Blast furnace working condition knowledge graph construction method and device

Abstract

The invention discloses a blast furnace working condition knowledge graph construction method and device, and belongs to the field of blast furnace ironmaking intelligent control. The method comprises the following steps: fusing multi-source data, integrating sensor real-time data, process parameters, operation logs and expert rules, and enhancing abnormal working condition data generation capability by using a time sequence generative adversarial network; a three-layer dynamic body structure with the core abnormal furnace condition as the center is constructed, and a dynamic edge weight updating mechanism is designed; extracting multi-region spatial-temporal characteristics of the blast furnace by adopting a spatial-temporal diagram convolutional network, and generating an interpretable knowledge triple in combination with a causal discovery algorithm FGES; and realizing real-time updating of the knowledge graph based on a streaming graph learning framework, generating a regulation and control instruction through multi-agent reinforcement learning, and issuing the regulation and control instruction to a blast furnace control system through a digital twin interface. The method can clearly represent the symptom incidence relation of the abnormal furnace condition of the blast furnace, rapidly position the abnormal source and generate the regulation and control strategy, and has important practical significance for improving the intelligent level of blast furnace ironmaking.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control of blast furnace ironmaking, and particularly relates to a method and device for constructing a blast furnace operating condition knowledge graph that integrates spatiotemporal characteristics and causal reasoning. Background Technology

[0002] The blast furnace is the core component of the long-process ironmaking process. The blast furnace smelting process is a highly complex thermodynamic and kinetic coupled process, characterized by nonlinearity, large time delay, high noise, and distributed parameters. Numerous complex chemical reactions occur inside the furnace, and the parameters are coupled with each other.

[0003] However, the internal condition of a blast furnace is difficult to directly detect, and operators have limitations in observing and judging its operating status, making it difficult to make timely and accurate judgments about the furnace condition. Fluctuations in blast furnace conditions and unknown anomalies seriously affect iron production, quality, and energy consumption. Therefore, ensuring stable and smooth blast furnace production is key to achieving high-quality, low-consumption, high-yield, and long-life blast furnaces.

[0004] Current blast furnace anomaly diagnosis mainly relies on expert systems, which diagnose based on predefined rules (such as "wind pressure > 250 kPa → risk of material suspension"). However, rule updates are lagging and cannot handle multi-parameter coupled scenarios, resulting in a certain false alarm rate in practical applications. Data-driven models, such as time-series models using LSTM, lack interpretability and fail to adequately model causal relationships. The heterogeneity of multi-source data is strong, with traditional methods showing time alignment errors > 500 ms, resulting in data dimension fragmentation and difficulty in uniformly representing the spatiotemporal coupling relationship of abnormal furnace conditions. Static knowledge models cannot adapt to the dynamic evolution of blast furnaces, resulting in large delays in control response, lack of causal reasoning ability, and difficulty in locating the root cause of anomalies (such as material suspension caused by multiple factors such as sudden changes in wind pressure or raw material deterioration).

[0005] Entering the 21st century, the globalization of resources and energy in the steel industry, the modernization, large-scale and automation of technology and equipment, the scientific research and technological innovation in the metallurgical industry, and the application of information technology have enabled the steel industry to achieve a historic leap. The rapid development of information technology and the continuous advancement of intelligent blast furnace diagnosis and construction have brought new changes and challenges to the traditional blast furnace ironmaking industry.

[0006] Therefore, it is urgent to deeply integrate expert knowledge, operational experience, and intelligent technology on the basis of existing automated production and informatization, so as to achieve in-depth information perception, optimized decision-making, and precise coordination, in order to improve the level of intelligence of blast furnaces. Summary of the Invention

[0007] The purpose of this invention is to provide a method and apparatus for constructing a blast furnace operating condition knowledge graph to solve the above-mentioned problems, thereby improving the speed and accuracy of blast furnace operating condition diagnosis, addressing abnormal causes in advance, ensuring the stable and smooth operation of the blast furnace, and assisting blast furnace operators in real-time monitoring of abnormal furnace conditions.

[0008] To achieve the above objectives, the present invention provides the following solution:

[0009] This invention first acquires multimodal data related to blast furnace operating conditions using various sensors and data acquisition methods, including real-time sensor data such as furnace top temperature, blast pressure, and material rate, infrared thermal imaging video, and process parameters (such as blast system and charging matrix) from a structured production database. In addition, it collects unstructured text data related to blast furnace operation, such as production logs, operation manuals, expert experience documents, and rule bases. To address the problem of imbalanced data samples in abnormal furnace conditions, this invention employs a Temporal Generative Adversarial Network (TimeGAN) to enhance the sensor data, thereby improving performance in cases with small sample sizes.

[0010] The multi-source heterogeneous data fusion and enhancement system includes a data acquisition system and a time-series data enhancement module. The data acquisition system includes a sensor network, an infrared monitoring system, and a text knowledge extraction module. Specifically, the sensor network deploys 328 high-precision sensors, including a furnace top temperature sensor (K-type thermocouple, ±0.5℃), a wind pressure sensor (piezoelectric, range 0-500kPa, ±0.01MPa), and a material velocity radar (millimeter-wave radar, ±0.05m / min); the infrared monitoring system uses a FLIR A65 infrared thermal imager (8-14μm band, 640×480 resolution, 10fps frame rate) to reconstruct the three-dimensional temperature field of the blast furnace wall (spatial resolution ≤5cm); the text knowledge extraction module extracts entities (F1≥0.92) and relationships (accuracy≥0.85) from the production logs based on the BERT-CRF model. Regarding time-series data augmentation, to address the problem of insufficient abnormal operating condition samples, this invention proposes a TimeGAN algorithm incorporating metallurgical constraints. By enhancing the temporal characteristics of the data, this further improves the model's ability and accuracy in detecting abnormal furnace conditions.

[0011] The ontological knowledge system design of the blast furnace defines 8 types of core abnormal furnace conditions, 12 types of process parameter nodes, and causal nodes. It abstracts the relationship between various process parameters and abnormal states into a graph structure, using real-time causal edges and statistical correlation edges to describe the dynamic connections between states. The half-life of real-time causal edges, the half-life of statistical correlation edges, and the dynamic edge weights are updated according to the exponential decay rule as follows:

[0012]

[0013] Where λ is the decay coefficient, determined by the half-life T 1 / 2 Decision; T 1 / 2 The time required for the weight to decay to half of its initial value is 30 ± 5 min for real-time causal edges and 8 hours ± 15 minutes for statistically related edges.

[0014] The real-time causal edges are weighted and decayed according to their half-life, reflecting the timeliness and real-time nature of blast furnace state changes. Examples of SWRL rules in the domain knowledge constraint library of this invention: The domain knowledge constraint library of this invention contains multiple causal rules in the metallurgical field to constrain the generation of relationships in the knowledge graph. For example, rule 1: If a sudden drop in blast temperature is detected (the drop exceeds 50°C / minute), and a cooling trend in furnace temperature is detected within the following 10-15 minutes (the iron [Si] content is below 0.3%), then the sudden drop in blast temperature is determined to be the cause of the cooling furnace temperature, with a softmax normalized probability value of 0.92; rule 2: If the oxygen enrichment rate increases by more than 0.5%, and a rising trend in furnace temperature is detected within 5-8 minutes (the iron [Si] content increases by more than 0.1%), then the increase in oxygen enrichment rate is determined to be the cause of the rising furnace temperature, with a softmax normalized probability value of 0.85. These rules avoid false associations by limiting the time window of causal relationships; by introducing domain expert knowledge, excluding causal relationships that do not conform to metallurgical laws, and assigning a softmax normalized probability value to each causal relationship, the accuracy of causal reasoning is improved by supporting uncertain reasoning.

[0015] The spatiotemporal knowledge extraction and triple generation in this invention utilizes a spatiotemporal graph convolutional network (ST-GCN) to extract spatial and temporal features from multimodal data, aggregates sensor data from different regions of the blast furnace, and generates causal triples in the blast furnace operating conditions to capture the evolution of parameters through gated temporal convolution (GTCM) formula (5) as shown below.

[0016] H t =σ(W g *X t +b g )⊙tanh(W c *X t +b g ) (5)

[0017] Where σ is the Sigmoid function; ⊙ represents element-wise multiplication; W g W c These are learnable weights.

[0018] Combining domain knowledge, the FGES causal discovery algorithm is adopted, and the formula is shown below:

[0019]

[0020] Where d is the dimension of the variable.

[0021] For example, in the case of "continuous decrease in material rate → material collapse", important causal relationships are evaluated and screened using probability values ​​normalized by softmax to ensure that the generated causal relationships have high credibility.

[0022] The dynamic knowledge graph update and reasoning described in this invention utilizes a streaming graph learning framework combined with the Neo4j graph database to achieve real-time updating and reasoning of the blast furnace operating condition knowledge graph. The streaming graph learning framework dynamically adjusts the correlation strength between nodes based on the Temporal Graph Attention (T-GAT) mechanism, and dynamically adjusts the correlation strength between abnormal furnace conditions and their causes, ensuring real-time reflection of the blast furnace status. During the graph update process, changes in temporal data and causal relationships continuously adjust the weights of nodes and edges, ensuring that the system's intelligent decision-making keeps pace with the real-time changes in the blast furnace.

[0023] The abnormal furnace condition diagnosis and control system of the present invention employs graph neural network embedding technology. Based on real-time data of the current blast furnace operating condition map, it generates an abnormal diagnosis report and outputs possible abnormal probability distributions. Combined with multi-agent reinforcement learning (MARL) technology, the system can select the optimal operation command (such as adjusting blast temperature, material line depth, etc.) based on the current operating condition and send the command to the blast furnace PLC system through a digital twin interface. By verifying the control effect of the system in real time, the stability of blast furnace operation is ensured.

[0024] The present invention also provides a blast furnace operating condition knowledge graph construction device based on the above method, comprising the following steps:

[0025] S1: Edge computing unit: Deployed at the blast furnace site, responsible for running the lightweight ST-GCN model to realize the real-time updating and processing of the blast furnace operating condition knowledge graph;

[0026] S2: Causal Reasoning Engine: Based on the FGES algorithm and combined with knowledge of the blast furnace field, it generates interpretable abnormal cause reports, provides causal reasoning support, and helps operators understand the causes of abnormal furnace conditions. Neo4jEnterprise 4.4: Supports temporal Cypher queries (such as MATCH(a)-[r]->(b)WHERE r.timestamp>datetime()-duration(`PT30M`)).

[0027] S3: Control and Decision Interface: Visually displays the softmax normalized probability curves of 8 core abnormal furnace conditions and recommended operations, supports manual intervention and correction, and provides decision support for operators.

[0028] Compared with the prior art, the present invention has the following advantages and technical effects:

[0029] This invention provides a method and apparatus for constructing a blast furnace operating condition knowledge graph. It integrates multi-source data fusion and spatiotemporal coupling modeling, effectively fusing multi-source data from the blast furnace using techniques such as Time-Generative Adversarial Network (TimeGAN) and Spatiotemporal Graph Convolutional Network (ST-GCN). This dynamically captures the spatiotemporal patterns of furnace condition evolution, overcoming the problem that static data modeling in traditional methods cannot adapt to dynamic changes. Through causal discovery algorithms and streaming graph learning frameworks, this invention can accurately uncover the causal relationships behind furnace condition anomalies, assisting intelligent control decisions and improving the intelligence and autonomy of blast furnace control. The apparatus of this invention has real-time monitoring, anomaly diagnosis, and self-healing control functions, enabling it to predict and handle potential abnormal furnace conditions in advance, ensuring the efficient and safe operation of the blast furnace. By integrating advanced artificial intelligence algorithms with multi-source data on blast furnace operating conditions, this invention proposes a method and apparatus for intelligent diagnosis of blast furnace operating conditions based on knowledge graph construction. This effectively improves the accuracy and real-time performance of blast furnace anomaly diagnosis, providing more intelligent and precise decision support for blast furnace operation. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a system architecture diagram of the present invention (data acquisition → knowledge graph → control closed loop);

[0032] Figure 2 Detailed diagram of the ST-GCN network structure provided by this invention;

[0033] Figure 3 This is a schematic diagram of the dynamic ontology model provided by the present invention;

[0034] Figure 4 This is a schematic diagram of the control interface provided by the present invention; Detailed Implementation

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Example 1

[0038] This embodiment provides a system architecture diagram (data acquisition → knowledge graph → control closed loop) as follows: Figure 1 As shown, the system includes a data acquisition system and time-series data enhancement. The knowledge acquisition system includes a sensor network deploying 328 high-precision sensors, including a furnace top temperature sensor (K-type thermocouple, ±0.5℃ accuracy, 1Hz sampling frequency), a piezoelectric wind pressure sensor (range 0-500kPa, ±0.01MPa), a material velocity radar, and a millimeter-wave radar (±0.05m / min accuracy); an infrared monitoring system using a FLIR A65 infrared thermal imager (wavelength 8-14μm, resolution 640×480, frame rate 10fps), reconstructing the three-dimensional temperature field of the furnace wall using a 3D convolutional network (C3D) (spatial resolution ≤5cm), outputting a 64×64×16 temperature tensor; and text knowledge extraction, extracting entities (F1≥0.92) and relationships (accuracy≥0.85) from production logs, for example, extracting the operation event from "15:30 increase wind temperature by 50℃". The time-series data augmentation employs an improved TimeGAN algorithm. The generated data is verified through metallurgical thermodynamics (heat conduction equation residual <5%), the outlier sample size is expanded to 5 times the original data, and the generated data distribution is highly consistent with the real data. The cross-modal alignment design uses a comparison loss function to constrain the semantic consistency between infrared video features and sensor data, with a temperature coefficient τ = 0.07 and an alignment error ≤5%.

[0039] Table 1 Sensor Parameter Table

[0040] Sensor type quantity Precision / Resolution Sampling frequency use K-type thermocouple 50 ±0.5℃ 1Hz Furnace top temperature monitoring piezoelectric wind pressure sensor 30 ±0.01MPa 1Hz Wind pressure monitoring Millimeter-wave material velocity radar 10 ±0.05m / min 0.5Hz Material rate monitoring FLIRA65 thermal imager 1 unit 640×480 pixels 10fps Furnace wall temperature field reconstruction

[0041] Example 2

[0042] This embodiment is based on Figure 2-3The model constructs a dynamic ontology model, using eight core abnormal furnace conditions as central nodes, and building a three-layer graph structure. It includes nodes for 12 process parameters such as furnace top temperature and air pressure, which are mapped to sensor values ​​in real time. Six types of control actions are defined, including air temperature adjustment and oxygen enrichment rate changes. Causal nodes link abnormal furnace conditions to their causes (e.g., "sudden increase in air pressure → risk of material suspension"). Dynamic relationships are described through real-time causal edges and statistical correlation edges. After inputting spatiotemporal feature vectors into the domain-constrained FGES algorithm, invalid causal relationships (e.g., "sudden drop in air temperature + furnace temperature cooling" time window constraint) are eliminated using SWRL rules (e.g., "increased oxygen enrichment rate → furnace temperature cooling"), generating interpretable triples (e.g., "continuous decrease in material rate, induction, material collapse"). Relationships with a probability value ≥ 0.78 after softmax normalization are written into the Neo4j graph database. Dynamic edge weights decay in real time according to a formula, where the half-life t of real-time causal edges is... 1 / 2 =30±5min, count the associated edges t 1 / 2 =8h±15min.

[0043] Example 3

[0044] The streaming graph learning framework dynamically updates node embeddings through a temporal graph attention mechanism (T-GAT). When an abnormal furnace top temperature gradient is detected, the model automatically strengthens the association weights between the nodes "furnace wall thickness" and "difficult furnace conditions". The anomaly diagnosis agent queries the current graph using Cypher statements. On the control interface, such as... Figure 4 As shown, the output displays the probability distribution of eight types of anomalies. The control agent generates a top-three operation plan based on multi-agent reinforcement learning (MARL). Operators can choose to "execute the intelligent plan immediately" or adjust the instructions through "manual intervention coverage." The instructions are sent to the blast furnace PLC system via the OPC UA protocol, with a delay of <100ms. Feedback is injected into the graph in real time, forming a closed-loop diagnosis and control system for blast furnace condition anomalies, encompassing "monitoring-inference-control."

[0045] The examples described above are merely typical application scenarios of the present invention. Based on the dynamic characteristics of metallurgical processes, those skilled in the art can adjust model parameters or expand the body nodes according to actual working conditions, all of which fall within the protection scope of the present invention. Without departing from the design spirit and concept of the present invention, all modifications and improvements made by those skilled in the art to the technical solutions of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method and apparatus for constructing a knowledge graph based on blast furnace operating conditions, characterized in that, The method includes: S1. Multi-source heterogeneous data fusion and enhancement: Collect data from 8 types of blast furnace sensors, integrate infrared thermal imaging video, structured process parameters with the following limits: air supply temperature 1100-1250℃, air volume fluctuation threshold ±3%, material distribution matrix ring width 200-500mm, tilt angle 30°-50°, and use a temporal adversarial generative network for data enhancement. The generator loss function is shown in the following formula (1), with metallurgical constraint terms added: in, ρc is the second spatial derivative of the temperature field. p κ is the volumetric heat capacity of the blast furnace masonry; κ is the thermal conductivity. This represents the rate of change of temperature over time. S2. Blast Furnace Body Knowledge System Design: A dynamic ontology model centered on eight core abnormal blast furnace conditions is constructed, defining three types of nodes: abnormal furnace condition nodes, process parameter nodes, and causal nodes; dynamic edge types and real-time causal edges are designed; and correlation edges are statistically analyzed. S3. Spatiotemporal knowledge extraction and triple generation: A spatiotemporal graph convolutional network is used to extract features from multimodal data, aggregating sensor data from different regions of the furnace body, and capturing parameter evolution patterns through gated temporal convolution. A domain-constrained FGES causal discovery algorithm is used to generate triples. The spatiotemporal features extracted by the spatiotemporal graph convolutional network are input into a streaming graph learning framework as input features for dynamic weight calculation. S4. Dynamic Knowledge Graph Update and Reasoning: Deploy a streaming graph learning framework and work in conjunction with the Neo4j graph database to process data streams in real time and update graph node attributes and edge weights; design a time-series graph attention mechanism to dynamically adjust the correlation strength between abnormal furnace conditions and their causes. S5. Abnormal Furnace Condition Diagnosis and Control: The diagnostic agent utilizes graph neural network embedding to query the current edge weights in Neo4j in real time via Cypher statements to obtain an abnormal probability distribution. The control agent outputs the optimal operation command (such as blast temperature adjustment) based on the current graph state and sends the control command to the blast furnace PLC system through a digital twin interface, and verifies the effect in real time.

2. The method according to claim 1, characterized in that, The multimodal data fusion in step S1 includes: A 3D convolutional network is used to extract the spatiotemporal features of the furnace wall temperature field from infrared thermal imaging video and align them with sensor data; a cross-modal contrastive learning loss function is designed to constrain the semantic consistency of text logs with sensor changes.

3. The method according to claim 1, characterized in that, The optimization process of the FGES causal discovery algorithm in step S3 is as follows: the FGES causal discovery algorithm with introduced domain constraints excludes causal relationships that do not conform to metallurgical logic in the prior stage.

4. The method according to claim 1, characterized in that, The streaming graph learning framework in step S4 includes: Dynamic nodes utilize Temporal Graph Transformer, incorporating the node's historical state {h t-T ,...,h t Generate the current embedding; define the weight update formula using the edge weight decay rule as follows: Where ω(t) is the edge weight at time t, reflecting the strength of the association between nodes; ω0 is the initial weight; λ is the decay coefficient, measuring the rate at which the weight decreases over time; T 1 / 2 t represents the half-life, the time required for the weight to decay to 50% of its initial value; t is time, in minutes.

5. The method according to claim 1, characterized in that, The multi-agent reinforcement learning (MARL) in step S5 includes: The state space consists of graph node embedding vectors plus real-time sensor values; the action space control commands are discretized; the reward function formula is shown below: R = α·(1-anomaly confidence level) + β·molten iron quality index - γ·operation frequency (3) α, β, and γ are adjustable hyperparameters used to balance anomaly suppression, quality improvement, and operational stability; the anomaly confidence is the current anomaly probability output by the map diagnosis module.

6. A device for constructing a knowledge graph of blast furnace operating conditions, characterized in that, include: The edge computing unit is deployed at the blast furnace site and runs a lightweight ST-GCN model to achieve real-time map updates. A causal reasoning engine based on the FGES algorithm to generate interpretable anomaly cause reports; The control decision-making visualization displays recommended actions for 8 abnormal furnace conditions, supporting manual intervention and correction.

Images