A method for real-time monitoring of synthetic reaction progress based on principal component analysis
By imposing mechanistic constraints and dynamic endpoint prediction on the principal component analysis model, the problem of false alarms caused by sensor drift was solved, improving the reliability and continuity of synthetic reaction monitoring and enabling efficient real-time monitoring of chemical reaction processes.
Patent Information
- Application Number
- CN202611220050.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-12
- Publication Date
- 2026-11-06
AI Technical Summary
Existing principal component analysis models are prone to misinterpreting data variations as abnormal reaction states when sensor measurement deviations occur due to factors such as changes in ambient temperature, probe contamination, or electron drift, leading to false alarms. Furthermore, they lack constraints on the intrinsic mechanisms of chemical reactions, resulting in low reliability of monitoring results in actual industrial production.
By acquiring asynchronous time-series signals from multiple sensor sources, performing time alignment and sampling frequency normalization, calculating the reaction heat release rate to divide the reaction stages, calling the local principal component analysis model and rotating it according to the mechanism direction constraint, and combining fuzzy membership strategy and dynamic endpoint prediction, an enhanced principal component trajectory is constructed for real-time monitoring.
It significantly reduces the false alarm rate, improves the reliability and interpretability of monitoring results, enables adaptive identification and correction of sensor drift, and provides continuous and stable intelligent monitoring decisions.
Smart Images
Figure CN123053273A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of chemical principal component analysis technology, specifically to a method for real-time monitoring of the synthesis reaction process based on principal component analysis. Background Technology
[0002] Principal component analysis, as a classic multivariate statistical process monitoring method, has been widely used in the field of online monitoring of chemical synthesis reactions in recent years because it can project high-dimensional process data into a low-dimensional feature space and effectively extract the main directions of process variation.
[0003] Currently, principal component analysis (PCA)-based synthetic reaction monitoring technologies typically employ a basic architecture of "global PCA model + fixed statistical control limits": first, a single PCA model is established using historical normal batch data; then, the principal component score or squared prediction error, etc., are calculated online at each sampling time; and alarms for abnormal states are triggered by determining whether these statistics exceed preset control limits. With the development of industrial IoT technology, some improved solutions have begun to introduce multi-sensor data fusion strategies, inputting multi-source signals such as temperature, pressure, near-infrared spectroscopy, and stirring torque into the PCA model after time alignment. Other researchers have proposed establishing multiple local PCA models for different stages of the reaction to address the challenge of global models failing to adapt to changes in the dynamic characteristics of the process.
[0004] However, the existing technologies still have the following shortcomings: the existing principal component analysis model is essentially a purely data-driven method, and its principal component extraction direction depends entirely on the direction of the maximum variance in the process data, lacking constraints on the intrinsic mechanism of chemical reactions.
[0005] When sensors experience measurement deviations due to factors such as changes in ambient temperature, probe contamination, or electronic drift, the data variance structure changes. In this case, principal component analysis models are prone to misinterpreting the data variations caused by sensor drift as abnormal fluctuations in the reaction state, resulting in a large number of false alarms. At the same time, because the physical meaning of principal components is unclear, operators find it difficult to distinguish whether the exceeding of statistical limits is due to real process abnormalities or sensor performance degradation, which seriously restricts the reliability of monitoring results in actual industrial production.
[0006] To address this issue, this application proposes a real-time monitoring method for the synthesis reaction process based on principal component analysis. Summary of the Invention
[0007] The purpose of this invention is to provide a real-time monitoring method for the synthesis reaction process based on principal component analysis (PCA). This addresses the problem in existing technologies where, due to factors such as changes in ambient temperature, probe contamination, or electron drift, the data variance structure changes. In such cases, PCA models are prone to misinterpreting data variations caused by sensor drift as abnormal fluctuations in the reaction state, resulting in numerous false alarms. Furthermore, because the physical meaning of principal components is unclear, operators find it difficult to distinguish whether the exceeding of statistical limits stems from genuine process anomalies or sensor performance degradation, severely limiting the reliability of monitoring results in actual industrial production.
[0008] To achieve the above objectives, the present invention provides the following technical solution:
[0009] A method for real-time monitoring of the synthesis reaction process based on principal component analysis, comprising:
[0010] Asynchronous time-series signals output by multiple sources of sensors during the synthesis reaction are acquired, and the asynchronous time-series signals are time-aligned and sample frequency normalized to obtain a synchronous multi-source data matrix.
[0011] The current heat release rate of the reaction is calculated based on the temperature and pressure data in the synchronous multi-source data matrix, and the reaction induction period, main reaction period and termination period are divided by the heat release rate of the reaction to obtain the stage category label at the current moment;
[0012] Based on the stage category labels, the corresponding local principal component analysis model is invoked to extract principal components from the synchronous multi-source data matrix, thereby obtaining a low-dimensional principal component score vector representing the reaction state.
[0013] Based on the low-dimensional principal component score vector and the mechanism direction vector constructed according to the reaction conversion rate and instantaneous exothermic rate calculated in real time from the synchronous multi-source data matrix, the load space of the low-dimensional principal component score vector is gradually rotated towards the mechanism direction vector using the Givens rotation method. The rotation step size is adaptively determined by the current angle deviation to obtain the enhanced principal component trajectory constrained by the reaction kinetic direction.
[0014] Based on the direction of movement and convergence speed of the enhanced principal component trajectory in space, the trajectory landing point within a preset time window is predicted, and a reaction endpoint warning signal is output based on the distance between the trajectory landing point and the safety boundary.
[0015] Furthermore, the step of performing time alignment and sampling frequency normalization on the asynchronous time-series signal includes:
[0016] Using the near-infrared spectral acquisition time as the global reference timestamp, cubic spline interpolation is performed on the temperature and stirring torque signals, and the interpolated data is weighted according to the preset instrument accuracy coefficients of each sensor to obtain the synchronous multi-source data matrix.
[0017] Furthermore, when dividing the reaction induction period, main reaction period, and termination period based on the reaction heat release rate, the stage criterion includes at least the reaction heat release rate criterion, supplemented by the near-infrared spectral conversion rate criterion for comprehensive determination, including:
[0018] The reaction heat release rate at the current moment is calculated based on temperature and pressure data, and the real-time conversion rate at the current moment is calculated based on the area of the characteristic absorption peaks in the near-infrared spectrum.
[0019] When the first derivative of the reaction heat release rate with respect to time changes from positive to negative and continues to decrease beyond the first threshold, and the real-time conversion rate reaches the preset conversion rate threshold, it is determined that the main reaction period has begun.
[0020] When the rate of heat release from the reaction drops to below a first percentage of the initial peak and the rate of change of the real-time conversion rate is less than a preset rate of change threshold, the process is determined to enter the termination period, and the stage category label is updated according to the determination result.
[0021] Furthermore, when the corresponding local principal component analysis model is invoked to extract principal components from the synchronous multi-source data matrix, if the stage category label is within the boundary between the induction period and the main reaction period, the fuzzy membership coefficient is determined by calculating the slope of the rate of change of the heat release rate of the reaction.
[0022] Based on the fuzzy membership coefficients, the induction period model and the main reaction period model are run simultaneously, and the initial principal component vectors output by the two models are weighted and summed to obtain the low-dimensional principal component score vector with a smooth transition, so as to eliminate the step jump in monitoring values caused by hard switching of models.
[0023] Furthermore, the kinetic characteristic parameters include the reaction conversion rate and the instantaneous heat release rate at the current moment, wherein the reaction conversion rate is obtained by converting the real-time reduction value of the area of the near-infrared spectral characteristic absorption peak, and the instantaneous heat release rate is calculated by combining the current in-vessel temperature and the jacket temperature difference with the heat transfer coefficient.
[0024] The product of the reaction conversion rate and the instantaneous exothermic rate is used as the rotation reference quantity, and the direction of the change gradient of the rotation reference quantity within a preset time window before the current moment is used as the mechanism direction vector. An orthogonal rotation matrix is constructed using the Givens rotation method.
[0025] The original coordinate axes of the low-dimensional principal component score vector are rotated by the orthogonal rotation matrix, so that the load direction of the first principal component is gradually aligned with the mechanism direction vector, with each rotation step not exceeding a preset angle threshold, to obtain the enhanced principal component trajectory constrained by the reaction mechanism.
[0026] Furthermore, during the construction of the orthogonal rotation matrix, the coordinate axis rotation is further limited to the principal component dimensions in the low-dimensional principal component score vector where the cumulative variance contribution rate is greater than a preset support threshold.
[0027] After rotation, the data reconstruction residual of the current enhanced principal component trajectory is calculated. If the reconstruction residual exceeds the dynamic tolerance band, the reconstruction residual contribution of each original variable dimension in the synchronous multi-source data matrix is further calculated, and the sensor channel corresponding to one or more variables with the largest contribution is determined as the drift source.
[0028] For the sensor channel corresponding to the identified drift source, calculate the drift compensation factor for that channel, and use the drift compensation factor to perform reverse correction on the original data of that channel.
[0029] Furthermore, when predicting the trajectory landing point within a future preset time window, a prediction model with a decay factor is constructed using the first and second differences of the enhanced principal component trajectory in the most recent historical sequence. When the second difference exceeds a preset amplitude threshold, the weight of the second term is automatically reduced to obtain the predicted trajectory point set.
[0030] Using the envelope of the enhanced principal component trajectory of the historical normal batch as the safety boundary, the shortest spatial distance from each point in the predicted trajectory point set to the safety boundary is calculated, and the time variation sequence of the shortest spatial distance is recorded.
[0031] Furthermore, the reaction endpoint warning signal includes a stage evolution prompt signal and an endpoint confirmation signal;
[0032] When the time change sequence of the shortest spatial distance shows a decreasing trend and the predicted trajectory moves in the direction of the safety boundary, the stage evolution prompt signal is output and the duration of the preset time window is shortened.
[0033] When the shortest spatial distance remains within the convergence quiescent threshold for multiple consecutive prediction periods, and the rate of change of the modulus of the enhanced principal component trajectory is less than the termination rate threshold, the endpoint confirmation signal is output.
[0034] Furthermore, after obtaining the synchronous multi-source data matrix and before calling the local principal component analysis model, the enrichment rate of the key intermediate at the current moment is calculated based on the characteristic absorption peak area corresponding to the near-infrared spectral data in the synchronous multi-source data matrix.
[0035] The boundary for the heat release rate of the reaction is corrected using the enrichment change rate:
[0036] When the deviation between the induction period termination time determined based on the reaction heat release rate and the induction period termination time determined based on the peak enrichment rate of the key intermediate exceeds a preset deviation threshold, the time corresponding to when the enrichment rate of the key intermediate reaches its peak is taken as the precise anchor point for the termination of the induction period, and the update time of the stage category label is forcibly corrected based on the precise anchor point.
[0037] Furthermore, after obtaining the low-dimensional principal component score vector, based on the dynamic change trend of the loading coefficients of different spectral wavelength variables in the synchronous multi-source data matrix, the subset of feature wavelengths with the highest sensitivity to the current stage category label is selected.
[0038] The variables within the subset of characteristic wavelengths are assigned covariance weights higher than the default values, and the principal component extraction is re-executed with the weighted covariance matrix, thereby enhancing the response sensitivity of the updated low-dimensional principal component score vector to intermediate structure transformation.
[0039] Compared with existing technologies, the present invention provides a real-time monitoring method for synthetic reaction processes based on principal component analysis. By introducing reaction kinetic characteristics as physical bias terms into the principal component extraction process, the data-driven principal components are subjected to coordinate rotation under mechanistic constraints. This allows the extracted enhanced principal component trajectories to reflect both statistical data variation and the fundamental laws of chemical reaction thermodynamics and kinetics. This solves the problem that traditional principal component analysis, due to the lack of physical constraints, misjudges sensor drift as process anomalies, and significantly improves the reliability of monitoring results.
[0040] This invention employs a soft-switching strategy for stage models based on fuzzy membership degrees, and combines the convergence trend of principal component trajectories in space for dynamic endpoint prediction. It can adaptively identify the fuzzy boundaries of the induction period, main reaction period, and termination period without relying on fixed time thresholds. At the same time, it issues phased early warning signals in advance by analyzing the distance evolution between the trajectory landing point and the safety boundary, effectively avoiding the problems of monitoring value jumps and endpoint misjudgments caused by hard model switching. It provides a continuous and stable intelligent monitoring and decision chain for the entire synthetic reaction process. Attached Figure Description
[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0042] Figure 1 A flowchart of a method for real-time monitoring of the synthesis reaction process based on principal component analysis is provided for embodiments of the present invention. Detailed Implementation
[0043] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.
[0044] As attached Figure 1 As shown:
[0045] Example 1:
[0046] This embodiment is applied to the batch polymerization process in a fine chemical workshop for producing high-solids-content acrylate copolymer intermediates (as pressure-sensitive adhesive matrix).
[0047] The reaction system uses azobisisobutyronitrile (AIBN, accounting for 0.6% of the total monomer mass) as the initiator and ethyl acetate as the solvent (solvent / monomer mass ratio of 0.3:1). The initial feed amount is 3.2 tons, the target reaction temperature is set at 78℃±1℃, and the agitator adopts a slanted blade disc turbine with a constant operating speed of 80 rpm.
[0048] The reactor is equipped with four types of online sensors:
[0049] Fourier transform near-infrared spectrometer (immersion fiber optic probe, 2mm optical path, spectral scanning range 4000~10000cm) -1 The system includes a spectrum acquisition every 1.0 second, a sheathed platinum resistance thermometer (Grade A Pt100, sampling frequency 50Hz, response time 0.8 seconds), a diaphragm pressure transmitter (range 0~1.6MPa, accuracy 0.075%, sampling frequency 2Hz), and a stirring shaft torque sensor (strain gauge type, sampling frequency 5Hz, resolution 0.1N·m).
[0050] The actual monitoring steps during operation are as follows:
[0051] I. Time Alignment and Reliability Fusion of Multi-Source Asynchronous Signals:
[0052] Using the acquisition time of the near-infrared spectrometer (e.g., the whole second t=2026-04-21 09:00:00.000) as the global reference timestamp, the original asynchronous sampling sequences of temperature, pressure and stirring torque are subjected to cubic spline interpolation and uniformly resampled to a fixed frequency of 1.0 seconds / frame.
[0053] After interpolation, a confidence weight is assigned based on the instrument accuracy coefficient in the factory calibration certificate of each sensor:
[0054] The weight of the temperature signal is 0.98 (compensated for thermal response lag), the weight of the pressure signal is 0.85 (the diaphragm is occasionally delayed due to polymer adhesion), the weight of the stirring torque is 0.92 (the friction torque at the shaft end fluctuates slightly with increasing temperature), and the overall weight of the near-infrared spectrum is 0.95 (the anti-contamination coating of the probe is intact).
[0055] The four weighted interpolation sequences are aligned and spliced according to the timestamp to form a synchronous multi-source data matrix X(t) containing 401 variables per frame (400 spectral wavenumber points + 1 temperature + 1 pressure + 1 torque, of which 400 points are extracted from the characteristic region of 1600~1850nm in the spectral band). The matrix dimension of each frame is (1×403).
[0056] II. Self-identification of the reaction phase and precise anchoring of the termination of the induction period:
[0057] The vessel's internal temperature T(t) and jacket temperature difference ΔT(t) for the current frame are extracted in real time from the synchronous multi-source data matrix, combined with the known heat transfer coefficient K (the vessel's calibration value is 215 W / (m³)). 2 ·K)) and heat exchange area S (18.5m 2 The reaction heat release rate q(t) (unit kW) at the current moment is calculated according to the formula q(t)=K·S·ΔT(t).
[0058] Regarding the main phase division:
[0059] Continuously calculate the sliding window first derivative dq / dt of q(t) with respect to time (window length is 30 frames, i.e. 30 seconds).
[0060] When the dq / dt value changes from positive to negative, and remains negative for 5 consecutive frames within the window with an absolute value exceeding the threshold of 0.015 kW / s, the current time t1 is determined to be the turning point from the end of the induction period to the beginning of the main reaction period; when q(t) drops to below 8% of the initial peak value qpeak observed throughout the process, and remains so for 120 seconds without rebounding, it is determined to enter the termination period.
[0061] Enhanced sensitivity to intermediates:
[0062] Meanwhile, at 1620 cm⁻¹ in the near-infrared spectrum -1 At the location (corresponding to the characteristic absorption peak of the C=C double bond in MMA), the difference of the absorption peak area A(t) with respect to time is calculated in real time to obtain the enrichment rate δA / δt of the key intermediates (initiating free radicals and active centers of the growth chain).
[0063] The above heat release boundary was corrected using this enrichment change rate:
[0064] The moment t2 when δA / δt reaches its positive peak (when the double bond opens fastest) is taken as the precise anchor point for the termination of the induction period.
[0065] If the deviation between t2 and t1 determined based on heat release exceeds ±10 seconds, then t2 will be used as the standard for the forced correction of the stage category label update time, ensuring that the initial state of monomer conversion rate has been accurately locked when entering the main reaction period.
[0066] III. Fuzzy Soft Handover PCA in the Boundary Interval:
[0067] When the stage category label indicates that the current time is within the boundary between the induction period and the main reaction period (defined as a window with a width of 100 seconds from 40 seconds before t2 to 60 seconds after t2), a hard model switch will no longer be performed.
[0068] Instead, it calculates the normalized value of the current reaction heat release rate slope k=dq / dt within the window (range mapped to [-1,1]), and uses the sigmoid membership function:
[0069] μ=1 / (1+exp(-β·k))
[0070] The value of β is 3.2, which makes μ≈0.5 when k=-0.5, thus obtaining the fuzzy membership coefficient μ.
[0071] In each frame, the "Local PCA Model for Induction Period" (retaining 3 principal components) and the "Local PCA Model for Main Reaction Period" (retaining 5 principal components) pre-stored in the industrial control computer are called simultaneously, and the initial principal component vectors PCind and PCmid are output respectively. Finally, the low-dimensional principal component score vector PCsmooth=(1-μ)·PCind+μ·PCmid is obtained by weighted summation.
[0072] Within this window, the monitored values (such as the overall principal component magnitude) remain continuous without step jumps, and the maximum single-frame jump variable is limited to within 0.05σ (σ is the historical standard deviation of the principal component).
[0073] IV. Mechanism-based constraint enhancement: self-compensation for dynamic reference rotation and drift:
[0074] (1) Coordinate rotation enhancement:
[0075] Once the reaction enters the main reaction period (μ=1), the real-time monomer conversion rate α(t) at the current moment is taken, and the result is obtained from the near-infrared 1620cm wavelength. -1 The normalized reduction value of the peak area is calculated, along with the instantaneous heat release rate q(t).
[0076] The product of the two, Φ(t) = α(t)·q(t), is used as the rotation reference quantity.
[0077] Using this benchmark, an orthogonal rotation matrix R is constructed. This matrix acts on the coordinate axes of the space containing the original low-dimensional principal component score vectors, with the following specific rules:
[0078] By using Givens rotation, the load vector direction of the first principal component is gradually brought closer to the gradient change direction of Φ(t) (i.e., the difference direction of Φ(t) in the past 10 frames), with each rotation step limited to ≤3° to avoid overcorrection.
[0079] After rotation, the enhanced principal component trajectory PCenhanced(t) constrained by the reaction mechanism is obtained.
[0080] (2) Reconstruction residual drift detection:
[0081] During the rotation, only the first three principal components whose cumulative variance contribution exceeds 85% are rotated along their axes; higher-order principal components (the 4th and 5th) below this support threshold remain unchanged. After the rotation, the synchronous data matrix is reconstructed using the rotated load matrix, and the reconstruction residual modulus RE(t) between the reconstructed data and the original synchronous multi-source data matrix is calculated.
[0082] The preset dynamic tolerance band is 3 times the standard deviation of the residual modulus of the historical normal batches (first 20 batches), i.e., ±3σRE.
[0083] The specific calculation method of the dynamic tolerance band is as follows: Take the reconstructed residual modulus sequence of the most recent consecutive M batches of normal reaction batches (M≥15) in the corresponding stage, calculate the mean μRE and standard deviation σRE of the sequence, the upper limit of the tolerance band is μRE+k·σRE, and the lower limit is μRE-k·σRE, where k is a preset multiple (value 2.5~4.0); when a new batch is completed, the mean of the reconstructed residual modulus of the batch is included in the historical database, and outlier batches that exceed the range of μRE±3σRE are removed. μRE and σRE are updated in a rolling window manner to realize the dynamic adaptive adjustment of the tolerance band.
[0084] During a certain actual operation, after 2 hours and 17 minutes, RE(t) exceeded the upper tolerance band by 4.1σRE for 3 consecutive frames. The system determined that the stirring torque sensor had temperature drift (caused by the strain gauge drifting due to the increase in shaft temperature). It immediately triggered the drift compensation factor γ=RE(t) / (REnormal). This factor was used to reverse the torque column vector in the synchronous multi-source data matrix (multiplied by 1 / γ), so that RE(t) fell back to within the tolerance band within 5 frames, realizing online self-calibration during long-term operation.
[0085] V. Trajectory Landing Point Prediction and Multi-Level Early Warning for Destination:
[0086] (1) Comparison between prediction and safety boundary:
[0087] Based on the historical sequence of the enhanced principal component trajectory PCenhanced over the most recent 60 frames (i.e., the past 60 seconds), its first-order difference δ1(t) and second-order difference δ2(t) are calculated, and a second-order extrapolation polynomial PCpred(t+Δt) = PCenhanced(t) + δ1(t)·Δt + 0.5·δ2(t)·Δt is constructed. 2 It is used to predict the set of trajectory points within a 60-second window.
[0088] By pre-storing the enhanced principal component trajectories of 20 historical normal batches, and using the convex hull algorithm to draw the safe envelope (with a Mahalanobis distance threshold of 3.2) on the two-dimensional plane of the first principal component vs. the second principal component.
[0089] The shortest spatial distance dmin(t) from each point in the predicted trajectory point set to the safety envelope is calculated in real time, and the time change sequence of this distance is recorded.
[0090] (2) Phased early warning output: Assuming that the current main reaction period has lasted for 80 minutes, if it is observed that dmin(t) shows a monotonically decreasing trend in 15 consecutive prediction cycles (60 seconds per cycle), and the direction vector of the prediction trajectory points to the outside of the safety envelope in the PC plane (with an angle of less than 15°), the system will immediately output a "phase evolution prompt signal" to remind the operator to prepare for cooling and adding the terminator; at the same time, the preset time window will be automatically shortened from 60 seconds to 30 seconds to improve the timeliness of subsequent predictions.
[0091] Subsequently, when dmin(t) gradually increases and converges again, and the rate of change of the modulus of the enhanced principal component trajectory remains below 0.002 (dimensionless) / frame for 20 consecutive prediction periods, and the modulus value is stable within ±1.2σ of the historical termination state center defined by the local PCA model at the termination period, it is determined that the rate of change of the modulus is less than the termination rate threshold (set to 0.005 / frame), and the system outputs an "end point confirmation signal".
[0092] At this point, the reaction conversion rate was verified to be 97.8% by offline gas chromatography. The system automatically notifies the DCS to open the cooling water valve and the terminator (hydroquinone) feed pump, completing the endpoint determination 4 minutes in advance, effectively avoiding the risk of widening of molecular weight distribution and gelation caused by excessive heat preservation.
[0093] VI. Dynamic wavelength weighting to enhance intermediate response:
[0094] Throughout the main reaction period, in addition to the above-mentioned fixed process, the system dynamically calculates the load coefficient change trend of each wavelength variable (400 wavenumber points in total) in the near-infrared spectrum of the synchronous multi-source data matrix under the current load matrix every 200 frames (200 seconds) as a cycle—specifically, the Euclidean deviation of the load coefficient in two adjacent cycles.
[0095] The top 5% of wavelengths with the largest deviations (i.e., 20 characteristic wavelengths, usually concentrated in the 1630~1650nm and 1720~1750nm ranges, corresponding to carbonyl overtones and hydrocarbon combination bands) are selected. The covariance weights of these wavelength variables are set to 3 times the default value (default value is 1), and the principal component extraction of the local PCA model is re-executed with the weighted covariance matrix.
[0096] The recalculated low-dimensional principal component score vector is updated. The response sensitivity of this vector to intermediates (especially the concentration fluctuations of terminal olefins during molecular weight growth) is improved by about 40% compared to the unweighted version. This allows the system to clearly capture the inflection point of the polymerization chain termination rate in the last 30 minutes before endpoint confirmation, providing more reliable preliminary features for endpoint confirmation.
[0097] To verify the technical effect of the method of the present invention, comparative tests were conducted on the same acrylate copolymerization reaction system (reaction conditions were the same as in Example 1) using the traditional global PCA method, the multimode hard switching PCA method, and the method of the present invention. Each method was run for 20 consecutive batches, and the statistical results are shown in Table 1 below.
[0098] Table 1
[0099] Evaluation metrics Traditional global PCA Multi-modal hard-switching PCA The method of the invention False alarm rate (events / batch) 4.2±1.1 2.8±0.9 0.6±0.3 Missed alarm rate (%) 12.5 7.8 1.2 Average early termination decision time (min) -1.5 (lag) 1.8 4.2 Maximum statistical quantity step (σ) at phase junctions — 1.8 0.05
[0100] Experimental results show that, compared with existing technologies, the false alarm rate of the method of the present invention is reduced by about 75% to 85%, the endpoint can be determined about 4 minutes earlier on average, and the monitoring statistics at the stage boundary remain continuous without step change. The fluctuation range of the endpoint conversion rate is reduced by about 60% to 75%, which verifies the significant advantages of the present invention in terms of monitoring continuity, endpoint prediction timeliness and result reliability.
[0101] As can be seen from the above, this invention introduces reaction kinetic characteristics as physical bias terms into the principal component extraction process, and performs coordinate rotation on the data-driven principal components under mechanistic constraints. This allows the extracted enhanced principal component trajectories to reflect both statistical variation in the data and the basic laws of chemical reaction thermodynamics and kinetics. This solves the problem that traditional principal component analysis, due to the lack of physical constraints, misjudges sensor drift as process anomalies, and significantly improves the interpretability and reliability of monitoring results.
[0102] This invention employs a soft-switching strategy for stage models based on fuzzy membership degrees, and combines the convergence trend of principal component trajectories in space for dynamic endpoint prediction. It can adaptively identify the fuzzy boundaries of the induction period, main reaction period, and termination period without relying on fixed time thresholds. At the same time, it issues phased early warning signals in advance by analyzing the distance evolution between the trajectory landing point and the safety boundary, effectively avoiding the problems of monitoring value jumps and endpoint misjudgments caused by hard model switching. It provides a continuous and stable intelligent monitoring and decision chain for the entire synthetic reaction process.
[0103] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.
Claims
1. A method for real-time monitoring of the synthesis reaction process based on principal component analysis, characterized in that, include: Asynchronous time-series signals output by multiple sources of sensors during the synthesis reaction are acquired, and the asynchronous time-series signals are time-aligned and sample frequency normalized to obtain a synchronous multi-source data matrix. The current heat release rate of the reaction is calculated based on the temperature and pressure data in the synchronous multi-source data matrix, and the reaction induction period, main reaction period and termination period are divided by the heat release rate of the reaction to obtain the stage category label at the current moment; Based on the stage category labels, the corresponding local principal component analysis model is invoked to extract principal components from the synchronous multi-source data matrix, thereby obtaining a low-dimensional principal component score vector representing the reaction state. Based on the low-dimensional principal component score vector and the mechanism direction vector constructed according to the reaction conversion rate and instantaneous exothermic rate calculated in real time from the synchronous multi-source data matrix, the load space of the low-dimensional principal component score vector is gradually rotated towards the mechanism direction vector using the Givens rotation method. The rotation step size is adaptively determined by the current angle deviation to obtain the enhanced principal component trajectory constrained by the reaction kinetic direction. Based on the direction of movement and convergence speed of the enhanced principal component trajectory in space, the trajectory landing point within a preset time window is predicted, and a reaction endpoint warning signal is output based on the distance between the trajectory landing point and the safety boundary.
2. The method for real-time monitoring of synthetic reaction progress based on principal component analysis according to claim 1, characterized in that, The process of time-aligning and sampling frequency normalizing the asynchronous time-series signal includes: Using the near-infrared spectral acquisition time as the global reference timestamp, cubic spline interpolation is performed on the temperature and stirring torque signals, and the interpolated data is weighted according to the preset instrument accuracy coefficients of each sensor to obtain the synchronous multi-source data matrix.
3. The method for real-time monitoring of synthetic reaction progress based on principal component analysis according to claim 1, characterized in that, When dividing the reaction induction period, main reaction period, and termination period based on the reaction heat release rate, the stage criterion includes at least the reaction heat release rate criterion, supplemented by the near-infrared spectral conversion rate criterion for comprehensive determination, including: The reaction heat release rate at the current moment is calculated based on temperature and pressure data, and the real-time conversion rate at the current moment is calculated based on the area of the characteristic absorption peaks in the near-infrared spectrum. When the first derivative of the reaction heat release rate with respect to time changes from positive to negative and continues to decrease beyond the first threshold, and the real-time conversion rate reaches the preset conversion rate threshold, it is determined that the main reaction period has begun. When the rate of heat release from the reaction drops to below a first percentage of the initial peak and the rate of change of the real-time conversion rate is less than a preset rate of change threshold, the process is determined to enter the termination period, and the stage category label is updated according to the determination result.
4. The method for real-time monitoring of synthetic reaction progress based on principal component analysis according to claim 1, characterized in that, When the corresponding local principal component analysis model is invoked to extract principal components from the synchronous multi-source data matrix, if the stage category label is within the boundary between the induction period and the main reaction period, the fuzzy membership coefficient is determined by calculating the slope of the rate of change of the heat release rate of the reaction. Based on the fuzzy membership coefficients, the induction period model and the main reaction period model are run simultaneously, and the initial principal component vectors output by the two models are weighted and summed to obtain the low-dimensional principal component score vector with a smooth transition, so as to eliminate the step jump in monitoring values caused by hard switching of models.
5. The method for real-time monitoring of synthetic reaction progress based on principal component analysis according to claim 1, characterized in that, The kinetic characteristic parameters include the reaction conversion rate and the instantaneous heat release rate at the current moment. The reaction conversion rate is obtained by converting the real-time reduction value of the area of the near-infrared spectral characteristic absorption peaks, and the instantaneous heat release rate is calculated by combining the current in-vessel temperature and the jacket temperature difference with the heat transfer coefficient. The product of the reaction conversion rate and the instantaneous exothermic rate is used as the rotation reference quantity, and the direction of the change gradient of the rotation reference quantity within a preset time window before the current moment is used as the mechanism direction vector. An orthogonal rotation matrix is constructed using the Givens rotation method. The original coordinate axes of the low-dimensional principal component score vector are rotated by the orthogonal rotation matrix, so that the load direction of the first principal component is gradually aligned with the mechanism direction vector, with each rotation step not exceeding a preset angle threshold, to obtain the enhanced principal component trajectory constrained by the reaction mechanism.
6. The method for real-time monitoring of synthetic reaction progress based on principal component analysis according to claim 5, characterized in that, During the construction of the orthogonal rotation matrix, the coordinate axis rotation is performed only on the principal component dimensions in the low-dimensional principal component score vector where the cumulative variance contribution rate is greater than a preset support threshold. After rotation, the data reconstruction residual of the current enhanced principal component trajectory is calculated. If the reconstruction residual exceeds the dynamic tolerance band, the contribution of the reconstruction residual of each original variable dimension in the synchronous multi-source data matrix is further calculated, and the sensor channel corresponding to one or more variables with the largest contribution is determined as the drift source. For the sensor channel corresponding to the identified drift source, calculate the drift compensation factor for that channel, and use the drift compensation factor to perform reverse correction on the original data of that channel.
7. The method for real-time monitoring of synthetic reaction progress based on principal component analysis according to claim 1, characterized in that, When predicting the trajectory landing point within a future preset time window, a prediction model with a decay factor is constructed using the first and second differences of the enhanced principal component trajectory in the most recent historical sequence. When the second difference exceeds a preset amplitude threshold, the weight of the second term is automatically reduced to obtain the predicted trajectory point set. Using the envelope of the enhanced principal component trajectory of the historical normal batch as the safety boundary, the shortest spatial distance from each point in the predicted trajectory point set to the safety boundary is calculated, and the time variation sequence of the shortest spatial distance is recorded.
8. The method for real-time monitoring of synthetic reaction progress based on principal component analysis according to claim 7, characterized in that, The reaction endpoint warning signal includes a stage evolution prompt signal and an endpoint confirmation signal; When the time change sequence of the shortest spatial distance shows a decreasing trend and the predicted trajectory moves in the direction of the safety boundary, the stage evolution prompt signal is output and the duration of the preset time window is shortened. When the shortest spatial distance remains within the convergence quiescent threshold for multiple consecutive prediction periods, and the rate of change of the modulus of the enhanced principal component trajectory is less than the termination rate threshold, the endpoint confirmation signal is output.
9. The method for real-time monitoring of synthetic reaction progress based on principal component analysis according to claim 1, characterized in that, After obtaining the synchronous multi-source data matrix and before calling the local principal component analysis model, the enrichment rate of the key intermediate at the current moment is calculated based on the characteristic absorption peak area corresponding to the near-infrared spectral data in the synchronous multi-source data matrix. The boundary for the heat release rate of the reaction is corrected using the enrichment change rate: When the deviation between the induction period termination time determined based on the reaction heat release rate and the induction period termination time determined based on the peak enrichment rate of the key intermediate exceeds a preset deviation threshold, the time corresponding to when the enrichment rate of the key intermediate reaches its peak is taken as the precise anchor point for the termination of the induction period, and the update time of the stage category label is forcibly corrected based on the precise anchor point.
10. The method for real-time monitoring of synthetic reaction progress based on principal component analysis according to claim 1, characterized in that, After obtaining the low-dimensional principal component score vector, based on the dynamic change trend of the loading coefficients of different spectral wavelength variables in the synchronous multi-source data matrix, the subset of feature wavelengths with the highest sensitivity to the current stage category label is selected. The variables within the subset of characteristic wavelengths are assigned covariance weights higher than the default values, and the principal component extraction is re-executed with the weighted covariance matrix, thereby enhancing the response sensitivity of the updated low-dimensional principal component score vector to intermediate structure transformation.