Intelligent chemical production real-time monitoring method, system and device

Abstract

The application provides an intelligent chemical production real-time monitoring method, system and equipment, relates to the technical field of production monitoring, and the method comprises the following steps: obtaining a historical working condition direction memory library of a reaction kettle circulating pipeline to perform training, obtaining an initial concentration analysis model to perform weight parameter space singular value decomposition, constituting a fixed main subspace, and fixing the same as an unupdateable part; when a new batch of raw materials is replaced, initial working condition data are collected to perform training, the model is updated without forgetting based on residual projection gradient, data of an online infrared spectrometer of the reaction kettle circulating pipeline are monitored and analyzed, and the technical problems that the existing technology has offline detection result feedback lag, the online analysis model is easy to forget historical working condition characteristics, and the concentration monitoring stability and reliability of the chemical production process are reduced are solved, and the technical effects that the model is resistant to forgetting and updating under multiple batches of working conditions and the long-term stability and reliability of the concentration measurement of the chemical process are improved are achieved.

Description

Technical Field

[0001] This invention relates to the field of production monitoring technology, specifically to intelligent real-time monitoring methods, systems, and equipment for chemical production. Background Technology

[0002] In chemical production processes, changes in material concentration in the reactor circulation pipeline are typically monitored by acquiring spectral data of the reaction system using an online near-infrared spectrometer. This data is then combined with a machine learning model to construct a mapping relationship between spectrum and concentration, enabling real-time prediction of the target component concentration. Simultaneously, process parameters such as online pH meters are used to further characterize the reaction state. Alternatively, offline high-performance liquid chromatography (HPLC) can be employed to detect samples and obtain highly accurate concentration or impurity content data.

[0003] However, while HPLC detection offers high precision, it suffers from long sampling cycles and results that lag behind the actual reaction process, making it difficult to meet the real-time process control requirements of continuous production. Although online near-infrared spectrometers can achieve continuous data acquisition, concentration prediction methods based on machine learning models typically rely on historical batch data to establish mapping relationships during the initial training phase. Over long-term operation, these models are susceptible to variations in raw material batches, fluctuations in process conditions, and spectral drift, leading to shifts in the distribution of input data and a gradual decline in model prediction performance.

[0004] Furthermore, in industrial applications, such predictive models typically require continuous incremental training or calibration updates using new batches of data to adapt to changing operating conditions. However, this update process can easily overwrite existing historical operating condition characteristics, causing the model to gradually lose its ability to recognize slowly evolving anomalies formed in the early stages. Simultaneously, conflicts between the distributions of new and old data can also lead to model parameter forgetting, resulting in a decrease in overall model stability and consequently affecting the long-term accuracy and reliability of concentration measurement results.

[0005] In summary, existing technologies suffer from technical problems such as delayed feedback of offline detection results, and the tendency for online analysis models to forget historical operating conditions during long-term operation and batch data updates, leading to a decline in the stability and reliability of concentration monitoring in chemical production processes. Summary of the Invention

[0006] The purpose of this application is to provide intelligent real-time monitoring methods, systems, and equipment for chemical production, in order to solve the technical problems of delayed feedback of offline detection results, and the easy forgetting of historical operating conditions during long-term operation and batch data updates of online analysis models, which leads to a decline in the stability and reliability of concentration monitoring in chemical production processes.

[0007] In view of the above problems, this application provides an intelligent method, system and equipment for real-time monitoring of chemical production.

[0008] The first aspect of this application provides an intelligent real-time monitoring method for chemical production. This method includes: acquiring a historical operating condition direction memory of the reactor circulation pipeline and training it based on the historical operating condition direction memory to obtain an initial concentration analysis model; performing singular value decomposition of the weight parameter space on the initial concentration analysis model, extracting principal singular vectors with a cumulative energy contribution rate greater than or equal to a preset threshold to form a fixed principal subspace, and fixing the fixed principal subspace as a non-updateable part; when a new batch of raw materials is replaced, collecting initial operating condition data of the reactor circulation pipeline in the initial production window, and training the initial concentration analysis model using the initial operating condition data to obtain a weight update gradient; orthogonally projecting the weight update gradient onto a pre-stored historical raw material batch direction memory to obtain a residual projection gradient, and performing forgetting-free adaptive updating of the initial concentration analysis model based on the residual projection gradient to obtain an updated concentration analysis model; monitoring and analyzing the data from an online infrared spectrometer of the reactor circulation pipeline based on the updated concentration analysis model and the fixed principal subspace, and triggering a monitoring anomaly alarm based on the analysis results.

[0009] Optionally, multiple absorbance data sets are obtained from multiple wavelength points output by an online near-infrared spectrometer on the reactor circulation pipeline during historical windows; multiple true values ​​of reaction liquid concentrations are obtained from offline high-performance liquid chromatography; the multiple absorbance data sets and the multiple true values ​​of reaction liquid concentrations are mapped and associated for storage to construct a historical operating condition direction memory.

[0010] Optionally, the weight matrix of each fully connected layer in the initial concentration analysis model is taken as the decomposition object, and singular value decomposition is performed on each weight matrix to obtain the singular value sequence and the left singular vector and right singular vector corresponding to each singular value. The sum of squares of the singular values ​​is accumulated in descending order. When the proportion of the accumulated sum of squares to the sum of squares of all singular values ​​is greater than or equal to a preset threshold, the accumulation is stopped, and the right singular vectors corresponding to the first few accumulated singular values ​​are extracted as the basis vectors of the fixed principal subspace. The subspace spanned by the basis vectors of the fixed principal subspace is taken as the fixed principal subspace.

[0011] Optionally, after each batch of raw materials is updated in the initial concentration analysis model, the weight change matrix of the low-rank update module during this update is extracted, wherein the weight change matrix is ​​the product of the upper projection matrix and the lower projection matrix; singular value decomposition is performed on the weight change matrix, and the left singular vector corresponding to the largest singular value is taken as the dominant direction vector of the corresponding batch of raw materials; the dominant direction vector, the supplier number of the batch of raw materials, the raw material warehousing date and the reactor batch number are added as a record to the historical raw material batch direction library.

[0012] Optionally, the multiple dominant direction vectors stored in the historical raw material batch direction library are subjected to Schmitt orthogonalization to obtain a set of orthonormal bases, wherein the orthonormal bases span the historical raw material subspace; the projection component of the weight update gradient in each basis direction in the orthonormal bases is calculated, and the projection component is multiplied by the corresponding basis and accumulated to obtain the total projection component; the total projection component is subtracted element by element from the weight update gradient to obtain the residual projection gradient.

[0013] Optionally, singular value decomposition is performed on the residual projected gradient, and the left singular vectors corresponding to the first four largest singular values ​​are extracted and shrunk into an N-row, 4-column lower projection matrix, where N is the dimension of the residual projected gradient; the value of the lower projection matrix is ​​fixed and does not participate in subsequent gradient descent optimization, while a 4-row, M-column upper projection matrix is ​​randomly initialized, where M is the model output dimension; the upper projection matrix is ​​optimized using the mean squared error loss function with the initial working condition data as the training set, and iterative training is performed until the loss converges; the trained upper projection matrix is ​​multiplied by the fixed lower projection matrix to obtain a low-rank update matrix, and the low-rank update matrix is ​​layered onto the corresponding weight matrix of the initial concentration analysis model to obtain the updated concentration analysis model.

[0014] Optionally, the real-time absorbance data of the online infrared spectrometer is input into the updated concentration analysis model to obtain real-time monitoring and analysis results; the projected energy on the fixed principal subspace is calculated based on the spectral data of the online infrared spectrometer; when the real-time monitoring and analysis results are greater than or equal to a preset concentration threshold and / or the projected energy does not meet the preset range constraint, a monitoring anomaly alarm is triggered.

[0015] Optionally, the preset range constraint is that the projected energy is greater than a first multiple threshold of the preset energy value and less than or equal to a second multiple threshold of the preset energy value.

[0016] A second aspect of this application provides an intelligent real-time monitoring system for chemical production. This system includes: a data training module for acquiring a historical operating condition direction memory of the reactor circulation pipeline and training an initial concentration analysis model based on the historical operating condition direction memory; a vector extraction module for performing singular value decomposition of the weight parameter space on the initial concentration analysis model, extracting principal singular vectors with a cumulative energy contribution rate greater than or equal to a preset threshold to form a fixed principal subspace, and fixing the fixed principal subspace as a non-updateable part; and a model training module for, when a new batch of raw materials is replaced, during the initial production window... The system collects initial operating condition data of the reactor circulation pipeline and uses this data to train an initial concentration analysis model to obtain a weight update gradient. A model update module orthogonally projects this weight update gradient onto a pre-stored historical raw material batch direction library to obtain a residual projection gradient. Based on this residual projection gradient, the initial concentration analysis model is adaptively updated without forgetting to obtain an updated concentration analysis model. A monitoring and analysis module monitors and analyzes the data from the online infrared spectrometer of the reactor circulation pipeline based on the updated concentration analysis model and a fixed principal subspace, triggering an alarm for any abnormalities based on the analysis results.

[0017] A third aspect of this application provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the aforementioned intelligent real-time monitoring method for chemical production.

[0018] One or more technical solutions provided in this application have at least the following technical effects or advantages: The method provided in this application obtains a historical operating condition direction memory of the reactor circulation pipeline and trains it based on this memory to obtain an initial concentration analysis model. The initial concentration analysis model undergoes weight parameter space singular value decomposition to extract principal singular vectors with a cumulative energy contribution rate greater than or equal to a preset threshold, forming a fixed principal subspace, which is then fixed as a non-updateable portion. When a new batch of raw materials is replaced, initial operating condition data of the reactor circulation pipeline is collected during the initial production window, and the initial concentration analysis model is trained using this data to obtain a weight update gradient. This weight update gradient is orthogonally projected onto a pre-stored historical raw material batch direction memory to obtain a residual projection gradient. Based on this residual projection gradient, the initial concentration analysis model is adaptively updated without forgetting to obtain an updated concentration analysis model. The updated concentration analysis model and the fixed principal subspace are used to monitor and analyze the data from the online infrared spectrometer of the reactor circulation pipeline. An alarm for abnormal monitoring is triggered based on the analysis results. This achieves the technical effect of constraining the model update direction through a fixed principal subspace, enabling model anti-forgetting updates under multiple batch operating conditions, and improving the long-term stability and reliability of concentration measurement in chemical processes.

[0019] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 A flowchart illustrating the intelligent real-time monitoring method for chemical production provided in this application.

[0022] Figure 2 A schematic diagram of the structure of the intelligent real-time monitoring system for chemical production provided in this application.

[0023] Figure 3 A schematic diagram of the structure of an exemplary electronic device provided in this application.

[0024] Figure labeling: Data training module 11, Vector extraction module 12, Model training module 13, Model update module 14, Monitoring and analysis module 15, Bus 300, Receiver 301, Processor 302, Transmitter 303, Memory 304, Bus interface 305. Detailed Implementation

[0025] This application provides an intelligent real-time monitoring method, system, and equipment for chemical production, addressing the technical problems of delayed feedback of offline detection results and the tendency of online analysis models to forget historical operating conditions during long-term operation and batch data updates, leading to a decline in the stability and reliability of concentration monitoring in chemical production processes. It achieves the technical effect of using a fixed principal subspace to constrain the model update direction, enabling model updates to resist forgetting under multiple batch operating conditions and improving the long-term stability and reliability of concentration measurement in chemical processes. This allows for real-time and accurate monitoring of concentrations under multiple batch raw material conditions, continuous adaptive model updates, and timely and accurate identification of anomalies during chemical production.

[0026] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0027] Example 1, as Figure 1 As shown, this application provides an intelligent real-time monitoring method for chemical production, which includes: A historical operating condition direction memory database of the reactor circulation pipeline is obtained, and an initial concentration analysis model is obtained by training based on the historical operating condition direction memory database.

[0028] Furthermore, a historical operating condition orientation memory library is obtained for the reactor circulation pipeline, including: obtaining multiple absorbance data sets of multiple wavelength points output by the online near-infrared spectrometer on the reactor circulation pipeline during multiple production batches in the historical window; obtaining multiple true value sets of reaction liquid concentrations simultaneously measured by offline high-performance liquid chromatography; mapping and associating the multiple absorbance data sets and the multiple true value sets of reaction liquid concentrations for storage, and constructing a historical operating condition orientation memory library.

[0029] Specifically, using an online near-infrared spectrometer deployed on the reactor's circulation pipeline, multiple absorbance data sets corresponding to multiple wavelength points from multiple production batches within a historical window are acquired. The historical window refers to the time range used to construct training data. This time range is selected as a stable operating cycle within a continuous industrial production process, such as 1 week, 3 weeks, or 1 month, to limit the data acquisition scope. This ensures that the selected data simultaneously covers the normal operation of the equipment and the switching processes between different batches, thereby guaranteeing the representativeness and diversity of the sample distribution. Multiple production batches refer to multiple independent feeding and reaction cycles produced continuously or intermittently within the historical window's reactor circulation pipeline. Each production batch corresponds to one complete process of raw material addition, reaction, and product formation. Different production batches typically exhibit differences in raw material sources, formulation ratios, or process parameters, thus resulting in different spectral response characteristic distributions. Multiple wavelength point sets refer to the set of discrete wavelength sequences obtained by an online near-infrared spectrometer sampling at fixed intervals within a preset wavelength range, such as 800nm ​​to 2500nm. For example, 128 or 256 equally or non-equally spaced wavelength points are selected. Each wavelength point corresponds to a specific optical absorption response dimension, which is used to characterize the absorption characteristics of the reaction system at different energy frequency bands. The absorbance data set refers to the data set consisting of the absorbance values ​​corresponding to each wavelength point at each sampling time point. The absorbance represents the logarithmic transformation result of the ratio of incident light intensity to transmitted light intensity, which is used to reflect the absorption capacity of materials for near-infrared light of a specific wavelength.

[0030] Meanwhile, under the same production batch and time synchronization conditions, multiple reaction solutions were sampled and analyzed by offline high performance liquid chromatography (HPLC). Offline HPLC involves injecting reaction solution samples into an HPLC instrument and using the differences in retention time of different components in the chromatographic column for separation and detection, thereby obtaining the true concentration values ​​of the corresponding batch of reaction solutions and forming a set of true concentration values ​​of multiple reaction solutions. It has high measurement accuracy and can be used as a supervised learning label.

[0031] Multiple absorbance data sets output from an online near-infrared spectrometer and corresponding time-synchronized sets of multiple true concentration values ​​of reaction solutions are mapped and stored to construct a historical operating condition orientation memory. Each record in the historical operating condition orientation memory contains at least: absorbance data, corresponding reaction solution concentration, and the identification information of the production batch to which it belongs. The historical operating condition orientation memory is used as training data for training a one-dimensional convolutional neural network (CNN). The CNN structure includes an input layer, three one-dimensional convolutional layers, two fully connected layers, and an output layer. The input layer receives a spectral absorbance vector of length 128 or 256. The first convolutional layer has 32 kernels with a kernel size of 5 and a stride of 1, using ReLU activation function to extract local spectral features. The second convolutional layer has 64 kernels with a kernel size of 3 to enhance the mid-layer feature representation capability. The third convolutional layer has 128 kernels and batch normalization is introduced to stabilize the training process. The features are flattened by the Flatten layer and then fed into the fully connected layer. The first fully connected layer has 128 nodes, the second fully connected layer has 64 nodes, and the activation function is ReLU. The final output layer has 1 node and is used to output continuous concentration prediction values.

[0032] During model training, mean squared error (MSE) is used as the loss function to measure the deviation between the predicted concentration value and the actual concentration value measured by offline high-performance liquid chromatography (HPLC). It is expressed as the average of the squared differences between the predicted and actual concentration values. The optimization algorithm employs the Adam optimizer, which has adaptive learning rate adjustment capabilities, improving the convergence stability of the nonlinear model on complex spectral data. The initial learning rate is set to 0.001, and an exponential decay strategy is used for dynamic adjustment with a decay coefficient of 0.95. A decay update is performed every few training epochs, e.g., 10 epochs. Batch gradient descent is used during training, with a batch size of 32 to balance training stability and computational efficiency. The model convergence criterion is set as follows: training stops when the validation set loss decreases by less than a set threshold (e.g., 0.001) within 10 consecutive epochs, or when the maximum number of training epochs reaches a preset upper limit (e.g., 100 epochs). Simultaneously, the model parameters corresponding to the minimum validation set error are saved as the initial concentration analysis model. This initial concentration analysis model can input online spectral data at any time and output the corresponding concentration prediction results.

[0033] By acquiring a historical operating condition direction memory database of the reactor circulation pipeline and training an initial concentration analysis model based on this database, a stable mapping relationship between spectral morphology changes and concentration can be learned, providing an initial reliable basis for accurately predicting the concentration of the reaction solution and timely detecting anomalies.

[0034] The initial concentration analysis model is subjected to weight parameter space singular value decomposition to extract the principal singular vectors with a cumulative energy contribution rate greater than or equal to a preset threshold to form a fixed principal subspace, and the fixed principal subspace is fixed as a non-updateable part.

[0035] Furthermore, the initial concentration analysis model is subjected to singular value decomposition of the weight parameter space to extract principal singular vectors with a cumulative energy contribution rate greater than or equal to a preset threshold to form a fixed principal subspace. This includes: taking the weight matrix of each fully connected layer in the initial concentration analysis model as the decomposition object, performing singular value decomposition on each weight matrix to obtain a sequence of singular values ​​and the left and right singular vectors corresponding to each singular value; accumulating the sum of squares of the singular values ​​in descending order, stopping the accumulation when the proportion of the accumulated sum of squares to the total sum of squares of all singular values ​​is greater than or equal to a preset threshold, and extracting the right singular vectors corresponding to the first few accumulated singular values ​​as the basis vectors of the fixed principal subspace; and using the subspace spanned by the basis vectors of the fixed principal subspace as the fixed principal subspace.

[0036] Specifically, the weight matrix of each fully connected layer in the initial concentration analysis model is taken as an independent decomposition object. The weight matrix of the fully connected layer comes from the parameter set in the neural network structure after the initial concentration analysis model is trained. Specifically, it is the linear mapping parameter formed when the one-dimensional convolutional neural network enters the fully connected layer after convolutional feature extraction and flattening. The weight matrix is ​​automatically learned by the backpropagation algorithm during the model training process. Its parameter dimension is determined by the number of input nodes n and the number of output nodes m of the layer, forming a weight matrix W with dimension m×n, where m is the output feature dimension and n is the input feature dimension.

[0037] Singular value decomposition (SVD) is performed on the weight matrix of each fully connected layer. SVD is a method that decomposes a matrix into a product of three matrices, i.e., W = UΣV. T Where U is an m×m left singular vector matrix, with each column representing a left singular vector; Σ is an m×n diagonal matrix, with the diagonal elements being singular values ​​arranged in descending order; and V is an n×n right singular vector matrix, with each column representing a right singular vector. The magnitude of each singular value reflects the importance of the corresponding singular vector in the weight matrix. The sum of squares of the singular values ​​is accumulated sequentially in descending order, starting from the first singular value, and the cumulative energy contribution rate is calculated. The numerator of the cumulative energy contribution rate is the sum of squares of the first i singular values, and the denominator is the sum of squares of all singular values. This energy contribution rate measures the interpretability of the currently selected subspace for the original weight information.

[0038] When the cumulative sum of squares accounts for a proportion of the sum of squares of all singular values, i.e., the cumulative energy contribution rate is greater than or equal to a preset threshold, such as 90% or 95%, accumulation stops, and the right singular vectors corresponding to the first few accumulated singular values ​​are extracted as the basis vectors of the fixed principal subspace. The fixed principal subspace refers to the linear subspace spanned by the selected right singular vectors. This fixed principal subspace includes the characteristic absorption peak positions of functional groups in the material or the physical conservation relationships of chemical units, representing stable and unchanging chemical mechanism information across batches. The subspace spanned by the basis vectors of the fixed principal subspace is defined as the fixed principal subspace and set as a non-updateable part. That is, in all subsequent model training or parameter update processes, gradient updates or reconstruction adjustments of the weight projection components in the direction of this subspace are prohibited to ensure that model updates do not destroy the learned stable chemical laws, thereby ensuring the accuracy and stability of the model's prediction of the reaction solution concentration.

[0039] By performing singular value decomposition of the weight parameter space on the initial concentration analysis model, the principal singular vectors with a cumulative energy contribution rate greater than or equal to a preset threshold are extracted to form a fixed principal subspace. This fixed principal subspace is then fixed as a non-updateable part, ensuring that the initial concentration analysis model does not lose features that are important representatives of historical operating conditions during the update process. This is beneficial for the long-term stable and reliable monitoring of the chemical production process by the model.

[0040] When a new batch of raw materials is replaced, the initial operating condition data of the reactor circulation pipeline is collected in the initial production window, and the initial concentration analysis model is trained using the initial operating condition data to obtain the weight update gradient.

[0041] Specifically, when a new batch of raw materials is introduced, a batch switching adaptation phase is entered. During this phase, an initial production window is defined. This initial production window refers to a short time interval, such as 0-30 minutes or 0-2 hours, after the new batch of raw materials is added to the reactor but before the process reaches steady state. Its purpose is to capture the early dynamic characteristics of the system's spectral response and concentration changes after the introduction of the new raw materials. Within this initial production window, initial operating condition data of the reactor's circulation pipeline is collected. This initial operating condition data includes multiple initial absorbance data and multiple initial true values ​​of the reaction solution concentration. Specifically, spectral data is continuously collected by an online near-infrared spectrometer deployed on the reactor's circulation pipeline, obtaining absorbance data corresponding to multiple time sampling points. Each sampling point consists of discrete wavelength points within a preset wavelength range, thus forming multiple initial absorbance data. This absorbance data reflects the changes in the absorption characteristics of the new batch of raw materials to near-infrared light in the early stages of the reaction. Simultaneously, corresponding reaction solution samples are obtained through offline sampling, and the actual concentration values ​​of the target components are determined using high-performance liquid chromatography (HPLC), forming multiple initial true values ​​of the reaction solution concentration. These are used as label data for model supervised learning, ensuring the accuracy of the training direction.

[0042] During the model update phase, initial absorbance data is used as input, and the initial true concentration value is used as a supervision signal. These are fed into the trained initial concentration analysis model for forward propagation to obtain the predicted output value. A loss function is calculated based on the deviation between the predicted and true concentration values, using mean squared error as the error metric to measure the degree of prediction deviation. Then, the gradient derivative of the initial concentration analysis model parameters is calculated using the backpropagation algorithm to obtain the gradient change of each layer's weight parameters relative to the loss function. This gradient is the weight update gradient, reflecting the direction and intensity of the initial concentration analysis model's adjustment under the influence of the new batch of data.

[0043] By collecting initial operating condition data of a new batch of raw materials and training the initial concentration analysis model to obtain the weight update gradient, the model can learn the new mapping relationship between the spectral morphology and the concentration of the reaction solution under the new batch of raw materials. This allows the concentration analysis model to adapt to the changes in the new batch of raw materials, improve the prediction accuracy and stability of the concentration analysis model under the new operating conditions, and provide a reliable guarantee for accurately predicting the concentration of the reaction solution and timely detecting anomalies.

[0044] The weight update gradient is orthogonally projected onto the pre-stored historical raw material batch direction library to obtain the residual projection gradient. Based on the residual projection gradient, the initial concentration analysis model is updated without forgetting to obtain the updated concentration analysis model.

[0045] Furthermore, the method includes: after each batch of raw materials is updated in the initial concentration analysis model, extracting the weight change matrix of the low-rank update module during the update process, wherein the weight change matrix is ​​the product of the upper projection matrix and the lower projection matrix; performing singular value decomposition on the weight change matrix, and taking the left singular vector corresponding to the largest singular value as the dominant direction vector of the corresponding batch of raw materials; adding the dominant direction vector, the supplier number of the batch of raw materials, the raw material warehousing date, and the reactor batch number as a record into the historical raw material batch direction library.

[0046] Specifically, after each batch of raw materials is trained and updated in the initial concentration analysis model, parameters are extracted from the low-rank update module at the model parameter saving and version switching node. The low-rank update module refers to a parameter decomposition update structure introduced when the original fully connected layer weight matrix is ​​not directly updated as a whole. That is, the weight update term is represented as the product of two low-dimensional matrices, ΔW=A×B, used to characterize the model change trend caused by the new batch of data with a lower number of parameters. The weight change matrix is ​​the product of the upper projection matrix and the lower projection matrix. The upper projection matrix B is an r×n dimensional parameter matrix used to map low-dimensional latent variables to the original feature space direction, and the lower projection matrix A is an m×r dimensional parameter matrix used to compress high-dimensional gradient information into a low-dimensional representation space. r is a low-rank dimension much smaller than m and n, thus achieving a compressed expression of the weight update.

[0047] During the extraction process, when a batch of training is completed and parameter solidification is triggered, the low-rank update module parameters before and after training for that batch are first read from the model parameter cache, i.e., (A0, B0) before the update and (A1, B1) after the update. Since the low-rank update module adopts an incremental update mechanism, the weight change matrix is ​​not directly stored, but is obtained by calculating the reconstruction difference of the low-rank decomposition parameters at two time points. That is, the weight contribution before the update ΔW0 = A0 × B0 and the weight contribution after the update ΔW1 = A1 × B1 are calculated respectively, and then the weight change matrix ΔW = ΔW1 - ΔW0 for this batch is obtained by matrix subtraction. The equivalent notation can be simplified to ΔW = A × B, where A and B represent the low-rank parameter combinations corresponding to the final update increment of this batch.

[0048] Singular value decomposition is performed on the weight change matrix. Singular values ​​are used to characterize the importance of information in different directions. The left singular vector corresponding to the largest singular value represents the most important change direction in the weight change of this batch, i.e., the dominant direction vector. The dominant direction vector represents the most significant change trend of the model parameters in the update process of this batch. In industrial semantics, it can be understood as the characteristic direction that has the strongest influence on the spectral concentration mapping relationship of the raw materials in this batch, such as the dominant perturbation direction caused by the enhancement of the absorption peak of a certain type of functional group or the change in reaction rate.

[0049] The obtained dominant direction vector is associated with the supplier number, raw material receipt date, and reactor batch number of that batch of raw materials. The supplier number uniquely identifies the raw material supplier, the raw material receipt date records the time the raw material entered the factory, and the reactor batch number identifies the batch of reactors using that raw material. This information is added as a record to the historical raw material batch direction vector database. This database stores information related to different batches of raw materials, and by continuously accumulating this data, a wealth of historical references is provided for analyzing raw material quality fluctuations and model performance changes.

[0050] By extracting the weight change matrix of the low-rank update module and performing singular value decomposition to obtain the dominant direction vector, and associating it with the supplier number, warehousing date, and reactor batch number of the raw materials, and storing it in the historical raw material batch direction database, a connection between raw material characteristics and model updates is established. This not only helps to trace the reasons for changes in model performance, such as when the model prediction accuracy decreases, by querying the historical raw material batch direction database, it can be analyzed whether the model is unsuitable due to raw materials from certain specific suppliers or raw materials that were warehoused at a specific time, but also provides data support for model optimization.

[0051] Furthermore, the weight update gradient is orthogonally projected onto a pre-stored historical raw material batch direction library to obtain a residual projection gradient. This includes: performing Schmitt orthogonalization on multiple dominant direction vectors stored in the historical raw material batch direction library to obtain a set of orthonormal bases, wherein the orthonormal bases span a historical raw material subspace; calculating the projection component of the weight update gradient in each basis direction of the orthonormal bases, and multiplying the projection component by the corresponding basis and summing them to obtain a total projection component; and subtracting the total projection component element by element from the weight update gradient to obtain the residual projection gradient.

[0052] Specifically, the historical raw material batch orientation database stores the dominant orientation vectors corresponding to multiple batches of raw materials, reflecting the main influence directions of different batches of raw materials on the weight adjustment of the low-rank update module during the model update process. The multiple dominant orientation vectors in the historical raw material batch orientation database are subjected to Schmitt orthogonalization. Schmitt orthogonalization is a mathematical method that transforms a linearly correlated vector group into a mutually orthogonal unit vector group, used to eliminate the linear correlation between orientations of different batches. For example, suppose the historical raw material batch orientation database stores K dominant orientation vectors, denoted as u1, u2, ..., u... K Each vector has the same dimension as the model parameter space, which is N-dimensional. First, the first dominant direction vector u1 is taken and normalized using its own norm to obtain the first orthonormal basis e1 = u1 / ||u1||. Then, for the second dominant direction vector u2, its projection component proj1(u2) = (u2·e1)e1 on the existing orthonormal basis e1 is calculated, and this projection component is subtracted from u2 to obtain the residual vector r2 = u2 - proj1(u2) orthogonal to e1. ​​r2 is then normalized to obtain the second orthonormal basis e2 = r2 / ||r2||. For the i-th dominant direction vector ui, its projection components on the first i-1 orthonormal bases e1 to e1 are calculated sequentially. i-1 The projection on the surface is subtracted item by item, i.e., r i =u i -∑[(u i ·e j )ej (j=1 to i-1), if the residual vector r i If the norm is not zero, then further normalization yields e. i =r i / ||r i ||.

[0053] Repeat the above process until all dominant direction vectors have been processed, resulting in a set of orthonormal bases that are orthogonal to each other and have a magnitude of 1. After obtaining the orthonormal bases, all linear combinations of the orthonormal bases are defined as the historical base subspace, used to represent the space of linear combinations of all operating condition change directions learned by the model in historical batches. The current weight update gradient is then projected onto the orthonormal bases one by one, i.e., the scalar projection coefficient α of the weight update gradient in each base direction is calculated respectively. i The calculation method is α i =g·e i , where “·” represents the vector dot product operation, e i For an orthonormal basis, the scalar projection coefficients represent the contribution of the current gradient in the i-th historical direction. Then, the scalar projection coefficients α... i Multiplying by the corresponding basis vectors yields the projection components p in each direction. i =α i e i Then, all projection components are summed to form a total projection component, which represents the portion of the current weight update gradient in the historical raw material subspace.

[0054] The residual projection gradient is obtained by subtracting the total projection component from the weight update gradient element by element. The residual projection gradient is orthogonal to all basis directions in the historical raw material subspace, that is, its inner product is zero. This means that the residual projection gradient does not contain the directional components learned in the historical batch, but only retains the newly added change information in the new batch that has not been covered by the historical data.

[0055] By using orthogonal projection, components related to historical operating conditions are removed from the current model update direction, retaining only independent information related to changes in new batches of raw materials or processes. This effectively avoids the model repeatedly fitting existing knowledge during continuous learning across multiple batches, further improving the model's adaptability to new batches of raw materials and its prediction accuracy.

[0056] Furthermore, based on the residual projection gradient, the initial concentration analysis model is updated adaptively without forgetting to obtain an updated concentration analysis model. This includes: performing singular value decomposition on the residual projection gradient, extracting the left singular vectors corresponding to the first four largest singular values ​​and shrinking them into an N-row, 4-column lower projection matrix, where N is the dimension of the residual projection gradient; fixing the value of the lower projection matrix, not participating in subsequent gradient descent optimization, and randomly initializing a 4-row, M-column upper projection matrix, where M is the model output dimension; using the initial working condition data as the training set, optimizing the upper projection matrix using the mean squared error loss function, iterating until the loss converges; multiplying the trained upper projection matrix with the fixed lower projection matrix to obtain a low-rank update matrix, and layer-by-layer superimposing the low-rank update matrix onto the corresponding weight matrix of the initial concentration analysis model to obtain the updated concentration analysis model.

[0057] Specifically, after obtaining the residual projection gradient, the model enters a forgetting-free adaptive update stage. This stage incrementally adapts the initial concentration analysis model using only the newly added change information from the current batch, without disrupting historical stable knowledge. Singular value decomposition is then performed on the residual projection gradient to obtain the gradient matrix G. res =UΣV T Let U be the left singular vector matrix, Σ be the singular value matrix, and V be the right singular vector matrix. After sorting the singular values ​​in descending order, the left singular vectors corresponding to the top four largest singular values ​​are extracted. These four vectors are then concatenated column-wise to form an N x 4 column downprojection matrix A, where N is the dimension of the residual projection gradient. The downprojection matrix is ​​a fixed mapping structure used to compress high-dimensional parameter changes into a low-dimensional subspace; its function is to lock the main direction of the current new change. By selecting the top four main directions, the features of the residual projection gradient are reduced in dimension and filtered, preserving the main information in the residual gradient while suppressing noise perturbations.

[0058] The values ​​of the downprojection matrix are fixed and do not participate in subsequent gradient descent optimization, meaning that its parameters are not allowed to change during training, thus ensuring that the direction of the new changes remains unchanged. Simultaneously, a 4xM upprojection matrix B is randomly initialized, where M is the output dimension of the corresponding layer. This upprojection matrix is ​​used to remap the low-dimensional update representation back to the original model output space, serving as the parameter adaptation learning function.

[0059] During the parameter optimization phase, the initial operating condition data obtained in the current new batch is used as the training set, where the input is the initial absorbance data and the supervision label is the corresponding ground truth concentration. Forward calculation is performed through the initial concentration analysis model, and the mean squared error is used as the loss function. Gradient optimization is performed only on the projection matrix B. The optimization algorithm can use the Adam optimizer. The initial learning rate is set to 0.001 for example, and the batch size is 32. During the training process, only B is updated while A is frozen, so that the initial concentration analysis model only learns the adaptation magnitude in the newly added change direction without changing the change direction itself.

[0060] Iterative training continues until the loss converges; for example, training stops when the validation set loss decreases by less than 0.001 for 10 consecutive rounds. After training, the optimized upper projection matrix B is multiplied by the fixed lower projection matrix A to generate a low-rank update matrix ΔW = A × B. This low-rank update matrix is ​​then layer-by-layer superimposed onto the weight matrices of the corresponding layers of the initial concentration analysis model to form updated weights, thus obtaining the updated concentration analysis model. Since the update is achieved only through the superposition of low-rank increments, and the original weight structure is not directly overwritten, the catastrophic forgetting problem caused by traditional full-parameter fine-tuning can be avoided.

[0061] By transforming the residual projection gradient into a low-rank parameter adaptation structure, and performing restricted updates only in the orthogonal directions of historical knowledge, rapid adaptation to changes in operating conditions of new batches can be achieved. At the same time, the stable operating conditions learned by the initial model are preserved, thereby achieving a forgetting-free adaptive update effect that learns new knowledge without forgetting old knowledge. This improves the stability of long-term online operation of the concentration analysis model and the reliability of continuous monitoring, providing a more reliable basis for decision-making in the production process.

[0062] Based on the updated concentration analysis model and the fixed principal subspace, the data from the online infrared spectrometer of the reactor circulation pipeline are monitored and analyzed, and an alarm for abnormal monitoring is triggered according to the analysis results.

[0063] Furthermore, based on the updated concentration analysis model and the fixed principal subspace, the data from the online infrared spectrometer in the reactor circulation pipeline are monitored and analyzed. An alarm for monitoring anomalies is triggered based on the analysis results, including: inputting the real-time absorbance data from the online infrared spectrometer into the updated concentration analysis model to obtain real-time monitoring and analysis results; calculating the projected energy in the fixed principal subspace based on the spectral data from the online infrared spectrometer; and triggering an alarm for monitoring anomalies when the real-time monitoring and analysis results are greater than or equal to a preset concentration threshold and / or the projected energy does not meet a preset range constraint.

[0064] Furthermore, the preset range constraint is that the projected energy is greater than a first multiple threshold of the preset energy value and less than or equal to a second multiple threshold of the preset energy value.

[0065] Specifically, the online infrared spectrometer continuously collects real-time absorbance data of the reaction liquid in the circulating pipeline of the reactor according to a preset sampling period. This real-time absorbance data is then input into the updated concentration analysis model to perform forward inference calculations, outputting the predicted real-time concentration of the target component as the real-time monitoring and analysis result. Simultaneously, the spectral data from the online infrared spectrometer at the same time are used to perform fixed principal subspace projection energy calculations. The fixed principal subspace originates from the set of principal singular vectors extracted from the singular value decomposition of the initial concentration analysis model's weight space, representing the main chemical mechanism characteristic directions under historical stable operating conditions. During the calculation, the spectral data is denoted as x, and the standard basis vectors of the fixed principal subspace are denoted as v1, v2, ..., v... k Calculate the projection coefficient α of x in each base direction. i =x·v i Then, calculate the sum of squares of each projection component to obtain the projected energy E=∑(α i 2 The projected energy is used to characterize the response intensity of the current spectrum in the steady-state characteristic space.

[0066] Pre-set concentration thresholds are established based on the target product's process specifications, such as upper and lower limits or alarm boundary values. When the real-time predicted concentration is greater than or equal to the preset concentration threshold, an abnormal concentration risk is identified. Simultaneously, preset range constraints are set. These constraints are defined as the projected energy being greater than a first multiple threshold and less than or equal to a second multiple threshold. The preset energy value is the historical average, calculated by averaging the projected energy from a large amount of historical spectral data under normal production conditions. The first and second multiple thresholds are set based on actual production conditions and data fluctuations to determine whether the current projected energy is within the normal range. The first multiple threshold is less than the second multiple threshold; for example, the first multiple threshold is set to 0.8, and the second multiple threshold is set to 1.2.

[0067] When the real-time monitoring and analysis results are greater than or equal to the preset concentration threshold and / or the projected energy does not meet the preset range constraints, a monitoring anomaly alarm is triggered to remind operators to take timely measures to avoid production accidents. The "and / or" clause indicates that an alarm can be triggered if either condition is met. This system can identify both concentration-exceeding anomalies and potential anomalies where the operating structure deviates from the normal pattern even if the concentration does not exceed the limit, thus achieving dual monitoring of both result-level and structural-level anomalies.

[0068] By utilizing an updated concentration analysis model and a fixed principal subspace to perform dual analysis on the data from the online infrared spectrometer, it is possible to more comprehensively and accurately monitor the changes in the concentration and spectral characteristics of substances in the reactor circulation pipeline over a long period of time. This improves the accuracy and reliability of real-time monitoring in chemical production, enables the timely detection of potential production anomalies, reduces the risk of false alarms and missed alarms, and provides strong support for the safe and stable operation of the production process.

[0069] Example 2, based on the same inventive concept as the intelligent real-time monitoring method for chemical production in the foregoing examples, such as... Figure 2 As shown, this application provides an intelligent real-time monitoring system for chemical production, wherein the intelligent real-time monitoring system for chemical production includes: The data training module 11 is used to acquire the historical operating condition direction memory of the reactor circulation pipeline and train the model based on the historical operating condition direction memory to obtain an initial concentration analysis model. The vector extraction module 12 is used to perform singular value decomposition of the weight parameter space on the initial concentration analysis model, extract the principal singular vectors with a cumulative energy contribution rate greater than or equal to a preset threshold to form a fixed principal subspace, and fix the fixed principal subspace as a non-updateable part. The model training module 13 is used to collect the initial operating condition data of the reactor circulation pipeline in the initial production window when a new batch of raw materials is replaced, and train the initial concentration analysis model using the initial operating condition data to obtain the weight update gradient. The model update module 14 is used to orthogonally project the weight update gradient onto the pre-stored historical raw material batch direction memory to obtain the residual projection gradient, and perform forgetting-free adaptive update of the initial concentration analysis model based on the residual projection gradient to obtain an updated concentration analysis model. The monitoring and analysis module 15 is used to monitor and analyze the data of the online infrared spectrometer of the reactor circulation pipeline based on the updated concentration analysis model and the fixed principal subspace, and trigger a monitoring anomaly alarm based on the analysis results.

[0070] Furthermore, the data training module 11 is also used to: acquire multiple absorbance data sets of multiple wavelength points output by the online near-infrared spectrometer on the reactor circulation pipeline in the historical window; acquire multiple true value sets of reaction liquid concentrations simultaneously measured by offline high performance liquid chromatography; map and associate the multiple absorbance data sets and the multiple true value sets of reaction liquid concentrations to construct a historical operating condition direction memory library.

[0071] Furthermore, the vector extraction module 12 is also used to: take the weight matrix of each fully connected layer in the initial concentration analysis model as the decomposition object, perform singular value decomposition on each weight matrix to obtain a sequence of singular values ​​and the left singular vector and right singular vector corresponding to each singular value; accumulate the sum of squares of the singular values ​​in descending order, and stop accumulating when the proportion of the accumulated sum of squares to the sum of squares of all singular values ​​is greater than or equal to a preset threshold, and extract the right singular vectors corresponding to the first few accumulated singular values ​​as the basis vectors of the fixed principal subspace; and take the subspace spanned by the basis vectors of the fixed principal subspace as the fixed principal subspace.

[0072] Furthermore, the model update module 14 is also used to: after the initial concentration analysis model completes the update of each batch of raw materials, extract the weight change matrix of the low-rank update module during the update process, wherein the weight change matrix is ​​the product of the upper projection matrix and the lower projection matrix; perform singular value decomposition on the weight change matrix, and take the left singular vector corresponding to the largest singular value as the dominant direction vector of the corresponding batch of raw materials; add the dominant direction vector, the supplier number of the batch of raw materials, the raw material warehousing date and the reactor batch number as a record into the historical raw material batch direction library.

[0073] Furthermore, the model update module 14 is also used to: perform Schmitt orthogonalization on multiple dominant direction vectors stored in the historical raw material batch direction library to obtain a set of orthonormal bases, wherein the orthonormal bases span the historical raw material subspace; calculate the projection component of the weight update gradient in each basis direction of the orthonormal bases, and multiply the projection component by the corresponding basis and accumulate it to obtain the total projection component; subtract the total projection component element by element from the weight update gradient to obtain the residual projection gradient.

[0074] Furthermore, the model update module 14 is also used to: perform singular value decomposition on the residual projection gradient, extract the left singular vectors corresponding to the first four largest singular values ​​and shrink them into an N-row, 4-column lower projection matrix, where N is the dimension of the residual projection gradient; fix the value of the lower projection matrix and not participate in subsequent gradient descent optimization, while randomly initializing a 4-row, M-column upper projection matrix, where M is the model output dimension; use the initial working condition data as the training set, optimize the upper projection matrix using the mean squared error loss function, and iterate training until the loss converges; multiply the trained upper projection matrix with the fixed lower projection matrix to obtain a low-rank update matrix, and stack the low-rank update matrix layer by layer onto the corresponding weight matrix of the initial concentration analysis model to obtain an updated concentration analysis model.

[0075] Furthermore, the monitoring and analysis module 15 is also used to: input the real-time absorbance data of the online infrared spectrometer into the updated concentration analysis model to obtain real-time monitoring and analysis results; calculate the projected energy on the fixed principal subspace based on the spectral data of the online infrared spectrometer; and trigger a monitoring anomaly alarm when the real-time monitoring and analysis results are greater than or equal to a preset concentration threshold and / or the projected energy does not meet the preset range constraint.

[0076] Furthermore, the monitoring and analysis module 15 is also used to: constrain the preset range to a first multiple threshold of the projected energy value that is greater than the preset energy value and a second multiple threshold of the preset energy value.

[0077] Example 3: Based on the same inventive concept as the intelligent real-time monitoring method for chemical production in the foregoing examples, this application also provides an electronic device, including: at least one processor; a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the intelligent real-time monitoring method for chemical production described in any one of Examples 1 above.

[0078] Appendix Figure 3 This is a schematic diagram of the structure of an exemplary electronic device of this application. Figure 3 In this document, the bus architecture is represented by bus 300. Bus 300 may include any number of interconnected buses and bridges, and bus 300 connects various circuits including one or more processors represented by processor 302 and memory represented by memory 304. Bus 300 may also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 305 provides an interface between bus 300 and receiver 301 and transmitter 303. Receiver 301 and transmitter 303 may be the same element, i.e., a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 302 is responsible for managing bus 300 and general processing, while memory 304 can be used to store data used by processor 302 during operation.

[0079] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0080] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. An intelligent real-time monitoring method for chemical production, characterized in that, The method includes: The historical operating condition direction memory of the reactor circulation pipeline is obtained, and the initial concentration analysis model is obtained by training based on the historical operating condition direction memory. The initial concentration analysis model is subjected to weight parameter space singular value decomposition to extract the principal singular vectors with a cumulative energy contribution rate greater than or equal to a preset threshold to form a fixed principal subspace, and the fixed principal subspace is fixed as a non-updateable part. When a new batch of raw materials is replaced, the initial operating condition data of the reactor circulation pipeline is collected in the initial production window, and the initial concentration analysis model is trained using the initial operating condition data to obtain the weight update gradient. The weight update gradient is orthogonally projected onto the pre-stored historical raw material batch direction library to obtain the residual projection gradient. Based on the residual projection gradient, the initial concentration analysis model is updated without forgetting to obtain the updated concentration analysis model. Based on the updated concentration analysis model and the fixed principal subspace, the data from the online infrared spectrometer of the reactor circulation pipeline are monitored and analyzed, and an alarm for abnormal monitoring is triggered according to the analysis results.

2. The intelligent real-time monitoring method for chemical production as described in claim 1, characterized in that, Retrieve the historical operating condition direction memory of the reactor circulation pipeline, including: Acquire multiple absorbance data sets from multiple wavelength points output by the online near-infrared spectrometer on the reactor circulation pipeline during the historical window for multiple production batches; Obtain the true set of concentration values ​​of multiple reaction solutions simultaneously determined by offline high performance liquid chromatography; The multiple absorbance data sets and multiple reaction solution concentration true value sets are mapped and associated for storage, and a historical operating condition direction memory is constructed.

3. The intelligent real-time monitoring method for chemical production as described in claim 1, characterized in that, Perform singular value decomposition of the weighted parameter space on the initial concentration analysis model, and extract the principal singular vectors with a cumulative energy contribution rate greater than or equal to a preset threshold to form a fixed principal subspace, including: The weight matrix of each fully connected layer in the initial concentration analysis model is taken as the decomposition object. Singular value decomposition is performed on each weight matrix to obtain the singular value sequence and the left singular vector and right singular vector corresponding to each singular value. The sum of squares of the singular values ​​is accumulated in descending order. When the proportion of the accumulated sum of squares to the sum of squares of all singular values ​​is greater than or equal to a preset threshold, the accumulation is stopped, and the right singular vectors corresponding to the first few accumulated singular values ​​are extracted as basis vectors of the fixed principal subspace. The subspace spanned by the basis vectors of the fixed principal subspace is taken as the fixed principal subspace.

4. The intelligent real-time monitoring method for chemical production as described in claim 1, characterized in that, include: After each batch of raw materials is updated in the initial concentration analysis model, the weight change matrix of the low-rank update module during this update is extracted, wherein the weight change matrix is ​​the product of the upper projection matrix and the lower projection matrix. The weight change matrix is ​​subjected to singular value decomposition, and the left singular vector corresponding to the largest singular value is taken as the dominant direction vector of the corresponding batch of raw materials. Add the dominant direction vector, the supplier number of the batch of raw materials, the date of raw material receipt, and the batch number of the reactor as a record to the historical raw material batch direction library.

5. The intelligent real-time monitoring method for chemical production as described in claim 4, characterized in that, The weight update gradient is orthogonally projected onto the pre-stored historical raw material batch direction library to obtain the residual projection gradient, including: The multiple dominant direction vectors stored in the historical raw material batch direction library are subjected to Schmidt orthogonalization to obtain a set of standard orthogonal bases, wherein the standard orthogonal bases span the historical raw material subspace; Calculate the projection component of the weight update gradient in each basis direction of the standard orthogonal basis, and multiply the projection component by the corresponding basis and sum them to obtain the total projection component; The residual projection gradient is obtained by subtracting the total projection component element by element from the weight update gradient.

6. The intelligent real-time monitoring method for chemical production as described in claim 1, characterized in that, The initial concentration analysis model is updated adaptively without forgetting based on the residual projection gradient to obtain an updated concentration analysis model, including: The residual projection gradient is subjected to singular value decomposition, and the left singular vectors corresponding to the first four largest singular values ​​are extracted and shrunk into an N-row, 4-column down projection matrix, where N is the dimension of the residual projection gradient. The value of the lower projection matrix is ​​fixed and does not participate in the subsequent gradient descent optimization. At the same time, a 4-row, M-column upper projection matrix is ​​randomly initialized, where M is the model output dimension. Using the initial working condition data as the training set, the mean square error loss function is used to optimize the upper projection matrix, and the training is iterated until the loss converges. The trained upper projection matrix is ​​multiplied by the fixed lower projection matrix to obtain a low-rank update matrix. The low-rank update matrix is ​​then layered onto the corresponding weight matrix of the initial concentration analysis model to obtain the updated concentration analysis model.

7. The intelligent real-time monitoring method for chemical production as described in claim 1, characterized in that, Based on the updated concentration analysis model and the fixed principal subspace, the data from the online infrared spectrometer of the reactor circulation pipeline are monitored and analyzed. Based on the analysis results, an alarm for any monitoring anomalies is triggered, including: The real-time absorbance data from the online infrared spectrometer is input into the updated concentration analysis model to obtain real-time monitoring and analysis results; Calculate the projected energy on the fixed principal subspace based on the spectral data from the online infrared spectrometer; When the real-time monitoring and analysis results are greater than or equal to the preset concentration threshold and / or the projected energy does not meet the preset range constraint, a monitoring anomaly alarm is triggered.

8. The intelligent real-time monitoring method for chemical production as described in claim 7, characterized in that, The preset range constraint is that the projected energy is greater than a first multiple threshold of the preset energy value and less than or equal to a second multiple threshold of the preset energy value.

9. An intelligent real-time monitoring system for chemical production, characterized in that, The steps for implementing the intelligent real-time monitoring method for chemical production according to any one of claims 1 to 8, wherein the intelligent real-time monitoring system for chemical production comprises: The data training module is used to acquire the historical operating condition direction memory of the reactor circulation pipeline, and to train the initial concentration analysis model based on the historical operating condition direction memory. The vector extraction module is used to perform weight parameter space singular value decomposition on the initial concentration analysis model, extract the principal singular vectors with a cumulative energy contribution rate greater than or equal to a preset threshold to form a fixed principal subspace, and fix the fixed principal subspace as a non-updateable part; The model training module is used to collect the initial operating condition data of the reactor circulation pipeline in the initial production window when a new batch of raw materials is replaced, and to train the initial concentration analysis model using the initial operating condition data to obtain the weight update gradient. The model update module is used to orthogonally project the weight update gradient onto the pre-stored historical raw material batch direction library to obtain the residual projection gradient, and perform forgetting-free adaptive update on the initial concentration analysis model based on the residual projection gradient to obtain the updated concentration analysis model. The monitoring and analysis module is used to monitor and analyze the data of the online infrared spectrometer in the circulation pipeline of the reactor based on the updated concentration analysis model and the fixed principal subspace, and trigger an alarm for monitoring anomalies based on the analysis results.

10. An electronic device, characterized in that, include: At least one processor; A memory that is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the steps of the intelligent real-time monitoring method for chemical production according to any one of claims 1 to 8.

Images