Working condition monitoring method and system for petrochemical industrial process and storage medium

Abstract

The invention provides a working condition monitoring method for a petrochemical industrial process, a working condition monitoring system for the petrochemical industrial process and a computer readable storage medium. The working condition monitoring method for the petrochemical industrial process comprises the following steps: screening key variables in acquired data of a catalytic cracking device based on an expert experience method and an XGBOOST ensemble learning method so as to construct a sample set; inputting the sample set into a monitoring model to obtain a monitoring statistic sequence, wherein the monitoring model fuses a gating circulation unit, an adversarial generative auto-encoder and co-integration analysis; and performing tail distribution scanning of the monitoring statistic sequence by using generalized Pareto distribution, and fitting tail features of the monitoring statistic sequence so as to determine a dynamic control threshold value for working condition anomaly identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of petrochemical industry process monitoring, and in particular to a working condition monitoring method for petrochemical industry process, a working condition monitoring system for petrochemical industry process, and a computer readable storage medium. BACKGROUND

[0002] As one of the core devices of oil refining process, the catalytic cracking process is affected by complex and changeable working conditions, numerous coupled process variables and strict performance indicators in actual operation, resulting in highly dynamic working condition state and difficult to be stably identified. In addition, the properties of the raw materials processed by the catalytic cracking device have strong volatility, resulting in that the process data of the catalytic cracking device presents a significant non-Gaussian distribution characteristic, therefore, the model based on the traditional Gaussian assumption is difficult to accurately depict the potential law of the process data of the catalytic cracking device.

[0003] At the same time, the catalytic cracking process has the characteristics of long-period continuous operation, and shows strong non-stationarity in the time scale. Many process variables not only change over time in statistical characteristics such as mean and variance, but also have potential cointegration relationship between multiple single integrated sequences of the same order. This long-term stable linear structure is often ignored by conventional modeling methods, thereby affecting the accuracy of working condition division and abnormal monitoring. Moreover, there are complex time sequence dependence and nonlinear coupling structure between a large amount of measurement point data, which further increases the difficulty of dynamic working condition modeling.

[0004] The dynamics of the catalytic cracking process data mainly reflects the time autocorrelation of the variables, that is, the statistical correlation degree between the current and historical time of a variable.

[0005] If the model fails to fully consider the non-Gaussian, non-stationarity and time correlation of the data, the accuracy and stability of the working condition monitoring will be significantly affected.

[0006] In order to overcome the above-mentioned defects existing in the prior art, there is an urgent need in the field for a working condition monitoring technology for petrochemical industry process, which can significantly improve the accuracy and stability of working condition monitoring by a monitoring model capable of fully considering the non-Gaussian, non-stationarity and time correlation of the data, has high false alarm control ability and fault recognition accuracy, and improves the dynamic monitoring ability of the catalytic cracking process under complex working conditions. SUMMARY

[0007] The following presents a simplified summary of one or more aspects in order to provide a basic understanding of such aspects. This summary is not an extensive overview of all contemplated aspects, and is intended to neither identify key or critical elements of all aspects nor delineate the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed description that is presented later.

[0008] In order to overcome the above-mentioned defects of the prior art, the present application provides a working condition monitoring method for petrochemical industrial processes, a working condition monitoring system for petrochemical industrial processes, and a computer-readable storage medium, which can significantly improve the accuracy and stability of working condition monitoring, and has higher false alarm control ability and fault identification accuracy by a monitoring model capable of fully considering the non-Gaussian, non-stationary and time correlation of data.

[0009] Specifically, the above-mentioned working condition monitoring method for petrochemical industrial processes according to the first aspect of the present application comprises the steps of: screening key variables in the data of a catalytic cracking unit obtained based on an expert experience method and an XGBOOST integrated learning method to construct a sample set; inputting the sample set into a monitoring model to obtain a monitoring statistic sequence, the monitoring model fusing a gated recurrent unit, an adversarial generative autoencoder and a cointegration analysis; and performing tail distribution scanning of the monitoring statistic sequence by using a generalized Pareto distribution, fitting tail features of the monitoring statistic sequence, to determine a dynamic control threshold for working condition anomaly identification.

[0010] Preferably, in an embodiment of the present application, the step of inputting the sample set into a monitoring model to obtain a monitoring statistic sequence comprises: determining a hidden state by using the gated recurrent unit to extract dynamic time sequence features of data in the sample set; inputting the hidden state into the adversarial generative autoencoder to capture the structure of dynamic non-Gaussian latent variables in the catalytic cracking process, wherein the adversarial training of the adversarial generative autoencoder is based on a total loss function of the monitoring model, the total loss function of the monitoring model including a reconstruction loss function of steady-state features determined by using the cointegration analysis; and determining the monitoring statistic sequence based on the steady-state features and the latent variables.

[0011] Preferably, in an embodiment of the present application, the total loss function is:

[0012] L total =L recon +λ sta L sta +L adv ,

[0013] wherein, L reconis a reconstruction loss function of the original data in the sample set, L sta is a reconstruction loss function of the steady-state feature, L adv is an adversarial loss term, λ sta is a weight coefficient.

[0014] Preferably, in an embodiment of the present application, the latent variable hidden distribution is:

[0015] Q φ (z|h t )=P(z),

[0016] wherein z is a latent variable, h t is a hidden state output by the gating recurrent unit at time t, Q φ is a conditional probability distribution under the encoder parameter φ, and P(z) is a prior probability distribution of the latent variable z.

[0017] Preferably, in an embodiment of the present application, the steady-state feature extraction formula is:

[0018] ξ t =β T x t =β1x 1t +β2x 2t +...+β m x mt ,

[0019] wherein ξ t is a steady-state feature sequence, x t is a cointegration sequence, and β is a cointegration vector.

[0020] Preferably, in an embodiment of the present application, the monitoring statistic sequence comprises one or more of a latent variable space statistic, a steady-state feature space offset statistic, a squared prediction error in the original space, and a reconstruction error of the steady-state feature, wherein,

[0021] The calculation formula of the latent variable space statistic is:

[0022]

[0023] wherein H 2 is a latent variable space statistic, and W is a weight matrix.

[0024] The calculation formula of the steady-state feature space offset statistic is:

[0025]

[0026] wherein, is a steady-state feature space offset statistic, ζ is a variable vector in the steady-state feature space, and μζ It is the mean vector of the variable vectors;

[0027] The formula for calculating the squared prediction error of the original space is:

[0028] SPE=||xx′|| 2 ,

[0029] Where SPE is the squared prediction error of the original space, x is the sample data, and x′ is the predicted data; and

[0030] The formula for calculating the reconstruction error of steady-state characteristics is:

[0031] SPE ζ =||ζ i -ζ′ i || 2 ,

[0032] Among them, SPE ζ It is the reconstruction error of the steady-state characteristics, ζ i ζ′ is the i-th actual feature variable in the steady-state feature space. i It is the i-th feature variable in the reconstructed steady-state feature space.

[0033] Preferably, in one embodiment of the present invention, the step of constructing the sample set includes: performing max-min normalization on the data of the sample set to eliminate the influence of dimensions and improve the convergence of the monitoring model.

[0034] Preferably, in one embodiment of the present invention, the formula for calculating the dynamic control threshold is:

[0035]

[0036] Among them, y α The threshold is dynamically controlled, θ is the initial threshold, σ is the scale parameter describing the scale characteristics of the tail distribution (σ>0), k is the shape parameter characterizing the tail distribution properties (when k>0, the tail distribution exhibits heavy tails), α is the given significance level, and N... θ It is the set of data points that exceed the initial threshold.

[0037] Furthermore, the above-described operating condition monitoring system for petrochemical industrial processes provided according to a second aspect of the present invention includes a memory and a processor. The memory stores computer instructions. The processor is connected to the memory and configured to execute the computer instructions stored in the memory to implement the operating condition monitoring method for petrochemical industrial processes provided in any of the above embodiments.

[0038] Furthermore, the computer-readable storage medium provided according to the third aspect of the present invention stores computer instructions. When the computer instructions are executed by a processor, the operating condition monitoring method for petrochemical industrial processes provided in any of the above embodiments is implemented. Attached Figure Description

[0039] The above-described features and advantages of the present invention will be better understood after reading the following detailed description of embodiments of the present disclosure in conjunction with the accompanying drawings. In the drawings, components are not necessarily drawn to scale, and components having similar related characteristics or features may have the same or similar reference numerals.

[0040] Figure 1 A schematic diagram of a condition monitoring system for petrochemical processes according to some embodiments of the present invention is shown.

[0041] Figure 2 A flowchart of a method for monitoring operating conditions in petrochemical processes according to some embodiments of the present invention is shown.

[0042] Figure 3 A structural diagram of a monitoring model provided according to some embodiments of the present invention is shown;

[0043] Figure 4 A flowchart illustrating the process of obtaining a monitoring statistics sequence through a monitoring model, according to some embodiments of the present invention, is shown.

[0044] Figure 5 A schematic diagram of the catalytic cracking unit is shown;

[0045] Figure 6 The diagram shows a comparison of various models provided according to embodiments of the present invention when monitoring technical and economic indicators and monitoring gasoline production; and

[0046] Figure 7 A schematic diagram of the Tennessee Eastman process is shown.

[0047] Figure label:

[0048] 100: Operating condition monitoring systems used in petrochemical processes;

[0049] 110: Memory;

[0050] 111: Computer-readable storage medium;

[0051] 120: Processor;

[0052] 200: Methods for monitoring operating conditions in petrochemical processes;

[0053] S210~S230: Steps; and

[0054] S221~S223: Steps. Detailed Implementation

[0055] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be noted that the aspects described below with reference to the accompanying drawings and specific embodiments are merely exemplary and should not be construed as limiting the scope of protection of the present invention in any way.

[0056] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0057] Furthermore, the terms "upper," "lower," "left," "right," "top," "bottom," "horizontal," and "vertical" used in the following description should be understood as the orientations shown in the relevant paragraphs and accompanying drawings. These relative terms are for illustrative purposes only and do not imply that the described apparatus must be manufactured or operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0058] It is understood that although terms such as "first," "second," and "third" may be used herein to describe various components, regions, layers, and / or parts, these components, regions, layers, and / or parts should not be limited by these terms, and these terms are only used to distinguish different components, regions, layers, and / or parts. Therefore, the first components, regions, layers, and / or parts discussed below may be referred to as second components, regions, layers, and / or parts without departing from some embodiments of the present invention.

[0059] As mentioned above, if the model fails to fully consider the non-Gaussianity, non-stationarity, and time correlation of the data, it will significantly affect the accuracy and stability of the operating condition monitoring.

[0060] To overcome the aforementioned deficiencies in the existing technology, this invention provides a method for monitoring operating conditions in petrochemical industrial processes, a system for monitoring operating conditions in petrochemical industrial processes, and a computer-readable storage medium. By using a monitoring model that fully considers the non-Gaussianity, non-stationarity, and time correlation of data, the accuracy and stability of operating condition monitoring are significantly improved, and it has high false alarm control capability and fault identification accuracy.

[0061] In some non-limiting embodiments, the above-described method for monitoring operating conditions in petrochemical processes provided in the first aspect of the present invention can be implemented via the above-described system for monitoring operating conditions in petrochemical processes provided in the second aspect of the present invention.

[0062] Please refer to Figure 1 , Figure 1 A schematic diagram of a condition monitoring system for petrochemical processes, provided according to some embodiments of the present invention, is shown.

[0063] like Figure 1 As shown, the operating condition monitoring system 100 for petrochemical industrial processes may be configured with a memory 110 and a processor 120. The memory 110 includes, but is not limited to, the computer-readable storage medium 111 described in the third aspect of the present invention, which stores computer instructions thereon. The processor 120 is connected to the memory 110 and configured to execute the computer instructions stored in the memory 110 to implement the operating condition monitoring method for petrochemical industrial processes provided in the first aspect of the present invention.

[0064] The working principle of the above-mentioned operating condition monitoring system for petrochemical processes will first be described with reference to some embodiments of operating condition monitoring methods for petrochemical processes. Those skilled in the art will understand that these embodiments of operating condition monitoring methods for petrochemical processes are merely non-limiting implementations provided by the present invention, intended to clearly demonstrate the main concepts of the invention and provide specific solutions convenient for public implementation, rather than limiting all functions or operating methods of the operating condition monitoring system for petrochemical processes. Similarly, the operating condition monitoring system for petrochemical processes is also only one non-limiting implementation provided by the present invention, and does not constitute a limitation on the executing entity and execution order of the steps in these operating condition monitoring methods for petrochemical processes.

[0065] Please refer to Figure 2 , Figure 2 A flowchart of a method for monitoring operating conditions in a petrochemical process, according to some embodiments of the present invention, is shown.

[0066] like Figure 2 As shown, the operating condition monitoring method 200 for petrochemical industrial processes may include step S210: screening key variables in the acquired data of the catalytic cracking unit based on expert experience and XGBOOST (eXtreme Gradient Boosting) ensemble learning method to construct a sample set.

[0067] The operating condition monitoring system can collect time-series process data and experimental analysis system data of the catalytic cracking unit as raw data of the catalytic cracking unit. Then, based on the expert experience method and the XGBOOST ensemble learning method, the system can screen the raw data to identify key variables and construct a sample set based on the key variables.

[0068] A sample set can be specifically defined as:

[0069] X = [x(1),x(2),...,x(n)] T ∈R n×m ,

[0070] Where X is the sample set input to the monitoring model, x(n) is the nth data sample in the sample set, and m represents the feature dimension of each sample.

[0071] In some embodiments, the step of constructing the sample set may further include data processing. Preferably, the operating condition monitoring system may perform max-min normalization on the data to eliminate the influence of dimensions and improve the convergence of subsequent monitoring models. The formula for calculating max-min normalization is:

[0072]

[0073] Where, x i It is the i-th data after normalization, x i ′ is the i-th original data, x min and x max These are the minimum and maximum values ​​in the sample set, respectively.

[0074] Then, the operating condition monitoring system can execute step S220: input the sample set into the monitoring model to obtain the monitoring statistics sequence. The monitoring model integrates gated cyclic units, adversarial generative autoencoders, and cointegration analysis.

[0075] Understandably, the sample set input to the monitoring model can be the sample set that has undergone the data processing described above.

[0076] Please refer to Figure 3 and Figure 4 , Figure 3 A structural diagram of a monitoring model provided according to some embodiments of the present invention is shown. Figure 4 A flowchart illustrating the process of obtaining a monitoring statistics sequence through a monitoring model, according to some embodiments of the present invention, is shown.

[0077] like Figure 3As shown, the monitoring model (GRU-AAE-CA) can integrate a gated recurrent unit (GRU), an adversarial autoencoder (AAE), and cointegration analysis (CA).

[0078] Specifically, such as Figure 4 As shown, the operating condition monitoring system can execute step S221: use the gated loop unit to determine the hidden state in order to extract the dynamic temporal characteristics of the data in the sample set.

[0079] The operating condition monitoring system uses a gated loop unit to obtain the hidden state. The formula for calculating the hidden state is as follows:

[0080] h t =GRU(x t ,h t-1 ),

[0081] Among them, h t It is the hidden state at time t, x t The input data at time t, h t-1 It represents the hidden state at time t-1. GRU is a gated recurrent unit network model.

[0082] Specifically, in the gated recurrent unit network model, the calculation process for each time step t can be as follows:

[0083] The gated recurrent unit network model can include an update gate. The update gate controls the hidden state h from the previous time step. t-1 The retention rate at the current moment determines how much past information the gated recurrent unit network model remembers. The expression for updating the gate can be:

[0084] z t =σ(W z x t +U z h t-1 +b z ),

[0085] Among them, z t It updates the gate value, σ is the activation function, and W... z It is the weight matrix, x t The input data at time t, h t-1 It is the hidden state at time t-1, U z It is the weight matrix corresponding to the previous time step, b z It is the bias term in the update gate calculation.

[0086] When updating the value z of the gate tAs the value approaches 1, the gated recurrent unit network model can retain more of the hidden state h from the previous time step. t-1 When updating the value z of the gate t If the value is close to 0, it is more likely to use the current input x. t Calculate the new state.

[0087] The gated loop unit network model can include a reset gate. The hidden state h at the previous control moment... t-1 Calculate candidate hidden state At that time, by resetting the gate r t The hidden state h that is applied in the previous time step t-1 This controls the influence of information from the previous moment. The expression for the reset gate can be:

[0088] r t =σ(W r x t +U r h t-1 +b r ),

[0089] Where, r t It resets the gate value, σ is the activation function, and W... r It is the weight matrix, x t The input data at time t, h t-1 It is the hidden state at time t-1, U r It is the weight matrix corresponding to the previous time step, b r It is the bias term in the reset gate calculation.

[0090] When the value of the door is reset r t When the value is close to 1, the gated recurrent unit network model can allow the hidden state h from the previous time step. t-1 Affects the current state; when the gate value r is reset t When the value is close to 0, the hidden state h from the previous time step is ignored. t-1 This causes the gated recurrent unit network model to tend to rely only on the current input x. t Calculate the new state.

[0091] The gated recurrent unit network model calculates the candidate hidden states at the current time step. At that time, by resetting the gate r t The hidden state h that is applied in the previous time step t-1 This is used to control the influence of information from the previous time step. The candidate hidden state at the current time step is calculated. The expression can be:

[0092]

[0093] in, It represents the candidate hidden state at the current time step, tanh is the hyperbolic tangent activation function, and W... h and U h It is the weight matrix, b h It is the bias term used to calculate the candidate hidden state.

[0094] When the value of the door is reset r t When the value is close to 1, the calculated candidate hidden state is... Simultaneously depends on the current input x t and the hidden state h from the previous moment t-1 When the value of the gate is reset r t When it is close to 0, the candidate hidden state Mainly determined by the current input x t Decide.

[0095] The final hidden state h t The value z of the updated gate can be used to update the gate. t Control is achieved by updating the gate to determine the merging of the current hidden state with the previous hidden state:

[0096]

[0097] Where, when updating the value z of the gate t When the value is close to 1, it relies more on the candidate hidden state. When updating the value z of the gate t When the value is close to 0, more of the hidden state h from the previous time step is retained. t-1 Information.

[0098] For each time step t of input data, the reset gate of the gated loop unit first determines whether to reset the hidden state; the update gate controls the degree to which the hidden state from the previous time step is retained; through the combined action of the two, the hidden state at the current time step is dynamically updated. By continuously updating the hidden state in this way, the gated loop unit can capture the dynamic characteristics of data changes over time, such as the trend and periodicity of data changes, thereby extracting the dynamic temporal characteristics of the data.

[0099] Please continue to refer to this. Figure 4 The operating condition monitoring system can execute step S222: input the hidden state into the adversarial generative autoencoder to capture the dynamic non-Gaussian latent variable structure during catalytic cracking.

[0100] like Figure 3 As shown, an adversarial generative autoencoder can include an encoder, a decoder, and a discriminator, which are used to implement data compression, data reconstruction, and constrained optimization of latent variable distribution, respectively.

[0101] The operating condition monitoring system can detect the hidden state h output by the gated loop unit. t The input is fed into an adversarial generative autoencoder, which then uses an encoder to process the hidden state h of the input. t The input data is mapped to a latent space (typically a low-dimensional representation), for example, as latent variables. Then, the decoder remaps the latent space representation back to the data space to reconstruct the input data. In other words, the latent variables output by the encoder are restored to a form similar to the original data.

[0102] Here, the distribution of this latent variable can be:

[0103] Q φ (z|h t )=P(z),

[0104] Where z is a latent variable, h t Q is the hidden state output by the gated loop unit at time t. φ P(z) is the conditional probability distribution under encoder parameters φ, and P(z) is the prior probability distribution of the latent variable z.

[0105] In some embodiments, the encoder of an adversarial generative autoencoder may attempt to make the latent variable z conform to a specified prior probability distribution P(z), preferably, the prior probability distribution P(z) may be a standard Gaussian distribution:

[0106]

[0107] Where I is the unit covariance matrix, representing that each dimension is independent of the others and has unit variance.

[0108] The discriminator can distinguish whether the input latent variable comes from the prior probability distribution P(z) or the distribution generated by the encoder.

[0109] Adversarial training of adversarial generative autoencoders can be achieved through dynamic game between the discriminator and the generator (encoder and decoder). Through multiple rounds of game, the encoder is forced to learn the latent structure (including non-Gaussian properties) in the data in order to capture the dynamic non-Gaussian latent variable structure in the catalytic cracking process. The latent variable distribution generated in the end will approximate the latent variable distribution of the real data (even if the latent variable distribution of the real data is non-Gaussian).

[0110] Here, adversarial training of the adversarial generative autoencoder (AAE) is based on the total loss function of the monitoring model, which includes the reconstruction loss function of the steady-state features, which are determined using cointegration analysis.

[0111] Steady-state features can be extracted using cointegration analysis from non-stationary variables in a sample set. Figure 3 In this context, non-stationary variables in the sample set can be... Where b is the number of non-stationary variables.

[0112] Preferably, non-stationary variables can be identified using the Augmented Dickey-Fuller (ADF) test. The ADF test is a classic unit root test widely used to determine the stationarity of time series data, particularly suitable for determining whether variables in industrial process data possess a unit root. The null hypothesis of the ADF test is the existence of a unit root (i.e., non-stationarity). Rejecting the null hypothesis indicates that the variable is stationary, while retaining the null hypothesis (i.e., the test statistic is greater than the critical value) indicates that the variable is non-stationary. The ADF test can identify unit roots for key variables in the input sample set. If a unit root exists, i.e., the null hypothesis is retained, it indicates that the key variable has a random trend and is a non-stationary variable. Therefore, the operating condition monitoring system can distinguish between stationary and non-stationary variables among the key variables.

[0113] The formula for extracting steady-state features can be:

[0114] ξ t =β T x t =β1x 1t +β2x 2t +...+β m x mt ,

[0115] Where, ξ t It is a steady-state characteristic sequence, x t The cointegrated sequence is based on non-stationary variables, and β is the cointegrating vector. Preferably, the operating condition monitoring system can determine the validity of the cointegrating vector β using stationarity tests such as the ADF test, thereby demonstrating the effectiveness of the steady-state characteristic sequence ξ. t With cointegrated sequence x t It is cointegrated, thus yielding the final steady-state characteristics.

[0116] In some embodiments, the operating condition monitoring system can represent and analyze the above-mentioned cointegration relationship using a vector error correction model, thereby describing the dynamic process of non-stationary variables adjusting to a long-term equilibrium state. The vector error correction model can be as follows:

[0117]

[0118] Where, △x t =x t -x t-1Γ represents the sequence after difference. i Here, μ is the coefficient matrix for short-term fluctuations, α is the error correction coefficient matrix, used to describe the speed at which the sequence adjusts towards long-term equilibrium. t The error term is represented, usually assumed to be white noise, and P is the maximum hysteresis order selected by the model, which is usually determined by the AIC or BIC criterion.

[0119] Among the parameters of the above vector error correction model, the cointegration vector β can be solved using the maximum likelihood estimation method. Specifically, the estimation problem of the cointegration vector β can be transformed into the following eigenvalue problem:

[0120]

[0121] Where λ is the eigenvalue, S 11 S 10 S 00 and S 01 These are the matrix elements constructed from time series data, which are based on non-stationary variables.

[0122] The operating condition monitoring system can solve the above eigenvalue problem to obtain several eigenvalues ​​and their corresponding eigenvectors. These eigenvectors are the candidate cointegrating vectors β. The candidate cointegrating vectors β are then verified using the ADF test, retaining the values ​​that make the above formula ξ... t =β T x t The cointegration vector β of the stationary sequence is an effective cointegration vector, thus yielding the final steady-state characteristic ξ. t .

[0123] like Figure 3 As shown, in some embodiments, the operating condition monitoring system can utilize cointegration analysis to extract reconstructed steady-state features from the reconstructed sample set data, thereby obtaining the reconstructed steady-state features. Obtaining the reconstructed steady-state features in this way can enhance the structural clarity and inference efficiency of the monitoring model.

[0124] In other embodiments, the operating condition monitoring system can utilize a dual-encoder mechanism to obtain reconstructed steady-state features. Because the distribution characteristics of dynamic temporal features and steady-state features differ significantly (the former is non-stationary and strongly non-Gaussian, while the latter is a stationary linear combination), setting only a single encoder in an adversarial generative autoencoder (AAE) may be insufficient to simultaneously adapt to both distributions. Therefore, an AAE can design two encoders to encode the dynamic temporal features formed from data based on the original sample set and the steady-state features extracted from non-stationary variables, respectively. This ensures that the latent variables of both features can approximate the prior distribution through adversarial training, ultimately improving the monitoring model's ability to monitor complex operating conditions.

[0125] In some embodiments, the operating condition monitoring system uses these steady-state features as another part of the input to the fusion model of the gated recurrent unit network (GRN) and the adversarial generative autoencoder (AAE). The steady-state features can be used to obtain latent variables through the GRN, and the structure of the latent variables obtained based on the steady-state features during catalytic cracking can be captured by the AAE. The reconstructed steady-state features are then obtained through a decoder. This dual-encoder mechanism can improve the robustness of the monitoring model, or it can be used in the evaluation stage after the monitoring model is deployed. This does not conflict with the aforementioned method of extracting steady-state features from the reconstructed sample set data using cointegration analysis. Those skilled in the art can choose to adopt one or both methods according to actual design requirements.

[0126] Subsequently, the operating condition monitoring system can introduce the reconstruction loss function of steady-state features into the total loss function of the monitoring model to achieve synchronous learning of key variables and steady-state features in the sample set.

[0127] The total loss function L of the monitoring model total It can be:

[0128] L total =L recon +λ sta L sta +L adv ,

[0129] Among them, L recon L is the reconstruction loss function of the original data. sta It is the reconstruction loss function of steady-state features, L adv It is the adversarial loss term, λ sta These are the weighting coefficients.

[0130] like Figure 3 As shown, the reconstruction loss function L of the original data recon It can be:

[0131]

[0132] Where, x j For the original data sample, x′ j The data sample is the reconstructed data sample, and n is the number of samples in the sample set.

[0133] Reconstruction loss function L of steady-state features sta It can be:

[0134]

[0135] Where, ζ q For the original steady-state characteristics, ζ′ q denoted as the reconstructed steady-state feature, and b represents the number of samples in the sample set.

[0136] Countermeasure loss term L adv It can be:

[0137]

[0138] Among them, D ψ It is the objective function of the discriminator D under parameter ψ. Let P(z) represent the expectation of a random latent variable z following a distribution P(z). This indicates the hidden state h t Expectation h t En is the hidden state of the gated loop unit output at time t. φ It is an encoder function with φ as a parameter.

[0139] By minimizing the total loss function, the monitoring model can simultaneously preserve the nonlinear temporal characteristics of the data and the implicit steady-state cointegration structure, thereby achieving a more accurate and robust operating condition identification capability when faced with data that has non-Gaussian properties and non-stationary dynamics.

[0140] Subsequently, based on the total loss function, adversarial training is achieved through the discriminator in the adversarial generative autoencoder. The parameters of the encoder, decoder and discriminator are continuously adjusted based on the optimization objective function so that the latent variable z generated by the encoder can better conform to the specified prior probability distribution P(z), thereby realizing nonlinear modeling of the latent space and capturing the latent variable structure with dynamic non-Gaussianity in the catalytic cracking process.

[0141] The aforementioned optimization objective function may include the optimization objective function of the encoder and decoder, and the optimization objective function of the discriminator.

[0142] The optimization objective functions for the encoder and decoder can be:

[0143]

[0144] Here, δ represents the decoder parameters, and Deδ is the decoder function with δ as a parameter. During training, the optimization objective function of the encoder and decoder is to minimize the reconstruction error, that is, to make the original input data and the reconstructed data as close as possible.

[0145] The objective function for optimizing the discriminator can be:

[0146]

[0147] The objective function of the discriminator is to maximize the discrimination accuracy, that is, to accurately distinguish between latent variables of the true distribution and latent variables generated by the encoder.

[0148] Here, since the total loss function of the monitoring model introduces a reconstruction loss function with steady-state features, the working condition monitoring system needs to optimize the adversarial loss term of the adversarial training in conjunction with the reconstruction loss function of the steady-state features, so as to achieve synchronous learning of dynamic features and steady-state features.

[0149] Thus, during the training of the monitoring model, the parameters of the gated recurrent unit and the adversarial generative autoencoder are optimized alternately. The monitoring model captures the temporal dynamic information of the process data through the gated recurrent unit and achieves effective distribution mapping of latent variables through the adversarial generative autoencoder, thereby accurately capturing the dynamic non-Gaussian latent variable structure in the catalytic cracking process and effectively improving the ability to identify abnormal operating conditions.

[0150] like Figure 4 As shown, the operating condition monitoring system continues to execute step S223: based on steady-state characteristics and latent variables, determine the monitoring statistics sequence.

[0151] Specifically, the determined sequence of monitoring statistics may include one or more of the following: latent variable spatial statistics, offset statistics of steady-state feature space, squared prediction error of the original space, and reconstruction error of steady-state features.

[0152] The determined monitoring statistics sequence can be flexibly selected according to the needs of the industrial system. For example, in some embodiments, when the industrial system has a significant cointegration relationship or needs to enhance the detection capability of long-term steady-state structural changes, the offset statistics of the steady-state space and the reconstruction error of the steady-state characteristics can be preferred as evaluation indicators.

[0153] The formula for calculating the spatial statistic of latent variables can be:

[0154]

[0155] Among them, H 2 It is the spatial statistics of latent variables, and W is the weight matrix.

[0156] The formula for calculating the offset statistics of the steady-state characteristic space can be:

[0157]

[0158] in, It is the offset statistic of the steady-state characteristic space, ζ is the vector of each variable in the steady-state characteristic space, and μ is the offset statistic. ζ It is the mean vector of all variable vectors.

[0159] The formula for calculating the squared prediction error of the original space can be:

[0160] SPE=||xx′|| 2 ,

[0161] Where SPE is the squared prediction error of the original space, x is the sample data, and x′ is the reconstructed sample data.

[0162] The formula for calculating the reconstruction error of steady-state characteristics can be:

[0163] SPE ζ =||ζ j -ζ′ j || 2 ,

[0164] Among them, SPE ζ It is the reconstruction error of the steady-state characteristics, ζ j ζ′ is the j-th actual feature variable in the steady-state feature space. j It is the j-th feature variable in the reconstructed steady-state feature space.

[0165] Please continue to refer to this. Figure 2 After determining the monitoring statistics sequence, the operating condition monitoring system can perform step S230: use the generalized Pareto distribution to scan the tail distribution of the monitoring statistics sequence, fit the tail features of the monitoring statistics sequence, and determine the dynamic control threshold for identifying operating condition anomalies.

[0166] Here, the formula for estimating the dynamic control threshold can be:

[0167]

[0168] Among them, y α The threshold is dynamically controlled, θ is the initial threshold, ρ is the scale parameter describing the scale characteristics of the tail distribution (ρ>0), k is the shape parameter characterizing the tail distribution characteristics (when k>0, the tail distribution exhibits heavy tails), τ is the given significance level, and N... θ This refers to the set of data points that exceed the initial threshold θ. Here, the initial threshold θ can be chosen based on the distribution characteristics of the sample set, for example, using the 95%, 97.5%, or 99% quantiles as the initial threshold. Alternatively, in other embodiments, the initial threshold θ can be chosen empirically to ensure that the number of tail samples meets the fitting requirements of the generalized Pareto distribution.

[0169] Thus, by using this dynamic control threshold, the operating condition monitoring system can identify abnormal conditions in the catalytic cracking unit.

[0170] In summary, the operating condition monitoring method for petrochemical processes provided by this invention utilizes a novel dynamic operating condition monitoring model (GRU-AAE-CA) that integrates a gated cyclic unit, an adversarial generative autoencoder, and cointegration analysis to identify anomalies in the operating conditions of catalytic cracking units. This monitoring model possesses superior capabilities in identifying and monitoring complex operating conditions. The gated cyclic unit effectively extracts dynamic features and time-series dependencies in the catalytic cracking process, while the introduced cointegration analysis method extracts linear combinations with long-term steady-state characteristics, effectively mitigating the interference and misleading effects of non-stationary variables on process monitoring. Therefore, the operating condition monitoring method for petrochemical processes provided by this invention can effectively address the problems of non-Gaussianity, non-stationarity, and time-series dependencies in catalytic cracking process data, overcoming the limitations of current traditional Gaussian modeling methods such as insufficient modeling capabilities and incomplete feature capture. It exhibits excellent false alarm control capabilities and accurate fault identification capabilities.

[0171] The following are some non-limiting embodiments and comparative examples, which illustrate the feasibility and effectiveness of the proposed method for monitoring operating conditions in petrochemical industrial processes.

[0172] Please refer to Figure 5 , Figure 5 A schematic diagram of the catalytic cracking unit is shown.

[0173] like Figure 5 As shown, the catalytic cracking process can include a reaction-regeneration process, a fractionation process, and an absorption-stabilization process.

[0174] The enlarged view on the left is a partial view of abnormal operating conditions of flue gas emission of pollutants (FEP) in the catalytic cracking unit. When the regulator PDICA105 is in positive-acting mode, its output is usually large. Therefore, the output value of the temperature regulator TICA101 is generally small. The low selector ORST101 mainly controls the regeneration slide valve SV101 based on the output value of the temperature regulator TICA101 to maintain appropriate temperature control. When the riser temperature decreases, the output of the temperature regulator TICA101 increases, the opening of the regeneration slide valve SV101 increases, the catalyst circulation rate increases, and the temperature recovers. The adjustment of the opening of the regeneration slide valve SV101 has a significant impact on energy consumption (EC). Furthermore, the control logic of the regeneration slide valve SV101 is closely related to temperature control and the flow rate of the C102 regenerator shown in the enlarged diagram on the left. This affects the overall thermal energy utilization efficiency in the catalytic cracking unit and the fuel consumption of the C102 regenerator. An imbalance in the ratio of catalyst to fuel in the C102 regenerator may lead to incomplete combustion or over-combustion, causing the flue gas pollutant emission FEP to deviate from the normal range, thus forming an abnormal operating condition of the flue gas pollutant emission FEP.

[0175] The enlarged view on the right is a partial diagram of abnormal operating conditions in the catalytic cracking unit's energy consumption (EC). Data sampling for the operating condition monitoring system used in petrochemical processes can be performed at 10-minute intervals. When abnormal fluctuations occur in the reaction temperature output by the temperature controller TICA101, the excessive opening of the regeneration slide valve SV101 leads to a decrease in the valve differential pressure, causing a rapid reduction in the output of the controller PDICA105, thus affecting energy consumption (EC). In this situation, timely adjustments to the regeneration slide valve SV101 can restore the catalytic cracking unit to normal operating conditions.

[0176] Operating condition monitoring methods for petrochemical processes can monitor energy consumption (EC), flue gas emissions (FEP), techno-economic indicators, and gasoline production in catalytic cracking processes. Based on the monitored targets, key variables in the acquired raw data of the catalytic cracking unit are screened using expert experience and XGBOOST ensemble learning methods to construct a sample set.

[0177] In this embodiment, the ADF test method is used to identify unit roots for each key variable. If a unit root exists, it indicates that the key variable has a random trend and may be a non-stationary variable, thereby distinguishing between stationary and non-stationary variables among the key variables.

[0178] When monitoring the energy consumption (EC) in the catalytic cracking process, the operating condition monitoring system can screen key variables in the raw data of the catalytic cracking unit based on expert experience and XGBOOST ensemble learning method, and distinguish between stationary and non-stationary variables among the key variables based on the ADF test method. The test results of the energy consumption EC modeling variables obtained by the ADF test method are shown in Table 1.

[0179]

[0180] Table 1

[0181] When monitoring the flue gas pollutant emissions (FEP) in the catalytic cracking process, the operating condition monitoring system can screen key variables in the raw data of the catalytic cracking unit based on expert experience and XGBOOST ensemble learning, and distinguish between stationary and non-stationary variables among the key variables based on the ADF test method. The test results of the flue gas pollutant emissions (FEP) modeling variables obtained by the ADF test method are shown in Table 2.

[0182]

[0183] Table 2

[0184] When monitoring the techno-economic indicators of the catalytic cracking process, the operating condition monitoring system can screen key variables in the original data of the catalytic cracking unit based on expert experience and XGBOOST ensemble learning method, and distinguish between stationary and non-stationary variables among the key variables based on the ADF test method. The test results of the techno-economic indicator modeling variables obtained by the ADF test method are shown in Table 3.

[0185]

[0186] Table 3

[0187] When monitoring gasoline production in the catalytic cracking process, the operating condition monitoring system can screen key variables in the raw data of the catalytic cracking unit based on expert experience and XGBOOST ensemble learning, and distinguish between stationary and non-stationary variables among the key variables based on the ADF test method. The test results of the gasoline production modeling variables obtained by the ADF test method are shown in Table 4.

[0188]

[0189] Table 4

[0190] Subsequently, the operating condition monitoring method for petrochemical industrial processes provided by this invention was used to monitor abnormal operating conditions of the above-mentioned catalytic cracking process.

[0191] In the operational condition monitoring method for petrochemical processes provided by this invention, the monitoring model GRU-AAE-CA can be configured as follows: the gated recurrent unit (GRU) adopts a single-layer structure with 64 hidden layer neurons and tanh activation function. The temporal features extracted by the GRU serve as the input to the adversarial generative autoencoder (AAE). The AAE encoder contains two fully connected layers with 64 and 16 neurons respectively. The decoder is symmetrical to the encoder structure, containing two fully connected layers with 16 and 64 neurons respectively and ReLU activation function. Another encoder in the AAE used to reconstruct steady-state features is a two-layer fully connected structure with 33 and 8 neurons respectively, and sigmoid activation function for each layer. The corresponding decoder structure is the opposite, containing two fully connected layers with 8 and 33 neurons respectively. The AAE discriminator is a two-layer fully connected structure with 16 and 1 neurons respectively, and the output layer uses the sigmoid function. Other hyperparameters of AAE are set as follows: batch_size is 512, gen_lr is 0.001, reg_lr is 0.00005, and seed is 1024. During training, the reconstruction loss of the original latent variable space and the steady-state feature space is comprehensively considered based on the total loss function, and an adversarial loss term with adversarial regularization is introduced to improve the fitting ability of the latent variable distribution.

[0192] Table 5 shows the fault monitoring performance and false alarm control effect of various neural network models in the above-mentioned catalytic cracking process operating condition monitoring scenarios.

[0193]

[0194] Table 5

[0195] As can be seen from Table 5, the monitoring model (GRU-AAE-CA) in the operating condition monitoring method provided by this invention has the lowest false alarm rate (FAR) (latent variable space statistic H) during the monitoring of energy consumption EC. 2 The squared prediction error (SPE) of the original space is 0.00%, while the fault detection rate (FDR) is the most prominent in the monitoring process of flue gas pollutant emissions (FEP) (latent variable spatial statistics H). 2 With a accuracy of 93.33% and a squared prediction error (SPE) of 86.67% in the original space, it can be seen that the monitoring model (GRU-AAE-CA) in the working condition monitoring method provided by this invention has the best performance among all neural network models.

[0196] Table 6 shows the monitoring results of the multivariate statistical method under the above-mentioned operating condition monitoring scenarios for catalytic cracking processes.

[0197]

[0198] Table 6

[0199] Table 6 shows that traditional multivariate statistical methods generally perform poorly. The PCA (Principal Component Analysis) method has a high false alarm rate in monitoring energy consumption (EC), and the monitoring index T... 2The false alarm rate was as high as 62.38%, while the false alarm rate for another monitoring indicator, SPE, was as high as 49.53%. Another ICA (Independent Component Analysis) method, while showing a high SPE (83.33%) in monitoring flue gas pollutant emissions (FEP), had a false alarm rate exceeding 65% in monitoring energy consumption (EC), significantly reducing the model's practicality. Among slow feature analysis methods, SFA (Slow Feature Analysis), ESFA (Enhanced Slow Feature Analysis), and IESFA (Improved Enhanced Slow Feature Analysis) showed improvements in some monitoring indicators, such as IESFA. In the process of monitoring flue gas pollutant emissions (FEP), the fault detection rate reached 70.70%. However, in the process of monitoring energy consumption (EC), the monitoring indicator used to reflect the false alarm rate was still in a high false alarm state (above 65%), and could not reliably and effectively identify abnormal operating conditions.

[0200] Please refer to Figure 6 , Figure 6 A comparative graph is shown showing the monitoring of technical and economic indicators and gasoline production using various models provided according to some embodiments of the present invention.

[0201] like Figure 6 As shown in (a), when monitoring technical and economic indicators, the monitoring model GRU-AAE-CA provided by this invention has the lowest false alarm rate (FAR), as follows: Figure 6 As shown in (b), when monitoring gasoline production, the fault detection rate (FDR) of the monitoring model GRU-AAE-CA is the most outstanding, with a higher detection rate.

[0202] Table 7 shows the anomaly detection time for each model for abnormal operating conditions.

[0203]

[0204] Table 7

[0205] Based on the anomaly detection times of each model in Table 7 for abnormal operating conditions, the GRU-AAE-CA monitoring model in the operating condition monitoring method provided by this invention can detect anomalies the fastest when monitoring energy consumption (EC) and flue gas pollutant emissions (FEP), responding promptly at the 102nd and 663rd sampling points respectively, outperforming other comparative models in Table 7. In contrast, while the ICA method also responds quickly (103rd) when monitoring energy consumption (EC), it lags significantly (666th) when monitoring flue gas pollutant emissions (FEP); the PCA method and SFA-type methods are significantly slower, especially the SFA method, which has a significantly delayed response time (678th) when monitoring flue gas pollutant emissions (FEP), failing to effectively meet the needs of real-time fault monitoring.

[0206] In another embodiment, the operating condition monitoring method for petrochemical industrial processes provided by the present invention can be used to monitor abnormal operating conditions of the Tennessee Eastman (TE) process.

[0207] Please refer to Figure 7 , Figure 7 A schematic diagram of the Tennessee Eastman process is shown.

[0208] The Tennessee Eastman (TE) process is a complex, dynamic process widely used to simulate real chemical production processes to support research in the field of process monitoring. The control system of the Tennessee Eastman process mainly consists of key equipment such as reactors, product condensers, gas-liquid separators, circulating compressors, and stripping towers, comprehensively reflecting common physical changes and chemical reactions in industrial production. Figure 7 The diagram shows the products G and H generated after the feeds A and C react with reactants D and E, respectively. The control system of this Tennessee Eastman process mainly consists of five process units: a two-phase reactor where the exothermic reaction occurs, a separator, a stripping tower, a compressor, and a condenser. In the diagram, FI is the flow indicator, PI is the pressure indicator, LI is the level indicator, and TI is the temperature indicator.

[0209] In the experimental dataset, each fault scenario for the Tennessee Eastman process contains 960 sampling data points. In each fault scenario, fault events are introduced starting at the 160th sampling point to simulate sudden abnormal conditions that may occur during industrial production. By analyzing the data from fault scenarios 1 and 13, the operating condition monitoring system can systematically evaluate the fault detection performance of the operating condition monitoring method provided by this invention under abnormal process conditions of different natures, further verifying the practical application value and robustness of this method. Here, the configuration of the monitoring model GRU-AAE-CA used in the operating condition monitoring system can be the same as in the above-described embodiment for abnormal operating condition monitoring of the catalytic cracking process.

[0210] Table 8 shows the Fault Detection Rate (FDR) results for each model in the operating condition monitoring scenario of the Tennessee Eastman process.

[0211]

[0212] Table 8

[0213] As shown in Table 8, the monitoring model GRU-AEE-CA in the operating condition monitoring method provided by this invention has a high detection rate in fault monitoring and can effectively identify different types of process faults.

[0214] In summary, the monitoring model GRU-AEE-CA in the working condition monitoring method provided by this invention not only significantly outperforms other neural network models and existing statistical methods in terms of false alarm control capability and fault identification accuracy, but also demonstrates a clear advantage in real-time response performance. This is mainly due to the strong sensitivity of the GRU structure in the monitoring model to short-term fluctuations in sequence data, which, combined with AAE, effectively captures the deep features and potential distribution characteristics of the data. Furthermore, CA accurately describes the long-term equilibrium relationships and collaborative change trends between variables, giving the model excellent generalization and stability. It is highly suitable for complex, multi-variable dynamic working condition monitoring scenarios and possesses high engineering application potential and promotional value.

[0215] Although the methods described above are illustrated and depicted as a series of actions for the sake of simplicity, it should be understood and appreciated that these methods are not limited by the order of the actions, as some actions may occur in a different order and / or concurrently with other actions from the illustrations and descriptions herein or not illustrated and described herein but which may be understood by those skilled in the art, according to one or more embodiments.

[0216] Those skilled in the art will understand that information, signals, and data can be represented using any of a variety of different techniques and arts. For example, the data, instructions, commands, information, signals, bits, symbols, and chips described throughout the above description can be represented by voltage, current, electromagnetic waves, magnetic fields or magnetic particles, light fields or optical particles, or any combination thereof.

[0217] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps are described above in a generalized manner in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art may implement the described functionality in different ways for each specific application, but such implementation decisions should not be construed as departing from the scope of the invention.

[0218] The various illustrative logic modules and circuits described in conjunction with the embodiments disclosed herein may be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0219] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor such that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.

[0220] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted as one or more instructions or code on or through a computer-readable medium. A computer-readable medium includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium accessible to a computer. By way of example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of a medium. As used in this article, disk and disc include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs. Disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of these should also be included within the scope of computer-readable media.

[0221] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for monitoring operating conditions in petrochemical industrial processes, characterized in that, Including the following steps: Key variables in the acquired catalytic cracking unit data were screened using expert experience and XGBOOST ensemble learning methods to construct a sample set; The sample set is input into the monitoring model to obtain a monitoring statistic sequence. The monitoring model integrates gated cyclic units, adversarial generative autoencoders, and cointegration analysis. as well as The tail distribution of the monitoring statistics sequence is scanned using a generalized Pareto distribution, and the tail features of the monitoring statistics sequence are fitted to determine the dynamic control threshold for identifying abnormal operating conditions.

2. The operating condition monitoring method as described in claim 1, characterized in that, The step of inputting the sample set into the monitoring model to obtain the monitoring statistics sequence includes: The hidden state is determined using the gated loop unit in order to extract the dynamic temporal features of the data in the sample set; The hidden state is input into the adversarial generative autoencoder to capture the structure of dynamic non-Gaussian latent variables during catalytic cracking. The adversarial training of the adversarial generative autoencoder is based on the total loss function of the monitoring model, which includes a reconstruction loss function for steady-state features, determined using the cointegration analysis; and Based on the steady-state characteristics and the latent variables, the monitoring statistics sequence is determined.

3. The operating condition monitoring method as described in claim 2, characterized in that, The total loss function is: L total L recon +λ sta L sta +L adv , Among them, L recon L is the reconstruction loss function of the original data in the sample set. sta It is the reconstruction loss function of steady-state features, L adv It is the adversarial loss term, λ sta These are the weighting coefficients.

4. The operating condition monitoring method as described in claim 2, characterized in that, The latent distribution of the latent variables is: Q φ (z|h t )=P(z), Where z is a latent variable, h t Q is the hidden state output by the gated loop unit at time t. φ P(z) is the conditional probability distribution under encoder parameters φ, and P(z) is the prior probability distribution of the latent variable z.

5. The operating condition monitoring method as described in claim 2, characterized in that, The formula for extracting the steady-state features is: x t =b T x t =β1x 1t +β2x 2t +...+b m x mt , Where, ξ t It is a steady-state characteristic sequence, x t It is a cointegrated sequence, and β is the cointegrated vector.

6. The operating condition monitoring method as described in claim 2, characterized in that, The monitoring statistic sequence includes one or more of the following: latent variable spatial statistics, steady-state feature space offset statistics, squared prediction error of the original space, and steady-state feature reconstruction error. The formula for calculating the spatial statistics of the latent variables is: Among them, H 2 It is the spatial statistic of latent variables, and W is the weight matrix; The formula for calculating the offset statistics of the steady-state feature space is: in, It is the steady-state characteristic space offset statistic, ζ is the variable vector in the steady-state characteristic space, and μ is the constant. ζ It is the mean vector of the variable vectors; The formula for calculating the squared prediction error of the original space is: SPE=||x-x′|| 2 , Where SPE is the squared prediction error of the original space, x is the sample data, and x′ is the predicted data; and The formula for calculating the reconstruction error of steady-state characteristics is: SPE ζ =||ζ i -ζ′ i || 2 , Among them, SPE ζ It is the reconstruction error of the steady-state characteristics, ζ i ζ′ is the i-th actual feature variable in the steady-state feature space. i It is the i-th feature variable in the reconstructed steady-state feature space.

7. The operating condition monitoring method as described in claim 1, characterized in that, The steps for constructing the sample set include: The data in the sample set are subjected to max-min normalization to eliminate the influence of dimensions and improve the convergence of the monitoring model.

8. The operating condition monitoring method as described in claim 1, characterized in that, The formula for calculating the dynamic control threshold is: Among them, y α The threshold is dynamically controlled, θ is the initial threshold, σ is the scale parameter describing the scale characteristics of the tail distribution (σ>0), k is the shape parameter characterizing the tail distribution properties (when k>0, the tail distribution exhibits heavy tails), α is the given significance level, and N... θ It is the set of data points that exceed the initial threshold.

9. A condition monitoring system for petrochemical industrial processes, characterized in that, include: Memory, on which computer instructions are stored; as well as A processor, connected to the memory, and configured to execute computer instructions stored in the memory to implement the operating condition monitoring method for petrochemical industrial processes as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed by the processor, the operating condition monitoring method for petrochemical industrial processes as described in any one of claims 1 to 8 is implemented.

Images