A continuous casting quality management and control method based on a meta-cognition coordination architecture intelligent agent cluster

Abstract

The application provides a continuous casting quality management and control method based on a meta-cognition coordination architecture intelligent agent cluster, and relates to the technical field of steel metallurgy intelligent manufacturing. The continuous casting quality management and control method comprises: data input and standardization, central coordination, and intelligent agent cluster operation and maintenance. In the central coordination process, the meta-cognition coordination intelligent agent is the core, and coordinates six types of functional intelligent agents, namely, semantic analysis intelligent agents, data perception intelligent agents, defect prediction intelligent agents, root cause analysis intelligent agents, process optimization intelligent agents, and digital twin intelligent agents, to realize task scheduling and state monitoring of the whole system, complete three key functions of task decomposition, data scheduling, and closed-loop management and control. In the closed-loop management and control step, risk early warning, defect prediction, root cause analysis, process optimization, and digital twin verification and feedback are carried out for the continuous casting quality. Compared with the traditional scheme, the application is optimized in the aspects of whole process, multi-modal, intelligentization, and collaboration, not only improves management efficiency, but also increases economic benefits.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steel metallurgical intelligent manufacturing, and particularly relates to a continuous casting quality management and control method based on a meta-cognition coordination architecture intelligent agent cluster. BACKGROUND

[0002] Chinese patent (publication number CN 119025412A) published on November 26, 2024 discloses an intelligent agent processing method: in the case of needing to inspect intelligent agents, a target intelligent agent to be processed is determined from multiple intelligent agents; at least one target test case required for simulating detection of the target intelligent agent is obtained; the target intelligent agent is simulated and detected based on the at least one target test case, and a detection result is obtained. The application of the main intelligent agent in this technology is related to virtual assistants and intelligent assistants, and the intelligent agent in the industrial environment cannot be effectively applied.

[0003] Chinese patent (publication number CN119903172A) published on April 29, 2025 discloses a large language model training method for generating intelligent agents, a generated agent template is screened by a labeled agent template, and the large language model is trained to form an automatic iterative closed loop, which can realize optimization of intelligent agent training corpus and continuous improvement of model performance, automatic feedback and adjustment, so that the intelligent agent can continuously learn and adapt, and the stability and reliability of the model are improved. This technology mainly uses a generative large language model to form a closed loop of the corresponding intelligent agent, and does not involve the corresponding cooperation and industrial application of the intelligent agent.

[0004] In the industrial application, especially in the field of steel metallurgy, there is an urgent need for a continuous casting quality intelligent agent system that integrates deep learning, reinforcement learning and large language model (LLM) to solve the following problems existing in this field:

[0005] (1) Defect tracing inefficiency problem, including: manual investigation across MES / L2 / point inspection systems, long time consumption (>48 hours), defect attribution accuracy <65%; traditional numerical models (such as two-cooling zone heat exchange coefficient optimization model) rely on empirical formulas, ignore multi-factor coupling effects, etc.

[0006] (2) Process adjustment lag problem, including: 72% of defects are exposed in the rolling process, causing huge downtime losses; manual parameter adjustment misadjustment rate is 35%; static optimization algorithms (such as genetic algorithm) do not integrate real-time working condition data.

[0007] (3) Serious data island problem, including: L1 / L2 / MES system data delay >72 hours, utilization rate <25%, etc. SUMMARY

[0008] In view of the deficiencies of the prior art, the present application proposes a continuous casting quality intelligent agent system that fuses deep learning, reinforcement learning and large language model (LLM). The system realizes the whole life cycle quality control of the continuous casting process from molten steel casting to slab cutting by building a "process perception-defect prediction-root cause tracing-dynamic optimization-digital twin verification" whole-process closed-loop management and control system. Through the three-dimensional semantic analysis engine, multi-modal feature fusion model, causal reasoning dual engine and self-adaptive optimization algorithm, the industry pain points such as data island, inefficient defect tracing and process adjustment lag in traditional continuous casting quality control are solved.

[0009] In a first aspect, the present application provides a continuous casting quality control method, comprising the following steps:

[0010] S1: data input and standardization; L1 / L2 / MES data are acquired, data conversion and standardization are performed on the basis of OPC UA combined with real-time data pipeline technology, data interaction between the intelligent agent cluster and the industrial field system is realized, and multi-source data fusion is completed;

[0011] S2: central coordination; a meta-cognition coordination intelligent agent is used as the core to coordinate the function intelligent agents, task scheduling and state monitoring of the whole system are realized, and three key functions of task decomposition, data scheduling and closed-loop control are completed; the function intelligent agents include: semantic analysis intelligent agent, data perception intelligent agent, defect prediction intelligent agent, root cause analysis intelligent agent, process optimization intelligent agent and digital twin intelligent agent;

[0012] The central coordination specifically includes:

[0013] S21: task decomposition; a complex quality control task is decomposed into subtasks such as semantic analysis, data acquisition and defect prediction, and then the semantic analysis intelligent agent is driven to understand the instructions;

[0014] S22: data scheduling; the meta-cognition coordination intelligent agent dynamically schedules the data perception intelligent agent to collect or preprocess data according to the load and task priority of each function intelligent agent;

[0015] S23: closed-loop control; under the core coordination of the meta-cognition coordination intelligent agent, the function intelligent agents work cooperatively to realize the whole-cycle control of quality problems; the specific steps include:

[0016] S231: risk warning; the meta-cognition coordination intelligent agent judges according to the fusion data output by the data perception intelligent agent and sends a risk warning signal to the defect prediction intelligent agent;

[0017] S232: defect prediction; the defect prediction intelligent agent predicts the slab defects based on multi-modal features using an LSTM-CNN hybrid model and outputs defect spatiotemporal features and risk levels;

[0018] S233: Root cause analysis; the root cause analysis agent receives the spatiotemporal feature data from the defect prediction agent, traces the defect causes by using a CART decision tree + Bayesian network, and generates root cause analysis results;

[0019] S234: Process optimization; the process optimization agent receives the output data from the root cause analysis agent, generates process parameter adjustment strategies by using an NSGA-III multi-objective algorithm;

[0020] S235: Digital twin verification and feedback; the digital twin agent receives the "parameter adjustment strategy" data from the process optimization agent, runs a thermal / stress field coupling model, simulates the quality of the casting blank after parameter adjustment, and feeds back the verification results to the meta-cognition coordination agent;

[0021] S3: Intelligent agent cluster operation and maintenance; specifically comprising the following steps:

[0022] S31: Intelligent agent optimization iteration; the running data of the intelligent agent is processed and analyzed, and the intelligent agent is continuously optimized through an iteration link of "monitoring -> evaluation -> optimization -> application" to improve the running effect;

[0023] S32: Intelligent agent performance monitoring; the performance data of the intelligent agent is collected and analyzed, and the health of the intelligent agent is evaluated and displayed.

[0024] As a further improvement of the application, data collaboration is carried out in the way of blackboard model knowledge sharing, the process state tensor is stored in the shared memory area, the whole process state is integrated by using the structured tensor, and the data collaboration between intelligent agents is realized by shared memory + SQL-like query.

[0025] As a further improvement of the application, in the intelligent agent communication protocol, the message format specification and performance guarantee mechanism are set, specifically,

[0026] The message format specification of the communication between intelligent agents is defined to ensure that different intelligent agents can accurately obtain the information transmitted by each other;

[0027] The performance guarantee mechanism setting includes: using Apache Kafka + RDMA network to limit the end-to-end data transmission delay; based on the RAFT consensus algorithm, the intelligent agent state synchronization is realized, so that when part of the intelligent agents fail, the standby intelligent agent can quickly take over the work; and the SM4 national encryption algorithm is used to guarantee the safety of industrial data.

[0028] As a further improvement of the application, the L1 / L2 / MES data is obtained, specifically including:

[0029] Obtain the real-time process data of sensors / actuators of the L1 basic automation layer;

[0030] Acquire the process model calculation results of the L2 process automation layer:

[0031] Acquire the production business data of the MES manufacturing execution system.

[0032] As a further improvement of the application, in the central coordination,

[0033] The metacognition coordination agent adopts a hierarchical task decomposition model:

[0034]

[0035] Wherein, A: task set, containing three types of task candidates "prediction, traceability, optimization"; T: target task to be determined finally, selected from set A through subsequent calculation; : is the weight of the corresponding task ; : is the score function of the task ;

[0036] In this model, dynamic allocation of weights is supported.

[0037] As a further improvement of the application, in the central coordination,

[0038] A real-time load balancing algorithm is used to guide dynamic scheduling of resources:

[0039]

[0040] Wherein, LoadIndex: load index, reflecting the load state of the system; QueueLen i : length of the i-th queue; k: total number of queues in the system; MaxQueue: maximum value of queue length;

[0041] When the load index exceeds the threshold value, trigger the Kubernetes container elastic expansion and contraction.

[0042] As a further improvement of the application, the process of the closed loop control further includes cognitive iteration of the metacognition coordination agent, and the specific steps are:

[0043] After the metacognition coordination agent in the "risk early warning→defect prediction→root cause analysis→process optimization→digital twin verification and feedback" link, receives the verification result feedback information, completes the cognitive iteration.

[0044] As a further improvement of the application, the defect prediction step adopts a multi-scale spatio-temporal feature fusion algorithm architecture, which accurately captures the time sequence dynamics and multi-scale correlation of process parameters, and provides strong feature characteristics for defect probability prediction; The specific steps include:

[0045] Dual-branch extraction: extract timing features F from the process parameter timing flow seq and spatial / multi-scale features F cnn ;

[0046] Attention fusion: model the cross-relation between F seq and F cnn , output attention-enhanced features F att ;

[0047] Gate decision: dynamically adjust the fusion weights of F seq and F cnn , get their dynamically weighted sum F fusion ;

[0048] Probability output: input F fusion into a fully connected layer + Softmax classifier, output defect probability distribution.

[0049] As a further improvement of the application, the specific steps of the dual-branch extraction include:

[0050] Input continuous timing data of the continuous casting process, and input these data into the timing feature branch and the spatial / multi-scale feature branch;

[0051] In the timing feature branch, a BiLSTM-Pro module is constructed, a bidirectional LSTM + enhanced version of time convolutional network (TCN) is used to extract long and short term timing dependencies, and timing features F seq are output; wherein the enhanced version of the time convolutional network (TCN) is that a convolution kernel covers a short time window to strengthen local timing patterns;

[0052] In the spatial / multi-scale feature branch, a multi-scale CNN module is constructed, a CNN with different sizes of convolution kernels is used to cover full-scale features from "short-period details" to "long-period trends", extract multi-scale spatio-temporal patterns, and output spatial / multi-scale features F cnn .

[0053] As a further improvement of the application, in the spatial / multi-scale feature branch, an EfficientNet-B4 model is used as a spatial feature extraction layer to convert a large resolution image into features containing semantics and spatial structure, and its general formula is expressed as:

[0054]

[0055] wherein F spatial : output spatial features, EfficientNet: efficient convolutional neural network, I 5120×3840 : input image data with a resolution of 5120x3840.

[0056] As a further improvement of the present invention, the specific steps of attention fusion include:

[0057] The spatiotemporal cross-attention module is constructed by using an attention mechanism to build a feature fusion layer, and by calculating F... seq With F cnn The association weights are modeled to determine the cross-correlation between the two features, thereby strengthening the most critical feature combination for defect prediction and outputting the attention-enhanced feature F. att .

[0058] As a further improvement of the present invention, the specific steps of the gating decision include:

[0059] F att The input gating feature fusion module dynamically determines F through the gating coefficient g. seq With F cnn The fusion weights; the gating mechanism formula is:

[0060]

[0061] Where σ: sigmoid activation function, W g b g : Gated weight matrix and bias, [F seq , F cnn ]: F seq With F cnn Perform feature splicing. Element-wise multiplication; F fusion :F seq With F cnn The dynamic weighted sum.

[0062] As a further improvement of the present invention, the gating decision also includes:

[0063] The W in the gating mechanism is learned using an online learning engine. g The parameters are updated incrementally, and the generalized formula for the parameter update is expressed as:

[0064]

[0065] Where, θ t : Model parameters at time t, η: learning rate, controlling the step size of gradient descent. New Data D new loss gradient on, : Regularization term.

[0066] As a further improvement of the present invention, in the root cause analysis step, different types of data are assigned weights and processed using different methods, specifically including:

[0067] The process parameters are processed by using a CART decision tree method;

[0068] The device vibration signal is processed by using a wavelet packet decomposition + energy entropy analysis method;

[0069] The operation log is processed by using a BERT semantic embedding method;

[0070] The environment temperature and humidity are processed by using a time series anomaly detection method.

[0071] As a further improvement of the application, in the root cause analysis step, the Bayesian network is used to dynamically calibrate the "root cause-defect" causal relationship, adjust the probability distribution using new samples, and adjust the weight of the causal relationship in real time. The formula of the Bayesian network is:

[0072]

[0073] Wherein, X: root cause event, P(crack|X): conditional probability of root cause X on "crack", : the count of root cause X occurrence and crack appearance in the sample, : the total count of root cause X occurrence in the sample.

[0074] As a further improvement of the application, the process optimization step further comprises a priority experience replay PER and a reward function, and the specific steps comprise:

[0075] A global optimal solution (v*, c*) is generated by using a genetic algorithm NSGA-III multi-objective optimization model;

[0076] A dynamic adjustment is made by using a reinforcement learning method to generate a dynamic adjustment amount (Δv, Δc);

[0077] The real-time system executes f(v*+Δv, c*+Δc) for verification, and feeds back the actual reward and experience data.

[0078] As a further improvement of the application, the specific way of the NSGA-III multi-objective optimization is:

[0079]

[0080] Wherein, v: process parameter, c: environment and control parameter, f1: defect rate, N defect : number of defective products, N total : total product quantity, f2: cooling energy consumption rate, E cool : current energy consumption of the cooling system, E max : maximum energy consumption of the cooling system, f3: vibration energy ratio, VibEnergy: current vibration energy, Vib max: maximum vibration energy.

[0081] As a further improvement of the application, the specific way of the reinforcement learning comprises: adopting priority experience replay (PER) combined with a reward function to optimize the policy, and generating dynamic adjustment quantities (Δv, Δc);

[0082] The formula of the priority experience replay (PER) is:

[0083]

[0084] Wherein, P(i): sampling probability of the i-th experience, δ i : deviation of the predicted value and the actual value, ε: error correction term, α: priority weight;

[0085] The reward function is:

[0086]

[0087] Wherein, r t : reward value at time t, RiskLevel: risk level, DefectRate: defect rate, Threshold: defect rate threshold, EnergySave: energy saving ratio, Vibration: current vibration energy, Vib max : maximum vibration energy.

[0088] As a further improvement of the application, the specific steps of the agent optimization iteration comprise:

[0089] Agent monitoring: real-time collection of operation data of the agent cluster as operation and maintenance evaluation data;

[0090] Construction of evaluation database: structured storage and index preprocessing of the collected operation data, conversion into structured indexes that can be evaluated and analyzed, and construction of an evaluation database;

[0091] Automatic optimization analysis: based on the indexes of the evaluation database, multi-objective optimization analysis is performed to obtain a parameter optimization adjustment scheme;

[0092] Optimization adjustment: the parameter optimization adjustment scheme is issued to the agent cluster to complete parameter adjustment.

[0093] As a further improvement of the application, in the agent health assessment, a trajectory deviation algorithm is used to quantify the deviation degree of the actual trajectory and the expected trajectory of the agent by using the average value of the relative deviation, and the algorithm formula is:

[0094]

[0095] Wherein, Deviation: trajectory deviation, N: total number of trajectory points, : the i th actual trajectory point, : the i th expected trajectory point;

[0096] When the trajectory deviation exceeds a set threshold, an intelligent agent behavior anomaly alarm is triggered.

[0097] As a further improvement of the present application, in the intelligent agent health assessment, a four-dimensional scoring model is adopted, and the overall performance of the intelligent agent is quantified by weighted summation of the accuracy, time, cost and interpretability four dimensions;

[0098] The interpretability dimension reflects the transparency of the intelligent agent decision-making process, including the clarity of root cause positioning and the traceability of the decision path, and the specific score is obtained by expert manual evaluation.

[0099] As a further improvement of the present application, the health instrument display converts the health status of the intelligent agent cluster into a visual form; its health status indicators include:

[0100] Task performance indicators, specifically including:

[0101] Defect prediction accuracy: measured by confusion matrix calculation method, target value > 95%;

[0102] Root cause positioning error: measured by coordinate comparison method, target value < ±1mm;

[0103] Response performance indicators, specifically including:

[0104] End-to-end delay: measured by a distributed tracking system, target value < 100ms;

[0105] Throughput: measured by stress testing method, target value > 800 QPS;

[0106] Resource performance indicators, specifically including:

[0107] CPU / GPU utilization: obtained by Prometheus monitoring, target value 60%-80%;

[0108] Memory usage: counted at container runtime, target value < 32GB / node.

[0109] In a second aspect, the present application provides a continuous casting quality management hardware system, comprising: a data acquisition component, an edge computing node, an intelligent processing and control component, a visualization and interaction component, a communication network component and a data storage component, the parts in the hardware system work together to realize the steps of the method in the first aspect.

[0110] The communication network component is used to realize data transmission and instruction interaction, and guarantee information connectivity between system levels.

[0111] The data storage component is used to store computer program files and data files.

[0112] As a further improvement of the present application, the data acquisition component comprises:

[0113] Parameter sensor: used to acquire process parameter data in industrial production process;

[0114] Visual acquisition device: used to acquire image or video data of production scene;

[0115] Vibration sensing device: used to acquire vibration signal data when the equipment is running.

[0116] As a further improvement of the present application, the edge computing node is in communication connection with the data acquisition component, and is used to perform real-time preprocessing on the acquired raw data.

[0117] As a further improvement of the present application, the intelligent processing and control component comprises:

[0118] Intelligent agent server cluster: in communication connection with the edge computing node, and deployed with functional intelligent agents, and used to perform deep analysis on the preprocessed data;

[0119] The intelligent agent server cluster comprises a GPU server.

[0120] Industrial control subsystem: in communication connection with the edge computing node, and used to receive control instructions output by the intelligent agent server cluster or the edge computing node, and to perform real-time adjustment on the running parameters of the industrial production equipment;

[0121] The industrial control subsystem comprises a PLC control system.

[0122] As a further improvement of the present application, the visualization and interaction component: in communication connection with the intelligent agent server cluster, and used to display real-time data, analysis results, equipment state and control instruction execution feedback of the production process, and to support operators to perform monitoring, intervention and decision-making;

[0123] The visualization and interaction component comprises a 3D monitoring large screen.

[0124] In a third aspect, the present application provides a computer program product, which, when executed by a processor, realizes the steps of the method in the first aspect.

[0125] Compared with the prior art, the application takes metacognition coordination agent as the center to form a whole cycle closed loop logic from "data collection" to "cognition iteration". The application realizes the whole life cycle quality control of the continuous casting process from molten steel casting to slab cutting by constructing a "process perception-defect prediction-cause tracing-dynamic optimization-digital twin verification" whole process closed loop management and control system. Through the three-dimensional semantic analysis engine, the multi-modal feature fusion model, the causal reasoning double engine and the self-adaptive optimization algorithm, the data island, the low efficiency of defect tracing and the lag of process adjustment in the traditional continuous casting quality control are solved. BRIEF DESCRIPTION OF DRAWINGS

[0126] Figure 1 An intelligent agent cluster topology structure for continuous casting quality control is disclosed in the application.

[0127] Figure 2 A multi-scale spatio-temporal feature fusion algorithm architecture diagram of a defect prediction intelligent agent.

[0128] Figure 3 A mixed optimization algorithm architecture diagram of a process optimization intelligent agent.

[0129] Figure 4 A continuous casting quality control hardware system topology structure diagram is disclosed in the application.

[0130] Figure 5 An intelligent agent cluster operation and maintenance framework diagram. DETAILED DESCRIPTION

[0131] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely in combination with the drawings. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. In the embodiments described by the application, steps S1, S2, … do not limit the unique execution steps of the application; various models, simulation environments and software described by the application are not the only limited way of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the application.

[0132] In the present application, computer device / system refers to the relevant entities applied to computers, such as hardware, combination of hardware and software, software or software in execution, etc. In detail, for example, software includes but is not limited to process running on processor, processor, object, executable software, execution thread, program and / or computer. Also, application programs or scripts running on servers, servers can be software. One or more software can be in the process and / or thread of execution, and the software can be localized on one computer and / or distributed between two or more computers, and can be run by various computer readable media.

[0133] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0134] In a first aspect, the present application provides a continuous casting quality management method embodiment, as shown in the following figure, the method takes metacognition coordination agent as the center to form a whole cycle closed loop of "data collection→defect prediction→root cause tracing→process optimization→digital twin verification→feedback optimization", and the basic logic structure is as follows: Figure 1 Data input: L1 / L2 / MES data flows into data perception agent through OPC UA, and multi-source fusion is completed;

[0135] Defect prediction: fusion data drives model to output defect risk, and triggers risk warning;

[0136] Root cause tracing: based on defect characteristics, using double engine (CART+Bayes) to locate the root cause of the problem;

[0137] Process optimization: generating parameter adjustment strategy under root cause orientation;

[0138] Physical verification: digital twin simulates the adjustment effect and verifies the effectiveness of the strategy;

[0139] Cognitive iteration: the verification result is fed back to the metacognition coordination agent, and the subsequent task scheduling and decision logic are optimized.

[0140] The present application has the following special design in data collaboration and communication:

[0141] (1) In terms of data collaboration, the present application adopts blackboard model knowledge sharing, and shared memory area stores process state tensor. The whole process state is integrated by structured tensor, and the data collaboration between intelligent agents is connected through shared memory+SQL-like query.

[0142] The expression of process state tensor is as follows:

[0143]

[0144]

[0145] ​where T meniscus : liquid surface temperature; v cast : drawing speed; : cooling pressure gradient; f vib : vibration frequency; sigma stress : stress distribution; delta crack : crack size; rho inclusion : inclusion density;

[0146] In an embodiment of the present application, the cross-agent SQL-like query mode is as follows:

[0147] SELECT vibration_data FROM Blackboard WHERE segment=7 AND time > '2025-03-15 14:00:00';

[0148] (2) In the agent communication protocol, the present application sets the message format specification and performance guarantee mechanism, specifically:

[0149] First, the message format specification for inter-agent communication is defined to ensure that different agents can accurately obtain the information transmitted by each other; the message format specification defined in an embodiment of the present application is as follows:

[0150] { "sender": "defect_predict_agent",

[0151] "receiver": "root_cause_agent",

[0152] "timestamp": "2025-03-15T14:23:18.123Z",

[0153] "priority": 3,

[0154] / / 1-5 level, 3 level and above trigger real-time response;

[0155] "content": { "event_id": "INC-20250315-1423",

[0156] "defect_type": "surface longitudinal crack",

[0157] "risk_level": 4,

[0158] "confidence": 0.92,

[0159] "data_ref": ["opcua: / / temp / segment7", "rtsp: / / cam12 / stream"]}};

[0160] Secondly, in terms of performance guarantee mechanism, the following three aspects are designed to guarantee the real-time and stability of the control method of the application:

[0161] Low delay transmission: Apache Kafka + RDMA network is adopted to ensure that the end-to-end data transmission delay is less than or equal to 50ms;

[0162] Fault-tolerant design: based on RAFT consensus algorithm, the state synchronization of intelligent agent is realized, and after the failure of a certain intelligent agent, the standby intelligent agent can quickly take over the work, and the fault switching time is less than 200ms;

[0163] Security encryption: SM4 national encryption algorithm is adopted to ensure the security of industrial data.

[0164] The specific implementation steps of the application can be as follows:

[0165] S1: data input and standardization, L1 / L2 / MES data is transmitted and standardized through OPC UA real-time data pipeline, input data sensing intelligent agent, and multi-source fusion is completed;

[0166] S11: collect data from L1 / L2 / MES system; wherein, L1 / L2 / MES system is the "source" of data in industrial field, specifically,

[0167] L1 (basic automation): collect real-time process data of sensors / actuators (such as crystallizer vibration frequency, secondary cooling water flow);

[0168] L2 (process automation): output process model calculation results (such as solidification endpoint prediction, temperature field distribution);

[0169] MES (manufacturing execution system): provide production business data (such as steel grade, pouring plan, quality traceability record).

[0170] S12: standardize the collected data by using OPC UA real-time data pipeline;

[0171] OPC UA is an open platform communication unified architecture, which guarantees the real-time (delay <5ms) and reliability (redundant transmission, data encryption) of data transmission in "publish-subscribe" mode; the application combines real-time data pipeline technology on the basis of OPC UA to realize data conversion and standardization, and realizes data interaction between intelligent agent cluster and industrial field system.

[0172] S2: central coordination;

[0173] The application adopts meta-cognition coordination agent as the core, coordinates 6 types of functional agents (semantic analysis agent, data perception agent, defect prediction agent, root cause analysis agent, process optimization agent, and digital twin agent), realizes task scheduling and state monitoring of the whole system, and completes three key functions of task decomposition, data scheduling, and closed-loop management and control;

[0174] The meta-cognition coordination agent adopts a hierarchical task decomposition model:

[0175]

[0176] Among them, A: task set, including three types of task candidates "prediction, tracing, optimization"; T: target task to be determined finally, selected from set A through subsequent calculation; : is the weight corresponding to the task ; : is the score function of the task ;

[0177] In this model, the weight can be dynamically allocated, for example, when the defect risk is greater than 3, the weight of the prediction task is set to .

[0178] At the same time, a real-time load balancing algorithm is used to guide dynamic scheduling of resources:

[0179]

[0180] Among them, LoadIndex: load index, reflecting the load state of the system; QueueLen i : length of the i-th queue; k: total number of queues in the system; MaxQueue: maximum value of queue length;

[0181] When the load index is less than 0.6, trigger Kubernetes container elastic expansion.

[0182] S21: task decomposition, decompose complex quality control tasks (such as "analyze the crack reason of sector 8 area") into semantic analysis, data collection, defect prediction, etc. Subtasks, and then understand the instructions through the semantic analysis agent.

[0183] In an embodiment of the application, the semantic analysis agent receives the "task decomposition" instruction from the meta-cognition coordination agent, and is required to analyze the industrial natural language instruction (such as "query the number of times of liquid level fluctuation over limit of No. 3 casting machine yesterday"), and map it into a system executable parameter; the semantic analysis agent analyzes and outputs a standardized query instruction.

[0184] S22: Data scheduling: The meta-cognitive coordination agent dynamically schedules the data perception agent to collect / preprocess data (such as vibration spectrum in high-risk areas) according to the load and task priority of each agent.

[0185] In an embodiment of the present application, the data perception agent receives the "data scheduling" instruction from the meta-cognitive coordination agent, requiring to extract multi-source data (process, image, vibration) from the OPC UA pipeline, perform standardization and feature extraction; the data perception agent is processed and outputs the fused spatio-temporal feature matrix.

[0186] S23: Closed-loop control: Under the core coordination of the meta-cognitive coordination agent, the defect prediction agent, root cause analysis agent, process optimization agent, and digital twin agent work together to realize the whole-cycle control of quality problems through the "risk warning-defect prediction-root cause analysis-process optimization-digital twin verification-feedback" link ("verification results" are fed back to the meta-cognitive coordination agent itself to complete cognitive iteration); the specific process includes.

[0187] S231: Risk warning;

[0188] The meta-cognitive coordination agent judges according to the fused data output by the data perception agent and sends the "risk warning" information to the defect prediction agent.

[0189] S232: Defect prediction;

[0190] The defect prediction agent predicts the casting blank defect (type, risk, location) based on multi-modal features using an LSTM-CNN hybrid model and outputs defect spatio-temporal features and risk levels.

[0191] As shown in Figure 2 The present application further improves the "multi-scale spatio-temporal feature fusion algorithm architecture" of the defect prediction agent: through double-branch parallel extraction + spatio-temporal attention fusion + gated dynamic decision, the time sequence dynamics and multi-scale correlation of process parameters are accurately captured, providing strong feature characteristics for defect probability prediction.

[0192] Compared with the prior art, the "single branch extraction", "simple splicing" and "fixed weight fusion" of the features cannot adapt to the situation of complex industrial scenes. The improvement of the present application realizes a substantial technical breakthrough, that is: (1) through the double branch design, the "time trend of process parameters" and "multi-scale correlation between parameters / time window" are covered at the same time, avoiding the information loss of single-dimensional features; (2) through the attention mechanism, the synergistic correlation of the two types of features is dynamically captured, so that the model pays more attention to "the feature combination most indicative of defects"; (3) through the gating coefficient, "dynamic adjustment of feature weight" is realized, and according to the current input process state, it is automatically determined whether to "rely more on time sequence features or spatial features", so that the fusion strategy is more in line with the dynamics of actual production.

[0193] The improved defect prediction specific steps include:

[0194] S2321: double branch extraction;

[0195] The process parameter time sequence flow is used as data input to provide continuous time sequence data of the continuous casting process, such as parameter sequences of time-varying parameters such as pulling speed, cooling water volume, liquid level, etc.; these data are input into the time sequence feature branch and the spatial / multi-scale feature branch for subsequent processing.

[0196] Time sequence feature branch: construct a BiLSTM-Pro module, adopt a bidirectional LSTM+time convolution network (TCN) enhanced version, extract long and short term time sequence dependencies, and output time sequence features F seq ; for example, capture the "cumulative effect" of liquid level fluctuation (such as 30-minute oscillation trend); wherein the enhanced version of the time convolution network (TCN) is to cover a short time window (such as 5 time steps) with a convolution kernel, and to strengthen local time sequence patterns (such as the oscillation period of liquid level every 2 minutes).

[0197] Wherein, the general formula of the time sequence feature extraction layer in the BiLSTM-Pro module is:

[0198]

[0199] Wherein, : the final feature after fusion, : extract time sequence dependent features using the current input x t and the state of the last moment, : extract local time sequence features using the input x t-k:t from the past k steps to the present, k: time convolution kernel size, k∈{3,5,7}, respectively for capturing short / medium / long term dependencies, : element-wise addition.

[0200] Spatial / multiscale feature branch: a multiscale CNN module is constructed, and a CNN with different sizes of convolution kernels (3×3, 5×5, 7×7, etc.) is adopted to cover full-scale features from “short-period details” to “long-period trends”, extract multiscale spatiotemporal patterns, and output spatial / multiscale features F cnn For example, capture the “distribution anomaly” (such as local overcooling) of the cooling water amount in the sector 7-9 area, and the difference in fluctuation between the 30-minute and 2-hour speeds.

[0201] In an embodiment of the present application, the EfficientNet-B4 model is used as a spatial feature extraction layer to convert a large resolution image into a feature containing semantics and spatial structure, and its general formula is expressed as:

[0202]

[0203] wherein F spatial : output spatial features, EfficientNet: efficient convolutional neural network, I 5120×3840 : input image data with a resolution of 5120×3840;

[0204] The inclusion detection result of this embodiment reaches mAP@0.5=0.97, that is, when the overlap ratio of the detection box and the real inclusion area is ≥50%, the average detection accuracy of all classes (such as different types of inclusions) is as high as 97%.

[0205] S2322: attention fusion;

[0206] A spatiotemporal cross-attention module is constructed, specifically an attention mechanism (Attention) is used to construct a feature fusion layer, the cross correlation between F seq and F cnn is modeled by calculating the correlation weight of F seq and F cnn , and the most critical feature combination for defect prediction is strengthened; for example, by calculating the correlation degree of “speed mutation” and “uneven distribution of cooling water amount”, the collaborative defect risk of the two is obtained;

[0207] The algorithm formula is:

[0208]

[0209] wherein Q, K, V: query matrix, key matrix, value matrix obtained by linear transformation of F seq and F cnn , d: feature dimension;

[0210] In this formula, the weight matrix obtained by processing with the softmax function is used for “weighted fusion” F seq and F cnn two types of features, and the attention-enhanced feature Fatt .

[0211] S2323: Gating decision;

[0212] The gating feature fusion module is constructed, a gating mechanism (Gated Mechanism) is used to construct a decision layer, and the fusion weights of two types of features are dynamically adjusted. Specifically, F att The gating feature fusion module is input, and the fusion weights of F seq and F cnn are dynamically determined by the gating coefficient g. The gating mechanism formula is:

[0213]

[0214] Where, σ: sigmoid activation function, W g , b g : gating weight matrix and bias, [F seq , F cnn ]: feature splicing of F seq and F cnn , ⊙: element-wise multiplication; F fusion : dynamic weighted sum of F seq and F cnn .

[0215] According to the dynamic decision logic of the above formula: when the liquid surface fluctuates violently (time sequence feature dominates the defect risk), g→1, the fusion feature is more biased towards F seq ; when the cooling water volume distribution is abnormal (spatial feature dominates the defect risk), g→0, the fusion feature is more biased towards F cnn .

[0216] Further, the online learning engine is used to incrementally update the W g parameters in the gating mechanism: incremental training every 24 hours, and the model iteration period is <30 minutes. The general formula for parameter updating is:

[0217]

[0218] Where, θ t : model parameters at time t, η: learning rate, control step size of gradient descent, : loss gradient on new data D new , : regularization term.

[0219] S2324: Probability output;

[0220] The defect probability output module is constructed, and the features F fusionAn input fully connected layer + a Softmax classifier outputs a defect probability distribution; for example, outputting prediction results such as "surface longitudinal crack probability 0.92" and "internal inclusion probability 0.15".

[0221] In an embodiment of the present application, the improved defect prediction method proposed in the present application achieves a leap-forward improvement in defect prediction accuracy, and better supports closed-loop management and control of continuous casting quality. Specifically, using a No. 1 continuous casting machine of a certain steel plant, the following results are obtained: (1) in terms of accuracy improvement: the accuracy of surface crack prediction is improved from 89% to 97%, and the false positive rate is reduced by 42%; the F1-Score of internal inclusion prediction is improved from 0.91 to 0.98 (the closer the F1-Score is to 1, the more balanced the precision and recall are); (2) in terms of closed-loop support: the accurate defect probability output provides reliable defect spatio-temporal features (such as "crack occurs in sector 7 zone, corresponding to a 2-minute fluctuation period of the pulling speed") for the subsequent "root cause analysis agent", and provides a basis for the parameter adjustment decision of the "process optimization agent" (such as "specifically reduce the cooling water amount in sector 7 zone").

[0222] S233: root cause analysis;

[0223] The root cause analysis agent receives the "spatio-temporal feature" data from the defect prediction agent, traces the defect causes using a CART decision tree + Bayesian network, and generates results such as root cause heat map, causal atlas, and root cause weight.

[0224] Further, the root cause analysis agent is upgraded in the present application, a multi-modal evidence fusion system is constructed, different methods are used to process different types of data input into the system, and different weights are assigned to the data, and the specific correspondence is shown in Table 1.

[0225] Table 1:

[0226]

[0227] The upgrade and improvement further include using a Bayesian network to dynamically calibrate the "root cause-defect" causal relationship, adjusting the probability distribution using new samples, and adjusting the weight of the causal relationship in real time. The formula of the Bayesian network is:

[0228]

[0229] wherein, X: root cause event, P(crack|X): conditional probability of "crack" given root cause X, : count of occurrence of root cause X and crack in the sample, : total count of occurrence of root cause X in the sample;

[0230] In an embodiment of the present application, the probability distribution is updated every time 100 new samples are received, and the "root-cause" causal relationship authority is adjusted accordingly. In the adjustment result, the weight of the liquid level fluctuation > ± 5 mm as the root cause of the crack is increased to 4.2.

[0231] S234: Process optimization;

[0232] The process optimization agent receives the "causal map" data from the root cause analysis agent, and uses the NSGA-III multi-objective algorithm to generate a process parameter adjustment strategy (simultaneously optimizing defect rate, energy consumption, and equipment life).

[0233] As shown in Figure 3 , further, the present application improves the process optimization agent by using NSGA-III, priority experience replay (PER), and reward function, and cooperates with the real-time system to jointly realize a complete closed loop of "theoretical optimal solution (NSGA-III) → dynamic adjustment (reinforcement learning + PER) → real-time system verification (feedback fitness)". The specific steps include:

[0234] S2341: Theoretical optimal solution;

[0235] A genetic algorithm NSGA-III multi-objective optimization model is used to generate a global optimal solution (v*, c*), which is used as the initial optimal solution.

[0236] The specific way of NSGA-III multi-objective optimization is as follows:

[0237]

[0238] Wherein, v: process parameters, c: environmental and control parameters, f1: defect rate, N defect : number of defective products, N total : total number of products, f2: cooling energy consumption rate, E cool : current energy consumption of the cooling system, E max : maximum energy consumption of the cooling system, f3: vibration energy proportion, VibEnergy: current vibration energy, Vib max : maximum vibration energy;

[0239] NSGA-III multi-objective optimization generates a Pareto frontier solution set.

[0240] In an embodiment of the present application, the defect rate is reduced by 18% using the NSGA-III multi-objective optimization method compared with the single-objective optimization method.

[0241] S2342: Dynamic adjustment;

[0242] The dynamic adjustment is performed by using reinforcement learning, specifically, priority experience replay (PER) is used in combination with a reward function to optimize the strategy, and a dynamic adjustment amount (Av, Ac) is generated to adapt the parameters to the dynamic changes of the actual system.

[0243] The formula of the priority experience replay (PER) is:

[0244]

[0245] wherein, P(i): the sampling probability of the ith experience, d i : the deviation between the predicted value and the actual value, e: error correction term, a: priority weight.

[0246] The reward function is

[0247]

[0248] wherein, r t : the reward value at time t, RiskLevel: risk level, DefectRate: defect rate, Threshold: defect rate threshold, EnergySave: energy saving ratio, Vibration: current vibration energy, Vib max : maximum vibration energy.

[0249] S2343: real-time system verification;

[0250] The real-time system executes f(v*+Av, c*+Ac), i.e. f(v, c), for verification, and feeds back the actual reward and experience data.

[0251] S235: digital twin verification and feedback;

[0252] The digital twin intelligent agent receives the "parameter adjustment strategy" data from the process optimization intelligent agent, runs the thermal / stress field coupling model, simulates the casting blank quality after parameter adjustment (verifies the effectiveness of the strategy), and feeds back the verification result (defect rate change) to the meta-cognition coordination intelligent agent.

[0253] S3: intelligent agent cluster operation and maintenance;

[0254] As shown in Figure 5 , the intelligent agent cluster operation and maintenance includes two links: the optimization iteration link continuously optimizes the intelligent agent to improve its operation effect; the performance monitoring link timely discovers the abnormal working condition of the intelligent agent to ensure the continuous and healthy work of the intelligent agent. The specific steps include:

[0255] S31: intelligent agent optimization iteration;

[0256] Through processing and analyzing the operation data of the intelligent agent, an optimized adjustment scheme of the intelligent agent is obtained and adjusted, and the optimization effect is verified in the actual task to form a closed loop of "monitoring -> evaluation -> optimization -> application". The specific steps include:

[0257] S311: Intelligent agent monitoring;

[0258] Real-time collection of operation data of the intelligent agent cluster as operation and maintenance evaluation data.

[0259] S312: Construction of evaluation database;

[0260] The collected operation data is stored and preprocessed, and is converted into structured indexes that can be evaluated and analyzed to build an evaluation database.

[0261] S313: Automatic optimization analysis;

[0262] Based on the indexes of the evaluation database, multi-objective optimization analysis is performed to obtain a parameter optimization adjustment scheme.

[0263] S314: Optimization adjustment;

[0264] The parameter optimization adjustment scheme is issued to the intelligent agent cluster to complete the parameter adjustment.

[0265] S32: Intelligent agent performance monitoring;

[0266] S321: Performance data collection and analysis;

[0267] The performance data of the intelligent agent cluster is collected for intelligent agent health evaluation.

[0268] Preferably, in the intelligent agent health evaluation, the trajectory deviation degree method is used to quantify the deviation degree of the actual trajectory and the expected trajectory of the intelligent agent by using the average value of the relative deviation, and the algorithm formula is:

[0269]

[0270] Wherein, Deviation: trajectory deviation degree, N: total number of trajectory points, : the i-th actual trajectory point, : the i-th expected trajectory point;

[0271] When the trajectory deviation degree exceeds 20%, an alarm of "intelligent agent behavior abnormal, possible fault" is triggered.

[0272] Preferably, in the intelligent agent health evaluation, a four-dimensional scoring model is used to quantify the overall performance of the intelligent agent by weighted summation of accuracy, time, cost and interpretability.

[0273] In an embodiment of the present application, different dimensions are set with corresponding weights, and the model formula is:

[0274]

[0275] Wherein, explainScore: comprehensive score, S acc : accuracy score, S time : time efficiency score, S cost : cost efficiency score, S explain : explainability score;

[0276] In the above formula, S explain can reflect the transparency of the agent decision-making process, such as the clarity of root cause positioning and the traceability of the decision-making path, and the specific score can be obtained by expert manual evaluation.

[0277] S322: health instrument display;

[0278] The health status of the agent cluster is converted into a visual form and displayed through the health instrument panel.

[0279] In an embodiment of the present application, the specific index information used in the operation and maintenance of the agent cluster is shown in Table 2.

[0280] Table 2:

[0281]

[0282] In a second aspect, as shown in Figure 4 , the present application provides a continuous casting quality management hardware system, which includes: a data acquisition component, an edge computing node, an intelligent processing and control component, a visualization and interaction component, a communication network component and a data storage component. The components in the hardware system work together to realize the steps of the method of the first aspect. Specifically:

[0283] 1. The data acquisition component includes multiple types of sensor acquisition units, specifically including:

[0284] (1) Parameter sensor; such as L1 sensor: used for collecting process parameter data in industrial production process;

[0285] (2) Visual acquisition device; such as industrial camera: used for collecting image or video data of production scene;

[0286] (3) Vibration sensing device; such as vibration sensor: used for collecting vibration signal data when the equipment is running.

[0287] 2. The edge computing node is in communication connection with the data acquisition component, and is used for real-time preprocessing of the collected raw data.

[0288] 3. Intelligent processing and control component, comprising:

[0289] (1) Intelligent agent server cluster: in communication connection with the edge computing node, deploying a functional intelligent agent, for deep analysis on pre-processed data;

[0290] Preferably, the intelligent agent server cluster comprises a GPU server;

[0291] (2) Industrial control subsystem: in communication connection with the edge computing node, receiving control instructions output by the intelligent agent server cluster or the edge computing node, for real-time adjustment on operation parameters of industrial production equipment.

[0292] 4. Visualization and interaction component: in communication connection with the intelligent agent server cluster, for displaying real-time data of production process, analysis results, equipment state and feedback of control instruction execution, supporting monitoring, intervention and decision-making by operating personnel;

[0293] In an embodiment of the present application, a 3D monitoring large screen is used as the visualization and interaction component.

[0294] 5. Communication network component: for realizing data transmission and instruction interaction among the data acquisition component, the edge computing node, the intelligent agent server cluster, the industrial control subsystem and the visualization terminal, guaranteeing information connectivity among all levels of the system; the industrial control subsystem comprises a PLC control system.

[0295] The communication network component can be in wired communication or wireless communication mode.

[0296] 6. Data storage component: for storing computer program files, original acquisition data, pre-processed data, analysis result data and system operation log, supporting historical backtracking of data, supplement of model training data and fault traceability analysis.

[0297] In an embodiment of the present application, parameters of the edge computing node, the GPU server, the industrial camera and the vibration sensor are shown in Table 3.

[0298] Table 3:

[0299]

[0300] In a third aspect, the present application provides a computer program product embodiment, which, when executed by a processor, realizes steps of the method of the first aspect.

[0301] Compared with the traditional scheme, the continuous casting quality management and control method, the hardware system and the program product have optimization and improvement in aspects of full process, multi-modal, intelligentization, collaboration and the like. Compared with a certain traditional scheme, the optimization effect obtained is as shown in Table 4.

[0302] Table 4:

[0303]

[0304] The economic benefits brought by the optimization are as shown in Table 5.

[0305] Table 5:

[0306]

[0307] The investment recovery period of the system modification is estimated as follows:

[0308] System investment 480 million yuan / year income 987.45 million yuan ≈ 0.49 years.

Claims

1. A continuous casting quality management method characterized by, Comprise the following steps: S1: data input and standardization; obtain L1 / L2 / MES data, combine real-time data pipeline technology on the basis of OPC UA, perform data conversion and standardization, realize data interaction of agent cluster and industrial field system, and complete multi-source data fusion; S2: central coordination; adopt meta-cognition coordination agent as the core to coordinate functional agents, realize task scheduling and state monitoring of the whole system, complete task decomposition, data scheduling, and closed-loop management and control of three key functions; the functional agents comprise: semantic analysis agent, data perception agent, defect prediction agent, root cause analysis agent, process optimization agent, and digital twin agent; The central coordination specifically comprises: S21: task decomposition; complex quality control tasks are decomposed into semantic analysis, data acquisition, and defect prediction subtasks, and then the semantic analysis agent is driven to understand the instructions; S22: data scheduling; the meta-cognition coordination agent dynamically schedules the data perception agent to collect or preprocess data according to the load and task priority of each functional agent; S23: closed-loop management and control; under the core coordination of the meta-cognition coordination agent, each functional agent cooperates through the following steps to realize the whole-cycle management and control of quality problems: Risk warning; the meta-cognition coordination agent judges according to the fusion data output by the data perception agent and sends risk warning information to the defect prediction agent; Defect prediction; the defect prediction agent predicts the casting blank defect based on multi-modal features using an LSTM-CNN hybrid model and outputs defect spatiotemporal features and risk levels; Root cause analysis; the root cause analysis agent receives spatiotemporal feature data from the defect prediction agent, traces the defect causes using a CART decision tree+Bayesian network, and generates root cause analysis results; Process optimization; the process optimization agent receives output data from the root cause analysis agent, generates process parameter adjustment strategies using an NSGA-III multi-objective algorithm; Digital twin verification and feedback; the digital twin agent receives the process parameter adjustment strategies output by the process optimization agent, runs a thermal / stress field coupling model, simulates the casting blank quality after parameter adjustment, and feeds back the verification results to the meta-cognition coordination agent; After the meta-cognition coordination agent receives the verification results in the "risk warning→defect prediction→root cause analysis→process optimization→digital twin verification and feedback" link, the cognitive iteration is completed; S3: agent cluster operation and maintenance; specifically comprising the following steps: S31: agent optimization iteration; the running data of the agent are processed and analyzed, and the agent is continuously optimized through the iteration link of "monitoring→evaluation→optimization→application" to improve the running effect; S32: agent performance monitoring; the performance data of the agent are collected and analyzed, and the health evaluation and health instrument display of the agent are performed.

2. The method of claim 1, characterized in that: Data collaboration is performed in the way of blackboard model knowledge sharing, a shared memory area stores process state tensors, structured tensors are used to integrate the whole process state, and data collaboration between agents is realized through shared memory+SQL-like query.

3. The method of claim 1, wherein: In the intelligent agent communication protocol, message format specification and performance guarantee mechanism setting are performed, specifically, the message format specification for inter-agent communication is defined to ensure that different intelligent agents can accurately obtain the information transmitted by each other; the performance guarantee mechanism setting includes: using Apache Kafka + RDMA network to limit the end-to-end data transmission delay; based on the RAFT consensus algorithm, the state synchronization of intelligent agents is realized, so that when some intelligent agents fail, the standby intelligent agents can quickly take over the work; and the SM4 national encryption algorithm is used to ensure the safety of industrial data.

4. The method of claim 1, wherein, The L1 / L2 / MES data is obtained, specifically including: obtaining real-time process data of sensors / actuators of the L1 basic automation layer; obtaining process model calculation results of the L2 process automation layer; obtaining production business data of the MES manufacturing execution system.

5. The method of claim 1, wherein, In the hub coordination, the meta-cognition coordination intelligent agent adopts a hierarchical task decomposition model: ； Wherein, A: task set, contains "prediction, trace, optimization" three kinds of task candidates; T: the target task to be determined finally, selected from the set A through subsequent calculation; : is the weight of the corresponding task ; : is the score function of the task ; In this model, dynamic allocation of weights is supported.

6. The method of claim 1, wherein, In the hub coordination, a real-time load balancing algorithm is used to guide dynamic scheduling of resources: ； Where, LoadIndex: load index, reflecting the system load state; QueueLen i : the length of the i-th queue; k: the total number of queues in the system; MaxQueue: the maximum value of the queue length; When the load index exceeds the threshold, the Kubernetes container elastic expansion and contraction is triggered.

7. The method of claim 1, wherein, The multi-scale spatio-temporal feature fusion algorithm architecture is used in the defect prediction step to accurately capture the time sequence dynamics and multi-scale correlation of process parameters, and to provide strong feature characteristics for defect probability prediction; The specific steps include: Dual-branch extraction: extract timing features F from the process parameter timing stream seq and spatial / multiscale features F cnn ; Attention fusion: modeling F seq cross-associations with F cnn , output attention-enhanced features F att ; Gating decision: adopt gating mechanism to dynamically adjust F seq and F cnn ; get the dynamic weighted sum F fusion ; Probability output: F fusion Input fully connected layer + Softmax classifier, output defect probability distribution.

8. The method of claim 7, wherein, The specific steps of the double-branch extraction include: inputting continuous time series data of the continuous casting process, and inputting these data into the time series feature branch and the spatial / multi-scale feature branch; In the timing feature branch, a BiLSTM-Pro module is constructed, a bidirectional LSTM+time convolution network TCN enhanced version is adopted to extract long and short term timing dependencies, and timing features F are output seq ; wherein the enhanced manner of the time convolution network TCN is that a convolution kernel covers a short time window to strengthen local timing patterns; In the space / multi-scale feature branch, a multi-scale CNN module is constructed, a CNN with different size convolution kernels is adopted, full-scale features from "short period details" to "long period trends" are covered, multi-scale space-time patterns are extracted, and space / multi-scale features F are output cnn .

9. The method of claim 8, wherein, in the spatial / multi-scale feature branch, an EfficientNet-B4 model is used as a spatial feature extraction layer to convert a large resolution image into a feature containing semantics and spatial structure, and the general formula is expressed as: ； wherein F spatial : output spatial feature, EfficientNet: Efficient Convolutional Neural Network, I 5120×3840 : input image data with a resolution of 5120x3840.

10. The method of claim 7, wherein, The specific steps of the attention fusion include: The spatio-temporal cross-attention module is constructed, specifically, a feature fusion layer is constructed by using an attention mechanism, cross correlation between F seq and F cnn is modeled by calculating the association weight of F seq and F cnn , cross correlation between the two is modeled, the most critical feature combination for defect prediction is strengthened, and attention-enhanced features F att are output.

11. The method of claim 7, wherein, The specific steps of the gating decision include: F att The input gate feature fusion module dynamically determines F seq The fusion weight of F cnn The gating mechanism formula is: ； where σ: sigmoid activation function, W g , b g : weight matrix and bias of the gate, [F seq , F cnn ]: feature concatenation of F seq and F cnn , ⊙: element-wise multiplication; F fusion : dynamic weighted sum of F seq and F cnn .

12. The method of claim 11, wherein, The gating decision also includes: The W in the gating mechanism is learned using an online learning engine. g The parameters are updated incrementally, and the generalized formula for the parameter update is expressed as: ； where θ t : model parameters at time t, η: learning rate, controlling the step size of gradient descent, : loss gradient on new data D new , : regularization term.

13. The method of claim 1, wherein, In the root cause analysis step, different types of data are respectively assigned weights, and different methods are used for processing, and the specific processing methods include: CART decision tree method is used to process process parameters; wavelet packet decomposition + energy entropy analysis method is used to process device vibration signals; BERT semantic embedding method is used to process operation logs; time series anomaly detection method is used to process environmental temperature and humidity.

14. The method of claim 13, wherein, In the root cause analysis step, a Bayesian network is used to dynamically calibrate the "root cause-defect" causal relationship, adjust the probability distribution using new samples, and real-time adjust the weight of the causal relationship; the formula of the Bayesian network is: ； where X: root cause event, P(Crack IX): conditional probability of "Crack" given root cause X, : count of root cause X occurrence and crack presence in the sample, : total count of root cause X occurrence in the sample.

15. The method of claim 1, wherein, The process optimization step also includes priority experience replay PER and reward function, and the specific steps include: a genetic algorithm NSGA-III multi-objective optimization model is used to generate a global optimal solution (v*, c*); a reinforcement learning method is used for dynamic adjustment to generate a dynamic adjustment amount (Δv, Δc); the real-time system executes f(v*+Δv, c*+Δc) for verification, and feeds back actual rewards and experience data.

16. The method of claim 15, wherein, The specific way of the NSGA-III multi-objective optimization is: ； wherein v: process parameter, c: environment, control parameter, : defect rate, : number of defective products, N total : total number of products, : cooling energy consumption rate, : current energy consumption of cooling system, E max : maximum energy consumption of cooling system, : vibration energy proportion, : current vibration energy, Vib max : maximum vibration energy.

17. The method of claim 15, wherein, The specific way of the reinforcement learning comprises: adopting a priority experience replay (PER) combined with a reward function to optimize the policy, and generating dynamic adjustment quantities (Δv, Δc); The formula of the priority experience replay (PER) is: ； Wherein, P(i): the i-th experience sampling probability, δ i : the deviation of the predicted value from the actual value, ε: error correction term, α: priority weight The reward function is: ； wherein, r t : reward value at time t, ΔRiskLevel: risk level, DefectRate: defect rate, Threshold: defect rate threshold, EnergySave: energy saving ratio, Vibration: current vibration energy, Vib max : maximum vibration energy.

18. The method of claim 1, wherein, The specific steps of the agent optimization iteration comprise: Agent monitoring: collecting running data of the agent cluster in real time as operation and maintenance evaluation data; Building an evaluation database: structuring and storing the collected running data and performing index preprocessing, converting the data into structured indexes that can be evaluated and analyzed, and building an evaluation database therefrom; Automatic optimization analysis: performing multi-objective optimization analysis based on the indexes of the evaluation database to obtain a parameter optimization adjustment scheme; Optimization adjustment: issuing the parameter optimization adjustment scheme to the agent cluster to complete parameter adjustment.

19. The method of claim 1, wherein, In the agent health assessment, a trajectory deviation algorithm is used to quantify the deviation degree of the actual trajectory of the agent from the expected trajectory by using the average value of the relative deviation, and the algorithm formula is: ； wherein Deviation: trajectory deviation degree, N: total number of trajectory points, : the i-th actual trajectory point, : the i-th expected trajectory point; When the trajectory deviation algorithm exceeds a set threshold, an agent behavior anomaly alarm is triggered.

20. The method of claim 1, wherein, In the agent health assessment, a four-dimensional scoring model is used to quantify the overall performance of the agent by weighted summation of the accuracy, time, cost, and interpretability of four dimensions; The interpretability dimension reflects the transparency of the agent decision-making process, including the clarity of root cause positioning and the traceability of decision path, and the specific score is obtained by manual evaluation by experts.

21. The method of claim 1, wherein, The health instrument display converts the health status of the agent cluster into a visual form; The health status indexes include: Task performance indexes, specifically including: Defect prediction accuracy: measured by a confusion matrix, with a target value of >95%; Root cause positioning error: measured by coordinate comparison, with a target value of <±1mm; Response performance indexes, specifically including: End-to-end delay: measured by a distributed tracing system, with a target value of <100ms; Throughput: measured by stress testing, with a target value of >800 QPS; Resource performance indexes, specifically including: CPU / GPU utilization: obtained by Prometheus monitoring, with a target value of 60%-80%; Memory usage: counted in container runtime, with a target value of <32GB / node.

22. A continuous-casting quality management hardware system characterized by comprising: It comprises: A data collection component, an edge computing node, an intelligent processing and control component, a visualization and interaction component, a communication network component, and a data storage component, and each part of the hardware system works cooperatively to realize the steps of the method of any one of claims 1-21; The communication network component is used to realize data transmission and instruction interaction, and ensure the information connectivity between the system levels; The data storage component is used to store computer program files and data files.

23. The hardware system of claim 22, wherein, The data collection component comprises: Parameter sensors: used to collect process parameter data in industrial production processes; Visual collection equipment: used to collect image or video data of production scenes; Vibration sensing devices: used to collect vibration signal data during equipment operation.

24. The hardware system of claim 22, wherein, The edge computing node is in communication connection with the data collection component and is used to perform real-time preprocessing on the collected raw data.

25. The hardware system of claim 22, wherein, The intelligent processing and control component comprises: An agent server cluster in communication connection with the edge computing node, deploying a functional agent for deep analysis on the preprocessed data; The agent server cluster comprises a GPU server; An industrial control subsystem in communication connection with the edge computing node, receiving the control instruction output by the agent server cluster or the edge computing node, and adjusting the operation parameters of the industrial production equipment in real time; The industrial control subsystem comprises a PLC control system.

26. The hardware system of claim 25, wherein, A visualization and interaction component in communication connection with the agent server cluster, used for displaying the real-time data of the production process, the analysis result, the equipment state and the control instruction execution feedback, and supporting the operation personnel to monitor, intervene and make decisions; The visualization and interaction component comprises a 3D monitoring large screen.

27. A computer program product, characterised in that, The computer program is executed by a processor to implement the steps of the method of any one of claims 1-21.

Images