Method for monitoring organic synthesis reaction based on near infrared spectrum and intelligent platform

Abstract

The application discloses a kind of organic synthesis reaction monitoring method and intelligent platform based on near infrared spectroscopy, the method continuously acquires reaction time series spectrum, extracts first principal component score in fixed or periodic renewal principal component model, forms distance sequence with the initial frame after homogenization as benchmark, and obtains time series slope by first-order least square fitting in rate observation window;After the first principal component dominance, time series fitting degree and high-order principal component fluctuation check, according to reaction order, initial concentration and metrological relationship conversion kinetics parameter, estimated conversion rate and target conversion time consumption, make the spectrum processing result act on the online display of specific reaction batch, abnormal prompt and process traceability.

Description

Technical Field

[0001] This invention relates to the field of online spectral monitoring and process analysis of organic synthesis processes, and in particular to an organic synthesis reaction monitoring method and intelligent platform that applies near-infrared time-series spectral acquisition, principal component validity verification, local rate feature extraction, and chemical kinetic conversion to a specific reaction system. Background Technology

[0002] In organic synthesis laboratories and small- to medium-batch production settings, operators typically monitor reaction progress using techniques such as thin-layer chromatography (TLC), high-performance liquid chromatography (HPLC), gas chromatography (GC), GC-MS, or nuclear magnetic resonance (NMR). Manual TLC requires periodic opening of the container for sampling and can only obtain discrete time-point information. While precise offline analysis can perform qualitative and quantitative analysis of components, the sample pretreatment and detection processes make it unsuitable for high-frequency, continuous observation throughout the entire process. For reactions requiring anhydrous, oxygen-free, or inert protection, frequent sampling increases material exposure, sample loss, and personnel operational risks.

[0003] Near-infrared spectroscopy can continuously acquire full-band signals formed by the overtone and combination absorption of hydrogen-containing groups in the reaction system using probes fixed to the reactor or bypass optical path. However, spectral band overlap, temperature and scattering perturbations are significant, making it difficult to directly correlate the original spectrum with the reaction progress. Existing schemes employ external concentration correction models, principal component trajectories, local absorbance slopes, rolling principal component monitoring, or kinetic model fitting. External correction models require corresponding standard samples and maintenance of applicable domains; relying solely on principal component trajectories or endpoint thresholds is usually insufficient for continuously outputting kinetic progress; directly treating single-point spectral differences or principal component slopes as chemical reaction rates easily overlooks the influence of scale changes caused by instrument response, optical path length, preprocessing, and latent variable coordinates.

[0004] Furthermore, the fixation and updating of principal component models are subject to technical constraints. Fixing the loads facilitates score comparability within the same model segment, but may exceed the applicability range of the original model in the later stages of the reaction or after changes in operating conditions; periodic updates can absorb new spectra, but change the loads, centering, and scale configuration, and the scores before and after the update cannot be unconditionally considered to be on the same physical scale. On the other hand, excessively short local observation windows can easily amplify bubbles, baseline jitter, and random noise into rate jumps, while excessively long windows can mask the true turning points in the reaction rate. Different reaction systems also require different pretreatment combinations, and the spectral drift of the pure solvent itself during temperature programming may be superimposed on reactant consumption and product formation.

[0005] Therefore, a complete monitoring scheme is needed, targeting the real reaction system, near-infrared acquisition link, and process conditions: first, verify whether the current spectrum can be represented by the first principal component to carry the main systematic changes of the main reaction; then, extract local spectral rates from multiple windows, and calculate and output kinetic parameters under chemical premises such as reaction order, stoichiometry, and initial concentration; at the same time, provide executable optimization or compensation paths for observation windows, preprocessing strategies, and temperature-programmed solvent drift; and stop the corresponding kinetic interpretation when the model premises are not met, instead of forcibly outputting continuous numbers as results. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides an organic synthesis reaction monitoring method and intelligent platform based on near-infrared spectroscopy. This method solves the problems of spectral changes throughout the reaction process being affected by measurement perturbations, lack of continuous verification of the dominance of the first principal component, difficulty in balancing noise suppression and time response in local rate observation windows, difficulty in selecting different pretreatment strategies, and interference of reaction signals by solvent spectral drift during temperature-programmed ascent. It also provides organic synthesis operators with continuous information on estimated conversion rate, estimated reaction time, and relative reaction rate.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for monitoring organic synthesis reactions based on near-infrared spectroscopy, which continuously acquires and stores time-series spectra of the entire reaction process, including: continuously acquiring time-series spectra of the reaction system; setting effective spectral bands, preprocessing methods, centering parameters, and scaling parameters; and preprocessing the time-series spectra using the effective spectral bands, preprocessing methods, centering parameters, and scaling parameters; performing principal component decomposition on the preprocessed time-series spectra to solve for the principal component loading matrix and principal component score matrix, wherein, after the spectra accumulate to a preset sample size or preset duration, a time-series spectrum set is constructed from the acquired preprocessed spectra and a single... Principal component decomposition is performed. Newly added spectra are not incorporated into the time-series spectral set; instead, the principal component loading matrix is ​​used to project these new spectra onto the time-series spectral set to expand the principal component score matrix. Alternatively, after accumulating spectra to a preset sample size or time period, an initial time-series spectral set is constructed and principal component decomposition is performed. Then, new spectra are incorporated into the time-series spectral set according to a preset sample threshold or time period, and principal component decomposition is performed again to update the principal component loading matrix and the principal component score matrix. Based on the principal component score matrix, the time-series first principal component score and the time-series first principal component variance contribution rate are extracted, and the model's effectiveness is determined based on the time-series first principal component variance contribution rate. The process involves performing a two-stage linearly coupled dynamic calculation, including rate feature extraction and kinetic parameter conversion. Specifically, a rate observation window with a duration shorter than the current cumulative monitoring duration is determined. The first principal component score corresponding to the spectrum at each moment within the rate observation window is retrieved. The reference distance between the first principal component score at each moment and the first principal component score of the initial frame spectrum after material homogenization is calculated. A reference distance sequence is formed from multiple reference distances. A first-order least-squares linear fit is performed with the reference distance as the ordinate and the corresponding monitoring time as the abscissa to obtain the temporal slope of the distance-time function. The goodness of fit and higher-order principal component fluctuation indices are then used to evaluate the temporal slope. The process involves verifying the time-series slope and associating the verified slope with the center time value of the corresponding rate observation window; obtaining the manually entered reaction order or pseudo-reaction order, or determining the reaction order based on the entered chemical reaction information; and, under the condition that there is a proportional relationship between the time-series slope and the reaction rate, transforming the time-series slope according to the kinetic function corresponding to the reaction order to obtain the time-series transformation quantity. A linear model of the time-series transformation quantity and the reaction time is constructed, and the kinetic constant is solved by regression. Then, combining the kinetic constant, the initial concentration, and the reaction stoichiometry relationship, the instantaneous conversion rate is output in real time, and the reaction time required to reach the target conversion rate is estimated.

[0008] Furthermore, three criteria for determining model effectiveness are set: the variance contribution rate of the first principal component is higher than a first threshold, the goodness of fit of the first principal component to the time series linearity is greater than a second threshold, and the goodness of fit of the second-order and higher principal components to the time series linearity is lower than a third threshold. When all three criteria are met, the model is deemed effective, integrating the original time series spectrum, effective preprocessing parameters, principal component scores, principal component contribution rates, kinetic calculation data, and state markers to construct a time series reaction evidence map for storing, visualizing, and tracing reaction process data. When any one criterion is not met, the model is deemed ineffective and an abnormal prompt message is output, responding to manual intervention for remodeling, or automatically executing an optimization strategy and remodeling. The optimization strategy includes adjusting the first, second, and third thresholds, and adjusting any one or more of the effective spectral bands, preprocessing methods, centering parameters, scaling parameters, preset sample size, preset duration, preset sample threshold, time period, and rate observation window. The thresholds are input through a human-computer interaction interface or generated by the system within a preset value range. The method also includes a reaction trajectory deviation early warning mechanism, which calculates the first principal component variance contribution rate within each rate observation window in real time and uses the fluctuation range obtained from multiple repeated measurements as the early warning threshold. When the first principal component variance contribution rate is lower than the early warning threshold for multiple consecutive frames, the reaction system is determined to be abnormal, and an abnormal prompt and interactive signal are output through the human-computer interaction interface, supporting message push or sending control commands to external related systems.

[0009] Furthermore, an adaptive optimization method is used to determine the length of the rate observation window, with a minimum window length of no less than three data points. Different window widths are traversed within a preset range. For each window, the instantaneous response rate is extracted by approximating it with a local window secant. After nonlinear transformation of the kinetic equation, a linear fit is performed to obtain the kinetic goodness of fit, and the variance contribution rate of the first principal component is checked to see if it reaches a threshold. A set of qualified candidate windows is obtained by screening, and the variation law of kinetic goodness of fit with window length is analyzed to determine a stable interval with small fluctuations in kinetic goodness of fit. The optimal observation window is selected within the stable interval. The optimization can be re-executed during system operation to dynamically adjust the window length.

[0010] Furthermore, an adaptive optimization of the spectral preprocessing strategy is performed, traversing multiple spectral preprocessing algorithms and their combinations to form candidate schemes. Principal component decomposition is then performed on the spectra processed by each candidate scheme, calculating the variance contribution rate of the first principal component and the goodness of fit of the kinetics. A preset threshold for the variance contribution rate of the first principal component is used as an admission criterion, retaining only candidate schemes that meet this threshold. Among the qualified candidate schemes, the optimal preprocessing scheme is selected based on the goodness of fit of the kinetics, and a self-learning database is constructed to store the correspondence between the reaction system, process conditions, and the optimal preprocessing strategy. The preprocessing algorithms include baseline correction, multivariate scattering correction, standard normal variable transformation, mean centering, vector normalization, Savitzky-Golay smoothing, sliding window smoothing, Gaussian smoothing, first derivative processing, second derivative processing, wavelet transform denoising, and combinations of the above methods.

[0011] Furthermore, blank solvent spectral compensation and correction were implemented for the temperature-programmed reaction process. Using a pure solvent consistent with the reaction system, time-series blank solvent spectra were collected according to the same temperature-programmed process as the formal monitoring, and a mapping relationship between temperature and solvent spectral drift was established. After collecting the measured spectra of the reaction system, baseline compensation was completed based on the aforementioned mapping relationship. The corrected spectral data were used to carry out subsequent principal component decomposition, rate extraction, and kinetic solution to suppress solvent spectral disturbances caused by temperature changes.

[0012] Furthermore, when the real-time conversion rate obtained from kinetic calculations reaches a preset conversion rate threshold, the reaction is determined to have reached its endpoint; the conversion rate threshold is set manually or automatically estimated by the system.

[0013] Furthermore, after a single reaction is completed, all the original spectra of this reaction are retrieved, the preprocessing parameters are reused or reconfigured, and principal component decomposition and kinetic fitting calculations are carried out again to achieve iterative comparison and review of multiple sets of parameters; the resulting review spectrum is only used for parameter verification and result optimization.

[0014] Furthermore, the monitoring method is adapted to non-one-time feeding conditions such as sequential dripping, continuous feeding, and intermittent feeding; it conducts time-series spectral acquisition, data processing, and modeling analysis during the initial feeding stage, and continuously applies the model parameters and kinetic characteristic data obtained during the initial feeding stage to subsequent reaction stages, thereby realizing dynamic monitoring and reaction process prediction throughout the non-one-time feeding process.

[0015] Furthermore, the monitoring method relies on an intelligent platform for monitoring organic synthesis reactions. This platform includes a reaction module, a near-infrared sampling module, a feed / discharge module, an inert atmosphere control module, a constant temperature control module, a central control unit, and a human-machine interface. The inert atmosphere control module is used to create an inert protective environment that isolates the reaction system from air. The constant temperature control module is used to regulate the temperature of the reaction system. The feed / discharge module is used to automate sample feeding and discharge control. The near-infrared sampling module is used to continuously collect time-series spectral data throughout the entire reaction process. The central control unit is used to execute spectral preprocessing, principal component decomposition, and kinetic calculation processes. The human-machine interface is used for parameter input, data display, and abnormal interaction control.

[0016] Secondly, this invention provides an intelligent platform for monitoring organic synthesis reactions, used to execute the aforementioned near-infrared spectroscopy-based organic synthesis reaction monitoring method, including a human-computer interaction interface; the human-computer interaction interface is used to display one or more of the following: the instantaneous conversion rate output in real time by the monitoring method, the reaction time estimated by the monitoring method to reach the target conversion rate, and the time-series relative reaction rate obtained by comparing the time-series slope at each time step with the first time-series slope. Compared with the prior art, this invention has the following beneficial effects: First, this invention addresses the nonlinear characteristics of the kinetic evolution of organic synthesis reactions by using the secant approximation of a local observation window combined with kinetic function transformation to transform the complex nonlinear kinetic problem into a linear regression problem. This achieves linearization of the nonlinear kinetic problem, reduces the difficulty of online real-time calculation, and adapts to the continuous computing needs of automated monitoring platforms.

[0017] Second, this invention constructs a multi-dimensional joint model effectiveness evaluation system, combining the variance contribution rate of the first principal component, the time-series goodness of fit of the first principal component, and the fluctuation index of higher-order principal components as the basis for model judgment. Relying on the applicable conditions of the first principal component carrying the evolutionary information of the reaction entity, it continuously tracks the state changes of the liquid-phase reaction system, achieving a quantitative evaluation of the reliability of kinetic modeling and improving the credibility of monitoring results under complex reaction systems. Unlike methods that rely solely on spectral signal-to-noise ratio for parameter optimization, this invention uses the kinetic linear fitting effect as the parameter optimization target, ensuring that the optimization results of spectral preprocessing and rate observation windows serve kinetic modeling.

[0018] Third, this invention integrates adaptive optimization of the rate observation window, adaptive optimization of the spectral preprocessing strategy, and spectral compensation of the blank solvent during programmed temperature rise, thereby suppressing noise, curvature errors caused by local secant approximation, and spectral drift caused by temperature in multiple dimensions to stably extract reaction kinetic characteristics.

[0019] Fourth, this invention uses the time-series signal of the first principal component as the core carrier for kinetic modeling, applicable to first-order and second-order reaction monitoring scenarios where the first principal component can carry the spectral evolution information of the main reaction. Since pharmaceutical small molecule and fine chemical research often employs liquid-phase reaction systems, liquid-phase reactions are simple to operate and conditions are easy to control. However, the solvent cage effect in the liquid-phase environment significantly reduces the probability of simultaneous effective collisions of three molecules, making third-order and higher-order elementary reactions extremely rare. Organic reaction elementary processes are mainly monomolecular and bimolecular transformations, making first-order or pseudo-first-order, second-order or pseudo-second-order kinetic models widely applicable; even autocatalytic systems can still be modeled based on combinations of the aforementioned elementary processes. If the system contains multiple independent reaction processes without stable coupling, causing the first principal component to fail to dominate spectral changes, the model is deemed invalid and outputs a warning to assist researchers in checking the reaction system and adjusting the synthesis process. Generally, researchers prioritize synthetic routes with simple reaction rules and easy control during the research process, avoiding reactions with complex kinetic behavior as much as possible. Therefore, this type of kinetic model is sufficient to support the kinetic analysis of most pharmaceutical small molecule liquid-phase synthesis processes.

[0020] This method can continuously monitor the reaction process and can be cross-validated with offline qualitative and quantitative methods such as high performance liquid chromatography, mass spectrometry and nuclear magnetic resonance. It can also reduce the frequency of offline sampling and the workload of thin-layer chromatography plate spotting, and improve the efficiency of synthesis process development and production monitoring. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the process for monitoring organic synthesis reactions based on near-infrared spectroscopy according to the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings. It should be understood that the described embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0023] In one implementation scenario, the reaction module is a liquid-phase reactor with stirring and temperature control. The near-infrared sampling module uses an immersion fiber optic probe or a bypass flow cell with a fixed optical path and continuous circulation with the reactor. The probe is connected to a near-infrared spectrometer and an acquisition controller. The acquisition controller adds a timestamp to each frame of spectrum and records the effective wavelength, integration time, averaging count, probe position, or bypass status. The inert atmosphere control module maintains the required nitrogen or other inert atmosphere through the reactor's gas inlet; the temperature control module regulates the reaction temperature through a jacket, heater, or cooling circuit; and the feed / discharge module records the start, end, and cumulative feed information. These modules provide status information to the central control unit via analog, digital, serial communication, or industrial Ethernet interfaces, without being limited to specific equipment models. The central control unit organizes data by specific reaction batches, receives time-series spectra, temperatures, and feed / discharge status, and performs spectral preprocessing, principal component decomposition, model validity assessment, rate feature extraction, and kinetic parameter conversion. The calculation results are sent to a human-machine interface for online display, anomaly alerts, and process traceability.

[0024] Combination Figure 1 In one embodiment, after the near-infrared sampling module is activated, the central control unit continuously acquires and stores the original time-series spectrum and acquisition status of the reaction system. Effective bands are used to eliminate bands with spectral saturation, no effective response, or excessive noise. Preprocessing methods are selected from baseline correction, multivariate scattering correction, standard normal variable transformation, mean centering, vector normalization, Savitzky-Golay smoothing, sliding window smoothing, Gaussian smoothing, first derivative, second derivative, and wavelet transform denoising, as well as combinations thereof. The centering parameter and scaling parameter are saved along with the preprocessing configuration. For the same real-time calculation segment, unless remodeling is triggered, the central control unit maintains consistency in the effective spectral bands, preprocessing methods, centering parameters, and scaling parameters, ensuring that the subsequently formed principal component scores and time-series slopes have a traceable and comparable computational basis.

[0025] In the case of single-feed operation, an initial frame is selected after the material is added and stirred to achieve a homogeneous state, or an initial reference spectrum is formed by averaging multiple frames near the homogenization moment. In the case of continuous feeding operation, an initial reference is selected in the stable observation interval during the first feeding stage, and the corresponding feeding status, cumulative feeding amount, and reaction temperature are recorded. The initial reference is used to establish the first principal component reference distance; whether to continue into the kinetic conversion is determined by the subsequent model validity conditions.

[0026] When performing principal component decomposition on the preprocessed time-series spectra, either a fixed projection method or a periodic update method is adopted. In the fixed projection method, after the spectra have accumulated to a preset sample size or a preset duration, a time-series spectral set is constructed using the acquired preprocessed spectra and principal component decomposition is performed once. Subsequent newly added spectra are not incorporated into the time-series spectral set, but are projected using the same centering parameter, scaling parameter, and principal component loading matrix to expand the principal component score matrix.

[0027] The periodic update method first establishes an initial principal component model. Then, according to a preset sample threshold or time period, newly added spectra are incorporated into the time-series spectral set and re-decomposed to update the principal component loading matrix and principal component score matrix. During the update, the spectral range currently involved in the decomposition, preprocessing configuration, centering parameters, scaling parameters, principal component loading matrix, and principal component score matrix are saved simultaneously. The first principal component scores before and after the update are generated according to their respective models and are not directly spliced ​​into the same slope sequence without alignment or re-benchmarking. When continuous display is required, the benchmark distance of the corresponding segment is recalculated using the updated homogenized reference frame, and the model update boundary is recorded in the time-series response evidence map.

[0028] The central control unit extracts the temporal first principal component score and the temporal first principal component variance contribution rate from the principal component score matrix, and determines whether the current spectral segment is suitable to continue to be interpreted as a temporal change dominated by the progress of a single main reaction based on the temporal first principal component variance contribution rate. Within the current spectral segment, if the systematic spectral evolution caused by the main reaction constitutes the main variance direction and no independent significant temporal changes occur in the higher-order principal components, the first principal component score can serve as the basic data for subsequent rate feature extraction. Once the model validity assessment indicates that the first principal component cannot bear the main evolution of the current main reaction, the central control unit stops the two-level linear coupling dynamics calculation for that segment and prompts for reselection of the band, preprocessing, observation segment, or monitoring model.

[0029] During rate feature extraction, the central control unit determines a rate observation window with a duration shorter than the current cumulative monitoring duration. It then retrieves the first principal component scores corresponding to the spectra at each time point within this window and calculates the baseline distance between the first principal component scores at each time point and the first principal component scores of the initial frame spectrum after material homogenization. Multiple baseline distances form a baseline distance sequence. The central control unit performs a first-order least-squares linear fit with the baseline distance as the ordinate and the corresponding monitoring time as the abscissa to obtain the temporal slope of the distance-time function. This is then validated using goodness-of-fit and higher-order principal component fluctuation indices. The validated temporal slope is then correlated with the center time value of the corresponding rate observation window. Thus, each temporal slope can be traced back to a specific window, original spectrum, preprocessing configuration, and model version, rather than being an isolated value detached from the acquisition conditions.

[0030] When performing kinetic parameter conversion, the reaction order is entered by the user based on known reaction mechanisms or process data, or the system provides a pseudo-reaction order based on the entered chemical reaction information and records the judgment criteria. Under the condition that there is a proportional relationship between the time-series slope and the corresponding reaction rate, the central control unit performs a function transformation on the time-series slope based on the kinetic function corresponding to the reaction order, obtaining the time-series transformation quantity; then, a linear model of the time-series transformation quantity and reaction time is constructed, and the kinetic constant is solved through regression. Combined with the initial concentration and reaction stoichiometry, the instantaneous conversion rate is output in real time, and the reaction time required to reach the target conversion rate is estimated. After changing the optical path, preprocessing scale, or principal component model, the proportional relationship is reconfirmed; when independent scale confirmation data is not obtained, the corresponding output is marked as the estimated conversion rate or apparent kinetic parameter, and the time-series relative reaction rate can be output simultaneously.

[0031] Within a fixed optical path, acquisition configuration, and applicable physical property range, near-infrared absorbance can be expressed as the superposition of the concentration contributions of each representative component in the reaction system and baseline perturbations: ; In the formula, wavelength and time The corresponding absorbance, For effective optical path, The number of representative components, For the first The effective absorption response function of a representative component For the first Representative components at time concentration, For the baseline, scattering, and unmodeled perturbation terms, this equation indicates that reactant consumption and product formation can form systematic spectral components that vary with the progress of the reaction across the entire wavelength range, but it does not presuppose that the first principal component of any reaction is necessarily equivalent to the chemical conversion rate.

[0032] Principal component decomposition was performed on the preprocessed time-series spectral matrix: ; In the formula, This is the preprocessed time-series spectral matrix. Principal component score matrix, Principal component loading matrix, The residual matrix is... Indicates matrix transpose. Fixed projection method locked. It includes its associated centralized and scaled configurations; the periodic update method re-obtains the matrix after the update conditions are met, and calculates the score for the corresponding time period using the updated matrix.

[0033] Let the single-frame preprocessed spectral vector be... The centered vector is The scale vector is The first principal component loading vector is Then the score of the first principal component can be expressed as: ; In the formula, For a moment The corresponding first principal component score, This is the preprocessed spectral vector at that moment. For a centered vector, The scale vector This indicates element-wise division. The first principal component loading vector, with superscript This represents the transpose of a vector.

[0034] The score of the first principal component of the spectrum of the initial frame after normalization Define the reference distance as the reference: ; In the formula, For a moment The corresponding baseline distance, For a moment The corresponding first principal component score, The score represents the first principal component of the initial frame spectrum after homogenization. The reference distance is the relative spectral change in the latent variable coordinates and does not directly have chemical concentration units.

[0035] For the center time is And includes The time-series slope is obtained by performing a first-order least-squares fitting on the rate observation window of 10 effective data points: ; In the formula, Center moment The corresponding time series slope, This represents the number of valid data points within the window. For the first Monitoring time for each data point This represents the average monitoring time within the window. For the first The baseline distance of each data point The average distance within the window is the baseline distance. The time-series slope represents the spectral rate characteristic, and its proportional relationship with the chemical reaction rate must hold true under the corresponding reaction, optical path, and pretreatment conditions.

[0036] The proportion of the projection variance of the current rate observation window along the direction of the first principal component can be expressed as: ; In the formula, For window The proportion of the projection variance of the first principal component within the matrix. The score for the first principal component. For the first Principal component scores, The number of principal components participating in the validation. This is the calculation of the variance of samples within the window. For periodic update mode, the proportion is calculated according to the current principal component model version; for fixed projection mode, the proportion is the proportion of the projection variance of the current window data in the existing principal component directions.

[0037] The three validity conditions of the model can be expressed as follows: ; In the formula, This represents the variance contribution rate of the first principal component or the proportion of the variance of the corresponding window projection. The first threshold, The goodness of fit of the first principal component to the time series linearity is given. The second threshold, For the first The goodness of fit of the time series linearity of the higher-order principal components. The third threshold is used. Only when all three conditions are met simultaneously will the first principal component be interpreted as the main reaction spectral evolution carrier within the currently applicable segment; if any one condition is not met, the kinetic conversion of that segment will be paused.

[0038] The goodness of linear fit is calculated using the coefficient of determination: ; In the formula, As the coefficient of determination, For the number of regression samples, For the first One observation value, For the first One fitted value, This is the mean of the observed values. This index is used for window slope verification, dynamic fitting evaluation, and comparison of candidate windows or preprocessing schemes.

[0039] For a first-order irreversible reaction, under the condition that the absolute value of the time slope remains in a constant proportion to the instantaneous reaction rate, a linear transformation quantity can be constructed: ; In the formula, This is a first-order dynamic time-series transformation quantity. The time series slope, For the regression intercept, This is a first-order apparent decay parameter. This refers to the reaction time. When the spectral scale has already been determined by independent data, it can be further... Convert to the corresponding kinetic constants; when not independently determined, the output value is the apparent kinetic parameter.

[0040] For an equimolar second-order irreversible reaction, under the condition that the absolute value of the time slope remains in a constant proportion to the square of the instantaneous concentration, a linear transformation quantity can be constructed: ; In the formula, For equimolar second-order dynamic time-series transformation quantities, The time series slope, For the regression intercept, These are second-order apparent parameters. The reaction time is given. The transformation avoids directly equating the reciprocal of the rate characteristic to an equimolar second-order integral; the determination of the intrinsic kinetic constant is still constrained by the spectral scale and initial concentration conditions.

[0041] The general definition of conversion rate is: ; In the formula, The conversion rate is based on the target reactant. This represents the initial concentration of the reactant at the start of the kinetic calculation section. For a moment The corresponding instantaneous residual concentration. In the absence of offline concentration or independent scale correction data, the interface output should be labeled as an estimated conversion rate to distinguish it from the precise chemical quantification results obtained by chromatography or NMR.

[0042] The time required for the first-order reaction to reach the target conversion rate can be estimated using the following formula: ; In the formula, To achieve the projected time to reach the target conversion rate, This represents the starting time of the dynamic calculation section. For the target conversion rate, These are the first-order kinetic constants or apparent constants obtained through effective regression and necessary scaling. The estimated times are output only if the model validity is consistently satisfied and the extrapolation range is reasonable.

[0043] The time required for an equimolar second-order reaction to reach the target conversion rate can be estimated using the following formula: ; In the formula, To achieve the projected time to reach the target conversion rate, This represents the starting time of the dynamic calculation section. For the target conversion rate, These are the second-order kinetic constants or apparent constants obtained through effective regression and necessary scaling. The initial concentration at the starting point of the segment.

[0044] It is worth noting that for non-equimolar second-order reaction systems, the real-time concentration of reactants at each moment can be deduced based on the fixed consumption ratio of reactants and the kinetic laws reflected by the changes in the temporal slope. By relying on the correspondence between temporal concentration changes and reactant consumption, a linear transformation equation for the kinetic variables can be constructed, transforming the original nonlinear kinetic model into a solvable linear expression. Then, the kinetic parameters can be fitted, the rate constant can be solved, and the reaction progress can be predicted using linear regression.

[0045] For typical bimolecular autocatalytic reactions, the products act as catalysts to drive the reaction forward. The time-series slope characterizing the reaction rate exhibits a single-peak pattern over time, showing a slow initial start, continuous acceleration, and gradual slowdown in the later stages. The complete kinetic evolution can be characterized by the time-series slope obtained through real-time monitoring, and the kinetic constants can be solved and the reaction progress predicted by combining the half-peak width characteristics of the curve.

[0046] For more complex reaction systems with higher order and multi-component coupling, this platform can achieve dynamic monitoring of the entire process and display of reaction curves. Such systems are currently difficult to fit linearly analytically through simple variable transformations; in practice, nonlinear numerical fitting methods can be used for kinetic analysis, parameter solving, and reaction process prediction.

[0047] The blank solvent spectral compensation under programmed temperature ramping conditions can be expressed as: ; In the formula, For a moment The compensated reaction spectrum, The measured spectrum of the reaction system is shown below. This is a compensation coefficient set based on optical path consistency or solvent ratio. For temperature The corresponding blank solvent spectrum, Reference temperature The corresponding blank solvent spectrum. Ideally, a blank mapping should be established under the same solvent, probe position, effective optical path, acquisition parameters, and temperature program; this compensation version should be discontinued if the conditions are inconsistent or the residual increases after compensation.

[0048] The complete process described above uses the time-series spectra of the same reaction batch as the main thread, and the acquisition configuration and process status as the calculation boundaries. Principal component scores, baseline distances, time-series slopes, time-series transformation quantities, kinetic constants, instantaneous conversion rates, and target conversion times are passed sequentially. The dominance of the first principal component and the time-series slope verification together define the applicable range of the kinetic interpretation, ensuring that online display results are generated only when the corresponding spectral status and kinetic premises are met. When the model premises are not met, the corresponding output stops and the original data is retained. Therefore, it is possible to continuously observe the reaction process while retaining evidence for anomaly checks and result traceability.

[0049] Furthermore, in the above embodiments, the model validity judgment is performed after extracting the first principal component score and the first principal component variance contribution rate, and these are used as the admission criteria for the rate feature in the kinetic conversion. The central control unit simultaneously checks that the first principal component variance contribution rate is higher than a first threshold, the first principal component temporal linear fit goodness of fit is greater than a second threshold, and the second-order and above principal component temporal linear fit goodness of fit is lower than a third threshold. When all three conditions are met, the model is deemed effective, and the original time-series spectrum, effective preprocessing parameters, principal component scores, principal component contribution rates, window start and end times, center time values, temporal slopes, kinetic calculation data, and state markers are organized into a time-series response evidence map. The time-series response evidence map is implemented using a relational database, a time-series database, or a data file with associated fields, so that the estimated conversion rate, estimated reaction time, or relative reaction rate displayed on the human-computer interaction interface can be traced back to the corresponding original spectrum and calculation parameters.

[0050] If any judgment condition is not met, the central control unit marks the current model as ineffective, stops the dynamic parameter conversion of that segment, outputs an anomaly prompt, and saves the failure type and the spectral and model states before and after the failure. In response to manual intervention, the thresholds or modeling parameters are reset and the model is rebuilt; during automatic optimization, the first, second, and third thresholds are adjusted, or one or more of the effective spectral bands, preprocessing methods, centering parameters, scaling parameters, preset sample size, preset duration, preset sample threshold, time period, and rate observation window are adjusted. Thresholds are input through the human-computer interface or generated by the system within a preset range; each automatic adjustment records the adjusted object, values ​​before and after adjustment, evaluation results, and status flags to prevent abnormal spectra from entering the dynamic conversion by unboundedly lowering the entry conditions.

[0051] The reaction trajectory deviation early warning uses the variance contribution rate of the first principal component within the current rate observation window, and uses the fluctuation range obtained from repeated measurements under the same optical path and similar process conditions as the early warning threshold. When the variance contribution rate of the first principal component is lower than the early warning threshold for multiple consecutive frames, the central control unit determines that the reaction system is abnormal, outputs an abnormal prompt of main reaction trajectory deviation or increased interference through the human-machine interface, and generates an interactive signal for message push or issuing control commands to external related systems. In the absence of chromatographic, mass spectrometric, or nuclear magnetic resonance evidence, specific side reactions or byproducts are not named separately based on the aforementioned indicators. This further scheme controls the kinetic conversion entry through three conditions and writes the failure, handling, and early warning processes into the evidence spectrum, so that the abnormal window is not mixed with the valid window for interpretation, while retaining traceable evidence for manual inspection, remodeling, and external interaction.

[0052] Furthermore, in the above embodiments, adaptive optimization of the rate observation window is performed before rate feature extraction. The central control unit first delineates an optimization segment containing only spectra acquired before the current time, and uses at least three data points with non-overlapping acquisition times as the minimum window length. Within a preset range, it traverses different window widths according to the number of data points, the actual duration, or a combination of both. For each candidate window, the central control unit uses the same spectral segment, the same reaction order, and the same evaluation rules, and extracts the instantaneous reaction rate using local window secant approximation or first-order least squares fitting. After kinetic function transformation, linear fitting is performed to obtain the kinetic fit goodness of fit, and the variance contribution rate of the first principal component and the number of effective data points of the corresponding window are checked. Candidate windows with insufficient effective points, zero time span, failure of the first principal component admission, or fitting residuals exceeding the allowable range are not included in the qualified candidate window set.

[0053] The central control unit determines a stable interval with minimal fluctuations in dynamic fit goodness based on the difference in dynamic fit goodness between adjacent candidate window widths, the range of dynamic fit goodness within a preset length range, or the degree of local variation. The parameters for determining the stable interval are derived from equipment noise records, acquisition frequency, and the process-allowed response time. An optimal observation window is selected from within the stable interval according to pre-saved selection rules and sent to the rate feature extraction process to form the subsequent baseline distance sequence and time-series slope. When the temperature program changes, a feeding event occurs, the sampling frequency changes, or the dynamic fit goodness of multiple consecutive windows decreases, the optimization process is re-executed to dynamically adjust the window length. This further scheme allows window selection to simultaneously undergo first principal component admission and dynamic fit evaluation, reducing the impact of excessively short windows amplifying single-frame perturbations and excessively long windows masking rate transitions, and enabling the selected window to be directly used for subsequent rate feature extraction.

[0054] Furthermore, in the above embodiments, the adaptive optimization of the spectral preprocessing strategy is performed before the original time-series spectrum enters principal component decomposition. The central control unit establishes a candidate preprocessing pool, traversing single algorithms and combinations of algorithms with a defined execution order among baseline correction, multivariate scattering correction, standard normal variable transformation, mean centering, vector normalization, Savitzky-Golay smoothing, sliding window smoothing, Gaussian smoothing, first derivative processing, second derivative processing, and wavelet transform denoising. The candidate parameters for wavelet transform denoising include wavelet basis type, number of decomposition levels, threshold rules, and reconstruction method; the candidate parameters for Savitzky-Golay smoothing include window width and polynomial order; the candidate parameters for derivative processing include derivative order and matching smoothing parameters. Each parameter is selected from a preset allowed set and stored together with the corresponding candidate scheme.

[0055] The central control unit processes each candidate preprocessing scheme using the same batch of original spectra, the same modeling segment, the same reaction order, and the same evaluation rules. Principal component decomposition (PCD) is performed on each group of preprocessed spectra, and the variance contribution rate of the first principal component and the kinetic goodness of fit are calculated. The variance contribution rate of the first principal component serves as an admission threshold; candidate schemes that do not meet the threshold are eliminated. Among the qualified candidate schemes, the optimal preprocessing scheme is selected based on the kinetic goodness of fit, and this optimal preprocessing scheme is used to process the time-series spectra subsequently entering PCD. A self-learning database stores reaction system identifiers, solvents, temperature ranges, feed methods, probe optical paths, acquisition configurations, candidate schemes, evaluation results, final preprocessing configurations, version numbers, and applicable statuses, enabling subsequent batches to call upon historical schemes and re-execute model validity verification. This further approach ensures that the selection of a preprocessing scheme depends not only on the smoothing or noise reduction results of the spectrum itself but also on its ability to generate time-series data that meets the first principal component admission criteria and serves kinetic linear fitting, thus providing a consistent processing configuration for subsequent PCD and rate extraction.

[0056] Furthermore, in the programmed temperature-increase reaction embodiment, blank solvent spectral compensation is performed before the original time-series spectrum enters the preprocessing process. Using a pure solvent consistent with the reaction system, and without adding reactants or catalysts, the time-series blank solvent spectrum is acquired according to the same temperature-increase program as the formal monitoring, using the same reactor or equivalent optical path, the same probe installation position, the same effective optical path length, integration time, sampling frequency, and stirring method. Temperature and time are recorded simultaneously, establishing a mapping relationship between temperature and the full-band blank solvent spectral drift. During formal reaction monitoring, adjacent blank temperature points are selected based on the measured temperature, and interpolation is performed wavelength by wavelength. When thermal hysteresis exists, the heating direction and time are used as additional indexes to distinguish between the heating path and the cooling path.

[0057] The central control unit uses spectral subtraction based on the difference between the blank drift and the reference temperature, or weighted compensation using a proportionality coefficient determined by the blank segment, to generate a compensated reaction spectrum. This compensated reaction spectrum is then preprocessed, undergoes principal component decomposition, rate feature extraction, and kinetic parameter conversion sequentially. After compensation, the baseline residual, the variance contribution rate of the first principal component, and fluctuations in higher-order principal components are checked. If the solvent composition, optical path, heating rate, or phase changes, or if higher-order interference is enhanced after compensation, the corresponding blank mapping is marked as inapplicable and discontinued. A further scheme compensates for solvent spectral drift caused by heating before it enters the principal component model, making subsequent temporal changes in the first principal component more focused on the systematic changes within the reaction system itself. The validity of the compensated model is then used to check whether the mapping is still usable.

[0058] Furthermore, in the above embodiments, the reaction endpoint determination is performed after the instantaneous conversion rate is formed. The conversion rate threshold is manually set by the human-machine interface or automatically estimated by the system based on the target conversion range of historical qualified batches; when the real-time conversion rate obtained from kinetic calculation reaches the conversion rate threshold, the central control unit determines that the reaction has reached the endpoint and forms an endpoint state. The endpoint state is displayed as a prompt and message push on the human-machine interface, or sent as a status signal to an external control system with process safety interlocks. The endpoint determination uses the real-time conversion rate formed when the model is in a valid state as input, thus converting continuous kinetic output into a clear process state and avoiding triggering endpoint prompts based on invalid conversion rates in the model failure section.

[0059] Furthermore, in the embodiment following a single reaction, the offline reconstruction debriefing optimization of the projection space is performed after the completion of the real-time monitoring process for this batch. The central control unit retrieves all the original spectra of this reaction, reuses the preprocessing parameters used in real-time monitoring, or reconfigures the preprocessing parameters, and then re-performs principal component decomposition and kinetic fitting calculations to form iterative comparative debriefing results corresponding to multiple sets of parameters. The debriefing spectra are only used for parameter verification and result optimization, and are saved separately from the conversion rate, abnormal state, and time-series reaction evidence spectra formed by real-time monitoring, without overwriting the real-time results. When the candidate parameters obtained from the debriefing enter the configuration reserve for the next batch, they still need to undergo a re-evaluation of the three model validity criteria in the next batch. This further scheme utilizes the original spectra of the entire batch to verify different processing and modeling configurations, while preserving the temporal authenticity and the integrity of the evidence chain of the real-time calculation results.

[0060] Furthermore, in embodiments involving sequential dripping, continuous feeding, or intermittent feeding, after the initial model and kinetic characteristics are formed during the first feeding stage, non-one-time feeding monitoring continues. During the first feeding stage, time-series spectral acquisition, preprocessing, principal component decomposition, and kinetic calculations are performed. The resulting model parameters and kinetic characteristic data are continuously applied to subsequent reaction stages to form dynamic monitoring and reaction progress prediction for the entire non-one-time feeding process. Subsequent stages continue to perform first principal component variance contribution rate, first principal component time-series goodness of fit, and higher-order principal component fluctuation verification to determine whether the model formed during the first feeding stage is still within its applicable range.

[0061] When the feed causes changes in system volume and concentration, the central control unit updates the initial concentration and reaction stoichiometric relationship based on the cumulative feed amount, feed components, and reactor volume, and then performs kinetic parameter conversion. If the material information is unavailable, only the relative spectral change, time-series relative reaction rate, and model validity status are output, without directly using the constant volume relationship from the initial feed stage as the quantitative basis for subsequent stages. When the model fails, it enters a manual handling or automatic optimization process for remodeling. This further approach ensures that the model formed in the initial stage is subject to the joint constraints of feed status and continuous validity verification, avoiding the unconditional continuation of the original quantitative conversion when the material relationships in the system have changed.

[0062] Furthermore, in the above embodiments, the monitoring method is executed based on an intelligent platform for monitoring organic synthesis reactions. The reaction module carries the specific organic synthesis reaction; the near-infrared sampling module continuously collects time-series spectral data of the entire reaction process and outputs the raw spectrum and acquisition timestamp to the central control unit; the feed and discharge module completes automated sample feeding and discharge control and outputs the sample feeding, dropping, or discharge status; the inert atmosphere control module isolates the reaction system from air and maintains an inert protective environment; the constant temperature control module regulates the temperature of the reaction system and outputs the temperature status; the central control unit receives the spectral data and platform status, and performs spectral preprocessing, principal component decomposition, model validity judgment, rate feature extraction, kinetic parameter conversion, evidence spectrum archiving, and corresponding optimization processes; the human-machine interface receives the reaction order, initial concentration, reaction stoichiometry, threshold, and parameter range, and is used for parameter input, data display, and abnormal interaction control.

[0063] Each module transmits data according to the monitoring process described above: the near-infrared sampling module provides the time-series spectrum to be processed; the temperature control module and the feed / discharge module provide the operating conditions for defining the spectral range and kinetic conversion; the central control unit generates the estimated conversion rate, estimated reaction time, time-series relative reaction rate, model validity status, and anomaly information; and the human-machine interface associates and displays these results with the reaction batch, acquisition timestamp, and time-series reaction evidence spectral records. This further scheme implements the abstract data processing steps into the actual reactor, near-infrared sampling link, and on-site control status, and enables the monitoring results to be used for online display, anomaly alerts, and process traceability for specific batches.

[0064] In one common implementation, the original spectrum is first compensated according to a programmed temperature-increasing blank mapping, then the preprocessing strategy adaptively optimizes and selects the processing configuration. Subsequently, a principal component model is constructed and three model validity judgments are performed. Only segments that meet the admission criteria enter the rate observation window for adaptive optimization and two-level linearly coupled kinetic calculations. The above processes form a progressive and complementary relationship, where blank compensation addresses solvent drift caused by temperature, preprocessing optimizes scattering, baseline, and noise, model validity judgment controls the kinetic interpretation entry point, and window optimization balances noise suppression and time response within the effective segment. Each process acts on different stages of the monitoring process and together defines the applicable state of the final time-series slope and kinetic output.

[0065] Reaction endpoint determination, offline replay, and non-disposable feed monitoring can also be implemented in conjunction with the above monitoring process. When the effective model continuously outputs instantaneous conversion rate, the instantaneous conversion rate is compared with the conversion rate threshold to determine the endpoint status; after the batch ends, all saved original spectra are used for offline replay without overwriting real-time results; in non-disposable feed conditions, the material relationships are updated according to the feed status and the applicability of the model is continuously verified. These three processes address online endpoint prompts, post-batch parameter verification, and changes in feed conditions, respectively, forming progressive support around the lifecycle of the same reaction batch.

[0066] In one embodiment of an intelligent platform for monitoring organic synthesis reactions, the platform includes the aforementioned human-machine interface. This interface receives one or more of the following from the central control unit: estimated conversion rate, estimated reaction time, and time-series relative reaction rate. The estimated conversion rate and estimated reaction time are calculated from kinetic parameters, and the time-series relative reaction rate is obtained by comparing the time-series slope at each time step with the first time-series slope. The interface simultaneously reads the reaction batch identifier, acquisition timestamp, model version, and model validity status, associating each displayed value with the corresponding window results in the time-series reaction evidence graph. When the model is not effective or the trajectory deviates, the interface marks the status of the corresponding displayed value or stops updating, and outputs abnormal interaction information. Therefore, the interface does not merely display calculated figures detached from the source, but rather integrates continuous monitoring results, model validity status, and platform data links into the display and interaction process of specific reaction batches.

[0067] The core technology of this invention is to acquire the time-series spectrum of a specific organic synthesis system using a near-infrared sampling module fixed in the actual reactor or bypass optical path, and to limit the applicability of the kinetic interpretation by reaction temperature, feed state, optical path conditions, model validity, and stoichiometry. The output results are used for online display of specific batches, anomaly alerts, sampling arrangements, operational interaction, and post-batch review. When the dominance of the first principal component or other model validity conditions are not met, the system exits the corresponding kinetic interpretation path and retains the evidence data, thereby forming a closed-loop relationship between the measured object, spectral signal, chemical premise, and platform action.

[0068] Although determining the kinetic constants of chemical reactions is not the core objective of this method, the prediction of reaction progress depends on solving the kinetic parameters during the reaction process. The accuracy of the calculated kinetic parameters directly determines the accuracy of the reaction progress prediction results. To examine the reliability and validity of this method and to verify its calculation accuracy, this invention selected several typical reactions with publicly reported kinetic constants in the literature for comparative verification experiments. The specific experimental information is shown in the table below.

[0069] Table 1 The formula for calculating the relative deviation is as follows: .

[0070] The experimental results in the table above show that the relative errors between the kinetic constants calculated using this method and the theoretical values ​​reported in the literature are all less than 5%. Most conventional methods for accurately determining kinetic parameters rely on sophisticated analytical techniques such as offline chromatographic external standard methods. While these methods offer excellent detection accuracy, they require periodic sampling and cumbersome sample pretreatment, making continuous real-time monitoring difficult. In contrast, this method requires neither external nor internal standards, eliminates the need for offline sampling and analysis, provides a rapid system response, and enables in-situ, continuous, real-time acquisition of information following the reaction system.

[0071] Further experiments show that, compared with the unbiased fitting system of the whole spectrum, the algorithm first performs principal component decomposition on the original spectrum and then carries out the second-stage linear fitting, which has lower computational cost and smaller deviation in output results; compared with the method of fitting time series curves based on fixed wavelength, the calculation accuracy of this method is significantly higher.

[0072] Considering the significant advantages of this method—its convenience, speed, and real-time online operation—while maintaining a low level of error in kinetic parameter calculations, it is sufficient to demonstrate that this method possesses good accuracy. This method has broad application value in the field of online monitoring and reaction process control for organic synthesis reactions.

Claims

1. A method for monitoring organic synthesis reactions based on near-infrared spectroscopy, comprising continuously acquiring and storing time-series spectra of the entire reaction process, characterized in that, The method includes: The time-series spectrum of the reaction system is continuously acquired, and the effective spectral band, preprocessing method, centering parameter and scaling parameter are set. The time-series spectrum is preprocessed using the effective spectral band, preprocessing method, centering parameter and scaling parameter. Principal component decomposition is performed on the preprocessed time-series spectrum to solve for the principal component loading matrix and the principal component score matrix. The principal component loading matrix and the principal component score matrix are obtained according to any of the following methods: After the spectra are accumulated to a preset sample size or preset duration, a time-series spectral set is constructed from the preprocessed spectra and a single principal component decomposition is performed. Subsequent newly added spectra are not incorporated into the time-series spectral set. The principal component loading matrix is ​​used to project the subsequent newly added spectra to expand the principal component score matrix. Alternatively, after the spectra have accumulated to a preset sample size or a preset duration, an initial time-series spectral set is constructed and principal component decomposition is performed. Newly added spectra are then incorporated into the time-series spectral set according to a preset sample threshold or time period, and principal component decomposition is performed again to update the principal component loading matrix and the principal component score matrix. Based on the principal component score matrix, extract the time series first principal component score and the time series first principal component variance contribution rate, and determine the model effectiveness based on the time series first principal component variance contribution rate; Perform two-stage linearly coupled dynamics calculations, including velocity feature extraction and dynamic parameter conversion; The rate feature extraction includes: determining a rate observation window with a duration shorter than the current cumulative total monitoring duration; retrieving the first principal component score corresponding to the spectrum at each moment within the rate observation window; calculating the reference distance between the first principal component score at each moment and the first principal component score of the initial frame spectrum after material homogenization; forming a reference distance sequence from multiple reference distances within the rate observation window; performing a first-order least squares linear fit with the reference distance as the ordinate and the corresponding monitoring time as the abscissa to obtain the time series slope; verifying the time series slope by combining the goodness of fit and the high-order principal component fluctuation index; and associating the verified time series slope with the center time value of the corresponding rate observation window. The kinetic parameter conversion includes: obtaining manually entered reaction order or pseudo-reaction order, or determining the reaction order based on entered chemical reaction information; Under the condition that the time slope and the reaction rate are proportional, the time slope is transformed by a kinetic function corresponding to the reaction order to obtain the time transformation quantity. A linear model of the time transformation quantity and the reaction time is constructed, and the kinetic constant is solved by regression. Combining the kinetic constant, the initial concentration and the reaction stoichiometry, the instantaneous conversion rate is output in real time and the reaction time required to reach the target conversion rate is estimated.

2. The method for monitoring organic synthesis reactions based on near-infrared spectroscopy according to claim 1, characterized in that, Three criteria for determining the effectiveness of the model are set, namely, the variance contribution rate of the first principal component is higher than a first threshold, the goodness of linear fit of the first principal component in time series is greater than a second threshold, and the goodness of linear fit of the second-order and above principal components in time series is lower than a third threshold. When the three judgment conditions are met simultaneously, the judgment model takes effect, integrating the original time-series spectrum, effective preprocessing parameters, principal component scores, principal component contribution rates, kinetic calculation data, and state markers to construct a time-series response evidence map for storing, visualizing, and tracing response process data. If any one of the three judgment conditions is not met, the judgment model is not effective and an abnormal prompt message is output. In response to manual handling, the model is remodeled, or the optimization strategy is automatically executed and the model is remodeled. The optimization strategy includes adjusting the first threshold, the second threshold, and the third threshold, as well as adjusting any one or more of the following: effective spectral band, preprocessing method, centering parameter, scale scaling parameter, preset sample size, preset duration, preset sample threshold, time period, and rate observation window. The first threshold, the second threshold, and the third threshold are input through a human-computer interaction interface, or are generated by the system through optimization within a preset value range; The system is configured with a response trajectory deviation early warning mechanism: the variance contribution rate of the first principal component within each rate observation window is calculated in real time, and the fluctuation range obtained from multiple repeated measurements is used as the early warning threshold; When the variance contribution rate of the first principal component within the window is lower than the warning threshold for multiple consecutive frames, the reaction system is judged to be abnormal. After determining that the system is abnormal, the human-machine interface outputs an abnormal prompt and generates an interactive signal, supporting message push or sending control commands to external related systems.

3. The method for monitoring organic synthesis reactions based on near-infrared spectroscopy according to claim 1, characterized in that, The rate observation window length is determined using adaptive optimization: The minimum window length is set to no less than 3 data points, and different window widths are traversed within a preset value range. For each window, the instantaneous reaction rate is extracted by approximating the local window secant. After nonlinear transformation of the kinetic equation, linear fitting is performed to obtain the kinetic fit goodness, and the variance contribution rate of the first principal component is checked to see if it reaches the threshold. The set of qualified candidate windows is obtained through filtering. Analyze the variation of goodness of fit with window length, determine the stable interval with small fluctuations in goodness of fit, and select the optimal observation window within the stable interval; The system can re-execute optimization during operation and dynamically adjust the window length.

4. The method for monitoring organic synthesis reactions based on near-infrared spectroscopy according to claim 1, characterized in that, Adaptive optimization of spectral preprocessing strategy: Traverse multiple spectral preprocessing algorithms and their combinations to form candidate solutions; Principal component decomposition was performed on the spectra of each group of candidate schemes, and the variance contribution rate and dynamic fit goodness of the first principal component were calculated. Using the first principal component's variance contribution rate reaching a preset threshold as the admission condition, only candidate solutions that meet the threshold condition are retained; The optimal pretreatment scheme is selected from the qualified candidate schemes based on the goodness of fit of the kinetics, and a self-learning database is constructed to store the correspondence between the reaction system, process conditions and the optimal pretreatment strategy. The preprocessing algorithms include baseline correction, multivariate scattering correction, standard normal variable transformation, mean centering, vector normalization, Savitzky-Golay smoothing, sliding window smoothing, Gaussian smoothing, first derivative processing, second derivative processing, wavelet transform denoising, and combinations of the above methods.

5. The method for monitoring organic synthesis reactions based on near-infrared spectroscopy according to claim 1, characterized in that, Blank solvent spectral compensation correction was performed for temperature-programmed reaction processes. Using the same pure solvent as the reaction system, time-series blank solvent spectra were collected according to the same temperature rise procedure as the formal monitoring, and the mapping relationship between temperature and solvent spectral drift was established. After collecting the measured spectra of the reaction system, baseline compensation is completed based on this mapping relationship; The corrected spectral data were used for subsequent principal component decomposition, rate extraction, and kinetic solution to suppress solvent spectral perturbations caused by temperature changes.

6. The method for monitoring organic synthesis reactions based on near-infrared spectroscopy according to claim 1, characterized in that, This includes a method for determining the reaction endpoint: when the real-time conversion rate obtained from kinetic calculations reaches a preset conversion rate threshold, the reaction is determined to have reached its endpoint. The conversion rate threshold can be set manually or automatically estimated by the system.

7. The method for monitoring organic synthesis reactions based on near-infrared spectroscopy according to claim 1, characterized in that, After a single reaction, it has the function of offline reconstruction, review and optimization of projection space: retrieve all the original spectra of this reaction, reuse or reconfigure the preprocessing parameters, and re-perform principal component decomposition and kinetic fitting calculations to achieve iterative comparison and review of multiple sets of parameters. The retrospective graphs are used only for parameter verification and result optimization.

8. The method for monitoring organic synthesis reactions based on near-infrared spectroscopy according to claim 1, characterized in that, The monitoring method is suitable for non-one-time feeding conditions such as sequential dripping, continuous feeding, and intermittent feeding. Time-series spectral acquisition, data processing, and modeling analysis are carried out during the initial feeding stage. The model parameters and kinetic characteristic data obtained during the initial feeding stage are continuously applied to subsequent reaction stages to achieve dynamic monitoring and reaction process prediction throughout the non-one-time feeding process.

9. The method for monitoring organic synthesis reactions based on near-infrared spectroscopy according to claim 1, characterized in that, The monitoring method is implemented using an intelligent platform for monitoring organic synthesis reactions. The intelligent platform for monitoring organic synthesis reactions includes a reaction module, a near-infrared sampling module, a material feeding and discharging module, an inert atmosphere control module, a constant temperature control module, a central control unit, and a human-machine interface. The inert atmosphere control module is used to achieve an inert protective environment that isolates the reaction system from air, and the constant temperature control module is used to regulate the temperature of the reaction system. The feeding and discharging module is used to complete automated sample feeding and discharging control; The near-infrared sampling module is used to continuously acquire time-series spectral data of the entire reaction process; The central control unit is used to execute the spectral preprocessing, principal component decomposition, and kinetic calculation processes. The human-computer interaction interface is used for parameter input, data display, and abnormal interaction control.

10. An intelligent platform for monitoring organic synthesis reactions, used to execute the near-infrared spectroscopy-based organic synthesis reaction monitoring method as described in any one of claims 1 to 9, characterized in that, Including human-computer interaction interfaces; The human-computer interaction interface is used to display one or more of the following: the instantaneous conversion rate output in real time by the monitoring method, the reaction time estimated by the monitoring method to reach the target conversion rate, and the time-series relative reaction rate obtained by comparing the time-series slope at each time step with the first time-series slope.

Images