A Multi-Unit Collaborative Safety Control Method for Process Industries Based on Spatiotemporal Variational Bayes

Abstract

A spatiotemporal variational Bayesian approach for multi-unit collaborative safety control in process industries is proposed. This approach involves collecting multivariate process data from multiple physical sensors and using mutual information clustering to group closely related components into sub-units. In the time dimension, SFA is used to monitor the rate of process change, and LSTM is used to mine temporal dependencies and deep temporal features. In the spatial dimension, GCN is used to model sub-units as graph nodes, mining anomaly propagation paths and spatial interaction features. A multi-dimensional analysis framework is constructed by integrating spatiotemporal features. A DBN model is built for each sub-unit, and variational inference is used to overcome the bottleneck of marginal likelihood calculation, constructing anomaly contribution scores to locate anomaly sources. An online anomaly data inference control scheme is input, validated by the industrial controller, and iterated with new data until the anomaly is eliminated. This method effectively addresses core technical challenges in process industries such as strong coupling among multi-unit components, complex anomaly propagation paths, and temporal dynamic nonlinearity, ensuring the safe and stable operation of industrial systems.

Description

Technical Field

[0001] This invention belongs to the field of industrial intelligent control technology, specifically relating to a multi-unit collaborative safety control method for process industries based on spatiotemporal variational Bayes. Background Technology

[0002] Modern industrial processes are generally characterized by large scale, dynamic spatiotemporal nature, and strong coupling, with close connections and mutual constraints between production units. An anomaly in a local unit can easily trigger a chain reaction through the coupling relationships between units, creating a butterfly effect that ultimately paralyzes the entire plant's production process, causing significant economic losses. Against this backdrop, the focus of industrial engineers and researchers has gradually shifted from traditional single discrete process control to plant-wide collaborative and safe operation control of industrial processes. Ensuring the safe and stable collaborative operation of multiple units across the entire plant's industrial processes has become a core issue that urgently needs to be addressed in the field of industrial intelligent control.

[0003] In the overall industrial process, operational units do not operate in isolation but rather collaborate to complete production tasks, with complex interdependencies between them. An anomaly in any unit can have a cascading effect on the entire production process, triggering wider systemic failures. Some anomalies can be mitigated locally through compensation or adjustments, while others spread along the production chain to other units as production progresses, severely impacting the overall stability and safety of the industrial process. This anomaly propagation characteristic makes it difficult for traditional single-process-based safety operation control methods to accurately pinpoint the root cause of anomalies and thoroughly repair them. Relying solely on local optimization or independent control of a single unit cannot fundamentally eliminate the impact of anomalies on the entire system, easily leading to fragmentation of plant-wide process control and significantly reducing the comprehensiveness and robustness of control effectiveness. Traditional plant-wide process safety operation control methods often treat each local unit as an independent operational object, neglecting the interactions and spatiotemporal information correlations between units, resulting in an inability to comprehensively grasp the overall operational status and potential risks of the entire plant process.

[0004] As industrial processes become increasingly complex, the spatiotemporal information correlation between local units becomes more critical. However, traditional control methods often ignore or only perform simplistic processing of this spatiotemporal information, further exacerbating the system's vulnerability to complex and abnormal operating conditions. Therefore, effectively integrating multi-unit collaborative control with spatiotemporal information correlation has become the core key to achieving safe operation control of the entire plant's industrial processes. The inherent limitations of traditional methods urgently require the exploration of a new control model to break through the bottleneck of local control, achieve collaborative response of various units to abnormal operating conditions across the entire plant, and thus ensure the stability and safety of the entire industrial system. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a multi-unit collaborative safety control method for process industries based on spatiotemporal variational Bayes. This method effectively addresses core challenges in process industries, such as strong coupling among multi-units, complex anomaly propagation paths, and temporal dynamic nonlinearity, through mutual information clustering, spatiotemporal feature collaborative extraction, GCN spatial propagation learning, DBN variational inference, and closed-loop control iteration. It can effectively ensure the safe and stable operation of industrial systems.

[0006] To achieve the above objectives, the present invention provides a multi-unit collaborative safety control method for process industries based on spatiotemporal variational Bayes, comprising the following steps; Step 1: Process data acquisition and sub-unit division based on physical sensors; Multi-variable data of each process link are collected in real time using multi-source physical sensors, and the dependencies between data are quantified by mutual information, and links with close information correlation are clustered into sub-units; Step 2: Slow Feature Analysis to Monitor the Rate of Change over Time; Introduce SFA to extract the essential features of slow changes from sensor time-series data, and identify rate abrupt changes or gradual change patterns caused by anomalies; Step 3: Use Long Short-Term Memory (LSTM) networks to mine deep time series features; use LSTM to learn the temporal dependencies of historical sensor data and mine deep temporal features. Step 4: Mining features in the spatial dimension using graph convolutional networks; Modeling with graph convolutional networks, sub-units are abstracted as graph nodes, and the material flow and energy flow between units are associated as edges. By aggregating information from neighboring nodes, the propagation path and impact range of anomalies in the spatial dimension are mined, and the spatial interaction features between units are extracted. Step 5: Integrate temporal and spatial features; combine the temporal features extracted by SFA and LSTM with the spatial features mined by GCN to construct a multi-dimensional analysis framework to comprehensively depict the dynamic evolution of industrial processes and the interaction between spatial structures. Step 6: Construct a dynamic Bayesian network model for each sub-unit; for each sub-unit, build a DBN model to describe the evolution of the sub-unit's state over time, as well as its interaction logic with other units, to achieve dynamic modeling at the sub-unit level. Step 7: Use variational inference to solve the dilemma of marginal likelihood calculation; To address the problem of the difficulty in calculating the marginal likelihood of DBN, we use variational inference to approximate the posterior distribution of latent variables, and use the KL divergence between the variational distribution and the true posterior distribution to derive the variational lower bound, thus breaking through the computational bottleneck. Step 8: Construct an anomaly contribution score model to quantify the contribution of anomalies and their degree of impact; build multi-dimensional indicators to quantify the contribution rate of sub-unit anomalies to the overall plant process, and accurately locate the source of anomalies and key propagation units; Step 9: Input online anomaly data and infer control scheme; Input the anomaly data collected by online sensors as evidence into the DBN model, locate the time slice and triggering variable of the anomaly, and output a targeted control scheme based on the model inference. Step 10: Execute the control scheme through the industrial controller and verify anomaly elimination; send the reasoned control scheme to the actuator through the industrial controller, monitor the process status, and verify whether the anomaly has been eliminated; Step 11: Iterative control scheme; guided by knowledge, supplement online data as new evidence, input the model again to reason about the control decision scheme, and continue iterating until the anomaly is eliminated.

[0007] As a preferred embodiment, in step 1, the process of acquiring process data and dividing sub-units based on physical sensors is as follows: S11: Sensor network deployment and data acquisition; installing multi-source physical sensors on key equipment in process industries for sub-units. variable sequence and sub-unit variable sequence The Pearson correlation coefficient is calculated according to formula (1). ; (1); In the formula, and Sub-units and The mean of a sequence of variables.

[0008] As a preferred embodiment, in step 2, the process of slow feature analysis monitoring the rate of change over time is as follows: S21: Construct the SFA optimization objective; assuming the data collected by the sensor is... , For data dimensions, Let be the number of time steps, and construct the optimization objective function according to formula (2); (2); S22: Extracting slow features; first, calculate the weight vector corresponding to the slow features. Then, the slow feature is obtained according to formula (3). Finally, a set of slow features that are crucial to the long-term evolution of the process was selected. ; (3).

[0009] As a preferred embodiment, in step 3, the process of mining deep features of time series using the Long Short-Term Memory network is as follows: S31: LSTM gating and state updates; a. Calculate the input gate according to formula (4) ; (4); b. Calculate the forgetting gate according to formula (5) ; (5); c. Calculate the candidate cell status according to formula (6) ; (6); d. Update the cell state according to formula (7); (7); e. Calculate the output gate according to formula (8) ; (8); f. Calculate the hidden state at the current moment according to formula (9). ; (9); In the formula, It is the sigmoid activation function; This is element-wise multiplication; This is the weight matrix; For bias terms; for Input at any moment; for The status is always hidden.

[0010] S32: Extracting the temporal feature vector; The sensor data after slow feature analysis is input into the LSTM network, and after multiple rounds of training, the temporal feature vector is extracted. , Let be the dimension of the hidden layer, denoted as . .

[0011] As a preferred embodiment, in step 4, the graph convolutional network mines features in the spatial dimension as follows: S41: Constructing spatial diagram structures; constructing spatial diagram structures for industrial processes. ,in It is a set of sub-unit nodes, and the feature of each node is the feature vector of the variables within the sub-unit after preliminary processing. , For node feature dimensions; The set of edges between nodes is defined, and the edge weights are quantized from the relationships between sub-units; an adjacency matrix is ​​constructed. , The number of sub-units, if sub-units and If there is a relationship The value corresponds to the correlation strength; otherwise, it is 0. S42: Extract spatial feature vectors through graph convolution operations; Graph convolution operations update the current node features by aggregating neighbor node information. First, the node feature matrix after the graph convolution operation is calculated according to formula (10). Furthermore, by stacking multiple layers of graph convolutional networks, the complex spatial propagation relationships between sub-units are automatically mined, and spatial feature vectors that can reflect spatial structure and association are extracted. ; (10); In the formula, It is a symmetric normalized adjacency matrix; It is the node feature matrix; , Output feature dimension; .

[0012] As a preferred embodiment, the process of fusing temporal and spatial features in step 5 is as follows: S51: Spatiotemporal feature fusion computation based on fully connected layers; a fusion module is constructed using fully connected layers to extract the temporal feature vectors. With spatial feature vectors Perform splicing and fusion; let the time feature dimension be... The spatial feature dimension is The fused feature dimensions are The calculation process of the fusion process is shown in formula (11); (11); In the formula, This represents the operation of fusing temporal and spatial features; It is a fusion weight matrix; It is a bias term; It is a linear rectified function; S52: Construction and Functional Implementation of a Multi-Dimensional Analysis Framework with Fusion Features; By using a spatiotemporal feature fusion method based on a fully connected layer, the model maps the originally independent temporal dynamic evolution information and spatial unit correlation information to a unified feature space, thus constructing a comprehensive multi-dimensional analysis framework.

[0013] As a preferred embodiment, in step 6, the process of constructing a dynamic Bayesian network model for each sub-unit is as follows: S61: Define the DBN structure; each sub-unit corresponds to a DBN, each DBN contains multiple time slices, and each time slice contains observed variables and latent variables; S62: Set transition probabilities and coupling; the transition probabilities of latent variables across time slices characterize state evolution, and the sub-units are coupled by sharing latent variables or conditional probability tables; the topology is set based on prior process knowledge, and the parameters are learned through historical data.

[0014] As a preferred option, in step 7, the process of using variational inference to solve the marginal likelihood calculation dilemma is as follows: S71: Introduce variational distribution; adopt parameterized variational distribution. Approximate true posterior distribution ; S72: Derive the lower bound of the evidence; based on formula (12), assess the marginal likelihood. Decompose; (12); In the formula, For edge likelihood; It is the likelihood function; It is the prior distribution; Indicates KL divergence; S73: Optimize variational parameters; maximize the ELBO to optimize the variational distribution. Approximate the true posterior distribution as closely as possible .

[0015] As a preferred option, in step 8, the process of constructing an anomaly contribution score model to quantify the anomaly contribution and its degree of influence is as follows: S81: Define the abnormal contribution score; set sub-unit The hidden variables are The variational posterior distribution is The abnormal state distribution is as follows The normal distribution is as follows ; S82: Calculate the KL divergence; calculate the KL divergence according to formula (13); (13); In the formula, Indicates KL divergence; and These represent the distributions of latent variables under abnormal and normal states, respectively. S83: Locate the source of the anomaly; combine contribution threshold and propagation path analysis to locate the source unit and key propagation nodes of the anomaly.

[0016] As a preferred option, in step 9, the process of inputting online anomaly data and reasoning about the control scheme is as follows: S91: Real-time sensor data acquisition and preprocessing; Real-time acquisition of industrial process data using multi-source physical sensors at the same acquisition frequency, followed by preprocessing. S92: Evidence Input and Posterior Inference; The preprocessed data is used as evidence nodes to input into the trained DBN, and variational inference is used to calculate the posterior probability of each latent variable to identify sub-units and variables with abnormally high probabilities. S93: Generate a control scheme; based on causal reasoning, infer the abnormal propagation path and generate a preliminary control scheme.

[0017] This invention addresses the challenges of large-scale processes, strong unit coupling, and complex anomaly propagation paths in the process industry. It proposes a multi-unit collaborative safe operation control method based on spatiotemporal variational Bayesian methods, which offers the following significant technical advantages compared to existing technologies: 1. The sub-unit partitioning based on mutual information fully considers nonlinear dependencies; Traditional methods often use linear correlation coefficients (such as Pearson's) for variable clustering, which are insufficient to characterize the inherent nonlinear coupling in industrial processes. This invention uses mutual information as a partitioning tool, which can quantify the linear and nonlinear dependencies between variables. It clusters closely related informational and logically coordinated processes into sub-units, enabling the rational decomposition of complex industrial processes and laying a foundation for subsequent multi-unit collaborative control that conforms to actual physical mechanisms.

[0018] 2. Slow Feature Analysis (SFA) and Long Short-Term Memory Network (LSTM) are used in synergy to extract multi-level temporal features; Industrial process anomalies are often accompanied by abrupt or gradual changes in parameters. This invention introduces SFA to capture the slow changing trends of process parameters, identify the rate abrupt or gradual change patterns caused by anomalies, and extract essential features reflecting long-term evolution. Simultaneously, it utilizes the gating mechanism of LSTM to learn the temporal dependence of historical data, mining deep temporal features such as anomaly development cycles and parameter fluctuation patterns. These two approaches complement each other, comprehensively characterizing the dynamic evolution of the process over time, and providing reliable temporal data for early anomaly warning.

[0019] 3. Graph Convolutional Networks (GCNs) automatically learn spatial propagation relationships between units; By abstracting sub-units as graph nodes and using material and energy flows between units as edges, this invention aggregates neighbor node information through graph convolution operations to automatically uncover the spatial propagation path and impact range of anomalies. Compared to traditional methods that rely on expert experience to manually define propagation paths, this invention can adaptively learn spatial interaction features from data, accurately extracting spatial dependencies within and between units. It is particularly suitable for scenarios with strong coupling of multiple units in complex process industries.

[0020] 4. Integrate temporal and spatial characteristics to construct a multi-dimensional analysis framework; The temporal features extracted by SFA and LSTM are concatenated and fused with the spatial features extracted by GCN to form a comprehensive representation that includes the dynamic evolution of the process and the interaction of spatial structure. This fused feature simultaneously reflects the cumulative effect of anomalies in time and the propagation effect in space, providing complete input for subsequent Bayesian inference and significantly improving the model's adaptability to complex operating conditions.

[0021] 5. Dynamic Bayesian Networks (DBN) and variational inference solve the problem of marginal likelihood calculation; A DBN is constructed for each sub-unit to describe its state evolution over time and its interactions with other units. Addressing the inherent challenge of computationally difficult marginal likelihood calculation for DBNs, variational inference is introduced. By approximating the latent variable posterior through a parameterized variational distribution, an evidence lower bound (ELBO) is derived as the optimization objective, transforming complex Bayesian inference into an optimizable variational parameter learning problem. This method overcomes the computational bottleneck of traditional Bayesian methods on large-scale industrial data, achieving efficient learning and inference for multi-unit time-series dynamic systems.

[0022] 6. Anomaly Contribution Score (ACS) quantifies the impact of anomalies in sub-units, accurately locating the source and propagation nodes; Based on the latent variable posterior distribution obtained from DBN variational inference, an anomaly contribution score (ACS) is constructed by calculating the KL divergence between the potential distribution of sub-units under abnormal conditions and the distribution under normal conditions. This score comprehensively considers model uncertainty, the cumulative effect of anomalies in the time dimension, and the chain reaction in the spatial dimension. It can accurately measure the contribution rate of each sub-unit to the abnormal operating conditions of the entire plant, identify the source of anomalies and key propagation nodes, and provide a quantitative basis for targeted control decisions.

[0023] 7. Establish a closed-loop safety operation control architecture encompassing perception, inference, decision-making, execution, and iteration; Online sensor data is input into the DBN as evidence to infer the time slice and triggering variables of anomalies, automatically generating control schemes. Commands are then issued to actuators (regulating valves, frequency converters, heaters, etc.) via industrial controllers (PLCs / DCS). The process response after control is monitored, and if the anomaly persists, new evidence is added for iterative reasoning. This closed-loop architecture overcomes the limitations of traditional methods that focus only on fragmented control of a single unit, achieving precise control through multi-unit collaboration and root cause analysis, ensuring the safe and stable operation of the entire plant's industrial processes.

[0024] 8. Supports dynamic model updates to adapt to changes in operating conditions; By using newly acquired industrial process data (normal or abnormal samples), the variational parameters of the DBN are fine-tuned online or periodically retrained, enabling the model to adapt to non-stationary operating conditions such as changes in raw material batches, equipment aging, and fluctuations in environmental temperature and humidity. This dynamic update mechanism ensures the robustness and accuracy of the model in long-term operation, meeting the adaptability requirements of continuous production control systems in process industries.

[0025] This method effectively addresses core challenges in process industries, such as strong coupling of multiple units, complex anomaly propagation paths, and temporal dynamic nonlinearity, through mutual information clustering, spatiotemporal feature collaborative extraction, GCN spatial propagation learning, DBN variational inference, and closed-loop control iteration. It achieves a technological breakthrough from single-point local control to global collaborative safe operation, providing intelligent and collaborative control solutions with practical engineering application value for modern complex industrial production. Attached Figure Description

[0026] Figure 1 This is a flowchart of the present invention; Figure 2 This is the raw coal unit DBN structure in this invention; Figure 3 This is the heavy media coal preparation DBN structure in this invention; Figure 4 This is the DBN structure of the flotation unit in this invention; Figure 5 This is the overflow ash content variation curve of the cyclone in the heavy medium coal preparation unit of this invention; Figure 6 This is the ash content variation curve of the overflow of the flotation cell in the flotation unit of this invention. Detailed Implementation

[0027] like Figures 1 to 6 As shown, the present invention provides a multi-unit collaborative safety control method for process industries based on spatiotemporal variational Bayes, including the following steps; Step 1: Process data acquisition and sub-unit division based on physical sensors; By utilizing multi-source physical sensors deployed at industrial sites to collect multivariate data from each process step in real time, and quantifying the dependencies between data through mutual information, clustering closely related information and collaborative process logic steps, breaking down complex processes into sub-units, and identifying the basic units for multi-unit collaborative control, thus laying the foundation for subsequent multi-unit collaborative control. As a preferred option, the process of acquiring process data and dividing sub-units based on physical sensors is as follows: S11: Sensor network deployment and data acquisition; Install multi-source physical sensors (including but not limited to temperature sensors (thermocouples / resistance temperature detectors), pressure transmitters (piezoresistive), flow meters (electromagnetic / vortex), component analyzers (online near-infrared / gas chromatographs), and vibration sensors (accelerometers)) on key equipment in the process industry (reactors, distillation columns, heat exchangers, pumps, valves, etc.); The multi-source physical sensors synchronously acquire variable data at a fixed sampling frequency (e.g., 1Hz) to form multivariate time series data; S12: Mutual information clustering to divide sub-units; by setting a mutual information threshold, variables with mutual information values ​​higher than the threshold are grouped into the same sub-unit, thus completing the sub-unit division; S13: Pearson correlation coefficient analysis of inter-unit correlation; using Pearson correlation coefficient Measurement sub-unit and The degree of linear correlation between the corresponding variable sequences, for sub-units variable sequence and sub-unit variable sequence The Pearson correlation coefficient is calculated according to formula (1). ; (1); In the formula, and Sub-units and The mean of the variable sequence; this coefficient is used to determine whether there is a positive correlation, negative correlation, or no significant linear relationship between sub-units, providing a basis for subsequent dependency construction.

[0028] This technical solution deploys multi-source physical sensors to synchronously collect multivariate time-series data from key equipment at a fixed frequency, ensuring the comprehensiveness, real-time nature, and synchronization of the data. Simultaneously, mutual information clustering is used to automatically divide the data into sub-units, effectively capturing linear and nonlinear dependencies between variables and avoiding the subjectivity of manual division. Furthermore, Pearson correlation coefficients are employed to quantify the degree of linear correlation between sub-units, providing quantifiable edge weights for the subsequent construction of the spatial graph structure, thus laying the data foundation for multi-unit collaborative control.

[0029] The mutual information method was used to divide the entire coal preparation process into sub-units to reveal the relationships and information transmission characteristics between the various operational units. This division method not only clarifies the functional positioning of each unit within the overall plant process but also provides a clear structural foundation and data support for subsequent multi-unit collaborative control and anomaly analysis. The following sections will explain each sub-unit and its corresponding variables in detail: Unit 1: Raw Coal Processing Unit; The raw coal unit is the starting point of the coal sorting process, primarily responsible for receiving, storing, and initially processing the raw coal mined by the coal preparation plant. The raw coal first enters storage facilities such as raw coal silos. During this process, some simple pre-treatment operations are performed, such as removing larger gangue or classifying the raw coal by particle size. This provides the raw coal feedstock that meets the initial requirements for subsequent core units such as heavy media coal preparation, making it the fundamental input link of the entire coal preparation process. The main equipment in the raw coal process includes: raw coal silos, double-layer screens, and single-layer screens. The raw coal unit contains three variables, and the detailed variable allocation is shown in Table 1.

[0030] Table 1: Variable Allocation for Raw Coal Processing Units Unit 2: Heavy Media Coal Preparation Unit; The heavy media coal preparation unit is a core component of modern coal preparation processes, crucial for achieving efficient and precise separation. Its theoretical basis stems from the significant density difference between the liquid medium and coal / gangue particles. In this unit, raw coal is first mixed with heavy media to form a suspension. Based on the density difference between coal and ore, stratified sedimentation occurs in a hydrocyclone or heavy media tank. The less dense coal particles float to the surface, forming a clean coal layer, while the denser, contaminated minerals sink to form tailings. The operating parameters of the heavy media coal preparation unit include medium density, flow rate, hydrocyclone pressure, and separation section length. These parameters directly affect the separation efficiency and clean coal grade. Simultaneously, the operating status of this unit significantly impacts the processing effect of subsequent flotation units, making it a key sub-unit in the multi-unit collaborative control of the entire plant's industrial process. The heavy media coal preparation unit contains 10 variables, detailed in Table 2.

[0031] Table 2: Variable Allocation for Heavy Medium Coal Preparation Unit Unit 3: Flotation Unit; The flotation unit is a fine separation stage in the coal preparation process, primarily used to recover fine-particle coal that cannot be completely separated in the heavy media coal preparation unit. This unit further separates coal from dirty minerals by adding flotation reagents to the slurry and using rising bubbles to carry hydrophobic coal particles to the froth layer. Key operating parameters of the flotation unit include reagent dosage, slurry concentration, bubble flow rate, and stirring rate, which directly affect flotation efficiency, clean coal grade, and recovery rate. In the overall plant process, the operating status of the flotation unit not only reflects the quality of the clean coal product but also plays a crucial role in the coordinated control of the upstream heavy media coal preparation unit and the downstream tailings treatment process, making it an indispensable core subunit in the multi-unit coordinated safe operation control. The flotation unit involves eight key variables, the specific allocation methods of which are shown in Table 3.

[0032] Table 3: Variable Allocation for Flotation Units Step 2: Slow Feature Analysis (SFA) monitors the rate of change of the process over time; SFA is introduced to extract slowly changing essential features from sensor time series data, identify rate abrupt changes or slow change patterns caused by anomalies, and provide a basis for early warning of anomalies; As a preferred method, the process of slow feature analysis to monitor the rate of change over time is as follows: S21: Construct the SFA optimization objective; Slow feature analysis aims to extract slowly changing features from high-dimensional time series data, assuming the data acquired by the sensor is... , For data dimensions, Let be the number of time steps, and construct the optimization objective function according to formula (2); (2); S22: Extracting slow features; first, calculate the weight vector corresponding to the slow features. Then, the slow feature is obtained according to formula (3). The rate of change of slow features can reflect the overall temporal dynamics of an industrial process. The slower the feature changes, the more it reflects the essential trend of the process. By monitoring the rate of change of the entire process over time, a set of slow features that are key to the long-term evolution of the process can be selected. ; (3).

[0033] In this technical solution, the process constructs a slow feature analysis optimization target to extract the slowest changing essential features from high-dimensional sensor time series data, effectively filtering out high-frequency noise and short-term fluctuations, enabling the model to focus on the long-term evolution trend of industrial processes; by using the rate of change of slow features to monitor process dynamics, it provides a time-dimensional indicator reflecting the essential laws for early warning of anomalies, overcoming the shortcomings of traditional methods that are sensitive to sudden changes but insensitive to slow changes.

[0034] Step 3: Use a Long Short-Term Memory (LSTM) network to mine deep features of the time series; By leveraging LSTM to learn the temporal dependencies of historical sensor data, we can uncover deep temporal features such as abnormal development cycles and parameter fluctuation patterns, thus supplementing the ability of slow feature analysis to characterize complex temporal patterns. As a preferred method, the process of mining deep features of time series data using Long Short-Term Memory (LSTM) networks is as follows: S31: LSTM Gating and State Update; LSTM networks effectively capture long-term dependencies through input gates, forget gates, and cell states; a. Calculate the input gate according to formula (4) ; (4); b. Calculate the forgetting gate according to formula (5) ; (5); c. Calculate the candidate cell status according to formula (6) ; (6); d. Update the cell state according to formula (7); (7); e. Calculate the output gate according to formula (8) ; (8); f. Calculate the hidden state at the current moment according to formula (9). ; (9); In the formula, It is the sigmoid activation function; This is element-wise multiplication; This is the weight matrix; For bias terms; for Input time here This is data after slow feature analysis processing; for The status is always hidden.

[0035] S32: Extracting the temporal feature vector; The sensor data after slow feature analysis is input into the LSTM network. After multiple rounds of training, a temporal feature vector that can accurately characterize the dynamic changes in the process over time is extracted. , Let be the dimension of the hidden layer, denoted as . .

[0036] This technical solution effectively solves the gradient vanishing and gradient exploding problems in long-term dependencies by utilizing the gating mechanism (input gate, forget gate, output gate) and cell state update of LSTM. It can learn complex temporal evolution patterns from sensor data after slow feature processing; and extracts temporal feature vectors through multiple rounds of training. It retains in-depth information such as abnormal development cycles and parameter fluctuation patterns, providing a high-quality temporal dimension representation for subsequent spatiotemporal feature fusion, and significantly enhancing the ability to characterize the dynamic changes of the process.

[0037] Step 4: Graph Convolutional Network (GCN) to mine features in the spatial dimension; Modeling with graph convolutional network, sub-units are abstracted as graph nodes, and the relationships between units such as material flow and energy flow are used as edges. By aggregating information from neighboring nodes, the propagation path and impact range of anomalies in the spatial dimension are mined, and the spatial interaction features between units are extracted. As a preferred method, the process of using graph convolutional networks to mine spatial dimension features is as follows: S41: Constructing spatial diagram structures; constructing spatial diagram structures for industrial processes. ,in It is a set of sub-unit nodes, and the feature of each node is the feature vector of the variables within the sub-unit after preliminary processing. , For node feature dimensions; The set of edges between nodes is defined, and the edge weights are quantized from the relationships between sub-units; an adjacency matrix is ​​constructed. , The number of sub-units, if sub-units and If there is a relationship The value corresponds to the correlation strength; otherwise, it is 0. S42: Extract spatial feature vectors through graph convolution operations; Graph convolution operations update the current node features by aggregating neighbor node information. First, the node feature matrix after the graph convolution operation is calculated according to formula (10). Furthermore, by stacking multiple layers of graph convolutional networks, the complex spatial propagation relationships between sub-units are automatically mined, and spatial feature vectors that can reflect spatial structure and association are extracted. This clearly presents the spatial dependencies within and between units; (10); In the formula, It is a symmetric normalized adjacency matrix ( (for degree matrix) It is the node feature matrix; , Output feature dimension; .

[0038] In this technical solution, the process constructs a spatial graph structure for the industrial process, abstracting sub-units as nodes and material-energy relationships as edges. It then uses graph convolution operations to aggregate neighbor node information and automatically learns the spatial dependencies and anomaly propagation paths between sub-units. The stacking of multiple graph convolutions can expand the receptive field layer by layer, capturing complex spatial interactions across units. The extracted spatial feature vectors clearly reflect the collaboration and coupling within and between units, providing a quantifiable spatial dimension representation for multi-unit collaborative control.

[0039] Step 5: Integrate temporal and spatial features; integrate the temporal features extracted by SFA and LSTM with the spatial features mined by GCN to construct a multi-dimensional analysis framework, comprehensively characterize the dynamic evolution and spatial structure interaction of industrial processes, and provide multi-dimensional feature support for collaborative control; As a preferred approach, the process of integrating temporal and spatial characteristics is as follows: S51: Spatiotemporal feature fusion computation based on fully connected layers; a fusion module is constructed using fully connected layers to extract the temporal feature vectors. With spatial feature vectors Perform splicing and fusion; let the time feature dimension be... The spatial feature dimension is The fused feature dimensions are The calculation process of the fusion process is shown in formula (11); (11); In the formula, This represents the operation of fusing temporal and spatial features; It is a fusion weight matrix; It is a bias term; It is a linear rectified function; S52: Construction and Functional Implementation of a Multi-Dimensional Analysis Framework with Fusion Features; Through a spatiotemporal feature fusion method based on a fully connected layer, the model maps the originally independent temporal dynamic evolution information and spatial unit correlation information to a unified feature space, constructing a comprehensive multi-dimensional analysis framework. This fusion mechanism enables the model to simultaneously consider the temporal trend changes of industrial processes and the inter-unit coupling propagation effects in the spatial dimension, thereby providing a data feature foundation with both temporal evolution and topological correlation characteristics for subsequent multi-unit collaborative safe operation control models.

[0040] This process integrates temporal and spatial features through a fully connected layer, introduces nonlinearity using the ReLU activation function, and constructs a unified multi-dimensional analysis framework. The fused features simultaneously contain the long-term temporal evolution trend of the industrial process and the spatial coupling and propagation relationships between sub-units, enabling subsequent models to collaboratively perceive anomalies from both spatiotemporal dimensions. This provides a more complete feature foundation for the collaborative safe operation control of multi-unit systems. Based on this fused feature, the model's ability to identify the collaborative response between units and anomaly propagation paths under complex operating conditions is enhanced, ultimately improving the accuracy and robustness of industrial process operation status evaluation and control.

[0041] Step 6: Build a dynamic Bayesian network (DBN) model for each sub-unit; For each sub-unit, a DBN model is built to describe the evolution of the sub-unit's state over time and its interaction logic with other units, thus realizing dynamic modeling at the sub-unit level. As a preferred approach, the process of constructing a dynamic Bayesian network model for each sub-unit is as follows: S61: Define the DBN structure; each sub-unit corresponds to a DBN, each DBN contains multiple time slices, and each time slice contains observed variables and latent variables; S62: Set transition probabilities and coupling; the transition probabilities of latent variables across time slices characterize state evolution, and the sub-units are coupled by sharing latent variables or conditional probability tables; the topology is set based on prior process knowledge, and the parameters are learned through historical data.

[0042] This process constructs an independent dynamic Bayesian network (DBN) for each sub-unit, explicitly models the evolution of latent and observed variables over time using a multi-time-slice structure, and captures the temporal dependencies of the internal states of sub-units by leveraging the cross-time-slice transition probabilities of latent variables. Simultaneously, it achieves coupling between sub-units by sharing latent variables or conditional probability tables, thereby accurately characterizing the causal interactions in the collaborative operation of multiple units in industrial processes. This method decomposes the complex plant-wide dynamic system into manageable sub-models, reducing modeling difficulty while preserving the spatiotemporal correlations between units, providing a structured and interpretable probabilistic graphical model foundation for subsequent variational inference and anomaly tracing.

[0043] These DBN structures visually demonstrate the causal relationships and dynamic dependencies among key variables within each subunit, revealing the interactions and potential impacts of these variables during operation. The DBN structure of the raw coal processing unit encompasses key variables such as the discharge flow rate of the double-layer screen, the discharge flow rate of the single-layer screen, and the coal feed rate into the raw coal silo. It reflects the material flow characteristics and time-dependent patterns during raw coal crushing, conveying, and preliminary classification, revealing the regulatory role of the raw coal unit in material supply and flow control throughout the entire coal preparation process, providing a fundamental guarantee for the stable operation of downstream subunits. The DBN structures of the raw coal processing unit, heavy media coal preparation unit, and flotation unit are respectively as follows: Figure 2 , Figure 3 and Figure 4 As shown.

[0044] Step 7: Use variational inference to solve the marginal likelihood calculation dilemma; to address the problem of the difficulty in calculating the marginal likelihood of DBN, we approximate the posterior distribution of latent variables through variational inference, and derive the variational lower bound by using the KL divergence between the variational distribution and the true posterior distribution, thus breaking through the computational bottleneck and achieving efficient model learning and inference. Variational likelihood (VB) is a core method in Bayesian inference that addresses the challenges of inferring complex models by parameterizing and approximating the posterior distribution. Its core logic involves introducing a parameterized variational distribution to approximate the posterior distribution, which is difficult to compute directly, thus transforming the complex inference problem into an optimization problem of a simple variational distribution. Specifically, VB constructs its optimization objective by maximizing the Evidence Lower Bound (ELBO). The process of using variational inference to solve the marginal likelihood calculation dilemma is as follows: S71: Introduce variational distribution; adopt parameterized variational distribution. Approximate the true posterior distribution that is difficult to solve directly. ; S72: Derive the lower bound of the evidence; based on formula (12), assess the marginal likelihood. Decompose; (12); In the formula, Marginal likelihood indicates that it does not depend on model parameters. At that time, the observed data The marginal probability; It is the likelihood function, representing the likelihood given model parameters. At that time, the observed data The probability of occurrence; The prior distribution represents the distribution of model parameters before the data is observed. Prior assumptions about the probability distribution; The KL divergence is used to quantify the degree of difference between two probability distributions; since the KL divergence is always non-negative, i.e. Therefore, the sum of the first two terms constitutes the marginal likelihood. The lower bound of the evidence is the ELBO (Evidence Lower Bound). Maximizing the ELBO is equivalent to minimizing the variational distribution. With the true posterior The KL divergence between them is used to achieve approximate inference of complex posterior distributions.

[0045] S73: Optimize variational parameters; maximize the ELBO to optimize the variational distribution. Approximate the true posterior distribution as closely as possible This completes the variational approximation inference of the model parameters; In this technical solution, a parameterized variational distribution is introduced to approximate the true posterior distribution, which is difficult to solve directly, transforming the complex Bayesian inference problem into an optimization problem. The evidence lower bound (ELBO) is derived by utilizing the non-negativity property of KL divergence. Maximizing the ELBO is equivalent to minimizing the difference between the variational distribution and the true posterior, effectively avoiding the dilemma of high-dimensional integral calculation of marginal likelihood. This method significantly reduces computational complexity while maintaining inference accuracy, and is suitable for dynamic Bayesian network learning of large-scale, high-dimensional data in the process industry, providing an efficient approximate inference basis for subsequent anomaly contribution score calculation and causal inference.

[0046] Step 8: Construct an Anomaly Contribution Score (ACS) model to quantify anomaly contributions and their impact. Construct multi-dimensional indicators to quantify the contribution rate of sub-unit anomalies to the overall plant process, and accurately locate the source of anomalies and key propagation units; combine contribution thresholds and propagation path analysis to accurately locate the source unit of the anomaly and the key node unit of fault propagation.

[0047] As a preferred approach, the anomaly contribution score is based on the posterior distribution of latent variables in a DBN (Depth-Neutral Variable Network), quantifying the degree of impact by comparing the differences between latent variables in sub-units under abnormal conditions and those under normal operating conditions. The process of constructing the anomaly contribution score model to quantify anomaly contributions and their degree of impact is as follows: S81: Define the abnormal contribution score; set sub-unit The hidden variables are The variational posterior distribution is The abnormal state distribution is as follows The normal distribution is as follows ; S82: Calculate KL divergence; Under abnormal operating conditions, the state of the sub-unit may deviate from the normal operating distribution. By calculating the difference between the abnormal state distribution and the normal state distribution, the contribution of the sub-unit to the abnormal propagation can be obtained. Specifically, calculate KL divergence according to formula (13). (13); In the formula, This represents the KL divergence, used to measure the difference between two distributions; and These represent the distributions of latent variables under abnormal and normal states, respectively. S83: Locate the source of the anomaly; combine contribution threshold and propagation path analysis to locate the source unit and key propagation nodes of the anomaly.

[0048] In this technical solution, an anomaly contribution score (ACS) is constructed by calculating the KL divergence between the latent variable distribution of sub-units under abnormal conditions and the distribution under normal conditions. This quantifies the contribution of sub-units to global anomalies into a comparable numerical indicator. The KL divergence can sensitively capture the degree of distribution deviation, thereby accurately reflecting the influence intensity of sub-units in anomaly propagation. Combined with contribution threshold and propagation path analysis, the source unit and key propagation nodes of anomalies can be automatically located, providing a clear quantitative basis for subsequent targeted control decisions. This overcomes the limitations of traditional methods that rely on manual experience and judgment, and improves the objectivity and efficiency of anomaly tracing.

[0049] Step 9: Input online anomaly data and infer control scheme; Abnormal data collected by online sensors is used as evidence to input into the DBN model to locate the time slice and triggering variables of the abnormality. Based on the model inference, a targeted control scheme is output to clarify the direction of abnormality control. As a preferred approach, the inference control scheme process, based on inputting online anomaly data, is as follows: S91: Real-time sensor data acquisition and preprocessing; Real-time acquisition of industrial process data using multi-source physical sensors at the same acquisition frequency, and preprocessing including normalization, feature extraction, and other preprocessing methods similar to those used for training data. S92: Evidence Input and Posterior Inference; The preprocessed data is used as evidence nodes to input into the trained DBN, and variational inference is used to calculate the posterior probability of each latent variable to identify sub-units and variables with abnormally high probabilities. S93: Generate control scheme; based on causal reasoning of abnormal propagation paths, generate preliminary control schemes (e.g., adjusting valve opening, changing feed rate, adjusting reactor temperature setpoint, etc.).

[0050] In this technical solution, the compatibility between online data and model input is ensured by real-time acquisition of multi-source physical sensor data and preprocessing it in accordance with the training data. Variational inference is used to quickly calculate the posterior probability of latent variables, which can identify sub-units and variables with increased anomaly probability in real time, realizing online accurate localization of abnormal states. Based on causal association reasoning of anomaly propagation paths, targeted control schemes (such as adjusting valve opening and changing feed rate) are generated, and the probability inference results are directly converted into executable operation instructions, providing a clear decision basis for subsequent closed-loop control and significantly improving the timeliness of anomaly response and the effectiveness of control measures.

[0051] Tables 4 and 5 are conditional probability tables for the mass variables M (overflow ash from the hydrocyclone) in the heavy media coal preparation unit and W (mass variable W) in the flotation unit, respectively. Abnormal quality indicators are input as evidence into the collaborative safe operation control model to deduce the adjustment strategy for abnormal control variables throughout the plant, as shown in Table 6. "This indicates that the direction of adjustment of the control variable needs to be increased." "" indicates that the direction of the control variable adjustment needs to be decreased, while "-" indicates that the direction of the control variable adjustment remains unchanged.

[0052] Table 4: Conditional Probability Table of Overflow Ash Content M in Heavy Medium Coal Preparation Unit Table 5: Conditional Probability Table of Overflow Ash Content W in Flotation Unit Table 6: Plant-wide Industrial Process Control Variable Adjustment Strategies Step 10: Execute the control scheme through the industrial controller and verify the elimination of the anomaly; The control scheme obtained through reasoning is sent to the actuator through the industrial controller (PLC / DCS) to monitor the process status and verify whether the abnormality has been eliminated.

[0053] As a preferred option, the process of implementing the control scheme and verifying the elimination of anomalies is as follows: S101: Control command issuance; converts the generated control scheme into standard control signals (such as 4-20mA current, Modbus TCP / IP messages), and outputs them to the corresponding actuators such as regulating valves, frequency converters, and heaters through programmable logic controllers (PLCs) or distributed control systems (DCS). S102: Process status monitoring; continuously monitor key process variables (temperature, pressure, flow rate, composition, etc.) in real time through multi-source physical sensors. If the monitored indicators return to normal and the abnormal contribution score decreases significantly, the abnormality is determined to be eliminated and the system returns to normal monitoring mode; otherwise, retain the current status and proceed to step eleven.

[0054] This process sends the control scheme generated by reasoning to a programmable logic controller or distributed control system through standardized signal conversion (such as 4-20mA current, Modbus TCP / IP), driving the actuators such as regulating valves and frequency converters to achieve closed-loop linkage from decision-making to physical execution. At the same time, it continuously monitors key process variables through multi-source physical sensors, and judges whether the anomaly has been eliminated by combining the downward trend of the anomaly contribution score (ACS). If it has not been eliminated, it automatically enters the iterative optimization stage, forming a closed-loop control mechanism, which significantly improves the reliability and automation level of anomaly handling.

[0055] Application results are as follows Figure 5 and Figure 6As shown in the data curve, when an abnormal condition occurs in the industrial process, this method can effectively bring the different overflow ash contents in the entire plant's industrial process back to the threshold range quickly. This not only significantly shortens the abnormal handling cycle, but also reduces the deviation in clean coal quality and capacity loss caused by index fluctuations.

[0056] Step 11: Iterative control scheme; guided by knowledge, supplement online data as new evidence, input the model again to reason about the control decision scheme, and continue to iterate until the anomaly is eliminated to ensure the safe and stable operation of the entire plant process.

[0057] The ash content threshold for the hydrocyclone overflow was set at 10.8%, and the ash content threshold for the flotation cell overflow was set at 6.5%. During the experiment, when the ash content index exceeded the threshold, the spatiotemporal variational Bayesian multi-unit collaborative safe operation control model automatically triggered the abnormal response mechanism, re-collected online abnormal evidence information, and input it into the spatiotemporal variational Bayesian multi-unit collaborative safe operation control model for secondary inference.

[0058] As a preferred option, the iterative control scheme proceeds as follows: S111: Supplementing with new evidence; The latest process response data (from sensors) is added to the DBN as a new evidence node to update the posterior distribution of latent variables; S112: Knowledge Guidance and Re-reasoning; By combining the expert rule base, the anomaly localization and propagation path reasoning are re-performed to generate an adjusted control scheme.

[0059] S113: Loop iteration; Repeated reasoning, execution, and monitoring cycles continue until abnormal conditions are completely eliminated, ensuring the safe and collaborative operation of multiple units in the process industry.

[0060] This process dynamically updates the posterior distribution of latent variables in the DBN using the latest process response data as new evidence. Combined with expert rule base-guided re-inference, it generates an adjusted control scheme and iteratively executes inference, execution, and monitoring until the anomaly is completely eliminated. This achieves adaptive iterative optimization of the control strategy and ensures the closed-loop convergence and robustness of the safe collaborative operation of multiple units.

[0061] To address the fragmentation and lack of coordination among multiple units in plant-wide industrial process control, this invention proposes a spatiotemporal variational Bayesian-based multi-unit collaborative safety control method for process industries, aiming to achieve coordinated and safe operation of the entire plant process. First, the industrial process is divided into sub-units and correlations are analyzed based on mutual information to uncover spatial dependencies within units and at the overall process level. In the temporal dimension, Slow Feature Analysis (SFA) is used to monitor the process change rate, combined with LSTM to extract temporal features. In the spatial dimension, graph convolutional networks are used to automatically learn anomaly propagation relationships between units, completing spatial feature extraction. By deeply fusing spatiotemporal features, a multi-dimensional dynamic analysis framework is constructed to accurately characterize the complex interactions between process dynamics and spatial structure, enhancing the model's multi-unit collaborative control capabilities. Second, a Deep Belief Network (DBN) is constructed for each sub-unit, using variational inference to estimate the posterior distribution of latent variables. A lower bound for evidence is constructed using KL divergence to approximate the posterior distribution of unobservable parameters and latent states, effectively solving the problem of difficult marginal likelihood calculation in traditional DBNs. Simultaneously, anomaly contribution scores are constructed for each sub-unit to quantify the impact of each unit on global anomalies, providing a decision-making basis for subsequent collaborative safety control. The multi-unit collaborative safety control model constructed by this method can effectively handle the uncertainty of spatiotemporal data, uncover causal relationships within and outside units and anomaly spatial propagation paths, and can be dynamically updated based on new data, ensuring collaborative handling of each sub-unit under abnormal operating conditions, ultimately achieving safe, stable, and collaborative operation of the entire plant's industrial processes. Finally, experimental verification in the entire plant's coal preparation production process demonstrates the effectiveness and feasibility of the proposed method, showcasing good engineering application potential.

Claims

1. A multi-unit collaborative safety control method for process industries based on spatiotemporal variational Bayesian methods, characterized in that, Includes the following steps; Step 1: Process data acquisition and sub-unit division based on physical sensors; Multi-source physical sensors are used to collect multivariate data of each process step in real time. The dependencies between data are quantified by mutual information, and closely related processes are clustered into sub-units. Step 2: Slow Feature Analysis to Monitor the Rate of Change over Time; Introduce SFA to extract the essential features of slow changes from sensor time-series data, and identify rate abrupt changes or gradual change patterns caused by anomalies; Step 3: Mining deep features of time series data using Long Short-Term Memory networks; By leveraging LSTM to learn the temporal dependencies of historical sensor data, deep temporal features can be extracted. Step 4: Mining features in the spatial dimension using graph convolutional networks; Modeling with graph convolutional networks, sub-units are abstracted as graph nodes, and the material flow and energy flow between units are associated as edges. By aggregating information from neighboring nodes, the propagation path and impact range of anomalies in the spatial dimension are mined, and the spatial interaction features between units are extracted. Step 5: Integrate temporal and spatial features; combine the temporal features extracted by SFA and LSTM with the spatial features mined by GCN to construct a multi-dimensional analysis framework to comprehensively depict the dynamic evolution of industrial processes and the interaction between spatial structures. Step 6: Construct a dynamic Bayesian network model for each sub-unit; for each sub-unit, build a DBN model to describe the evolution of the sub-unit's state over time, as well as its interaction logic with other units, to achieve dynamic modeling at the sub-unit level. Step 7: Use variational inference to solve the dilemma of marginal likelihood calculation; To address the problem of the difficulty in calculating the marginal likelihood of DBN, we use variational inference to approximate the posterior distribution of latent variables, and use the KL divergence between the variational distribution and the true posterior distribution to derive the variational lower bound, thus breaking through the computational bottleneck. Step 8: Construct an anomaly contribution score model to quantify the contribution of anomalies and their degree of impact; build multi-dimensional indicators to quantify the contribution rate of sub-unit anomalies to the overall plant process, and accurately locate the source of anomalies and key propagation units; Step 9: Input online anomaly data and infer control scheme; Input the anomaly data collected by online sensors as evidence into the DBN model, locate the time slice and triggering variable of the anomaly, and output a targeted control scheme based on the model inference. Step 10: Execute the control scheme through the industrial controller and verify anomaly elimination; send the reasoned control scheme to the actuator through the industrial controller, monitor the process status, and verify whether the anomaly has been eliminated; Step 11: Iterative control scheme; guided by knowledge, supplement online data as new evidence, input the model again to reason about the control decision scheme, and continue iterating until the anomaly is eliminated.

2. The method for multi-unit collaborative safety control in process industries based on spatiotemporal variational Bayes as described in claim 1, characterized in that, In step 1, the process of acquiring process data and dividing sub-units based on physical sensors is as follows: S11: Sensor network deployment and data acquisition; installing multi-source physical sensors on key equipment in process industries for sub-units. variable sequence and sub-unit variable sequence The Pearson correlation coefficient is calculated according to formula (1). ; （1）； In the formula, and Sub-units and The mean of a sequence of variables.

3. A multi-unit collaborative safety control method for process industries based on spatiotemporal variational Bayes, as described in claim 1 or 2, characterized in that... In step 2, the process of slow feature analysis monitoring the rate of change over time is as follows: S21: Construct the SFA optimization objective; assuming the data collected by the sensor is... , For data dimensions, Let be the number of time steps, and construct the optimization objective function according to formula (2); （2）； S22: Extracting slow features; First, solve for the weight vector corresponding to the slow feature. Then, the slow feature is obtained according to formula (3). Finally, a set of slow features that are crucial to the long-term evolution of the process was selected. ; （3）。 4. The method for multi-unit collaborative safety control in process industries based on spatiotemporal variational Bayes, as described in claim 3, is characterized in that... In step 3, the process of mining deep features of time series using the Long Short-Term Memory network is as follows: S31: LSTM gating and state updates; a. Calculate the input gate according to formula (4) ; （4）； b. Calculate the forgetting gate according to formula (5) ; （5）； c. Calculate the candidate cell status according to formula (6) ; （6）； d. Update the cell state according to formula (7); （7）； e. Calculate the output gate according to formula (8) ; （8）； f. Calculate the hidden state at the current moment according to formula (9). ; （9）； In the formula, It is the sigmoid activation function; This is element-wise multiplication; This is the weight matrix; For bias terms; for Input at any moment; for Always hide your status; S32: Extracting the temporal feature vector; The sensor data after slow feature analysis is input into the LSTM network, and after multiple rounds of training, the temporal feature vector is extracted. , Let be the dimension of the hidden layer, denoted as . .

5. The method for multi-unit collaborative safety control in process industries based on spatiotemporal variational Bayes as described in claim 4, characterized in that, In step 4, the process of the graph convolutional network mining spatial dimension features is as follows: S41: Constructing spatial diagram structures; constructing spatial diagram structures for industrial processes. ,in It is a set of sub-unit nodes, and the feature of each node is the feature vector of the variables within the sub-unit after preliminary processing. , For node feature dimensions; The set of edges between nodes is defined, and the edge weights are quantized from the relationships between sub-units; an adjacency matrix is ​​constructed. , The number of sub-units, if sub-units and If there is a relationship The value corresponds to the correlation strength; otherwise, it is 0. S42: Extract spatial feature vectors through graph convolution operations; Graph convolution operations update the current node features by aggregating neighbor node information. First, the node feature matrix after the graph convolution operation is calculated according to formula (10). Furthermore, by stacking multiple layers of graph convolutional networks, the complex spatial propagation relationships between sub-units are automatically mined, and spatial feature vectors that can reflect spatial structure and association are extracted. ; （10）； In the formula, It is a symmetric normalized adjacency matrix; It is the node feature matrix; , Output feature dimension; .

6. The method for multi-unit collaborative safety control in process industries based on spatiotemporal variational Bayes as described in claim 5, characterized in that, In step 5, the process of fusing temporal and spatial features is as follows: S51: Spatiotemporal feature fusion computation based on fully connected layers; a fusion module is constructed using fully connected layers to extract the temporal feature vectors. With spatial feature vectors Perform splicing and fusion; let the time feature dimension be... The spatial feature dimension is The fused feature dimensions are The calculation process of the fusion process is shown in formula (11); (11)； In the formula, This represents the operation of fusing temporal and spatial features; It is a fusion weight matrix; It is a bias term; It is a linear rectified function; S52: Construction and Functional Implementation of a Multi-Dimensional Analysis Framework with Fusion Features; By using a spatiotemporal feature fusion method based on a fully connected layer, the model maps the originally independent temporal dynamic evolution information and spatial unit correlation information to a unified feature space, thus constructing a comprehensive multi-dimensional analysis framework.

7. The method for multi-unit collaborative safety control in process industries based on spatiotemporal variational Bayes as described in claim 2, characterized in that, In step 6, the process of constructing a dynamic Bayesian network model for each sub-unit is as follows: S61: Define the DBN structure; each sub-unit corresponds to a DBN, each DBN contains multiple time slices, and each time slice contains observed variables and latent variables; S62: Set transition probabilities and coupling; the transition probabilities of latent variables across time slices characterize state evolution, and the sub-units are coupled by sharing latent variables or conditional probability tables; the topology is set based on prior process knowledge, and the parameters are learned through historical data.

8. The method for multi-unit collaborative safety control in process industries based on spatiotemporal variational Bayes as described in claim 1, characterized in that, In step 7, the process of using variational inference to solve the marginal likelihood calculation dilemma is as follows: S71: Introduce variational distribution; adopt parameterized variational distribution. Approximate true posterior distribution ; S72: Derive the lower bound of the evidence; based on formula (12), assess the marginal likelihood. Decompose; （12）； In the formula, For edge likelihood; It is the likelihood function; It is the prior distribution; Indicates KL divergence; S73: Optimize variational parameters; maximize the ELBO to optimize the variational distribution. Approximate the true posterior distribution as closely as possible .

9. A multi-unit collaborative safety control method for process industries based on spatiotemporal variational Bayes as described in claim 1, characterized in that, In step 8, the process of constructing an anomaly contribution score model to quantify anomaly contributions and their degree of influence is as follows: S81: Define the abnormal contribution score; set sub-unit The hidden variables are The variational posterior distribution is The abnormal state distribution is as follows The normal distribution is ; S82: Calculate the KL divergence; calculate the KL divergence according to formula (13); (13)； In the formula, Indicates KL divergence; and These represent the distributions of latent variables under abnormal and normal states, respectively. S83: Locate the source of the anomaly; combine contribution threshold and propagation path analysis to locate the source unit and key propagation nodes of the anomaly.

10. A multi-unit collaborative safety control method for process industries based on spatiotemporal variational Bayes as described in claim 1, characterized in that, In step 9, the process of inputting online anomaly data and reasoning about the control scheme is as follows: S91: Real-time sensor data acquisition and preprocessing; Real-time acquisition of industrial process data using multi-source physical sensors at the same acquisition frequency, followed by preprocessing. S92: Evidence Input and Posterior Inference; The preprocessed data is used as evidence nodes to input into the trained DBN. Variational inference is used to calculate the posterior probability of each latent variable and to identify sub-units and variables with abnormally high probabilities. S93: Generate a control scheme; based on causal reasoning, infer the abnormal propagation path and generate a preliminary control scheme.

Images