Continuous casting quality control method based on meta-cognitive coordination architecture agent cluster

Abstract

The invention provides a continuous casting quality control method based on a meta-cognitive coordination architecture agent cluster, and relates to the technical field of ferrous metallurgy intelligent manufacturing. The continuous casting quality control method comprises the steps of data input and standardization, center coordination and intelligent agent cluster operation and maintenance. In the central coordination process, the meta-cognitive coordination agent serves as a core to coordinate six kinds of functional agents including a semantic analysis agent, a data perception agent, a defect prediction agent, a root cause analysis agent, a process optimization agent and a digital twinborn agent, and task scheduling and state monitoring of the whole system are achieved. Three key functions of task decomposition, data scheduling and closed-loop management and control are completed; in the closed-loop management and control step, risk early warning, defect prediction, root cause analysis, process optimization and digital twinborn verification and feedback are carried out for continuous casting quality. Compared with a traditional scheme, optimization is carried out in the aspects of whole process, multiple modes, intelligence, collaboration and the like, the management efficiency is improved, and economic benefits can also be increased.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent manufacturing of steel metallurgy, and in particular to a continuous casting quality control method based on a metacognitive coordination architecture intelligent agent cluster. Background Art

[0002] A Chinese patent (publication number CN 119025412A), published on November 26, 2024, discloses a method for processing an intelligent agent. When an intelligent agent needs to be inspected, a target intelligent agent to be processed is identified from multiple intelligent agents; at least one target test case is obtained for simulating the detection of the target intelligent agent; and based on the at least one target test case, the target intelligent agent is simulated and tested, and a test result is obtained. The primary intelligent agent application in this technology involves virtual assistants, which are generally not effectively applicable to intelligent assistants in industrial environments.

[0003] A Chinese patent (publication number CN119903172A), published on April 29, 2025, discloses a large language model training method for generating intelligent agents. This method uses the agent's proxy template to filter generated proxy templates and train the large language model, forming an automatic iterative closed loop. This method optimizes the agent training corpus and continuously improves model performance. Automatic feedback and adjustment enable the agent to continuously learn and adapt, improving the model's stability and reliability. This technology primarily uses a generative large language model to close the loop for the corresponding intelligent agent and does not involve corresponding intelligent agent collaboration or industrial applications.

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

[0005] (1) Inefficient defect tracing, including: manual troubleshooting across MES / L2 / inspection systems, which takes a long time (>48 hours), and defect attribution accuracy is less than 65%; traditional numerical models (such as the secondary cooling zone heat transfer coefficient optimization model) rely on empirical formulas and ignore the coupling effects of multiple factors.

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

[0007] (3) Serious data silo problems, including: L1 / L2 / MES system data delay > 72 hours, utilization rate < 25%, etc. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this paper proposes a continuous casting quality intelligent agent system that integrates deep learning, reinforcement learning, and a large language model (LLM). This system implements a closed-loop control system encompassing process perception, defect prediction, root cause tracing, dynamic optimization, and digital twin verification, enabling full-lifecycle quality control of the continuous casting process, from molten steel pouring to billet cutting. By leveraging a three-dimensional semantic parsing engine, a multimodal feature fusion model, a dual causal reasoning engine, and an adaptive optimization algorithm, it addresses industry pain points in traditional continuous casting quality control, such as data silos, inefficient defect tracing, and delayed process adjustments.

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

[0010] S1: Data input and standardization: Acquire L1 / L2 / MES data and, based on OPC UA and combined with real-time data pipeline technology, perform data conversion and standardization to enable data interaction between intelligent agent clusters and industrial field systems, completing multi-source data fusion.

[0011] S2: Central coordination: Using a metacognitive coordination agent as the core, it coordinates functional agents to implement task scheduling and status monitoring for the entire system, completing the three key functions of task decomposition, data scheduling, and closed-loop control. The functional agents include: semantic parsing agent, data perception agent, defect prediction agent, root cause analysis agent, process optimization agent, and digital twin agent.

[0012] The central coordination specifically includes:

[0013] S21: Task decomposition: Decompose complex quality control tasks into subtasks such as semantic parsing, data collection, and defect prediction, and then use semantic parsing agents to drive instruction understanding;

[0014] S22: Data Scheduling: The metacognitive coordination agent dynamically schedules the data perception agent to collect or pre-process data based on the load and task priority of each functional agent.

[0015] S23: Closed-loop control: Under the core coordination of the metacognitive coordination agent, various functional agents work together to achieve full-cycle control of quality issues. Specific steps include:

[0016] S231: Risk warning: The metacognitive coordination agent makes a judgment based on the fused data output by the data perception agent and sends a risk warning letter to the defect prediction agent.

[0017] S232: Defect prediction: The defect prediction agent uses a LSTM-CNN hybrid model based on multimodal features to predict slab defects and outputs the spatiotemporal characteristics and risk level of the defects.

[0018] S233: Root cause analysis: The root cause analysis agent receives the spatiotemporal feature data from the defect prediction agent, uses the CART decision tree + Bayesian network to trace the cause of the defect, and generates a root cause analysis result.

[0019] S234: Process optimization: The process optimization agent receives the output data from the root cause analysis agent and generates a process parameter adjustment strategy using the 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 the thermal / stress field coupling model, simulates the quality of the cast billet after parameter adjustment, and feeds back the verification results to the metacognitive coordination agent.

[0021] S3: Agent cluster operation and maintenance; specifically includes the following steps:

[0022] S31: Agent optimization iteration: Process and analyze the agent's operating data, and continuously optimize the agent through the iterative chain of "monitoring → evaluation → optimization → application" to improve its operating performance;

[0023] S32: Intelligent agent performance monitoring; collect and analyze performance data of the intelligent agent, and perform intelligent agent health assessment and health instrument display.

[0024] As a further improvement of the present invention, data collaboration is carried out by adopting the knowledge sharing method of blackboard model, the process state tensor is stored in the shared memory area, the whole process state is integrated with the structured tensor, and the data collaboration between intelligent agents is achieved through shared memory + SQL-like query.

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

[0026] Define the message format specifications for communication between agents to ensure that different agents can accurately obtain the information transmitted to each other;

[0027] The performance guarantee mechanism settings include: using Apache Kafka + RDMA network to limit end-to-end data transmission delays; based on the RAFT consensus algorithm, achieving intelligent agent state synchronization, so that when some intelligent agents fail, the backup intelligent agents can quickly take over the work; using the SM4 national encryption algorithm to ensure industrial data security.

[0028] As a further improvement of the present invention, the obtaining of L1 / L2 / MES data specifically includes:

[0029] Obtaining L1 basic automation layer: real-time process data from sensors / actuators;

[0030] Obtain the L2 process automation layer: process model calculation results;

[0031] Obtain production business data from the MES manufacturing execution system.

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

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

[0034]

[0035] Where A is the task set, which includes three types of task candidates: prediction, tracing, and optimization. T is the target task to be determined, which is selected from set A through subsequent calculations. :It corresponds to the task The weight of :It is a task The scoring function of

[0036] In this model, support for weights Perform dynamic allocation.

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

[0038] Use real-time load balancing algorithms to guide dynamic resource scheduling:

[0039]

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

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

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

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

[0044] As a further improvement of the present invention, the defect prediction step adopts a multi-scale spatiotemporal feature fusion algorithm architecture to accurately capture the temporal dynamics and multi-scale correlation of process parameters, providing strong characterization features for defect probability prediction; the specific steps include:

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

[0046] Attention Fusion: Modeling F seq With F cnn Cross-correlation of the output attention enhancement feature F att ;

[0047] Gating decision: Use gating mechanism to dynamically adjust F seq With F cnn The fusion weight of the two is obtained by dynamic weighted sum F fusion ;

[0048] Probability output: F fusion Input the fully connected layer + Softmax classifier and output the defect probability distribution.

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

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

[0051] In the temporal feature branch, a BiLSTM-Pro module is constructed, which uses a bidirectional LSTM+time convolutional network TCN enhanced version to extract long-term and short-term temporal dependencies and output temporal features F seq Among them, the enhancement method of the temporal convolutional network (TCN) is to use convolution kernels to cover short time windows and strengthen local temporal patterns;

[0052] In the spatial / multi-scale feature branch, a multi-scale CNN module is constructed, using CNNs with different sizes of convolution kernels to cover the full-scale features from "short-term details" to "long-term trends", extracting multi-scale spatiotemporal patterns and outputting spatial / multi-scale features F cnn .

[0053] As a further improvement of the present invention, in the spatial / multi-scale feature branch, the EfficientNet-B4 model is used as the spatial feature extraction layer to convert the large-resolution image into features containing semantics and spatial structures. The generalized formula is expressed as follows:

[0054]

[0055] Among them, F spatial : Spatial features of output, EfficientNet: Efficient Convolutional Neural Network, I 5120×3840 : The input image data has a resolution of 5120×3840.

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

[0057] Construct a spatiotemporal cross attention module, specifically using the attention mechanism to construct a feature fusion layer, by calculating F seq With F cnn The association weight of the two is used to model the cross-correlation between them, strengthen the most critical feature combination for defect prediction, and output the attention enhancement feature F att .

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

[0059] F att Input gated feature fusion module, dynamically determines F through the gate coefficient g seq With F cnn The fusion weight of ; the gating mechanism formula is:

[0060]

[0061] Where, σ: sigmoid activation function, W g 、b g : The 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 of .

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

[0063] An online learning engine is used to learn the W g The parameters are updated incrementally, and the generalized formula for the parameter update is expressed as:

[0064]

[0065] Among them, θ t : Model parameters at time t, η: learning rate, which controls the step size of gradient descent, :New data D new The 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 weighted and processed using different methods. The specific processing methods include:

[0067] The CART decision tree method was used to process the process parameters;

[0068] The vibration signal of the equipment is processed by wavelet packet decomposition + energy entropy analysis;

[0069] The operation logs are processed using the BERT semantic embedding method;

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

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

[0072]

[0073] Where X is the root cause event, P(crack|X) is the conditional probability of the root cause X to the “crack”, : The number of cracks that occurred due to root cause X in the sample, : The total number of root cause X occurrences in the sample.

[0074] As a further improvement of the present invention, the process optimization step further includes a priority experience replay PER and a reward function, and the specific steps include:

[0075] The genetic algorithm NSGA-III multi-objective optimization model is used to generate the global optimal solution (v*, c*);

[0076] Use reinforcement learning to make dynamic adjustments and generate dynamic adjustment values ​​(Δv, Δc);

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

[0078] As a further improvement of the present invention, the specific method of the NSGA-III multi-objective optimization is:

[0079]

[0080] Among them, v: process parameters, c: environment, 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 ratio, VibEnergy: Current vibration energy, Vib max: Maximum vibration energy.

[0081] As a further improvement of the present invention, the specific method of reinforcement learning includes: using priority experience replay PER in combination with a reward function optimization strategy to generate a dynamic adjustment amount (Δv, Δc);

[0082] The formula for the priority experience replay PER is:

[0083]

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

[0085] The reward function is:

[0086]

[0087] Among them, 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 present invention, the specific steps of the agent optimization iteration include:

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

[0090] Build an evaluation database: Structure the collected operating data and perform indicator preprocessing to convert it into structured indicators that can be evaluated and analyzed, thereby building an evaluation database;

[0091] Automatic optimization analysis: Based on the indicators of the evaluation database, perform multi-objective optimization analysis to obtain parameter optimization adjustment plans;

[0092] Optimization and adjustment: The parameter optimization and adjustment plan is sent to the intelligent agent cluster to complete the parameter adjustment.

[0093] As a further improvement of the present invention, in the intelligent agent health assessment, a trajectory deviation algorithm is used to quantify the degree of deviation between the actual trajectory of the intelligent agent and the expected trajectory using the average value of the relative deviation. The algorithm formula is:

[0094]

[0095] Among them, 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 the set threshold, an abnormal behavior alarm of the intelligent agent is triggered.

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

[0098] The explainability dimension reflects the transparency of the intelligent agent's decision-making process, including the clarity of root cause location and the traceability of the decision-making path. The specific score is obtained through manual evaluation by experts.

[0099] As a further improvement of the present invention, the health meter displays the health status of the agent cluster in a visual form; its health status indicators include:

[0100] Mission effectiveness indicators include:

[0101] Defect prediction accuracy: measured using confusion matrix calculation, with a target value of >95%;

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

[0103] Response performance indicators, including:

[0104] End-to-end latency: measured using a distributed tracing system, with a target value of <100ms.

[0105] Throughput: measured using stress testing, with a target value of >800 QPS;

[0106] Resource efficiency indicators include:

[0107] CPU / GPU utilization: Prometheus monitoring is used to obtain the target value of 60%-80%;

[0108] Memory usage: Statistics are collected while the container is running, with a target value of <32 GB per node.

[0109] In a second aspect, the present invention provides a continuous casting quality control 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, wherein the various components of the hardware system work together to implement the steps of the method described in the first aspect;

[0110] The communication network component is used to realize data transmission and instruction interaction, and ensure information connectivity between various levels of the system;

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

[0112] As a further improvement of the present invention, the data acquisition component includes:

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

[0114] Visual acquisition equipment: used to collect image or video data of production scenes;

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

[0116] As a further improvement of the present invention, the edge computing node is communicatively connected to the data acquisition component for performing real-time preprocessing on the collected raw data.

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

[0118] Agent server cluster: communicates with the edge computing nodes and deploys functional agents for in-depth analysis of pre-processed data;

[0119] The intelligent agent server cluster includes a GPU server;

[0120] Industrial control subsystem: communicates with the edge computing node, receives control instructions from the intelligent server cluster or edge computing node, and adjusts the operating parameters of industrial production equipment in real time;

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

[0122] As a further improvement of the present invention, the visualization and interaction component is communicatively connected to the intelligent server cluster to display real-time data, analysis results, equipment status and control instruction execution feedback of the production process, supporting operators in monitoring, intervention and decision-making;

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

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

[0125] Compared with existing technologies, the present invention uses a metacognitive coordination agent as its core, forming a full-cycle closed-loop logic from "data collection" to "cognitive iteration." By constructing a full-process closed-loop control system of "process perception-defect prediction-root cause tracing-dynamic optimization-digital twin verification," the present invention achieves full-lifecycle quality control of the continuous casting process, from molten steel casting to ingot cutting. Through a three-dimensional semantic parsing engine, a multimodal feature fusion model, a dual causal reasoning engine, and an adaptive optimization algorithm, the present invention addresses industry pain points in traditional continuous casting quality control, such as data silos, inefficient defect tracing, and delayed process adjustments. BRIEF DESCRIPTION OF THE DRAWINGS

[0126] Figure 1 This is a topological structure diagram of an intelligent agent cluster for continuous casting quality control disclosed in the present invention.

[0127] Figure 2 This is the architecture diagram of the multi-scale spatiotemporal feature fusion algorithm for the defect prediction agent.

[0128] Figure 3 Architecture diagram of the hybrid optimization algorithm for the process optimization agent.

[0129] Figure 4 This is a topological structure diagram of the continuous casting quality control hardware system disclosed in the present invention.

[0130] Figure 5 This is a diagram of the intelligent agent cluster operation and maintenance framework. DETAILED DESCRIPTION

[0131] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the present invention will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, wherein steps S1, S2... in the embodiments described in the present invention do not limit the only execution steps of the present invention; the various models, simulation environments, and software described in the present invention are not the only way to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0132] In the present invention, a computer device / equipment / system refers to a computer-related entity, such as hardware, a combination of hardware and software, software, or software in execution. Specifically, for example, software includes, but is not limited to, a process running on a processor, a processor, an object, executable software, an execution thread, a program, and / or a computer. Furthermore, an application or script running on a server, or a server, can also be software. One or more software programs can be in an execution process and / or thread, and software can be localized on a single computer and / or distributed between two or more computers, and can be executed from various computer-readable media.

[0133] It should be noted that, unless there is any conflict, the embodiments and features in the embodiments of this application can be combined with each other.

[0134] In a first aspect, the present invention provides an embodiment of a continuous casting quality management method, such as Figure 1 As shown in the figure, this method uses a metacognitive coordination agent as the core to form a full-cycle closed loop of "data collection → defect prediction → root cause tracing → process optimization → digital twin verification → feedback optimization". Its basic logical structure is as follows:

[0135] Data input: L1 / L2 / MES data flows into the data perception agent via OPC UA, completing multi-source fusion;

[0136] Defect prediction: Fusion data drives the model to output defect risks and trigger risk warnings;

[0137] Root cause tracing: Based on defect characteristics, dual engines (CART + Bayesian) are used to locate the root cause of the problem;

[0138] Process optimization: Generate parameter adjustment strategies based on root causes;

[0139] Physical verification: Digital twin simulation parameter adjustment effect to verify the effectiveness of the strategy;

[0140] Cognitive iteration: Verification results are fed back to the metacognitive coordination agent to optimize subsequent task scheduling and decision-making logic.

[0141] The present invention has the following special designs in terms of data collaboration and communication:

[0142] (1) In terms of data collaboration, the present invention adopts blackboard model knowledge sharing, stores process state tensors in shared memory, integrates all process states with structured tensors, and opens up data collaboration among intelligent agents through shared memory + SQL-like query.

[0143] The expression of the process state tensor is as follows:

[0144]

[0145] Among them, T meniscus : liquid surface temperature; v cast : throwing speed; : cooling pressure gradient; f vib : vibration frequency; σ stress : stress distribution; δ crack : crack size; ρ inclusion : inclusion density;

[0146] In one embodiment of the present invention, the cross-agent SQL-like query method is:

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

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

[0149] First, define the message format specification for communication between agents to ensure that different agents can accurately obtain the information transmitted to each other; the message format specification defined in one embodiment of the present invention 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] / / Levels 1-5, level 3 and above trigger real-time response;

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

[0156] "defect_type": "Surface longitudinal cracks",

[0157] "risk_level": 4,

[0158] "confidence": 0.92,

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

[0160] Second, in terms of performance assurance mechanism, the following three aspects are designed to ensure the real-time performance and stability of the control method of the present invention:

[0161] Low-latency transmission: Using Apache Kafka + RDMA network to ensure end-to-end data transmission latency ≤ 50ms;

[0162] Fault-tolerant design: Based on the RAFT consensus algorithm, it achieves agent state synchronization, ensuring that if an agent fails, the backup agent can quickly take over the work, and the failover time is less than 200ms;

[0163] Secure encryption: Using SM4 national encryption algorithm to ensure industrial data security.

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

[0165] S1: Data input and standardization: L1 / L2 / MES data is transmitted and standardized via the OPC UA real-time data pipeline, and input into the data perception agent to complete multi-source fusion;

[0166] S11: Collect data from L1 / L2 / MES systems; L1 / L2 / MES systems are the data "source" of industrial sites. Specifically,

[0167] L1 (Basic Automation): Collecting real-time process data from sensors / actuators (e.g., 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): Provides production business data (such as steel grade, casting plan, and quality traceability records).

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

[0171] OPC UA is a unified architecture for open platform communications that uses a publish-subscribe model to ensure real-time data transmission (latency <5ms) and reliability (redundant transmission, data encryption). This invention combines real-time data pipeline technology with OPC UA to perform data conversion and standardization, enabling data interaction between intelligent agent clusters and industrial field systems.

[0172] S2: central coordination;

[0173] This invention uses a metacognitive coordination agent as its core to coordinate six types of functional agents (semantic parsing agent, data perception agent, defect prediction agent, root cause analysis agent, process optimization agent, and digital twin agent) to achieve task scheduling and status monitoring for the entire system, completing the three key functions of task decomposition, data scheduling, and closed-loop control.

[0174] The metacognitive coordination agent adopts a hierarchical task decomposition model:

[0175]

[0176] Where A is the task set, which includes three types of task candidates: prediction, tracing, and optimization. T is the target task to be determined, which is selected from set A through subsequent calculations. :It corresponds to the task The weight of :It is a task The scoring function of

[0177] In this model, weights can be dynamically assigned. For example, when the defect risk is greater than 3, the prediction task corresponding to .

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

[0179]

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

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

[0182] S21: Task decomposition: breaking down complex quality control tasks (such as "analyzing the cause of cracks in sector 8") into subtasks such as semantic parsing, data collection, and defect prediction, and then using semantic parsing agents to drive instruction understanding.

[0183] In one embodiment of the present invention, the semantic parsing agent receives a "task decomposition" instruction from the metacognitive coordination agent, requiring it to parse industrial natural language instructions (such as "query the number of times the liquid level fluctuation of casting machine No. 3 exceeded the limit yesterday") and map them into system executable parameters; after parsing, the semantic parsing agent outputs a standardized query instruction.

[0184] S22: Data Scheduling: The metacognitive coordination agent dynamically schedules the data perception agent to collect / preprocess data (e.g., prioritize the acquisition of vibration spectra in high-risk areas) based on the load and task priority of each agent.

[0185] In one embodiment of the present invention, the data-aware agent receives a "data scheduling" instruction from the metacognitive coordination agent, requiring the extraction of multi-source data (process, image, vibration) from the OPC UA pipeline and the performance of standardization and feature extraction; after processing, the data-aware agent outputs a fused spatiotemporal feature matrix.

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

[0187] S231: Risk warning;

[0188] The metacognitive coordination agent makes judgments based on 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 uses the LSTM-CNN hybrid model based on multimodal features to predict ingot defects (type, risk, location) and outputs the spatiotemporal characteristics and risk level of the defects.

[0191] like Figure 2 As shown, the present invention further improves the "multi-scale spatiotemporal feature fusion algorithm architecture" of the defect prediction agent: through dual-branch parallel extraction + spatiotemporal attention fusion + gated dynamic decision-making, the temporal dynamics and multi-scale correlation of process parameters are accurately captured, providing strong characterization features for defect probability prediction.

[0192] Compared with the existing technologies that perform "single branch extraction", "simple splicing" and "fixed weight fusion" on features, which cannot adapt to complex industrial scenarios, the improvements of the present invention have achieved substantial technological breakthroughs, namely: (1) through the dual-branch design, it covers the "time trend of process parameters" and "multi-scale correlation between parameters / time windows" at the same time, avoiding the information loss of single-dimensional features; (2) through the attention mechanism, it dynamically captures the synergistic correlation of the two types of features, allowing the model to pay more attention to "the feature combination that is most indicative of defects"; (3) through the gating coefficient, it realizes "dynamic adjustment of feature weights", and automatically determines "whether to rely more on temporal features or spatial features" according to the current input process status, so that the fusion strategy is more in line with the dynamics of actual production.

[0193] The specific steps of the improved defect prediction include:

[0194] S2321: dual branch extraction;

[0195] The process parameter time series stream is used as data input to provide continuous time series data of the continuous casting process, such as the parameter sequence that changes with time, such as casting speed, cooling water volume, and liquid level height; this data is input into the time series feature branch and the spatial / multi-scale feature branch for subsequent processing.

[0196] Time series feature branch: Build the BiLSTM-Pro module, using the enhanced version of bidirectional LSTM + temporal convolutional network (TCN) to extract long-term and short-term time series dependencies and output the time series feature F seq For example, it can capture the “cumulative effect” of liquid level fluctuations (such as the oscillation trend within 30 minutes). The temporal convolutional network (TCN) is enhanced by covering a short time window (such as 5 time steps) with a convolution kernel to strengthen the local time series pattern (such as the oscillation period of the liquid level every 2 minutes).

[0197] The generalized formula of the temporal feature extraction layer in the BiLSTM-Pro module is expressed as:

[0198]

[0199] in, : The final feature after fusion, :Use the current input x t and the previous state Extracting temporal dependency features, : Use the past k steps to the current input x t-k:t Extract local temporal features, k: temporal convolution kernel size, k∈{3,5,7}, used to capture short / medium / long term dependencies respectively, : Element-wise addition.

[0200] Spatial / multi-scale feature branch: Construct a multi-scale CNN module, using CNNs with different convolution kernel sizes (3×3, 5×5, 7×7, etc.), covering full-scale features from "short-term details" to "long-term trends", extracting multi-scale spatiotemporal patterns, and outputting spatial / multi-scale features F cnn For example, it can capture the “abnormal distribution” of cooling water volume in sectors 7-9 (such as local overcooling), and the fluctuation difference in casting speed between 30 minutes and 2 hours.

[0201] In one embodiment of the present invention, the EfficientNet-B4 model is used as the spatial feature extraction layer to convert large-resolution images into features containing semantics and spatial structures. The generalized formula is:

[0202]

[0203] Among them, F spatial : Spatial features of output, EfficientNet: Efficient Convolutional Neural Network, I 5120×3840 : The input image data has 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 between the detection box and the actual inclusion area is ≥50%, the average detection accuracy of all categories (such as different types of inclusions) is as high as 97%.

[0205] S2322: Attention Fusion;

[0206] Construct a spatiotemporal cross attention module, specifically using the attention mechanism to construct a feature fusion layer, by calculating F seq With F cnn The correlation weight of the two is modeled, and the cross-correlation between the two is strengthened to strengthen the most critical feature combination for defect prediction. For example, by calculating the correlation between "pulling speed mutation" and "uneven distribution of cooling water", the synergistic defect risk of the two is obtained.

[0207] The algorithm formula is:

[0208]

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

[0210] In this formula, the weight matrix obtained by the softmax function is used for "weighted fusion" F seq With F cnn Two types of features, output attention enhancement feature Fatt .

[0211] S2323: Gating decision;

[0212] Construct a gated feature fusion module, use the gated mechanism to build a decision layer, and dynamically adjust the fusion weights of the two types of features. att Input gated feature fusion module, dynamically determines F through the gate coefficient g seq With F cnn The fusion weight of the gate mechanism is:

[0213]

[0214] Where, σ: sigmoid activation function, W g 、b g : The gated weight matrix and bias, [F seq , F cnn ]:F seq With F cnn Perform feature concatenation, ⊙: element-wise multiplication; F fusion :F seq With F cnn Dynamic weighted sum of

[0215] According to the dynamic decision logic of the above formula: when the liquid level fluctuates violently (the timing feature dominates the defect risk), g→1, and the fusion feature is more inclined to F seq When the cooling water distribution is abnormal (spatial features dominate the defect risk), g→0, and the fusion feature is more inclined to F cnn .

[0216] Furthermore, an online learning engine is used to learn the W g Parameters are updated incrementally: incremental training is performed every 24 hours, and the model iteration cycle is less than 30 minutes. The generalized formula for parameter update is expressed as:

[0217]

[0218] Among them, θ t : Model parameters at time t, η: learning rate, which controls the step size of gradient descent, :New data D new The loss gradient on : Regularization term.

[0219] S2324: probability output;

[0220] Construct a defect probability output module and integrate the feature F after the gated feature fusion fusionInput the fully connected layer + Softmax classifier and output the defect probability distribution; for example, output prediction results such as "probability of surface longitudinal crack 0.92" and "probability of internal inclusion 0.15".

[0221] In one embodiment of the present invention, the improved defect prediction method proposed in the present invention achieves a leapfrog improvement in defect prediction accuracy, better supporting the closed-loop control of continuous casting quality. Specifically, the No. 1 continuous casting machine of a certain steel plant is used for verification: (1) In terms of accuracy improvement: the accuracy of surface crack prediction is improved from 89% to 97%, and the false alarm 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 accuracy and recall rate); (2) In terms of closed-loop support: accurate defect probability output provides reliable defect spatiotemporal characteristics for the subsequent "root cause analysis intelligent agent" (such as "cracks occur in sector 7, corresponding to a casting speed fluctuation period of 2 minutes"), and provides a basis for the parameter adjustment decision of the "process optimization intelligent agent" (such as "targeted reduction of cooling water volume in sector 7").

[0222] S233: Root cause analysis;

[0223] The root cause analysis agent receives the "spatiotemporal feature" data from the defect prediction agent, uses the CART decision tree + Bayesian network to trace the cause of the defect, and generates root cause heat maps, causal maps, root cause weights and other results.

[0224] Furthermore, the present invention upgrades the root cause analysis agent and constructs a multimodal evidence fusion system. Different methods are used to process different types of input data and different weights are assigned. The specific corresponding relationships are shown in Table 1.

[0225] Table 1:

[0226]

[0227] The upgrade and improvement also includes the use of Bayesian networks to dynamically calibrate the "root cause-defect" causal relationship, using new samples to adjust the probability distribution and adjust the weight of the causal relationship in real time. The formula of the Bayesian network is:

[0228]

[0229] Where X is the root cause event, P(crack|X) is the conditional probability of the root cause X to the “crack”, : The number of cracks that occurred due to root cause X in the sample, : The total number of root cause X occurrences in the sample;

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

[0231] S234: Process optimization;

[0232] The process optimization agent receives the "cause-effect graph" 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] like Figure 3 As shown, the present invention further improves the process optimization agent, using NSGA-III, Prioritized Experience Replay (PER), and a reward function to collaborate with the real-time system to achieve 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] The genetic algorithm NSGA-III multi-objective optimization model is used to generate the global optimal solution (v*, c*) as the initial optimal solution.

[0236] The specific method of NSGA-III multi-objective optimization is:

[0237]

[0238] Among them, v: process parameters, c: environment, 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 ratio, VibEnergy: Current vibration energy, Vib max : Maximum vibration energy;

[0239] NSGA-III multi-objective optimization generates Pareto front solution sets.

[0240] In one embodiment of the present invention, the defect rate is further reduced by 18% by using the NSGA-III multi-objective optimization method compared to the single-objective optimization method.

[0241] S2342: Dynamic adjustment;

[0242] Dynamic adjustment is performed using reinforcement learning. Specifically, Prioritized Experience Replay (PER) is used in conjunction with the reward function optimization strategy to generate dynamic adjustment values ​​(Δv, Δc) to adapt the parameters to the dynamic changes of the actual system. Specifically,

[0243] The formula for Prioritized Experience Replay (PER) is:

[0244]

[0245] Among them, P(i): sampling probability of the i-th experience, δ i : The deviation between the predicted value and the actual value, ε: error correction term, α: priority weight.

[0246] The reward function is

[0247]

[0248] Among them, 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.

[0249] S2343: Real-time system verification;

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

[0251] S235: Digital Twin Verification and Feedback;

[0252] The digital twin agent receives the "parameter adjustment strategy" data from the process optimization agent, runs the thermal / stress field coupling model, simulates the quality of the ingot after parameter adjustment (verifies the effectiveness of the strategy), and feeds back the verification results (defect rate changes) to the metacognitive coordination agent.

[0253] S3: Agent cluster operation and maintenance;

[0254] like Figure 5 As shown in the figure, the operation and maintenance of the intelligent agent cluster is divided into two links: the optimization and iteration link continuously optimizes the intelligent agent to improve its operating performance; the performance monitoring link promptly detects abnormal working conditions of the intelligent agent to ensure its continued healthy operation. The specific steps include:

[0255] S31: Agent optimization iteration;

[0256] By processing and analyzing the agent's operating data, we can obtain and adjust the agent's optimization plan, verify the optimization effect in actual tasks, and form a closed loop of "monitoring → evaluation → optimization → application". The specific steps include:

[0257] S311: Agent Monitoring;

[0258] Collect the operation data of the intelligent agent cluster in real time as operation and maintenance evaluation data.

[0259] S312: Building an evaluation database;

[0260] The collected operating data is structured and stored, and indicators are pre-processed to convert them into structured indicators that can be evaluated and analyzed, thereby building an evaluation database.

[0261] S313: Automatic optimization analysis;

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

[0263] S314: Optimization and adjustment;

[0264] The parameter optimization and adjustment plan is sent to the intelligent agent cluster to complete the parameter adjustment.

[0265] S32: Agent performance monitoring;

[0266] S321: Performance data collection and analysis;

[0267] Collect performance data of the intelligent agent cluster and conduct intelligent agent health assessment.

[0268] Preferably, in the health assessment of the intelligent agent, a trajectory deviation algorithm is used to quantify the degree of deviation between the actual trajectory of the intelligent agent and the expected trajectory using the average value of the relative deviation. The algorithm formula is:

[0269]

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

[0271] When the trajectory deviates from the calculated value by more than 20%, an alarm is triggered: "the intelligent agent behaves abnormally and there may be a fault."

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

[0273] In one embodiment of the present invention, corresponding weights are set for different dimensions, and the model formula is:

[0274]

[0275] Among them, 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 It can reflect the transparency of the intelligent agent's decision-making process, such as the clarity of root cause location and the traceability of the decision-making path. The specific score can be obtained by manual evaluation by experts.

[0277] S322: Health meter display;

[0278] Convert the health status of the agent cluster into a visual form and display it through the health dashboard.

[0279] In one embodiment of the present invention, the specific indicator information used in the operation and maintenance of the intelligent agent cluster is shown in Table 2.

[0280] Table 2:

[0281]

[0282] Second, as Figure 4 As shown, the present invention provides a continuous casting quality control 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 implement the steps of the method described in the first aspect. Specifically:

[0283] 1. Data acquisition components, including multiple types of sensor acquisition units, specifically including:

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

[0285] (2) Visual acquisition equipment, such as industrial cameras: used to collect images or video data of production scenes;

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

[0287] 2. An edge computing node, which is in communication with the data acquisition component and is used to perform real-time preprocessing on the collected raw data.

[0288] 3. Intelligent processing and control components, including:

[0289] (1) Agent server cluster: communicates with the edge computing nodes and deploys functional agents for in-depth analysis of pre-processed data;

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

[0291] (2) Industrial control subsystem: communicates with the edge computing node, receives control instructions output by the intelligent server cluster or edge computing node, and adjusts the operating parameters of industrial production equipment in real time.

[0292] 4. Visualization and interaction component: Communicates with the intelligent server cluster to display real-time data, analysis results, equipment status, and control instruction execution feedback of the production process, supporting operators in monitoring, intervention, and decision-making;

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

[0294] 5. Communication network component: used to realize data transmission and command interaction between the data acquisition component, edge computing nodes, intelligent server cluster, industrial control subsystem and visualization terminal, and ensure information connectivity between all levels of the system; the industrial control subsystem includes a PLC control system.

[0295] The communication network components can use wired communication or wireless communication.

[0296] 6. Data storage component: used to store computer program files, original collected data, preprocessed data, analysis result data and system operation logs, supporting historical data backtracking, model training data supplementation and fault tracing analysis.

[0297] The parameters of the edge computing node, GPU server, industrial camera, and vibration sensor used in one embodiment of the present invention are shown in Table 3.

[0298] Table 3:

[0299]

[0300] In a third aspect, the present invention provides an embodiment of a computer program product, which implements the steps of the method described in the first aspect when executed by a processor.

[0301] Compared with traditional solutions, the continuous casting quality control method, hardware system, and program product proposed in this invention have optimization improvements in terms of full process, multimodality, intelligence, and collaboration. Compared with a certain traditional solution, the optimization effect obtained is shown in Table 4.

[0302] Table 4:

[0303]

[0304] The economic benefits brought by these optimizations are shown in Table 5.

[0305] Table 5:

[0306]

[0307] The investment recovery period for system transformation is estimated as follows:

[0308] The system investment is 4.8 million yuan / year and the profit is 9.8745 million yuan ≈ 0.49 years.

Claims

1. A continuous casting quality control method, characterized in that: The following steps are involved: S1: Data input and standardization: Acquire L1 / L2 / MES data and, based on OPC UA and combined with real-time data pipeline technology, perform data conversion and standardization to enable data interaction between intelligent agent clusters and industrial field systems, completing multi-source data fusion. S2: Central coordination: Using a metacognitive coordination agent as the core, it coordinates functional agents to implement task scheduling and status monitoring for the entire system, completing the three key functions of task decomposition, data scheduling, and closed-loop control. The functional agents include: semantic parsing agent, data perception agent, defect prediction agent, root cause analysis agent, process optimization agent, and digital twin agent. The central coordination specifically includes: S21: Task decomposition: Decompose complex quality control tasks into semantic parsing, data collection, and defect prediction subtasks, and then use semantic parsing agents to drive instruction understanding; S22: Data Scheduling: The metacognitive coordination agent dynamically schedules the data perception agent to collect or pre-process data based on the load and task priority of each functional agent. S23: Closed-loop control: Under the core coordination of the metacognitive coordination agent, various functional agents work together to achieve full-cycle control of quality issues; S3: Agent cluster operation and maintenance; specifically includes the following steps: S31: Agent optimization iteration: Process and analyze the agent's operating data, and continuously optimize the agent through the iterative chain of "monitoring → evaluation → optimization → application" to improve its operating performance; S32: Intelligent agent performance monitoring; collect and analyze performance data of the intelligent agent, and perform intelligent agent health assessment and health instrument display.

2. The method according to claim 1, wherein: Data collaboration is achieved by using blackboard model knowledge sharing. The process status tensor is stored in the shared memory area, and the entire process status is integrated with structured tensors. Data collaboration between intelligent agents is achieved through shared memory + SQL-like query.

3. The method according to claim 1, wherein: In the intelligent agent communication protocol, the message format specification and performance guarantee mechanism are set. Specifically, Define the message format specifications for communication between agents to ensure that different agents can accurately obtain the information transmitted to each other; The performance guarantee mechanism settings include: using Apache Kafka + RDMA network to limit end-to-end data transmission latency; Based on the RAFT consensus algorithm, the state synchronization of intelligent agents is achieved, so that when some intelligent agents fail, the backup intelligent agents can quickly take over the work; the SM4 national encryption algorithm is used to ensure the security of industrial data.

4. The method according to claim 1, wherein The acquisition of L1 / L2 / MES data specifically includes: Obtaining L1 basic automation layer: real-time process data from sensors / actuators; Obtain the L2 process automation layer: process model calculation results; Obtain production business data from the MES manufacturing execution system.

5. The method according to claim 1, wherein In the central coordination, The metacognitive coordination agent adopts a hierarchical task decomposition model: Where A is the task set, which includes three types of task candidates: prediction, tracing, and optimization. T is the target task to be determined, which is selected from set A through subsequent calculations. :It corresponds to the task The weight of :It is a task The scoring function of In this model, support for weights Perform dynamic allocation.

6. The method according to claim 1, wherein In the central coordination, Use real-time load balancing algorithms to guide dynamic resource scheduling: Among them, LoadIndex: load index, reflecting the system load status; QueueLen i : the length of the i-th queue; k: the total number of queues in the system; MaxQueue: the maximum queue length; When the load index exceeds the threshold, the Kubernetes container elastic expansion and contraction is triggered.

7. The method according to claim 1, characterized in that The specific steps of the closed-loop control include: Risk warning: The metacognitive coordination agent makes judgments based on the fused data output by the data perception agent and sends a risk warning letter to the defect prediction agent; Defect prediction: The defect prediction agent uses the LSTM-CNN hybrid model based on multimodal features to predict slab defects and outputs the spatiotemporal characteristics and risk levels of the defects. Root cause analysis: The root cause analysis agent receives the spatiotemporal feature data from the defect prediction agent, uses the CART decision tree + Bayesian network to trace the cause of the defect, and generates a root cause analysis result. Process optimization: The process optimization agent receives the output data from the root cause analysis agent and uses the NSGA-III multi-objective algorithm to generate a process parameter adjustment strategy. Digital twin verification and feedback: The digital twin agent receives the "parameter adjustment strategy" data from the process optimization agent, runs the thermal / stress field coupling model, simulates the quality of the cast billet after parameter adjustment, and feeds the verification results back to the metacognitive coordination agent; After the metacognitive coordination agent completes the "risk warning → defect prediction → root cause analysis → process optimization → digital twin verification and feedback" link, it receives the verification results and completes the cognitive iteration.

8. The method according to claim 7, characterized in that The defect prediction step adopts a multi-scale spatiotemporal feature fusion algorithm architecture to accurately capture the temporal dynamics and multi-scale correlations of process parameters, providing strong characterization features for defect probability prediction; The specific steps include: Dual-branch extraction: extracting timing features F from process parameter timing streams seq and spatial / multi-scale features F cnn ; Attention Fusion: Modeling F seq With F cnn Cross-correlation of the output attention enhancement feature F att ; Gating decision: Use gating mechanism to dynamically adjust F seq With F cnn The fusion weight of the two is obtained by dynamic weighted sum F fusion ; Probability output: F fusion Input the fully connected layer + Softmax classifier and output the defect probability distribution.

9. The method according to claim 8, characterized in that The specific steps of the dual-branch extraction include: Input the continuous time series data of the continuous casting process and input these data into the time series feature branch and the spatial / multi-scale feature branch; In the temporal feature branch, a BiLSTM-Pro module is constructed, which uses a bidirectional LSTM+time convolutional network TCN enhanced version to extract long-term and short-term temporal dependencies and output temporal features F seq Among them, the enhancement method of the temporal convolutional network (TCN) is to use convolution kernels to cover short time windows and strengthen local temporal patterns; In the spatial / multi-scale feature branch, a multi-scale CNN module is constructed, using CNNs with different sizes of convolution kernels to cover full-scale features from "short-term details" to "long-term trends", extract multi-scale spatiotemporal patterns, and output spatial / multi-scale features F cnn .

10. The method according to claim 9, characterized in that In the spatial / multi-scale feature branch, the EfficientNet-B4 model is used as the spatial feature extraction layer to convert large-resolution images into features containing semantics and spatial structures. Its generalized formula is expressed as: Among them, F spatial : Spatial features of output, EfficientNet: Efficient Convolutional Neural Network, I 5120×3840 : The input image data has a resolution of 5120×3840.

11. The method according to claim 8, characterized in that The specific steps of attention fusion include: Construct a spatiotemporal cross attention module, specifically using the attention mechanism to construct a feature fusion layer, by calculating F seq With F cnn The association weight of the two is used to model the cross-correlation between them, strengthen the most critical feature combination for defect prediction, and output the attention enhancement feature F att .

12. The method according to claim 8, characterized in that The specific steps of the gating decision include: F att Input gated feature fusion module, dynamically determines F through the gate coefficient g seq With F cnn The fusion weight of ; the gating mechanism formula is: Where, σ: sigmoid activation function, W g 、b g : The 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 of .

13. The method according to claim 12, characterized in that The gating decision also includes: An online learning engine is used to learn the W g The parameters are updated incrementally, and the generalized formula for the parameter update is expressed as: Among them, θ t : Model parameters at time t, η: learning rate, which controls the step size of gradient descent, :New data D new The loss gradient on : Regularization term.

14. The method according to claim 7, wherein: In the root cause analysis step, different types of data are weighted and processed using different methods. The specific processing methods include: The CART decision tree method was used to process the process parameters; The vibration signal of the equipment is processed by wavelet packet decomposition + energy entropy analysis; The operation logs are processed using the BERT semantic embedding method; The ambient temperature and humidity are processed using time series anomaly detection method.

15. The method according to claim 14, characterized in that 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 adjust the weight of the causal relationship in real time. The formula of the Bayesian network is: Where X is the root cause event, P(crack|X) is the conditional probability of the root cause X to the "crack", : The number of cracks that occurred due to root cause X in the sample, : The total number of root cause X occurrences in the sample.

16. The method according to claim 7, characterized in that The process optimization step also includes a priority experience replay PER and a reward function, and the specific steps include: The genetic algorithm NSGA-III multi-objective optimization model is used to generate the global optimal solution (v*, c*); Use reinforcement learning to make dynamic adjustments and generate dynamic adjustment values ​​(Δv, Δc); The real-time system executes f(v*+Δv, c*+Δc), performs verification, and feeds back actual rewards and experience data.

17. The method according to claim 16, characterized in that The specific method of the NSGA-III multi-objective optimization is: Among them, v: process parameters, c: environment, 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 ratio, VibEnergy: Current vibration energy, Vib max : Maximum vibration energy.

18. The method according to claim 16, characterized in that The specific method of reinforcement learning includes: using priority experience replay PER in conjunction with a reward function optimization strategy to generate dynamic adjustment values ​​(Δv, Δc); The formula for the priority experience replay PER is: Among them, P(i): sampling probability of the i-th experience, δ i : the deviation between the predicted value and the actual value, ε: error correction term, α: priority weight; The reward function is: Among them, 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.

19. The method according to claim 1, wherein The specific steps of the agent optimization iteration include: Agent monitoring: real-time collection of agent cluster operation data as operation and maintenance evaluation data; Build an evaluation database: Structure the collected operating data and perform indicator preprocessing to convert it into structured indicators that can be evaluated and analyzed, thereby building an evaluation database; Automatic optimization analysis: Based on the indicators of the evaluation database, perform multi-objective optimization analysis to obtain parameter optimization adjustment plans; Optimization and adjustment: The parameter optimization and adjustment plan is sent to the intelligent agent cluster to complete the parameter adjustment.

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

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

22. The method according to claim 1, wherein The health meter display converts the health status of the agent cluster into a visual form; Its health status indicators include: Mission effectiveness indicators include: Defect prediction accuracy: measured using confusion matrix calculation, with a target value of >95%; Root cause positioning error: measured by coordinate comparison, the target value is <±1mm; Response performance indicators, including: End-to-end latency: measured using a distributed tracing system, with a target value of <100ms. Throughput: measured using stress testing, with a target value of >800 QPS; Resource efficiency indicators include: CPU / GPU utilization: Prometheus monitoring is used to obtain the target value of 60%-80%; Memory usage: Statistics are collected while the container is running, with a target value of <32 GB per node.

23. A continuous casting quality control hardware system, characterized in that: include: 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, wherein the various components of the hardware system work together to implement the steps of the method according to any one of claims 1 to 22; The communication network component is used to realize data transmission and instruction interaction, and ensure information connectivity between various levels of the system; The data storage component is used to store computer program files and data files.

24. The hardware system according to claim 23, wherein: The data acquisition component includes: Parameter sensor: used to collect process parameter data in industrial production processes; Visual acquisition equipment: used to collect image or video data of production scenes; Vibration sensor device: used to collect vibration signal data when the equipment is running.

25. The hardware system according to claim 23, wherein: The edge computing node is in communication with the data acquisition component and is used to perform real-time preprocessing on the collected raw data.

26. The hardware system according to claim 23, characterized in that: The intelligent processing and control component includes: Agent server cluster: communicates with the edge computing nodes and deploys functional agents for in-depth analysis of pre-processed data; The intelligent agent server cluster includes a GPU server; Industrial control subsystem: communicates with the edge computing node, receives control instructions from the intelligent server cluster or edge computing node, and adjusts the operating parameters of industrial production equipment in real time; The industrial control subsystem includes a PLC control system.

27. The hardware system according to claim 26, characterized in that: The visualization and interaction component is connected to the intelligent server cluster to display real-time data, analysis results, equipment status and control instruction execution feedback of the production process, supporting operators in monitoring, intervention and decision-making; The visualization and interaction component includes a 3D monitoring screen.

28. A computer program product, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 22 are implemented.

Images