A machine learning prediction system for chemical product quality

Abstract

The application relates to a machine learning prediction system for chemical product quality, and relates to the technical field of product quality prediction, and comprises a collection module: collecting multidimensional operation parameters, and screening parameter items that are strongly associated with product quality; a division module: constructing a time sequence association matrix based on the associated parameters, and obtaining time sequence window segments through sliding window division; an extraction module: extracting parameter fluctuation characteristics in the window, and constructing a fluctuation characteristic vector; a mapping module: inputting the fluctuation characteristic vector into a prediction learning model for feature mapping, and obtaining a quality prediction value; and a judgment module: when the prediction value exceeds a preset threshold value, outputting a quality early warning signal. The technical problem of how to realize early and accurate prediction of product quality based on multidimensional operation parameters of a chemical production line is solved.

Description

Technical Field

[0001] This invention relates to the field of product quality prediction technology, and in particular to a machine learning prediction system for the quality of chemical products. Background Technology

[0002] In existing technologies, chemical production processes typically rely on offline sampling and laboratory testing to assess product quality. This approach suffers from significant lag and struggles to reflect the real-time impact of production fluctuations on the final product. Meanwhile, distributed control systems (DCS) and process analysis techniques (PAT) in production lines accumulate massive amounts of multi-dimensional operating parameters (such as temperature, pressure, flow rate, material ratio, and reaction time), but the nonlinear and strongly coupled relationships between these parameters and quality indicators have not been fully explored and utilized. Traditional statistical process control or univariate threshold early warning methods often ignore the interactive effects between variables, failing to accurately predict quality deviations in advance. This leads to reactive adjustments after non-conforming products are produced, resulting in wasted raw materials and energy. Therefore, establishing a model based on the multi-dimensional operating parameters of chemical production lines that can accurately predict product quality in advance has become a pressing technical challenge. Summary of the Invention

[0003] The purpose of this invention is to at least partially solve one of the technical problems existing in the prior art.

[0004] To achieve the above objectives, the present invention provides a machine learning prediction system for chemical product quality, comprising: The data acquisition module is used to collect multi-dimensional parameters from the chemical production line to obtain a set of operating parameters, and to filter out parameter items that are strongly correlated with the quality of chemical products from the set of operating parameters to obtain a set of correlated parameters. The partitioning module is used to construct a time-series correlation matrix based on the correlation parameter set, and to partition the time-series correlation matrix into a sliding window to obtain time-series window segments. An extraction module is used to extract parameter fluctuation features within the time window segment and construct a fluctuation feature vector based on the parameter fluctuation features; The mapping module is used to input the fluctuation feature vector into a preset prediction learning model, perform feature mapping on the fluctuation feature vector, and obtain the quality prediction value. The judgment module is used to output a quality warning signal when the predicted quality value exceeds a preset quality threshold.

[0005] Furthermore, the multi-dimensional parameter collection of the chemical production line to obtain the operating parameter set includes: The module is used to pre-install temperature sensors, pressure sensors, and flow sensors in the reactors, heat exchangers, and conveying pipelines of the chemical production line. The reading module is used to synchronously read the instantaneous measurement values ​​of each sensor according to a preset sampling period to obtain the raw sampling sequence including temperature sequence, pressure sequence and flow sequence; The alignment module is used to align the timestamps of each measurement value in the original sampling sequence, and then horizontally splice the aligned temperature sequence, pressure sequence, and flow sequence according to the same time to obtain the set of operating parameters.

[0006] Furthermore, the step of filtering parameters strongly correlated with chemical product quality from the set of operating parameters to obtain a set of correlated parameters includes: Historical chemical product quality test data are obtained and arranged in time series to obtain historical quality series; Time-frequency domain mapping is performed on each parameter sequence in the set of operating parameters and the historical quality sequence to obtain the parameter time spectrum corresponding to each parameter sequence and the quality time spectrum corresponding to the historical quality sequence. The cross-spectral density is analyzed by combining the time spectrum of the parameters and the time spectrum of the mass to extract the average cross-amplitude spectrum curve; Extract the target frequency band curve located within the preset chemical response frequency band from the average cross amplitude spectrum curve; The target frequency band curve is fitted with an envelope to extract the spectral peak envelope area. The parameter sequences corresponding to the spectral peak envelope areas that are greater than a preset area threshold are separated from the running parameter set. The separated parameter sequences are then horizontally spliced ​​together to obtain the associated parameter set.

[0007] Furthermore, the step of performing cross-spectral density analysis on the parameter time spectrum and the mass time spectrum to extract the average cross-amplitude spectrum curve includes: The cross power matrix is ​​obtained by performing conjugate multiplication of the complex values ​​of the parameter time spectrum and the quality time spectrum; The phase angle data of each time-frequency node is extracted from the cross power matrix, and the phase angle data is unwrapped. The partial derivative of the unwrapped phase angle data in the time domain is calculated to obtain the phase fluctuation gradient. The matrix elements in the cross power matrix that correspond to phase fluctuation gradients greater than a preset gradient threshold are set to zero to obtain a purified cross matrix. The amplitude sequence of the purified cross matrix is ​​extracted, and the average amplitude of each frequency point in the amplitude sequence is calculated along the time dimension to obtain the average cross amplitude spectrum curve.

[0008] Furthermore, the step of performing conjugate multiplication of the complex values ​​of the parameter time spectrum and the quality time spectrum to obtain the cross power matrix includes: The complex values ​​of the parameter time spectrum and the mass time spectrum are decomposed in polar coordinates to extract the parameter amplitude matrix and parameter phase matrix of the parameter time spectrum, and the mass amplitude matrix and mass phase matrix of the mass time spectrum, respectively. The parameter amplitude matrix and the mass amplitude matrix are multiplied by the Hadamard product to obtain the joint energy matrix, and the parameter phase matrix and the mass phase matrix are interpolated by the element-wise difference to obtain the time-delay phase matrix. Substitute each element of the time-delay phase matrix into Euler's formula to generate a complex phase matrix, and multiply the joint energy matrix and the complex phase matrix element by element to obtain the cross power matrix.

[0009] Furthermore, the step of performing polar coordinate decomposition on the complex values ​​of the parameter time spectrum and the quality time spectrum to extract the parameter amplitude matrix and parameter phase matrix of the parameter time spectrum, respectively, includes: Extract the node complex values ​​corresponding to each time-frequency coordinate from the time-frequency spectrum of the parameters, perform vector magnitude analysis on the node complex values ​​to obtain the node magnitude sequence, and reconstruct the node magnitude sequence according to the time-frequency coordinates to obtain the parameter amplitude matrix; Argument features are extracted from the complex values ​​of the nodes to obtain the wrapped phase sequence. The phase is then continuously expanded along the time dimension of the wrapped phase sequence to obtain the deflection phase sequence. The deflection phase sequence is then reconstructed into a matrix according to the time-frequency coordinates to obtain the parameter phase matrix.

[0010] Furthermore, the step of constructing a time-series correlation matrix based on the correlation parameter set, and then dividing the time-series correlation matrix into sliding window segments to obtain time-series window fragments, includes: The parameter sequence is separated from the set of associated parameters, and the cross-correlation integral is performed on the parameter sequence to obtain the cross-correlation delay. The parameter sequence is shifted along the time axis based on the cross-correlation delay to obtain a lag-compensated sequence, and the lag-compensated sequence is then horizontally concatenated to obtain a time-series correlation matrix. The time series correlation matrix is ​​subjected to first-order difference calculation along the time axis to obtain the difference gradient matrix, and each column of the difference gradient matrix is ​​normalized by maximum absolute value to obtain the dimensionless gradient matrix. The modulus of each row is extracted from the dimensionless gradient matrix to obtain a modulus sequence. The extreme points of the modulus are extracted from the modulus sequence, and the time difference between adjacent extreme points of the modulus is calculated to obtain the dynamic sliding step size. According to the preset fixed window length, and using the dynamic sliding step size as the starting point interval of adjacent windows, the temporal correlation matrix is ​​segmented to obtain temporal window segments.

[0011] Furthermore, the step of extracting parameter fluctuation features within the time-series window segment and constructing a fluctuation feature vector based on the parameter fluctuation features includes: The second derivative of the time window segment is calculated along the time dimension to obtain a local curvature sequence, which is a two-dimensional matrix where rows represent time points and columns represent parameter dimensions. The local curvature sequence is subjected to element-wise difference operation with the preset linear baseline matrix to obtain the curvature deviation matrix. The standard deviation of each column of the curvature deviation matrix is ​​normalized to obtain the dimensionless curvature deviation matrix. Singular value decomposition is performed on the dimensionless curvature deviation matrix to extract the set of non-zero singular values. Calculate the second and third central moments of the set of non-zero singular values, and then concatenate the second and third central moments numerically to obtain the parameter fluctuation characteristics; The fluctuation characteristics of the parameters are subjected to range standardization to obtain normalized feature elements, and the normalized feature elements are arranged in a preset dimension order to construct the fluctuation feature vector.

[0012] Furthermore, the calculation of the second and third central moments of the set of non-zero singular values ​​includes: The distribution expectation is extracted from the set of non-zero singular values ​​to obtain the mean of the singular values; Based on the mean of the singular values, a centering mapping is performed on the set of non-zero singular values ​​to obtain a centered discrete sequence; The second-order distribution analysis is performed on the centered discrete sequence to obtain the second-order central moments, and the third-order distribution analysis is performed on the centered discrete sequence to obtain the third-order central moments.

[0013] Furthermore, the step of inputting the fluctuation feature vector into a preset prediction learning model and performing feature mapping on the fluctuation feature vector to obtain a quality prediction value includes: The fluctuation feature vector is input into a preset prediction learning model, which includes a sparse autoencoder module, a cross-dimensional attention regression module, and a quantile residual correction module. The wave feature vector is sparsely constrained and encoded by a sparse autoencoder module to obtain a hidden layer activation feature map. Based on the cross-dimensional attention regression module, the hidden layer activation feature map is subjected to multi-head attention decomposition and nonlinear regression to obtain a quality-sensitive score vector. The quality-sensitive score vector is input into the quantile residual correction module for quantile regression and residual compensation to obtain the quality prediction value.

[0014] This invention provides a machine learning prediction system for chemical product quality, comprising: a data acquisition module for acquiring multi-dimensional parameters of a chemical production line to obtain an operational parameter set, and selecting parameters strongly correlated with chemical product quality from the operational parameter set to obtain a correlation parameter set; a partitioning module for constructing a time-series correlation matrix based on the correlation parameter set, and partitioning the time-series correlation matrix into a sliding window to obtain time-series window segments; an extraction module for extracting parameter fluctuation features within the time-series window segments, and constructing fluctuation feature vectors based on the parameter fluctuation features; a mapping module for inputting the fluctuation feature vectors into a preset prediction learning model, performing feature mapping on the fluctuation feature vectors to obtain a quality prediction value; and a judgment module. This system outputs a quality warning signal when the predicted quality value exceeds a preset quality threshold. It solves the technical problem of how to accurately predict product quality in advance based on the multi-dimensional operating parameters of a chemical production line. By performing strong correlation screening and time-series sliding window division on the multi-dimensional operating parameters of the chemical production line, it effectively extracts feature vectors reflecting the dynamic fluctuations of the production process. Combined with a preset predictive learning model for nonlinear feature mapping, it can accurately predict the trend of product quality deviation in advance before the quality indicators are detected offline. This avoids the waste of raw materials and energy consumption caused by traditional delayed detection, significantly improves the timeliness and accuracy of quality warnings, reduces the defect rate, and ensures the stability and economy of the chemical production process. Attached Figure Description

[0015] 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a structural block diagram of a machine learning prediction system for chemical product quality according to an embodiment of the present invention; Figure 2 This is one embodiment of the present invention. Figure 1 Structural block diagram of the data acquisition module; The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0018] Please see Figure 1 An embodiment of the present invention provides a machine learning prediction system for the quality of chemical products, comprising: The data acquisition module 1 is used to collect multi-dimensional parameters of the chemical production line to obtain a set of operating parameters, and to filter out parameter items that are strongly correlated with the quality of chemical products from the set of operating parameters to obtain a set of correlated parameters.

[0019] Specifically, the data acquisition module 1 acquires multi-dimensional parameters such as temperature, pressure, flow rate, pH value, and reaction time in real time through a sensor network deployed on the chemical production line, forming an initial set of operating parameters. Subsequently, Pearson correlation coefficients or mutual information methods are used to analyze the correlation between each parameter and the final chemical product quality indicators (such as purity and impurity content). A threshold is set (e.g., the absolute value of the correlation coefficient is greater than 0.7) to filter out strongly correlated parameters, forming a set of associated parameters. For example, on an esterification reaction production line, a significant negative correlation was found between reflux temperature fluctuations and product acid value; this temperature parameter is then included in the associated parameter set. In this embodiment, the above scheme effectively eliminates redundant interference variables, improving the input signal-to-noise ratio for subsequent modeling.

[0020] The partitioning module 2 is used to construct a time-series correlation matrix based on the correlation parameter set, and to partition the time-series correlation matrix by sliding window to obtain time-series window segments.

[0021] Specifically, module 2 first aligns the associated parameter set according to sampling time, forming a time-series correlation matrix where rows correspond to time steps and columns correspond to each associated parameter. Then, it sets the sliding window length to 60 sampling points and the step size to 10 points, successively extracting sub-matrices along the time axis to generate continuously overlapping time-series window segments. Taking the three associated parameters selected in the esterification reaction—reflux temperature, feed rate, and reactor pressure—as an example, sampling is performed every 5 seconds, thus one window covers 5 minutes of process data. In this embodiment, the above scheme effectively preserves the dynamic coupling structure between multiple parameters, providing standardized input units for subsequent fluctuation feature extraction.

[0022] Extraction module 3 is used to extract parameter fluctuation features within the time window segment and construct a fluctuation feature vector based on the parameter fluctuation features.

[0023] Specifically, extraction module 3 calculates the first-order difference for each parameter sequence in each time window segment to characterize the instantaneous rate of change, and then statistically analyzes its standard deviation, peak-to-peak value, and local extreme point density as fluctuation features. Subsequently, the fluctuation features corresponding to all associated parameters are concatenated in a fixed order to form a fluctuation feature vector of a unified dimension. For example, within the 60-point window of the esterification reaction, the standard deviation of the reflux temperature is 1.2℃, extreme points occur 4 times, and the peak-to-peak feed rate reaches 0.8 L / min; these values ​​collectively constitute the feature vector of this window. In this embodiment, the above scheme effectively quantifies the dynamic perturbation characteristics of process parameters, providing physically meaningful criteria for model input.

[0024] The mapping module 4 is used to input the fluctuation feature vector into a preset prediction learning model, perform feature mapping on the fluctuation feature vector, and obtain the quality prediction value.

[0025] Specifically, mapping module 4 feeds the fluctuation feature vector into a pre-trained multilayer perceptron model. This model contains two hidden layers with 64 and 32 nodes respectively, uses ReLU as the activation function, and has a single-neuron linear unit as the output layer. After layer-by-layer weighting and nonlinear transformation, the input vector finally outputs the predicted quality value of the corresponding chemical product, such as the product's acid value or purity percentage. Taking esterification reaction as an example, the 12-dimensional fluctuation feature vector generated in a certain window, after model mapping, outputs a predicted acid value of 0.18 mgKOH / g. In this embodiment, the above scheme achieves end-to-end nonlinear mapping from dynamic process features to quality indicators, significantly improving prediction accuracy and generalization ability.

[0026] The judgment module 5 is used to output a quality warning signal when the predicted quality value exceeds a preset quality threshold.

[0027] Specifically, the judgment module 5 compares the predicted quality value output by the mapping module 4 with a preset quality threshold, which is set according to the process specifications, for example, the upper limit of the acid value of the esterification reaction product is 0.20 mgKOH / g. Once the predicted value exceeds this limit, the system immediately triggers a Boolean judgment condition, generates a high-level signal as a quality warning signal, and outputs it to the DCS operation station through the industrial communication interface. In this embodiment, the above scheme realizes the real-time capture of potential quality deterioration trends, avoiding the batch generation of non-conforming products.

[0028] In a specific embodiment, such as Figure 2 As shown, the multi-dimensional parameter collection of the chemical production line to obtain the operating parameter set includes: The module 11 is used to pre-install temperature sensors, pressure sensors and flow sensors in the reactor, heat exchanger and conveying pipeline of the chemical production line. The reading module 12 is used to synchronously read the instantaneous measurement values ​​of each sensor according to a preset sampling period to obtain the original sampling sequence including temperature sequence, pressure sequence and flow sequence; Alignment module 13 is used to align the timestamps of each measurement value in the original sampling sequence, and to horizontally splice the aligned temperature sequence, pressure sequence and flow sequence according to the same time to obtain the set of operating parameters.

[0029] Specifically, when implementing multi-dimensional parameter acquisition to construct an operating parameter set in a chemical production line, temperature sensors, pressure sensors, and flow sensors must first be installed in key process units such as reactors, heat exchangers, and conveying pipelines. The selection of these sensors must match their respective operating conditions. For example, a Pt100 resistance thermometer is installed in the reactor to monitor the reaction temperature, piezoresistive pressure transmitters are configured at the inlet and outlet of the heat exchanger to capture pressure drop changes, and electromagnetic flow meters are used on the conveying pipeline to obtain the material flow rate in real time. Subsequently, the entire sensor network is connected to the same data acquisition system, and a fixed sampling period (e.g., 1 second) is set to synchronously trigger and read all sensors, thereby obtaining the temperature, pressure, and flow sequences arranged in chronological order, forming the original sampling sequence.

[0030] Because different sensors may have microsecond-level response delays or communication transmission differences, the measured values ​​in the original sampling sequence may not strictly correspond to the same physical moment. Therefore, a timestamp alignment operation needs to be performed. Specifically, using the system master clock as a reference, the data reported by each sensor is time-stamped to the millisecond. Then, through interpolation or nearest neighbor matching, the three types of data—temperature, pressure, and flow rate—are aligned on a unified time grid. After alignment, the temperature, pressure, and flow rate values ​​corresponding to the same moment are horizontally concatenated into a record containing three dimensions. This process is repeated to form a complete multivariate time series matrix, i.e., the set of operating parameters.

[0031] For example, in a certain esterification reaction process, the data aligned at t=125 seconds might be: reactor temperature 86.3℃, heat exchanger outlet pressure 0.32 MPa, and feed line flow rate 4.7 L / min. This set of data is stored in the database as a row in the operating parameter set for subsequent processing. In this embodiment, the above scheme effectively solves the problem of asynchronous multi-source heterogeneous sensor data in the time dimension, ensuring the consistency of the operating parameter set in terms of time sequence and the integrity of its structure, providing a high-fidelity input basis for subsequent fluctuation feature extraction and quality prediction.

[0032] In a specific embodiment, the step of filtering parameters strongly correlated with chemical product quality from the set of operating parameters to obtain a set of correlated parameters includes: Historical chemical product quality test data are obtained and arranged in time series to obtain historical quality series; Time-frequency domain mapping is performed on each parameter sequence in the set of operating parameters and the historical quality sequence to obtain the parameter time spectrum corresponding to each parameter sequence and the quality time spectrum corresponding to the historical quality sequence. The cross-spectral density is analyzed by combining the time spectrum of the parameters and the time spectrum of the mass to extract the average cross-amplitude spectrum curve; Extract the target frequency band curve located within the preset chemical response frequency band from the average cross amplitude spectrum curve; The target frequency band curve is fitted with an envelope to extract the spectral peak envelope area. The parameter sequences corresponding to the spectral peak envelope areas that are greater than a preset area threshold are separated from the running parameter set. The separated parameter sequences are then horizontally spliced ​​together to obtain the associated parameter set.

[0033] Specifically, historical chemical product quality testing data needs to be acquired and arranged in chronological order of sampling time to form a historical quality sequence. This sequence typically originates from periodic laboratory testing results, such as acid value, color, or purity indicators for each batch of final products. Although its temporal resolution is lower than that of online sensor data, it can be extended to the same time granularity as the operating parameter set through linear interpolation or spline interpolation methods. Next, time-frequency domain mapping is performed on each dimension of the parameter sequence in the operating parameter set (such as reactor temperature sequence, heat exchanger pressure sequence, feed flow rate sequence, etc.) and the interpolated historical quality sequence. Specifically, a short-time Fourier transform (STFT) is used, with a Hanning window as the window function, a window length of 256 sampling points, and an overlap rate of 50%, thereby generating the parameter time-frequency spectrum corresponding to each parameter sequence and the quality time-frequency spectrum corresponding to the historical quality sequence.

[0034] The cross-spectral density (CSD) of each parameter's time-frequency spectrum and the quality's time-frequency spectrum is calculated point-by-point in the frequency domain. The CSD results from multiple time periods are then averaged to obtain the average cross-spectral density curve, which reflects the coupling strength between a specific parameter and product quality fluctuations at different frequencies. Considering that the dynamic response of chemical processes is typically concentrated in the 0.001 Hz to 0.1 Hz frequency band (corresponding to a time scale of approximately 10 seconds to 15 minutes), a target frequency band curve within this preset chemical response frequency band is extracted. Envelope fitting is then performed on this target frequency band curve. The Hilbert transform can be used to extract the modulus of the analytical signal, or an upper envelope can be constructed by connecting local maxima points, followed by integration to obtain the spectral peak envelope area. If the spectral peak envelope area corresponding to a parameter sequence exceeds a preset area threshold (such as the 3σ upper limit determined by statistical analysis), it is determined that the parameter has a significant frequency domain correlation with product quality, and it is separated from the original set of operating parameters.

[0035] All parameter sequences that meet the conditions are horizontally concatenated at the same time point to form a set of associated parameters with reduced dimensionality but focused information. Taking esterification reaction as an example, if the envelope area of ​​the reactor temperature sequence at 0.02 Hz is 0.85, which is higher than the threshold of 0.7, while the envelope area of ​​the delivery pipeline flow sequence is only 0.32, then only the temperature sequence is retained in the associated parameter set. In this embodiment, the above scheme accurately identifies key process variables that have a dynamic impact on product quality through time-frequency coupling analysis, effectively eliminating redundant or weakly correlated parameters, and significantly improving the efficiency and predictive stability of subsequent modeling.

[0036] In a specific embodiment, the step of performing cross-spectral density analysis on the parameter time spectrum and the mass time spectrum to extract the average cross-amplitude spectrum curve includes: The cross power matrix is ​​obtained by performing conjugate multiplication of the complex values ​​of the parameter time spectrum and the quality time spectrum; The phase angle data of each time-frequency node is extracted from the cross power matrix, and the phase angle data is unwrapped. The partial derivative of the unwrapped phase angle data in the time domain is calculated to obtain the phase fluctuation gradient. The matrix elements in the cross power matrix that correspond to phase fluctuation gradients greater than a preset gradient threshold are set to zero to obtain a purified cross matrix. The amplitude sequence of the purified cross matrix is ​​extracted, and the average amplitude of each frequency point in the amplitude sequence is calculated along the time dimension to obtain the average cross amplitude spectrum curve.

[0037] Specifically, conjugate multiplication needs to be performed on the complex values ​​of both on the same time-frequency grid; more specifically, suppose a parameter sequence, after a short-time Fourier transform, at time... ,frequency The complex number at the location is The complex number corresponding to the position in the historical quality sequence is... Then the element of that node in the cross power matrix is ​​calculated as follows: ,in express The complex conjugate of the cross power matrix is ​​obtained. The resulting cross power matrix contains magnitude and phase information, but it is contaminated with spurious phase jumps caused by noise or non-steady-state disturbances.

[0038] The phase angle data of all time-frequency nodes needs to be extracted from this matrix. And perform phase unwrapping processing on it—that is, when the phase difference between adjacent time points exceeds Automatic addition or subtraction when dealing with radians. To restore continuous phase trajectories and avoid model-related issues Gradient distortion caused by folding. Subsequently, the first-order partial derivative of the unwrapped phase sequence along the time dimension is calculated, i.e. The phase fluctuation gradient is obtained, which reflects the stability of the dynamic coupling between parameters and mass at a specific frequency. If the gradient is too large, it indicates that the phase relationship oscillates violently, which may be due to random disturbances rather than the actual process. Therefore, a preset gradient threshold (e.g., 0.5 rad / s obtained from statistical analysis of measured data) is set, and all elements in the cross-power matrix whose corresponding phase fluctuation gradient exceeds this threshold are forced to zero, thereby obtaining a purified cross-power matrix and effectively suppressing non-physically related components.

[0039] Take each frequency point in the purification cross matrix corresponding amplitudes at all time points Calculate the arithmetic mean along the time axis, i.e. This forms a one-dimensional sequence dependent only on frequency, which is the average cross-amplitude spectrum curve. Taking the reactor temperature and product acid value in the esterification reaction as an example, if the original cross-spectrum at 0.02 Hz experiences severe phase jitter due to pump start-stop interference, with a gradient reaching 0.8 rad / s, exceeding the threshold, then this frequency point is set to zero; while at 0.015 Hz, the gradient is 0.3 rad / s, and its amplitude is retained for mean calculation. In this embodiment, the above scheme purifies the cross-spectrum by introducing a phase stability criterion, significantly improving the noise resistance of frequency domain correlation identification, avoiding misjudging transient interference as quality correlation features, and thus providing a more reliable basis for subsequent parameter screening.

[0040] In a specific embodiment, the step of performing conjugate multiplication of the complex values ​​of the parameter time spectrum and the quality time spectrum to obtain the cross power matrix includes: The complex values ​​of the parameter time spectrum and the mass time spectrum are decomposed in polar coordinates to extract the parameter amplitude matrix and parameter phase matrix of the parameter time spectrum, and the mass amplitude matrix and mass phase matrix of the mass time spectrum, respectively. The parameter amplitude matrix and the mass amplitude matrix are multiplied by the Hadamard product to obtain the joint energy matrix, and the parameter phase matrix and the mass phase matrix are interpolated by the element-wise difference to obtain the time-delay phase matrix. Substitute each element of the time-delay phase matrix into Euler's formula to generate a complex phase matrix, and multiply the joint energy matrix and the complex phase matrix element by element to obtain the cross power matrix.

[0041] Specifically, in industrial process monitoring, suppose in an esterification reaction system, the complex numerical matrix of reactor temperature (as a process parameter) and product acid value (as a quality indicator) on a time-frequency grid is obtained through short-time Fourier transform (STFT). For example, at time point... ,frequency At that point, the complex value of the spectrum of the parameter is The complex value of the spectrum at mass time is First, we perform polar coordinate decomposition on both complex numbers: Calculate its modulus This is the element of the parameter magnitude matrix at that position; simultaneously, its argument is calculated. This constitutes the corresponding terms of the parametric phase matrix. Similarly, for... , module length Record the mass amplitude matrix; argument Stored in the mass phase array.

[0042] The parameter magnitude matrix and the mass magnitude matrix are in the same position. Perform the Hadamard product (i.e., multiply element-wise) on the position to obtain the value of that point in the joint energy matrix: At the same time, the parameter phase matrix and the mass phase matrix are calculated as a successive difference at this point, i.e. This forms the corresponding elements of the time-delay phase matrix. Here, "time-delay phase" does not refer to the actual time delay, but rather to the relative phase shift reflecting the dynamic response of the two signals at that frequency.

[0043] Substituting the time delay phase into Euler's formula Calculated , Therefore, the phase complex matrix at this point is Finally, the joint energy matrix and the phase complex matrix are multiplied element-wise at the same positions, i.e. This is the cross-power matrix in The complex value at the given position. If we directly verify this using the conjugate multiplication of the original complex numbers: The results showed a high degree of agreement, and the joint energy matrix and the phase complex matrix were matrices in which the number of columns of the first matrix was equal to the number of rows of the second matrix.

[0044] The process is repeated at all time-frequency nodes, and finally the complete cross-power matrix is ​​assembled. In this embodiment, the above scheme explicitly separates amplitude and phase information, and reconstructs the cross spectrum using the Hadamard product and phase difference. This not only preserves the mathematical essence of conjugate multiplication, but also provides a structured data path for the subsequent introduction of phase stability criteria (such as gradient threshold screening), significantly enhancing the robustness and physical interpretability of cross-spectral analysis in noisy industrial scenarios.

[0045] In a specific embodiment, the step of performing polar coordinate decomposition on the complex values ​​of the parameter time spectrum and the quality time spectrum to extract the parameter amplitude matrix and parameter phase matrix of the parameter time spectrum, respectively, includes: Extract the node complex values ​​corresponding to each time-frequency coordinate from the time-frequency spectrum of the parameters, perform vector magnitude analysis on the node complex values ​​to obtain the node magnitude sequence, and reconstruct the node magnitude sequence according to the time-frequency coordinates to obtain the parameter amplitude matrix; Argument features are extracted from the complex values ​​of the nodes to obtain the wrapped phase sequence. The phase is then continuously expanded along the time dimension of the wrapped phase sequence to obtain the deflection phase sequence. The deflection phase sequence is then reconstructed into a matrix according to the time-frequency coordinates to obtain the parameter phase matrix.

[0046] Specifically, in industrial process data analysis, performing polar coordinate decomposition on the complex values ​​of the parameter time-frequency spectrum and the quality time-frequency spectrum to extract the parameter amplitude matrix and parameter phase matrix is ​​a crucial signal preprocessing operation. This process begins by reading the time-frequency coordinates point by point from the parameter time-frequency spectrum. The corresponding node complex values. For example, in a certain esterification reaction process, the reactor temperature signal at time point is obtained through short-time Fourier transform (STFT). ,frequency The complex value at is Perform vector magnitude parsing on this complex value, that is, calculate its absolute value. This operation is repeated at all time and frequency points, forming a one-dimensional sequence of node magnitudes. This sequence is then rearranged into a two-dimensional matrix structure according to the row and column order of the original time-frequency grid—typically, time indices correspond to rows and frequency indices to columns—to obtain the parameter amplitude matrix. This matrix fully preserves the energy distribution characteristics of the original time-frequency spectrum, with each element representing the local intensity of the signal at the corresponding time and frequency point.

[0047] Argument feature extraction is performed on the complex values ​​of the same set of nodes, i.e., using the four-quadrant arctangent function. Calculate the principal value interval The phase angles within the range form a wrapped phase sequence. For example, the aforementioned complex number Argument is However, due to the periodicity constraint of the arctangent function, when the true phase spans... At the boundary, discontinuous jumps occur, causing phase trajectory distortion. To solve this problem, it is necessary to process each fixed frequency along the time dimension of the enclosing phase sequence. The corresponding phase trajectory is then subjected to phase continuity unrolling separately. Specifically, this involves iterating through adjacent phase points on the time axis; if the absolute value of the phase difference between two points is greater than... Then add or subtract at subsequent points. This causes the phase change to exhibit a monotonically continuous trend. For example, suppose at frequency... At a certain location, the package phase sequence at three consecutive time points is as follows: radians, where from arrive The jump exceeds This is determined to be an encapsulation effect; at this point, the second point is corrected to... The third point is corrected accordingly. The final deflection phase sequence is as follows: This eliminates spurious transitions. After expanding all frequency channels, the resulting deflection phase sequence is reconstructed using the original time-frequency coordinates to form a parametric phase matrix with the same dimension as the parametric amplitude matrix.

[0048] In this embodiment, the above scheme effectively restores the true amplitude and phase structure of the parameter time spectrum by explicitly performing modulus calculation and phase unwrapping operations, avoiding gradient distortion caused by phase folding, and providing a reliable data foundation for subsequent construction of a high-precision cross power matrix and identification of dominant influencing frequencies.

[0049] In a specific embodiment, the step of constructing a time-series correlation matrix based on the correlation parameter set, and dividing the time-series correlation matrix into sliding window segments to obtain time-series window fragments, includes: The parameter sequence is separated from the set of associated parameters, and the cross-correlation integral is performed on the parameter sequence to obtain the cross-correlation delay. The parameter sequence is shifted along the time axis based on the cross-correlation delay to obtain a lag-compensated sequence, and the lag-compensated sequence is then horizontally concatenated to obtain a time-series correlation matrix. The time series correlation matrix is ​​subjected to first-order difference calculation along the time axis to obtain the difference gradient matrix, and each column of the difference gradient matrix is ​​normalized by maximum absolute value to obtain the dimensionless gradient matrix. The modulus of each row is extracted from the dimensionless gradient matrix to obtain a modulus sequence. The extreme points of the modulus are extracted from the modulus sequence, and the time difference between adjacent extreme points of the modulus is calculated to obtain the dynamic sliding step size. According to the preset fixed window length, and using the dynamic sliding step size as the starting point interval of adjacent windows, the temporal correlation matrix is ​​segmented to obtain temporal window segments.

[0050] Specifically, in practical applications of industrial process monitoring and modeling, parameter sequences are separated from the associated parameter set. For example, in a continuous esterification reaction system, the associated parameter set may include two key process variables: reactor temperature and product acid value, each forming a parameter sequence that evolves over time. To reveal the potential causal delay relationship between the two, cross-correlation integral calculations are performed on the parameter sequences. The essence of this calculation is to measure the similarity of the overall waveforms of the two sequences at different time offsets, thereby locating the offset that maximizes the similarity, which is the cross-correlation delay. In actual operation, if it is found that the cross-correlation value reaches its peak when the product acid value sequence lags behind the temperature sequence by approximately 420 seconds, then this 420-second delay can be determined as the actual process response delay.

[0051] The parameter sequence is shifted along the time axis based on the cross-correlation delay, that is, the acid value sequence is shifted backward by 420 seconds (or equivalently, the temperature sequence is truncated forward by 420 seconds). This makes the two originally misaligned variables physically synchronized in time. The resulting two aligned data streams are called hysteresis-compensated sequences. Subsequently, the hysteresis-compensated sequences are horizontally concatenated—that is, at the same time scale, the temperature value is used as the first column and the aligned acid value as the second column, and they are stacked row by row to form a two-dimensional structure. This is the time-series correlation matrix, which completely preserves the co-evolution trajectory of the two key variables in the dynamic process.

[0052] To further highlight the relative drastic changes in variables and eliminate dimensional interference, the time-series correlation matrix needs to be calculated using first-order difference along the time axis. This involves subtracting the value from the previous time step from the current value, thus obtaining a difference gradient matrix reflecting the instantaneous rate of change. Since temperature changes are measured in °C / s while acid value changes are measured in mgKOH / g·s, the numerical scales differ significantly. Therefore, each column of the difference gradient matrix must be normalized to its maximum absolute value. Specifically, the change with the largest absolute value in the temperature difference column (e.g., 1.8 °C / s) is identified, and all elements in that column are divided by this value. Similarly, the acid value difference column (e.g., a maximum change of 0.05 mgKOH / g·s) is processed. The resulting dimensionless gradient matrix compresses all elements to the [-1, 1] interval, making the intensity of changes in different variables comparable.

[0053] The modulus of each row is extracted from the dimensionless gradient matrix—that is, the two normalized gradients at each moment are treated as two components of a two-dimensional vector, and their Euclidean lengths are calculated, thus obtaining a modulus sequence reflecting the overall dynamic activity of the system. The peaks of this sequence correspond to critical moments when the process state undergoes significant changes; therefore, modulus extrema (usually local maxima) need to be extracted from it. For example, in an 8-hour period of operating data, several extrema may be identified, such as at the 1800th second, 4200th second, and 6500th second. Then, the time difference between adjacent modulus extrema (e.g., 4200th second) is calculated. 1800 = 2400 seconds, 6500 (4200 = 2300 seconds), take the average or median as the dynamic sliding step size, which is approximately 2350 seconds in this case.

[0054] The time-series correlation matrix is ​​segmented according to a preset fixed window length (e.g., 3000 seconds) and with the dynamic sliding step size as the starting point interval between adjacent windows: the first window is segmented from second 0 to second 3000, the second window from second 2350 to second 5350, the third from second 4700 to second 7700, and so on. This results in time-series window segments that maintain a fixed analysis length to accommodate subsequent model inputs while ensuring that the starting position of each window is close to the dynamic inflection point of the process through dynamic step sizes. This effectively avoids the stage fragmentation problem that may be caused by traditional equal-interval segmentation, significantly improving the representativeness and effectiveness of the obtained time-series window segments in fault diagnosis or soft measurement modeling.

[0055] In a specific embodiment, the step of extracting parameter fluctuation features within the time-series window segment and constructing a fluctuation feature vector based on the parameter fluctuation features includes: The second derivative of the time window segment is calculated along the time dimension to obtain a local curvature sequence, which is a two-dimensional matrix where rows represent time points and columns represent parameter dimensions. The local curvature sequence is subjected to element-wise difference operation with the preset linear baseline matrix to obtain the curvature deviation matrix. The standard deviation of each column of the curvature deviation matrix is ​​normalized to obtain the dimensionless curvature deviation matrix. Singular value decomposition is performed on the dimensionless curvature deviation matrix to extract the set of non-zero singular values. Calculate the second and third central moments of the set of non-zero singular values, and then concatenate the second and third central moments numerically to obtain the parameter fluctuation characteristics; The fluctuation characteristics of the parameters are subjected to range standardization to obtain normalized feature elements, and the normalized feature elements are arranged in a preset dimension order to construct the fluctuation feature vector.

[0056] Specifically, the second derivative of the time-series window segment is calculated along the time dimension. This is not a simple difference calculation, but rather an estimation of the rate of change of acceleration of each parameter at each moment using discrete numerical methods (such as central difference), thereby revealing the "curvature" of its trajectory. For example, in a 3000-second window, if the temperature and acid value curves show a rapid rise followed by a rapid flattening, their second derivatives will exhibit a pulse shape of first positive and then negative. This shape directly reflects the intensity of the dynamic transition in the process. The resulting local curvature sequence is a two-dimensional matrix, where rows represent time points (e.g., 3000 rows) and columns represent parameter dimensions (e.g., 2 columns, corresponding to temperature and acid value respectively). Each element characterizes the local curvature of a specific variable at a specific moment.

[0057] Even with similar curvature values ​​under different operating conditions, their deviations from "normal and stable operation" can vary drastically. Therefore, it is necessary to perform element-by-element difference operations between the local curvature sequence and a preset linear baseline matrix. This preset linear baseline matrix refers to the curvature values ​​that each parameter should have under ideal steady-state or linear response conditions; it is usually set as a zero matrix (because the second derivative of linear change is zero). Therefore, the difference operation essentially preserves the original curvature itself, but this step logically emphasizes the concept of "deviation." The resulting curvature deviation matrix highlights the nonlinear perturbation of the actual process relative to the linear baseline. Since the curvature amplitudes of temperature and acid value are naturally not on the same order of magnitude, standard deviation normalization must be performed on each column of the curvature deviation matrix—that is, dividing the temperature curvature column by its own standard deviation, and performing the same treatment on the acid value curvature column—to eliminate the difference in variable scales and obtain a dimensionless curvature deviation matrix, allowing subsequent analysis to focus on the relative structure of the fluctuations rather than their absolute magnitude.

[0058] Singular Value Decomposition (SVD) is performed on the dimensionless curvature deviation matrix. SVD is a powerful matrix factorization technique that represents the original matrix as a product of three matrices, where the middle diagonal matrix contains a number of singular values ​​arranged in descending order, reflecting the energy concentration of the data in different orthogonal directions. In this context, the set of non-zero singular values ​​essentially encodes the number and intensity of the main modes of multivariate curvature co-oscillations within the window.

[0059] For example, if two variables fluctuate in a highly synchronous manner, there may only be one dominant singular value; if complex asynchronous oscillations exist, multiple significant singular values ​​may appear. After extracting these non-zero singular values, the system does not directly use their original values, but further calculates the second and third central moments of the set of non-zero singular values: the second central moment, i.e., the variance, reflects the dispersion of the singular value distribution and characterizes the diversity of fluctuation patterns; the third central moment reflects the skewness of the distribution, revealing whether there is a dominant pattern that is much stronger than other patterns. By numerically concatenating these two statistics (e.g., writing the variance value first, then the skewness value), the parameter fluctuation characteristics are constituted, which summarize the nonlinear dynamic complexity of the entire window in a very low-dimensional (only two-dimensional) but high-information-density manner.

[0060] Since the range of parameter fluctuation features extracted from different windows may vary significantly, range standardization is required. This involves subtracting the historical minimum value from the current value for each feature dimension, then dividing by the difference between the historical maximum and minimum values, thereby compressing all normalized feature elements into the [0,1] interval. Subsequently, these normalized feature elements are arranged according to a preset dimensional order (e.g., second moment first, then third moment), ultimately constructing a uniformly formatted fluctuation feature vector. This vector can be directly input into a classifier or clustering model to identify abnormal operating conditions, divide operational stages, or assess process stability, thus achieving in-depth characterization and intelligent analysis of the inherent fluctuation characteristics of the esterification reaction system without relying on a physical model.

[0061] In a specific embodiment, calculating the second and third central moments of the set of non-zero singular values ​​includes: The distribution expectation is extracted from the set of non-zero singular values ​​to obtain the mean of the singular values; Based on the mean of the singular values, a centering mapping is performed on the set of non-zero singular values ​​to obtain a centered discrete sequence; The second-order distribution analysis is performed on the centered discrete sequence to obtain the second-order central moments, and the third-order distribution analysis is performed on the centered discrete sequence to obtain the third-order central moments.

[0062] Specifically, in the aforementioned esterification reaction system, the expected distribution of the set of non-zero singular values ​​is extracted, i.e., the arithmetic mean of all singular values ​​in the set is calculated; this is the singular value mean. For example, if a time window segment yields three non-zero singular values ​​after SVD: [5.2, 1.8, 0.6], then its singular value mean is (5.2 + 1.8 + 0.6) / 3 = 2.533. This mean represents the "center of gravity" level of the overall fluctuation energy within the current window and serves as the benchmark for subsequent centralization processing.

[0063] The set of non-zero singular values ​​is centered based on the mean of the singular values. This involves subtracting the mean from each singular value, thereby eliminating the overall bias and retaining only the deviation information relative to the average energy. Taking the above example, the centered result is [5.2]. 2.533, 1.8 2.533, 0.6 2.533] ≈ [2.667, 0.733, [1.933] This result constitutes a centered discrete sequence. This sequence is no longer affected by the absolute energy magnitude, but purely reflects the relative strength of each fluctuation mode compared to the average level—positive numbers indicate that the mode is stronger than the average, and negative numbers indicate that it is weaker. This centering operation is a prerequisite for statistical moment analysis, ensuring that subsequent calculations are of higher-order characteristics around the distribution center, rather than appearances disturbed by the overall scale.

[0064] A second-order distribution analysis is performed on the centered discrete sequence, that is, the square of each element in the sequence is calculated and then averaged. The result is the second-order central moment (i.e., variance). Continuing with the numerical example, (2.667² + ( 0.733)² +( 1.933)²) / 3 ≈ (7.113 + 0.537 + 3.737) / 3 ≈ 3.796. This second-order central moment measures the dispersion of the singular value distribution: if all singular values ​​are close to the mean (e.g., [2.6, 2.5, 2.5]), the variance is small, indicating that the fluctuation energy is evenly distributed across multiple modes, and the process is relatively stable; conversely, if the variance is large (as in this example), it indicates the existence of a dominant mode (the first singular value is much higher than the others), and the system is in a highly nonlinear or abruptly changing state. Meanwhile, a third-order distribution analysis is performed on the centered discrete sequence, i.e., the cube of each element is calculated and then averaged to obtain the third-order central moment. Continuing with the previous example, (2.667³ + ( 0.733)³ + ( 1.933)³) / 3 ≈ (18.97 0.394 7.226) / 3 ≈ 3.783. This third-order central moment reflects the direction and degree of skewness of the distribution: a positive value indicates that the distribution is right-skewed, that is, there are a few extremely singular values ​​that raise the overall shape, which is common in sudden disturbances or start-up / termination stages; a negative value indicates left-skewed, but this is relatively rare in actual industrial processes.

[0065] By first extracting the mean of singular values, then centering them, and finally performing second- and third-order power-wise averaging, a rigorous high-order statistical characterization of the set of non-zero singular values ​​is achieved. Furthermore, the resulting second- and third-order central moments possess clear engineering interpretability—the former characterizes the diversity of fluctuation patterns, while the latter reveals the prominence of the dominant pattern. These two quantities together constitute a compact and profound description of the cooperative nonlinear behavior of parameters within the time series window, laying a solid statistical foundation for the subsequent construction of highly discriminative fluctuation feature vectors.

[0066] In a specific embodiment, the step of inputting the fluctuation feature vector into a preset prediction learning model and performing feature mapping on the fluctuation feature vector to obtain a quality prediction value includes: The fluctuation feature vector is input into a preset prediction learning model, which includes a sparse autoencoder module, a cross-dimensional attention regression module, and a quantile residual correction module. The wave feature vector is sparsely constrained and encoded by a sparse autoencoder module to obtain a hidden layer activation feature map. Based on the cross-dimensional attention regression module, the hidden layer activation feature map is subjected to multi-head attention decomposition and nonlinear regression to obtain a quality-sensitive score vector. The quality-sensitive score vector is input into the quantile residual correction module for quantile regression and residual compensation to obtain the quality prediction value.

[0067] Specifically, in the quality prediction task of the esterification reaction process, the fluctuation feature vector is input into a preset prediction learning model for feature mapping to finally obtain the quality prediction value. Its implementation includes three sequentially connected core modules. First, the fluctuation feature vector is fed into a sparse autoencoder module, which consists of an encoder and a decoder. The encoder adopts a single hidden layer fully connected structure with ReLU activation function. An L1 regularization term is introduced into the loss function to impose sparse constraints on the activation values ​​of the hidden layer neurons. For example, if the input fluctuation feature vector has a dimension of 2 (corresponding to the second and third central moments respectively), the encoder maps it to an 8-dimensional latent space. By optimizing the weighted sum of the reconstruction error and the sparse penalty term, the model is forced to activate only the few neurons most discriminative to the current condition, thereby outputting a hidden layer activation feature map with compact representation capabilities.

[0068] The hidden activation feature map is passed to the cross-dimensional attention regression module. This module first performs a linear projection on the 8-dimensional feature map to generate three sets of vectors: query, key, and value, and configures four parallel attention heads. Each head independently calculates the attention weight, that is, by normalizing the dot product of query and key using Softmax, and then weighting and summing the values ​​to capture the interaction dependencies between different hidden units. After the outputs of each head are concatenated, they are passed through a non-linear fully connected layer (containing tanh activation) for regression mapping, and finally output a scalar quality-sensitive score vector. Although it is called a "vector" here, in this embodiment, because the prediction target is a single acid value index, the actual dimension is 1, and its value directly reflects the intensity of the influence of dynamic fluctuations within the current window on the final product acid value.

[0069] The quality-sensitive score vector is input into the quantile residual correction module. Instead of using traditional mean squared error regression, this module constructs a 0.5 quantile (median) regression loss function and introduces a residual compensation mechanism: specifically, the model synchronously records the historical prediction residual sequence during training and, based on the empirical distribution of residuals within a sliding window, adds a dynamic offset to the current prediction value. For example, if recent residuals are mostly positive (i.e., the model systematically underestimates the acid value), the output is automatically adjusted upwards during the correction phase. The final output value is the quality prediction value, representing a real-time estimate of the acid value of the current batch of products.

[0070] In this embodiment, the above scheme suppresses redundant feature responses through sparse autoencoder, explicitly models the nonlinear coupling relationship between hidden layer dimensions by means of cross-dimensional attention mechanism, and improves prediction robustness by combining quantile regression and residual feedback. It effectively overcomes the problem of insufficient adaptability of traditional linear models in strongly nonlinear chemical processes and significantly improves the accuracy and stability of acid value prediction.

Claims

1. A machine learning prediction system for the quality of chemical products, characterized in that, include: The data acquisition module is used to collect multi-dimensional parameters of the chemical production line to obtain a set of operating parameters, and to filter out parameter items that are strongly correlated with the quality of chemical products from the set of operating parameters to obtain a set of correlated parameters. The partitioning module is used to construct a time-series correlation matrix based on the correlation parameter set, and to partition the time-series correlation matrix into a sliding window to obtain time-series window segments. An extraction module is used to extract parameter fluctuation features within the time window segment and construct a fluctuation feature vector based on the parameter fluctuation features; The mapping module is used to input the fluctuation feature vector into a preset prediction learning model, perform feature mapping on the fluctuation feature vector, and obtain the quality prediction value. The judgment module is used to output a quality warning signal when the predicted quality value exceeds a preset quality threshold.

2. The machine learning prediction system for chemical product quality according to claim 1, characterized in that, The process of collecting multi-dimensional parameters from the chemical production line to obtain an operating parameter set includes: The module is used to pre-install temperature sensors, pressure sensors, and flow sensors in the reactors, heat exchangers, and conveying pipelines of the chemical production line. The reading module is used to synchronously read the instantaneous measurement values ​​of each sensor according to a preset sampling period to obtain the raw sampling sequence including temperature sequence, pressure sequence and flow sequence; The alignment module is used to align the timestamps of each measurement value in the original sampling sequence, and then horizontally splice the aligned temperature sequence, pressure sequence, and flow sequence according to the same time to obtain the set of operating parameters.

3. The machine learning prediction system for chemical product quality according to claim 1, characterized in that, The step of filtering parameters strongly correlated with chemical product quality from the set of operating parameters yields a set of correlated parameters, including: Historical chemical product quality test data are obtained and arranged in time series to obtain historical quality series; Time-frequency domain mapping is performed on each parameter sequence in the set of operating parameters and the historical quality sequence to obtain the parameter time spectrum corresponding to each parameter sequence and the quality time spectrum corresponding to the historical quality sequence. The cross-spectral density is analyzed by combining the time spectrum of the parameters and the time spectrum of the mass to extract the average cross-amplitude spectrum curve; Extract the target frequency band curve located within the preset chemical response frequency band from the average cross amplitude spectrum curve; The target frequency band curve is fitted with an envelope to extract the spectral peak envelope area. The parameter sequences corresponding to the spectral peak envelope areas that are greater than a preset area threshold are separated from the running parameter set. The separated parameter sequences are then horizontally spliced ​​together to obtain the associated parameter set.

4. The machine learning prediction system for chemical product quality according to claim 3, characterized in that, The step of performing cross-spectral density analysis on the time-frequency spectrum of the parameters and the time-frequency spectrum of the mass to extract the average cross-amplitude spectrum curve includes: The cross power matrix is ​​obtained by performing conjugate multiplication of the complex values ​​of the parameter time spectrum and the quality time spectrum; The phase angle data of each time-frequency node is extracted from the cross power matrix, and the phase angle data is unwrapped. The partial derivative of the unwrapped phase angle data in the time domain is calculated to obtain the phase fluctuation gradient. The matrix elements in the cross power matrix that correspond to phase fluctuation gradients greater than a preset gradient threshold are set to zero to obtain a purified cross matrix. The amplitude sequence of the purified cross matrix is ​​extracted, and the average amplitude of each frequency point in the amplitude sequence is calculated along the time dimension to obtain the average cross amplitude spectrum curve.

5. The machine learning prediction system for chemical product quality according to claim 4, characterized in that, The step of performing conjugate multiplication of the complex values ​​of the parameter time spectrum and the quality time spectrum to obtain the cross power matrix includes: The complex values ​​of the parameter time spectrum and the mass time spectrum are decomposed in polar coordinates to extract the parameter amplitude matrix and parameter phase matrix of the parameter time spectrum, and the mass amplitude matrix and mass phase matrix of the mass time spectrum, respectively. The parameter amplitude matrix and the mass amplitude matrix are multiplied by the Hadamard product to obtain the joint energy matrix, and the parameter phase matrix and the mass phase matrix are interpolated by the element-wise difference to obtain the time-delay phase matrix. Substitute each element of the time-delay phase matrix into Euler's formula to generate a complex phase matrix, and multiply the joint energy matrix and the complex phase matrix element by element to obtain the cross power matrix.

6. The machine learning prediction system for chemical product quality according to claim 5, characterized in that, The step of performing polar coordinate decomposition on the complex values ​​of the parameter time spectrum and the quality time spectrum to extract the parameter amplitude matrix and parameter phase matrix of the parameter time spectrum, respectively, includes: Extract the node complex values ​​corresponding to each time-frequency coordinate from the time-frequency spectrum of the parameters, perform vector magnitude analysis on the node complex values ​​to obtain the node magnitude sequence, and reconstruct the node magnitude sequence according to the time-frequency coordinates to obtain the parameter amplitude matrix; Argument features are extracted from the complex values ​​of the nodes to obtain the wrapped phase sequence. The phase is then continuously expanded along the time dimension of the wrapped phase sequence to obtain the deflection phase sequence. The deflection phase sequence is then reconstructed into a matrix according to the time-frequency coordinates to obtain the parameter phase matrix.

7. The machine learning prediction system for chemical product quality according to claim 1, characterized in that, The step of constructing a time-series correlation matrix based on the correlation parameter set, and then dividing the time-series correlation matrix into sliding window segments to obtain time-series window fragments includes: The parameter sequence is separated from the set of associated parameters, and the cross-correlation integral is performed on the parameter sequence to obtain the cross-correlation delay. The parameter sequence is shifted along the time axis based on the cross-correlation delay to obtain a lag-compensated sequence, and the lag-compensated sequence is then horizontally concatenated to obtain a time-series correlation matrix. The time series correlation matrix is ​​subjected to first-order difference calculation along the time axis to obtain the difference gradient matrix, and each column of the difference gradient matrix is ​​normalized by maximum absolute value to obtain the dimensionless gradient matrix. The modulus of each row is extracted from the dimensionless gradient matrix to obtain a modulus sequence. The extreme points of the modulus are extracted from the modulus sequence, and the time difference between adjacent extreme points of the modulus is calculated to obtain the dynamic sliding step size. The temporal correlation matrix is ​​segmented according to a preset fixed window length and with the dynamic sliding step size as the starting point interval between adjacent windows to obtain temporal window segments.

8. The machine learning prediction system for chemical product quality according to claim 1, characterized in that, The step of extracting parameter fluctuation features within the time-series window segment and constructing a fluctuation feature vector based on the parameter fluctuation features includes: The second derivative of the time window segment is calculated along the time dimension to obtain a local curvature sequence, which is a two-dimensional matrix where rows represent time points and columns represent parameter dimensions. The local curvature sequence is subjected to element-wise difference operation with the preset linear baseline matrix to obtain the curvature deviation matrix. The standard deviation of each column of the curvature deviation matrix is ​​normalized to obtain the dimensionless curvature deviation matrix. Singular value decomposition is performed on the dimensionless curvature deviation matrix to extract the set of non-zero singular values. Calculate the second and third central moments of the set of non-zero singular values, and then concatenate the second and third central moments numerically to obtain the parameter fluctuation characteristics; The fluctuation characteristics of the parameters are subjected to range standardization to obtain normalized feature elements, and the normalized feature elements are arranged in a preset dimension order to construct the fluctuation feature vector.

9. The machine learning prediction system for chemical product quality according to claim 8, characterized in that, The calculation of the second and third central moments of the set of non-zero singular values ​​includes: The distribution expectation is extracted from the set of non-zero singular values ​​to obtain the mean of the singular values; Based on the mean of the singular values, a centering mapping is performed on the set of non-zero singular values ​​to obtain a centered discrete sequence; The second-order distribution analysis is performed on the centered discrete sequence to obtain the second-order central moments, and the third-order distribution analysis is performed on the centered discrete sequence to obtain the third-order central moments.

10. The machine learning prediction system for chemical product quality according to claim 1, characterized in that, The step of inputting the fluctuation feature vector into a preset prediction learning model and performing feature mapping on the fluctuation feature vector to obtain a quality prediction value includes: The fluctuation feature vector is input into a preset prediction learning model, which includes a sparse autoencoder module, a cross-dimensional attention regression module, and a quantile residual correction module. The wave feature vector is sparsely constrained and encoded by a sparse autoencoder module to obtain a hidden layer activation feature map. Based on the cross-dimensional attention regression module, the hidden layer activation feature map is subjected to multi-head attention decomposition and nonlinear regression to obtain a quality-sensitive score vector. The quality-sensitive score vector is input into the quantile residual correction module for quantile regression and residual compensation to obtain the quality prediction value.

Images