Method, device and equipment for time-lag causal report generation of process industry control system and medium

Abstract

The application discloses a time delay causal report generation method, device and equipment of a process industry control system and a medium, relates to the field of artificial intelligence technology, and comprises time delay unfolding of time series data of the process industry control system to obtain a historical tensor; single-step state prediction of a delay action vector, a mechanism vector and a parent variable aggregate quantity to obtain a state prediction result of the next period, and multi-step rolling prediction of the time series data; target lag order of each causal edge is determined based on a world model, a causal diagram and a time delay hierarchical diagram are generated; control actions in the time series data are replaced with target control actions, state variables corresponding to the target control actions are deduced based on the world model, and a state change curve is obtained; based on the causal diagram, the time delay hierarchical diagram, the target lag order and the state change curve, a time delay causal report of the process industry control system is generated, the generalization ability of the world model under different working conditions is enhanced, and the accuracy of the time delay causal report generation is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, equipment, and medium for generating time-delay causal reports for process industry control systems. Background Technology

[0002] One of the challenges in process industry control optimization is the widespread and unstable nature of time delays. Common problems encountered in the field include: (1) The fixed time delay assumption is too crude: the response lag is different under different variables, different control actions and different operating conditions. The same action may also have different delays for multiple target variables. This difference cannot be expressed by a uniform fixed time delay window.

[0003] (2) The prediction model lacks causal structure constraints: ordinary recurrent neural networks or sequence models can fit time series, but they are usually difficult to explain which historical variables, which historical actions and which lag order mainly determine a future state. For field control, the prediction accuracy is not enough.

[0004] (3) The process of discovering causal factors first and then predicting and modeling is prone to breakage: If a causal graph is obtained first using a separate algorithm and then the graph is used as a fixed input to train the world model, the causal graph cannot be further corrected based on prediction error and action intervention error.

[0005] (4) Action intervention and external operating conditions need to be distinguished: Process industry data contains controllable and uncontrollable variables. If actions are not encoded in an intervention manner, the world model will treat historical actions as ordinary observations and cannot reliably answer "how the future state will change after a certain control setting is changed".

[0006] As can be seen from the above, how to improve the precision of the fixed time delay assumption, enhance the generalization ability of the world model under different operating conditions, and improve the accuracy of the time delay causal report generation are problems to be solved in this field. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for generating time-delay causal reports in process industry control systems, which can improve the precision of fixed time-delay assumptions, enhance the generalization ability of the world model under different operating conditions, and improve the accuracy of time-delay causal report generation. The specific solution is as follows: In a first aspect, this application discloses a method for generating time-delay causality reports for process industry control systems, including: Acquire time-series data of the process industry control system, determine target state variables based on the time-series data, and perform time-delay expansion on the time-series data to obtain historical tensors; the target state variable is the state variable of the next cycle of the current cycle. Based on the historical tensor, side-gating parameters, and time-delay kernel parameters, a parent variable aggregation quantity is generated for each target state variable. The time series data is encoded to obtain a delayed action vector and a mechanism vector. Single-step state prediction is performed on the delayed action vector, the mechanism vector, and the parent variable aggregation quantity to obtain the state prediction result for the next cycle. The state prediction result for the next cycle is used as the initial state to perform multi-step rolling prediction on the state variables in the time series data to obtain multi-step rolling prediction results. Using the target state variable, the side gating parameter, the time delay kernel parameter, the state prediction result of the next cycle, and the multi-step rolling prediction result, a loss function is constructed, and the initial world model is trained using the loss function to obtain the world model; Based on the world model, each causal edge between each variable in the time series data and the target state variable, as well as the target lag order of each causal edge, are determined. A causal graph and a time-lag layered graph are generated using each causal edge. The target lag order is the lag order corresponding to the maximum time-lag weight of the causal edge. The control actions in the time series data are replaced with target control actions, and the state variables corresponding to the target control actions are deduced based on the world model to obtain state change curves; Based on the causal graph, the time delay hierarchical graph, the target lag order, and the state change curve, a time delay causal report for the process industry control system is generated so that the process industry control system can be controlled based on the time delay causal report.

[0008] Optionally, the step of acquiring time-series data of the process industry control system, determining target state variables based on the time-series data, and performing time-delay expansion on the time-series data to obtain a historical tensor includes: Acquire initial timing data for the process industry control system; The initial time series data is aligned according to a preset sampling period to obtain time series data; the time series data includes state variables, control actions, and external operating conditions. The state variable of the next cycle in the current cycle of the time series data is taken as the target state variable; The state variable, the control action, and the external operating condition are concatenated to obtain a concatenated variable; The maximum time delay order is determined by using the residence time, heat transfer time, conveying distance, and the preset sampling period of the process industry control system. Based on the maximum time delay order, the spliced ​​variables are expanded by time delay to obtain the historical tensor.

[0009] Optionally, based on the historical tensor, edge gating parameters, and time-delay kernel parameters, a parent variable aggregation quantity is generated for each target state variable, including: Based on the side-gating parameters, determine the side-gating weights between each variable and each target state variable in the time series data; At different lag orders, the time delay kernel parameters are exponentially normalized to obtain normalized parameters; Using the edge gating weights and the normalization parameters, the time delay weights between each variable in the time series data and each target state variable at different lag orders are determined; Based on the time delay weights, parent variable aggregates are generated for each target state variable.

[0010] Optionally, the time-series data is encoded to obtain a delayed action vector and a mechanism vector. Single-step state prediction is then performed on the delayed action vector, the mechanism vector, and the aggregated parent variable to obtain the state prediction result for the next cycle, including: Based on the control actions in the time series data, construct an action queue and an action difference queue; The action queue and the action differential queue are encoded to obtain a delayed action vector; Construct a variable window based on external operating conditions in time series data; The variables in the variable window are encoded to obtain the mechanism vector; Using structural equation network, single-step state prediction is performed on the delayed action vector, the mechanism vector, and the aggregated parent variable to obtain the state prediction result for the next cycle.

[0011] Optionally, the step of constructing a loss function using the target state variable, the side gating parameter, the time delay kernel parameter, the state prediction result of the next cycle, and the multi-step rolling prediction result includes: Based on the target state variable and the state prediction result of the next period, determine the single-step prediction error; Based on the multi-step rolling prediction results and the state variables in the time series data corresponding to the multi-step rolling prediction results, the multi-step rolling error is determined. Determine graph sparsity constraints based on edge gating parameters; Based on the time delay kernel parameters, determine the time delay kernel smoothing constraint and the time delay peak constraint; Construct a loss function that includes single-step prediction error, multi-step rolling error, graph sparsity constraint, time-delay kernel smoothing constraint, and time-delay peak constraint; The loss function is used to update the edge-gated parameters, time-delay kernel parameters, and model parameters in the initial world model, resulting in updated edge-gated parameters, updated time-delay kernel parameters, and the world model.

[0012] Optionally, the step of determining, based on the world model, each causal edge between each variable in the time series data and the target state variable, and the target lag order of each causal edge, and generating a causal graph and a time-lag layered graph using each causal edge, includes: Based on the updated edge gating parameters, determine the edge gating weights between each variable in the time series data and the target state variable; The weights with edge gating weights greater than the preset edge gating weight threshold are used as the target edge gating weights; The edge between the variable corresponding to the target edge gating weight in the time series data and the target state variable is taken as a causal edge; Based on the updated edge gating parameters and the updated time delay kernel parameters, the time delay weights of each causal edge under different time delay orders are determined; The lag order corresponding to the maximum value of the time delay weight is taken as the target lag order; The causal graph is generated using each of the causal edges; Based on the target lag order, each causal edge is displayed in a separate column to generate a time-lag layered diagram.

[0013] Optionally, the control actions in the time-series data are replaced with target control actions, and the state variables corresponding to the target control actions are deduced based on the world model to obtain state change curves; based on the causal graph, the time-delay hierarchical graph, the target lag order, and the state change curves, a time-delay causal report for the process industry control system is generated, including: The state variables, control actions, and external conditions in the time series data are input into the world model for simulation, and the baseline trajectory is output. The state variables and external conditions are fixed, and one or more control actions in the time series data are replaced with target control actions. Based on the world model, the state variables corresponding to the target control actions are deduced to obtain state change curves. Based on the baseline trajectory and the state change curve, the magnitude and direction of the impact of the target control action under different lag orders are determined; Based on the cause-effect graph, time delay hierarchical graph, target lag order, state change curve, the magnitude of the influence, and the direction of the influence, a time delay cause-effect report for the process industry control system is generated.

[0014] Secondly, this application discloses a time-delay causality report generation device for a process industry control system, comprising: The time-delay expansion module is used to acquire time-series data of the process industry control system, determine the target state variable based on the time-series data, and perform time-delay expansion on the time-series data to obtain the historical tensor; the target state variable is the state variable of the next cycle of the current cycle. The rolling prediction module is used to generate parent variable aggregation quantities for each target state variable based on the historical tensor, side gating parameters, and time delay kernel parameters; encode the time series data to obtain delayed action vectors and mechanism vectors; perform single-step state prediction on the delayed action vectors, the mechanism vectors, and the parent variable aggregation quantities to obtain the state prediction result for the next cycle; use the state prediction result of the next cycle as the initial state; and perform multi-step rolling prediction on the state variables in the time series data to obtain multi-step rolling prediction results. The training module is used to construct a loss function using the target state variable, the side gating parameter, the time delay kernel parameter, the state prediction result of the next cycle, and the multi-step rolling prediction result, and to train the initial world model using the loss function to obtain the world model. The causal graph and time-delay layered graph generation module is used to determine, based on the world model, each causal edge between each variable in the time series data and the target state variable, as well as the target lag order of each causal edge, and to generate a causal graph and a time-delay layered graph using each causal edge; the target lag order is the lag order corresponding to the maximum time-delay weight of the causal edge; The deduction module is used to replace the control actions in the time series data with the target control actions, and to deduce the state variables corresponding to the target control actions based on the world model to obtain the state change curves. The time-delay causality report generation module is used to generate a time-delay causality report for the process industry control system based on the causality graph, the time-delay hierarchical graph, the target lag order, and the state change curve, so as to control the process industry control system based on the time-delay causality report.

[0015] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned method for generating time-delay causal reports for process industry control systems.

[0016] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed method for generating time-delay causal reports for a process industry control system.

[0017] Beneficial Effects: This application provides a method for generating time-delay causal reports for process industry control systems. The method includes acquiring time-series data of the process industry control system; determining target state variables based on the time-series data; performing time-delay expansion on the time-series data to obtain a historical tensor, thus avoiding information leakage; generating parent variable aggregates for each target state variable based on the historical tensor, side-gating parameters, and time-delay kernel parameters; encoding the time-series data to obtain delayed action vectors and mechanism vectors; performing single-step state prediction on the delayed action vectors, mechanism vectors, and parent variable aggregates to obtain the state prediction result for the next cycle; using the state prediction result of the next cycle as the initial state, performing multi-step rolling prediction on the state variables in the time-series data to obtain multi-step rolling prediction results, thus improving the stability and accuracy of single-step state prediction and multi-step rolling prediction; constructing a loss function using the target state variables, side-gating parameters, time-delay kernel parameters, the state prediction result of the next cycle, and the multi-step rolling prediction results; and using the loss function to analyze the initial world... The model is trained to obtain a world model, thus improving the fitting accuracy of the world model. Based on the world model, the causal edges between each variable in the time series data and the target state variable, as well as the target lag order of each causal edge, are determined. Causal graphs and time-delay hierarchical graphs are generated using these causal edges, thus improving the accuracy of determining the causal relationships between each variable and the target state variable and their target lag orders, and enhancing the intuitiveness of causal and lag relationships. Control actions in the time series data are replaced with target control actions, and state change curves are derived from the state variables corresponding to the target control actions based on the world model, thus improving the accuracy of analyzing the state change process caused by the target control actions. Based on the causal graph, time-delay hierarchical graph, target lag order, and state change curves, a time-delay causal report for the process industry control system is generated, enabling control of the process industry control system based on the time-delay causal report, thus improving the accuracy of the time-delay causal report generation and ensuring the completeness of the relevant analysis results in the time-delay causal report. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This application discloses a flowchart of a method for generating time-delay causal reports for a process industry control system. Figure 2 A specific flowchart is generated for the time-delay causality report of the process industry control system disclosed in this application; Figure 3 This application discloses a system architecture diagram for generating time-delay causal reports for a process industry control system. Figure 4 This is a schematic diagram of the structure of a time-delay causality report generation device for a process industry control system disclosed in this application; Figure 5 This is a schematic diagram of the structure of an electronic device disclosed in this application. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] One of the difficulties in process industry control optimization is the widespread and unstable existence of time delays. Common problems on-site include: (1) The fixed time delay assumption is too coarse: the response lag is different under different variables, different control actions and different operating conditions. The same action may also have different delays for multiple target variables. A uniform fixed lag window cannot express this difference. (2) The prediction model lacks causal structure constraints: ordinary recurrent neural networks or sequence models can fit time series, but it is usually difficult to explain which historical variables, which historical actions and which lag order determine a certain future state. For on-site control, the prediction accuracy is not enough. (3) Causal discovery first and prediction modeling later are prone to breakage: if a causal graph is obtained first using a separate algorithm and then the graph is used as a fixed input to train the world model, the causal graph cannot continue to be corrected according to the prediction error and action intervention error. (4) Action intervention and external operating conditions need to be distinguished: process industry data contains controllable variables and uncontrollable variables. If the action is not encoded in an intervention manner, the world model will treat the historical action as an ordinary observation and cannot reliably answer "how the future state will change after a certain control setting is changed". As can be seen from the above, how to improve the precision of the fixed time delay assumption, enhance the generalization ability of the world model under different operating conditions, and improve the accuracy of the time delay causal report generation are problems to be solved in this field.

[0022] See Figure 1 As shown in the figure, this invention discloses a method for generating time-delay causality reports for process industry control systems, which may specifically include: Step S11: Obtain the time series data of the process industry control system, determine the target state variable based on the time series data, and perform time delay expansion on the time series data to obtain the historical tensor; the target state variable is the state variable of the next cycle of the current cycle.

[0023] In this embodiment, initial time-series data of the process industry control system is acquired; the initial time-series data is aligned according to a preset sampling period to obtain time-series data; the time-series data includes state variables, control actions, and external operating conditions; the state variables of the next cycle of the current cycle in the time-series data are used as target state variables; the state variables, control actions, and external operating conditions are concatenated to obtain concatenated variables; the maximum time delay order is determined using the residence time, heat transfer time, conveying distance of the process industry control system, and the preset sampling period; based on the maximum time delay order, the concatenated variables are time-delayed expanded to obtain historical tensors.

[0024] In this step, the state variable is The control action is External operating conditions are The system concatenates state variables, control actions, and external operating conditions to form concatenated variables. The maximum time delay order is The history tensor constructed by the system based on the maximum time delay order includes: ; in, The concatenation variable corresponding to the previous period of the current period. This is a concatenation variable corresponding to the two previous periods of the current period. For the previous period The splicing variables corresponding to each period.

[0025] Maximum time delay order It is determined based on residence time, heat transfer time, transport distance, and sampling cycle.

[0026] When generating the historical tensor, the system first creates a list of variables and a list of target state variables. The list of variables includes state variables, control actions, and external conditions, while the target state variables are the state variables for the next cycle of the current cycle. Each sample window is saved as a three-dimensional tensor, with the three dimensions corresponding to the batch size, lag order, and variable, respectively. Each target state variable corresponds to a set of learnable side-gated parameters and time-delay kernel parameters, thus preserving the correspondence between variables, target state variables, and lag orders in the data structure.

[0027] Sample window length The time delay order should be no less than the maximum delay order and should be selectable based on the verification error. Each sample window must not cross irrelevant batch, maintenance switch, or trajectory boundaries. The model output should return to the original process state space, and action variables must retain the interventionist identifier.

[0028] Regarding sample validity, the system records trajectory boundaries, shutdown flags, operating condition switching flags, and key variable validity flags for each sample window. When a sample window spans shutdown, maintenance, batch switching, or sampling interruption, that sample window is not used for continuous dynamics training; when retaining the corresponding data, it is modeled as a separate operating condition segment to avoid treating discontinuous events as normal time-delay causal relationships in training.

[0029] Step S12: Based on the historical tensor, edge gating parameters, and time delay kernel parameters, generate parent variable aggregation quantities for each target state variable, encode the time series data to obtain delayed action vectors and mechanism vectors, perform single-step state prediction on the delayed action vectors, the mechanism vectors, and the parent variable aggregation quantities to obtain the state prediction result for the next cycle, use the state prediction result of the next cycle as the initial state, and perform multi-step rolling prediction on the state variables in the time series data to obtain multi-step rolling prediction results.

[0030] In this embodiment, based on the side-gating parameters, the side-gating weights between each variable in the time-series data and each target state variable are determined; the time-delay kernel parameters are exponentially normalized at different lag orders to obtain normalized parameters; using the side-gating weights and the normalized parameters, the time-delay weights between each variable in the time-series data and each target state variable at different lag orders are determined; based on the time-delay weights, parent variable aggregation quantities are generated for each target state variable; based on the control actions in the time-series data, an action queue and an action difference queue are constructed; the action queue and the action difference queue are encoded to obtain a delayed action vector; based on the external operating conditions in the time-series data, a variable window is constructed; the variables in the variable window are encoded to obtain a mechanism vector; using a structural equation network, a single-step state prediction is performed on the delayed action vector, the mechanism vector, and the parent variable aggregation quantity to obtain the state prediction result for the next cycle; using the state prediction result of the next cycle as the initial state, multi-step rolling prediction is performed on the state variables in the time-series data to obtain a multi-step rolling prediction result.

[0031] Let the target state variable be numbered as The variable number is The lag order is The time delay weights for system learning are: ; in, For variables In lag order Below the target state variable The time delay weight, This is a function for processing the side-gating parameters. For variables to target state variable The side gating parameters, For variables to target state variable The causal edge in the lag order The time-delay kernel parameters on, This refers to the lag order in the exponential normalization process. For variables to target state variable The causal edge in the lag order The time-delay kernel parameters.

[0032] The former determines whether an edge exists, while the latter determines the main effective delay of that edge.

[0033] For the target state variable The aggregate size of the parent variable is: ; in, For the current period Next target state variable The amount of aggregation of the parent variable, For variables Before the current cycle The variable values ​​for each period.

[0034] Action queue retains the most recent one The system sets the control action settings for each cycle, and records the changes in control actions between adjacent cycles in the action differential queue to distinguish the different impacts of stable and abrupt settings on future states. The system encodes the action queue and action differential queue, and the delayed action vector is: ; in, For the current period The corresponding delayed action vector, An encoder for encoding action queues and action difference queues. The control action corresponding to the previous cycle in the current cycle. The control actions corresponding to the two previous cycles of the current cycle. For the previous period The control actions corresponding to each cycle This represents the change in control action corresponding to the previous cycle in the current cycle. For the previous period The change in control action corresponding to each cycle.

[0035] For external operating conditions, the system constructs a variable window and encodes the variables within it. The variable window can generate mechanism labels based on shift, equipment mode, raw material type, or load range. The mechanism vector is: ; in, For the current period The corresponding mechanism vector, An encoder that encodes external operating conditions in a variable window. The first before the current cycle External operating conditions corresponding to each cycle The first before the current cycle -1 cycle corresponds to the external operating conditions This refers to the external operating conditions corresponding to the current cycle.

[0036] The state transition of the target state variable is as follows: ; in, For the target state variable The state prediction results for the next period For the target state variable In the current cycle The state variable value, For the target state variable The corresponding residual scaling factor, For the target state variable The corresponding structural equation network, For the current period Next target state variable The amount of aggregated parent variables.

[0037] The structural equation network (SEM) uses a residual form, adding the current cycle's state variable value to the SEM output's state change to obtain the state prediction result for the next cycle. Subsequently, the system uses the next cycle's state prediction result as the initial state and continues to use the prediction result as subsequent input to perform multi-step rolling predictions on the state variables in the time series data, obtaining multi-step rolling prediction results. The model output is the original process state, thus enabling comparison between the state prediction results and the state variables in the time series data.

[0038] Step S13: Construct a loss function using the target state variable, the side gating parameter, the time delay kernel parameter, the state prediction result of the next cycle, and the multi-step rolling prediction result. Use the loss function to train the initial world model to obtain the world model.

[0039] In this embodiment, the single-step prediction error is determined based on the target state variable and the state prediction result of the next cycle; the multi-step rolling error is determined based on the multi-step rolling prediction result and the state variable corresponding to the multi-step rolling prediction result in the time series data; the graph sparsity constraint is determined based on the edge gating parameter; the time delay kernel smoothing constraint and the time delay peak constraint are determined based on the time delay kernel parameter; a loss function is constructed including the single-step prediction error, multi-step rolling error, graph sparsity constraint, time delay kernel smoothing constraint, and time delay peak constraint; the edge gating parameter, time delay kernel parameter, and model parameters in the initial world model are updated using the loss function to obtain the updated edge gating parameter, updated time delay kernel parameter, and world model.

[0040] The loss function is: ; in, For loss function, This represents the single-step prediction error. For multi-step rolling error, To achieve sparse constraints, For time-delay kernel smoothing constraints, For time delay peak constraint, For multi-step rolling error weights, To impose sparse constraint weights on the graph, For time-delay kernel smoothing constraint weights, The time-delay peak constraint weight.

[0041] The graph sparsity constraint weight is used to control the number of edges retained; the time delay kernel smoothing constraint weight is used to avoid meaningless jitter of the time delay kernel; the time delay peak constraint weight is used to encourage the formation of a clear dominant time delay; and the multi-step rolling error weight is used to improve the stability of long-term prediction.

[0042] During training, the system executes the following sequence: First, the system initializes the side-gating parameters and time-delay kernel parameters, and applies variable role masks to prevent uninterrupted external conditions from being used as control action outputs or incorrectly fed back to the past as target state variables; Second, the system calculates the aggregated amount of parent variables for each batch of sample windows and sets the aggregated amount of parent variables... Delayed action vector and mechanism vector First, input the structural equation network corresponding to each target state variable together; second, supervise the single-step prediction with the real next state, and continue to roll with the model's own prediction results to supervise multi-step stability.

[0043] The loss function participates in backpropagation, updating the edge-gated parameters, time-delay kernel parameters, and model parameters in the initial world model. This ensures that the weights of causal edges and lag orders, control action responses, external operating conditions, and state transitions are all within the same trainable computational graph. The edge-gated parameters and time-delay kernel parameters are updated together with the single-step prediction error, multi-step rolling error, graph sparsity constraints, time-delay kernel smoothing constraints, and time-delay peak constraints, resulting in updated edge-gated parameters, updated time-delay kernel parameters, and the world model.

[0044] If the time delay kernel of a causal edge exhibits peaks simultaneously at multiple mutually distant lag orders, the system retains the causal edge and marks it as a multi-peak response edge to indicate the existence of multiple process channels or multiple device response times. Thus, while retaining the causal edge, the system can display its weight distribution at different lag orders in subsequent reports.

[0045] To improve feasibility, the system performs edge stability recalculation after training. Specifically, the target lag order is calculated on the pre-training, post-training, and validation sets. If the target lag order of the same causal edge varies beyond the allowable range in different time segments, the reported confidence of that causal edge is reduced; if the target lag order is stable and the prediction error is sensitive to that causal edge, then the causal edge is added to the high-confidence time-delay causal edge table.

[0046] Step S14: Based on the world model, determine each causal edge between each variable in the time series data and the target state variable, as well as the target lag order of each causal edge, and generate a causal graph and a time-delay layered graph using each causal edge; the target lag order is the lag order corresponding to the maximum time-delay weight of the causal edge.

[0047] In this embodiment, based on the updated edge gating parameters, the edge gating weights between each variable in the time series data and the target state variable are determined; the weights with edge gating weights greater than a preset edge gating weight threshold are used as target edge gating weights; the edges between the variables in the time series data corresponding to the target edge gating weights and the target state variables are used as causal edges; based on the updated edge gating parameters and the updated time delay kernel parameters, the time delay weights of each causal edge at different time delay orders are determined; the time delay order corresponding to the maximum time delay weight is used as the target time delay order; the causal graph is generated using each causal edge; and the causal edges are displayed in columns according to the target time delay order to generate a time delay layered graph.

[0048] In this step, edge gating weights are filtered according to a preset edge gating weight threshold, retaining the causal edges between the variables corresponding to the target edge gating weights and the target state variables. The edge gating weight threshold is used to filter causal edges in the report, and the updated edge gating parameters and the updated time delay kernel parameters jointly determine the time delay weights of each causal edge at different lag orders.

[0049] The target lag order of each causal edge is: ; in, For variables to target state variable The objective lag order corresponding to the causal edge. This indicates the lag order that determines the time delay weight to achieve its maximum value.

[0050] In this embodiment, a causal graph is generated using the selected causal edges, and each causal edge is displayed separately according to the target lag order to generate a time-lag layered graph. The time-lag layered graph displays the corresponding causal edges separately under different historical orders, thereby presenting the influence path between variables in past periods and the target state variable.

[0051] This application generates a time-delayed causal edge table based on the causal edges and the target lag order. The fields of the time-delayed causal edge table include the target state variable, variable, target lag order, edge strength, edge confidence, and direction of influence. A time-delayed kernel table is also generated to record the time-delay weight of each causal edge at different lag orders, so as to analyze the response time of the control action and the propagation time of the state variable.

[0052] When the time delay kernel of a causal edge does not have a significant peak, the system marks the corresponding causal edge as a "time delay unstable edge". When the difference in the target lag order of a causal edge between the training set and the test set exceeds a threshold, the system prompts in the time delay causality report that the corresponding causal relationship needs to be reviewed in segments according to external operating conditions.

[0053] Step S15: Replace the control action in the time series data with the target control action, and deduce the state variables corresponding to the target control action based on the world model to obtain the state change curve.

[0054] In this embodiment, the state variables, control actions, and external conditions in the time series data are input into the world model for deduction, and a baseline trajectory is output. The state variables and external conditions are fixed, and one or more control actions in the time series data are replaced with target control actions. Based on the world model, the state variables corresponding to the target control actions are deduced to obtain the state change curve.

[0055] When extrapolating the target control action, the system fixes the historical state variables and external conditions, replacing only one or more control actions in the action queue. The world model recalculates the delayed action vector based on the replaced action queue and outputs the state change curve corresponding to the target control action. The baseline trajectory corresponds to the state variables, control actions, and external conditions in the time-series data, while the state change curve corresponds to the extrapolation result after fixing the state variables and external conditions and replacing the control actions with the target control action.

[0056] Step S16: Based on the cause-effect graph, the time delay hierarchical graph, the target lag order, and the state change curve, generate a time delay cause-effect report for the process industry control system so as to control the process industry control system based on the time delay cause-effect report.

[0057] In this embodiment, based on the baseline trajectory and the state change curve, the influence amplitude and direction of the target control action under different lag orders are determined; based on the cause-effect graph, time delay layer graph, target lag order, state change curve, influence amplitude, and influence direction, a time delay causality report of the process industry control system is generated so that the process industry control system can be controlled based on the time delay causality report.

[0058] The time-delay causality report includes a time-delay causality side table, a time-delay kernel table, a causal graph, a time-delay hierarchical graph, world model prediction results, control action intervention results, and an acceptance report. World model prediction results include single-step state predictions, multi-step rolling predictions, error curves, and comparisons of key variable trends. Control action intervention results include state change curves after changing valve, pump frequency, reagents, aeration rate, or temperature control settings. The acceptance report includes graph sparsity, target lag order stability, control action intervention consistency, and test set rolling prediction error.

[0059] The post-processing stage includes parameter storage and the generation of a verifiable report. The system saves the time-delay kernel as a table, saves the main variables, target lag orders, and edge strengths corresponding to each target state variable as structured files, and saves the state change curves of the control actions as a traceable time series. The time-delay causality report includes both a global causal graph and local variable descriptions. The local variable descriptions include the first three types of variables corresponding to the target state variables, the corresponding target lag orders, the contribution of the control actions, and the mechanism contribution of the external operating conditions.

[0060] If the prediction error of the target state variable is significantly higher than that of other target state variables, the system re-performs local diagnosis based on the variable contribution. When the high-contribution variable is an external operating condition, the time-delay causality report prompts for supplementary external operating condition data collection; when the high-contribution variable comes from the action queue corresponding to the control action, the time-delay causality report prompts that the effective time of the control action may exceed the current maximum time delay order; when the high-contribution variable comes from a state variable but the target lag order is unstable, the time-delay causality report prompts for segmented modeling of the corresponding target state variable according to the external operating condition.

[0061] In this embodiment, the specific process for generating time-delay causality reports in the process industry control system is as follows: Figure 2 As shown, this invention employs time-delay expansion, causal aggregation, intervention coding, mechanism modulation, and rolling prediction steps. Specifically: The first and second steps involve preparing time-series data. The system aligns the initial time-series data of the process industry control system according to a preset sampling period to obtain time-series data including state variables, control actions, and external operating conditions. The state variables of the next cycle of the current cycle are used as the target state variables. The state variables, control actions, and external operating conditions are concatenated to obtain concatenated variables. The concatenated variables are then expanded based on the maximum time delay order to obtain the historical tensor. The purpose of this is to ensure that the model can only use past information to predict future states and to avoid future information leakage caused by random window partitioning.

[0062] The third and fourth steps are the differentiable time-delay causal modeling stage. The system first generates parent variable aggregation quantities for each target state variable based on the historical tensor, side-gating parameters, and time-delay kernel parameters. Then, the time series data is encoded to obtain delayed action vectors and mechanism vectors. Steps five through eight are the joint training and output phase of the world model. Single-step state prediction is performed on the delayed action vector, mechanism vector, and aggregated parent variable to obtain the state prediction result for the next cycle. Using the state prediction result of the next cycle as the initial state, multi-step rolling prediction is performed on the state variables in the time series data to obtain the multi-step rolling prediction result. Then, the single-step prediction error, multi-step rolling error, graph sparsity constraint, and time delay kernel smoothing constraint are used for backpropagation to update the time delay graph, time delay kernel, and world model parameters. After training is completed, the system outputs a time delay causality report for the process industry control system to explain at which lag order each causal edge is mainly effective.

[0063] The system architecture diagram for the time-delay causality report generation system of the process industry control system is as follows: Figure 3As shown, the system includes a data and window module, a differentiable time-delay causality module, an intervention and mechanism module, and a world model and reporting module. The data and window module constructs historical sequences corresponding to state variables, control actions, and external conditions; the differentiable time-delay causality module determines time-delay weights based on side-gating parameters and time-delay kernel parameters, and generates the aggregate quantity of parent variables and the target lag order; the intervention and mechanism module encodes control actions and external conditions; and the world model and reporting module performs single-step state prediction, multi-step rolling prediction, and target control action deduction, and generates a time-delay causality report.

[0064] This application utilizes edge-gating weights and normalized parameters to determine the time-delay weights between each variable in the time-series data and each target state variable at different lag orders. The time-delay weights are composed of three dimensions: the target state variable, the variables in the time-series data, and the lag order, reflecting the influence of each variable on the target state variable at different lag orders. The edge-gating parameters and time-delay kernel parameters are located in a trainable computation graph and can be updated based on single-step prediction error and multi-step rolling error, avoiding the use of a fixed causal graph. Single-step state prediction is performed on the delayed action vector, mechanism vector, and parent variable aggregation to obtain the state prediction result for the next cycle. Using the state prediction result of the next cycle as the initial state, multi-step rolling prediction is performed on the state variables in the time-series data to obtain multi-step rolling prediction results, utilizing the target state variable... The model constructs a loss function based on gating parameters, time delay kernel parameters, state prediction results for the next cycle, and multi-step rolling prediction results. This loss function is then used to train the initial world model, resulting in a world model capable of answering the question "If a certain action is changed, how will the future state change?" rather than simply passively predicting. By encoding slowly changing factors such as raw materials, load, and equipment modes, the model gains better generalization ability under different operating conditions. Based on causal graphs, time delay hierarchical graphs, target lag order, and state change curves, a time delay causal report for the process industry control system is generated, facilitating process engineers' review of action effectiveness times and variable propagation paths.

[0065] Taking a flue gas denitrification process as an example, the process industry control system collects inlet nitrogen oxide concentration, flue gas flow rate, flue gas temperature, oxygen content, ammonia injection rate, fan status, outlet nitrogen oxide concentration, and ammonia slip. There is a transport delay between the adjustment of the ammonia injection rate and the change in the outlet indicators. The system determines the maximum time lag order and updates the time lag weights of the ammonia injection rate on the outlet nitrogen oxide concentration and ammonia slip under different lag orders during training.

[0066] After training, the system can determine the target lag order of ammonia injection rate with respect to outlet nitrogen oxide concentration and ammonia slip, and also determine different target lag orders of inlet nitrogen oxide concentration and flue gas flow rate with respect to outlet indicators. After replacing the ammonia injection rate in the time-series data with the target ammonia injection rate, the world model extrapolates the outlet nitrogen oxide concentration and ammonia slip, obtaining the corresponding state change curves. The time-delay causality report displays the state change curve of the ammonia injection rate, the magnitude of its influence, the direction of its influence, and the target lag order, used to verify the delayed impact of ammonia injection rate adjustments.

[0067] In addition, the above methods can also be applied to water aeration, chemical reaction temperature control, metallurgical furnace temperature control, pump station water level regulation, fermentation process feeding, and cement kiln system air-fuel material regulation.

[0068] As can be seen, this application provides a method for generating time-delay causal reports for process industry control systems. This method includes acquiring time-series data of the process industry control system; determining target state variables based on the time-series data; performing time-delay expansion on the time-series data to obtain a historical tensor, thus avoiding information leakage; generating parent variable aggregates for each target state variable based on the historical tensor, side-gating parameters, and time-delay kernel parameters; encoding the time-series data to obtain delayed action vectors and mechanism vectors; performing single-step state prediction on the delayed action vectors, mechanism vectors, and parent variable aggregates to obtain the state prediction result for the next cycle; using the state prediction result of the next cycle as the initial state, performing multi-step rolling prediction on the state variables in the time-series data to obtain multi-step rolling prediction results, thus improving the stability and accuracy of single-step state prediction and multi-step rolling prediction; constructing a loss function using the target state variables, side-gating parameters, time-delay kernel parameters, the state prediction result of the next cycle, and the multi-step rolling prediction results; and using the loss function to adjust the initial world model. Training is performed to obtain a world model, thus improving the fitting accuracy of the world model. Based on the world model, the causal edges between each variable in the time series data and the target state variable, as well as the target lag order of each causal edge, are determined. Causal graphs and time-delay hierarchical graphs are generated using these causal edges, thus improving the accuracy of determining the causal relationships between each variable and the target state variable and their target lag orders, and enhancing the intuitiveness of causal and lag relationships. Control actions in the time series data are replaced with target control actions, and state change curves are derived from the state variables corresponding to the target control actions based on the world model, thus improving the accuracy of analyzing the state change process caused by the target control actions. Based on the causal graph, time-delay hierarchical graph, target lag order, and state change curves, a time-delay causal report for the process industry control system is generated, enabling control of the process industry control system based on the time-delay causal report, thus improving the accuracy of the time-delay causal report generation and ensuring the completeness of the relevant analysis results in the time-delay causal report.

[0069] See Figure 4 As shown in the figure, an embodiment of the present invention discloses a time-delay causality report generation device for a process industry control system, which may specifically include: The time-delay expansion module 11 is used to acquire time-series data of the process industry control system, determine the target state variable based on the time-series data, and perform time-delay expansion on the time-series data to obtain the historical tensor; the target state variable is the state variable of the next cycle of the current cycle. The rolling prediction module 12 is used to generate parent variable aggregation quantities for each target state variable based on the historical tensor, side gating parameters, and time delay kernel parameters; encode the time series data to obtain delayed action vectors and mechanism vectors; perform single-step state prediction on the delayed action vectors, the mechanism vectors, and the parent variable aggregation quantities to obtain the state prediction result for the next cycle; use the state prediction result of the next cycle as the initial state; and perform multi-step rolling prediction on the state variables in the time series data to obtain multi-step rolling prediction results. Training module 13 is used to construct a loss function using the target state variable, the side gating parameter, the time delay kernel parameter, the state prediction result of the next cycle, and the multi-step rolling prediction result, and to train the initial world model using the loss function to obtain the world model; The causal graph and time-delay layered graph generation module 14 is used to determine, based on the world model, each causal edge between each variable in the time series data and the target state variable, and the target lag order of each causal edge, and to generate a causal graph and a time-delay layered graph using each causal edge; the target lag order is the lag order corresponding to the maximum time-delay weight of the causal edge; The deduction module 15 is used to replace the control action in the time series data with the target control action, and to deduce the state variables corresponding to the target control action based on the world model to obtain the state change curve. The time-delay causality report generation module 16 is used to generate a time-delay causality report for the process industry control system based on the causality graph, the time-delay hierarchical graph, the target lag order, and the state change curve, so as to control the process industry control system based on the time-delay causality report.

[0070] In some specific embodiments, the time delay unrolling module 11 may specifically include: The initial timing data acquisition module is used to acquire the initial timing data of the process industry control system. The data alignment module is used to align the initial time series data according to a preset sampling period to obtain time series data; the time series data includes state variables, control actions, and external operating conditions. The target state variable determination module is used to take the state variable of the next period of the current period in the time series data as the target state variable; The variable splicing module is used to splice the state variable, the control action, and the external operating condition to obtain the spliced ​​variable; The maximum time delay order determination module is used to determine the maximum time delay order by utilizing the residence time, heat transfer time, conveying distance, and the preset sampling period of the process industry control system. The time-delay expansion submodule is used to perform time-delay expansion on the spliced ​​variables based on the maximum time-delay order to obtain the historical tensor.

[0071] In some specific embodiments, the rolling prediction module 12 may specifically include: The edge gating weight determination module is used to determine the edge gating weights between each variable and each target state variable in the time series data based on the edge gating parameters. The normalization module is used to perform exponential normalization on the time delay kernel parameters at different lag orders to obtain normalized parameters; The time delay weight determination module is used to determine the time delay weights between each variable in the time series data and each target state variable under different lag orders by using the side gating weights and the normalization parameters; The parent variable aggregation quantity generation module is used to generate parent variable aggregation quantities for each target state variable based on the time delay weight; The action queue construction module is used to build action queues and action difference queues based on control actions in time-series data. The delayed action vector generation module is used to encode the action queue and the action differential queue to obtain the delayed action vector; The variable window building module is used to build variable windows based on external operating conditions in time series data; The mechanism vector generation module is used to encode the variables in the variable window to obtain the mechanism vector; The single-step state prediction module is used to perform single-step state prediction on the delayed action vector, the mechanism vector, and the parent variable aggregation using a structural equation network, so as to obtain the state prediction result for the next cycle. The multi-step rolling prediction module is used to perform multi-step rolling prediction on the state variables in the time series data, taking the state prediction result of the next cycle as the initial state, to obtain the multi-step rolling prediction result.

[0072] In some specific embodiments, the training module 13 may specifically include: The single-step prediction error determination module is used to determine the single-step prediction error based on the target state variable and the state prediction result of the next period. The multi-step rolling error determination module is used to determine the multi-step rolling error based on the multi-step rolling prediction result and the state variables in the time series data corresponding to the multi-step rolling prediction result; The graph sparsity constraint determination module is used to determine graph sparsity constraints based on edge gating parameters. The time delay constraint determination module is used to determine the time delay kernel smoothing constraint and the time delay peak constraint based on the time delay kernel parameters; The loss function construction module is used to construct loss functions including single-step prediction error, multi-step rolling error, graph sparsity constraint, time-delay kernel smoothing constraint, and time-delay peak constraint. The parameter update module is used to update the edge-gated parameters, time-delay kernel parameters, and model parameters in the initial world model using the loss function, so as to obtain the updated edge-gated parameters, updated time-delay kernel parameters, and world model.

[0073] In some specific embodiments, the cause-effect graph and time-delay hierarchical graph generation module 14 may specifically include: The edge gating weight calculation module is used to determine the edge gating weights between each variable in the time series data and the target state variable based on the updated edge gating parameters. The target edge gating weight determination module is used to take the weight that is greater than the preset edge gating weight threshold as the target edge gating weight; The causal edge determination module is used to identify the edge between the variable corresponding to the target edge gating weight in the time series data and the target state variable as a causal edge. The time delay weight calculation module is used to determine the time delay weight of each causal edge under different time delay orders based on the updated edge gating parameters and the updated time delay kernel parameters. The target lag order determination module is used to take the lag order corresponding to the maximum value of the time delay weight as the target lag order; The causal graph generation module is used to generate the causal graph using each of the causal edges; The time-delay layered graph generation module is used to display each of the causal edges in columns according to the target lag order, and generate a time-delay layered graph.

[0074] In some specific embodiments, the deduction module 15 may specifically include: The baseline trajectory generation module is used to input the state variables, control actions, and external conditions in the time series data into the world model for deduction and output the baseline trajectory. The state change curve generation module is used to fix the state variables and external operating conditions, replace one or more control actions in the time series data with target control actions, and deduce the state variables corresponding to the target control actions based on the world model to obtain the state change curve.

[0075] In some specific embodiments, the time-delay causality report generation module 16 may specifically include: The action influence determination module is used to determine the influence magnitude and direction of the target control action under different hysteresis orders based on the reference trajectory and the state change curve. The report generation module is used to generate a time-delay causal report for the process industry control system based on the causal graph, time-delay hierarchical graph, target lag order, state change curve, the magnitude of the influence, and the direction of the influence.

[0076] Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the time-delay causal report generation method for a process industrial control system executed by an electronic device as disclosed in any of the foregoing embodiments.

[0077] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0078] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0079] The operating system 221 manages and controls the various hardware devices on the electronic device 20 and the computer program 222 to enable the processor 21 to perform calculations and processing on the data 223 in the memory 22. It can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the time-delay causal report generation method of the process industry control system executed by the electronic device 20 as disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the time-delay causal report generation device of the process industry control system from external devices, and may also include data collected by its own input / output interface 25.

[0080] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0081] Furthermore, embodiments of this application also disclose a computer-readable storage medium storing a computer program. When the computer program is loaded and executed by a processor, it implements the steps of the time-delay causal report generation method for the process industrial control system disclosed in any of the foregoing embodiments.

[0082] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0083] The above provides a detailed description of the method, apparatus, equipment, and medium for generating time-delay causal reports for a process industry control system provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for time-lag causal reporting generation for a process industry control system, the method comprising: include: Acquire time-series data of the process industry control system, determine target state variables based on the time-series data, and perform time-delay expansion on the time-series data to obtain historical tensors; the target state variable is the state variable of the next cycle of the current cycle. Based on the historical tensor, side-gating parameters, and time-delay kernel parameters, a parent variable aggregation quantity is generated for each target state variable. The time series data is encoded to obtain a delayed action vector and a mechanism vector. Single-step state prediction is performed on the delayed action vector, the mechanism vector, and the parent variable aggregation quantity to obtain the state prediction result for the next cycle. The state prediction result for the next cycle is used as the initial state to perform multi-step rolling prediction on the state variables in the time series data to obtain multi-step rolling prediction results. Using the target state variable, the side gating parameter, the time delay kernel parameter, the state prediction result of the next cycle, and the multi-step rolling prediction result, a loss function is constructed, and the initial world model is trained using the loss function to obtain the world model; Based on the world model, each causal edge between each variable in the time series data and the target state variable, as well as the target lag order of each causal edge, are determined. A causal graph and a time-lag layered graph are generated using each causal edge. The target lag order is the lag order corresponding to the maximum time-lag weight of the causal edge. The control actions in the time series data are replaced with target control actions, and the state variables corresponding to the target control actions are deduced based on the world model to obtain state change curves; Based on the causal graph, the time delay hierarchical graph, the target lag order, and the state change curve, a time delay causal report for the process industry control system is generated so that the process industry control system can be controlled based on the time delay causal report.

2. The time-lag causal report generation method of a process industry control system according to claim 1, characterized by, The process involves acquiring time-series data from the process industry control system, determining target state variables based on the time-series data, and performing time-delay expansion on the time-series data to obtain a historical tensor, including: Acquire initial timing data for the process industry control system; The initial time series data is aligned according to a preset sampling period to obtain time series data; the time series data includes state variables, control actions, and external operating conditions. The state variable of the next cycle in the current cycle of the time series data is taken as the target state variable; The state variable, the control action, and the external operating condition are concatenated to obtain a concatenated variable; The maximum time delay order is determined by using the residence time, heat transfer time, conveying distance, and the preset sampling period of the process industry control system. Based on the maximum time delay order, the spliced ​​variables are expanded by time delay to obtain the historical tensor.

3. The time-lag causal report generation method of a process industry control system according to claim 1, characterized by, Based on the historical tensor, edge gating parameters, and time-delay kernel parameters, parent variable aggregates are generated for each target state variable, including: Based on the side-gating parameters, determine the side-gating weights between each variable and each target state variable in the time series data; At different lag orders, the time delay kernel parameters are exponentially normalized to obtain normalized parameters; Using the edge gating weights and the normalization parameters, the time delay weights between each variable in the time series data and each target state variable at different lag orders are determined; Based on the time delay weights, parent variable aggregates are generated for each target state variable.

4. The time-lag causal report generation method of a process industry control system according to claim 1, characterized by, The time-series data is encoded to obtain a delayed action vector and a mechanism vector. Single-step state prediction is then performed on the delayed action vector, the mechanism vector, and the aggregated parent variable to obtain the state prediction result for the next cycle, including: Based on the control actions in the time series data, construct an action queue and an action difference queue; The action queue and the action differential queue are encoded to obtain a delayed action vector; Construct a variable window based on external operating conditions in time series data; The variables in the variable window are encoded to obtain the mechanism vector; Using structural equation network, single-step state prediction is performed on the delayed action vector, the mechanism vector, and the aggregated parent variable to obtain the state prediction result for the next cycle.

5. The time-lag causal report generation method of a process industry control system according to claim 1, characterized by, The loss function is constructed using the target state variable, the side gating parameter, the time delay kernel parameter, the state prediction result of the next cycle, and the multi-step rolling prediction result, including: Based on the target state variable and the state prediction result of the next period, determine the single-step prediction error; Based on the multi-step rolling prediction results and the state variables in the time series data corresponding to the multi-step rolling prediction results, the multi-step rolling error is determined. Determine graph sparsity constraints based on edge gating parameters; Based on the time delay kernel parameters, determine the time delay kernel smoothing constraint and the time delay peak constraint; Construct a loss function that includes single-step prediction error, multi-step rolling error, graph sparsity constraint, time-delay kernel smoothing constraint, and time-delay peak constraint; The loss function is used to update the edge-gated parameters, time-delay kernel parameters, and model parameters in the initial world model, resulting in updated edge-gated parameters, updated time-delay kernel parameters, and the world model.

6. The time-lag causal report generation method of a process industry control system according to claim 5, characterized in that, Based on the world model, the process of determining each causal edge between each variable in the time series data and the target state variable, as well as the target lag order of each causal edge, and generating a causal graph and a time-lag layered graph using each causal edge includes: Based on the updated edge gating parameters, determine the edge gating weights between each variable in the time series data and the target state variable; The weights with edge gating weights greater than the preset edge gating weight threshold are used as the target edge gating weights; The edge between the variable corresponding to the target edge gating weight in the time series data and the target state variable is taken as a causal edge; Based on the updated edge gating parameters and the updated time delay kernel parameters, the time delay weights of each causal edge under different time delay orders are determined; The lag order corresponding to the maximum value of the time delay weight is taken as the target lag order; The causal graph is generated using each of the causal edges; Based on the target lag order, each causal edge is displayed in a separate column to generate a time-lag layered diagram.

7. The method for generating time-delay causality reports for process industry control systems according to any one of claims 1 to 6, characterized in that, The control actions in the time series data are replaced with target control actions, and the state variables corresponding to the target control actions are deduced based on the world model to obtain state change curves; Based on the causal graph, the time-delay hierarchical graph, the target lag order, and the state change curve, a time-delay causal report for the process industry control system is generated, including: The state variables, control actions, and external conditions in the time series data are input into the world model for simulation, and the baseline trajectory is output. The state variables and external conditions are fixed, and one or more control actions in the time series data are replaced with target control actions. Based on the world model, the state variables corresponding to the target control actions are deduced to obtain state change curves. Based on the baseline trajectory and the state change curve, the magnitude and direction of the impact of the target control action under different lag orders are determined; Based on the cause-effect graph, time delay hierarchical graph, target lag order, state change curve, the magnitude of the influence, and the direction of the influence, a time delay cause-effect report for the process industry control system is generated.

8. A time-delay causality report generation device for a process industry control system, characterized in that, include: The time-delay expansion module is used to acquire time-series data of the process industry control system, determine the target state variable based on the time-series data, and perform time-delay expansion on the time-series data to obtain the historical tensor; the target state variable is the state variable of the next cycle of the current cycle. The rolling prediction module is used to generate parent variable aggregation quantities for each target state variable based on the historical tensor, side gating parameters, and time delay kernel parameters; encode the time series data to obtain delayed action vectors and mechanism vectors; perform single-step state prediction on the delayed action vectors, the mechanism vectors, and the parent variable aggregation quantities to obtain the state prediction result for the next cycle; use the state prediction result of the next cycle as the initial state; and perform multi-step rolling prediction on the state variables in the time series data to obtain multi-step rolling prediction results. The training module is used to construct a loss function using the target state variable, the side gating parameter, the time delay kernel parameter, the state prediction result of the next cycle, and the multi-step rolling prediction result, and to train the initial world model using the loss function to obtain the world model. The causal graph and time-delay layered graph generation module is used to determine, based on the world model, each causal edge between each variable in the time series data and the target state variable, as well as the target lag order of each causal edge, and to generate a causal graph and a time-delay layered graph using each causal edge; the target lag order is the lag order corresponding to the maximum time-delay weight of the causal edge; The deduction module is used to replace the control actions in the time series data with the target control actions, and to deduce the state variables corresponding to the target control actions based on the world model to obtain the state change curves. The time-delay causality report generation module is used to generate a time-delay causality report for the process industry control system based on the causality graph, the time-delay hierarchical graph, the target lag order, and the state change curve, so as to control the process industry control system based on the time-delay causality report.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the time-delay causal report generation method for a process industry control system as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when the computer programs are executed by a processor, they implement the time-delay causal report generation method for a process industry control system as described in any one of claims 1 to 7.

Images