Continuous casting quality root cause self-healing method, system and program product based on causal atlas

Abstract

The invention provides a continuous casting quality root cause self-healing method and system based on a causal atlas and a program product. The method comprises the steps that an exception handling instruction is read, and a task scheduling result is output; vibration original data, defect image original data and process parameter original data are collected, features are cleaned and extracted, and a multi-modal feature sequence is output; performing Lagrange space-time alignment on the multi-modal feature sequence, and outputting a Lagrange feature tensor; constructing a target dynamic space-time causal graph based on the Lagrange feature tensor and screening a candidate root cause set; performing anti-fact verification and causal entropy sorting on the candidate root cause set based on the target dynamic space-time causal graph, and outputting a true root cause judgment result; generating a candidate self-healing strategy based on the true root cause judgment result; and performing digital twinborn verification on the candidate self-healing strategy, outputting and executing a target self-healing strategy, and collecting and executing a feedback data set and writing back.

Description

Technical Field

[0001] This invention relates to the field of intelligent quality control technology for continuous casting processes, and in particular to a method, system, and program product for self-healing of the root causes of continuous casting quality based on causal graphs. Background Technology

[0002] In continuous casting production, quality problems such as longitudinal cracks, slag entrainment, and segregation on the billet surface are usually related to a variety of factors, including casting speed, liquid level fluctuations, temperature changes, cooling water distribution, roll gap condition, equipment vibration, and the condition of the protective slag. Current on-site handling methods largely rely on manual experience, single-parameter alarms, or post-incident investigation. Typically, the cause of the anomaly is determined based on defect detection results and the shift's process records, followed by adjustments to the casting speed, cooling water volume, or protective slag scheme. Since the defect formation time is usually earlier than the exit detection time, directly using the current process state corresponding to the detection time for analysis can easily misjudge unrelated current environmental conditions as the causative environment, leading to inaccurate root cause localization.

[0003] Furthermore, vibration data, defect image data, and process parameter data from the continuous casting site exhibit differences in sampling frequency, temporal granularity, and spatial location. Existing analysis methods typically involve simple splicing or coarse alignment of multi-source data, followed by the use of empirical rules, statistical analysis, or conventional models to determine the cause of anomalies. While these methods can identify some correlations, they struggle to accurately reflect the true historical process environment during defect formation and to distinguish between the true root cause and associated factors. Therefore, this invention proposes a self-healing method, system, and program product for continuous casting quality root cause analysis based on causal mapping.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may contain prior art information that is not common knowledge to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a self-healing method, system, and program product for the root causes of continuous casting quality based on causal graphs, thereby solving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first part of this invention provides a self-healing method for the root causes of continuous casting quality based on causal graphs, comprising the following steps:

[0008] S1. Read the exception handling instructions, generate a task request record, perform task decomposition and resource adjudication on the task request record, and output the task scheduling result.

[0009] S2. Based on the task scheduling results, collect raw vibration data, raw defect image data, and raw process parameter data, clean and extract features, and output a multimodal feature sequence;

[0010] S3. Perform Lagrange spatiotemporal alignment on the multimodal feature sequence, trace back the defect generation time and motion segment, and output the Lagrange feature tensor;

[0011] S4. Construct a dynamic spatiotemporal causal graph of the target based on the Lagrange feature tensor, and filter the candidate root cause set;

[0012] S5. Based on the target dynamic spatiotemporal causal graph, perform counterfactual verification and causal entropy sorting on the candidate root cause set, output the true root cause determination result, and generate candidate self-healing strategies based on the true root cause determination result.

[0013] S6. Perform digital twin verification on the candidate self-healing strategy, output the target self-healing strategy and execute it, collect the execution feedback dataset and write it back.

[0014] S1 specifically includes: reading anomaly handling instructions, flow number identifiers, steel type identifiers, defect type identifiers, target quality indicators, and production plan identifiers; generating task request records and deleting task request records with missing or duplicate fields; outputting a set of valid task requests; inputting the set of valid task requests into the metacognitive coordination agent; combining the task parsing knowledge graph to perform semantic parsing, object extraction, and task decomposition; generating task orchestration records and deleting task orchestration records with broken dependency chains; outputting a set of valid task orchestrations; generating a set of candidate scheduling schemes based on the set of valid task orchestrations and the current computing power, planned production value, and energy consumption budget value; calculating the total system utility value; deleting candidate scheduling scheme records below the preset lower limit value; and outputting the task scheduling result.

[0015] S2 specifically includes: reading the original vibration data, original defect image data, and original process parameter data corresponding to the target furnace number, target flow number, and target time window based on the task scheduling results; performing time base verification and basic cleaning; deleting records with incomplete time coverage, consecutive missing points reaching a preset threshold, exceeding the equipment range, or abnormal image size; and outputting a preprocessed observation dataset; extracting equipment health status features, defect image features, and process status features from the preprocessed observation dataset; outputting a multimodal feature sequence; performing Lagrange spatiotemporal alignment based on the multimodal feature sequence; calculating the defect generation time; reconstructing the material slice motion trajectory segment; aligning various features according to the segment and time step; and outputting a Lagrange feature tensor.

[0016] S3 specifically includes: reading the Lagrange feature tensor, performing variable mapping according to the preset variable dictionary, generating a system state vector sequence, an observation state slice set, and a causal modeling sample set, and deleting records with inconsistent variable dimensions or missing current defect risk value fields; inputting the effective causal modeling sample set into the spatiotemporal variational causal discovery engine, constructing a spatiotemporal structural equation model, and outputting an initial dynamic spatiotemporal causal graph and an effective initial edge weight set; performing continuous constraint optimization on the initial dynamic spatiotemporal causal graph and the effective initial edge weight set to obtain the target dynamic spatiotemporal causal graph, and filtering ancestor nodes that have path connectivity with the target result variable, and outputting a candidate root cause set.

[0017] S4 specifically includes: reading the target dynamic spatiotemporal causal graph, candidate root cause set, and effective observation state slice set; constructing the fact world network and counterfactual world network; inferring the posterior distribution of exogenous noise; deleting records with noise variance exceeding a preset upper limit; and outputting the effective exogenous noise inference set; performing counterfactual intervention on candidate variables based on the effective exogenous noise inference set; calculating pure causal effects; deleting candidate variables with pure causal effects less than a preset threshold; and outputting the true root cause candidate set; fusing pure causal effects, defect classification confidence, equipment anomaly degree, and mechanism consistency identifier; calculating and sorting the holographic causal entropy; and outputting the true root cause determination result.

[0018] S5 specifically includes: reading the root cause determination results, the target dynamic spatiotemporal causal graph, and the effective observation state slice set; generating a reinforcement learning state set and a constrained action set; constructing an action mask based on the ancestor node set of the target outcome variable; deleting action records with a mask value of 0; and outputting the effective action set. Based on the reinforcement learning state set and the effective action set, performing causal constraint reinforcement learning search, calculating the immediate reward value of each action combination, and deleting action combinations that exceed the preset action boundary; and outputting the policy scoring result. Reading the policy scoring result, selecting the top k action combinations with the highest immediate reward values ​​as candidate self-healing policies, deleting candidate self-healing policy records with duplicate action combinations, and outputting the candidate self-healing policy set.

[0019] S6 specifically includes: reading candidate self-healing strategy records with a verification identifier field of 1 from the candidate self-healing strategy set, and simultaneously reading the steel grade process library record corresponding to the steel grade field; using a digital twin verification agent to perform virtual trial and error on the candidate self-healing strategies, and outputting digital twin prediction result records; performing physical safety boundary verification on the digital twin prediction result records; generating rejection feedback records and sending back candidate self-healing strategies with a safety boundary judgment field of "not passed"; generating target self-healing strategy records for candidate self-healing strategies with a safety boundary judgment field of "passed"; outputting a target self-healing strategy set; reading the target self-healing strategy set and sending it to the programmable logic controller for execution, generating strategy execution records, collecting execution feedback data to form an execution feedback dataset, and writing the complete execution feedback dataset back to the historical time series database.

[0020] The second part of this invention provides a self-healing system for the root causes of continuous casting quality based on causal graphs, comprising:

[0021] The task scheduling module is used to read exception handling instructions, generate task request records, decompose tasks and make resource decisions on the task request records, and output task scheduling results.

[0022] The data perception module collects raw vibration data, raw defect image data, and raw process parameter data based on task scheduling results. It then cleans and extracts features from these raw vibration data, raw defect image data, and raw process parameter data, and outputs a multimodal feature sequence.

[0023] The spatiotemporal alignment module is used to perform Lagrange spatiotemporal alignment on multimodal feature sequences, trace back the defect generation time and material slice motion segment, and output the Lagrange feature tensor;

[0024] The causal modeling module constructs a dynamic spatiotemporal causal graph of the target based on the Lagrange feature tensor and filters the set of candidate root causes.

[0025] The root cause determination module performs counterfactual verification and causal entropy sorting on the candidate root cause set based on the target dynamic spatiotemporal causal graph, and outputs the true root cause determination result.

[0026] The strategy generation module generates candidate self-healing strategies based on the root cause determination results.

[0027] The verification and execution module is used to perform digital twin verification of candidate self-healing strategies, output the target self-healing strategy and execute it, collect the execution feedback dataset and write it back.

[0028] The third part of this invention provides a computer program product, which includes computer program code. When the computer program code is run on a computer, the computer implements the above-described self-healing method for the root causes of continuous casting quality based on causal graphs.

[0029] The beneficial effects of this invention are as follows:

[0030] This invention achieves unified collection, cleaning, feature extraction, and Lagrange spatiotemporal alignment of vibration data, defect image data, and process parameter data, thus correlating defect detection results with the actual historical process environment at the time of defect formation. This solves the problem of inaccurate root cause localization caused by the misalignment between the detection time and the defect formation time in the prior art, thereby improving the accuracy of continuous casting quality anomaly diagnosis.

[0031] This invention constructs a target dynamic spatiotemporal causal graph and combines it with a candidate root cause screening mechanism to jointly analyze the instantaneous and time-delayed influence relationships between process variables, equipment status, and quality results. This enables the identification of candidate variables with genuine causal relationships to the target result variables under multi-factor coupling conditions, thereby reducing the risk of misjudgment based solely on correlation analysis. By constructing a factual world network and a counterfactual world network, counterfactual intervention is performed on candidate variables. Combined with pure causal effect calculation and holographic causal entropy ranking, the true root cause is further determined, distinguishing between the true root cause and associated related factors, thus improving the credibility and interpretability of the root cause identification results.

[0032] This invention generates a set of reinforcement learning states and a set of constrained actions based on the results of true cause determination. It then uses an action mask to limit the search range, performing policy searches only within the adjustable parameter space related to the target outcome variable. This reduces invalid searches and blind trial-and-error, improving the efficiency of self-healing policy generation. Before the formal execution of candidate self-healing policies, a digital twin verification agent is introduced to physically verify the safety boundaries of the candidate policies. Policies that fail are returned for re-search, while verified policies are then issued for execution. This reduces on-site trial-and-error costs, minimizes the introduction of new quality risks, and achieves closed-loop self-healing handling of continuous casting quality anomalies. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the self-healing method for the root causes of continuous casting quality based on causal graphs according to the present invention.

[0034] Figure 2 This is a schematic diagram of the self-healing system framework for continuous casting quality based on causal graphs, as proposed in this invention. Detailed Implementation

[0035] 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.

[0036] Example 1: As Figure 1 As shown, this embodiment provides a self-healing method for the root causes of continuous casting quality based on causal graphs, including the following steps:

[0037] S1. Read the exception handling instructions, generate a task request record, perform task decomposition and resource adjudication on the task request record, and output the task scheduling result.

[0038] S2. Based on the task scheduling results, collect raw vibration data, raw defect image data, and raw process parameter data, clean and extract features, and output a multimodal feature sequence;

[0039] S3. Perform Lagrange spatiotemporal alignment on the multimodal feature sequence, trace back the defect generation time and motion segment, and output the Lagrange feature tensor;

[0040] S4. Construct a dynamic spatiotemporal causal graph of the target based on the Lagrange feature tensor, and filter the candidate root cause set;

[0041] S5. Based on the target dynamic spatiotemporal causal graph, perform counterfactual verification and causal entropy sorting on the candidate root cause set, output the true root cause determination result, and generate candidate self-healing strategies based on the true root cause determination result.

[0042] S6. Perform digital twin verification on the candidate self-healing strategy, output the target self-healing strategy and execute it, collect the execution feedback dataset and write it back.

[0043] S1 specifically includes the following sub-steps:

[0044] S110. Read the abnormal handling instructions input by the operator, and simultaneously read the current flow number identifier, steel type identifier, defect type identifier, target quality index, and current production plan identifier to generate a task request record. The task request record is used to convert unstructured language input into a structured object that can be called by subsequent steps; therefore, the record name, field order, and field type are fixed when it is generated.

[0045] The task request record includes at least the task number field, the flow number field, the steel grade field, the defect type field, the target quality index field, the instruction timestamp field, and the production constraint field. Among them, the task number field is used to uniquely identify the current abnormal handling task. In one embodiment, it is generated by concatenating "furnace number-flow number-timestamp", such as "Q345B-3-143000". The production constraint field includes at least the lower limit of production capacity, the upper limit of energy consumption, and the equipment safety status indicator.

[0046] Perform integrity verification on task request records. The verification rules are as follows: the serial number field cannot be empty, the steel type field cannot be empty, the defect type field cannot be empty, the target quality index field must fall within the preset quality index dictionary, and the instruction timestamp field must meet the standard time format. If any field is missing or the format is invalid, the corresponding task request record will be written into the invalid task set and deleted.

[0047] For task request records that pass the integrity verification, a duplicate verification is performed. If multiple task request records with instruction timestamp intervals less than 30 seconds exist under the same batch, serial number, and defect type, only the task request record with the latest instruction timestamp is retained, and the remaining task request records are added to the duplicate task set and then deleted. After verification and deletion, a valid task request set is output, and each record in the valid task request set serves as input for step S120.

[0048] S120. Read the set of valid task requests and input each valid task request record into the task semantic parsing module in the metacognitive coordination agent. The task semantic parsing module, in conjunction with the task parsing knowledge graph, performs semantic parsing, object extraction, and task decomposition on the anomaly handling instructions to generate a task orchestration record. The task parsing knowledge graph is used to provide semantic mapping relationships between defect types, equipment objects, process variables, and handling actions to ensure that the task decomposition results are consistent with the on-site objects.

[0049] The task parsing knowledge graph is pre-built by performing named entity recognition and relation extraction on historical continuous casting operation procedure texts and expert experience bases. Its nodes include defect entities, equipment entities, and parameter entities, and its edges contain control, association, and constraint relationships.

[0050] The task orchestration record adopts a hierarchical structure of "one main task corresponding to multiple sub-tasks". The sub-tasks include at least data perception tasks, causal discovery tasks, counterfactual verification tasks, process optimization tasks, and digital twin verification tasks.

[0051] The task orchestration record includes at least the following fields: main task number, subtask number, target agent identifier, input object, preceding subtask, output object, priority, and time limit.

[0052] Dependency verification is performed on the task orchestration records. The verification rules are as follows: each counterfactual verification task must correspond to at least one causal discovery task as a prerequisite subtask; each process optimization task must correspond to at least one counterfactual verification task as a prerequisite subtask; and each digital twin verification task must correspond to at least one process optimization task as a prerequisite subtask. If the prerequisite subtask field is empty or the dependency chain is broken, the corresponding task orchestration record is written to the broken chain task set and deleted. After verification and deletion, a valid task orchestration set is output, and each record in the valid task orchestration set is used as input for step S130.

[0053] S130. Read the effective task scheduling set, and simultaneously read the currently available GPU computing power, edge node computing power, current production plan value, and current media energy consumption budget value to generate a candidate scheduling scheme set. Each candidate scheduling scheme record includes at least the following fields: task execution order field, target computing power allocation field, estimated execution time field, conflict object field, and conflict resolution field. The candidate scheduling scheme set describes how each subtask is allocated to the corresponding agent for execution under the current production line resource conditions. Calculate the total system utility value for each candidate scheduling scheme in the set:

[0054] ;

[0055] In the formula, The total utility value of the system is represented by α; the quality weighting coefficient is represented by α. β represents the quality benefit term; β represents the capacity weighting coefficient. Represents the capacity benefit item; γ represents the energy consumption weighting coefficient; This represents the energy consumption cost item; α+β+γ=1. During calculation, the quality benefit item is normalized according to the expected reduction in defective billet quantity, the capacity benefit item is normalized according to the billet output per unit time, and the energy consumption cost item is normalized according to the converted costs of water, electricity, and gas.

[0056] In one embodiment, when the object to be dealt with is a longitudinal crack defect, α can be set to 0.6, β to 0.3, and γ to 0.1 to increase the weight of the quality objective in the scheduling decision. Candidate scheduling scheme records with a total system utility value less than a preset lower limit are directly deleted. The scheduling scheme with the largest total system utility value is selected from the remaining candidate scheduling scheme records as the task scheduling result output, and the task scheduling result is used as the input for step S210.

[0057] S2 specifically includes the following sub-steps:

[0058] S210. Read the task scheduling results and extract the raw data corresponding to the target furnace number, target flow number, and target time window from the high-frequency vibration monitoring array, machine vision inspection system, and L1 / L2 (basic level / process level) basic automation system according to the task scheduling results, respectively, to generate a multimodal raw dataset. The multimodal raw dataset includes at least vibration raw data records, defect image raw data records, and process parameter raw data records;

[0059] Among them, the vibration raw data record includes at least the sampling point number field, sampling time field, sensor number field and vibration amplitude field; the defect image raw data record includes at least the image number field, acquisition time field, camera number field, defect image matrix field and defect physical coordinate field; and the process parameter raw data record includes at least the acquisition time field, pulling speed field, liquid level field, temperature field, cooling water flow rate field and roll gap value field.

[0060] Time baseline verification and basic cleaning are performed on the multimodal raw dataset. The verification rules are as follows: the sampling time of the vibration raw data record must cover the detection time window corresponding to the defect image raw data record, and the acquisition time of the process parameter raw data record must cover the historical time window required for defect backtracking; if the time coverage of any mode is incomplete, the record corresponding to that mode is written into the time missing set and deleted. The deletion rules are as follows: process parameter raw data records with 5 or more consecutive missing points are directly deleted, and process parameter raw data records with less than 5 consecutive missing points are compensated by linear interpolation; data points exceeding the upper or lower limit of the equipment range are directly deleted; defect image raw data records with damaged image matrix or image size inconsistent with camera calibration size are directly deleted. After verification and deletion, a preprocessed observation dataset is output, which serves as the input for step S220.

[0061] S220. Read the preprocessed observation dataset, perform adaptive wavelet packet decomposition on the original vibration data records, extract energy values ​​and energy entropy in different frequency bands, and generate equipment health status feature records. The equipment health status feature records include at least the sensor number field, frequency band number field, energy value field, energy entropy field, equipment anomaly degree field, and feature time window field; among them, the equipment anomaly degree field is calculated from the energy entropy field and the frequency band energy distribution field, and is used to characterize the degree to which the current equipment state deviates from the normal state. Writing the equipment anomaly degree field into the equipment health status feature records ensures that a unified field can be directly called during subsequent multimodal fusion, without having to rename it temporarily.

[0062] Read the original data records of defect images from the preprocessed observation dataset, and use a pre-trained convolutional neural network (such as ResNet or VGG backbone network) to extract high-dimensional feature vectors of surface defects of the billet, and generate defect image feature records.

[0063] The defect image feature record includes at least an image number field, a defect classification confidence field, a feature vector field, a vertical coordinate field, and a horizontal coordinate field; wherein, the defect classification confidence field is used to indicate the probability that the current defect image belongs to the target defect category. The original process parameter data records in the preprocessed observation dataset are read and resampled at a uniform time step to generate process state feature records; the process state feature records include at least a resampling time field, a pulling speed field, a liquid level field, a temperature field, a cooling water flow rate field, and a roll gap value field. Time alignment and number association are performed on the above three types of feature records, and a multimodal feature sequence is output, which serves as the input for step S230.

[0064] S230. Read the multimodal feature sequence and extract the vertical coordinate field and acquisition time field from the defect image feature record. Calculate the generation time of the defect slice at the meniscus of the crystallizer using the Lagrange spatiotemporal alignment mechanism to generate a defect backtracking record. The defect backtracking record includes at least the defect number field, vertical coordinate field, detection time field, generation time field, and backtracking confidence field. The generation time field is obtained by solving the following formula:

[0065] ;

[0066] In the formula, Indicates the longitudinal location of the defect on the billet; This indicates the initial time when the defect slice was generated inside the crystallizer; This indicates the absolute time at which the machine vision system detected the defect; Representing time variables Corresponding real-time pulling speed; This represents the integral variable. During calculation, the pulling speed field in the process state characteristic record is constructed into a piecewise constant function and the numerical inverse calculation is performed; in one embodiment, when the longitudinal position of the defect is 15m, the detection time is 14:30:00, and the average pulling speed is 1.0m / min, the defect generation time is obtained by inverse calculation as approximately 14:15:00.

[0067] Read the generation time field from the defect backtracking record, perform segment reconstruction on the motion trajectory of the material slice, and generate trajectory segment records. The trajectory segment records include at least a segment number field, a segment entry time field, a segment exit time field, and a segment identifier field; the segment identifier field includes at least the identifiers for the crystallizer area, foot roller area, and secondary cooling area. Perform time coverage verification on the trajectory segment records; if any segment's entry time or exit time field is missing, delete the corresponding defect backtracking record. Subsequently, using the trajectory segment records as indexes, extract process parameter fragments within the corresponding time interval from the process status feature records, and align the equipment health status feature records, defect image feature records, and process status feature records by segment and time step, outputting a Lagrange feature tensor. The Lagrange feature tensor includes at least a defect number dimension, a segment identifier dimension, a time step dimension, and a feature dimension. The feature dimension includes at least the pulling speed, liquid level, temperature, cooling water flow rate, roll gap value, equipment anomaly degree, and defect image feature vector. The Lagrange feature tensor serves as the input for step S310.

[0068] S3 specifically includes the following sub-steps:

[0069] S310. Read the Lagrange feature tensor and perform variable mapping on each field in the feature dimension according to the preset variable dictionary to generate a system state vector sequence and an observation state slice set. The variable dictionary is used to ensure that the object names in subsequent causal modeling, counterfactual inference, and policy optimization are completely consistent. Therefore, each field is mapped to a unique variable name. For example, the pulling speed field is mapped to the pulling speed variable, the liquid level field is mapped to the liquid level variable, the cooling water flow rate field is mapped to the cooling water flow rate variable, the roll gap value field is mapped to the roll gap value variable, the equipment anomaly degree field is mapped to the equipment anomaly degree variable, and the defect classification confidence field is mapped to the defect characterization variable. Further, the target quality index field is mapped to the target outcome variable Y, and the value of the target outcome variable Y under the current abnormal scenario is written into the current defect risk value field. The current defect risk value field serves as a recorded expression of the target outcome variable Y and is used for subsequent counterfactual verification and policy optimization.

[0070] The system state vector sequence is segmented to generate a causal modeling sample set. This set includes at least the current state vector field, the next state vector field, the historical lag state vector field, and the sample weight field. The observed state slice set includes at least the slice number field, the current state vector field, the current defect risk value field, and the scenario identifier field, providing a unified input source for subsequent counterfactual inference and strategy optimization. The verification rules are as follows: the current state vector field, the next state vector field, and the historical lag state vector field must be completely consistent in terms of variable dimensions, and the current defect risk value field cannot be empty; if the variable dimensions are inconsistent or the current defect risk value field is missing, the corresponding causal modeling sample or observed state slice is deleted. After verification and deletion, a valid causal modeling sample set and a valid observed state slice set are output. The valid causal modeling sample set serves as the input for step S320, and the valid observed state slice set serves as the input source for steps S410 and S510.

[0071] S320. Read the set of valid causal modeling samples and input them into the spatiotemporal variational causal discovery engine to construct a spatiotemporal structural equation model, generating an initial dynamic spatiotemporal causal graph and initial edge weight records. The spatiotemporal structural equation model is used to learn the underlying dynamic causal structure from observational data, and its mathematical expression is:

[0072] ;

[0073] In the formula, This represents the system state vector at time t+1; This represents the system state vector at time t; This represents the system state vector at a time d steps behind. The parameter is The nonlinear function; A represents the instantaneous causal adjacency matrix, used to characterize the strength of the causal relationship between variables at the same moment; W represents the lagged causal adjacency matrix, used to characterize the causal influence of historical states on the current state; d represents the Hadamard product, which is the element-wise multiplication of matrices; d represents the time lag step. Let represent the exogenous noise vector at time t.

[0074] Initial edge filtering is performed on the output of the spatiotemporal structure equation model to generate initial edge weight records. Each initial edge weight record includes at least a source variable field, a target variable field, an edge type field, an edge weight value field, and an occurrence round field; the edge type field distinguishes between instantaneous causal edges and lagged causal edges. The deletion rules are as follows: initial edge weight records with an absolute edge weight value less than 0.05 are deleted, and initial edge weight records with an occurrence round less than 3 are deleted. After filtering, an initial dynamic spatiotemporal causal graph and a set of valid initial edge weights are output, serving as inputs to step S330.

[0075] S330. Read the initial dynamic spatiotemporal causal graph and the effective initial edge weight set. Perform continuous constraint optimization on the instantaneous causal adjacency matrix and the lag causal adjacency matrix, and output the target dynamic spatiotemporal causal graph and the candidate root cause set. The continuous constraint optimization uses the augmented Lagrange multiplier method, and its optimization objective function is:

[0076] ;

[0077] In the formula, This represents the lower bound loss of evidence and is used to measure the model's ability to reconstruct observed data. This represents the sparsity penalty coefficient; Denotes the norm of the instantaneous causal adjacency matrix; This represents the norm of the lagged causal adjacency matrix; Network parameters representing nonlinear functions; Represents the Lagrange multipliers; Let h(A) denote the acyclic constraint function, and h(A) = 0 if and only if A is a directed acyclic graph; This represents the coefficient of the quadratic penalty term. During calculation, it is first fixed... and Update A, W and Then update based on the current value of h(A) and This process continues until h(A) converges below a preset threshold. The reason for using this optimization process is that counterfactual reasoning requires the causal graph to satisfy a clear parent-child relationship. If the acyclic constraint is not satisfied, the subsequent intervention path will not be unique.

[0078] Read the optimized target dynamic spatiotemporal causal graph, and perform candidate root cause filtering on ancestor nodes that have path connectivity with the target outcome variable Y, generating a candidate root cause set. The candidate root cause set includes at least the candidate variable field, the path length field to the target outcome variable Y, the edge weight summary field, and the occurrence stability field. The verification rule is: only retain ancestor nodes with an edge weight summary value greater than 0.15 that appear in all three consecutive optimization rounds; candidate variables that do not meet this rule are deleted. After filtering and deletion, output the target dynamic spatiotemporal causal graph and the candidate root cause set, where the target dynamic spatiotemporal causal graph, the candidate root cause set, and the effective observation state slice set are used as inputs for step S410.

[0079] S4 specifically includes the following sub-steps:

[0080] S410. Read the target dynamic spatiotemporal causal graph, candidate root cause set, and effective observation state slice set, and construct the fact-world network and counterfactual-world network to generate exogenous noise inference results. The fact-world network and counterfactual-world network adopt a shared parameter structure; the fact-world network is used to recover the background environmental conditions based on the observation state vector corresponding to the current abnormal scenario, and the counterfactual-world network is used to change only a single candidate variable and calculate the result change while keeping the background environmental conditions unchanged. To ensure the consistency of the source of subsequent inference objects, a fact-reasoning record is first generated for each candidate root cause variable. The fact-reasoning record includes at least a candidate variable field, an observation state vector field, a current defect risk value field, a scene identifier field, and a fact-world network input field; the fact-world network input field is directly assigned by the current moment state vector field in the effective observation state slice set.

[0081] The fact-world network input field is read from the fact-reasoning record, and the variational autoencoder is based on the observed state vector. Inferring the posterior distribution of exogenous noise The output is the inference result of exogenous noise. Where, Represents the observation state vector corresponding to the current abnormal scenario; U represents the exogenous noise variable; This represents the conditional probability distribution function. The exogenous noise inference result includes at least the candidate variable field, the noise mean field, the noise variance field, and the scene identifier field. The verification rule is: if the noise variance field is greater than a preset upper limit, the current background environment is determined to be unstable, and the corresponding exogenous noise inference result is deleted. After verification and deletion, a valid exogenous noise inference set is output, which serves as the input for step S420.

[0082] S420. Read the effective exogenous noise inference set and perform counterfactual intervention on each candidate variable in the candidate root cause set to generate a pure causal effect record. The counterfactual intervention is performed as follows: keep the exogenous noise variable U unchanged in the counterfactual world network, disconnect the candidate variable from its parent node, and force the candidate variable to be assigned a value within the normal process range. Then, calculate the difference in predicted values ​​of the target outcome variable Y under the factual scenario and the counterfactual scenario respectively. The pure causal effect record includes at least the candidate variable field, factual value field, normal value field, factual scenario prediction result field, counterfactual scenario prediction result field, and pure causal effect field.

[0083] The pure causal effect field is calculated using the following formula:

[0084] ;

[0085] In the formula, Representing variables The pure causal effect value on the target outcome variable Y; This represents the i-th candidate root dependent variable; Represents the mathematical expectation operator; Indicates causal interference; Representing variables Values ​​within the normal process range; Representing variables The actual abnormal values ​​are recorded. A threshold check is performed on the pure causal effect records. If the pure causal effect field is less than 0.2, the corresponding candidate variable is determined to be a non-true cause and deleted; if the pure causal effect field is greater than or equal to 0.2, it is retained as a true cause candidate variable. In one embodiment, if the improvement of the target outcome variable Y is approximately 10% after intervention on the pulling speed variable, then this variable is retained as a low-priority candidate; if the improvement of the target outcome variable Y is approximately 90% after intervention on the protective slag-related variables, then this variable is retained as a high-priority candidate. After verification and deletion, a true cause candidate set is output, which serves as the input for step S430.

[0086] S430. Read the candidate set of root causes and simultaneously read the pure causal effect field, defect classification confidence field, and equipment anomaly field corresponding to the candidate variables to generate a multimodal causal evidence record. The defect classification confidence field is derived from the defect image feature record output in step S220, and the equipment anomaly field is derived from the equipment health status feature record output in step S220. Therefore, no new synonymous field names are introduced in this step. The multimodal causal evidence record includes at least the candidate variable field, pure causal effect field, defect classification confidence field, equipment anomaly field, and mechanism consistency identifier field. The mechanism consistency identifier field is obtained by performing an edge-level comparison between the target dynamic spatiotemporal causal graph and the metallurgical mechanism knowledge graph, and is used to determine whether the data-driven results conform to the physical laws of metallurgy.

[0087] Read the multimodal causal evidence records and use the holographic causal entropy formula to calculate the causal certainty of each candidate variable, generating a causal entropy ranking record:

[0088] ;

[0089] In the formula, Let represent the holographic causal entropy of candidate root cause c; p(c) represent the posterior probability distribution of candidate root cause c; λ represents the knowledge constraint weight coefficient. Represents the relative entropy function; Represents the dynamic spatiotemporal causal graph of the target; This represents a knowledge graph of metallurgical mechanisms. During calculation, p(c) is first obtained by normalizing the pure causal effect field, defect classification confidence field, and equipment anomaly field. Then, the relative entropy term is calculated based on the difference between the target dynamic spatiotemporal causal graph and the knowledge graph of metallurgical mechanisms. The holographic causal entropy values ​​are sorted from smallest to largest. If the difference between the first and second ranked holographic causal entropies is greater than or equal to 0.05, only the first ranked candidate variable is output as the true root cause; if the difference is less than 0.05, both the primary root cause and the secondary root cause are output. After sorting, the true root cause determination result is output. The true root cause determination result and the set of effective observation state slices are used as input for step S510.

[0090] S5 specifically includes the following sub-steps:

[0091] S510: Read the root cause determination result, the target dynamic spatiotemporal causal graph, and the set of effective observation state slices. Map the observation state vector corresponding to the current abnormal scenario to the node space of the causal graph to generate a reinforcement learning state set and a set of constrained actions. The reinforcement learning state set includes at least the current pulling speed field, liquid level fluctuation field, heat flux density fluctuation field, cooling water flow rate field for each zone, roll gap value field, equipment anomaly degree field, and current defect risk value field; among which, the current defect risk value field comes from the set of effective observation state slices output in step S310. The set of constrained actions includes at least the pulling speed adjustment action, cooling water flow rate adjustment action, roll gap compensation action, and protective slag switching suggestion action.

[0092] Read the target's dynamic spatiotemporal causal graph and generate an action mask based on the set of ancestor nodes of the target's outcome variable Y. Output the action mask record:

[0093] ;

[0094] In the formula, This represents the action masking value corresponding to the j-th process variable; Let represent the j-th process variable; Anc(Y) represents the set of ancestor nodes of the target outcome variable Y. When, the corresponding process variables can be adjusted; when At this time, adjusting the corresponding process variables is prohibited. The deletion rule is: all corresponding... The action records are directly deleted from the constrained action set. After deletion, the valid action set is output, and the valid action set is used as the input for step S520.

[0095] S520: Read the reinforcement learning state set, the effective action set, and the rejection feedback record returned in step S620. If no rejection feedback record exists, directly use the effective action set as the current search space. If a rejection feedback record exists, update the action boundary according to the feedback constraint field in the rejection feedback record, and then use the updated effective action set as the current search space. The rejection feedback record includes at least a policy number field, a rejection reason field, a boundary violation field, and a feedback constraint field. Input the reinforcement learning state set and the current search space into the process optimization agent, perform causal constraint reinforcement learning search, and generate a policy score record. The policy score record includes at least a policy number field, an action combination field, a predicted defect rate decrease value field, a cost increase value field, a potential risk value field, and an immediate reward value field.

[0096] Calculate the instant reward value for each action combination:

[0097] ;

[0098] In the formula, Represents the instantaneous reward value at time t; This represents the quality benefit weighting coefficient; Indicates the amount of quality improvement; This represents the cost penalty weighting coefficient; Indicates operating costs; This represents the risk penalty weighting coefficient; This indicates potential risk penalties. The verification rules are: the speed adjustment range must not exceed 15% of the current speed, the cooling water flow rate adjustment range must not exceed 10% of the current section's rated flow rate, and the roll gap compensation range must not exceed 0.3mm; combinations of actions exceeding these boundaries are directly deleted. In one embodiment, the single-step increase / decrease of the speed adjustment action can be set to 0.05 m / min, the single-step increase / decrease of the secondary cooling water flow rate in a certain section can be set to 3% of the rated flow rate, and the single-step increase / decrease of the roll gap compensation action can be set to 0.1mm. After verification, deletion, and sorting, the strategy score result is output, and the strategy score result serves as the input for step S530.

[0099] S530. Read the strategy scoring results and select the top k action combinations with the largest immediate reward values ​​as candidate self-healing strategies, where k represents the number of candidate strategies to retain, which is 3 in one embodiment. Generate a candidate self-healing strategy record. The candidate self-healing strategy record includes at least the following fields: strategy number field, steel grade field, target casting speed adjustment value field, target cooling water allocation value field, target roll gap compensation value field, protective slag switching suggestion field, predicted defect rate reduction value field, potential risk value field, and verification identifier field. The verification identifier field is always written as "1" to indicate that the candidate self-healing strategy has not yet entered the digital twin verification process.

[0100] Perform duplicate policy deletion and priority sorting on the candidate self-healing policy records. If the action combination fields of two candidate self-healing policy records are exactly the same, only the one with the higher immediate reward value is retained, and the rest of the records are deleted. After deletion, sort the candidate self-healing policies from high to low according to their immediate reward values, and output the candidate self-healing policy set. The records in the candidate self-healing policy set with the verification identifier field of "1" are used as the input for step S610.

[0101] S6 specifically includes the following sub-steps:

[0102] S610. Read the candidate self-healing strategy records with the verification identifier field set to "1" in the candidate self-healing strategy set, and simultaneously read the steel grade process library records corresponding to the steel grade field. Input the candidate self-healing strategy records one by one into the digital twin verification agent, call the physical field proxy model of the continuous casting solidification process to perform virtual trial and error, and generate digital twin prediction result records.

[0103] The physical field proxy model for the continuous casting solidification process is a reduced-order neural network model (or: a numerical calculation model based on the second-order partial differential heat transfer equation) pre-trained based on historical continuous casting finite element thermo-mechanical coupling simulation data. It is used to quickly solve the temperature field and stress field distribution according to the input process conditions.

[0104] The steel grade process library records include at least the brittle temperature range field, stress upper limit field, and thermophysical property parameter field, which are used to provide steel grade-related physical boundaries for digital twin verification. The digital twin prediction result records include at least the strategy number field, corner temperature prediction value field, center temperature prediction value field, surface stress prediction value field, brittle zone identifier field, and safety boundary determination field.

[0105] Physical safety boundary verification is performed on the digital twin prediction results. The verification rules are as follows: when the corner temperature prediction value falls within the brittle temperature range field in the steel grade process library record, or the surface stress prediction value exceeds the stress upper limit field, the brittle zone identifier field is set to "1", and the safety boundary judgment field is set to "fail"; when neither of the above two conditions is met, the brittle zone identifier field is set to "0", and the safety boundary judgment field is set to "pass". In one embodiment, for Q345B steel grade, the brittle temperature range parameter and stress upper limit parameter in its corresponding steel grade process library record are directly called. After verification, the set of determined digital twin results is output, and the set of determined digital twin results is used as the input for step S620.

[0106] S620: Read the set of determined digital twin results, generate a rejection feedback record for candidate self-healing strategies whose safety boundary determination field is "not passed", and send the rejection feedback record back to step S520. The rejection feedback record includes at least a strategy number field, a rejection reason field, a boundary violation field, and a feedback constraint field; among them, the boundary violation field is used to indicate whether the rejection is triggered by the corner temperature boundary or the surface stress boundary, and the feedback constraint field is used to constrain the search space for subsequent actions. For example, when a candidate self-healing strategy causes an increased risk of freezing due to an excessive decrease in pulling speed, the feedback constraint field is written with "the lower limit of pulling speed adjustment is increased to 95% of the current pulling speed"; when a candidate self-healing strategy causes the corner temperature to enter the brittle zone due to excessive cooling water in a certain area, the feedback constraint field is written with "the upper limit of water flow adjustment in the corresponding section is reduced by 5%".

[0107] For candidate self-healing strategy records whose safety boundary determination field is "passed", a target self-healing strategy record is generated. The target self-healing strategy record includes at least a strategy number field, a steel grade field, a target casting speed adjustment value field, a target cooling water allocation value field, a target roll gap compensation value field, a protective slag switching suggestion field, and an execution priority field. If no verified candidate self-healing strategy record is found in the determined digital twin result set, the rejection feedback record is re-entered into step S520 to trigger constraint re-search; in one embodiment, if no verified candidate self-healing strategy record is obtained in three consecutive re-searches, automatic issuance is stopped and a manual review prompt is output. After screening, a target self-healing strategy set is output, which serves as the input for step S630.

[0108] S630: Read the target self-healing strategy set and, according to the execution priority field, record and send the target self-healing strategy to the programmable logic controller for execution, generating a strategy execution record. The strategy execution record includes at least a strategy number field, an execution start time field, an execution completion time field, an execution object field, and an execution status field; wherein, the execution object field includes at least one of the following: speed control module, secondary cooling water control module, roll gap control module, and operator prompt module. During execution, the target speed adjustment value field, the target cooling water allocation value field, and the target roll gap compensation value field are mapped to register write values ​​recognizable by the programmable logic controller, respectively, to ensure that the target self-healing strategy record can be directly executed by the field control system.

[0109] After the target self-healing strategy is executed, execution feedback data is collected at preset feedback time points to generate an execution feedback dataset. The execution feedback dataset includes at least the following fields: feedback time, pull speed after execution, liquid level after execution, temperature after execution, cooling water flow rate after execution, roll gap value after execution, equipment anomaly degree after execution, and current defect risk value after execution.

[0110] In one embodiment, the feedback time field corresponds to three sampling time points: 30s, 3min, and 10min. An integrity check is performed on the execution feedback dataset; incomplete records are deleted; complete records are written back to the historical time-series database and used again as input for step S210, thus forming a data and control closed loop from anomaly handling instruction input, multimodal data perception, target dynamic spatiotemporal causal graph construction, counterfactual root cause determination, causal constraint strategy generation, digital twin verification to on-site execution feedback. Taking the longitudinal crack scenario as an example, when the root cause is determined to be a mismatch between the protective slag-related variables and the high-speed casting process, the target self-healing strategy record can be written as "temporarily reduce the casting speed to 1.0m / min and switch to low-viscosity protective slag in the next batch." The execution feedback dataset continues to be used to verify whether the current defect risk value field has decreased.

[0111] Example 2: Figure 2 As shown, this embodiment provides a self-healing system for the root causes of continuous casting quality based on causal graphs, including:

[0112] The task scheduling module is used to read exception handling instructions, generate task request records, decompose tasks and make resource decisions on the task request records, and output task scheduling results.

[0113] The data perception module collects raw vibration data, raw defect image data, and raw process parameter data based on task scheduling results. It then cleans and extracts features from these raw vibration data, raw defect image data, and raw process parameter data, and outputs a multimodal feature sequence.

[0114] The spatiotemporal alignment module is used to perform Lagrange spatiotemporal alignment on multimodal feature sequences, trace back the defect generation time and material slice motion segment, and output the Lagrange feature tensor;

[0115] The causal modeling module constructs a dynamic spatiotemporal causal graph of the target based on the Lagrange feature tensor and filters the set of candidate root causes.

[0116] The root cause determination module performs counterfactual verification and causal entropy sorting on the candidate root cause set based on the target dynamic spatiotemporal causal graph, and outputs the true root cause determination result.

[0117] The strategy generation module generates candidate self-healing strategies based on the root cause determination results.

[0118] The verification and execution module is used to perform digital twin verification of candidate self-healing strategies, output the target self-healing strategy and execute it, collect the execution feedback dataset and write it back.

[0119] All the above formulas are performed using dimensionless numerical calculations; the relevant formulas are based on empirical models that approximate the real situation, obtained through extensive data collection and software simulation fitting. The preset parameters and thresholds involved in the formulas can be conventionally set and adjusted by those skilled in the art according to the physical constraints of the actual application scenario.

[0120] Example 3: This example provides a computer program product, which includes computer program code. When the computer program code is run on a computer, the computer implements the self-healing method for the root cause of continuous casting quality based on causal graphs as described in Example 1.

[0121] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0122] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A self-healing method for the root causes of continuous casting quality based on causal graphs, characterized in that, Includes the following steps: S1. Read the exception handling instructions, generate a task request record, perform task decomposition and resource adjudication on the task request record, and output the task scheduling result. S2. Based on the task scheduling results, collect raw vibration data, raw defect image data, and raw process parameter data, clean and extract features, and output a multimodal feature sequence; S3. Perform Lagrange spatiotemporal alignment on the multimodal feature sequence, trace back the defect generation time and motion segment, and output the Lagrange feature tensor; S4. Construct a dynamic spatiotemporal causal graph of the target based on the Lagrange feature tensor, and filter the candidate root cause set; S5. Based on the target dynamic spatiotemporal causal graph, perform counterfactual verification and causal entropy sorting on the candidate root cause set, output the true root cause determination result, and generate candidate self-healing strategies based on the true root cause determination result.

2. The self-healing method for the root causes of continuous casting quality based on causal graphs according to claim 1, characterized in that, Also includes: S6. Perform digital twin verification on the candidate self-healing strategy, output the target self-healing strategy and execute it, collect the execution feedback dataset and write it back.

3. The self-healing method for continuous casting quality root causes based on causal graphs according to claim 1, characterized in that, S1 specifically includes: Read the exception handling instruction, flow number identifier, steel type identifier, defect type identifier, target quality index and production plan identifier, generate task request records, delete task request records with missing or duplicate fields, and output the set of valid task requests. Input the set of valid task requests into the metacognitive coordination agent, combine the task parsing knowledge graph to perform semantic parsing, object extraction and task decomposition, generate task orchestration records, delete task orchestration records with broken dependency chains, and output the set of valid task orchestrations. Based on the effective task orchestration set and the current computing power, production plan value, and energy consumption budget value, a candidate scheduling scheme set is generated. The total system utility value is calculated, candidate scheduling scheme records below the preset lower limit value are deleted, and the task scheduling result is output.

4. The self-healing method for continuous casting quality root causes based on causal graphs according to claim 1, characterized in that, S2 specifically includes: Based on the task scheduling results, the original vibration data, defect image data and process parameter data corresponding to the target furnace number, target flow number and target time window are read. Time reference verification and basic cleaning are performed. Records with incomplete time coverage, consecutive missing points reaching the preset threshold, exceeding the equipment range or abnormal image size are deleted. The preprocessed observation dataset is output. Extract equipment health status features, defect image features, and process status features from the preprocessed observation dataset, and output a multimodal feature sequence; Lagrange spatiotemporal alignment is performed based on multimodal feature sequences to calculate defect generation time, reconstruct material slice motion trajectory segments, align various features according to segments and time steps, and output Lagrange feature tensors.

5. The self-healing method for continuous casting quality root causes based on causal graphs according to claim 1, characterized in that, S3 specifically includes: Read the Lagrange feature tensor, perform variable mapping according to the preset variable dictionary, generate a system state vector sequence, an observation state slice set and a causal modeling sample set, and delete records with inconsistent variable dimensions or missing current defect risk value fields; Input the effective causal modeling sample set into the spatiotemporal variational causal discovery engine to construct a spatiotemporal structural equation model and output the initial dynamic spatiotemporal causal graph and the effective initial edge weight set. Perform continuous constraint optimization on the initial dynamic spatiotemporal causal graph and the effective initial edge weight set to obtain the target dynamic spatiotemporal causal graph, and filter ancestor nodes that have path connectivity with the target outcome variable to output the candidate root cause set.

6. The self-healing method for the root causes of continuous casting quality based on causal graphs according to claim 1, characterized in that, S4 specifically includes: Read the target dynamic spatiotemporal causal graph, candidate root cause set and effective observation state slice set, construct the fact world network and counterfactual world network, infer the posterior distribution of exogenous noise, delete records with noise variance exceeding the preset upper limit value, and output the effective exogenous noise inference set; Based on the effective exogenous noise inference set, counterfactual intervention is performed on candidate variables, pure causal effect is calculated, candidate variables with pure causal effect less than a preset threshold are deleted, and the true root cause candidate set is output. By integrating pure causal effects, defect classification confidence, equipment anomaly degree, and mechanism consistency indicators, the holographic causal entropy is calculated and sorted, and the true root cause determination result is output.

7. The self-healing method for continuous casting quality root causes based on causal graphs according to claim 1, characterized in that, S5 specifically includes: Read the root cause determination results, the target dynamic spatiotemporal causal graph, and the effective observation state slice set, generate the reinforcement learning state set and the constrained action set, construct the action mask based on the ancestor node set of the target outcome variable, and output the effective action set; Based on the set of reinforcement learning states and the set of effective actions, perform causal constraint reinforcement learning search, calculate the immediate reward value of each action combination, delete action combinations that exceed the preset action boundaries, and output the policy score result. Read the strategy scoring results, select candidate self-healing strategies, delete candidate self-healing strategy records with duplicate action combinations, and output the candidate self-healing strategy set.

8. The self-healing method for the root causes of continuous casting quality based on causal graphs according to claim 1, characterized in that, S6 specifically includes: Read the candidate self-healing strategy records in the candidate self-healing strategy set, and simultaneously read the steel grade process library records corresponding to the steel grade field. Use the digital twin verification agent to perform virtual trial and error on the candidate self-healing strategies, and output the digital twin prediction result records. Physical security boundary verification is performed on the digital twin prediction results record. For candidate self-healing strategies whose security boundary judgment field is not passed, a rejection feedback record is generated and returned. For candidate self-healing strategies whose security boundary judgment field is passed, a target self-healing strategy record is generated and the target self-healing strategy set is output. The system reads the target self-healing strategy set and sends it to the programmable logic controller for execution, generates strategy execution records, collects execution feedback data to form an execution feedback dataset, and writes the complete execution feedback dataset back to the historical time series database.

9. A self-healing system for the root causes of continuous casting quality based on causal graphs, employing the self-healing method for the root causes of continuous casting quality based on causal graphs as described in any one of claims 1 to 8, characterized in that, include: The task scheduling module is used to read exception handling instructions, generate task request records, decompose tasks and make resource decisions on the task request records, and output task scheduling results. The data sensing module collects raw vibration data, raw defect image data, and raw process parameter data based on task scheduling results, cleans and extracts features from the raw data, and outputs a multimodal feature sequence. The spatiotemporal alignment module is used to perform Lagrange spatiotemporal alignment on multimodal feature sequences, trace back the defect generation time and material slice motion segment, and output the Lagrange feature tensor; The causal modeling module constructs a dynamic spatiotemporal causal graph of the target based on the Lagrange feature tensor and filters the set of candidate root causes. The root cause determination module performs counterfactual verification and causal entropy sorting on the candidate root cause set based on the target dynamic spatiotemporal causal graph, and outputs the true root cause determination result. The strategy generation module generates candidate self-healing strategies based on the root cause determination results. The verification and execution module is used to perform digital twin verification of candidate self-healing strategies, output the target self-healing strategy and execute it, collect the execution feedback dataset and write it back.

10. A computer program product, characterized in that, The computer program product includes computer program code, which, when run on a computer, causes the computer to implement the continuous casting quality root cause self-healing method based on causal graphs as described in any one of claims 1 to 8.

Images