Chemical process risk identification method based on double-branch time series prediction model
By using a bi-branch time-series prediction model, the long-term time-series characteristics of single variables and the delayed coupling characteristics between variables in chemical processes are modeled separately. This solves the problem of incomplete modeling of coupling between variables in chemical processes and enables accurate identification of potential risks and safety monitoring of chemical production systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2026-05-19
- Publication Date
- 2026-07-14
AI Technical Summary
Existing methods for identifying risks in chemical processes have imperfect coupling modeling between variables, resulting in low accuracy in risk identification and difficulty in accurately identifying potential risks in complex chemical production processes.
A dual-branch time series prediction model is adopted, which models the long-term time series features within a single variable and the delay coupling features between variables through time branch and variable branch respectively. A high-order feature representation is generated through an adaptive gating fusion mechanism, and risk identification is performed by combining residual connection and classification output head.
It significantly improves the accuracy of risk identification in chemical processes, enabling precise identification of potential operational risks caused by complex variable coupling and nonlinearity, thereby enhancing the safety and reliability of chemical production systems.
Smart Images
Figure CN122390475A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of chemical process monitoring and risk identification technology, and in particular to a method for chemical process risk identification based on a dual-branch time series prediction model. Background Technology
[0002] Chemical processes are the core of modern chemical industrial production, typically operating under high temperature and pressure conditions, and exhibiting significant characteristics such as strong coupling between variables, dynamic characteristics across multiple time scales, and highly nonlinear behavior. Risks in chemical processes often exhibit a gradual nature, with early warning signs being relatively weak. If these risks are not identified and detected in a timely manner, they can easily spread throughout the entire production system, leading to production interruptions or even serious safety accidents. Therefore, accurate and timely risk identification of chemical processes is a fundamental prerequisite for ensuring safe production operations and improving production reliability.
[0003] Chemical processes are complex dynamic processes with coupled process parameters. In actual industrial operation, multidimensional process features extracted from on-site monitoring data (flow rate, pressure, temperature, liquid level) constitute typical multivariate heterogeneous time series. The dynamic behavior of this type of time series data exhibits both slow, gradual trends (such as operating condition drift, equipment aging, and parameter baseline shift) and instantaneous abrupt disturbances (such as operating condition fluctuations, sudden parameter changes, and equipment failures). Therefore, accurately and synchronously capturing these two types of dynamic features within a limited observation window remains a core challenge in chemical process risk modeling and identification. Existing chemical process risk identification methods are mainly divided into two categories: Multivariate Statistical Process Control (MSPC) and deep learning. MSPC-based models have strong physical interpretability, low computational cost, and are easy to implement in engineering, but they generally struggle to characterize the nonlinear coupling relationships between variables and the dynamic evolution laws over long time series, and their adaptability to complex time-varying operating conditions is limited. On the other hand, deep learning methods based on recurrent neural networks (RNNs) are prone to gradient vanishing defects when processing time series chemical monitoring data, which is not conducive to training stable and reliable models. While deep learning methods based on the Transformer architecture can effectively model long-distance dependencies, the number of parameters and computational complexity of this architecture increase quadratically with the sequence length. When applied to long-term chemical engineering scenarios, the computational overhead is significantly high, making it difficult to meet the high real-time requirements of industrial sites.
[0004] State-space models (SSMs), with Mamba as a typical example, have opened up a new research path for time series modeling in recent years. Mamba, relying on its innovative selective state-space scanning mechanism, maintains its ability to model long-distance temporal dependencies, while its parameter count and computational complexity increase linearly with the sequence length, resulting in computational efficiency far superior to the traditional Transformer architecture. However, the native Mamba architecture is primarily designed for univariate time series and lacks the ability to jointly model the internal correlations and delayed coupling characteristics between variables in multivariate time series data. It cannot fully exploit the complex dynamic correlations between multi-sensor variables in chemical processes, making it difficult to directly apply to risk identification tasks in multivariate coupled scenarios in chemical processes. Summary of the Invention
[0005] This invention provides a chemical process risk identification method based on a dual-branch time series prediction model, which solves the problem of low risk identification accuracy caused by imperfect modeling of inter-variable coupling in existing chemical process risk identification methods. This method can achieve accurate identification of potential risks in complex chemical production processes.
[0006] A first aspect of this invention provides a method for identifying risks in chemical processes based on a dual-branch time-series prediction model, comprising the following steps: Step 1: Obtain online operating data of the chemical production system, preprocess and standardize the online operating data to obtain a multivariate time series; Step 2: Input the multivariate time series into the pre-constructed dual-branch time series prediction model, and use an adaptive gating fusion mechanism to fuse the time series features output by the time branch with the delay coupling features output by the variable branch to generate a high-order feature representation; wherein, the time branch performs time series modeling along the time series dimension to extract long-range time series features within a single variable, and the variable branch reorganizes the input data to construct a time series prediction model based on the variable adaptive state space model and extracts the delay coupling features between different process variables; Step 3: Perform residual fusion on the fused high-order features to output the global features of the model. Connect the global features to the classification output head to complete the identification of the risk status of the chemical process.
[0007] In one embodiment of the present invention, the method further includes: Step 4: Construct a classification loss function and update the parameters of the two-branch time series prediction model through backpropagation to obtain a two-branch time series prediction model adapted to industrial time series scenarios.
[0008] In one embodiment of the present invention, step 1 specifically includes: Acquire online operating data of the chemical production system, including temperature, pressure, flow rate, and equipment operating status; Identify and remove invalid and missing data from the online running data, and perform z-score standardization on the remaining valid data.
[0009] In one embodiment of the present invention, step 2 specifically includes: The multivariate time series is fed into the embedding layer, and the feature dimension mapping is completed through linear projection, as shown in the following formula: in, For standardized online operation data The first in One sample, Indicates the number of variables being measured. This represents the number of time steps for each sample. This represents the learnable weight matrix. This represents a learnable bias term. For activation function, For embedding features, For embedding feature dimensions; In the time branch, input-dependent dynamic parameters are first generated through an independent projection layer, and the discretized parameter matrix is then calculated. Input matrix Output matrix Among them, the discretization parameter matrix through The activation function constraint is positive: in, For activation functions that produce positive results, The weight matrix is a learnable matrix. These are learnable bias terms; Update the hidden state step by step over time to extract long-term time series features: in, Indicates time step The hidden state at that time This represents the natural exponential function. For a learnable transfer matrix, Indicates time step The hidden state at that time Represents the identity matrix. Representing embedded features The OK, This represents the intermediate features extracted from the time branch; Long-range time-series features are obtained through feature gating and linear projection. The formula is as follows: in, The weight matrix is a learnable matrix. For learnable bias terms; In the variable branch, first check the input. Perform dimensional reconstruction, that is: in, This represents the input after the dimensional reconstruction. Based on embedding features Input matrix Output matrix The formula yields the embedding features of the variable branches. Input matrix Output matrix ; Independent discretization parameters are generated for each measured variable, as shown in the following formula: in, Indicates the first Discretization parameters of the measured variables, Embedding features representing variable branches The OK, and They respectively represent the corresponding number Learnable weight matrix and bias terms for each discretized parameter; The hidden state is updated for each variable individually, and the delayed coupling features between variables are extracted. The formula is as follows: in, Indicates the first The hidden state of each measured variable The discretized parameter matrix for the variable branch. For a learnable transfer matrix, Indicates the first The hidden state of each measured variable The input matrix for the variable branch, Embedding features representing variable branches The OK, This represents the intermediate features extracted from the variable branch. The output matrix for the variable branch; By using feature gating and linear projection, the delay coupling features between variables are obtained. The formula is as follows: in, The weight matrix is a learnable matrix. For learnable bias terms; Long-range time series characteristics of time branches Delayed coupling between variables in variable branches Concatenate along the feature dimension, that is: in, Indicates the features after splicing; Generate learnable gating weights : in, The range of values is Activation function, The weight matrix is a learnable matrix. Learnable biases; Weighted fusion of bi-branch features is performed based on learnable weights to generate higher-order fused features. The formula is as follows: .
[0010] In one embodiment of the present invention, step 3 specifically includes: A residual connection is constructed, which adds the fused higher-order features to the original embedded features, preserving the original low-level information and preventing gradient vanishing. The formula is as follows: in, For the global features of the two-branch time series prediction model, For the higher-order features after fusion, For embedded features; global features Input risk identification and classification head, and then linearly project and Function activation outputs the predicted probabilities of various risk states in a chemical process. : in, The weight matrix is a learnable matrix. This is a learnable weight matrix.
[0011] The chemical process risk identification method based on a dual-branch time-series prediction model in this invention can improve the accuracy of chemical process risk identification, accurately identify potential operational risks caused by complex variable coupling and nonlinearity, and provide technical support for the safe operation and intelligent monitoring of chemical production systems.
[0012] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0013] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart of a chemical process risk identification method based on a dual-branch time-series prediction model provided by the present invention; Figure 2 This invention provides an architecture diagram for a chemical process risk identification method based on a dual-branch time-series prediction model. Detailed Implementation
[0014] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0015] Figure 1 The flowchart of a chemical process risk identification method based on a dual-branch time series prediction model provided by the present invention is shown.
[0016] The chemical process risk identification method based on a two-branch time-series prediction model includes the following steps: Step 1: Obtain online operating data of the chemical production system, preprocess and standardize the online operating data to obtain a multivariate time series.
[0017] Step 1 specifically includes: Obtain online operating data from chemical production systems ,in, Represents the total sample size. This indicates the number of variables being measured. Variables being measured are those in the online dataset; a sample consists of time series data of multiple variables. This indicates the time step for each sample. Online operational data includes temperature, pressure, flow rate, and equipment operating status.
[0018] Identify and remove invalid and missing data from the online dataset to ensure the integrity and validity of the input data. Perform z-score standardization on the remaining valid data to eliminate interference from differences in units and scales between different types of variables. The z-score standardization formula is: in, For online operation data The Middle The first sample A data sequence of the measured variable, For online operation data The standardized value of the data sequence of the m-th measured variable in the n-th sample. Let be the mean of the m-th measured variable sequence. Let be the standard deviation of the m-th measured variable sequence.
[0019] Preprocessing and standardizing online operational data from chemical production systems can suppress the effects of sensor noise, measurement scale differences, and outliers, thereby ensuring numerical consistency of input data across all variables. This prevents the model from learning spurious patterns due to differences in data measurement scales, thus improving model stability.
[0020] Step 2: Input the multivariate time series into the pre-built dual-branch time series prediction model, and use an adaptive gating fusion mechanism to fuse the time series features output by the time branch with the delay coupling features output by the variable branch to generate a high-order feature representation. Specifically, the time branch performs time series modeling along the time series dimension to extract long-range time series features within a single variable, while the variable branch reorganizes the input data to construct a time series prediction model based on the variable adaptive SSM Mamba encoder and extracts the delay coupling features between different process variables.
[0021] In an embodiment of the present invention, step 2 specifically includes: The multivariate time series is fed into the embedding layer, and the feature dimension mapping is completed through linear projection, as shown in the following formula: in, For standardized online operation data The first in One sample, Indicates the number of variables being measured. This represents the number of time steps for each sample. This represents the learnable weight matrix. This represents a learnable bias term. For activation function, For embedding features, For embedding feature dimensions; In the time branch, input-dependent dynamic parameters are first generated through an independent projection layer, and the discretized parameter matrix is then calculated. Input matrix Output matrix Among them, the discretization parameter matrix through The activation function constraint is positive: in, For activation functions that produce positive results, The weight matrix is a learnable matrix. These are learnable bias terms; Update the hidden state step by step over time to extract long-term time series features: in, Indicates time step The hidden state at that time This represents the natural exponential function. For a learnable transfer matrix, Indicates time step The hidden state at that time Represents the identity matrix. Representing embedded features The OK, This represents the intermediate features extracted from the time branch; Long-range time-series features are obtained through feature gating and linear projection. The formula is as follows: in, The weight matrix is a learnable matrix. For learnable bias terms; In the variable branch, first check the input. Perform dimensional reconstruction, that is: in, This represents the input after the dimensional reconstruction. Based on embedding features Input matrix Output matrix The formula yields the embedding features of the variable branches. Input matrix Output matrix ; Independent discretization parameters are generated for each measured variable, as shown in the following formula: in, Indicates the first Discretization parameters of the measured variables, Embedding features representing variable branches The OK, and They respectively represent the corresponding number Learnable weight matrix and bias terms for each discretized parameter; The hidden state is updated for each variable individually, and the delayed coupling features between variables are extracted. The formula is as follows: in, Indicates the first The hidden state of each measured variable The discretized parameter matrix for the variable branch. For a learnable transfer matrix, Indicates the first The hidden state of each measured variable The input matrix for the variable branch, Embedding features representing variable branches The OK, This represents the intermediate features extracted from the variable branch. The output matrix for the variable branch; By using feature gating and linear projection, the delay coupling features between variables are obtained. The formula is as follows: in, The weight matrix is a learnable matrix. For learnable bias terms; Long-range time series characteristics of time branches Delayed coupling between variables in variable branches Concatenate along the feature dimension, that is: in, Indicates the features after splicing; Generate learnable gating weights : in, The range of values is Activation function, The weight matrix is a learnable matrix. For learnable bias terms; Weighted fusion of bi-branch features is performed based on learnable weights to generate higher-order fused features. The formula is as follows: .
[0022] By structurally decoupling univariate time-series feature modeling from multivariate delay-coupled feature modeling, a dual-parallel modeling branch architecture is designed. Compared to single-path models, this architecture can extract more discriminative deep representational features from multivariate sensor time-series data, significantly improving the accuracy and robustness of chemical process risk state identification. Simultaneously, variable-specific learnable discretization parameters are introduced. This enables the model to adaptively capture the time-delay correlation characteristics between different process variables, enhances the ability to extract features of the risk propagation and evolution laws of industrial systems, and effectively improves the accuracy of risk identification with variable delay coupling.
[0023] Step 3: Perform residual fusion on the fused high-order features to output the global features of the model. Connect the global features to the classification output head to complete the identification of the risk status of the chemical process.
[0024] In an embodiment of the present invention, step 3 specifically includes: A residual connection is constructed, which adds the fused higher-order features to the original embedded features, preserving the original low-level information and preventing gradient vanishing. The formula is as follows: in, For the global features of the two-branch time series prediction model, For the higher-order features after fusion, For embedded features; global features Input risk identification and classification head, and then linearly project and Function activation outputs the predicted probabilities of various risk states in a chemical process. : in, The weight matrix is a learnable matrix. This is a learnable bias term.
[0025] Step 4: Construct a classification loss function and update the parameters of the two-branch time series prediction model through backpropagation to obtain a two-branch time series prediction model adapted to industrial time series scenarios.
[0026] The embodiments of the present invention use cross-entropy as the loss function and update all learnable parameters of the model through the backpropagation algorithm.
[0027] This invention effectively preserves shallow original feature information by introducing a residual connection mechanism, alleviates the gradient vanishing problem in deep networks, stabilizes feature distribution, and improves the integrity and stability of global representation. It also achieves accurate mapping from features to risk categories by relying on a dedicated classification output head, thus meeting the needs of chemical process risk identification tasks.
[0028] like Figure 2 As shown, this invention includes an offline process and an online process. The offline process uses offline data to train a constructed bi-branch time-series prediction model. The trained bi-branch time-series prediction model is then applied to the online process for chemical process risk identification. During the online identification process, the parameters of the bi-branch time-series prediction model are updated through backpropagation, continuously optimizing the model.
[0029] This invention provides a method for identifying risks in chemical processes based on a dual-branch time-series prediction model. The dual-branch time-series prediction model is built based on the Mamba encoder. This method constructs parallel modeling branches for time and variables, combines adaptive learnable discretization parameters for variables with an adaptive gating fusion mechanism, and relies on the sequence modeling capabilities of the Mamba model to achieve deep risk feature mining of multivariate time-series monitoring data in chemical processes. Thus, while retaining the linear complexity of the Mamba model, it improves the accuracy of the model in identifying process risks in long-term time-series chemical data.
[0030] The technical solution proposed in this invention effectively enhances the risk identification and modeling capabilities in chemical engineering scenarios characterized by long-term time-series dependence and multivariate delayed coupling. The dual-branch separation modeling architecture avoids the shortcomings of traditional single-time-series modeling networks, which cannot simultaneously characterize the internal temporal evolution of variables and cross-variable time-delay correlations. It can extract temporal features and variable coupling features separately. The introduction of variable-specific learnable discretization parameters enables the model to adaptively match the dynamic response patterns of different variables, accurately capturing delayed coupling and risk propagation behavior between variables, thus improving the modeling capability for strongly coupled time-series variables such as temperature, pressure, and flow rate. This significantly improves the accuracy of risk identification.
[0031] This embodiment uses the production process of base asphalt as the verification object to demonstrate the feasibility and accuracy of the proposed dual-branch Mamba-based chemical process risk identification method in complex, multi-variable coupled industrial scenarios. Base asphalt preparation mainly involves processes such as electro-desalination, initial distillation, atmospheric distillation, vacuum distillation, and oxidative blending, each completed by a corresponding equipment unit and heating furnace unit. The material and energy flows between these processes are highly correlated, and the process variables exhibit strong nonlinearity, long-term time dependence, and time-delay coupling characteristics, placing high demands on the feature mining and modeling capabilities of the risk identification method. This embodiment uses industrial measured time-series data collected by distributed sensors in the base asphalt production line, containing 263 process monitoring variables covering key chemical process parameters such as temperature, pressure, and flow rate. After data cleaning, denoising, and standardization preprocessing, a total of 12,362 valid samples were obtained, with a single sample time series length of 3,001. The dataset comprehensively covers normal operating conditions and various typical chemical process risks.
[0032] To objectively evaluate the risk identification performance of the method of this invention, this embodiment selects the traditional MSPC method Dynamic Principal Component Analysis (DPCA), the classic deep learning model Transformer, and the original single-branch Mamba model as benchmarks. The risk identification effect of this invention is verified on the industrial test dataset using the above comparison methods. The effectiveness of the methods is compared using three widely used metrics: accuracy, recall, and F1 score. The experimental comparison results are shown in the table below: Table 1 Comparison of Experimental Results The experimental results show that the traditional MSPC method DPCA performs poorly in this scenario, with an accuracy of only 73.43%, a recall of 72.26%, and an F1 score of 73.93%. This indicates that traditional statistical methods are difficult to adapt to the characteristics of multivariate nonlinear coupling and long-term time dependence in complex chemical processes, resulting in insufficient risk identification capabilities.
[0033] The Transformer model offers a significant performance improvement over DPCA, achieving an accuracy of 93.26%. However, it is limited by the lack of inductive bias for time-series data, which restricts its ability to characterize the continuous operating conditions of chemical systems and capture the delayed coupling characteristics between variables, thus its performance is not optimal.
[0034] The single-branch Mamba model, with its powerful state-space modeling capabilities, further improves performance compared to the Transformer, achieving an accuracy of 94.88%, thus verifying the adaptability of the Mamba architecture in long-term chemical engineering data modeling. However, due to the use of only single-path modeling and the lack of targeted design of variable coupling modeling branches, the mining of delayed interaction features between multiple variables is still insufficient, resulting in a performance bottleneck.
[0035] The proposed two-branch Mamba method achieves optimal results across three core metrics: accuracy of 97.54%, recall of 97.71%, and F1 score of 97.17%. These results represent improvements over the single-branch Mamba method.
[0036] The above results fully verify the technical advantages of the dual-branch architecture of this invention: the time branch accurately captures the long-term time-series evolution characteristics within process variables, the variable branch specifically models the delay coupling characteristics between multiple variables, and the adaptive gating fusion mechanism realizes the organic combination of multi-dimensional features. This effectively solves the problems of insufficient extraction of complex coupling features and low risk identification accuracy of traditional methods, and demonstrates stronger risk identification capabilities under complex chemical working conditions, which can better meet the actual needs of high-reliability risk monitoring in industrial sites.
[0037] This invention proposes a chemical process risk identification method based on a dual-branch time-series prediction model. Using a Mamba network as the temporal modeling architecture, it constructs two parallel modeling branches to represent and learn multivariate sensor data from both temporal and variable dimensions. The temporal branch models complex temporal dependencies within a single variable, while the variable branch characterizes delayed coupling relationships between different process variables. The output features of the two branches are dynamically fused via an adaptive gating fusion mechanism, and the resulting unified representation is input into a classifier to identify and determine the risk status of the chemical process. The proposed method can simultaneously model the temporal evolution features within variables and the delayed coupling relationships between multiple variables, effectively improving the accuracy of risk identification under complex chemical operating conditions. Simultaneously, the model maintains linear computational complexity, making it suitable for real-time online monitoring applications in actual industrial settings.
[0038] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0039] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.
Claims
1. A method for identifying risks in chemical processes based on a dual-branch time-series prediction model, characterized in that, Includes the following steps: Step 1: Obtain online operating data of the chemical production system, preprocess and standardize the online operating data to obtain a multivariate time series; Step 2: Input the multivariate time series into the pre-constructed dual-branch time series prediction model, and use an adaptive gating fusion mechanism to fuse the time series features output by the time branch with the delay coupling features output by the variable branch to generate a high-order feature representation; wherein, the time branch performs time series modeling along the time series dimension to extract long-range time series features within a single variable, and the variable branch reorganizes the input data to construct a time series prediction model based on the variable adaptive state space model and extracts the delay coupling features between different process variables; Step 3: Perform residual fusion on the fused high-order features to output the global features of the model. Connect the global features to the classification output head to complete the identification of the risk status of the chemical process.
2. The method according to claim 1, characterized in that, The method further includes: Step 4: Construct a classification loss function and update the parameters of the two-branch time series prediction model through backpropagation to obtain a two-branch time series prediction model adapted to industrial time series scenarios.
3. The method according to claim 1, characterized in that, Step 1 specifically includes: Acquire online operating data of the chemical production system, including temperature, pressure, flow rate, and equipment operating status; Identify and remove invalid and missing data from the online running data, and perform z-score standardization on the remaining valid data.
4. The method according to claim 1, characterized in that, Step 2 specifically includes: The multivariate time series is fed into the embedding layer, and the feature dimension mapping is completed through linear projection, as shown in the following formula: in, For standardized online operation data The first in One sample, Indicates the number of variables being measured. This represents the number of time steps for each sample. This represents the learnable weight matrix of the embedding layer. This represents the learnable bias term of the embedding layer. For activation function, For embedding features, For embedding feature dimensions; In the time branch, input-dependent dynamic parameters are first generated through an independent projection layer, and the discretized parameter matrix is then calculated. Input matrix Output matrix Among them, the discretization parameter matrix through The activation function constraint is positive: in, For activation functions that produce positive results, The weight matrix is a learnable matrix. For learnable bias terms; Update the hidden state step by step over time to extract long-term time series features: in, Indicates time step The hidden state at that time This represents the natural exponential function. For a learnable transfer matrix, Indicates time step The hidden state at that time Represents the identity matrix. Representing embedded features The OK, This represents the intermediate features extracted from the time branch; Long-range time-series features are obtained through feature gating and linear projection. The formula is as follows: in, The weight matrix is a learnable matrix. For learnable bias terms; In the variable branch, first check the input. Perform dimensional reconstruction, that is: in, This represents the input after the dimensional reconstruction. Based on embedding features Input matrix Output matrix The formula yields the embedding features of the variable branches. Input matrix Output matrix ; Independent discretization parameters are generated for each measured variable, as shown in the following formula: in, Indicates the first Discretization parameters of the measured variables, Embedding features representing variable branches The OK, and They respectively represent the corresponding number Learnable weight matrix and bias terms for each discretized parameter; The hidden state is updated for each variable individually, and the delayed coupling features between variables are extracted. The formula is as follows: in, Indicates the first The hidden state of each measured variable The discretized parameter matrix for the variable branch. For a learnable transfer matrix, Indicates the first The hidden state of each measured variable The input matrix for the variable branch, Embedding features representing variable branches The OK, This represents the intermediate features extracted from the variable branch. The output matrix for the variable branch; By using feature gating and linear projection, the delay coupling features between variables are obtained. The formula is as follows: in, The weight matrix is a learnable matrix. For learnable bias terms; Long-range time series characteristics of time branches Delayed coupling between variables in variable branches Concatenate along the feature dimension, that is: in, Indicates the features after splicing; Generate learnable gating weights : in, The range of values is Activation function, The weight matrix is a learnable matrix. For learnable bias terms; Weighted fusion of bi-branch features is performed based on learnable weights to generate higher-order fused features. The formula is as follows: 。 5. The method according to claim 1, characterized in that, Step 3 specifically includes: A residual connection is constructed, which adds the fused higher-order features to the original embedded features, preserving the original low-level information and preventing gradient vanishing. The formula is as follows: in, For the global features of the two-branch time series prediction model, For the higher-order features after fusion, For embedded features; global features Input risk identification and classification head, and then linearly project and Function activation outputs the predicted probabilities of various risk states in a chemical process. : in, The weight matrix is a learnable matrix. This is a learnable bias term.