A fault warning method and system for a fixed-bed reactor industrial process

Abstract

This application relates to a fault early warning method and system for industrial processes in a fixed-bed reactor, comprising: acquiring historical process data of different spaces inside the fixed-bed reactor under normal operating conditions; preprocessing the historical process data to obtain offline input data; establishing a process monitoring model based on a convolutional autoencoder based on the offline input data to obtain offline results; acquiring process data of different spaces inside the fixed-bed reactor under operating conditions online in real time; preprocessing the process data to obtain online input data; obtaining online results based on the online input data and the process monitoring model; comparing the online results with the offline results to determine whether a fault has occurred in the fixed-bed reactor; the fixed-bed reactor process monitoring method based on spatial feature extraction can effectively extract the spatial correlation between measured variables, realize early warning of overdue faults in the production process, and provide operators with more time to eliminate faults.

Description

Technical Field

[0001] This application relates to the technical field of fault early warning for chemical equipment, and in particular to a fault early warning method and system for a fixed-bed reactor industrial process. Background Technology

[0002] Fixed-bed reactors are commonly used in multiphase reaction processes and are widely applied in the petroleum refining, inorganic chemical, and organic chemical industries. As a critical core equipment unit, a process failure in a fixed-bed reactor can potentially cause significant losses of personnel and property. Therefore, achieving early warning of process failures in fixed-bed reactors and ensuring their safe and stable operation is of paramount importance.

[0003] With the widespread use of distributed control systems (DCS) in the petrochemical industry, a large amount of production process data is collected and stored, laying a solid foundation for the development of data-driven process monitoring methods. By analyzing and processing historical process data, extracting data features under normal operating conditions to establish process monitoring models enables proactive identification of process faults, providing operators with more time to eliminate them. Traditional process monitoring methods, such as principal component analysis (PCA) and partial least squares (PLS), primarily extract data features from process data through linear orthogonal transformations. However, modern chemical processes are typically complex nonlinear processes, making it difficult to extract process features using linear methods. To address this issue, several nonlinear process monitoring methods have been proposed, which can be mainly divided into three categories: linear approximation methods, kernel methods, and nonlinear process monitoring methods based on artificial neural networks. Linear approximation methods, such as locally weighted principal component analysis (LTPC) and locally weighted partial least squares (PLS), are the simplest and most effective modeling methods among nonlinear process monitoring methods. However, they are not suitable for chemical processes with strong nonlinearity. Kernel methods, such as kernel PCA and kernel PLS, are nonlinear process monitoring methods developed based on kernel techniques. However, building and computing kernel functions is often very time-consuming, which can affect the performance of real-time process monitoring.

[0004] In recent years, with the rapid development of artificial neural networks, several nonlinear process monitoring methods based on artificial neural networks have been proposed. Currently, common artificial neural networks include multilayer perceptrons, recurrent neural networks, convolutional neural networks, and autoencoders. Among them, autoencoders, as an unsupervised artificial neural network, are widely used in process monitoring modeling. Based on autoencoders, various improved autoencoders have been proposed to improve the performance of process monitoring models. For example, process monitoring methods based on stacked autoencoders are used to improve the model's generalization ability; process monitoring methods based on variational autoencoders are proposed to address the situation where the data features captured by autoencoders do not conform to a normal distribution; convolutional autoencoders and long short-term memory autoencoders are used to extract dynamic features of the process. For autoencoder-based process monitoring methods, it is usually necessary to preprocess the process data into the input form required by the autoencoder. However, regardless of the input data matrix form used, the influence of the arrangement of process variables on data feature extraction, especially the spatial arrangement of variables, is not considered.

[0005] In chemical processes, fixed-bed reactors are three-dimensional devices where reactions and transport occur, and certain key performance indicators are spatially distributed. Spatial correlations exist between process variables obtained from measuring instruments at different locations, significantly impacting the early warning performance of fixed-bed reactors for industrial failures. However, this has not been adequately considered in previous process monitoring methods, reducing the accuracy of early warning for fixed-bed reactor failures and lowering work efficiency. Summary of the Invention

[0006] This application provides a fault early warning method and system for fixed-bed reactor industrial processes. The fixed-bed reactor process monitoring method based on spatial feature extraction can effectively extract the spatial correlation between measured variables, realize early warning of overdue production process faults, and especially realize early identification of small deviations in the early stage of fault occurrence, providing operators with more time to eliminate faults.

[0007] In a first aspect, this application provides a fault early warning method for an industrial process in a fixed-bed reactor. The method includes: offline modeling and online monitoring. The offline modeling includes the following steps: acquiring historical process data of different spaces inside the fixed-bed reactor under normal operating conditions; preprocessing the historical process data to obtain offline input data; establishing a process monitoring model based on a convolutional autoencoder according to the offline input data, and obtaining offline results. The online monitoring includes the following steps: real-time online acquisition of process data of different spaces inside the fixed-bed reactor under operating conditions; preprocessing the process data to obtain online input data; obtaining online results based on the online input data and the process monitoring model; and comparing the online results with the offline results to determine whether a fault has occurred in the fixed-bed reactor.

[0008] Optionally, before performing data preprocessing on the historical process data to obtain offline input data, the method further includes: normalizing the historical process data, as described by formula (1):

[0009]

[0010] Optionally, the step of preprocessing the historical process data to obtain offline input data includes: arranging the historical process data into a two-dimensional data matrix according to the different spatial distributions of the historical process data inside the fixed-bed reactor; stacking the two-dimensional data matrices at different times into a three-dimensional time series data matrix to obtain offline input data; and extracting the spatial correlation between the historical process data based on the three-dimensional time series data matrix.

[0011] Optionally, the step of preprocessing the historical process data to obtain offline input data further includes: preprocessing the two-dimensional data matrix using data augmentation techniques based on the Laplace operator to enhance the distinction between data under normal and fault conditions, as described by formula (2.1):

[0012]

[0013] Where I(p,q) and V(p,q) represent the variable values ​​in the original two-dimensional data matrix and the enhanced two-dimensional data matrix, respectively.

[0014] Optionally, the step of establishing a process monitoring model based on a convolutional autoencoder based on the offline input data and obtaining offline results includes: using a portion of the offline input data as a training set; using the training set as input data, extracting the autocorrelation, cross-correlation, and spatial correlation between the variables in the training set based on the convolutional autoencoder, and obtaining the data features extracted by the convolutional autoencoder, which are described by formula (3):

[0015]

[0016] Where x represents the input data, w k represents the weight parameters of the convolution kernel, b represents the bias parameters of the convolution kernel, and f() represents the activation function;

[0017] The data features extracted by the convolutional autoencoder are reconstructed to obtain reconstructed data, which is described by formula (4):

[0018] c'=f(∑ k c k *w k +b) (4),

[0019] Where c' represents the output of data reconstruction, c k w represents the input to the deconvolutional layer. k represents the weight parameters of the deconvolutional layer kernel, and b represents the bias parameters;

[0020] The reconstruction error is calculated based on the input data and the reconstructed data, and is described by formula (5):

[0021]

[0022] Where, x i Indicates the input data, y i This indicates that the output data is also the reconstructed data.

[0023] Optionally, establishing a process monitoring model based on a convolutional autoencoder based on the offline input data and obtaining offline results further includes: using a portion of the offline input data as a test set; using the test set as input data for the process monitoring model; calculating the reconstruction error of the process monitoring model under normal operating conditions based on the process monitoring model; selecting a threshold of the reconstruction error within a 99% confidence interval as the control limit of the process monitoring model, described by formulas (6) and (7):

[0024]

[0025]

[0026] Wherein, the control limit refers to the threshold of reconstruction error obtained under normal operating conditions, which is used to verify whether the collected online data is safe, h represents the smoothing parameter of the kernel density estimation function, n represents the number of reconstruction error samples, and MSE represents the reconstruction error.

[0027] Optionally, the process data preprocessing to obtain online input data includes: real-time online acquisition of process data in different spaces inside the fixed-bed reactor, and normalization of the process data; and preprocessing the normalized process data to obtain an enhanced three-dimensional time series data matrix, which is used as online input data for the process monitoring model.

[0028] Optionally, obtaining online results based on the online input data according to the process monitoring model includes: inputting the online input data into the process monitoring model to obtain the reconstruction error.

[0029] Optionally, comparing the obtained online results with the offline results to determine whether the fixed-bed reactor has malfunctioned includes: comparing the reconstruction error obtained from the online input data with the control limit; if the reconstruction error is less than the control limit, it indicates that the process is still under normal operating conditions; otherwise, it indicates that the fixed-bed reactor has malfunctioned.

[0030] Secondly, this application provides a fault early warning system for a fixed-bed reactor industrial process. The system includes: a training module for offline modeling; and a monitoring module for online monitoring. The training module includes the following units: an acquisition unit for acquiring historical process data of different spaces inside the fixed-bed reactor under normal operating conditions; a first processing unit for preprocessing the historical process data to obtain offline input data; and a first calculation unit for establishing a process monitoring model based on a convolutional autoencoder according to the offline input data and obtaining offline results. The monitoring module includes the following units: an acquisition unit for online real-time acquisition of process data of different spaces inside the fixed-bed reactor under operating conditions; a second processing unit for preprocessing the process data to obtain online input data; a second calculation unit for obtaining online results based on the online input data and the process monitoring model; and a result judgment unit for comparing the online results with the offline results to determine whether a fault has occurred in the fixed-bed reactor.

[0031] Thirdly, this application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the steps of the method described above when executing the computer program.

[0032] Fourthly, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method described above.

[0033] This application has at least the following advantages:

[0034] According to the technical content provided in the embodiments of this application, process data is rearranged into a three-dimensional time series data matrix based on the spatial distribution characteristics of measuring instruments on a fixed-bed reactor. A process monitoring model based on a convolutional autoencoder is used to more effectively extract the cross-correlation, autocorrelation, and spatial correlation between variables. Furthermore, data augmentation technology based on the Laplace operator can further enhance the distinction between process data under normal and abnormal operating conditions, improving the process monitoring model's ability to identify process faults. Data obtained through multi-dimensional considerations and data augmentation processing yields multi-dimensional online input data. The reconstruction error calculated by the process monitoring model is used to accurately analyze whether a fault has occurred in the fixed-bed reactor. This allows for the early identification of process faults in the fixed-bed reactor, preventing further development of process faults and avoiding more serious production accidents. It also provides operators with more time to eliminate process faults and improves work efficiency. Attached Figure Description

[0035] Figure 1 This is a diagram illustrating the application environment of a fault early warning method for a fixed-bed reactor industrial process in one embodiment;

[0036] Figure 2 This is a schematic diagram illustrating offline modeling in a fault early warning method for a fixed-bed reactor industrial process in one embodiment;

[0037] Figure 3 This is a schematic diagram illustrating the online monitoring process in a fault early warning method for a fixed-bed reactor industrial process in one embodiment;

[0038] Figure 4 This is a flowchart illustrating a fault early warning method for a fixed-bed reactor industrial process in one embodiment;

[0039] Figure 5 This is a schematic diagram illustrating the data preprocessing process in one embodiment;

[0040] Figure 6 This is a schematic diagram illustrating the arrangement of variables in a fixed-bed reactor in one embodiment;

[0041] Figure 7 This is a schematic diagram of the convolution operation of a convolutional autoencoder in one embodiment;

[0042] Figure 8 This is a schematic diagram showing the structure of the Laplacian operator template in one embodiment;

[0043] Figure 9 This is a schematic diagram showing the structure of a convolutional autoencoder in one embodiment;

[0044] Figure 10This is a flowchart illustrating the process of establishing a process monitoring model in one embodiment;

[0045] Figure 11 This is a schematic diagram illustrating the process of establishing control limits in one embodiment;

[0046] Figure 12 This is a schematic diagram illustrating the online monitoring process in a fault early warning method for a fixed-bed reactor industrial process in one embodiment;

[0047] Figure 13 This is a schematic diagram showing the structure of a fixed-bed reactor in a hydrogen production unit in one embodiment;

[0048] Figure 14 One embodiment shows a three-dimensional time series data matrix after the monitoring variables have been rearranged;

[0049] Figure 15 One embodiment shows a spatial correlation analysis plot between the monitored variables;

[0050] Figure 16 This is a schematic diagram illustrating monitoring results based on six different process monitoring methods in one embodiment;

[0051] Figure 17 This is a schematic diagram showing a comparison between the data augmentation matrix and the original data matrix in one embodiment;

[0052] Figure 18 One embodiment shows a comparison chart of the raw test dataset and the data augmentation matrix STD;

[0053] Figure 19 This is a schematic diagram showing the mean square error of the test dataset in one embodiment;

[0054] Figure 20 This is a schematic diagram showing the peak signal-to-noise ratio of a test dataset in one embodiment;

[0055] Figure 21 Here is a system architecture block diagram showing offline modeling in one embodiment;

[0056] Figure 22 This is a block diagram of the system architecture for online monitoring, shown in one embodiment.

[0057] Figure 23 This is a schematic structural diagram of a computer device in one embodiment. Detailed Implementation

[0058] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the scope of the present application.

[0059] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when used in this specification, the words “comprising” and / or “including” indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0060] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the various embodiments of this application to facilitate a better understanding of the application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0061] For ease of understanding, the system to which this application applies will first be described. This application provides a fault early warning method for a fixed-bed reactor industrial process, which can be applied to, for example... Figure 1 The system architecture shown includes a user-space file server 103 and a terminal device 101. The terminal device 101 communicates with the user-space file server 103 via a network. The user-space file server 103 can be a file server based on the NFSv3 / v4 protocol, running in a Linux environment. NFS (Network File System) is a network abstraction on top of a file system, allowing remote clients running on the terminal device 101 to access the file system over the network in a manner similar to a local file system. The terminal device 101 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, etc. The user-space file server 103 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0062] Figure 2 , Figure 3 This application provides a schematic flowchart of a fault early warning method for a fixed-bed reactor industrial process, which can be implemented by, for example... Figure 1 The user-space file server in the system shown executes this. For example... Figure 2 , Figure 3 As shown, the method may include the following steps:

[0063] Offline modeling and online monitoring, wherein offline modeling specifically includes the following steps:

[0064] S201. Obtain historical process data of different spaces inside the fixed-bed reactor under normal operating conditions;

[0065] S202. Perform data preprocessing on the historical process data to obtain offline input data;

[0066] S203. Establish a process monitoring model based on convolutional autoencoder according to the offline input data and obtain offline results;

[0067] Online monitoring specifically includes the following steps:

[0068] S301. Real-time online acquisition of process data from different spaces inside the fixed-bed reactor under operating conditions;

[0069] S302. Perform data preprocessing on the process data to obtain online input data;

[0070] S303. Based on the online input data and the process monitoring model, obtain online results;

[0071] S304. Compare the online results with the offline results to determine whether the fixed-bed reactor has malfunctioned.

[0072] The following is a detailed explanation of each step:

[0073] S201. Obtain historical process data of different spaces inside the fixed-bed reactor under normal operating conditions;

[0074] like Figure 2 , Figure 4 As shown in this embodiment, it should be noted that the fixed-bed reactor is a three-dimensional device, with the raw material flowing through the reactor from top to bottom, and both the reaction and transport occurring in a three-dimensional space. In chemical processes, the measured variables of composition, temperature, and pressure can be used to determine the progress of the reaction. Typically, compositional information directly reflects the reaction progress. However, due to limitations in measurement technology, compositional data is difficult to measure online. To observe the reaction progress in real time, measuring instruments are usually installed at different locations within the fixed-bed reactor to monitor key performance indicators distributed in different spaces, thereby obtaining historical process data from different spaces within the fixed-bed reactor. These key performance indicators can be temperature fields, pressure fields, etc., which can be used to indirectly reflect the reaction progress.

[0075] In one example, to obtain historical process data of a fixed-bed reactor, the temperature variable of the fixed-bed reactor is collected as a modeling variable. Measuring instruments are installed at different locations inside the fixed-bed reactor to monitor the temperature variable in different spaces.

[0076] Specifically, based on the chemical process in the fixed-bed reactor, 18 temperature measurement points are set up, distributed across 6 different bed heights, with 3 temperature measurement points at each bed height. Monitoring data is obtained at different spatial locations based on the spatial distribution of the measuring instruments.

[0077] S202. Perform data preprocessing on the historical process data to obtain offline input data;

[0078] In this embodiment, it should be noted that, based on the spatial distribution of the measuring instruments, the spatial distribution characteristics of the temperature measurement points are analyzed, feature variables are extracted based on the spatial distribution characteristics, and the data is preprocessed to enhance the distinction between data under normal and fault conditions, thereby further improving the fault identification capability of the process monitoring method and obtaining offline input data to provide a data foundation for subsequent operations.

[0079] S203. Establish a process monitoring model based on convolutional autoencoder according to the offline input data and obtain offline results;

[0080] In this embodiment, it should be noted that an autoencoder is a typical unsupervised artificial neural network, consisting of five parts: an input layer, an encoding layer, an encoding space, a decoder, and an output layer. As an unsupervised artificial neural network, an autoencoder is actually a feedforward neural network with identical input and output data. It extracts data features by minimizing the reconstruction error between the input and output data. A convolutional autoencoder is an important variant of the autoencoder, commonly used for feature extraction from image or video data. The encoder of a convolutional autoencoder consists of convolutional layers and pooling layers, while the decoder consists of deconvolutional layers and unpooling layers. When processing input data, the convolutional kernels in the convolutional layers slide along different dimensions of the data matrix, extracting autocorrelation, cross-correlation, and spatial correlation between variables through convolution operations. This establishes a process monitoring model and yields offline results after analysis and computation.

[0081] S301. Real-time online acquisition of process data from different spaces inside the fixed-bed reactor under operating conditions;

[0082] like Figure 3 , Figure 4 As shown in this embodiment, it should be noted that process data such as temperature data are collected in real time by measuring instruments installed at different locations inside the fixed-bed reactor to obtain the process data of the fixed-bed reactor for subsequent analysis.

[0083] S302. Perform data preprocessing on the process data to obtain online input data;

[0084] like Figure 3 , Figure 4As shown in this embodiment, it should be noted that, based on the spatial distribution location of the measuring instruments, the spatial distribution characteristics of the temperature measuring points are analyzed, feature variables are extracted based on the spatial distribution characteristics, and the process data is preprocessed to enhance the distinction between data under normal and fault conditions, thereby further improving the fault identification capability of the process monitoring method and obtaining online input data to provide a data foundation for subsequent operations.

[0085] S303. Based on the online input data and the process monitoring model, obtain online results;

[0086] like Figure 3 , Figure 4 As shown in this embodiment, it should be noted that the preprocessed online input data is input into the process monitoring model trained based on the process history data to obtain online results, providing a data basis for subsequent analysis of whether the fixed-bed reactor has failed.

[0087] S304. Compare the online results with the offline results to determine whether the fixed-bed reactor has malfunctioned.

[0088] like Figure 3 , Figure 4 As shown in this embodiment, it should be noted that by comparing the online results with the offline results obtained from the analysis and calculation under normal operating conditions, the cross-correlation, autocorrelation, and spatial correlation between variables obtained from the offline results are analyzed from multiple dimensions to analyze the historical data of the process, thereby enabling a more accurate determination of whether the fixed bed reactor has malfunctioned.

[0089] The following is a detailed explanation of each step in the offline modeling process:

[0090] Reference Figure 2 , Figure 4 As shown, in some embodiments, before performing data preprocessing on the process history data to obtain offline input data in S202, the method further includes:

[0091] The historical process data is normalized and described by formula (1):

[0092]

[0093] In this embodiment, it should be noted that the acquired historical process data is normalized to unify the historical process data, so as to facilitate subsequent analysis and calculation.

[0094] Reference Figure 2 , Figure 5 As shown, in some embodiments, in S202, the process history data is preprocessed to obtain offline input data, including:

[0095] S2021. Based on the different spatial distributions of process history data inside the fixed-bed reactor, arrange them into a two-dimensional data matrix;

[0096] In this embodiment, it should be noted that, referring to Figure 6 As shown, the collected historical data can be viewed as a two-dimensional matrix, where rows represent the variable dimension and columns represent the time dimension. A moving window of length l is used to select the multivariate time series data as the process monitoring data at time t. To meet the input data format requirements of the autoencoder, the multivariate time series data is usually rearranged into the following format: Figure 6 Part A and Part B show two types of data matrices. Thanks to the ability of convolutional kernels to extract local features from input data, CAE has been successfully applied in the field of process monitoring. One-dimensional and two-dimensional convolutional kernels are two commonly used types of convolutional autoencoders.

[0097] like Figure 7 As shown in Part A, a single-channel two-dimensional convolution kernel slides along the variable and time dimensions to perform convolution operations, which can be used to simultaneously extract cross-correlation and autocorrelation between adjacent variables. Furthermore, multi-channel one-dimensional convolution kernels can be used to extract process data features such as... Figure 7 As shown in Part B, each kernel matrix of the convolution kernel is used to extract the autocorrelation of the variables. Then, the sum of the convolution results of each kernel matrix is ​​used as the data feature extracted by that convolution kernel. From... Figure 7 As can be seen from Part A and Part B, both convolutional kernels are used to extract data features from adjacent variables. However, the order of the variables affects the extraction of data features, which has not been considered in previous process monitoring methods based on convolutional autoencoders.

[0098] The fixed-bed reactor described in this application is a three-dimensional device, with the feedstock flowing through it from top to bottom, and both the reaction and transport occurring in a three-dimensional space. To monitor the reaction process in real time, measuring instruments are typically installed at different locations within the fixed-bed reactor. Some key performance indicators (such as the temperature field) are spatially distributed, thus exhibiting spatial correlations among these measured variables. Therefore, if a process monitoring model can be established based on the extracted spatial correlations of these variables, more effective identification of process faults can be achieved. Therefore, as... Figure 7 As shown in Part C, in this application, the measured variables are arranged into a two-dimensional data matrix according to the spatial distribution characteristics of the measuring instrument. For example, arranging 9 measured variables into a 3*3 two-dimensional data matrix can be regarded as a two-dimensional "image".

[0099] S2022: Stack the two-dimensional data matrices at different times into a three-dimensional time series data matrix to obtain offline input data;

[0100] In this embodiment, it should be noted that, referring to Figure 6 , Figure 7 As shown, this application is used in a fixed-bed reactor, and considering the highly dynamic characteristics of the chemical process, therefore... Figure 7 As shown in Part C, two-dimensional data matrices at time l are stacked into a three-dimensional time series data matrix as offline input data for the process monitoring model at time t.

[0101] S2023. Extract the spatial correlation between historical data in the process based on the three-dimensional time series data matrix.

[0102] In this embodiment, it should be noted that, taking advantage of the local feature extraction of the convolution kernel, the three-dimensional time series data matrix obtained by the new variable arrangement is locally extracted. In this application, the two-dimensional data matrices at l time points are stacked and arranged into a three-dimensional time series data matrix, and a moving window of length L is used to extract the spatial correlation between the monitoring variables on the three-dimensional time series data matrix.

[0103] Reference Figure 2 , Figure 5 As shown, in some embodiments, S202, data preprocessing is performed on the historical data of the process to obtain offline input data, which further includes:

[0104] Data augmentation techniques based on the Laplace operator are used to preprocess a two-dimensional data matrix to enhance the distinction between data under normal and fault conditions, as described by the following formula (2.1):

[0105]

[0106] Where I(p,q) and V(p,q) represent the variable values ​​in the original two-dimensional data matrix and the enhanced two-dimensional data matrix, respectively.

[0107] For ease of calculation, the discrete equation of the following formula is usually used to approximate formula (2.1).

[0108]

[0109] like Figure 8The image shows a Laplace operator template, which can be considered a two-dimensional convolution kernel. The Laplace transform of a two-dimensional data matrix can be further understood as a convolution operation between the Laplace operator template and the data matrix. It can be seen that through the Laplace operation, the differences between intermediate and surrounding variables are further amplified, making the data features that represent the intrinsic characteristics of the process more prominent and easier to extract. Especially when a fault occurs, some process variables undergo abnormal changes, and the use of the Laplace operator further highlights this change. Therefore, the data discrimination features between normal and fault conditions are further enhanced, thereby significantly improving the fault identification capability of process monitoring methods.

[0110] In this embodiment, it should be noted that the Laplacian operator-based data augmentation technique, where image enhancement is a common image preprocessing method in the field of image processing, can be used to improve the visual effect of images and provide better image data for other image processing techniques. The Laplacian operator is a spatial augmentation technique. Based on the spatial distribution characteristics of the measuring instruments, the process variables of a fixed-bed reactor are arranged into a two-dimensional data matrix, which can be viewed as a two-dimensional "image." Therefore, by using the Laplacian operator-based data augmentation technique, the two-dimensional data matrix is ​​preprocessed using the Laplacian operation, enhancing the distinction between data under normal and fault conditions, and further improving the fault identification capability of the process monitoring method.

[0111] In this embodiment, it should also be noted that, in addition to data augmentation techniques based on the Laplacian operator, other data augmentation techniques can also be used to increase the difference in data features between normal and abnormal operating conditions, such as data augmentation techniques based on grayscale transformation and data augmentation techniques based on histogram equalization.

[0112] Let's take data augmentation technology based on grayscale transformation as an example:

[0113] Grayscale transformation is a point-by-point operation, as shown in formula (2.2).

[0114] V(p,q)=T[I(p,q)];(2.2)

[0115] In the formula, T[] represents the point operation method, which can be divided into linear transformation and nonlinear transformation methods.

[0116] Linear transformation:

[0117] As shown in formula (2,3), the input and output data are amplified or reduced in a linear relationship.

[0118] V(p,q)=T[I(p,q)]=a(I(p,q))+b; (2,3)

[0119] When coefficient a > 1, the output value increases; when coefficient a < 1, the output value decreases; when coefficient a = 1 and b ≠ 0, the values ​​of all points are shifted up or down, making the overall data darker or brighter; when coefficient a = 1 and b = 0, the input and output data are the same; when coefficient a < 0, the dark areas of the input data become brighter, and the bright areas become darker.

[0120] Nonlinear transformation:

[0121] Nonlinear transformation refers to the nonlinear transformation between input and output data. Depending on the transformation method, it can be further divided into logarithmic transformation and exponential transformation, etc. The logarithmic transformation is shown in formula (2.4).

[0122] V(p,q)=T[I(p,q)]=clog(1+I(p,q)); (2.4)

[0123] Where c is the coefficient. The logarithmic transformation compresses the dynamic range of the original data; the larger the input data value, the smaller the output data value; and vice versa.

[0124] The exponential transformation is shown in equation (2.5).

[0125] V(p,q)=T[I(p,q)]=ce I(p,q) (2.5)

[0126] Where c is the coefficient. In contrast to the logarithmic transformation, the larger the input data value, the larger the output data value; and the smaller the input data value, the smaller the output data value.

[0127] Let's take histogram equalization-based data augmentation techniques as an example:

[0128] The main idea behind histogram-based data augmentation is to remap the data value levels of each point in the input data based on a uniform expansion of its probability density function (PDF). First, it constructs a cumulative histogram from the original histogram. Second, this cumulative histogram is normalized to the intensity values ​​of the output data. Then, the normalized cumulative histogram is used as a transfer mapping function to the original data. Its transformation formula mainly consists of the mapping interval and the mapping function, namely the cumulative distribution function (CDF). The formulas for calculating the PDF and CDF of the histogram are shown in equations (2.6) and (2.7).

[0129]

[0130]

[0131] Where g represents a data value level on the histogram, n g represents the number of statistical data at the g data value level, N represents the total number of data in the input image, and start represents the minimum data value level of a certain interval.

[0132] According to equation (2.7), the histogram equilibrium transformation formula can be derived as follows:

[0133] F(g)=start+(end-start)×CDF(g);

[0134] Where end-start represents the mapping interval.

[0135] Of course, there may be other data augmentation techniques, which are not listed in this embodiment.

[0136] In some embodiments, refer to Figure 2 , Figure 10 As shown in S203, a process monitoring model based on a convolutional autoencoder is established based on offline input data, and offline results are obtained, including:

[0137] S2031. Use a portion of the offline input data as the training set;

[0138] S2032. Using the training set as input data, the autocorrelation, cross-correlation, and spatial correlation between variables in the training set are extracted based on the convolutional autoencoder to obtain the data features extracted by the convolutional autoencoder, which are described by formula (3):

[0139] c k =f(∑ k x*w k +b) (3),

[0140] Where x represents the input data, w k represents the weight parameters of the convolution kernel, b represents the bias parameters of the convolution kernel, and f() represents the activation function;

[0141] S2033. Reconstruct the data features extracted by the convolutional autoencoder to obtain reconstructed data, which is described by formula (4):

[0142] c'=f(Σ k c k *w k +b) (4),

[0143] Where c' represents the output of data reconstruction, c k w represents the input to the deconvolutional layer. k represents the weight parameters of the deconvolutional layer kernel, and b represents the bias parameters;

[0144] S2034. Calculate the reconstruction error based on the input data and the reconstructed data, and describe it using formula (5):

[0145]

[0146] Where x i Indicates the input data, y i This indicates that the output data is also the reconstructed data.

[0147] In this embodiment, it should be noted that, referring to Figure 9 As shown, the encoding layer of a convolutional autoencoder can consist of multiple convolutional and pooling layers. The obtained offline input data is used as the training set; here, 70% of the offline input data is used as the training set. When processing the input data, i.e., the training set, the convolutional kernels in the convolutional layers slide along different dimensions of the data matrix, extracting autocorrelation, cross-correlation, and spatial correlation between variables through convolution operations. After multiple convolution and pooling operations, the data features extracted by the convolutional autoencoder are finally obtained, i.e., the encoding. The decoding layer is used to reconstruct the extracted data features. This data feature is input into the depooling layer to restore the dimensionality of the data features compressed by the pooling layer. The deconvolution layer is used to reconstruct the input data from the extracted data features, ultimately obtaining the reconstructed data. The mean squared error function is often used as the loss function of the autoencoder to evaluate the reconstruction error. Based on the input data and the obtained reconstructed data, the reconstruction error is obtained through the mean squared error function.

[0148] By using a training set that considers the autocorrelation, cross-correlation, and spatial correlation among variables, and based on the reconstruction error calculated by the process monitoring model, we can accurately analyze whether a stationary reactor has failed.

[0149] In some embodiments, refer to Figure 2 , Figure 4 , Figure 11 As shown in S203, the process monitoring model based on a convolutional autoencoder is established based on offline input data, and offline results are obtained. This also includes:

[0150] S2035. Use a portion of the offline input data as the test set;

[0151] In this embodiment, it should be noted that the remaining portion of the offline input data is used as validation data to form a test set, thereby determining the key parameters of the process monitoring model to obtain the optimal model. During the model iterative training process, when the validation data is input into the initial convolutional autoencoder model built based on 70% of the offline input data, and the mean square error between the model input and output values ​​no longer decreases, and the mean square error increases in 30 consecutive iterations, it is considered that the iterative calculation of the model can be stopped. At this time, the model parameters with the smallest mean square error in the last 30 iterations are selected as the final optimal model parameters.

[0152] S2036. The test set is used as input data for the process monitoring model. The reconstruction error of the process monitoring model under normal operating conditions is calculated based on the process monitoring model.

[0153] In this embodiment, it should be noted that the test set is input into the process monitoring model of the convolutional autoencoder obtained in the above offline modeling process, and the reconstruction error MSE of the model under normal working conditions is calculated.

[0154] S2037. Select the threshold of reconstruction error within the 99% confidence interval as the control limit of the process monitoring model, and describe it using formulas (6) and (7):

[0155]

[0156]

[0157] Among them, the control limit refers to the threshold of reconstruction error obtained under normal operating conditions, which is used to verify whether the collected online data is safe. h represents the smoothing parameter of the kernel density estimation function, n represents the number of reconstruction error samples, and MSE represents the reconstruction error.

[0158] In this embodiment, it should be noted that the MSE (Mean Separation Error) can be used to evaluate the difference between the original data matrix and the data augmentation matrix; a larger MSE indicates a greater difference between the two. By calculating the reconstruction error under normal operating conditions and using its threshold within the 99% confidence interval as the control limit, a standard for judging whether a fixed-bed reactor has failed is obtained for subsequent analysis and judgment.

[0159] The following is a detailed explanation of each step in the online monitoring process:

[0160] In some embodiments, refer to Figure 2 , Figure 12 As shown in step S302, data preprocessing is performed on the process data to obtain online input data, including:

[0161] S3021. Online real-time acquisition of process data in different spaces inside the fixed-bed reactor under operating conditions, and normalization processing of the process data;

[0162] S3022. Perform data preprocessing on the normalized process data to obtain an enhanced three-dimensional time series data matrix, which will be used as online input data for the process monitoring model.

[0163] In this embodiment, it should be noted that, firstly, process data of the current fixed-bed reactor under operating conditions is collected and normalized. Then, the normalized process data is preprocessed in the manner described in the offline modeling process to obtain an enhanced three-dimensional time series data matrix, which is processed into an input form that the process monitoring model can recognize, so as to obtain online input data for subsequent analysis.

[0164] In some embodiments, refer to Figure 3 , Figure 4 As shown in S303, based on the online input data and the process monitoring model, online results are obtained, including: inputting the online input data into the process monitoring model to obtain the reconstruction error.

[0165] In this embodiment, it should be noted that by analyzing the autocorrelation, cross-correlation, and spatial correlation of the collected process data, multi-dimensional online input data is obtained. Based on the reconstruction error calculated by the process monitoring model, the failure of the stationary reactor is accurately analyzed.

[0166] In some embodiments, refer to Figure 3 , Figure 4 As shown in S304, the obtained online results are compared with the offline results to determine whether the fixed-bed reactor has malfunctioned, including:

[0167] The reconstruction error obtained from online input data is compared with the control limit; if the reconstruction error is less than the control limit, it indicates that the process is still under normal operating conditions; otherwise, it indicates that the fixed-bed reactor has malfunctioned.

[0168] In this embodiment, it should be noted that the online process data, including data on malfunctions, is compared with the control limit by the reconstruction error obtained from the online input data. Since the control limit is the reconstruction error obtained from normal process history data during offline modeling, the reconstruction error obtained from the online input data is compared with the control limit. If the reconstruction error is less than the control limit, it indicates that the process is still under normal operating conditions; otherwise, it indicates that the fixed-bed reactor has malfunctioned, thereby enabling accurate prediction of whether the fixed-bed reactor has malfunctioned.

[0169] In this embodiment, it should be noted that, in order to verify the monitoring effect of the process monitoring model, in one example, such as Figure 13 As shown, a fixed-bed reactor in a hydrogen production unit is a vertical adiabatic fixed-bed reactor. Inside the reactor, a mixture of dilution steam and light naphtha flows from top to bottom and reacts under the action of a catalyst to generate methane-rich gas. By comparing the monitoring and early warning method (IET-CAE) of the process monitoring model in this application with process monitoring methods based on principal component analysis (PCA), autoencoder-based process monitoring (AE), one-dimensional convolutional autoencoder-based process monitoring (1D-CAE), two-dimensional convolutional autoencoder-based process monitoring (2D-CAE), and spatial correlation extraction-based process monitoring (3D-CAE), the fault identification effects of various monitoring and early warning methods on fixed-bed reactors are obtained. Specifically, 1D-CAE and 2D-CAE refer to convolutional autoencoders using one-dimensional and two-dimensional convolutional kernels, respectively, while 3D-CAE refers to a convolutional autoencoder that only uses a variable arrangement method based on the spatial distribution characteristics of measuring instruments and does not employ data augmentation techniques based on the Laplacian operator.

[0170] Reference Figure 13 As shown, this fixed-bed reactor has 18 temperature measurement points distributed across 6 different bed heights, with 3 measurement points at each bed height. Temperature variables are collected from these measurement points, and the spatial distribution characteristics of these points are analyzed. In this application, the temperature data at time t are arranged as follows: Figure 14 The two-dimensional data matrix shown in Part A is then further processed using the data augmentation techniques proposed in this application to obtain a two-dimensional augmented data matrix as shown in Part A. Figure 14 As shown in Part B, finally, following the method shown in step S2023 of the offline modeling above, the two-dimensional augmented data matrices at times t-2, t-1, and t are arranged as follows: Figure 14 The three-dimensional time series data matrix shown in Part C.

[0171] By collecting 2200 sample points under normal operating conditions, including a segment of process fault data, this data was used as test data to evaluate the effectiveness of the process monitoring model in fault identification. First, the spatial correlation between variables was analyzed. Considering the nonlinear relationship between variables, the maximum mutual information coefficient (MIC) was used to evaluate the correlation between variables. For example... Figure 15 As shown, the maximum mutual information coefficient between each variable and its surrounding variables is greater than the maximum mutual information coefficient between each variable and other variables. Spatial correlations were found among the variables. The variable arrangement method based on the spatial distribution characteristics of the measuring instruments can group highly correlated variables together, which is more conducive to the extraction of spatial correlations between variables by the convolutional autoencoder.

[0172] like Figure 16 The figures shown are monitoring results based on six different process monitoring methods. Figure a represents the PCA monitoring results, figure b represents the AE monitoring results, figure c represents the 2D-CAE monitoring results, figure d represents the 1D-CAE monitoring results, figure e represents the 3D-CAE monitoring results, and figure f represents the IET-CAE monitoring results. Figure 16 As shown in Figure f, the process monitoring model based on convolutional autoencoders detected a process fault at the 1350th sample point. Reviewing the unit's historical production records revealed that operators observed abnormal temperature changes in the C and D beds of the fixed-bed reactor at the 2196th sample point. Analysis of the raw materials revealed excessive sulfur content in the dilution steam, confirming a catalyst sulfur poisoning fault in the fixed-bed reactor. In contrast, the process monitoring model based on convolutional autoencoders detected the fault 846 minutes earlier than operators, providing more time for fault handling and effectively reducing the impact on unit production. Furthermore, the monitoring results show that PCA, AE, 2D-CAE, 1D-CAE, and 3D-CAE detected the process fault at the 1967th, 1471st, 1460th, 1462nd, and 1447th sample points, respectively. The comparison revealed that the fault warning method based on the process monitoring model of convolutional autoencoder can detect the fault at least 2 hours earlier than the other five process monitoring methods.

[0173] In this embodiment, it should also be noted that, in order to more intuitively compare the differences between the data augmentation matrix and the original data matrix, the two types of data matrices are also converted into images for comparison, such as... Figure 17 As shown in the figure, under normal operating conditions, there are no clear boundaries between the temperatures of each bed in the original data matrix. Under fault conditions, the temperature differences between beds C, D, E, and F are also very small, almost indistinguishable. In contrast, in the enhanced data matrix, there are clear differences in temperature between the beds. When a fault occurs, the temperature anomalies in reactor beds C and D can be clearly detected in the enhanced data matrix, which are difficult to detect in the original data. This means that the data augmentation technique in this application can effectively enhance the distinction between process data under normal and fault conditions, making it easier for process monitoring methods to detect process faults.

[0174] In this embodiment, it should also be noted that, in order to objectively compare the differences between the original data matrix and the data augmentation matrix, several objective evaluation methods are used in this application, such as standard deviation (STD), mean squared error (MSE), and peak signal-to-noise ratio (PNSR). STD refers to the variance of the variable values ​​in the data matrix, and its calculation formula is as follows:

[0175]

[0176]

[0177] Where H and W represent the height and width of the data matrix, respectively, and μ represents the average value of the data matrix variables.

[0178] MSE can also be used to evaluate the difference between the original data matrix and the data augmentation matrix, and its calculation formula is shown below. As can be seen from the formula, a larger MSE indicates a greater difference between the original data matrix and the data augmentation matrix.

[0179]

[0180] PSNR can also be used to evaluate the difference between the original data matrix and the data augmentation matrix. A lower PSNR value indicates a greater difference between the original data matrix and the data augmentation matrix. Its mathematical expression is as follows:

[0181]

[0182] Where I max This represents the maximum value of the variable in the original data matrix and the data augmentation matrix.

[0183] The STD comparison between the original data matrix and the data augmentation matrix in the test dataset is as follows: Figure 18 As shown in the figure, the STD of the data augmentation matrix increases significantly when a fault occurs, indicating that the information contained in the data augmentation matrix changes significantly when a fault occurs. Furthermore, from... Figure 19 and Figure 20 As can be seen, when a fault occurs, the MSE and PSNR of the test dataset show significant increases and decreases, respectively. This demonstrates that, compared to the original data matrix, the data augmentation technique proposed in this application can significantly enhance the distinction between normal and abnormal process data, which will be more beneficial for the process monitoring model to proactively identify process faults.

[0184] In summary, the method presented in this application is the first to rearrange process data into a three-dimensional time-series data matrix based on the spatial distribution characteristics of measuring instruments. This allows for more effective extraction of cross-correlation, autocorrelation, and spatial correlation among variables using convolutional autoencoders. Furthermore, data augmentation techniques based on the Laplace operator further enhance the distinction between process data under normal and abnormal operating conditions, improving the process monitoring model's ability to identify process faults. Test results from a fixed-bed reactor in a hydrogen production unit demonstrate its effective application in actual industrial processes. Compared to operator intervention and other traditional process monitoring methods, this method can proactively identify process faults in fixed-bed reactors, preventing further development of faults and more serious production accidents, and providing operators with more time to eliminate process faults.

[0185] This application proposes a fixed-bed reactor process monitoring method based on spatial feature extraction. Compared with previous process monitoring methods, this method can effectively extract the spatial correlation between variables using a convolutional autoencoder. Simultaneously, data augmentation technology based on the Laplace operator can further enhance the distinction between process data under normal and abnormal operating conditions, which is beneficial for the early identification of process faults. The application of this method in a fixed-bed reactor of a hydrogen production unit, and its comparison with other process monitoring methods, confirm that this method can detect fixed-bed reactor process faults earlier in practical industrial applications, providing operators with more time to eliminate process faults and thus avoiding further economic losses caused by the further development of process faults.

[0186] Reference Figure 21 , Figure 22 As shown, this application also provides a fault early warning system for a fixed-bed reactor industrial process. This system may include a training module and a monitoring module, wherein the detailed classification units and main functions of each component module are as follows:

[0187] The training module, used for offline modeling, specifically includes:

[0188] Acquisition unit 401 is used to acquire historical process data of monitored variables of the fixed-bed reactor under normal operating conditions;

[0189] The first processing unit 402 is used to preprocess the process history data of the monitored variables to obtain offline input data in a form that conforms to the requirements of the convolutional autoencoder.

[0190] The first computing unit 403 is used to establish a process monitoring model based on a convolutional autoencoder by inputting data offline and to obtain offline results.

[0191] The monitoring module, used for online monitoring, specifically includes:

[0192] The data acquisition unit 501 is used for online real-time acquisition of process data of monitored variables;

[0193] The second processing unit 502 is used to preprocess the process data of the monitored variables to obtain online input data in a form that conforms to the requirements of the convolutional autoencoder.

[0194] The second calculation unit 503 is used to obtain online results based on the online input data and the process monitoring model;

[0195] The result judgment unit 504 compares the obtained online results with the offline results to determine whether the fixed bed reactor has malfunctioned.

[0196] According to embodiments of this application, this application also provides a computer device and a computer-readable storage medium.

[0197] like Figure 23 The diagram shown is a block diagram of a computer device according to an embodiment of this application. The term "computer device" is intended to represent various forms of digital computers or mobile devices. The digital computer may include a desktop computer, a portable computer, a workbench, a personal digital assistant, a server, a mainframe computer, and other suitable computers. The mobile device may include a tablet computer, a smartphone, a wearable device, etc.

[0198] like Figure 23 As shown, device 600 includes a computing unit 601, a ROM 602, a RAM 603, a bus 604, and an input / output (I / O) interface 605. The computing unit 601, ROM 602, and RAM 603 are interconnected via the bus 604. The input / output (I / O) interface 605 is also connected to the bus 604.

[0199] The computing unit 601 can execute various processes in the method embodiments of this application according to computer instructions stored in the read-only memory (ROM) 602 or computer instructions loaded from the storage unit 608 into the random access memory (RAM) 603. The computing unit 601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. The computing unit 601 can include, but is not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. In some embodiments, the methods provided in the embodiments of this application can be implemented as computer software programs, which are tangibly contained in a computer-readable storage medium, such as the storage unit 608.

[0200] RAM 603 may also store various programs and data required for the operation of device 600. Part or all of the computer program may be loaded and / or installed on device 600 via ROM 602 and / or communication unit 609.

[0201] The input unit 606, output unit 607, storage unit 608, and communication unit 609 in device 600 can be connected to I / O interface 605. The input unit 606 can be, for example, a keyboard, mouse, touchscreen, or microphone; the output unit 607 can be, for example, a display, speaker, or indicator light. Device 600 can exchange information and data with other devices through the communication unit 609.

[0202] It should be noted that the device may also include other components necessary for normal operation. It may also include only the components necessary for implementing the solution of this application, without necessarily including all the components shown in the figures.

[0203] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SOCs), payload programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof.

[0204] The computer instructions used to implement the methods of this application may be written in any combination of one or more programming languages. These computer instructions may be provided to the computing unit 601 such that when executed by the computing unit 601, such as a processor, the computer instructions cause the execution of the steps involved in the embodiments of the methods of this application.

[0205] The computer-readable storage medium provided in this application can be a tangible medium that can contain or store computer instructions for performing the steps involved in the method embodiments of this application. The computer-readable storage medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, and other forms of storage media.

[0206] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A fault early warning method for a fixed-bed reactor industrial process, characterized in that, include: Offline modeling and online monitoring; among them, The offline modeling includes the following steps: Acquire historical process data of different spaces inside a fixed-bed reactor under normal operating conditions; The historical data of the process is preprocessed to obtain offline input data; The process of preprocessing the historical data to obtain offline input data includes: Based on the different spatial distributions of the process history data inside the fixed-bed reactor, a two-dimensional data matrix is ​​arranged. Two-dimensional data matrices at different times are stacked into a three-dimensional time series data matrix to obtain offline input data; The spatial correlation between the historical process data is extracted based on the three-dimensional time series data matrix; the two-dimensional data matrix is ​​preprocessed using data augmentation techniques based on the Laplace operator to enhance the distinction between data under normal and fault conditions, as described by formula (2.1): Where I(p,q) and V(p,q) represent the variable values ​​in the original two-dimensional data matrix and the enhanced two-dimensional data matrix, respectively; A process monitoring model based on a convolutional autoencoder is established based on the offline input data, and offline results are obtained. The process monitoring model based on the convolutional autoencoder, established according to the offline input data, and the offline results obtained, include: Use a portion of the offline input data as the training set; Using the training set as input data, the autocorrelation, cross-correlation, and spatial correlation between the variables in the training set are extracted based on the convolutional autoencoder, and the data features extracted by the convolutional autoencoder are obtained, which are described by formula (3): c k =f(∑ k x*w k +b) (3), Where x represents the input data, w k represents the weight parameters of the convolution kernel, b represents the bias parameters of the convolution kernel, and f() represents the activation function; The data features extracted by the convolutional autoencoder are reconstructed to obtain reconstructed data, which is described by formula (4): c'=f(∑ k c k *w k +b) (4), Where c' represents the output of data reconstruction, c k w represents the input to the deconvolutional layer. k represents the weight parameters of the deconvolutional layer kernel, and b represents the bias parameters; The reconstruction error is calculated based on the input data and the reconstructed data, and is described by formula (5): x i Indicates the input data, y i This indicates that the output data is also the reconstructed data; Among them, a single-channel two-dimensional convolution kernel slides along the variable and time dimensions to perform convolution operations, which can be used to simultaneously extract the cross-correlation and autocorrelation between adjacent variables. Multi-channel one-dimensional convolution kernels are used to extract process data features, and each kernel matrix of the convolution kernel is used to extract the autocorrelation of variables. The fixed-bed reactor is a three-dimensional device in which raw materials flow from top to bottom through the reactor. The reaction and transfer both occur in a three-dimensional space. Measuring instruments are installed at different positions in the fixed-bed reactor. According to the spatial distribution characteristics of the measuring instruments, the measured variables are arranged into a two-dimensional data matrix. The two-dimensional data matrices at l time points are stacked into a three-dimensional time series data matrix as the offline input data of the process monitoring model at time t. Taking advantage of the local feature extraction of the convolution kernel, the three-dimensional time series data matrix obtained by the new variable arrangement is locally extracted. The two-dimensional data matrices at l time points are stacked into a three-dimensional time series data matrix. A moving window of length L is used to extract the spatial correlation between monitoring variables on the three-dimensional time series data matrix. The online monitoring includes the following steps: Online real-time acquisition of process data from different spaces inside a fixed-bed reactor under operating conditions; The process data is preprocessed to obtain online input data; Based on the online input data and the process monitoring model, online results are obtained; The online results are compared with the offline results to determine whether the fixed-bed reactor has malfunctioned.

2. The fault early warning method for a fixed-bed reactor industrial process according to claim 1, characterized in that, Before performing data preprocessing on the historical process data to obtain offline input data, the process also includes: The historical data of the process are normalized and described by formula (1):

3. The fault early warning method for a fixed-bed reactor industrial process according to claim 1, characterized in that, The process monitoring model based on the convolutional autoencoder is established based on the offline input data, and offline results are obtained, including: A portion of the offline input data was used as the test set. The test set is used as input data for the process monitoring model, and the reconstruction error of the process monitoring model under normal operating conditions is calculated based on the process monitoring model. The threshold value of the reconstruction error within the 99% confidence interval is selected as the control limit of the process monitoring model, and is described by formulas (6) and (7): Wherein, the control limit refers to the threshold of reconstruction error obtained under normal operating conditions, which is used to verify whether the collected online data is safe, h represents the smoothing parameter of the kernel density estimation function, n represents the number of reconstruction error samples, and MSE represents the reconstruction error.

4. The fault early warning method for a fixed-bed reactor industrial process according to any one of claims 1-3, characterized in that, The process data preprocessing to obtain online input data includes: The process data of different spaces inside the fixed bed reactor are collected online in real time, and the process data are normalized. The normalized process data is preprocessed to obtain an enhanced three-dimensional time series data matrix, which is used as the online input data for the process monitoring model.

5. The fault early warning method for a fixed-bed reactor industrial process according to claim 4, characterized in that, Based on the online input data and the process monitoring model, online results are obtained, including: The online input data is input into the process monitoring model to obtain the reconstruction error.

6. The fault early warning method for a fixed-bed reactor industrial process according to claim 5, characterized in that, The step of comparing the obtained online results with the offline results to determine whether the fixed-bed reactor has malfunctioned includes: The reconstruction error obtained from the online input data is compared with the control limit; If the reconstruction error is less than the control limit, it indicates that the process is still under normal operating conditions; otherwise, it indicates that the fixed-bed reactor has malfunctioned.

7. A fault early warning system for a fixed-bed reactor industrial process, characterized in that, The system includes: The training module is used for offline modeling. The monitoring module is used for online monitoring. The training module includes the following units: The acquisition unit is used to acquire historical process data of different spaces inside the fixed-bed reactor under normal operating conditions. The first processing unit is configured to preprocess the historical process data to obtain offline input data; the preprocessing of the historical process data to obtain offline input data includes: Based on the different spatial distributions of the process history data inside the fixed-bed reactor, a two-dimensional data matrix is ​​arranged. Two-dimensional data matrices at different times are stacked into a three-dimensional time series data matrix to obtain offline input data; The spatial correlation between the historical process data is extracted based on the three-dimensional time series data matrix; the two-dimensional data matrix is ​​preprocessed using data augmentation techniques based on the Laplace operator to enhance the distinction between data under normal and fault conditions, as described by formula (2.1): Where I(p,q) and V(p,q) represent the variable values ​​in the original two-dimensional data matrix and the enhanced two-dimensional data matrix, respectively; The first computing unit is used to establish a process monitoring model based on a convolutional autoencoder based on the offline input data and obtain offline results; the establishment of the process monitoring model based on the convolutional autoencoder based on the offline input data and obtaining offline results includes: Use a portion of the offline input data as the training set; Using the training set as input data, the autocorrelation, cross-correlation, and spatial correlation between the variables in the training set are extracted based on the convolutional autoencoder, and the data features extracted by the convolutional autoencoder are obtained, which are described by formula (3): c k =f(∑ k x*w k +b) (3), Where x represents the input data, w k represents the weight parameters of the convolution kernel, b represents the bias parameters of the convolution kernel, and f() represents the activation function; The data features extracted by the convolutional autoencoder are reconstructed to obtain reconstructed data, which is described by formula (4): c'=f(∑ k c k *w k +b) (4), Where c' represents the output of data reconstruction, c k w represents the input to the deconvolutional layer. k represents the weight parameters of the deconvolutional layer kernel, and b represents the bias parameters; The reconstruction error is calculated based on the input data and the reconstructed data, and is described by formula (5): x i Indicates the input data, y i This indicates that the output data is also the reconstructed data; Among them, a single-channel two-dimensional convolution kernel slides along the variable and time dimensions to perform convolution operations, which can be used to simultaneously extract the cross-correlation and autocorrelation between adjacent variables. Multi-channel one-dimensional convolution kernels are used to extract process data features, and each kernel matrix of the convolution kernel is used to extract the autocorrelation of variables. The fixed-bed reactor is a three-dimensional device in which raw materials flow from top to bottom through the reactor. The reaction and transfer both occur in a three-dimensional space. Measuring instruments are installed at different positions in the fixed-bed reactor. According to the spatial distribution characteristics of the measuring instruments, the measured variables are arranged into a two-dimensional data matrix. The two-dimensional data matrices at l time points are stacked into a three-dimensional time series data matrix as the offline input data of the process monitoring model at time t. Taking advantage of the local feature extraction of the convolution kernel, the three-dimensional time series data matrix obtained by the new variable arrangement is locally extracted. The two-dimensional data matrices at l time points are stacked into a three-dimensional time series data matrix. A moving window of length L is used to extract the spatial correlation between monitoring variables on the three-dimensional time series data matrix. The monitoring module includes the following units: The data acquisition unit is used to collect process data in different spaces inside the fixed-bed reactor under operating conditions in real time online. The second processing unit is used to preprocess the process data to obtain online input data; The second calculation unit is used to obtain online results based on the online input data and the process monitoring model; The result determination unit compares the online results with the offline results to determine whether the fixed-bed reactor has malfunctioned.

Images