Dry-process acetylene generator material component soft measurement method

Abstract

The application discloses a kind of dry-process acetylene generator material component soft measurement method, belong to chemical production process parameter detection technical field.The method is first based on calcium carbide hydrolysis reaction mechanism screening can be detected on-line auxiliary variable, match historical assay data and corresponding time period process parameters, after data cleaning and time alignment processing construction soft measurement training data set;Secondly, using mobile time window and time decomposition processing extracts multiple data features, separates the steady component and dynamic component in data to obtain multi-scale feature set;Again, combined with the two types of calculation results of mechanism balance and data fitting, the material component prediction result is obtained by weighted fusion;Finally, the parameter maintenance mechanism of periodic correction and incremental update is established.The method can realize the continuous on-line detection of core material components, eliminate the time lag of artificial assay, adapt to complex closed production conditions, reduce the input cost of detection link, and provide stable data support for real-time control of production process.

Description

Technical Field

[0001] This invention relates to the field of chemical production process parameter detection technology, and in particular to a soft measurement method for material components in a dry acetylene generator. Background Technology

[0002] Dry acetylene production is an important form of acetylene preparation via the calcium carbide process. This process centers on the hydrolysis of calcium carbide with water, and the resulting acetylene gas is a fundamental raw material for various chemical products such as polyvinyl chloride (PVC) and acetylene black, possessing broad industrial application value in chlor-alkali and fine chemical industries. Currently, the operation and control of the dry acetylene production process mainly rely on various online-acquisitionable process parameters. Parameter acquisition systems are widely deployed at production sites, enabling real-time acquisition of various operational data such as feed flow rate, reaction temperature, equipment pressure, and operating current, providing a data foundation for the digital management and control of the production process. In the field of industrial process parameter detection, soft measurement technology has become a mature technology. Mechanism balance, time-series feature extraction, and data fitting calculation techniques have been thoroughly studied, providing indirect calculation pathways for process parameters that are difficult to detect directly. These methods have been implemented in multiple chemical production scenarios.

[0003] Currently, the detection of core material components in the dry acetylene production process faces several limitations. Manual sampling and analysis require periodic sample collection and laboratory analysis, resulting in significant time intervals in data acquisition and failing to provide continuous input for real-time process control. Dedicated online analytical equipment involves high procurement and maintenance costs; in high-temperature, high-dust, and enclosed reaction conditions, critical components are susceptible to environmental influences, leading to insufficient operational stability and difficulty in long-term reliable operation. Existing soft-sensor calculation methods often employ single-mechanism calculations or single data fitting paths, failing to simultaneously consider the physical rationality of the calculation and its dynamic adaptability to operating conditions, and insufficiently exploring the steady-state and dynamic characteristics of process parameters. Furthermore, as production progresses, equipment status and operating conditions gradually change; calculation parameters cannot automatically adjust to these changes, leading to increasing deviations in calculation results and making it difficult to maintain stable calculation accuracy over the long term. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a soft measurement method for material components in a dry acetylene generator.

[0005] The objective of this invention is achieved through the following technical solution: A soft measurement method for material composition in a dry acetylene generator is provided, the method comprising the following steps: S1. Based on the mechanism of calcium carbide hydrolysis reaction, auxiliary variables were screened, including calcium carbide feed rate, water injection flow rate, reaction temperature, stirring current, generator pressure, and acetylene gas production flow rate; historical test data and corresponding process parameters were matched, and after data cleaning and time-series alignment, a soft measurement training dataset was constructed. S2. The moving time window algorithm is used to process the time series data of the soft measurement training dataset to extract parameter trend features, fluctuation features and rate of change features; the ensemble empirical mode decomposition method is used to separate the steady-state components and dynamic components of the data to obtain a multi-scale feature set; S3. Calculate the basic values ​​of the components based on the material balance and energy balance of the calcium carbide hydrolysis reaction; use a long short-term memory network and support vector regression to fit the mapping relationship between the auxiliary variables and the target components, and calculate the fitted values ​​of the components; perform a weighted summation of the basic values ​​of the components and the fitted values ​​of the components to obtain the predicted results of the material components. S4. Set the calibration cycle and deviation threshold, and correct the calculation parameters after obtaining new manual test data; set the update cycle, accumulate running data and update the fitting calculation parameters to maintain the deviation of the calculation results within the set range.

[0006] Furthermore, step S1 includes the following sub-steps: S1.1. Based on the analysis of the influencing factors of material composition according to the mechanism of calcium carbide hydrolysis reaction, and combined with the process operation records, screen out online detectable parameters that are related to the material composition to form an initial set of auxiliary variables; S1.2. Collect historical test data and corresponding process operation parameters within a set time range, and perform time sequence alignment according to the timestamp; S1.3. Perform outlier removal and missing value imputation on the aligned dataset. Outlier removal uses the Laida criterion to identify and remove outlier data that exceeds the set range. Missing value imputation uses linear interpolation to fill in missing data, resulting in a soft measurement training dataset.

[0007] Furthermore, step S2 includes the following sub-steps: S2.1. Set the window width and sliding step size of the moving time window, and segment the time series data of the soft measurement training dataset according to the set parameters; S2.2. Perform statistical calculations on the parameter data within each time window, extract the trend characteristics, fluctuation characteristics, and rate of change characteristics of the corresponding parameters, and form a single-scale feature set; S2.3. The ensemble empirical mode decomposition method is used to decompose the single-scale feature set, separate the steady-state components and dynamic components in the data, and combine them to form a multi-scale feature set.

[0008] Furthermore, step S3 includes the following sub-steps: S3.1. Based on the material balance and energy balance of the calcium carbide hydrolysis reaction, establish material balance calculation formulas and energy balance calculation formulas. The material balance calculation formula outputs the mass change data of the reactants and inputs it into the energy balance calculation formula. The energy balance calculation formula outputs the enthalpy change data of the reaction and inputs it into the material balance calculation formula. The basic values ​​of the components are calculated. S3.2. Input the multi-scale feature set into the long short-term memory network to extract the time-dependent features, input the time-dependent features into support vector regression for nonlinear mapping fitting, and calculate the component fitting values; S3.3. Set the weighting coefficients for the basic component values ​​and the fitted component values, and perform a weighted summation of the basic component values ​​and the fitted component values ​​according to the weighting coefficients to obtain the predicted material composition results.

[0009] Furthermore, step S4 includes the following sub-steps: S4.1. Set the calibration cycle and deviation threshold, obtain new manual test data according to the calibration cycle, compare the manual test data with the material composition prediction results of the corresponding time period, and calculate the prediction deviation value. S4.2. When the prediction deviation exceeds the deviation threshold, the material balance calculation parameters, energy balance calculation parameters, and support vector regression parameters are corrected by a correction algorithm. S4.3. Set the update cycle, accumulate running data according to the update cycle, add the newly added running data and corresponding test data to the soft measurement training dataset, and perform incremental fitting and updating of the long short-term memory network and support vector regression.

[0010] Furthermore, in step S1.1, the data of each parameter in the initial auxiliary variable set is standardized to eliminate the influence of dimensional differences on the calculation results; the correlation degree between each initial auxiliary variable and the material component is calculated using the Pearson correlation analysis method, and the multicollinearity coefficient between the initial auxiliary variables is also calculated. Based on the correlation degree and multicollinearity coefficient, redundant variables are removed from the initial auxiliary variable set, and parameters whose correlation degree and multicollinearity coefficient meet the set conditions are retained to form the final auxiliary variable set.

[0011] Furthermore, in step S2.1, the moving time window adopts a variable window width setting method; at least two sets of window width values ​​and corresponding fluctuation degree judgment conditions are preset, the standard deviation of the process parameters in the current time period is calculated in real time as the fluctuation degree characterization value, the calculated fluctuation degree characterization value is matched with each fluctuation degree judgment condition, the window width value corresponding to the successfully matched judgment condition is selected, and the time series data of the soft measurement training dataset is segmented and truncated.

[0012] Furthermore, in step S2.3, the processing procedure of the set empirical mode decomposition method is as follows: each dimension feature in the single-scale feature set is decomposed sequentially, white noise with a set amplitude is added during the decomposition process, the decomposition operation is repeated at least twice and the average value of the calculation results is taken to obtain the steady-state component and dynamic component of the corresponding dimension, and the steady-state component and dynamic component of all dimensions are combined to form a multi-scale feature set.

[0013] Furthermore, in step S3.2, the Long Short-Term Memory (LSTM) network includes an input layer, a set number of hidden layers, and an output layer. The hidden layers are configured with a forget gate, an input gate, and an output gate. Support Vector Regression (SVR) uses a radial basis function kernel to construct a mapping space. The output of the LSM network is connected to the input of SVR. The multi-scale feature set is divided into a training subset and a validation subset. The training subset is used to train the LSM network for temporal feature extraction. The temporal dependency features output by the LSM network are input into SVR for mapping training. The validation subset is used to verify the accuracy of the combined structure of the LSM network and SVR. The parameters of the LSM network and SVR are adjusted until the validation results meet the set conditions.

[0014] Furthermore, in step S4.3, the incremental fitting update uses a sliding window method to update the soft measurement training dataset; historical data exceeding the set time range are removed according to the set time step, and newly added running data and manual test data of the corresponding time period are added simultaneously to maintain the sample size of the soft measurement training dataset within the set range; during the update process, the underlying network parameters of the Long Short-Term Memory Network and the kernel function parameters of the Support Vector Regression are retained, and only the weight parameters of the top output layer of the Long Short-Term Memory Network and the bias parameters of the Support Vector Regression are adjusted and updated to complete the incremental fitting update process.

[0015] The beneficial effects of this invention are: (1) By screening process parameters that can be detected online, a training dataset is constructed. The fusion calculation logic of mechanism balance and data fitting is combined to realize the continuous online calculation of material components, which replaces the detection mode of manual sampling and testing and eliminates the time lag problem caused by manual testing. (2) By adopting a multi-scale data feature extraction and weighted fusion processing method, the steady-state component and dynamic component in the data are separated, which can be adapted to different operating conditions with different fluctuation levels and improve the adaptability and stability of the component calculation results in the process of changing operating conditions. (3) Establish a parameter maintenance mechanism for regular calibration and incremental updates, correct the calculation parameters based on the newly added test data, compensate for the calculation deviation caused by working condition drift and equipment aging, continuously maintain the reasonableness of the calculation results, and reduce the manual input in the testing process. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of a soft measurement method for material components in a dry acetylene generator; Figure 2 The following is a flowchart illustrating the specific steps of a soft measurement method for material components in a dry acetylene generator, provided as an example. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 See Figure 1 This embodiment provides a soft measurement method for material components in a dry acetylene generator, which includes the following steps: S1. Based on the mechanism of calcium carbide hydrolysis reaction, auxiliary variables were screened, including calcium carbide feed rate, water injection flow rate, reaction temperature, stirring current, generator pressure, and acetylene gas production flow rate; historical test data and corresponding process parameters were matched, and after data cleaning and time-series alignment, a soft measurement training dataset was constructed. S2. The moving time window algorithm is used to process the time series data of the soft measurement training dataset to extract parameter trend features, fluctuation features and rate of change features; the ensemble empirical mode decomposition method is used to separate the steady-state components and dynamic components of the data to obtain a multi-scale feature set; S3. Calculate the basic values ​​of the components based on the material balance and energy balance of the calcium carbide hydrolysis reaction; use a long short-term memory network and support vector regression to fit the mapping relationship between the auxiliary variables and the target components, and calculate the fitted values ​​of the components; perform a weighted summation of the basic values ​​of the components and the fitted values ​​of the components to obtain the predicted results of the material components. S4. Set the calibration cycle and deviation threshold, and correct the calculation parameters after obtaining new manual test data; set the update cycle, accumulate running data and update the fitting calculation parameters to maintain the deviation of the calculation results within the set range.

[0019] In some embodiments, step S1 includes the following sub-steps: S1.1. Based on the analysis of the influencing factors of material composition according to the mechanism of calcium carbide hydrolysis reaction, and combined with the process operation records, screen out online detectable parameters that are related to the material composition to form an initial set of auxiliary variables; S1.2. Collect historical test data and corresponding process operation parameters within a set time range, and perform time sequence alignment according to the timestamp; S1.3. Perform outlier removal and missing value imputation on the aligned dataset. Outlier removal uses the Laida criterion to identify and remove outlier data that exceeds the set range. Missing value imputation uses linear interpolation to fill in missing data, resulting in a soft measurement training dataset.

[0020] In some embodiments, step S2 includes the following sub-steps: S2.1. Set the window width and sliding step size of the moving time window, and segment the time series data of the soft measurement training dataset according to the set parameters; S2.2. Perform statistical calculations on the parameter data within each time window, extract the trend characteristics, fluctuation characteristics, and rate of change characteristics of the corresponding parameters, and form a single-scale feature set; S2.3. The ensemble empirical mode decomposition method is used to decompose the single-scale feature set, separate the steady-state components and dynamic components in the data, and combine them to form a multi-scale feature set.

[0021] In some embodiments, step S3 includes the following sub-steps: S3.1. Based on the material balance and energy balance of the calcium carbide hydrolysis reaction, establish material balance calculation formulas and energy balance calculation formulas. The material balance calculation formula outputs the mass change data of the reactants and inputs it into the energy balance calculation formula. The energy balance calculation formula outputs the enthalpy change data of the reaction and inputs it into the material balance calculation formula. The basic values ​​of the components are calculated. S3.2. Input the multi-scale feature set into the long short-term memory network to extract the time-dependent features, input the time-dependent features into support vector regression for nonlinear mapping fitting, and calculate the component fitting values; S3.3. Set the weighting coefficients for the basic component values ​​and the fitted component values, and perform a weighted summation of the basic component values ​​and the fitted component values ​​according to the weighting coefficients to obtain the predicted material composition results.

[0022] In some embodiments, step S4 includes the following sub-steps: S4.1. Set the calibration cycle and deviation threshold, obtain new manual test data according to the calibration cycle, compare the manual test data with the material composition prediction results of the corresponding time period, and calculate the prediction deviation value. S4.2. When the prediction deviation exceeds the deviation threshold, the material balance calculation parameters, energy balance calculation parameters, and support vector regression parameters are corrected by a correction algorithm. S4.3. Set the update cycle, accumulate running data according to the update cycle, add the newly added running data and corresponding test data to the soft measurement training dataset, and perform incremental fitting and updating of the long short-term memory network and support vector regression.

[0023] In some embodiments, in step S1.1, the data of each parameter in the initial auxiliary variable set is standardized to eliminate the influence of dimensional differences on the calculation results; the correlation degree between each initial auxiliary variable and the material component is calculated using the Pearson correlation analysis method, and the multicollinearity coefficient between the initial auxiliary variables is calculated at the same time. Based on the correlation degree value and the multicollinearity coefficient, redundant variables are removed from the initial auxiliary variable set, and parameters whose correlation degree value and multicollinearity coefficient meet the set conditions are retained to form the final auxiliary variable set.

[0024] In some embodiments, in step S2.1, the moving time window adopts a variable window width setting method; at least two sets of window width values ​​and corresponding fluctuation degree judgment conditions are preset, the standard deviation of the process parameters in the current time period is calculated in real time as the fluctuation degree characterization value, the calculated fluctuation degree characterization value is matched with each fluctuation degree judgment condition, the window width value corresponding to the successfully matched judgment condition is selected, and the time series data of the soft measurement training dataset is segmented and truncated.

[0025] In some embodiments, in step S2.3, the processing procedure of the set empirical mode decomposition method is as follows: each dimension feature in the single-scale feature set is decomposed sequentially, white noise of a set amplitude is added during the decomposition process, the decomposition operation is repeated at least twice and the average value of the calculation results is taken to obtain the steady-state component and dynamic component of the corresponding dimension, and the steady-state component and dynamic component of all dimensions are combined to form a multi-scale feature set.

[0026] In some embodiments, in step S3.2, the Long Short-Term Memory (LSTM) network includes an input layer, a set number of hidden layers, and an output layer. The hidden layers are configured with a forget gate, an input gate, and an output gate. Support Vector Regression (SVR) uses a radial basis function kernel to construct a mapping space. The output of the LSM network is connected to the input of SVR. The multi-scale feature set is divided into a training subset and a validation subset. The training subset is used to train the LSM network for temporal feature extraction. The temporal dependent features output by the LSM network are input into SVR for mapping training. The validation subset is used to verify the accuracy of the combined structure of the LSM network and SVR. The parameters of the LSM network and SVR are adjusted until the validation results meet the set conditions.

[0027] In some embodiments, in step S4.3, the incremental fitting update uses a sliding window method to update the soft measurement training dataset; historical data exceeding the set time range are removed according to the set time step, and newly added running data and manual test data of the corresponding time period are added simultaneously to maintain the sample size of the soft measurement training dataset within the set range; during the update process, the underlying network parameters of the Long Short-Term Memory Network and the kernel function parameters of the Support Vector Regression are retained, and only the weight parameters of the top output layer of the Long Short-Term Memory Network and the bias parameters of the Support Vector Regression are adjusted and updated to complete the incremental fitting update process.

[0028] Example 2 This embodiment provides a specific implementation process for a soft measurement method of material components in a dry acetylene generator. This method continuously calculates and outputs the material components in the dry acetylene production process by combining mechanistic calculation rules and data mapping rules, obtaining component data without relying on additional online analysis equipment. Figure 2 As shown, the specific implementation process is as follows: Step 1. Selection of auxiliary variables and construction of dataset: Step 1.1 Initial selection of auxiliary variables: Based on the analysis of the influencing factors of material composition according to the calcium carbide hydrolysis reaction mechanism, and combined with process operation records, online detectable parameters correlated with material composition were screened to form an initial set of auxiliary variables. The auxiliary variables include calcium carbide feed rate, water injection flow rate, reaction temperature, stirring current, generator pressure, and acetylene gas production flow rate. The data for each parameter in the initial set of auxiliary variables were standardized to eliminate the influence of dimensional differences on the calculation results. Pearson correlation analysis was used to calculate the correlation degree between each initial auxiliary variable and the material composition. Pearson correlation analysis is a statistical analysis method that measures the degree of linear correlation between two variables; in this embodiment, it is used to quantify the correlation strength between process parameters and material composition, thereby determining whether a parameter meets the conditions for serving as an auxiliary calculation input.

[0029] Simultaneously, the multicollinearity coefficient among the initial auxiliary variables is calculated. Multicollinearity refers to the phenomenon where a high correlation exists between input variables in a linear calculation relationship, leading to distorted calculation results. In this embodiment, redundant relationships between parameters are identified by calculating the multicollinearity coefficient. Redundant variables are removed from the initial auxiliary variable set based on the correlation degree value and the multicollinearity coefficient, retaining parameters whose correlation degree value and multicollinearity coefficient both meet the set conditions, forming the final auxiliary variable set. Standardization and variable screening can reduce the interference of irrelevant and redundant parameters on the calculation process, improving the efficiency and stability of subsequent calculations.

[0030] Step 1.2 Timing data alignment processing: Historical laboratory data and corresponding process operating parameters within a defined time range are collected and time-series aligned according to timestamps. Time-series alignment is a process of matching time-series data from different sources to a unified time benchmark. In this embodiment, it ensures that each set of laboratory data corresponds to process operating parameters within the same time period, establishing a correspondence between process parameters and material components, and providing an accurate matching basis for subsequent dataset construction. During the alignment process, the sampling time of the laboratory data is used as the benchmark, and process operating parameters within the corresponding time interval are selected for matching to ensure that the time range of each set of data remains consistent.

[0031] Step 1.3 Dataset Cleaning and Construction: The aligned dataset undergoes outlier removal and missing value imputation for data cleaning. Outlier removal uses the Laida criterion to identify and remove outliers exceeding a set range. The Laida criterion is an outlier determination criterion based on the assumption of normal distribution. It calculates the mean and standard deviation of the data, classifying data deviating from the mean by more than a set multiple of the standard deviation as outliers. In this embodiment, it is used to identify and remove outlier sample values ​​in process parameters and laboratory data, preventing outliers from interfering with subsequent calculations. Missing value imputation uses linear interpolation to fill in missing data. Linear interpolation is a data imputation method that calculates the value of the missing point based on the values ​​of two adjacent known data points using a linear function. In this embodiment, it is used to fill in missing data caused by equipment communication fluctuations during data acquisition, maintaining the temporal continuity of the soft measurement training dataset. After data cleaning, the soft measurement training dataset is obtained.

[0032] In some embodiments, missing value imputation can be performed by averaging neighboring values, selecting the average of a set number of valid data before and after the missing position as the imputation value. This method is suitable for situations where the proportion of missing data is small and the parameter operation is stable.

[0033] In some specific implementations, the soft measurement training dataset is stored in a structured table format. Each row of data corresponds to a complete set of parameters for a sampling time, and each column of data corresponds to a type of auxiliary variable or test component data. During the time-series alignment process, the sampling time of the test data is used as the reference time point. Process operating parameters within a set time range before and after the reference time point are selected, and the average value of the process operating parameters within the corresponding time range is calculated as the process parameter value for the corresponding reference time point, thus completing the matching of a single set of test data with process parameters.

[0034] During data cleaning, all parameter data are first traversed, and the overall mean and standard deviation of each parameter category are calculated one by one. Data points whose absolute difference from the mean is greater than three times the standard deviation are marked as outliers, and these outliers are removed. After outlier removal, missing data locations are identified in the data sequence, and the number of consecutively missing data points is counted. When the number of consecutively missing data points is below a set threshold, linear interpolation is used to fill in the missing data; when the number of consecutively missing data points is above the set threshold, the corresponding data sequence is discarded and not included in the soft measurement training dataset. The cleaned dataset is sorted chronologically and divided into training, validation, and test subsets according to a set ratio. The training subset is used for parameter adjustment during the fitting calculation process, the validation subset is used for accuracy verification during training, and the test subset is used for verification of the final calculation results. After the dataset is constructed, the value range of each parameter category is recorded for input data validity verification during subsequent online calculations.

[0035] In some specific implementations, a secondary verification and normalization process for the input data is added to address calculation biases caused by random spike noise and dimensional differences in the field-acquired data. After data acquisition, a first-order moving average filter is first performed to smooth continuous process parameter data and suppress random spike noise generated by electromagnetic interference from the acquisition equipment. After filtering, a minimum-maximum normalization method is used to perform a linear transformation on each type of parameter data, mapping all parameter values ​​to a unified range and eliminating the impact of differences in the dimensions and value ranges of different parameters on feature extraction and mapping calculations.

[0036] The parameter range for normalization is set based on the extreme values ​​of historical operating data, while reserving a certain margin to accommodate parameter out-of-bounds situations during fluctuations in operating conditions. During data verification, parameter data exceeding the historical extreme value range are marked. If the number of out-of-bounds parameters in a single sample group is less than a set proportion, the out-of-bounds parameters are replaced using boundary values; if the number of out-of-bounds parameters in a single sample group is greater than a set proportion, the sample group is determined to be invalid data and directly removed from the dataset. This processing flow can improve the quality of input data, reduce the interference of noisy data and dimensional differences on subsequent calculation results, and improve the stability of component calculation results.

[0037] In some specific implementations, as shown in Table 1, the soft measurement training dataset is divided into five categories according to function. Each category corresponds to different sample proportions, batch sample size, feature dimensions, and sampling interval parameters. The training subset has a sample proportion of 70%, a batch sample size of 16 groups, a feature dimension of 12, and a sampling interval of 1 minute. The validation subset has a sample proportion of 20%, a batch sample size of 8 groups, a feature dimension of 12, and a sampling interval of 1 minute. The test subset has a sample proportion of 10%, uses full sample validation, a feature dimension of 12, and a sampling interval of 1 minute. The incremental update set adds 30 groups of samples at a time, has a feature dimension of 12, and a sampling interval of 1 minute. The validation dataset has a fixed sample size of 50 groups, a feature dimension of 12, and a sampling interval of 1 minute. After the dataset is divided according to the above parameters, the training subset and the validation subset jointly participate in the iterative training process of the model. Each training round inputs data in batches according to the single batch sample size. After training, the validation subset is used to calculate the current accuracy. Training is terminated when the accuracy does not improve for several consecutive rounds to avoid overfitting. The test subset is used independently after the entire training process is completed to verify the overall computational structure's adaptability to data that did not participate in the training. The incremental update set is gradually accumulated as production runs and is added to the training set at fixed intervals to support dynamic parameter adjustment. The validation dataset does not participate in any training or updates. It is only used to verify the correction effect after the correction operation is completed to ensure that the corrected parameters have universal adaptability.

[0038] Table 1 Soft Measurement Dataset Configuration Parameter Table Step 2. Multi-scale feature extraction processing: Step 2.1 Time series data segmentation: The moving time window algorithm sets the window width and sliding step size, and segments the time-series data of the soft measurement training dataset according to the set parameters. The moving time window algorithm is a segmentation algorithm for long time-series data. It divides continuous long time-series data into multiple short, contiguous data segments by sliding a window of a set width along the time axis with a fixed step size. In this embodiment, it is used to segment the time-series data of the soft measurement training dataset to extract data features within different time periods. The moving time window adopts a variable window width setting. At least two sets of window width values ​​and corresponding fluctuation degree judgment conditions are preset. The standard deviation of the process parameters in the current time period is calculated in real time as the fluctuation degree characterization value. The calculated fluctuation degree characterization value is matched with each fluctuation degree judgment condition. The window width value corresponding to the successfully matched judgment condition is selected to perform segmentation and truncation operations on the time-series data of the soft measurement training dataset. The variable window width setting can adjust the segmentation granularity according to the parameter fluctuation degree. When the parameter fluctuation degree is large, a narrower window width is used to improve feature resolution; when the parameter fluctuation degree is small, a wider window width is used to improve feature stability.

[0039] In some specific implementations, as shown in Table 2, the moving time window sets corresponding parameters for window width, sliding step size, number of decompositions, and number of features for five types of operating conditions. Under the basic operating condition, the window width is 30 sampling points, the sliding step size is 5 sampling points, the number of decompositions is 10, and the number of extracted features is 18. Under the fluctuating operating condition, the window width is 15 sampling points, the sliding step size is 3 sampling points, the number of decompositions is 10, and the number of extracted features is 18. Under the stable operating condition, the window width is 60 sampling points, the sliding step size is 10 sampling points, the number of decompositions is 10, and the number of extracted features is 18. Under the start-stop operating condition, the window width is 20 sampling points, the sliding step size is 2 sampling points, the number of decompositions is 8, and the number of extracted features is 12. Under the verification operating condition, the window width is 30 sampling points, the sliding step size is 5 sampling points, the number of decompositions is 10, and the number of extracted features is 18. During feature extraction, the fluctuation level of process parameters for the current time period is first calculated in real time. The time-series data is then segmented by matching the window and step size parameters to the corresponding operating conditions. Next, three types of basic features are statistically extracted from the data within each time window. Subsequently, the steady-state and dynamic components are separated according to the corresponding number of decomposition steps, ultimately yielding a corresponding number of multi-scale features. Differentiated operating condition parameter configurations can reduce the window size to improve time resolution during fluctuation phases, expand the window to improve statistical stability during stable phases, and simplify feature dimensions to reduce computational load during start-up and shutdown phases, thus balancing feature extraction accuracy and computational efficiency under different operating conditions.

[0040] Table 2 Parameter Table for Moving Time Window Step 2.2 Single-scale feature extraction calculation: Statistical calculations are performed on the parameter data within each time window to extract the corresponding trend characteristics, fluctuation characteristics, and rate of change characteristics, forming a single-scale feature set. Trend characteristics reflect the overall trend of parameter change within the time window; the slope of the linear fit is used as the numerical representation of the trend characteristic. Fluctuation characteristics reflect the dispersion of the parameter within the time window; the standard deviation of the data within the time window is used as the numerical representation of the fluctuation characteristic. Rate of change characteristics reflect the magnitude of parameter change per unit time; the ratio of the difference between the first and last data points within the time window to the width of the time window is used as the numerical representation of the rate of change characteristic. These three types of features reflect the operating status of process parameters from different dimensions, providing multi-dimensional input data for subsequent mapping calculations and improving the adaptability of the mapping relationship.

[0041] Step 2.3 Multi-scale feature decomposition and combination: A ensemble empirical mode decomposition (EMD) method is employed to decompose a single-scale feature set, separating the steady-state and dynamic components of the data, and combining them to form a multi-scale feature set. EMD is an adaptive time-series signal decomposition method. It suppresses mode aliasing during decomposition by adding white noise to the original signal and repeatedly performing the decomposition and averaging the results. This decomposes the original signal into multiple intrinsic mode components and residual components. In this embodiment, it is used to separate the steady-state operating components and dynamic fluctuation components in the single-scale feature data. The steady-state components correspond to the feature baseline under stable process operation, while the dynamic components correspond to the feature fluctuations during the process of changing operating conditions. The EMD process involves sequentially decomposing each dimension of the single-scale feature set, adding white noise of a set amplitude during the decomposition process, repeating the decomposition operation at least twice, and averaging the results to obtain the steady-state and dynamic components of the corresponding dimension. The steady-state and dynamic components of all dimensions are then combined to form a multi-scale feature set. This multi-scale feature set can simultaneously reflect the baseline and fluctuation states of the process operation, improving the adaptability of subsequent component calculations to different operating conditions.

[0042] In some embodiments, multi-scale feature decomposition can be performed using wavelet decomposition. By setting wavelet basis functions of a certain order, the single-scale feature set is decomposed into multiple layers to obtain approximate components and detail components of the corresponding scale. The approximate components are used as steady-state components, and the detail components are used as dynamic components, which are then combined to form a multi-scale feature set.

[0043] In some specific implementations, the soft measurement training dataset is constructed by matching process data with corresponding batches of laboratory data at a fixed sampling frequency. The sampling interval for process operating parameters is set to 1 minute, and the amount of process data samples that can be collected per day meets the input requirements for continuous computation. Laboratory data is sampled according to fixed shifts, and each set of laboratory data corresponds to process operating parameters within a 30-minute range before and after it. The arithmetic mean of the process operating parameters within this time range is calculated as the input feature data for the corresponding laboratory sample.

[0044] During the data cleaning phase, linear interpolation is used to fill in no more than 5 consecutive missing data points. If the number of consecutive missing data points exceeds 5, the complete data sequence for the corresponding time period is discarded to avoid interpolation errors affecting the dataset quality. After cleaning, the dataset is divided into training, validation, and test subsets in a 7:2:1 ratio. The training subset is used for iterative adjustment of fitting calculation parameters, the validation subset is used for accuracy monitoring and early stopping detection during training, and the test subset is used for independent verification of the final computational performance.

[0045] The basic window width of the moving time window is set to 30 sampling points, and the sliding step size is set to 5 sampling points. When the fluctuation level of the process parameters exceeds the first judgment threshold, the window width is adjusted to 15 sampling points, and the sliding step size is adjusted to 3 sampling points to improve the feature time resolution under fluctuating conditions. When the fluctuation level of the process parameters is lower than the second judgment threshold, the window width is adjusted to 60 sampling points, and the sliding step size is adjusted to 10 sampling points to improve the feature statistical stability under stable conditions. The white noise amplitude added during the ensemble empirical mode decomposition process is set to 0.2 times the standard deviation of the original signal, and the number of decomposition repetitions is set to 10. The average value of the multiple decomposition results is taken as the final component output to suppress the influence of mode mixing on the feature separation effect.

[0046] Step 3. Component fusion calculation output: Step 3.1 Numerical calculation of the mechanism basis: Based on the material and energy balance relationships in the calcium carbide hydrolysis reaction, material and energy balance calculation formulas are established. The material balance calculation formula outputs the mass change data of the reactants and inputs it into the energy balance calculation formula, while the energy balance calculation formula outputs the enthalpy change data of the reaction and inputs it into the material balance calculation formula, thus calculating the basic component values. Material balance is a method for balancing the input, output, and accumulation of each material in the reaction system based on the law of conservation of mass. In this embodiment, it is used to calculate the theoretical component values ​​of the reaction products based on the calcium carbide feed rate and water injection flow rate, establishing the basic framework for component calculation.

[0047] Energy balance is a method for balancing heat input, heat output, and heat accumulation within a reaction system based on the law of conservation of energy. In this embodiment, it is used to calculate the actual reaction progress based on reaction temperature data and to correct the results of the material balance calculation. The material balance and energy balance calculations are coupled. The mass change data of reactants output by the material balance calculation is input into the energy balance calculation to calculate the heat released during the reaction process; the enthalpy change data of the reaction output by the energy balance calculation is input into the material balance calculation to correct the reaction rate coefficient. This achieves iterative coupled calculation between the material balance and energy balance calculations, improving the matching degree between the basic component values ​​and the actual reaction state.

[0048] Step 3.2 Data mapping and fitting calculation: The multi-scale feature set is input into a long short-term memory network to extract time-dependent features. The time-dependent features are then input into support vector regression for nonlinear mapping fitting, and the component fitting values ​​are calculated.

[0049] Long Short-Term Memory (LSTM) networks are recurrent neural network structures with long-sequence memory capabilities. An LTM network includes an input layer, a predetermined number of hidden layers, and an output layer. The hidden layers employ gating structures such as a forget gate, an input gate, and an output gate. The forget gate filters out historical information to be discarded, the input gate filters out new information to be retained, and the output gate outputs the feature results at the current time. In this embodiment, the LTM network is used to extract temporal dependencies from a multi-scale feature set, capturing the regular characteristics of process parameters changing over time. Support Vector Regression (SVR) is a nonlinear regression method based on statistical learning theory. SVR uses a radial basis function (RBF) kernel to construct a mapping space. The RBF kernel is a commonly used kernel function in SVR, which can map low-dimensional input data to a high-dimensional feature space, constructing a linear regression function in the high-dimensional space to fit the nonlinear relationship. In this embodiment, SVR is used to establish the mapping relationship between temporal dependent features and target components, outputting the fitted values ​​of the components.

[0050] The output of the Long Short-Term Memory (LSTM) network is connected to the input of Support Vector Regression (SVR), and the temporal dependency features output by the LTM network are directly used as input data for SVR. The multi-scale feature set is divided into training and validation subsets. The training subset is used to train the LTM network for temporal feature extraction. The temporal dependency features output by the LTM network are then input into SVR for mapping training. The validation subset is used to validate the accuracy of the combined LTM and SVR structure. The parameters of both the LTM network and SVR are adjusted until the validation results meet the set conditions.

[0051] In some embodiments, time-series dependency feature extraction can be performed using a gated recurrent unit network. The gated recurrent unit network includes a gated structure with update gates and reset gates. The update gate controls the degree of retention of the previous time-state information, and the reset gate controls the degree of fusion of the current time-input information, thereby extracting dependency features from the time-series data.

[0052] In some implementations, the training process of the Long Short-Term Memory (LSTM) network is performed using a combination of forward and backward propagation. The number of nodes in the input layer of the LSM network is consistent with the number of feature dimensions in the multi-scale feature set. The hidden layers are configured with a two-layer structure, with the number of nodes in each hidden layer set to a predetermined value. The number of nodes in the output layer is consistent with the number of categories of the target component. During training, the mean squared error loss function is used to calculate the deviation between the predicted and actual values. The mean squared error loss function is a commonly used loss calculation function in regression calculation tasks. It calculates the loss value by averaging the squared differences between the predicted and actual values, and is used to measure the degree of deviation between the predicted and actual results.

[0053] The optimizer employs an adaptive moment estimation optimizer (IME) to iteratively update the network parameters. IEM is a commonly used parameter optimization algorithm in deep learning training. IEM improves the convergence speed and stability of the training process by adaptively adjusting the learning rate of each parameter. After each complete training subset iteration, a validation subset is used to calculate the accuracy of the combined structure of the Long Short-Term Memory (LSTM) network and Support Vector Regression (SVR). If the validation accuracy fails to improve for several consecutive iterations, the training process is terminated early, and the current network and regression parameters are saved to avoid overfitting. After training, a test subset is used to verify the generalization ability of the overall computational structure. Once the deviation of the calculation results is confirmed to be within a set range, the corresponding parameters are used in the online computation process.

[0054] In some specific implementations, the Long Short-Term Memory (LSTM) network has 12 input layer nodes, corresponding to all feature dimensions of the multi-scale feature set; two hidden layers with 32 nodes each; and two output layer nodes, corresponding to the outputs of the two target material components. During training, the batch size is set to 16, the initial learning rate to 0.001, the learning rate decay factor to 0.9 times every 10 epochs, and the total number of training epochs to 200. An early stopping mechanism is triggered when the validation set loss does not decrease for 15 consecutive epochs, terminating the training process and saving the network parameters for the optimal epoch. The penalty coefficient for Support Vector Regression is set to 100, the width parameter of the radial basis function kernel is set to 0.5, and the insensitive loss coefficient is set to 0.01. A grid search is used to traverse and optimize within the set parameter range, selecting the parameter combination with the smallest validation set error as the final calculation parameters. During the incremental fitting and update phase, the number of new samples added in a single update is set to 30 groups, and the total number of samples retained in the sliding window is set to 1000 groups. Each update removes the 30 oldest historical data groups and simultaneously adds 30 new valid data groups to maintain a stable total number of samples in the dataset. When updating parameters, only the weights and bias parameters of the Long Short-Term Memory network output layer and the bias parameters of the Support Vector Regression are adjusted; the parameters of the underlying feature extraction layer remain fixed. The number of iterations per update is set to 20, ensuring parameter adaptability while reducing computational resource consumption during the update process.

[0055] Step 3.3 Weighted output of fusion results: Weighting coefficients are set for the baseline and fitted component values. These coefficients are then weighted and summed to obtain the predicted material composition. The weighting coefficients can be preset based on the stability of the process operation. When the process is in a steady-state phase, the weight of the baseline component values ​​is increased to ensure reliable results based on the stability of the mechanism calculation. When the process is in a dynamic adjustment phase, the weight of the fitted component values ​​is increased to adapt to changes in operating conditions based on the flexibility of data mapping. This weighted fusion method combines the physical rationality of the mechanism calculation with the adaptability of the data mapping to different operating conditions, improving the reasonableness of the component calculation results under different operating states.

[0056] In some embodiments, weighted fusion can be performed using a dynamic weight allocation method, adjusting the weight coefficients of the basic component values ​​and the fitted component values ​​in real time according to the fluctuation level of the process parameters in the current period. When the fluctuation level of the parameters increases, the weight coefficient of the fitted component values ​​is increased accordingly, and when the fluctuation level of the parameters decreases, the weight coefficient of the basic component values ​​is increased accordingly.

[0057] In some specific implementations, a multi-component parallel computing architecture is adopted to simultaneously calculate two components: the gas generation rate of calcium carbide and the moisture content of calcium carbide slag, thus solving the computational delay problem caused by calculating a single component sequentially. In the parallel computing architecture, the mechanism calculation part uses a unified material balance and energy balance calculation formula as the basic framework. Corresponding correction coefficients and calculation output branches are set for the two types of target components, sharing the same set of input process parameters, and synchronously outputting the basic calculation values ​​of the two types of components.

[0058] The data mapping part adopts a structure of shared bottom-level feature extraction network and independent top-level output branches. The input layer and hidden layer of the long short-term memory network are shared by two types of components, and the temporal-dependent feature extraction of multi-scale feature sets is completed in a unified manner. The output layer is set as two independent branch structures, corresponding to the feature output of the two types of components respectively, and then connected to their respective support vector regression calculation modules to complete the final mapping calculation.

[0059] During training, a multi-task loss function is used for joint training. The predicted losses of the two components are weighted and summed according to set weights, serving as the optimization objective for the overall network. This improves overall training efficiency while ensuring the computational accuracy of individual components. The parallel computing architecture can obtain the results of multiple components simultaneously in a single input computation, reducing repetitive feature extraction and mechanism calculation processes, lowering computational resource consumption, and improving the response speed of online computation.

[0060] Step 4. Calculate parameter correction and update: Step 4.1 Prediction bias calculation and processing: Set a calibration period and deviation threshold. Acquire new manual testing data according to the calibration period, compare the manual testing data with the material composition prediction results for the corresponding time period, and calculate the prediction deviation value. The calibration period can be set according to the stability of production operation; shorten the calibration period when production conditions fluctuate frequently, and extend the calibration period when production conditions are stable. During the deviation calculation process, select the material composition prediction results that completely correspond to the time range of the manual testing data for comparison, and calculate the difference between the two sets of data as the prediction deviation value. The deviation calculation result is used to determine the suitability of the current calculation parameters.

[0061] Step 4.2 Calculate parameter deviation correction: When the prediction deviation exceeds the deviation threshold, a correction algorithm is used to adjust the parameters of the material balance calculation formula, energy balance calculation formula, and support vector regression parameters. This correction algorithm is a method for correcting calculation parameters based on measured data. In this embodiment, it is used to adjust the correction coefficients in the material balance and energy balance calculation formulas, as well as the bias parameters of support vector regression, based on manual analysis data, to compensate for calculation deviations caused by operating condition drift and equipment aging. During the correction process, the correction coefficients of the mechanism calculation are adjusted first. If the deviation still exceeds the deviation threshold after correction, the bias parameters of support vector regression are further adjusted to correct the overall calculation deviation while ensuring the rationality of the mechanism calculation. Deviation correction ensures that the calculation results continuously match the actual production state, avoiding deviations in calculation results caused by changes in equipment status and operating condition drift.

[0062] In some specific implementations, as shown in Table 3, deviation correction is set with corresponding adjustment step size, verification sample size, correction rounds, and trigger threshold parameters according to the deviation level. For Level 1 deviation, the adjustment step size is 0.01, the verification sample size is 50 sets, the correction rounds are 5 rounds, and the trigger threshold is the upper limit of the corresponding deviation. For Level 2 deviation, the adjustment step size is 0.03, the verification sample size is 50 sets, the correction rounds are 10 rounds, and the trigger threshold is the upper limit of the corresponding deviation. For Level 3 deviation, the adjustment step size is 0.05, the verification sample size is 80 sets, the correction rounds are 15 rounds, and the trigger threshold is the upper limit of the corresponding deviation. For critical deviation, the adjustment step size is 0.08, the verification sample size is 100 sets, the correction rounds are 20 rounds, and the trigger threshold is the upper limit of the corresponding deviation. The rollback threshold corresponds to the lower limit of the deviation, with a verification sample size of 50 sets and 0 correction rounds, triggering the restoration of parameters that cause correction failure. During calibration, the corresponding deviation level is first matched based on the predicted deviation value. For a level one deviation, only the correction coefficient of the mechanistic balance is adjusted, and parameter iterations are completed according to the corresponding step size and number of rounds. After adjustment, the effect is verified using a corresponding number of validation samples. For higher deviation levels, the regression parameters are gradually adjusted, increasing the adjustment step size and calibration rounds. If the deviation exceeds the backoff threshold after calibration, parameter restoration is triggered, restoring the effective parameters of the previous version. The tiered calibration mode matches the adjustment intensity according to the degree of deviation. Small adjustments ensure computational stability for small deviations, while larger adjustments quickly correct the deviation for large deviations. Combined with validation and backoff mechanisms, this avoids accuracy loss due to calibration errors.

[0063] Table 3 Deviation Correction Grading Parameter Table Step 4.3 Incremental Fitting Update Processing: An update cycle is set, and operational data is accumulated according to the update cycle. New operational data and corresponding test data are added to the soft measurement training dataset, and incremental fitting updates are performed on the Long Short-Term Memory Network and Support Vector Regression. The incremental fitting update uses a sliding window method to update the soft measurement training dataset. Historical data exceeding the set time range is removed according to the set time step, and new operational data and manual test data of the corresponding time period are added simultaneously to maintain the sample size of the soft measurement training dataset within the set range.

[0064] The update process retains the underlying network parameters of the Long Short-Term Memory (LSTM) network and the kernel function parameters of Support Vector Regression (SVR), adjusting and updating only the weight parameters of the top output layer of the LSM network and the bias parameters of SVR, thus completing the incremental fitting update process. Incremental fitting updates do not require retraining all parameters, allowing for adaptation to new operating data while preserving existing computational capabilities, reducing computational resource consumption and improving the efficiency of parameter updates.

[0065] In some embodiments, incremental fitting updates can be performed using a full training method, which merges the newly added running data with the corresponding test data and the original historical data to form a complete training dataset, and retrains all parameters of the long short-term memory network and support vector regression. This method is suitable for scenarios where the working conditions change significantly and the adaptability of the original parameters decreases significantly.

[0066] In some specific implementations, the incremental fitting update process is set with update trigger conditions and update restriction conditions. Update trigger conditions include two types: periodic trigger and deviation trigger. Periodic trigger automatically starts the update process according to the set update cycle, while deviation trigger starts the update process when the prediction deviation of multiple consecutive sets of test data exceeds the set threshold.

[0067] Update constraints include limits on the number of samples per update and the number of updates per unit of time. The number of new samples used in a single update must not exceed the set range to avoid excessive impact of single data fluctuations on overall parameters. The number of updates per unit of time must not exceed the set upper limit to avoid frequent updates causing oscillations in calculated parameters. During the update process, the newly added running data and laboratory data undergo the same cleaning process as in the dataset construction phase, removing outliers and filling in missing data, before being added to the training set. After parameter adjustment, the updated calculation effect is verified using validation data. If the deviation is confirmed to be within the set range, the update is completed. If the deviation exceeds the set range, the system reverts to the previous version of parameters and re-executes the data verification and parameter adjustment process.

[0068] In some specific implementations, an automatic operating condition identification and branch calculation process is added to address the issue of insufficient adaptability of a single calculation rule under different production operating conditions. Operating condition identification is based on the characteristic distribution of process parameters. Feature templates for four typical operating conditions—start-up, stable, load adjustment, and shutdown—are pre-defined based on historical operating data. Each operating condition template corresponds to a set of feature value ranges and distribution characteristics. During online calculation, the statistical characteristics of the process parameters for the current time period are first extracted and matched with the four operating condition templates for similarity. The operating condition type with the highest similarity is selected as the current operating condition determination result.

[0069] For each type of operating condition, corresponding weight coefficients and calculation parameters are set. The weight ratio of the mechanism calculation is increased under startup and shutdown conditions, relying on the physical constraints of the reaction mechanism to ensure the rationality of the calculation results. Under stable operating conditions, the weight of mechanism calculation and data mapping is balanced, taking into account both calculation accuracy and stability. Under load adjustment conditions, the weight ratio of the data mapping is increased to adapt to the nonlinear relationships in the dynamic changes of operating conditions. Operating condition identification and branch calculation enable the calculation rules to automatically match the current production state, improving the adaptability of component calculation results under different operating stages and reducing calculation deviations during operating condition switching.

[0070] The soft measurement method for material components in a dry acetylene generator provided in this embodiment achieves continuous calculation of material components through a combination of mechanistic calculation and data fitting. It obtains continuous component data without the need for additional online analysis equipment and is adaptable to high-temperature, high-dust, closed production conditions. This method, through multi-scale feature extraction and fusion calculation, can adapt to operating conditions with varying degrees of fluctuation and maintain stable calculation output even when conditions change. Through periodic correction and incremental update mechanisms, it can compensate for deviations caused by operating condition drift and equipment aging, continuously maintaining the reasonableness of the calculation results, reducing the frequency of manual testing, and decreasing the workload of manual operation. This method uses online detectable process parameters as input, obtains component results through data processing and calculation rules, and can be directly integrated into existing process control systems, providing continuous data support for the adjustment and control of the production process and improving the stability of production operation. The overall solution is data processing-centric, requires no modification to existing production equipment, has low implementation costs, and good applicability, making it applicable to component detection scenarios in similar closed reaction equipment.

[0071] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A soft measurement method for material components in a dry acetylene generator, characterized in that, Includes the following steps: S1. Based on the mechanism of calcium carbide hydrolysis reaction, auxiliary variables were screened, including calcium carbide feed rate, water injection flow rate, reaction temperature, stirring current, generator pressure, and acetylene gas production flow rate; historical test data and corresponding process parameters were matched, and after data cleaning and time-series alignment, a soft measurement training dataset was constructed. S2. The moving time window algorithm is used to process the time series data of the soft measurement training dataset to extract parameter trend features, fluctuation features and rate of change features; the ensemble empirical mode decomposition method is used to separate the steady-state components and dynamic components of the data to obtain a multi-scale feature set; S3. Calculate the basic values ​​of the components based on the material balance and energy balance of the calcium carbide hydrolysis reaction; use a long short-term memory network and support vector regression to fit the mapping relationship between the auxiliary variables and the target components, and calculate the fitted values ​​of the components; perform a weighted summation of the basic values ​​of the components and the fitted values ​​of the components to obtain the predicted results of the material components. S4. Set the calibration cycle and deviation threshold, and correct the calculation parameters after obtaining new manual test data; set the update cycle, accumulate running data and update the fitting calculation parameters to maintain the deviation of the calculation results within the set range.

2. The method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.

1. Based on the analysis of the influencing factors of material composition according to the mechanism of calcium carbide hydrolysis reaction, and combined with the process operation records, screen out online detectable parameters that are related to the material composition to form an initial set of auxiliary variables; S1.

2. Collect historical test data and corresponding process operation parameters within a set time range, and perform time sequence alignment according to the timestamp; S1.

3. Perform outlier removal and missing value imputation on the aligned dataset; Outlier removal uses the Raida criterion to identify and remove outlier data that exceeds the set range. Missing value imputation uses linear interpolation to fill in missing data, resulting in a soft measurement training dataset.

3. The method according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.

1. Set the window width and sliding step size of the moving time window, and segment the time series data of the soft measurement training dataset according to the set parameters; S2.

2. Perform statistical calculations on the parameter data within each time window, extract the trend characteristics, fluctuation characteristics, and rate of change characteristics of the corresponding parameters, and form a single-scale feature set; S2.

3. The ensemble empirical mode decomposition method is used to decompose the single-scale feature set, separate the steady-state components and dynamic components in the data, and combine them to form a multi-scale feature set.

4. The method according to claim 1, characterized in that, Step S3 includes the following sub-steps: S3.

1. Based on the material balance and energy balance of the calcium carbide hydrolysis reaction, establish material balance calculation formulas and energy balance calculation formulas. The material balance calculation formula outputs the mass change data of the reactants and inputs it into the energy balance calculation formula. The energy balance calculation formula outputs the enthalpy change data of the reaction and inputs it into the material balance calculation formula. The basic values ​​of the components are calculated. S3.

2. Input the multi-scale feature set into the long short-term memory network to extract the time-dependent features, input the time-dependent features into support vector regression for nonlinear mapping fitting, and calculate the component fitting values; S3.

3. Set the weighting coefficients for the basic component values ​​and the fitted component values, and perform a weighted summation of the basic component values ​​and the fitted component values ​​according to the weighting coefficients to obtain the predicted material composition results.

5. The method according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.

1. Set the calibration cycle and deviation threshold, obtain new manual test data according to the calibration cycle, compare the manual test data with the material composition prediction results of the corresponding time period, and calculate the prediction deviation value. S4.

2. When the prediction deviation exceeds the deviation threshold, the material balance calculation parameters, energy balance calculation parameters, and support vector regression parameters are corrected by a correction algorithm. S4.

3. Set the update cycle, accumulate running data according to the update cycle, add the newly added running data and corresponding test data to the soft measurement training dataset, and perform incremental fitting and updating of the long short-term memory network and support vector regression.

6. The method according to claim 2, characterized in that, In step S1.1, the data of each parameter in the initial auxiliary variable set is standardized to eliminate the influence of dimensional differences on the calculation results; the correlation degree between each initial auxiliary variable and the material component is calculated using the Pearson correlation analysis method, and the multicollinearity coefficient between the initial auxiliary variables is also calculated. Based on the correlation degree and multicollinearity coefficient, redundant variables are removed from the initial auxiliary variable set, and parameters whose correlation degree and multicollinearity coefficient meet the set conditions are retained to form the final auxiliary variable set.

7. The method according to claim 3, characterized in that, In step S2.1, the moving time window adopts a variable window width setting method; at least two sets of window width values ​​and corresponding fluctuation degree judgment conditions are preset, the standard deviation of the process parameters in the current time period is calculated in real time as the fluctuation degree characterization value, the calculated fluctuation degree characterization value is matched with each fluctuation degree judgment condition, the window width value corresponding to the successfully matched judgment condition is selected, and the time series data of the soft measurement training dataset is segmented and truncated.

8. The method according to claim 3, characterized in that, In step S2.3, the processing procedure of the set empirical mode decomposition method is as follows: each dimension feature in the single-scale feature set is decomposed sequentially. White noise with a set amplitude is added during the decomposition process. After repeating the decomposition operation at least twice, the average value of the calculation results is taken to obtain the steady-state component and dynamic component of the corresponding dimension. The steady-state component and dynamic component of all dimensions are combined to form a multi-scale feature set.

9. The method according to claim 4, characterized in that, In step S3.2, the Long Short-Term Memory (LSTM) network includes an input layer, a set number of hidden layers, and an output layer. The hidden layers are configured with a forget gate, an input gate, and an output gate. Support Vector Regression (SVR) uses a radial basis function kernel to construct a mapping space. The output of the LSM network is connected to the input of SVR. The multi-scale feature set is divided into a training subset and a validation subset. The training subset is used to train the LSM network for temporal feature extraction. The temporal dependency features output by the LSM network are input into SVR for mapping training. The validation subset is used to verify the accuracy of the combined structure of the LSM network and SVR. The parameters of the LSM network and SVR are adjusted until the validation results meet the set conditions.

10. The method according to claim 5, characterized in that, In step S4.3, the incremental fitting update uses a sliding window method to update the soft measurement training dataset; historical data exceeding the set time range are removed according to the set time step, and newly added running data and manual test data of the corresponding time period are added simultaneously to maintain the sample size of the soft measurement training dataset within the set range; during the update process, the underlying network parameters of the Long Short-Term Memory Network and the kernel function parameters of the Support Vector Regression are retained, and only the weight parameters of the top output layer of the Long Short-Term Memory Network and the bias parameters of the Support Vector Regression are adjusted and updated to complete the incremental fitting update process.

Images