Automatic precise control method and system for macrolide derivative production process

Abstract

The invention discloses an automatic precise control method and system for a macrolide derivative production process, and relates to the technical field of intelligent control. Multidimensional process parameter time sequence data and raw material spectral characteristic data of a reaction kettle are collected in real time, dynamic self-adaptive filtering processing is performed on the multidimensional process parameter time sequence data, a denoising process parameter sequence is generated, and the denoising process parameter sequence is calculated; performing baseline correction on the spectral feature data of the raw materials, generating standard spectral data, performing feature fusion, constructing a process parameter feature matrix, inputting an improved regression model constrained by process knowledge, outputting a key parameter sensitivity vector and a product yield predicted value, extracting a dominant parameter sub-matrix from the process parameter feature matrix, and calculating the product yield according to the dominant parameter sub-matrix. A time-varying state space model is constructed by adopting a process time sequence alignment algorithm, the time-varying state space model is dynamically compared with a predicted value of the product yield, an optimization control parameter set is generated through a multi-objective optimization model, and the yield of the macrocyclic acid derivative product can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control technology, and specifically to an automated precise control method and system for a macrolide derivative production process. Background Art

[0002] As an important chemical additive, macrolide derivatives are widely used in fine chemicals, petrochemicals, papermaking, light industry and other fields. Their production process usually involves multi-step reactions, strict process conditions and complex reaction kinetics. During the production process, it is necessary not only to monitor the conventional process parameters such as temperature, pressure, pH value in the reactor in real time, but also to combine the spectral information of raw materials, flow rate, temperature and other multi-dimensional data to accurately control the reaction state. Due to the high nonlinearity and time-varying nature of the reaction process, coupled with factors such as raw material quality fluctuations, equipment response lag and environmental interference, traditional manual control and static model-based control methods are difficult to meet the requirements of product quality and production efficiency. Therefore, intelligent, automated and precise control has become an inevitable trend to achieve high-quality and stable production.

[0003] Existing process control methods are mostly based on linear or static parameter correction, which makes it difficult to cope with the complex situation where process parameters change dynamically over time. They also lack deep fusion processing of multi-sensor and multi-dimensional data, making it difficult to simultaneously utilize information such as process parameters and spectral features.

[0004] To this end, the present invention proposes a method and system for automated and precise control of the production process of macrolide derivatives. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes an automated and precise control method and system for the production process of macrolide derivatives, which can significantly improve the yield of macrolide derivative products.

[0006] To achieve the above-mentioned purpose, a method for automated and precise control of the production process of macrolide derivatives is proposed, comprising the following steps: Step 1: Use the distributed sensing system to collect the multi-dimensional process parameter time series data of the reactor in real time and simultaneously obtain the spectral characteristic data of the raw materials; Step 2: Implement dynamic adaptive filtering on the multi-dimensional process parameter time series data to generate a denoised process parameter sequence, and perform baseline correction on the raw material spectral feature data to generate standard spectral data; Step 3: Feature fusion of the denoised process parameter sequence and the standard spectral data, constructing a process parameter feature matrix, inputting an improved regression model constrained by process knowledge, and outputting key parameter sensitivity vectors and product yield prediction values; Step 4: Based on the key parameter sensitivity vector, the dominant parameter sub-matrix is ​​extracted from the process parameter feature matrix, and the time-varying state space model is constructed using the process timing alignment algorithm; Step 5: dynamically compare the time-varying state space model with the product yield prediction value. When the deviation exceeds the preset threshold, generate an optimized control parameter set through a multi-objective optimization model; Step 6: Input the optimized control parameter set into the adaptive controller to generate execution instructions to control the process parameters.

[0007] The real-time acquisition of multi-dimensional process parameter time series data of the reactor by the distributed sensing system comprises the following steps: Step 101: Arrange a distributed sensor network for collecting various process parameters at key process nodes of the reactor; Step 102: synchronously obtain the process parameter time series corresponding to each process parameter from the sensor network, and use the time protocol to time-domain align the time of the process parameter time series to generate an original data set with a unified timestamp to form multi-dimensional process parameter time series data.

[0008] The method of synchronously acquiring the raw material spectral characteristic data is: The raw material flow is scanned in real time by an online near-infrared spectrometer with a wavelength range of 900-1700 nm, and the raw spectrum matrix is ​​obtained in each sampling period, and the raw material flow rate and raw material temperature are collected synchronously; The Savitzky-Golay filter is used to perform spectral denoising on the original spectral matrix to generate a standard spectral matrix; The standard spectrum matrix, raw material flow rate and raw material temperature constitute the raw material spectrum characteristic data.

[0009] The method of implementing dynamic adaptive filtering on the multi-dimensional process parameter time series data to generate a denoising process parameter sequence comprises the following steps: Step 211: constructing the original multi-dimensional process parameter time series data with a unified time stamp into the form of a data matrix X(t); Step 212: Set the window length to Tw and calculate the local mean of each process parameter within the window. With local variance Constitute the local noise statistics and define the local signal-to-noise ratio As a quantitative indicator of noise level; Step 213: designing an adaptive filter model based on local noise statistical information; Step 214: after the dynamic adaptive filter model is constructed, point-by-point filtering is performed on the multi-dimensional parameters at each time t; Step 215: Collect the time series data sequence after the denoising of each process parameter after dynamic adaptive filtering processing, and express it as a denoising process parameter sequence.

[0010] The method of simultaneously performing baseline correction on the raw material spectral characteristic data to generate standard spectral data comprises the following steps: Step 221: record the standard spectrum matrix as Sraw, where each column represents the spectrum response within a wavelength range within a sampling period; Step 222: first, extract the local minimum value of the wavelength for the spectrum curve within each sampling period; Step 223: Based on the extracted local minimum, a baseline model of the entire spectral curve is constructed using a low-order polynomial fitting or an adaptive smoothing method; Step 224: Using the constructed baseline model, perform baseline correction on the standard spectrum matrix.

[0011] The process of fusing the denoising process parameter sequence with the standard spectral data to construct a process parameter feature matrix includes the following steps: Step 311: Mark the collected denoising process parameter sequence as , the collected standard spectral data are marked as ; Step 312: performing correlation analysis on each parameter in the denoising process parameter sequence, removing redundant highly correlated parameters, and extracting the main components of the spectral data by using a dimensionality reduction method; Step 313: splicing the denoising process parameter sequence and the standard spectrum data after dimension reduction along the time dimension to form a new fusion feature matrix; Step 314: Standardize the fusion feature matrix generated after fusion to obtain a process parameter feature matrix.

[0012] The improved regression model with input process knowledge constraints and outputting key parameter sensitivity vectors and product yield prediction values ​​comprises the following steps: Step 321: label the process parameter feature vector of each row in the process parameter feature matrix as X(t); label the target variable product yield as Y(t); Step 322: Label the improved regression model as , and introduce a priori constraints for some key parameters based on the process experience of each process stage in the production of macrolide derivatives; Step 323: pre-collecting several process parameter feature vectors and corresponding product yields at various moments in various process stages during the production of macrolide derivatives as a sample set; Step 324: using the process parameter feature vectors at each moment in the sample set as the input of the improved regression model, using the mean square error between the product yield corresponding to each moment in the sample set and the predicted value calculated by the improved regression model expression as the loss function, using the prior constraints of various key parameters as inequality constraints, and using the inequality constraint optimization algorithm to solve the intercept term and sensitivity coefficient of the improved regression model; Step 325: The sensitivity coefficients of the solved key parameters are used to form a key parameter sensitivity vector; for any new moment in the production process, the process parameter characteristic vector of the new moment is input into the improved regression model to obtain a predicted value of the product yield.

[0013] The extracting of the dominant parameter sub-matrix from the process parameter feature matrix based on the key parameter sensitivity vector comprises the following steps: Step 411: Mark the key parameter sensitivity vector as , calculate the absolute value of each sensitivity coefficient | |; where d=1,2,…,D; Step 412: Determine the dominant parameter set by a threshold method or a sorting method; Step 413: extracting the parameter value corresponding to each dominant parameter in the dominant parameter set from the process parameter feature matrix to form a dominant parameter sub-matrix.

[0014] The method of constructing the time-varying state space model using the process timing alignment algorithm is as follows: The dynamic time warping algorithm is selected as the process timing alignment algorithm; Select a dominant parameter timing template that is stable and represents a normal process state from historical data; Get the current real-time dominant parameter data sequence; The dynamic time warping algorithm is used to compare the two time series data, and then the cost matrix is ​​constructed to calculate the matching cost between the dominant parameter time series template and the dominant parameter data sequence to find the optimal alignment path π to achieve nonlinear correction on the time axis; The time-varying state-space model describing the dynamic evolution of the process is constructed using the dominant parameter data aligned by the dynamic time warping algorithm.

[0015] The method of dynamically comparing the time-varying state space model with the product yield prediction value is: Inputting a real-time process parameter characteristic matrix into the improved regression model to obtain a predicted product yield prediction value; Inputting a real-time dominant parameter sub-matrix into the time-varying state-space model to obtain an expected yield prediction value output by the time-varying state-space model;

[0016] Compare the difference between the product yield prediction value and the expected yield prediction value to determine whether the difference is greater than a preset difference threshold.

[0017] The method of generating the optimization control parameter set through the multi-objective optimization model is: The difference threshold between the predicted value of the comparison product yield and the predicted value of the expected yield is marked as h; The expected yield prediction value is marked as H; Based on the expected yield prediction value and the improved regression model, a multi-objective optimization function Yu is constructed; Based on the expected yield prediction value, the improved regression model and the difference threshold, a constraint condition set U is constructed; Taking minimization of multi-objective optimization function as optimization objective and constraint condition set as constraint condition, a multi-objective optimization model is constructed. By using the optimization algorithm to solve the multi-objective optimization model, the solution values ​​of the parameter variables to be solved for each key parameter are obtained, and the solution values ​​of all parameter variables constitute the optimization control parameter set.

[0018] An automated precision control system for the production process of macrolide derivatives is proposed, including a raw data collection module, a preprocessing module, a yield prediction module, a state space model training module, and a control parameter optimization module; wherein each module is electrically connected; The raw data collection module collects the multi-dimensional process parameter time series data of the reactor in real time through the distributed sensing system, synchronously obtains the raw material spectral characteristic data, and sends the multi-dimensional process parameter time series data and the raw material spectral characteristic data to the preprocessing module; The preprocessing module implements dynamic adaptive filtering processing on the multi-dimensional process parameter time series data to generate a denoising process parameter sequence, and at the same time performs baseline correction on the raw material spectral feature data to generate standard spectral data, and sends the denoising process parameter sequence and the standard spectral data to the yield prediction module; The yield prediction module performs feature fusion of the denoised process parameter sequence and the standard spectral data, constructs a process parameter feature matrix, inputs an improved regression model constrained by process knowledge, outputs key parameter sensitivity vectors and product yield prediction values, and sends the key parameter sensitivity vectors to the state space model training module, and sends the improved regression model and product yield prediction values ​​to the control parameter optimization module; The state space model training module extracts the dominant parameter sub-matrix from the process parameter feature matrix based on the key parameter sensitivity vector, constructs the time-varying state space model using the process timing alignment algorithm, and sends the time-varying state space model to the control parameter optimization module; The control parameter optimization module dynamically compares the time-varying state space model with the product yield prediction value. When the deviation exceeds the preset threshold, the multi-objective optimization model generates an optimized control parameter set, which is then input into the adaptive controller to generate execution instructions to control the process parameters.

[0019] Compared with the prior art, the present invention has the following beneficial effects: The present invention collects process parameters such as temperature, pressure, pH and online near-infrared spectral data in the reactor through a distributed sensing system, and uses dynamic adaptive filtering and baseline correction to denoise the original data, thereby improving the accuracy of the original data and reducing noise pollution. The denoised process parameters and standardized spectral data are then reduced in dimension and weighted fused to construct a high-dimensional feature matrix reflecting the process state. Based on the fused feature matrix and product yield label data, an improved regression model is trained through a number of prior constraints, and a key parameter sensitivity vector is output, thereby achieving accurate prediction of future yields. The sensitivity vector selects the dominant parameters from the fused feature matrix, uses the dynamic time warping algorithm to align the process timing, and constructs a time-varying state space model to capture the dynamic characteristics of the process, so as to obtain accurate expected yields. Finally, the predicted value of the time-varying state space model is dynamically compared with the product yield collected in real time, and the deviation is calculated. The multi-objective optimization model is used to automatically generate an optimized control parameter set, which realizes intelligent monitoring and real-time optimization control of key parameters in the process, thereby significantly improving product yield and product quality, reducing energy consumption and safety risks, and improving the level of production automation and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of the method for automated and precise control of the production process of macrolide derivatives in Example 1 of the present invention; Figure 2 This is a diagram showing an example of a configuration of a sensor network in Embodiment 1 of the present invention; Figure 3 This is a module connection relationship diagram of the automated precision control system for the production process of macrolide derivatives in Example 2 of the present invention. DETAILED DESCRIPTION

[0021] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] In large-scale production, due to differences in raw material batches, equipment aging, and environmental factors, key parameters such as temperature, pressure, and pH value in the reactor often experience dynamic fluctuations, making it difficult to maintain stable product yield and quality. At the same time, traditional process control methods do not respond to these dynamic changes in a timely manner, and relying on manual intervention is not only inefficient and error-prone, but also prone to safety hazards. Therefore, how to use intelligent systems to capture process changes in real time, dynamically adjust control parameters, and ensure the accurate and stable operation of the entire production process is a key technical issue that needs to be solved urgently.

[0023] Example 1

[0024] like Figure 1 As shown, the method for automated and precise control of the production process of macrolide derivatives comprises the following steps: Step 1: Use the distributed sensing system to collect the multi-dimensional process parameter time series data of the reactor in real time and simultaneously obtain the spectral characteristic data of the raw materials; Step 2: Implement dynamic adaptive filtering on the multi-dimensional process parameter time series data to generate a denoised process parameter sequence, and perform baseline correction on the raw material spectral feature data to generate standard spectral data; Step 3: Feature fusion of the denoised process parameter sequence and the standard spectral data, constructing a process parameter feature matrix, inputting an improved regression model constrained by process knowledge, and outputting key parameter sensitivity vectors and product yield prediction values; Step 4: Based on the key parameter sensitivity vector, the dominant parameter sub-matrix is ​​extracted from the process parameter feature matrix, and the time-varying state space model is constructed using the process timing alignment algorithm; Step 5: dynamically compare the time-varying state space model with the product yield prediction value. When the deviation exceeds the preset threshold, generate an optimized control parameter set through a multi-objective optimization model; Step 6: Input the optimized control parameter set into the adaptive controller to generate execution instructions to control the process parameters.

[0025] In an embodiment of the present invention, the real-time acquisition of multi-dimensional process parameter time series data of the reactor by a distributed sensing system comprises the following steps: Step 101: Arrange a distributed sensor network for collecting various process parameters at key process nodes of the reactor; Specifically, the process parameters include but are not limited to the temperature, pressure and pH value in the reactor; Specifically, if Figure 2 As shown in the structural diagram of the reactor with a distributed sensor network, the arrangement of the distributed sensor network can be: The temperature sensing groups are evenly distributed along the axis of the reactor, and each sensing group contains three PT100 platinum resistors distributed at 120°; Figure 2 T1, T2 and T3 in; Three pressure sensing units are installed at the top, middle and bottom of the reactor, respectively, and piezoelectric sensors are used as pressure sensing units; that is, Figure 2 P1, P2 and P3 in Two pH value detection probes are symmetrically arranged above and below the stirring impeller and are equipped with an automatic cleaning device; Figure 2 pH1 and pH2; All sensors are connected to the central control station via industrial Ethernet, the sampling frequency is uniformly set to 2Hz, and hardware-level synchronization is established with the stirring speed signal of the reactor; Step 102: synchronously obtain the process parameter time series corresponding to each process parameter from the sensor network through the Modbus-TCP protocol, and use the IEEE 1588 precision time protocol to perform time domain alignment on the time of the process parameter time series, and generate a raw data set with a unified timestamp to form multi-dimensional process parameter time series data; specifically, including temperature time series, pressure time series and pH time series; Furthermore, the method of synchronously acquiring the raw material spectral characteristic data is: The raw material flow is scanned in real time by an online near-infrared spectrometer with a wavelength range of 900-1700 nm, and the raw spectrum matrix is ​​obtained in each sampling period, and the raw material flow rate and raw material temperature are collected synchronously; The Savitzky-Golay filter is used to perform spectral denoising on the original spectral matrix to generate a standard spectral matrix, wherein the window width of the Savitzky-Golay filter is 15 points, and a quadratic polynomial fitting method is used for fitting; The standard spectrum matrix, raw material flow rate and raw material temperature constitute the raw material spectrum characteristic data.

[0026] In an embodiment of the present invention, the step of performing dynamic adaptive filtering on the multi-dimensional process parameter time series data to generate a denoising process parameter sequence comprises the following steps: Step 211: constructing the original multi-dimensional process parameter time series data with a unified time stamp into the form of a data matrix X(t); Specifically, the formula form of X(t) is: , i=1,2,3...I; where I is the number of process parameters in the collected multidimensional process parameter time series data; i is the number of the process parameter, represents the measured value of the ith parameter at time t; t is any time when the process parameter data is collected through the distributed sensor network; At the same time, each process stage in the production process of macrolide derivatives is predefined according to the process flow, and the number of the process stage is marked as k; specifically, in the production process of macrolide derivatives, the process stages can be divided into a raw material pretreatment stage, an initiation and cyclization stage, a side chain modification stage, a post-treatment and purification stage, and a stability enhancement stage, wherein the raw material pretreatment stage is used for raw material activation and impurity control, the initiation and cyclization stage is used for macrocyclic skeleton construction, the side chain modification stage is used for the directional introduction of functional groups, and the stability enhancement stage is used for product separation and purification; Step 212: After the matrix is ​​constructed, in order to make the filter have dynamic adaptive capabilities, it is necessary to estimate the noise characteristics of each parameter within the sliding time window; for this purpose, the window length is set to Tw, and the local mean of each process parameter is calculated within the window. With local variance Constitute the local noise statistics and define the local signal-to-noise ratio As a quantitative indicator of noise level; It should be noted that, combined with the characteristics of the current process stage, this step provides a basis for the dynamic adjustment of the adaptive parameters in the subsequent filter, so that the response speed and filtering strength of the filter can be flexibly changed at different stages (such as startup, stabilization or end stage); Step 213: designing an adaptive filter model based on local noise statistical information; Specifically, the adaptive Kalman filter is used as the adaptive filter model, and for the i-th parameter, the following state update equation is used: ;in, represents the denoised estimate after filtering, is the adaptive Kalman gain, and its calculation formula is: ,in is the error covariance at the previous moment, and The dynamic adjustment is performed based on the noise variance calculated in the current sliding window and the kth process stage. For example, ,in, is the preset adjustment coefficient corresponding to the kth process stage; Step 214: after the dynamic adaptive filter model is constructed, point-by-point filtering is performed on the multi-dimensional parameters at each time t; Specifically, at each sampling time t, the error covariance is updated in real time according to the current process stage and the local noise statistical characteristics. , and recalculate the Kalman gain , to ensure that the filter can respond quickly to sudden noise or process fluctuations; It can be understood that this dynamic update mechanism ensures that the filter can automatically increase the smoothing force when the noise is large or the process startup phase, and when the process is stable, the filter allows more response to capture the real signal changes.

[0027] The whole process can be recorded as:

[0028] in, is the process noise covariance, which is also dynamically set according to the process stage; Step 215: Collect the time series data sequence after the denoising of each process parameter after dynamic adaptive filtering processing, and express it as a denoising process parameter sequence.

[0029] In an embodiment of the present invention, the simultaneous baseline correction of the raw material spectral characteristic data to generate standard spectral data comprises the following steps: Step 221: record the standard spectrum matrix as Sraw, where each column represents the spectrum response within a wavelength range within a sampling period; Step 222: Considering that in the raw spectral data after preprocessing, there is often a non-ideal baseline drift due to factors such as instrument response, ambient temperature changes, and raw material flow rate fluctuations. Therefore, firstly, the local minimum value of the wavelength is extracted for the spectral curve within each sampling period; Specifically, the local minimum detection refers to selecting a window interval of a preset sliding window size within the wavelength interval, calculating the local minimum value within the window interval as a preliminary estimation point of the baseline within the window interval, so as to preliminarily capture the low value trend in the spectral curve; Step 223: Based on the extracted local minimum, a baseline model of the entire spectral curve is constructed using a low-order polynomial fitting or an adaptive smoothing method; Specifically, taking the polynomial fitting method as an example, the baseline model is set as: Where n is the fitting order, is the wavelength, and the coefficients are solved by the least squares method , using the extracted local minimum as the label value of the spectral curve at the corresponding wavelength, an adaptive algorithm such as adaptive iterative weighted least squares or ALS method is used to compensate for the nonlinear drift caused by the raw material flow rate and temperature while ensuring smooth fitting of the low signal-to-noise ratio area; this process ensures that the generated baseline can fully reflect the background trend of the actual spectrum; Step 224: using the constructed baseline model, performing baseline correction on the standard spectrum matrix; Specifically, the method of performing baseline correction on the standard spectrum matrix is: Correction is performed using the following correction formula: ,in, is the spectral value after baseline calibration, is the wavelength in the standard spectral matrix The corresponding spectral value is The lowest spectral value at the fitted wavelength can be understood as the background spectral value. The background spectral value is subtracted to achieve benchmark correction.

[0030] In an embodiment of the present invention, the process of fusing the denoising process parameter sequence with the standard spectral data to construct a process parameter feature matrix includes the following steps: Step 311: Mark the collected denoising process parameter sequence as , the collected standard spectral data are marked as ; Specifically, is the time series data of multidimensional process parameters in the form of:

[0031] Among them, each element corresponds to a process parameter, which represents the denoised sequence of the corresponding process parameter at time t; The standard spectral dataset is the standard spectrum matrix in each sampling period p, and its form is:

[0032] Wherein, N represents the number of wavelength points, Q represents the number of sampling periods, and each element in the standard spectrum matrix represents the standard spectrum data value of the corresponding wavelength point in the corresponding sampling period; Further preferably, since the time intervals of the denoising process parameter sequence and the standard spectral data may be different, the denoising process parameter sequence and the standard spectral data may be time-aligned first, so an interpolation algorithm such as linear interpolation or spline interpolation is used to time-align one of the data sets to make it consistent with the sampling time point of the process parameters. After the alignment is completed, the obtained denoising process parameter sequence and the standard spectral data will be aligned at the same timestamp; As an example, in an actual reaction process, a set of denoising process parameter sequences were collected The expression is: ; Each row corresponds to a time point (t=1 to 5), which is temperature, pressure, and pH value in turn; Standard spectral data collected It is expressed as: , which means that the spectrum data is collected at 4 points in the wavelength direction (wavelength 1000, 1200, 1400, 1600nm) at 5 sampling moments to form a 5×4 matrix; Step 312: Since both the denoising process parameter sequence and the standard spectral data may contain redundant or highly correlated features, feature selection and dimensionality reduction processing are required to improve the fusion effect and reduce the computational burden. To this end, correlation analysis is performed on each parameter in the denoising process parameter sequence to remove redundant highly correlated parameters, and for the spectral data, the main components are extracted through dimensionality reduction methods. Specifically, the correlation analysis uses the Pearson correlation coefficient or the mutual information method to quantify the correlation between various process parameters and selects the main parameters that are highly correlated with the yield of the target product; The dimensionality reduction methods include, but are not limited to, principal component analysis (PCA), weighted average or independent component analysis (ICA) and other dimensionality reduction methods; For the instance of step 311, the process of dimensionality reduction is as follows: In this example, since the process parameters such as temperature, pressure and pH are all key parameters, they are directly retained without dimensionality reduction; As for standard spectral data, the original 4-dimensional spectral data usually contains strong correlation, and simple weighted average is used for dimensionality reduction, and the weight vector of the dimension is set to [0.25, 0.25, 0.25, 0.25]; For each sampling moment, the data of the four wavelength points are projected into a scalar. For example, for the first moment, its spectral feature is 0.25×(0.80+0.75+0.70+0.65)=0.25×2.90=0.725; The dimensionality reduction results at each moment are calculated in sequence: 0.725, 0.745, 0.735, 0.755, 0.765; Thus, a feature vector with a dimension of 5×1 is obtained; Step 313: splicing the denoising process parameter sequence and the standard spectrum data after dimension reduction along the time dimension to form a new fusion feature matrix; Specifically, in the feature process, each wavelength point of the standard spectral data is regarded as an independent feature, and each parameter of the denoising process parameter sequence is also regarded as an independent feature. The weighted fusion method is used to configure the weights of the process parameters and spectral data, and each weight is set in advance; For the example of step 312, the process of feature fusion is as follows: Set the weight of each process parameter to 1; Then at t=1, the process parameters are: [80, 2.0, 7.0]; the spectral feature is 0.725, and the fused feature vector after fusion is [80, 2.0, 7.0, 0.725]. Similarly, for t=2, 3, 4, 5, the corresponding fused feature vectors are: [82, 2.1, 7.1, 0.745], [83, 2.0, 7.0, 0.735], [84, 2.2, 7.2, 0.755], [85, 2.1,7.3, 0.765] respectively; The fusion feature matrix formed by concatenating all fusion feature vectors in chronological order is: ; Step 314: normalizing the fusion feature matrix generated after fusion to obtain a process parameter feature matrix; For the example of step 313, the process parameter characteristic matrix obtained by the standardization process is approximately: .

[0033] In an embodiment of the present invention, the improved regression model with input process knowledge constraints and outputting key parameter sensitivity vectors and product yield prediction values ​​comprises the following steps: Step 321: label the process parameter feature vector of each row in the process parameter feature matrix as X(t); label the target variable product yield as Y(t); in, Where D represents the feature dimension obtained after fusion; Step 322: Label the improved regression model as , and introduce a priori constraints for some key parameters based on the process experience of each process stage in the production of macrolide derivatives; The specific expression of the improved regression model is set as: ;in is the intercept term, is the sensitivity coefficient of the dth key parameter; Specifically, the a priori constraints can have a positive impact on certain key parameters, that is, the corresponding >0, you can also set the impact range for some key parameters, that is, the corresponding within a certain preset range; Step 323: pre-collecting several process parameter feature vectors and corresponding product yields at various moments in various process stages during the production of macrolide derivatives as a sample set; Step 324: using the process parameter feature vectors at each moment in the sample set as the input of the improved regression model, using the mean square error between the product yield corresponding to each moment in the sample set and the predicted value calculated by the improved regression model expression as the loss function, using the prior constraints of various key parameters as inequality constraints, and using the inequality constraint optimization algorithm to solve the intercept term and sensitivity coefficient of the improved regression model; Step 325: The sensitivity coefficients of the solved key parameters are used to form a key parameter sensitivity vector; for any new moment in the production process, the process parameter characteristic vector of the new moment is input into the improved regression model to obtain a predicted value of the product yield.

[0034] In an embodiment of the present invention, extracting a dominant parameter sub-matrix from a process parameter feature matrix based on a key parameter sensitivity vector comprises the following steps: Step 411: Mark the key parameter sensitivity vector as , calculate the absolute value of each sensitivity coefficient | |; where d=1,2,…,D; Understandably, | | Characterizes the importance of the dth key parameter and is a direct indicator of the impact of the dth key parameter on the target variable; Step 412: Determine the dominant parameter set by a threshold method or a sorting method; Specifically, the threshold method refers to presetting an importance threshold, and forming a dominant parameter set with key parameters whose absolute values ​​are greater than the importance threshold; The ranking method mentioned above refers to the absolute values ​​of all sensitivity coefficients| | Sort by numerical value, and select D' key parameters from large to small to form a dominant parameter set, where D' is the preset number of dominant parameters; Step 413: extracting the parameter value corresponding to each dominant parameter in the dominant parameter set from the process parameter feature matrix to form a dominant parameter sub-matrix; It can be understood that by extracting the dominant parameter submatrix mentioned above, obtaining the sensitivity vector from the model training, and calculating the absolute influence index of each parameter, the dominant parameter index that needs attention is determined, which greatly reduces the dimension of the process parameter feature matrix.

[0035] Furthermore, the method of constructing the time-varying state space model using the process timing alignment algorithm is as follows: The dynamic time warping algorithm is selected as the process timing alignment algorithm; Select a dominant parameter timing template that is stable and represents a normal process state from historical data; Get the current real-time dominant parameter data sequence; The dynamic time warping algorithm is used to compare the two time series data, and then the cost matrix is ​​constructed to calculate the matching cost between the dominant parameter time series template and the dominant parameter data sequence to find the optimal alignment path π to achieve nonlinear correction on the time axis; It can be understood that, through the above-mentioned nonlinear correction, each moment j in the real-time sequence can be mapped to a certain moment i in the reference sequence template, thereby eliminating the time distortion in the process; Using the dominant parameter data aligned by the dynamic time warping algorithm, a time-varying state space model describing the dynamic evolution of the process is constructed; Specifically, the time-varying state space model for describing the dynamic evolution of the process is constructed in the form of:

[0036] Among them, x(t) represents the state vector, which is the value of the aligned dominant parameter submatrix at time t, u(t) is the controllable input such as adjustment parameters and compensation variables, y(t) is the observed output, which corresponds to the key process indicators measured in real time, such as the predicted value of product yield, A(t), B(t) and C(t) are time-varying system matrices, which reflect the dynamic characteristics of the process in different time periods; w(t) and v(t) are process noise and measurement noise, respectively; Finally, by using the system identification method of recursive least squares or Kalman filter to update parameters on historical alignment data, the above system matrix can be updated online in real time, so that the state space model can dynamically reflect the current process status and evolution trend; It should be noted that the constructed time-varying state-space model is used to capture the dynamic changes of the process. By comparing the deviation between the real-time predicted product yield data and the product yield in the template, process anomalies or process drift can be discovered in a timely manner, so that correction instructions can be produced in a timely manner and the controllable parameters in the production process can be adjusted in a timely manner.

[0037] In an embodiment of the present invention, the method of dynamically comparing the time-varying state space model with the product yield prediction value is: Inputting a real-time process parameter characteristic matrix into the improved regression model to obtain a predicted product yield prediction value; Inputting a real-time dominant parameter sub-matrix into the time-varying state-space model to obtain an expected yield prediction value output by the time-varying state-space model; Compare the difference between the product yield prediction value and the expected yield prediction value to determine whether the difference is greater than a preset difference threshold; It is understandable that the improved regression model makes predictions based on the parameter values of current process parameters, while the time-varying state space model makes predictions based on the empirical template of the yield in the historical production process. Generally speaking, the yield value predicted by the improved regression model is based on data fitting for the entire current production cycle and is the predicted future yield, while the time-varying state space model makes predictions based on the historical template, so the predicted value is an expected value, that is, the yield that is expected to be achieved when producing according to the current production parameters. If the predicted values of the two differ greatly, it indicates that there may be some abnormalities in the current production parameters and adjustments are needed.

[0038] In the embodiment of the present invention, the method for generating the optimized control parameter set through the multi-objective optimization model is as follows: Mark the difference threshold between the predicted value of the by-product yield and the predicted value of the expected yield as h; Mark the predicted value of the expected yield as H; Construct a multi-objective optimization function Yu; The expression of Yu is , is the parameter variable to be solved for the d-th key parameter, that is, it represents the adjusted parameter value, is the parameter value of the current d-th key parameter; it is understandable that, measures the difference between the predicted value of the by-product yield and the predicted value of the expected yield. When is 0, it means that the current process parameters are most in line with the expectations, measures the adjustment range of each key parameter; among them, g0 and gd are both preset proportionality coefficients; Construct a set of constraint conditions U; Specifically, the set of constraint conditions U at least includes <h, indicating that after the key parameters are adjusted, the difference between the predicted yield and the predicted value of the expected yield is less than the difference threshold, so as to meet the expectations; Furthermore, the set of constraint conditions U can also include the allowable range and the change range of each key parameter; the allowable range is the range in which each key parameter should be in the production process. For example, the temperature should be maintained within a certain range, and the change range refers to the difference between the adjusted parameter value of each key parameter and the current parameter value should be within a certain range to reduce the amplitude of parameter changes and avoid drastic changes in the environment in the reaction kettle, resulting in damage to the raw materials; Taking the minimization of the multi-objective optimization function as the optimization goal and the set of constraint conditions as the constraints, a multi-objective optimization model is constructed; By using an optimization algorithm to solve the multi-objective optimization model, the solution values ​​of the parameter variables to be solved for each key parameter are obtained, and the solution values ​​of all parameter variables constitute an optimization control parameter set;

[0039] The optimized control parameter set is input into the adaptive controller to generate execution instructions, that is, the solution values ​​of the parameter variables to be solved for each key parameter are used as control instructions, and the parameter values ​​of each key parameter are adjusted by the production program background to the corresponding solution values ​​to improve the matching degree between the yield of the macrocyclic acid derivative and the expected yield.

[0040] Example 2

[0041] like Figure 3 As shown, the automated precision control system for the production process of macrolide derivatives includes a raw data collection module, a preprocessing module, a yield prediction module, a state space model training module, and a control parameter optimization module; wherein each module is electrically connected; The raw data collection module collects the multi-dimensional process parameter time series data of the reactor in real time through the distributed sensing system, synchronously obtains the raw material spectral characteristic data, and sends the multi-dimensional process parameter time series data and the raw material spectral characteristic data to the preprocessing module; The preprocessing module implements dynamic adaptive filtering processing on the multi-dimensional process parameter time series data to generate a denoising process parameter sequence, and at the same time performs baseline correction on the raw material spectral feature data to generate standard spectral data, and sends the denoising process parameter sequence and the standard spectral data to the yield prediction module; The yield prediction module performs feature fusion of the denoised process parameter sequence and the standard spectral data, constructs a process parameter feature matrix, inputs an improved regression model constrained by process knowledge, outputs key parameter sensitivity vectors and product yield prediction values, and sends the key parameter sensitivity vectors to the state space model training module, and sends the improved regression model and product yield prediction values ​​to the control parameter optimization module; The state space model training module extracts the dominant parameter sub-matrix from the process parameter feature matrix based on the key parameter sensitivity vector, constructs the time-varying state space model using the process timing alignment algorithm, and sends the time-varying state space model to the control parameter optimization module; The control parameter optimization module dynamically compares the time-varying state space model with the product yield prediction value. When the deviation exceeds the preset threshold, the multi-objective optimization model generates an optimized control parameter set, which is then input into the adaptive controller to generate execution instructions to control the process parameters.

[0042] In addition, the parts of the above-mentioned technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0043] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0044] The above preset parameters or preset thresholds are all set by technicians in this field according to actual conditions or obtained through large-scale data simulation.

[0045] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A method for automated and precise control of the production process of macrolide derivatives, characterized in that: The following steps are involved: Step 1: Use the distributed sensing system to collect the multi-dimensional process parameter time series data of the reactor in real time and simultaneously obtain the spectral characteristic data of the raw materials; Step 2: Implement dynamic adaptive filtering on the multi-dimensional process parameter time series data to generate a denoised process parameter sequence, and perform baseline correction on the raw material spectral feature data to generate standard spectral data; Step 3: Feature fusion of the denoised process parameter sequence and the standard spectral data, constructing a process parameter feature matrix, inputting an improved regression model constrained by process knowledge, and outputting key parameter sensitivity vectors and product yield prediction values; Step 4: Based on the key parameter sensitivity vector, the dominant parameter sub-matrix is ​​extracted from the process parameter feature matrix, and the time-varying state space model is constructed using the process timing alignment algorithm; Step 5: dynamically compare the time-varying state space model with the product yield prediction value. When the deviation exceeds the preset threshold, generate an optimized control parameter set through a multi-objective optimization model; Step 6: Input the optimized control parameter set into the adaptive controller to generate execution instructions to control the process parameters.

2. The method for automated and precise control of the production process of macrolide derivatives according to claim 1, characterized in that: The real-time acquisition of multi-dimensional process parameter time series data of the reactor by the distributed sensing system comprises the following steps: Step 101: Arrange a distributed sensor network for collecting various process parameters at key process nodes of the reactor; Step 102: synchronously obtain the process parameter time series corresponding to each process parameter from the sensor network, and use the time protocol to time-domain align the time of the process parameter time series to generate an original data set with a unified timestamp to form multi-dimensional process parameter time series data.

3. The method for automated and precise control of the production process of macrolide derivatives according to claim 2, characterized in that: The method of synchronously acquiring the raw material spectral characteristic data is: The raw material flow is scanned in real time by an online near-infrared spectrometer with a wavelength range of 900-1700 nm, and the raw spectrum matrix is ​​obtained in each sampling period, and the raw material flow rate and raw material temperature are collected synchronously; The Savitzky-Golay filter is used to perform spectral denoising on the original spectral matrix to generate a standard spectral matrix; The standard spectrum matrix, raw material flow rate and raw material temperature constitute the raw material spectrum characteristic data.

4. The method for automated and precise control of the production process of macrolide derivatives according to claim 3, characterized in that: The method of implementing dynamic adaptive filtering on the multi-dimensional process parameter time series data to generate a denoising process parameter sequence comprises the following steps: Step 211: constructing the original multi-dimensional process parameter time series data with a unified time stamp into the form of a data matrix X(t); Step 212: Set the window length to Tw and calculate the local mean of each process parameter within the window. With local variance Constitute the local noise statistics and define the local signal-to-noise ratio As a quantitative indicator of noise level; Step 213: designing an adaptive filter model based on local noise statistical information; Step 214: after the dynamic adaptive filter model is constructed, point-by-point filtering is performed on the multi-dimensional parameters at each time t; Step 215: Collect the time series data sequence after the denoising of each process parameter after dynamic adaptive filtering processing, and express it as a denoising process parameter sequence.

5. The method for automated and precise control of the production process of macrolide derivatives according to claim 4, characterized in that: The method of simultaneously performing baseline correction on the raw material spectral characteristic data to generate standard spectral data comprises the following steps: Step 221: record the standard spectrum matrix as Sraw, where each column represents the spectrum response within a wavelength range within a sampling period; Step 222: first, extract the local minimum value of the wavelength for the spectrum curve within each sampling period; Step 223: Based on the extracted local minimum, a baseline model of the entire spectral curve is constructed using a low-order polynomial fitting or an adaptive smoothing method; Step 224: Using the constructed baseline model, perform baseline correction on the standard spectrum matrix.

6. The method for automated and precise control of the production process of macrolide derivatives according to claim 5, characterized in that: The process of fusing the denoising process parameter sequence with the standard spectral data to construct a process parameter feature matrix comprises the following steps: Step 311: Mark the collected denoising process parameter sequence as , the collected standard spectral data are marked as ; Step 312: performing correlation analysis on each parameter in the denoising process parameter sequence, removing redundant highly correlated parameters, and extracting the main components of the spectral data by using a dimensionality reduction method; Step 313: splicing the denoising process parameter sequence and the standard spectrum data after dimension reduction along the time dimension to form a new fusion feature matrix; Step 314: Standardize the fusion feature matrix generated after fusion to obtain a process parameter feature matrix.

7. The method for automated and precise control of the production process of macrolide derivatives according to claim 6, characterized in that: The improved regression model with input process knowledge constraints and outputting key parameter sensitivity vectors and product yield prediction values ​​comprises the following steps: Step 321: label the process parameter feature vector of each row in the process parameter feature matrix as X(t); label the target variable product yield as Y(t); Step 322: Label the improved regression model as , and introduce a priori constraints for some key parameters based on the process experience of each process stage in the production of macrolide derivatives; Step 323: pre-collecting several process parameter feature vectors and corresponding product yields at various moments in various process stages during the production of macrolide derivatives as a sample set; Step 324: using the process parameter feature vectors at each moment in the sample set as the input of the improved regression model, using the mean square error between the product yield corresponding to each moment in the sample set and the predicted value calculated by the improved regression model expression as the loss function, using the prior constraints of various key parameters as inequality constraints, and using the inequality constraint optimization algorithm to solve the intercept term and sensitivity coefficient of the improved regression model; Step 325: The sensitivity coefficients of the solved key parameters are used to form a key parameter sensitivity vector; for any new moment in the production process, the process parameter characteristic vector of the new moment is input into the improved regression model to obtain a predicted value of the product yield.

8. The method for automated and precise control of the production process of macrolide derivatives according to claim 7, characterized in that: The extracting of the dominant parameter sub-matrix from the process parameter feature matrix based on the key parameter sensitivity vector comprises the following steps: Step 411: Mark the key parameter sensitivity vector as , calculate the absolute value of each sensitivity coefficient | |; where d=1,2,…,D; Step 412: Determine the dominant parameter set by a threshold method or a sorting method; Step 413: extracting the parameter value corresponding to each dominant parameter in the dominant parameter set from the process parameter feature matrix to form a dominant parameter sub-matrix.

9. The method for automated and precise control of the production process of macrolide derivatives according to claim 8, characterized in that: The method of constructing the time-varying state space model using the process timing alignment algorithm is as follows: The dynamic time warping algorithm is selected as the process timing alignment algorithm; Select a dominant parameter timing template that is stable and represents a normal process state from historical data; Get the current real-time dominant parameter data sequence; The dynamic time warping algorithm is used to compare the two time series data, and then the cost matrix is ​​constructed to calculate the matching cost between the dominant parameter time series template and the dominant parameter data sequence to find the optimal alignment path π to achieve nonlinear correction on the time axis; The time-varying state-space model describing the dynamic evolution of the process is constructed using the dominant parameter data aligned by the dynamic time warping algorithm.

10. The method for automated and precise control of the production process of macrolide derivatives according to claim 9, characterized in that: The method of dynamically comparing the time-varying state space model with the product yield prediction value is: Inputting a real-time process parameter characteristic matrix into the improved regression model to obtain a predicted product yield prediction value; Inputting a real-time dominant parameter sub-matrix into the time-varying state-space model to obtain an expected yield prediction value output by the time-varying state-space model; Compare the difference between the product yield prediction value and the expected yield prediction value to determine whether the difference is greater than a preset difference threshold.

11. The method for automated and precise control of the production process of macrolide derivatives according to claim 10, characterized in that: The method of generating the optimization control parameter set through the multi-objective optimization model is: The difference threshold between the predicted value of the comparison product yield and the predicted value of the expected yield is marked as h; The expected yield prediction value is marked as H; Based on the expected yield prediction value and the improved regression model, a multi-objective optimization function Yu is constructed; Based on the expected yield prediction value, the improved regression model and the difference threshold, a constraint condition set U is constructed; Taking minimization of multi-objective optimization function as optimization objective and constraint condition set as constraint condition, a multi-objective optimization model is constructed. By using the optimization algorithm to solve the multi-objective optimization model, the solution values ​​of the parameter variables to be solved for each key parameter are obtained, and the solution values ​​of all parameter variables constitute the optimization control parameter set.

12. An automated and precise control system for the production process of macrolide derivatives, which is used to implement the automated and precise control method for the production process of macrolide derivatives according to any one of claims 1 to 11, characterized in that: It includes a raw data collection module, a preprocessing module, a yield prediction module, a state space model training module, and a control parameter optimization module; wherein each module is electrically connected; The raw data collection module collects the multi-dimensional process parameter time series data of the reactor in real time through the distributed sensing system, synchronously obtains the raw material spectral characteristic data, and sends the multi-dimensional process parameter time series data and the raw material spectral characteristic data to the preprocessing module; The preprocessing module implements dynamic adaptive filtering processing on the multi-dimensional process parameter time series data to generate a denoising process parameter sequence, and at the same time performs baseline correction on the raw material spectral feature data to generate standard spectral data, and sends the denoising process parameter sequence and the standard spectral data to the yield prediction module; The yield prediction module performs feature fusion of the denoised process parameter sequence and the standard spectral data, constructs a process parameter feature matrix, inputs an improved regression model constrained by process knowledge, outputs key parameter sensitivity vectors and product yield prediction values, and sends the key parameter sensitivity vectors to the state space model training module, and sends the improved regression model and product yield prediction values ​​to the control parameter optimization module; The state space model training module extracts the dominant parameter sub-matrix from the process parameter feature matrix based on the key parameter sensitivity vector, constructs the time-varying state space model using the process timing alignment algorithm, and sends the time-varying state space model to the control parameter optimization module; The control parameter optimization module dynamically compares the time-varying state space model with the product yield prediction value. When the deviation exceeds the preset threshold, the multi-objective optimization model generates an optimized control parameter set, which is then input into the adaptive controller to generate execution instructions to control the process parameters.

Images