in correction intensity at the same wavelength point in the boundary junction region between adjacent corrected spectral segments is then arithmetically averaged. For example, when two corrected spectral segments contain the same wavelength point in the boundary overlap region, the arithmetic mean of the two correction intensity values ​​is directly taken as the correction intensity value for that wavelength point. When there is a wavelength gap at the boundary junction, cubic spline interpolation is used to generate a smooth transition interpolated correction intensity value within the wavelength gap, thereby eliminating the spectral intensity jumps caused by segmented correction and obtaining the preliminary spliced ​​spectrum. Secondly, using the pre-constructed residual connection identity mapping branch in the segmented autoencoded spectral transfer network, the preliminary spliced ​​spectrum and the preprocessed spectrum are subjected to wavelength-by-wavelength residual superposition processing. That is, the correction intensity value at each wavelength point in the preliminary spliced ​​spectrum is added point-by-point to the original intensity value of that wavelength point in the preprocessed spectrum. Through residual superposition processing, the correction intensity value at each wavelength point in the preprocessed spectrum is obtained. The original spectral baseline contour features are linearly fused with the instrument nonlinear difference correction details learned by the segmented autoencoder spectral transfer network. This results in a reconstructed spectrum that retains the overall shape information of the original spectrum while correcting local response differences between different instruments, thus obtaining a residual-enhanced spectrum. Then, a weighted moving average filter is applied to the reconstructed spectrum within a preset width range near the segment boundary wavelength points. The window width of the weighted moving average filter is taken as the range of a preset number of wavelength points on both sides of the adjacent segment boundary when the segmented autoencoder spectral transfer network segments. The filter weights are distributed using a Gaussian kernel function based on the distance between the wavelength points and the boundary center to attenuate possible spectral splicing artifacts near the boundary. For example, when the segment boundary is located near the wavenumber position of 1200 nm, a Gaussian weighted smoothing process is applied to the range of five wavelength points on both sides of the boundary to eliminate local spectral discontinuities caused by independent segment correction, resulting in a smooth reconstructed spectrum that is aligned with the spectral space of the main instrument.

[0071] The property prediction module 300 performs segmented feature extraction on the reconstructed spectral data to obtain the local spectral features of each spectral segment, and determines the predicted values ​​of various key quality indicators in the refined oil material to be tested based on all the local spectral features.

[0072] In this embodiment, segmented feature extraction of the reconstructed spectral data to obtain the local spectral features of each spectral segment can be achieved through the following steps:

[0073] The reconstructed spectral data is segmented into equal-length segments to obtain multiple reconstructed spectral segments;

[0074] Each reconstructed spectral segment is input into its corresponding independent fully connected hidden subnetwork for absorption response feature extraction, resulting in a local feature vector for each reconstructed spectral segment.

[0075] The local spectral features of each spectral band are determined based on all local feature vectors.

[0076] It should be noted that, in this application, the reconstructed spectral segment is a continuous and non-overlapping sub-interval of spectral data in the reconstructed spectral data. This allows subsequent feature extraction to focus on the wavelength range of independent absorption peaks corresponding to specified molecular groups, avoiding mutual interference between absorption features of different bands, thereby improving the spectral resolution accuracy of complex components in refining materials. The local feature vector is a low-dimensional abstract feature characterizing the vibrational transition intensity of molecular groups within the reconstructed spectral segment. It can filter out irrelevant noise and redundant variables while retaining key component information, reducing the complexity of subsequent fusion calculations. The local spectral feature is a standardized feature representation of the reconstructed spectral segment with consistent data distribution and generalization robustness. It can eliminate dimensional differences between the outputs of different sub-network segments, suppress the risk of overfitting caused by spectral fluctuations in the industrial online environment, and ensure the stability and reliability of the predicted values ​​of key quality indicators in the refining process during long-term operation.

[0077] In practice, firstly, the reconstructed spectral data can be segmented into equal-length segments with a fixed window width (e.g., every fifty consecutive wavelengths). This ensures that each segment covers the local wavelength range in the near-infrared spectrum corresponding to the independent vibrational transitions of different molecular groups such as methyl, methylene, and aromatic hydrocarbons, resulting in multiple continuous and non-overlapping reconstructed spectral segments. Secondly, each reconstructed spectral segment is input into an independent fully connected hidden subnetwork uniquely corresponding to that segment. The input reconstructed spectral segment undergoes nonlinear feature transformation and feature encoding using a nonlinear activation function (e.g., a linear unit function with leakage correction or a hyperbolic tangent function) within the hidden subnetwork. This extracts the characteristic absorption peak positions of the corresponding molecular groups within the reconstructed spectral segment. The local absorption response features at each location are used as local feature vectors, thus obtaining the local feature vectors for each reconstructed spectral segment. Then, batch normalization and random deactivation sparsity constraint processing are applied to the local feature vectors corresponding to all reconstructed spectral segments. Batch normalization is used to unify the data distribution scale of each local feature vector to accelerate network convergence. Random deactivation sparsity constraint processing is used to randomly disconnect some neuron connections in the hidden subnetwork at a preset ratio (e.g., 20% to 50%) to suppress the overfitting trend during the training process of the subnetwork and enhance the robustness of the extracted local features to spectral noise in the online industrial environment, thus obtaining the standardized local spectral features of each reconstructed spectral segment.

[0078] In this embodiment, the predicted values ​​of various key quality indicators in the refined oil material to be tested can be determined based on all local spectral characteristics using the following steps:

[0079] All local spectral features are spliced ​​together in segmented order to obtain a global feature representation of the complete band;

[0080] Based on the global features, the local spectral features of each spectrum segment are adaptively weighted and fused to obtain the fused features;

[0081] The fusion features are scaled to obtain predicted values ​​of various key quality indicators in the refined oil materials to be tested.

[0082] It should be noted that in this application, the global feature representation is a joint feature carrier characterizing the complete molecular vibration information of the refined oil material to be tested. It can transform the independently extracted spectral features into complete band feature inputs that can be uniformly processed by the fully connected layer, avoiding the loss of information on the coordinated expression of chemical groups across the entire band due to feature fragmentation between segmented spectra. The fusion feature is a nonlinear combination result that differentiates the contribution of each band spectrum to the prediction of the target quality index. It can overcome the limitations of the traditional partial least squares method in equally weighting each wavelength point, so that the aromatic characteristic peak bands have higher weights in octane number prediction and the carbon-hydrogen bond stretching vibration bands have higher weights in density prediction, thereby significantly improving the accuracy of simultaneous prediction of multiple indicators under complex refined oil material systems. The predicted value is a real-time quality feedback value with physical dimensions that matches the actual refining process control parameters. It can provide downstream distributed control systems with high-frequency, multi-dimensional, and synchronous refined oil material property data that can be directly used for deviation comparison and quality edge control, replacing the lag mode of traditional laboratory manual test results input control.

[0083] In practical implementation, firstly, a one-dimensional feature vector concatenation technique can be used to concatenate the local spectral features of each spectrum segment. For example, the local spectral features of the first spectrum segment correspond to octane number prediction, the local spectral features of the second spectrum segment correspond to density prediction, and the local spectral features of the third spectrum segment correspond to vapor pressure prediction, etc., and these are sequentially concatenated along the feature dimension to obtain a complete global feature representation of the band. The dimension of this global feature representation vector is the sum of the dimensions of the local spectral features of each segment. Secondly, the trainable weight parameters of the fully connected neural network layer are used to adaptively weight and fuse the local absorption feature vectors of each spectrum segment in the global feature representation. That is, based on the physicochemical mechanisms such as the band where the octane number is mainly associated with the characteristic absorption peaks of olefins and aromatic groups, the band where the density is mainly associated with the overall molecular structure information of hydrocarbons, and the band where the vapor pressure is mainly associated with the volatility characteristics of light components, the weight coefficient values ​​of each band are automatically adjusted during the training process of historical sample spectral data through the backpropagation algorithm. For example, when predicting the octane number, the adaptive weighted fusion layer automatically assigns higher weight coefficients to the aromatic characteristic peak bands; when predicting density, the adaptive weighted fusion layer automatically assigns higher weight coefficients to the bands with richer C-H bond stretching vibration information across the entire band, thus obtaining fused features. Then, the numerical scale mapping transformation of the fused features is performed through the linearly activated output layer, transforming each feature value within the fused features into a value that matches the actual physical dimensions of various key quality indicators of the refined oil material to be tested as the predicted value output. For example, for the octane number, the corresponding feature value is linearly mapped to the dimensionless octane number range between 90 and 100; for density, the corresponding feature value is linearly mapped to the density range between 0.72 and 0.78 grams per cubic centimeter; and for vapor pressure, the corresponding feature value is linearly mapped to the vapor pressure range between 45 and 80 kPa, thus obtaining the predicted values ​​of various key quality indicators in the refined oil material to be tested; wherein, the key quality indicators include octane number, density, and vapor pressure.

[0084] The closed-loop control module 400 compares the predicted value of each key quality indicator with the preset target quality indicator range. When the deviation exceeds the control threshold, it generates a control signal to adjust the operating parameters of the corresponding refining unit.

[0085] In this embodiment, the predicted value of each key quality indicator is compared with the preset target quality indicator range. When the deviation exceeds the control threshold, a control signal for adjusting the corresponding refining unit operating parameters is generated. This can be achieved through the following steps:

[0086] For each key quality indicator, the predicted value of the key quality indicator is compared with the preset target quality indicator range to obtain the quality deviation signal.

[0087] When the absolute value of the quality deviation signal exceeds the control threshold, the quality deviation signal is marked as a valid deviation signal, thereby obtaining the valid deviation signal for each key quality indicator.

[0088] Multivariate decoupling calculations are performed on all valid deviation signals to obtain the target correction amount for the operating parameters of each refining unit.

[0089] Based on all the target correction values, control signals are generated to adjust the operating parameters of the corresponding refining unit.

[0090] It should be noted that, in this application, the quality deviation signal is a deviation measurement parameter characterizing the degree of deviation of the predicted value of a key quality indicator from the preset target quality indicator range. It can transform abstract quality fluctuations into calculable numerical signals, providing a unified quantitative input benchmark for subsequent control decisions. The effective deviation signal is an effective quality deviation signal that eliminates small and meaningless fluctuations. It can avoid frequent adjustments to the operating parameters of the refining unit triggered by normal random jitter of the predicted value near the target boundary, thereby reducing the frequency of actuator actions and wear, and improving the control stability of the online monitoring system. The target correction amount is the magnitude value that the operating parameters of the refining unit need to be adjusted. It can collaboratively convert the deviation information of one or more quality indicators into the coordinated adjustment amount of multiple mutually coupled operating parameters, solving the technical problem that traditional single-loop control cannot take into account the complex coupling relationship between multiple quality indicators and multiple operating parameters. The control signal is the final adjustment command signal used to drive the action of the refining unit actuator. It can seamlessly embed the quality information sensed in real time by near-infrared spectroscopy into the existing automatic control loop of the refining unit, realizing a complete closed loop from spectral acquisition to unit adjustment, and upgrading the quality control of the refining process from hysteresis feedback to real-time feedforward compensation.

[0091] In practice, firstly, for each key quality indicator, the predicted value is compared with the preset target quality indicator range. That is, the absolute difference between the predicted value and the upper and lower boundary values ​​of the target quality indicator range is calculated to obtain the quality deviation signal. Secondly, the quality deviation signal is processed by a control threshold. That is, when the absolute value of the quality deviation signal exceeds the preset control threshold, the quality deviation signal is marked as a valid deviation signal. The control threshold can be preset based on the historical process control data of the corresponding key quality indicators of the refining unit. For example, for octane number, the preset control threshold can be 0.2 octane number units. The valid deviation signal of each key quality indicator can be obtained through the above steps. Then, all valid deviation signals are transmitted to the advanced process control module of the distributed control system. The advanced process control module performs multivariate decoupling calculations on all valid deviation signals, comprehensively considering the mutual coupling relationship between various key quality indicators, and combining the current operating parameter status of the refining unit to obtain the target correction amount for the corresponding refining unit operating parameters. The refining unit operating parameters include reaction temperature, feed flow rate, catalyst circulation rate, hydrogen-to-oil ratio, and reaction pressure. Finally, all target correction amounts are integrated according to the type of the corresponding refining unit operating parameters, and the distributed control system generates control signals to adjust the corresponding refining unit operating parameters as the final output signal for the online closed-loop monitoring of the refining unit.

[0092] It should be noted that the reference Figure 3As shown in the figure, this is a module interaction diagram of an online monitoring system for oil refining processes based on near-infrared spectroscopy according to this embodiment of the present application. It includes: the interaction between the spectral acquisition module and the oil refining unit / the oil material to be tested, whereby the spectral acquisition module acquires the raw spectrum of the oil material to be tested from the instrument at the oil refining unit, performs mode coupling baseline correction and convolutional smoothing filtering, and outputs the preprocessed spectrum to the model transfer module to complete the noise reduction and drift removal preprocessing of the raw spectrum, providing a clean data foundation for subsequent spectral correction; and the interaction between the model transfer module and the spectral acquisition module, whereby the model transfer module receives the preprocessed spectrum, performs spectral segmentation and nonlinear correction through a pre-trained segmented autoencoder spectral transfer network to obtain the corrected spectral segment, and then performs wavelength sequential stitching, residual superposition, and weighted moving average filtering to output reconstructed spectral data precisely aligned with the spectral space of the main instrument. The property prediction module addresses the nonlinear spectral differences between the slave instrument and the master instrument. The interaction between the property prediction module and the model transfer module involves the property prediction module receiving reconstructed spectral data, segmenting it into equal-length segments, extracting local spectral features through an independent fully connected hidden subnetwork, and then outputting predicted values ​​for key quality indicators such as octane number, density, and vapor pressure to the closed-loop control module after feature splicing, adaptive weighted fusion, and scale mapping. This enables simultaneous and accurate prediction of multiple indicators. The interaction between the closed-loop control module and the property prediction module / refining unit involves the closed-loop control module receiving predicted values ​​for key quality indicators, comparing them with preset target ranges, marking valid deviation signals, and generating control signals for adjusting parameters such as reaction temperature and feed flow rate through multivariate decoupling operations. This feedback is then sent to the refining unit, forming a complete online closed loop of "spectral acquisition - data processing - indicator prediction - unit control."

[0093] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0094] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0095] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. An online monitoring system for oil refining processes based on near-infrared spectroscopy, characterized in that, The monitoring system includes: The spectral acquisition module acquires the raw spectrum of the oil refining material to be tested during the oil refining process using an online near-infrared spectrometer, and performs baseline correction and smoothing and noise reduction on the raw spectrum to obtain a preprocessed spectrum. The model transfer module segments the preprocessed spectrum using a pre-trained segmented autoencoder spectral transfer network, then performs nonlinear correction on each segmented spectrum to obtain multiple corrected spectral segments, and finally stitches together all the corrected spectral segments to obtain reconstructed spectral data aligned with the spectral space of the main instrument. The property prediction module performs segmented feature extraction on the reconstructed spectral data to obtain the local spectral features of each spectral segment, and determines the predicted values ​​of various key quality indicators in the refined oil material to be tested based on all the local spectral features. The closed-loop control module compares the predicted value of each key quality indicator with the preset target quality indicator range. When the deviation exceeds the control threshold, it generates a control signal to adjust the operating parameters of the corresponding refining unit.

2. The online monitoring system for oil refining processes based on near-infrared spectroscopy as described in claim 1, characterized in that, The preprocessed spectrum obtained by performing baseline correction and smoothing / denoising on the raw spectrum from the instrument specifically includes: The original spectrum from the instrument is subjected to mode coupling baseline correction processing to obtain a preliminary corrected spectrum; By performing convolution smoothing filtering on the preliminary corrected spectrum, a smoothed corrected spectrum is obtained; The preprocessed spectrum is obtained by performing a first-order derivative transformation based on the smoothed corrected spectrum.

3. The online monitoring system for oil refining processes based on near-infrared spectroscopy as described in claim 1, characterized in that, The preprocessed spectrum is segmented using a pre-trained segmented autoencoder spectral transfer network, and then each segmented spectrum is nonlinearly corrected to obtain multiple corrected spectral segments, specifically including: The preprocessed spectrum is uniformly divided into non-overlapping spectral sub-intervals according to a preset equal wavelength interval to obtain multiple original segmented spectra. Each original segmented spectrum is input into a pre-trained segmented autoencoder spectral transfer network for compression mapping and nonlinear reconstruction dimensionality upscaling to obtain the preliminary corrected spectral segments for each spectrum. Each original segmented spectrum is added point by point to the corresponding preliminary corrected spectrum segment to obtain multiple corrected spectrum segments.

4. The online monitoring system for oil refining processes based on near-infrared spectroscopy as described in claim 1, characterized in that, After stitching together all the calibrated spectral bands, the reconstructed spectral data aligned with the spectral space of the main instrument is obtained, specifically including: All the corrected spectral bands are arranged end to end according to the original acquisition wavelength to obtain the preliminary spliced ​​spectrum; The initial spliced ​​spectrum and the preprocessed spectrum are subjected to wavelength-by-wavelength residual superposition processing to obtain a reconstructed spectrum with enhanced residuals; By applying a weighted moving average filter to the reconstructed spectrum, reconstructed spectral data aligned with the spectral space of the main instrument is obtained.

5. The online monitoring system for oil refining processes based on near-infrared spectroscopy as described in claim 1, characterized in that, The instrument in question is a near-infrared spectrometer installed at the oil refining unit site for real-time online spectral acquisition.

6. The online monitoring system for oil refining processes based on near-infrared spectroscopy as described in claim 1, characterized in that, The main instrument is a near-infrared spectrometer used to establish a spectral data correction model.

7. The online monitoring system for oil refining processes based on near-infrared spectroscopy as described in claim 1, characterized in that, The segmented autoencoded spectral transfer network is a neural network specifically designed for nonlinear difference correction between near-infrared spectroscopy instruments in the oil refining process. It consists of multiple independent autoencoded blocks that correspond one-to-one with spectral segments and residual connection identity mapping branches.

8. The online monitoring system for oil refining processes based on near-infrared spectroscopy as described in claim 1, characterized in that, The segmented feature extraction of the reconstructed spectral data to obtain the local spectral features of each spectral segment specifically includes: The reconstructed spectral data is segmented into equal-length segments to obtain multiple reconstructed spectral segments; Each reconstructed spectral segment is input into its corresponding independent fully connected hidden subnetwork for absorption response feature extraction, resulting in a local feature vector for each reconstructed spectral segment. The local spectral features of each spectral band are determined based on all local feature vectors.

9. The online monitoring system for oil refining processes based on near-infrared spectroscopy as described in claim 1, characterized in that, The predicted values ​​of various key quality indicators in the refined oil material to be tested are determined based on all local spectral characteristics, including: All local spectral features are spliced ​​together in segmented order to obtain a global feature representation of the complete band; Based on the global features, the local spectral features of each spectrum segment are adaptively weighted and fused to obtain the fused features; The fusion features are scaled to obtain predicted values ​​of various key quality indicators in the refined oil materials to be tested.

10. The online monitoring system for oil refining processes based on near-infrared spectroscopy as described in claim 1, characterized in that, The predicted value of each key quality indicator is compared with the preset target quality indicator range. When the deviation exceeds the control threshold, a control signal is generated to adjust the corresponding refining unit operating parameters. Specifically, this includes: For each key quality indicator, the predicted value of the key quality indicator is compared with the preset target quality indicator range to obtain the quality deviation signal. When the absolute value of the quality deviation signal exceeds the control threshold, the quality deviation signal is marked as a valid deviation signal, thereby obtaining the valid deviation signal for each key quality indicator. Multivariate decoupling calculations are performed on all valid deviation signals to obtain the target correction amount for the operating parameters of each refining unit. Based on all the target correction values, control signals are generated to adjust the operating parameters of the corresponding refining unit.

Images