Continuous stirring reaction kettle operation state monitoring method based on time sequence data analysis

Abstract

The invention discloses a continuous stirring reaction kettle operation state monitoring method based on time sequence data analysis, and relates to the technical field of process state monitoring, and the method comprises the following steps: collecting multi-source operation time sequence data comprising stirring motor driving signals, material components, temperature and pressure, and product quality; performing cross-domain correlation analysis on the driving signal and the material components, and identifying a coupling correlation mode; dynamic process segmentation is performed on the temperature and pressure data based on the mode, and feature fragments of all reaction stages are extracted; constructing a causal inference network model between the feature fragments and the product quality; reversely tracing key process intervals and features influencing quality through the model, and generating a process traceability report; and judging the comprehensive operation state of the reaction kettle. According to the method, the accuracy and interpretability of state monitoring are improved by revealing process internal association and realizing causal traceability.

Description

A Method for Monitoring the Operating Status of Continuous Stirred Reactors Based on Time-Series Data Analysis Technical Field

[0001] This invention belongs to the field of process status monitoring technology, specifically a method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis. Background Technology

[0002] In chemical production processes, monitoring the operational status of continuous stirred tank reactors is crucial for ensuring product quality and safety. Current technologies primarily rely on setting independent thresholds for parameters such as temperature, pressure, and material composition to trigger alarms, or employing multivariate statistical methods to monitor overall parameter deviations. These methods treat data on stirring drive, materials, and reaction conditions as independent variables or perform joint statistical monitoring, lacking in-depth analysis of the inherent physicochemical relationships within the process.

[0003] Existing technical solutions have shortcomings. They fail to analyze the dynamic coupling relationship between mechanical stirring and chemical reaction, making it impossible to effectively distinguish whether the root cause of product quality fluctuations stems from mechanical abnormalities, material changes, or reaction control failures. Traditional methods cannot construct a causal chain between process status and final quality from multi-source time-series data, resulting in delayed problem diagnosis, ambiguous source tracing, and reliance on experience and statistical bias for status determination, lacking precise process mechanism support.

[0004] A state monitoring method is needed that can reveal the inherent dynamic correlations within a process and enable precise traceability from quality anomalies to specific process stages. This method needs to address the lack of cross-domain signal correlation analysis and establish an interpretable causal model between reaction process characteristics and product quality, thereby achieving accurate and in-depth determination of the reactor's operating status. Summary of the Invention

[0005] This invention aims to at least solve one of the technical problems existing in the prior art; to this end, this invention proposes a method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis, comprising: collecting a multi-source operating time-series data set of the continuous stirred reactor within a set monitoring period, wherein the multi-source operating time-series data set includes stirring motor drive signals, reactant component sensor readings, reactor internal temperature and pressure monitoring data, and online analysis results of reaction product quality; performing cross-domain signal correlation analysis on the stirring motor drive signals and reactant component sensor readings to identify coupling correlation patterns between the drive signals and material changes, and generating a set of coupling correlation patterns; based on the coupling... The associated pattern set performs dynamic process segmentation on the temperature and pressure monitoring data inside the reactor, dividing it into multiple process intervals reflecting different chemical reaction stages, and extracting characteristic data fragments within each process interval. Utilizing the mapping relationship between these characteristic data fragments and the online analysis results of the reaction product quality, a causal inference network model between the reaction state and product quality is constructed. Through this causal inference network model, the key process intervals and their characteristic data leading to the current online analysis results of product quality are traced back, generating a reaction process traceability report. Based on the reaction process traceability report, the comprehensive operating status of the continuous stirred reactor during the current monitoring period is determined, and an operating status monitoring conclusion is generated.

[0006] Furthermore, cross-domain signal correlation analysis is performed on the stirring motor drive signal and the reaction material composition sensor readings to identify the coupling correlation patterns between the drive signal and material changes, generating a set of coupling correlation patterns. This includes: performing a spectral transformation on the stirring motor drive signal to convert it from a time-domain signal to a frequency-domain energy distribution signal; performing principal component dimensionality reduction on the reaction material composition sensor readings to extract the principal component change trajectory reflecting the trend of material composition changes; calculating the dynamic correlation coefficient sequence between the frequency-domain energy distribution signal and the principal component change trajectory over multiple preset time windows; performing pattern clustering analysis on the dynamic correlation coefficient sequence to identify different patterns where the correlation coefficient is consistently high, consistently low, or periodically fluctuating; and combining the identified different patterns with the original stirring motor drive signal features and the original reaction material composition sensor reading features within the corresponding time windows to form multiple coupling correlation patterns, constituting the set of coupling correlation patterns.

[0007] Furthermore, based on the set of coupling correlation patterns, dynamic process segmentation is performed on the in-vessel temperature and pressure monitoring data to divide it into multiple process intervals reflecting different chemical reaction stages, and feature data fragments are extracted from each process interval. This includes: identifying key coupling correlation patterns that indicate the switching of chemical reaction stages from the set of coupling correlation patterns; using the time point of occurrence of the key coupling correlation pattern as a candidate segmentation point to perform preliminary segmentation of the continuous time series of in-vessel temperature and pressure monitoring data; performing trend fitting on the in-vessel temperature and pressure monitoring data within each preliminary segment to extract feature combinations of temperature change rate and pressure change rate; comparing the feature combinations of two adjacent preliminary segments, and if the difference exceeds a preset threshold, confirming the candidate segmentation point as a valid process boundary, otherwise merging adjacent segments; and finally dividing the in-vessel temperature and pressure monitoring data of the entire monitoring period into multiple continuous process intervals based on all valid process boundaries, and extracting the data corresponding to each process interval from the original data as the feature data fragment.

[0008] Furthermore, utilizing the mapping relationship between the feature data segments and the online analysis results of the product quality, a causal inference network model between the reaction state and the product quality is constructed. This includes: for each process interval, pairing its corresponding feature data segment with the online analysis results of the product quality at the end of the reaction or at the end of the interval to form process-quality pairing samples; analyzing multiple process-quality pairing samples to calculate the partial correlation between the statistical characteristic values ​​of different feature data segments and various indicators in the online analysis results of the product quality; based on the partial correlation calculation results, selecting feature data segments whose absolute value of the partial correlation coefficient with at least one quality indicator in the online analysis results of the product quality is greater than a preset correlation threshold, and using the selected feature data segment features and various quality indicators in the online analysis results of the product quality as feature nodes and quality indicator nodes of the causal inference network model, respectively; and using the temporal Granger causality test method to analyze the predictive ability of the temporal changes of the feature data segment features on the subsequent changes of the product quality indicators, establishing directed causal edges from feature nodes to quality indicator nodes, thereby constructing the causal inference network model.

[0009] Furthermore, through the causal inference network model, the key process intervals and their characteristic data leading to the current product quality online analysis results are traced backward to generate a reaction process traceability report, including: obtaining the final reaction product quality online analysis results for the current monitoring period; locating the quality indicator node corresponding to the final reaction product quality online analysis results in the causal inference network model; starting from the quality indicator node, traversing backward along the directed causal edges pointing to the quality indicator node to find all characteristic data fragment feature nodes that directly or indirectly point to the quality indicator node; locating the key process intervals that contribute to the final quality results based on the relationship between the characteristic data fragment feature nodes and the process intervals; aggregating the characteristic data fragment features extracted from the key process intervals and combining them with their causal contribution in the causal inference network model to generate the reaction process traceability report describing how each key process interval affects the final product quality.

[0010] Furthermore, based on the relationship between feature nodes of feature data segments and process intervals, key process intervals that contribute to the final quality result are located. This includes: obtaining a list of all feature nodes of feature data segments pointing to the final quality indicator node in the causal inference network model; querying the mapping relationship table between feature data segment features and original process intervals to determine the process interval from which each feature node of feature data segment in the list originates; counting the frequency of each process interval being referenced by nodes in the feature node list of feature data segments, and summarizing the total causal contribution of all feature nodes of feature data segments from the same process interval to the final quality indicator; and marking process intervals whose total causal contribution exceeds a preset threshold as key process intervals.

[0011] Furthermore, the feature data fragments extracted from the key process intervals are aggregated and combined with their causal contribution in the causal inference network model to generate the reaction process traceability report describing how each key process interval affects the quality of the final product. This includes: creating a report entry for each marked key process interval; recording the start and end times of the key process interval and the core feature values ​​of the in-vessel temperature and pressure monitoring data within the key process interval in the report entry; extracting the weights of the directed causal edges pointing from the feature nodes of each feature data fragment from the key process interval to the final quality indicator nodes from the causal inference network model, as the causal contribution; listing each feature data fragment feature and its corresponding causal contribution in detail in the report entry, and sorting them from high to low according to the causal contribution; organizing all the report entries for the key process intervals in chronological order and adding an overall summary to form the reaction process traceability report.

[0012] Further, based on the reaction process traceability report, the overall operating status of the continuous stirred reactor during the current monitoring period is determined, including: parsing the reaction process traceability report and extracting the deviation between the features of characteristic data segments in each key process interval and their standard reference values; calculating the weighted average of the deviations of all key process intervals, where the weights are determined by the causal contribution of each key process interval; comparing the weighted average with a preset operating status threshold range; if the weighted average is within the preset normal operating status threshold range, the overall operating status of the continuous stirred reactor is determined to be normal; if the weighted average exceeds the normal operating status threshold range but does not reach the warning status threshold, the overall operating status is determined to be requiring attention; if the weighted average reaches or exceeds the warning status threshold, the overall operating status is determined to be abnormal.

[0013] Further, the calculation of the weighted average of the deviations of all critical process intervals includes: for each critical process interval in the reaction process traceability report, calculating the average of the absolute deviations of all characteristic data fragments listed in its report entry from their respective standard reference values, as the characteristic deviation of the critical process interval; obtaining the total causal contribution of the critical process interval to the final quality result from the reaction process traceability report; using the total causal contribution of each critical process interval as its weight, performing a weighted summation of the characteristic deviations of all critical process intervals; and dividing the result of the weighted summation by the sum of the weights of all critical process intervals to obtain the weighted average.

[0014] Furthermore, the method also includes a step of continuously updating the causal inference network model, specifically including: using the multi-source operating time series data set collected in the new monitoring period and its corresponding reaction process traceability report and operating status monitoring conclusion as a new training sample; using the new training sample, recalculating the partial correlation between the features of the feature data fragments and the online analysis results of the quality of the reaction products, and re-performing the time-series Granger causality test; based on the recalculation and test results, incrementally updating or correcting the nodes, directed edges, and weights in the existing causal inference network model; and using the updated causal inference network model for operating status monitoring in subsequent monitoring periods.

[0015] Compared with existing technologies, the advantages of this invention are: by analyzing cross-domain signal correlation, it identifies the coupling correlation pattern between the stirring motor drive signal and the sensor readings of the reactant composition. This technique establishes a dynamic quantitative relationship between mechanical operating parameters and changes in chemical reactants. In traditional monitoring methods, these two types of data are treated in isolation, while correlation analysis can capture how changes in drive load instantly affect the reaction process, or how changes in material composition are fed back to the motor operating status through factors such as viscosity and density. Based on this coupling pattern, dynamic process segmentation of temperature and pressure data is performed, so that the segmented intervals no longer depend on fixed time windows or simple statistical clustering, but correspond to different reaction stages defined by changes in intrinsic mechanisms. This provides a basic unit with clear chemical and physical meaning for state feature extraction and evaluation, enabling subsequent analysis to accurately focus on the state evolution of specific reaction stages.

[0016] A causal inference network model was constructed using feature data fragments and online quality analysis results. This model quantifies the influence path and intensity of process characteristics at different reaction stages on various quality indicators of the final product. Through the model's backtracking function, it is possible to locate one or more specific process intervals that have a decisive impact on the quality data obtained in real time, and to pinpoint the specific key feature data within those intervals. This allows the diagnosis of quality anomalies or fluctuations to move beyond the traditional phenomenon reporting level and delve into the level of localizable and attributable process mechanisms. The conclusions of condition monitoring can thus clearly identify the underlying process causes leading to the current operating state, realizing a shift from "monitoring anomalies" to "explaining the root causes of anomalies," enhancing the accuracy of condition determination and the effectiveness of decision support. Attached Figure Description

[0017] Figure 1 is a flowchart of the continuous stirred reactor operation status monitoring method based on time series data analysis according to the present invention; Figure 2 is a flowchart of cross-domain signal correlation analysis to generate a set of coupled correlation patterns; Figure 3 is a flowchart of constructing a causal inference network model; Figure 4 is a graph of temperature and pressure change trends at each stage of the continuous stirred reactor; Figure 5 is a heat map of the causal contribution of key process characteristics of the continuous stirred reactor to product quality indicators. Detailed Implementation

[0018] 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.

[0019] Referring to Figure 1, a multi-source time-series data set was collected from the continuous stirred reactor during a set monitoring period. This multi-source time-series data set specifically includes the stirring motor drive signal, reactant component sensor readings, reactor internal temperature monitoring data, reactor internal pressure monitoring data, and online analysis results of the reaction product quality. Cross-domain signal correlation analysis was performed on the collected stirring motor drive signal and reactant component sensor readings to identify the coupling correlation patterns between the drive signal and material changes, generating a set of coupling correlation patterns containing multiple modes. Based on the generated set of coupling correlation patterns, dynamic process segmentation was performed on the reactor internal temperature and pressure monitoring data, dividing the data of the entire monitoring period into multiple process intervals that reflect different chemical reaction stages, and extracting characteristic data fragments corresponding to each process interval from the raw data. The mapping relationship between these characteristic data fragments and the final online analysis results of the reaction product quality was analyzed, constructing a causal inference network model describing the causal relationship between the reaction process state and the final product quality. Through this causal inference network model, starting from the final online analysis results of the product quality, the key process intervals and their characteristic data were traced back to determine which key process intervals and their characteristic data led to the quality result, generating a reaction process traceability report. Based on the generated reaction process traceability report, the overall operating status of the continuous stirred reactor during the current monitoring period is determined, and an operating status monitoring conclusion is generated.

[0020] Referring to Figure 2, in one embodiment of the present invention, cross-domain signal correlation analysis is performed on the stirring motor drive signal and the readings of the reactant component sensor to identify the coupling correlation pattern between the drive signal and the material changes, and a set of coupling correlation patterns is generated. The specific implementation is as follows: Taking a continuous stirring reactor operation monitoring period with a monitoring duration of four hours as an example, the collected stirring motor drive signal is a time-domain sequence containing drive current and rotation speed, and the collected reactant component sensor readings include concentration reading sequences of multiple raw material components. The original time-domain sequence of the stirring motor drive signal is processed by Fast Fourier Transform (FFT) to convert the stirring motor drive signal into a series of frequency-domain energy distribution signals over a series of time windows. This frequency-domain energy distribution signal can be represented as a series of energy distributions at different frequency components. Principal component analysis is performed on the reactant component sensor readings, i.e., the concentration sequences of multiple components, to extract the top few principal components with a cumulative variance contribution rate exceeding 95%. These principal components are then connected together with their numerical changes over time to generate principal component change trajectories.

[0021] In practical implementation, the correlation between the frequency domain energy distribution signal and the principal component change trajectory is calculated. A preset time window of five minutes is used to calculate the Spearman rank correlation coefficient between the energy envelope of the frequency domain energy distribution signal and the principal component change trajectory within each time window, thus forming a dynamic correlation coefficient sequence. Pattern clustering analysis is performed on the dynamic correlation coefficient sequence. In practice, the density-based DBSCAN clustering algorithm is applied to identify different patterns. This identifies high-value patterns with correlation coefficients consistently above 0.8, low-value patterns with correlation coefficients consistently below 0.3, and periodic patterns with correlation coefficients fluctuating regularly between 0.3 and 0.8. It can be understood that during the stirring reaction process, high-value patterns may correspond to a stable reaction stage where materials are uniformly mixed and the motor drive and component changes are highly synchronized; low-value patterns may correspond to the initial feeding or material stratification state where the drive and component changes are disconnected; and periodic patterns may correspond to the intermittent feeding or cyclic reaction stage.

[0022] In some embodiments, the formation of coupled correlation patterns requires combining original signal features. For example, for an identified high-value pattern, the average drive power feature within the corresponding time window is extracted from the original stirring motor drive signal, and the variation amplitude feature of the key component concentration within the corresponding time window is extracted from the original reactant component sensor readings. This high-value pattern, average drive power feature, and variation amplitude feature are combined to form a coupled correlation pattern. For an identified low-value pattern, the speed fluctuation variance feature within the corresponding time window is extracted from the original stirring motor drive signal, and the covariance matrix feature of the concentrations of multiple components within the corresponding time window is extracted from the original reactant component sensor readings. This low-value pattern, speed fluctuation variance feature, and covariance matrix feature are combined to form another coupled correlation pattern. In specific implementations, similar feature extraction and combination operations are performed for each identified pattern, ultimately gathering all pattern instances to form a complete set of coupled correlation patterns for subsequent dynamic process segmentation.

[0023] In practical implementation, the dynamic correlation coefficient can be calculated using a specific quantization formula. In some embodiments, let the energy vector of the frequency domain energy distribution signal within the i-th time window be... Where m is the number of frequency components, and the numerical vector of the principal component variation trajectory within the same time window is... k is the number of principal components. The ranking vector of the energy vector is denoted as . The ranking vector of the principal component numerical vectors is denoted as... . In the formula, express and The difference in ranking rank along the corresponding dimension j. The dynamic correlation coefficient for this time window. It can be calculated using a formula. The dynamic correlation coefficient calculated by this formula can measure the degree of monotonic correlation between the frequency domain energy distribution signal and the trajectory of principal component changes.

[0024] Referring to Figure 3, in one embodiment of the present invention, dynamic process segmentation is performed on the in-vessel temperature and pressure monitoring data based on a set of coupled correlation patterns. This divides the data into multiple process intervals reflecting different chemical reaction stages, and extracts feature data fragments from each process interval. This process is combined with the process of constructing a causal inference network model using the feature data fragments. From the generated set of coupled correlation patterns, key coupled correlation patterns that indicate possible switching of chemical reaction stages are identified. For example, in a specific reaction monitoring scenario, the coupled correlation pattern formed by the driving power characteristics of the stirring motor drive signal and the variation amplitude characteristics of the material composition sensor readings has the largest mode transition amplitude in the coupled correlation pattern set. When the amplitude of this coupled correlation pattern drops sharply from a high-value mode to a low-value mode, exceeding a preset threshold, this coupled correlation pattern is identified as a key coupled correlation pattern. The time point when the key coupled correlation pattern occurs—that is, the moment when the power characteristic and variation amplitude characteristics drop sharply—is used as a candidate segmentation point to initially segment the continuous time series formed by the in-vessel temperature monitoring data and the in-vessel pressure monitoring data. The initial segmentation divides the 4-hour monitoring period into two segments at the candidate segmentation point, forming segment 1 and segment 2.

[0025] In the specific implementation, trend fitting is performed on the vessel internal temperature monitoring data and vessel internal pressure monitoring data within each preliminary segment to extract the characteristic combination of temperature change rate and pressure change rate. For the first preliminary segment, linear regression is performed on the vessel internal temperature monitoring data, calculating that the average temperature change rate for the first segment is an increase of 10 degrees Celsius per hour. Linear regression is also performed on the vessel internal pressure monitoring data for the first preliminary segment, calculating that the average pressure change rate for the first segment is an increase of 0.5 MPa per hour. The characteristic combination for the first segment is the combination of an average temperature change rate of 10 degrees Celsius per hour and an average pressure change rate of 0.5 MPa per hour. For the second preliminary segment, linear regression is performed on the vessel internal temperature monitoring data, calculating that the average temperature change rate for the second segment is a decrease of 2 degrees Celsius per hour. Linear regression is also performed on the vessel internal pressure monitoring data for the second preliminary segment, calculating that the average pressure change rate for the second segment is a decrease of 0.1 MPa per hour. The characteristic combination for the second segment is the combination of an average temperature change rate of 2 degrees Celsius per hour and an average pressure change rate of 0.1 MPa per hour. It is understandable that the rate of change of temperature and the rate of change of pressure are characteristics of the kinetics of chemical reactions.

[0026] In the specific implementation, the feature combinations of two adjacent preliminary segments are compared. The feature combination of the first preliminary segment is an average temperature change rate increasing by 10 degrees Celsius per hour and an average pressure change rate increasing by 0.5 MPa per hour. The feature combination of the second preliminary segment is an average temperature change rate decreasing by 2 degrees Celsius per hour and an average pressure change rate decreasing by 0.1 MPa per hour. The difference between the two feature combinations is calculated as the sum of the absolute differences in the temperature change rate and the absolute differences in the pressure change rate, which is 12.1 degrees Celsius and MPa per hour. This difference exceeds the preset threshold of 5 degrees Celsius and MPa per hour, and the candidate segmentation point is confirmed as the effective process boundary. Based on the effective process boundary, the in-vessel temperature monitoring data and in-vessel pressure monitoring data for the entire 4-hour monitoring period are finally divided into two continuous process intervals. Data corresponding to the first and second process intervals are extracted from the original monitoring data and used as feature data segments for the first and second process intervals, respectively.

[0027] In practical implementation, a causal inference network model between reaction state and product quality is constructed by utilizing the mapping relationship between characteristic data segments and online analysis results of reaction product quality. For the first process interval, the characteristic data segments of the first process interval are paired with the online analysis results of reaction product quality at the end of the entire reaction, forming the first process-quality paired sample. For the second process interval, the characteristic data segments of the second process interval are paired with the online analysis results of reaction product quality at the end of the entire reaction, forming the second process-quality paired sample. The first and second process-quality paired samples are analyzed. The eigenvalues ​​in the characteristic data segments of the first process interval, such as the variance of the in-vessel temperature monitoring data within the first process interval, are calculated, along with the partial correlation between these eigenvalues ​​and the yield index in the online analysis results of the reaction product quality. Similarly, the eigenvalues ​​in the characteristic data segments of the second process interval, such as the mean of the in-vessel pressure monitoring data within the second process interval, are calculated, along with the partial correlation between these eigenvalues ​​and the purity index in the online analysis results of the reaction product quality. Based on the partial correlation calculation results, feature data segments with absolute values ​​of partial correlation coefficients greater than the preset correlation threshold of 0.3 in the yield index of the online analysis results of the reaction product quality were selected. These selected feature data segments, namely the variance of the in-vessel temperature monitoring data within the first process interval, and the yield index in the online analysis results of the reaction product quality, were used as the feature node and quality index node of the causal inference network model, respectively. It can be understood that the partial correlation coefficient reflects the strength of the association between a single feature and the quality index after excluding the influence of other features.

[0028] In some embodiments, the temporal Granger causality test is used to analyze the predictive ability of temporal changes in characteristic data segments on changes in subsequent product quality indicators. In the causal inference network model, the temporal data of the variance of the in-vessel temperature monitoring data within the first process interval of the characteristic node are compared with the temporal data of the yield indicator at the quality indicator node, using a Granger causality test. In the test formula, ... This represents the change in the yield index over time t. This represents the change in variance of the in-vessel temperature monitoring data within the first process interval at the previous time point. The formula determines whether the change in the characteristic node is a Granger cause of the change in the quality indicator node by comparing the sum of squared residuals of the model with and without lag terms. If the F-statistic of the Granger causality test is greater than the critical value, a directed causal edge is established from the variance of the in-vessel temperature monitoring data within the first process interval of the characteristic node to the yield indicator of the quality indicator node, thus constructing the basic structure of the causal inference network model.

[0029] In one embodiment of the present invention, a causal inference network model is used to trace back the key process intervals and their characteristic data that lead to the current online product quality analysis results, generating a reaction process traceability report. Specifically, this process involves: obtaining the online quality analysis results of the final reaction product during the current monitoring period. In an example scenario, the online quality analysis results show that the yield of the final product is 95.3% and the purity is 99.1%. In the constructed causal inference network model, the quality indicator nodes corresponding to the online quality analysis results of the final reaction product are located, specifically including yield indicator nodes and purity indicator nodes. Starting from the yield indicator node, a reverse traversal is performed along all directed causal edges pointing to the yield indicator node to find all characteristic data segment feature nodes that directly or indirectly point to the yield indicator node. In the example causal inference network model, the characteristic data segment feature nodes "variance of the in-vessel temperature monitoring data in the first process interval" and "mean value of the in-vessel pressure monitoring data in the second process interval" are found to point to the yield indicator node through directed causal edges. Based on the relationship between the feature nodes of the feature data segment and the process interval, the key process intervals that contribute to the final quality result are located. The mapping relationship table between the feature data segment features and the original process interval is queried to determine that the feature node "variance of the temperature monitoring data in the first process interval" of the feature data segment originates from the first process interval, and the feature node "mean value of the pressure monitoring data in the second process interval" of the feature data segment originates from the second process interval.

[0030] In the specific implementation, a list of feature nodes for all feature data segments pointing to the final quality indicator node is obtained from the causal inference network model. For the yield indicator node, the feature node list for the feature data segments contains two nodes: "variance of in-vessel temperature monitoring data within the first process interval" and "mean of in-vessel pressure monitoring data within the second process interval". The mapping relationship table between the feature data segment features and the original process intervals is queried to determine the process interval from which each feature node in the list originates. The node "variance of in-vessel temperature monitoring data within the first process interval" is determined to originate from the first process interval, and the node "mean of in-vessel pressure monitoring data within the second process interval" is determined to originate from the second process interval. The frequency of each process interval being referenced by nodes in the feature node list for the feature data segments is counted; the first process interval is referenced once, and the second process interval is referenced once. The total causal contribution of all feature data segments and feature nodes from the same process interval to the final quality index is summarized. From the causal inference network model, the weight of the directed causal edge pointing to the yield index node from the "variance of the in-vessel temperature monitoring data in the first process interval" is set to 0.65, and the weight of the directed causal edge pointing to the yield index node from the "mean of the in-vessel pressure monitoring data in the second process interval" is set to 0.35. Therefore, the total causal contribution of the first process interval to the yield index is 0.65, and the total causal contribution of the second process interval to the yield index is 0.35. Process intervals with a total causal contribution exceeding a preset threshold are marked as critical process intervals. If the preset threshold is 0.3, then both the first and second process intervals are marked as critical process intervals.

[0031] In the specific implementation, a report entry is created for each marked key process interval. In the report entry for the first process interval, the start and end times of the first process interval are recorded as 0 to 2 hours after the start of monitoring. The core characteristic of the vessel temperature monitoring data within the first process interval is recorded as an average temperature increase of 10 degrees Celsius per hour, and the core characteristic of the vessel pressure monitoring data is recorded as an average pressure increase of 0.5 MPa per hour. From the causal inference network model, the weight of the directed causal edge pointing to the yield index node from the feature node "variance of vessel temperature monitoring data within the first process interval" of the feature data segment from the first process interval is extracted, and 0.65 is used as its causal contribution. The report entry details the feature "variance of vessel temperature monitoring data within the first process interval" and its corresponding causal contribution of 0.65. In the report entry for the second process interval, the start and end times of the second process interval are recorded as 2 to 4 hours after the start of monitoring. The core characteristic of the vessel temperature monitoring data within the second process interval is recorded as an average temperature decrease of 2 degrees Celsius per hour, and the core characteristic of the vessel pressure monitoring data is recorded as an average pressure decrease of 0.1 MPa per hour. From the causal inference network model, the weight of the directed causal edge pointing to the yield index node from the feature node "mean value of pressure monitoring data inside the vessel within the second process interval" of the feature data segment originating from the second process interval is extracted, and 0.35 is used as its causal contribution. The report entry lists in detail the feature data segment feature "mean value of pressure monitoring data inside the vessel within the second process interval" and its corresponding causal contribution of 0.35.

[0032] In practice, the characteristic data segments in the report entries for all critical process intervals are sorted from highest to lowest causal contribution. The causal contribution of the characteristic data segment "variance of the vessel temperature monitoring data within the first process interval" in the first process interval report entry is 0.65, and the causal contribution of the characteristic data segment "mean of the vessel pressure monitoring data within the second process interval" in the second process interval report entry is 0.35. After sorting, the entries for the first process interval are listed first, followed by those for the second process interval. All report entries for critical process intervals are organized chronologically, with the first process interval report entries appearing first, followed by those for the second process interval, and an overall summary is added to form a complete reaction process traceability report. The overall summary can include the total number of critical process intervals and information on the process intervals with the greatest impact on the final quality indicators. For example, the calculation of the total causal contribution can use a specific quantitative formula. ;in: This represents the overall impact of the critical process interval on the final quality indicator. It is the number of critical process intervals. It is the total causal contribution of the nth critical process interval. It is the comprehensive deviation score between the features of the nth key process interval feature data segment and the standard reference value. This formula is used to quantify the overall measure of the impact of the process interval on the final result.

[0033] In one embodiment of the present invention, the comprehensive operating status of the continuous stirred reactor during the current monitoring period is determined based on the reaction process traceability report. This process is specifically implemented as follows: The reaction process traceability report is analyzed, and the deviation between the characteristic data segments listed in each key process interval of the report and their corresponding standard reference values ​​is extracted. For example, two key process intervals are analyzed from a specific reaction process traceability report. The first key process interval report entry lists the characteristic data segment "variance of the reactor temperature monitoring data within the first process interval," with an actual value of 2.5 and a standard reference value of 2.0. The second key process interval report entry lists the characteristic data segment "mean value of the reactor pressure monitoring data within the second process interval," with an actual value of 1.8 MPa and a standard reference value of 1.9 MPa. The absolute deviation between these two characteristic data segments and their respective standard reference values ​​is calculated. The absolute deviation of the characteristic "variance" of the first key process interval is |2.5 - 2.0| = 0.5, and the absolute deviation of the characteristic "mean value" of the second key process interval is |1.8 - 1.9| = 0.1.

[0034] In specific implementation, a weighted average of the deviations of all critical process intervals is calculated, where the weights are determined by the causal contribution of each critical process interval. For each critical process interval in the reaction process traceability report, the average of the absolute deviations of all feature data segments listed in its report entry from their respective standard reference values ​​is calculated as the feature deviation of that critical process interval. For example, the first critical process interval contains only one feature data segment, "variance," with a feature deviation of 0.5; the second critical process interval contains only one feature data segment, "mean," with a feature deviation of 0.1. The total causal contribution of these two critical process intervals to the final quality result is obtained from the reaction process traceability report. As calculated in the example, the total causal contribution of the first critical process interval is 0.65, and the total causal contribution of the second critical process interval is 0.35. Using the total causal contribution of each critical process interval as its weight, the feature deviations of all critical process intervals are weighted and summed. The calculation process is (0.65... 0.5)+(0.35 0.1) = 0.325 + 0.035 = 0.36. Dividing the weighted sum by the sum of the weights of all critical process intervals, i.e., 0.36 / (0.65 + 0.35) = 0.36, we get a weighted average of 0.36.

[0035] In specific implementation, the calculated weighted average value is compared with a preset operating state threshold range. The preset normal operating state threshold range is [0, 0.2), the threshold range for states requiring attention is [0.2, 0.5), and the threshold range for abnormal states is [0.5, ∞). If the weighted average value of 0.36 falls within the threshold range of [0.2, 0.5), then the overall operating state of the continuous stirred reactor is determined to be requiring attention. In some embodiments, referring to Table 1, the determination logic can be summarized into the following operating state determination comparison table.

[0036] Table 1. Operation Status Judgment Comparison Table: It can be understood that when multiple feature data segments exist, the formula for calculating the feature deviation of the critical process interval is: ;in: This represents the characteristic deviation of the k-th critical process interval. This represents the number of features in the feature data segment within the k-th critical process interval. This represents the actual value of the feature of the i-th feature data segment within the k-th critical process interval. This represents the standard reference value corresponding to the i-th feature data segment within the k-th critical process interval. The formula calculates the average deviation of all features within the critical process interval.

[0037] Referring to Figure 4, this is a trend chart of temperature and pressure changes in a continuous stirred reactor at different stages, clearly showing the characteristics of different stages of the reaction process. From 0 to 60 minutes, the temperature fluctuates drastically between 80 and 110°C, and the pressure oscillates significantly between 0.8 and 1.5 MPa, reflecting the unstable state during equipment startup and initial material mixing. From 60 to 240 minutes, the temperature rapidly rises to 180–200°C and remains high, while the pressure fluctuates between 1.4 and 2.0 MPa, representing the main range of chemical reaction. A significant pressure drop occurs around 180 minutes, possibly corresponding to a critical operation or stage transition. From 240 to 300 minutes, the temperature gradually drops from 190°C to 140°C, and the pressure decreases from 1.7 MPa to 1.1 MPa, indicating that the reaction is nearing completion and the system is entering a depressurization and cooling process. Pressure and temperature exhibit a strong coupling relationship within 60–240 minutes, with pressure increasing with temperature, reflecting typical characteristics of reaction kinetics.

[0038] In one embodiment of the present invention, the step of continuously updating the causal inference network model is specifically implemented as follows: The multi-source operating time series data set collected during a new monitoring period, along with its corresponding reaction process tracing report and operating status monitoring conclusions, is used as a new training sample. For example, during the next four-hour monitoring period, new stirring motor drive signals, new reactant component sensor readings, new reactor temperature and pressure monitoring data, and new online analysis results of reaction product quality are collected, constituting a new multi-source operating time series data set. The data for the new monitoring period is analyzed according to the method of the previous embodiment to generate a new reaction process tracing report, and a new operating status monitoring conclusion is determined. The new multi-source operating time series data set, the new reaction process tracing report, and the new operating status monitoring conclusions are packaged together as a new training sample to update the existing causal inference network model.

[0039] In specific implementation, the partial correlation between the features of the characteristic data segments and the online analysis results of the reaction product quality is recalculated using new training samples. Feature data segments are extracted from the new training samples, such as the average in-vessel temperature in the first process interval of the new monitoring period and the variance of in-vessel pressure in the second process interval of the new monitoring period. Simultaneously, new online analysis results of the reaction product quality are obtained, such as new yield index values. The partial correlation coefficient between the new feature data segments and the new yield index is calculated to analyze whether the correlation strength between the features and the quality index changes under the new data conditions. In some embodiments, the formula for calculating the partial correlation can be expressed as: ;in: This represents the partial correlation coefficient between variables x and y, with control over variable z. , , These represent the simple correlation coefficients between variables x and y, x and z, and y and z, respectively. This formula measures the net correlation between two variables after removing the influence of other variables. If the absolute value of the recalculated partial correlation coefficient is greater than a preset correlation threshold, the feature node of that data segment is retained in the causal inference network model; otherwise, it may be necessary to consider removing or adjusting the feature node.

[0040] In practice, the Granger causality test is re-performed using new training samples. Data from the new monitoring period is combined with historical monitoring data to form a longer time-series data sequence, and the existing directed causal edges in the causal inference network model are re-examined. For example, after adding data points from the new monitoring period to the time-series data of the feature node "variance of in-vessel temperature monitoring data within the first process interval" and the quality indicator node "yield," the Granger causality test is re-executed to determine whether the original causal relationship still holds in the new, longer time-series data. If the F-statistic of the Granger causality test is no longer significant in the new time-series data, the directed edge from that feature node to the quality indicator node is removed from the causal inference network model. If a new feature data segment emerges and a significant Granger causal relationship is found between it and the quality indicator, a new feature node and a new directed edge are added to the causal inference network model.

[0041] In practice, based on the results of recalculation and retesting, the nodes, directed edges, and weights in the existing causal inference network model are incrementally updated or corrected. If the recalculated partial correlation coefficient value differs from the original partial correlation coefficient value in the model, the weights of the causal edges between the corresponding feature nodes and quality index nodes are updated to the recalculated partial correlation coefficient value or its function value. If the re-performed Granger causality test indicates that a certain original causal relationship no longer holds, the corresponding directed edge is deleted from the causal inference network model. If new training samples reveal feature-quality index relationships not included in the original model that satisfy the conditions of partial correlation and Granger causality test, new nodes and directed edges are added to the causal inference network model, and initial weights are assigned to the newly added directed edges. It can be understood that this incremental update is performed gradually, with each adjustment only made locally based on new training samples, rather than a complete reconstruction of the model. The updated causal inference network model reflects more comprehensive data information, including new monitoring periods, in terms of the structure and weights of its nodes and edges. This allows it to be used for monitoring the operational status of subsequent monitoring periods, enabling the monitoring conclusions to adapt to possible drifts or changes in the process.

[0042] Referring to Figure 5, this is a heatmap showing the causal contribution of key process characteristics in a continuous stirred reactor to product quality indicators. Darker colors represent higher causal contributions. Temperature change rate → Yield has the highest causal contribution, indicating that the rate of temperature change within the reactor is the most critical factor affecting product yield. Material composition → Purity has an extremely high causal contribution, indicating that precise control of raw material composition is crucial for product purity. Reaction time → Stability has an extremely high causal contribution, indicating that reaction time is a core variable determining process stability. Pressure change rate → Purity has a relatively high causal contribution, as pressure fluctuations significantly affect product purity. Stirring frequency → Stability has a relatively high causal contribution, and stable control of the stirring frequency helps improve process stability. This heatmap is a visual output of the causal inference network model and can be directly used to guide process optimization, prioritizing adjustments to high-contribution operational variables to improve product quality and process stability.

[0043] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis, characterized in that, include: A multi-source operating time series data set of a continuous stirred reactor is collected within a set monitoring period. The multi-source operating time series data set includes stirring motor drive signal, reactant component sensor readings, reactor internal temperature and pressure monitoring data, and online analysis results of reaction product quality. Cross-domain signal correlation analysis is performed on the stirring motor drive signal and reactant component sensor readings to identify the coupling correlation pattern between the drive signal and material changes, and a set of coupling correlation patterns is generated. Based on the aforementioned set of coupled correlation patterns, dynamic process segmentation is performed on the temperature and pressure monitoring data inside the reactor to divide it into multiple process intervals reflecting different chemical reaction stages, and feature data fragments within each process interval are extracted. Utilizing the mapping relationship between these feature data fragments and the online analysis results of the reaction product quality, a causal inference network model between the reaction state and product quality is constructed. Through this causal inference network model, the key process intervals and their feature data leading to the current online analysis results of the product quality are traced back to generate a reaction process traceability report. Based on the reaction process traceability report, the comprehensive operating status of the continuous stirred reactor during the current monitoring period is determined, and an operating status monitoring conclusion is generated.

2. The method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis according to claim 1, characterized in that, Cross-domain signal correlation analysis is performed on the stirring motor drive signal and the reaction material composition sensor readings to identify the coupling correlation patterns between the drive signal and material changes, generating a set of coupling correlation patterns. This includes: performing a spectral transformation on the stirring motor drive signal to convert it from a time-domain signal to a frequency-domain energy distribution signal; performing principal component dimensionality reduction on the reaction material composition sensor readings to extract the principal component change trajectory reflecting the trend of material composition changes; calculating the dynamic correlation coefficient sequence between the frequency-domain energy distribution signal and the principal component change trajectory over multiple preset time windows; performing pattern clustering analysis on the dynamic correlation coefficient sequence to identify different patterns where the correlation coefficient is consistently high, consistently low, or fluctuating periodically; and combining the identified different patterns with the original stirring motor drive signal characteristics and the original reaction material composition sensor reading characteristics within the corresponding time windows to form multiple coupling correlation patterns, constituting the set of coupling correlation patterns.

3. The method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis according to claim 2, characterized in that, Based on the set of coupling correlation patterns, dynamic process segmentation is performed on the in-vessel temperature and pressure monitoring data to divide it into multiple process intervals reflecting different chemical reaction stages. Feature data fragments within each process interval are extracted, including: identifying key coupling correlation patterns from the set of coupling correlation patterns that indicate the switching of chemical reaction stages; using the time point of occurrence of the key coupling correlation pattern as candidate segmentation points to initially segment the continuous time series of in-vessel temperature and pressure monitoring data; performing trend fitting on the in-vessel temperature and pressure monitoring data within each initial segment to extract feature combinations of temperature and pressure change rates; comparing the feature combinations of two adjacent initial segments; if the difference exceeds a preset threshold, the candidate segmentation point is confirmed as a valid process boundary; otherwise, adjacent segments are merged; based on all valid process boundaries, the in-vessel temperature and pressure monitoring data for the entire monitoring period is finally divided into multiple continuous process intervals, and data corresponding to each process interval is extracted from the original data as the feature data fragment.

4. The method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis according to claim 3, characterized in that, By utilizing the mapping relationship between the feature data segments and the online analysis results of the product quality, a causal inference network model between the reaction state and the product quality is constructed. This includes: for each process interval, pairing its corresponding feature data segment with the online analysis results of the product quality at the end of the reaction or the end of the interval to form process-quality pairing samples; analyzing multiple process-quality pairing samples to calculate the partial correlation between the statistical characteristic values ​​of different feature data segments and various indicators in the online analysis results of the product quality; based on the partial correlation calculation results, selecting feature data segments whose absolute value of the partial correlation coefficient with at least one quality indicator in the online analysis results of the product quality is greater than a preset correlation threshold; using the selected feature data segments and various quality indicators in the online analysis results of the product quality as feature nodes and quality indicator nodes of the causal inference network model, respectively; and using the temporal Granger causality test to analyze the predictive ability of the temporal changes of the feature data segment features on subsequent changes in the product quality indicators, establishing directed causal edges from feature nodes to quality indicator nodes, thereby constructing the causal inference network model.

5. The method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis according to claim 4, characterized in that, Using the causal inference network model, the key process intervals and their characteristic data leading to the current product quality online analysis results are traced backward to generate a reaction process traceability report. This includes: obtaining the final reaction product quality online analysis results for the current monitoring period; locating the quality indicator node corresponding to the final reaction product quality online analysis results in the causal inference network model; starting from the quality indicator node, traversing backward along the directed causal edges pointing to the quality indicator node to find all characteristic data fragment feature nodes that directly or indirectly point to the quality indicator node; locating the key process intervals that contribute to the final quality results based on the relationship between the characteristic data fragment feature nodes and the process intervals; aggregating the characteristic data fragment features extracted from the key process intervals and combining them with their causal contribution in the causal inference network model to generate the reaction process traceability report describing how each key process interval affects the final product quality.

6. The method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis according to claim 5, characterized in that, Based on the relationship between feature nodes of feature data segments and process intervals, the key process intervals that contribute to the final quality result are located. This includes: obtaining a list of all feature nodes of feature data segments that point to the final quality indicator node in the causal inference network model; querying the mapping relationship table between feature data segment features and original process intervals to determine the process interval from which each feature node of feature data segment in the list originates; counting the frequency of each process interval being referenced by nodes in the feature node list of feature data segments, and summarizing the total causal contribution of all feature nodes of feature data segments from the same process interval to the final quality indicator; and marking process intervals whose total causal contribution exceeds a preset threshold as key process intervals.

7. The method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis according to claim 6, characterized in that, The process involves aggregating feature data fragments extracted from key process intervals and combining their causal contribution in a causal inference network model to generate a reaction process traceability report describing how each key process interval affects the quality of the final product. This includes: creating a report entry for each marked key process interval; recording the start and end times of the key process interval and the core feature values ​​of the in-vessel temperature and pressure monitoring data within the key process interval in the report entry; extracting the weights of the directed causal edges pointing from each feature data fragment feature node from the key process interval to the final quality indicator node from the causal inference network model, as the causal contribution; listing each feature data fragment feature and its corresponding causal contribution in detail in the report entry, and sorting them from high to low causal contribution; organizing all report entries for all key process intervals in chronological order and adding an overall summary to form the reaction process traceability report.

8. The method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis according to claim 7, characterized in that, Based on the reaction process traceability report, the overall operating status of the continuous stirred reactor during the current monitoring period is determined, including: parsing the reaction process traceability report and extracting the deviation between the features of characteristic data segments in each key process interval and their standard reference values; calculating the weighted average of the deviations of all key process intervals, where the weights are determined by the causal contribution of each key process interval; comparing the weighted average with a preset operating status threshold range; if the weighted average is within the preset normal operating status threshold range, the overall operating status of the continuous stirred reactor is determined to be normal; if the weighted average exceeds the normal operating status threshold range but does not reach the warning status threshold, the overall operating status is determined to be requiring attention; if the weighted average reaches or exceeds the warning status threshold, the overall operating status is determined to be abnormal.

9. The method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis according to claim 8, characterized in that, The calculation of the weighted average of the deviations of all critical process intervals includes: for each critical process interval in the reaction process traceability report, calculating the average of the absolute deviations of all characteristic data segments listed in its report entry from their respective standard reference values, as the characteristic deviation of the critical process interval; obtaining the total causal contribution of the critical process interval to the final quality result from the reaction process traceability report; using the total causal contribution of each critical process interval as its weight, performing a weighted summation of the characteristic deviations of all critical process intervals; and dividing the result of the weighted summation by the sum of the weights of all critical process intervals to obtain the weighted average.

10. The method for monitoring the operating status of a continuous stirred reactor based on time-series data analysis according to claim 9, characterized in that, It also includes the step of continuously updating the causal inference network model, specifically including: using the multi-source operating time series data set collected in the new monitoring period and its corresponding reaction process traceability report and operating status monitoring conclusion as a new training sample; using the new training sample, recalculating the partial correlation between the features of the feature data fragments and the online analysis results of the quality of the reaction products, and re-performing the time-series Granger causality test; based on the recalculation and test results, incrementally updating or correcting the nodes, directed edges and weights in the existing causal inference network model; and using the updated causal inference network model for operating status monitoring in subsequent monitoring periods.

Images