Pre-ironmaking intelligent ore blending system and device, storage medium and program product

Abstract

The invention provides an intelligent ore blending system and device before iron making, a storage medium and a program product, and relates to the technical field of metallurgical engineering. The pre-ironmaking intelligent ore blending system comprises a raw material knowledge graph module, a raw material distribution module and a raw material distribution module, wherein the raw material knowledge graph module is used for constructing ore attributes and a raw material knowledge graph and realizing multi-dimensional coding of the raw material attributes; the metallurgy prediction module is used for realizing joint optimization of the model through fusion of an ore blending multi-objective optimization model and a large-model multi-modal model; the robust optimization module is used for fusing a transfer learning algorithm and constructing an anti-interference ore blending optimization decision model; and the closed-loop control module is used for integrating an online detection instrument and a dynamic parameter correction model and realizing hour-level ratio closed-loop adjustment. The deep neural network, the multi-objective optimization algorithm and the metallurgical process mechanism model are integrated, collaborative optimization of raw material cost saving and stable operation of the blast furnace is achieved, and the method is suitable for ore proportion decision making at the front end of sintered ore production.

Description

Technical Field

[0001] The present invention relates to the fields of metallurgical engineering technology, artificial intelligence control technology, and Internet of Things technology, and in particular to an intelligent ironmaking optimization system, device, storage medium, and program product based on industrial big data. Background Art

[0002] Traditional pre-iron ore blending operations face three major technical bottlenecks. First, strong reliance on experience: ore blending plans rely heavily on the expertise of process experts, making it impossible to quantify and evaluate nonlinear relationships between parameters. Second, limited computational dimensionality: existing multi-objective optimization models can only handle up to 20 raw material variables and fail to dynamically track all elements. Third, insufficient response speed: manual plan adjustments take more than 48 hours, making them unable to adapt to real-time fluctuations in raw material market prices.

[0003] For example, Chinese patent application number CN202111228305.4, published on March 1, 2022, uses a data management system to collect, analyze, and process sintering process parameters. The sintering ore blending module predicts sintered ore quality, develops ore blending plans, evaluates sintering production, adjusts process parameters, and optimizes ore blending plans. Another example is Chinese patent application number CN201610062405.7, published on May 25, 2016. Through production exploration geological data, ore powder sampling and testing, and rock and mineral identification, it guides ore blending technicians in quantitatively blending iron ores with multiple ore bodies and ore types.

[0004] Although these existing technologies have proposed blast furnace ironmaking cost optimization models, their core algorithms have not yet broken through the limitations of existing technologies: the objective function only includes a single dimension of cost, and does not consider technical indicators such as alkalinity balance and metallurgical properties; the constraint equation uses static linear programming and fails to establish the influence function of chemical composition fluctuations on sintered ore strength; the input parameters do not integrate the raw material composition fluctuation prediction module, resulting in poor adaptability of the optimization results to supply chain changes. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention proposes an intelligent ironmaking optimization system, device, storage medium and program product based on industrial big data, which integrates deep neural networks, multi-objective optimization algorithms and metallurgical process mechanism models to achieve coordinated optimization of raw material cost savings and stable blast furnace operation, and is suitable for ore proportioning decisions at the front end of sintered ore production.

[0006] In a first aspect, the present invention provides an intelligent ore blending system for iron ore production, comprising the following modules:

[0007] Raw material knowledge graph module: Construct a three-dimensional feature matrix to describe ore properties, build a dynamic knowledge graph for raw materials, and build an ore-origin-supplier association database based on knowledge extraction technology to achieve multi-dimensional encoding of raw material properties.

[0008] Metallurgical prediction module: Builds a sintering performance prediction network to adjust the attention mechanism, and uses a spatiotemporal graph convolutional network (ST-GCN) to model the complex mapping relationship between ore proportions and sinter properties. By integrating the ore proportioning multi-objective optimization model with the large-scale multimodal model, joint model optimization is achieved.

[0009] Robust optimization module: Integrates robust optimization theory and transfer learning algorithm to build an interference-resistant ore allocation optimization decision model.

[0010] Closed-loop control module: Integrates online detection instruments and dynamic parameter correction models to achieve hourly ratio closed-loop control; the cyclic operation steps of the closed-loop control include: online detection, dynamic correction, execution control, production feedback, and online detection.

[0011] As a further improvement of the present invention, the raw material knowledge graph module includes:

[0012] Construct a three-dimensional feature matrix describing ore properties;

[0013] By defining quintuples to describe ore entities, a dynamic ore knowledge graph is constructed.

[0014] As a further improvement of the present invention, the quintuple includes:

[0015] Use hash values as unique ore identifiers, ore chemical composition vectors, ore physical property sets, ore supply chain relationship matrices, and dynamic feature tensors.

[0016] As a further improvement of the present invention, the metallurgical prediction module includes:

[0017] Using a dual-path attention mechanism to build a deep sintering performance prediction network;

[0018] Large model multimodal model: multimodal fusion of large model and material characteristics;

[0019] Fusion of multi-objective optimization model for ore distribution and multi-modal model of large model.

[0020] As a further improvement of the present invention, the dual-path attention includes: ore parameter attention and time series feature attention.

[0021] As a further improvement of the present invention, the multimodal fusion of the large model and material characteristics includes the following fusion framework:

[0022] Obtain raw material modalities based on material testing data;

[0023] Characterize the raw material modality;

[0024] Cross-modal alignment of feature encodings via hypersphere embedding;

[0025] Knowledge-enhanced reasoning, including fusing cross-modal aligned features with a knowledge base;

[0026] Decision optimization via physics-guided attention weight modification techniques.

[0027] As a further improvement of the present invention, the multimodal fusion of the large model and material characteristics includes the optimization design of spatiotemporal convolution operators, wherein:

[0028] The spatial domain convolution operator is optimized through an improved residual graph convolution layer;

[0029] The time domain convolution operator is optimized through multi-scale dilated causal convolution.

[0030] As a further improvement of the present invention, the multimodal fusion of the large model and material properties includes process knowledge distillation, specifically by incorporating sintering operation expert experience into the pre-training stage through physically guided attention weight correction technology.

[0031] As a further improvement of the present invention, the multimodal fusion of the large model and material properties, including key constraints of the sintering process, specifically includes:

[0032] Thermodynamic constraints: Add a residual calculation module after the model output layer to compare the predicted temperature with the solution of the thermodynamic equation in real time, and regularly update the Q_loss value through an online thermal imager;

[0033] Chemical reaction constraints: by calculating the mass ratio of basic oxides to acidic oxides, the amount of liquid phase generated is limited;

[0034] Equipment operation constraints: The total mass of materials passing through the sintering machine per unit time is used as a trigger parameter for dynamic adjustment.

[0035] As a further improvement of the present invention, the fusion of the ore blending multi-objective optimization model and the large model multi-modal model includes:

[0036] Define a multi-objective optimization function that comprehensively considers economy, metallurgical performance balance, and process constraints;

[0037] The improved NSGA-III algorithm is used to solve the optimization function of S31 and define the reference point adaptive adjustment strategy.

[0038] As a further improvement of the present invention, the robust optimization module includes:

[0039] Robust dual transformation;

[0040] Explicitly transform the dual problem, including obtaining strong duality conditions via Legendre-Fenchel transformation;

[0041] Robust duality gap analysis, including: defining approximation error bounds and uncertainty modeling;

[0042] Transfer learning, including: feature alignment loss and rule extraction.

[0043] As a further improvement of the present invention, the closed-loop control module includes the following operating steps:

[0044] Online detection: data collection through sensors;

[0045] Dynamic correction: Integrates the LSTM time series prediction model with the adaptive PID algorithm to automatically generate parameter adjustment suggestions every hour;

[0046] Execution control: control the executive agencies in the system;

[0047] Production feedback;

[0048] Return to online testing.

[0049] As a further improvement of the present invention, the closed-loop control module includes an hourly closed-loop control mode, and its operating steps include:

[0050] Data collection: obtain 200+ parameters such as composition, particle size, temperature, etc.

[0051] Anomaly detection: Screening abnormal indicators based on the 3σ principle;

[0052] Model calculation: generate three sets of candidate solutions for ratio adjustment;

[0053] Execution verification: issue instructions and verify the effects;

[0054] The following dynamic adjustment strategies are also included:

[0055] Normal mode: fine-tune parameters every hour;

[0056] Emergency mode: When a component fluctuation >2% is detected, adjustments will be made within 5 minutes.

[0057] In a second aspect, the present invention provides a multimodal fusion method of a large model and material characteristics, characterized in that the fusion framework includes:

[0058] Obtain raw material modalities based on material testing data;

[0059] Characterize the raw material modality;

[0060] Cross-modal alignment of feature encodings via hypersphere embedding;

[0061] Knowledge-enhanced reasoning, including fusing cross-modal aligned features with a knowledge base;

[0062] Decision optimization via physics-guided attention weight modification techniques.

[0063] As a further improvement of the present invention, the multimodal fusion method of the large model and material properties further includes:

[0064] The spatiotemporal convolution operators are optimized by using improved residual graph convolution layers and multi-scale dilated causal convolutions respectively;

[0065] Through the physics-guided attention weight correction technology, the experience of sintering operation experts is incorporated into the pre-training stage to perform process knowledge distillation;

[0066] The key constraints of the sintering process include comprehensive thermodynamic constraints, chemical reaction constraints and equipment operation constraints.

[0067] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the system or method described in the first and second aspects.

[0068] In a fourth aspect, the present invention provides a device system, comprising the computer device described in the third aspect, and also comprising a raw material storage device, a feeding device, a crushing and screening device, a grinding device, a beneficiation device, a batching device, and a conveying device involved in the pre-iron ore blending system. The device system operates in coordination to realize the system or method described in the first and second aspects.

[0069] In a fifth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the system or method described in the first or second aspect.

[0070] In a sixth aspect, the present invention provides a computer program product, which, when executed by a processor, implements the system or method described in the first and second aspects.

[0071] The present invention proposes an intelligent ironmaking optimization system, device, storage medium and program product based on industrial big data, which integrates deep neural networks, multi-objective optimization algorithms and metallurgical process mechanism models to achieve coordinated optimization of raw material cost savings and stable blast furnace operation. It is suitable for ore proportioning decisions at the front end of sintered ore production. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1This is a structural diagram of an intelligent ore blending system before iron ore production disclosed in the present invention.

[0073] Figure 2 This is a closed-loop control operation mode diagram of the iron ore intelligent ore blending system disclosed in the present invention. DETAILED DESCRIPTION

[0074] 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.

[0075] 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.

[0076] 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.

[0077] like Figure 1 As shown, the present invention provides an intelligent ore blending system for iron ore, which includes four core modules:

[0078] Raw material knowledge graph module: Based on knowledge extraction technology, a database of ore-origin-supplier association is constructed to achieve multi-dimensional coding of raw material attributes.

[0079] Metallurgical prediction module: The spatiotemporal graph convolutional network (ST-GCN) is used to model the complex mapping relationship between ore proportions and sinter properties.

[0080] Robust optimization module: Integrates robust optimization theory and transfer learning algorithm to build an interference-resistant ore allocation optimization decision model.

[0081] Closed-loop control module: Integrates online detection instruments and dynamic parameter correction models to achieve hourly ratio closed-loop adjustment.

[0082] Module 1: Raw Material Knowledge Graph Module

[0083] By constructing a three-dimensional feature matrix to describe the ore properties, and then building a dynamic knowledge graph of raw materials, the internal connection of the dynamic knowledge graph is formed. The detailed method is as follows:

[0084] S1: Construct a three-dimensional feature matrix describing ore properties:

[0085]

[0086] in,

[0087] N: number of ore varieties (N≥100);

[0088] T: time window length (T = 30 days);

[0089] D: Feature dimension (D=4), including:

[0090] (1) Chemical composition: {w(Fe),w(SiO2),w(Al2O3)}∈[0,1],

[0091] (2) Grinding power consumption index: P_grind∈R+,

[0092] (3) Granulation efficiency coefficient: η_gran∈[0.7,1.0],

[0093] (4) Base price and transportation cost: {C_base,C_trans}∈R+;

[0094] R+: the set of positive real numbers;

[0095] w: mass percentage.

[0096] S2: Describe the ore entity and then build a dynamic ore knowledge graph.

[0097] The ore entity is described by defining a five-tuple:

[0098] O=<ID,C,A,R,F>

[0099] in,

[0100] (1) ID: Use hash value as the unique identifier of the ore;

[0101] (2) C: ore chemical composition vector;

[0102] For example, the vector is composed of elements such as iron, iron oxide, silicon dioxide, aluminum oxide, phosphorus, and sulfur, that is,

[0103]

[0104] (3) A: A set of physical properties of the ore, exemplary,

[0105]

[0106] in,

[0107] d 50 (μm): The particle size corresponding to when the cumulative particle size distribution reaches 50%;

[0108] HGI: Hardgrove Grindability Index;

[0109] ρ: density, in tons per cubic meter (t / m 3 );

[0110] (4) R: Supply Chain Relationship Matrix

[0111]

[0112] (5) F: dynamic feature tensor, exemplary,

[0113]

[0114] Module 2: Metallurgical Prediction Module

[0115] A spatiotemporal graph convolutional network is used to model the complex mapping relationship between ore proportions and sintered ore properties, and a sintering performance prediction network is constructed to adjust the attention mechanism. At the same time, the large model is multimodally fused with material properties, and the Transformer architecture is integrated to achieve deep coupling with the metallurgical mechanism model to achieve joint optimization of the model.

[0116] S1: Sintering performance prediction network

[0117] This example uses a depth prediction model with a dual-path attention mechanism:

[0118]

[0119] in:

[0120] (1)α i : The attention weight of the i-th ore, calculated as:

[0121]

[0122] in,

[0123] e i,j ∈R^d: embedding vector of ore (i,j),

[0124] W∈R^(d×d): learnable parameter matrix,

[0125] q∈R^d: global context query vector;

[0126] (2)g φ : Temporal feature extraction sub-network, using TCN architecture with gating mechanism:

[0127]

[0128] in,

[0129] t: time step; l: number of layers; W: weight matrix; “*”: convolution operation; “⊙”: Hadamard product; σ: Sigmoid activation function.

[0130] S2: Multimodal Fusion of Large Models and Material Characteristics

[0131] Deeply couple the Transformer architecture with the metallurgical mechanism model to achieve data-driven and knowledge-driven joint optimization. Specific technical paths include:

[0132] S21: Multimodal Fusion Framework

[0133] The system adopts a phased and progressive architecture, and the workflow includes:

[0134] S211: obtaining raw material modal based on material detection data;

[0135] For example, the input material detection data includes five types of data modalities:

[0136] (1) Chemical analysis: XRF element content (25+ dimensions)

[0137] (2) Physical testing: particle size distribution curve (128 dimensions), drum index

[0138] (3) Metallurgical parameters: reduction degree (η), softening melting range (ΔT)

[0139] (4) Economic indicators: ex-factory price including tax, ocean freight volatility

[0140] (5) Process constraints: alkalinity target (R = CaO / SiO2), equipment capacity

[0141] S212: feature encoding of raw material modalities;

[0142] Feature Encoder: Based on the improved Transformer architecture, a modality-specific encoding network is designed:

[0143]

[0144] in,

[0145] x m : the mth vector in the input data;

[0146] :x m After conversion, the vector input to the first layer network;

[0147] : Modality-specific projection matrix, adapting to the dimensional differences of each modality;

[0148] b m : learnable bias;

[0149] GeLU: Gaussian Error Linear Unit activation function.

[0150] S213: Cross-modal alignment of feature encodings via hypersphere embedding;

[0151] The six-dimensional hypersphere embedding technology is used to map heterogeneous data such as the physical properties (particle size, moisture), chemical indicators (Fe, SiO2 content, etc.), and economic parameters (ex-factory price including tax) of the ore into a unified representation space.

[0152] For example, given an n-dimensional data point set The method of mapping to the 6-dimensional unit hypersphere is:

[0153]

[0154] in,

[0155] : n-dimensional real number space; : 6-dimensional unit hypersphere; : 7-dimensional real number space.

[0156] The geometric meaning of hypersphere embedding includes:

[0157] Preserve the original data manifold structure (such as local neighborhood relationships)

[0158] Achieve feature dimension compression (n→6, usually n≥10)

[0159] Supports hypersphere metric learning (cosine similarity)

[0160] S214: Perform knowledge-enhanced reasoning;

[0161] By combining industry knowledge (such as smelting theory and ore characteristics), expert experience (such as expert rule base) and historical data to optimize the reasoning process, the intelligence of the model and the accuracy of ore allocation decisions can be improved.

[0162] For example, the cross-modal aligned features are stored in a knowledge base and fused with the existing knowledge in the knowledge base.

[0163] S215: Decision optimization via physics-guided attention weight correction techniques.

[0164] S22: Spatiotemporal Convolution Operator Design

[0165] (1) Spatial domain convolution operator design: The spatial domain convolution operator is optimized through the improved residual graph convolution layer:

[0166]

[0167] in,

[0168] H (l+1) : The node feature matrix of the l+1th layer in the spatial convolutional neural network;

[0169] H (l) : The node feature matrix of the lth layer in the spatial convolutional neural network;

[0170] σ: activation function, which introduces nonlinearity. For example, the ReLU (Rectified Linear Unit) function is used;

[0171] : degree matrix Perform element-by-element -1 / 2 power operation;

[0172] To add a self-connected adjacency matrix, A represents the original adjacency matrix and I represents the identity matrix, that is, a square matrix with the main diagonal elements being 1 and the rest being 0;

[0173] Wr: residual weight.

[0174] (2) Operator design for time domain convolution and optimization of multi-scale dilated causal convolution operators:

[0175]

[0176] in,

[0177] T (l+1) : The time series feature matrix of the l+1th layer in the time domain convolutional neural network;

[0178] T (l) : The time series feature matrix of the lth layer in the time domain convolutional neural network;

[0179] DilatedConv(T (l) ,d=2 k ):To T (l)Perform a dilated causal convolution operation, where d is the dilation factor;

[0180] : Feature fusion operation;

[0181] LSTM: A recurrent neural network that controls the flow of information through a gating mechanism, remembering long-term information and selectively forgetting unimportant information.

[0182] For example, the expansion factors d=1, 2, and 4 are set for three groups of parallel processing, and the time window length is set to 60 minutes (corresponding to the duration of the key sintering stage in the sintering process). Through the above-mentioned time domain convolution operator design, the multi-scale time dependency and long-term dependency features in the time series data of the sintering process are extracted.

[0183] S23: Process knowledge distillation, through physics-guided attention weight correction technology, incorporates sintering operation expert experience in the pre-training stage:

[0184]

[0185] in,

[0186] : The corrected attention weight between node i and node j;

[0187] : The attention weight between node i and node j in the original Transformer model;

[0188] h i : Feature representation of node i in the model;

[0189] : Feature representation of node j incorporating expert experience;

[0190] σ: Variance hyperparameter, used to control the scaling of the exponential term.

[0191] S24: Key constraints of sintering process include the following three aspects:

[0192] S241: Thermodynamic Constraints,

[0193] The basis for constraint calculation is the energy conservation equation:

[0194]

[0195] in,

[0196] m i : the mass of the i-th ore;

[0197] c p,i: Specific heat capacity at constant pressure of the i-th ore;

[0198] T out : system output temperature;

[0199] T in : system input temperature;

[0200] Q comb : the total energy released during the sintering process;

[0201] Q loss : Energy lost during sintering.

[0202] Exemplary methods for performing thermodynamic constraints may be:

[0203] A residual calculation module is added after the model output layer to compare the predicted temperature with the solution of the thermodynamic equation in real time; the Q_loss value is updated every 8 hours using an online thermal imager (error <3%).

[0204] S242: Chemical Reaction Constraints

[0205] The amount of liquid phase generated is limited by calculating the mass ratio of basic oxides to acidic oxides.

[0206] For example, the following chemical composition ratios are set:

[0207]

[0208] Example: Inserting a Constraint Satisfaction Layer (CS-Layer) into the penultimate layer of a neural network

[0209] S243: Equipment Operation Constraints

[0210] Maximum throughput of sintering machine:

[0211]

[0212] in,

[0213] vi: volume flow rate of type i ore, i.e., the volume passing through the sintering machine per unit time;

[0214] ρi: density of the i-th ore;

[0215] 650t / h: The design upper limit of the sintering machine. Here, 650 means that the total mass of materials passing through the sintering machine per hour cannot exceed 650 tons.

[0216] For example, the equipment operation constraint can be used as a trigger parameter for dynamic adjustment: when the predicted value exceeds the limit, the gradient truncation mechanism is triggered.

[0217] S3: Fusion model of multi-objective optimization model for ore distribution and large-scale multi-modal model

[0218] S31: Define the three-objective optimization function:

[0219]

[0220] x i ∈[0,1]: the ratio coefficient of the i-th ore;

[0221] Δp i ~N(μ,σ 2 ): price fluctuation random item;

[0222] f k (x): confidence level of the k-th metallurgical performance index, output by the prediction network;

[0223] a ji : Contribution coefficient of ore i in the j-th constraint;

[0224] S32: Use the improved NSGA-III algorithm to solve the optimization function of S31 and define the reference point adaptive adjustment strategy:

[0225]

[0226] in,

[0227] : The new value of the j-th reference point after adaptive adjustment;

[0228] : The old value before the j-th reference point adjustment;

[0229] η: learning rate;

[0230] H(P) is the population distribution entropy.

[0231] Module 3: Robust Optimization Module

[0232] By integrating robust optimization theory with transfer learning algorithm, an interference-resistant ore allocation optimization model is constructed.

[0233] S1: Robust Dual Transformation for Robust Optimization Theory

[0234]

[0235] in,

[0236] f0(x): main objective function;

[0237] f i(x)(i=1,2,…,m): represents the other m objective functions;

[0238] u i :For the objective function f i The weight coefficient of (x);

[0239] Uncertainty set:

[0240]

[0241] in,

[0242] Γ≥0: budget parameter to control uncertainty;

[0243] σ i : the standard deviation of the i-th interference term;

[0244] S2: Explicitly transform the dual problem

[0245] (1) The strong duality condition is obtained through Legendre-Fenchel transformation:

[0246]

[0247] in,

[0248] : Lagrangian function that integrates the objective function and constraints; p: objective function coefficient vector, p∈R^n;

[0249] A: constraint matrix, A∈R^{m×n};

[0250] η: Dual gap tolerance (taken as 0.01);

[0251] ∈: positive constant; b: constant vector;

[0252] : The gradient of the Lagrangian function L with respect to the variable x.

[0253] S3: Robust Duality Gap Analysis

[0254] (1) Define the approximation error bound:

[0255]

[0256] L: Lipschitz constant of the objective function (calculated based on sintering data)

[0257] δ: Primal-dual variable spacing, δ = |x k -x * |

[0258] μ: strong convexity coefficient (take 0.1)

[0259] (2) Uncertainty modeling:

[0260]

[0261] in,

[0262] U: Uncertainty set (price fluctuation ±15%, component fluctuation ±3%)

[0263] c i (ξ i ): Dynamic cost coefficient (including 23 factors such as transportation and tariffs)

[0264] a ij (ξ i ):Process constraint matrix (sintering temperature, basicity ratio and other related parameters) S4:Transfer learning

[0265] (1) Feature alignment loss function:

[0266]

[0267] in,

[0268] MMD (Maximum Mean Difference): measures the source domain P s With the target domain P t Distribution difference; φ s : Feature extraction function of source domain;

[0269] φ t : Feature extraction function of target domain;

[0270] γ: a hyperparameter that controls the alignment strength, for example, γ = 0.5;

[0271] (2) Rule extraction function:

[0272]

[0273] II: Indicator function, returns 1 if the condition is met, otherwise returns 0;

[0274] R k : kth decision rule;

[0275] Path i : The i-th decision tree path.

[0276] Use this function to refine rules and retain core rules with importance > 0.8.

[0277] Module 4: Closed-loop control module

[0278] Integrate online detection instruments and dynamic parameter correction models to achieve hourly ratio closed-loop adjustment.

[0279] S1: System Architecture and Core Links

[0280] like Figure 2 As shown, the closed-loop control operation mode includes: online detection → dynamic correction → execution control → production feedback → online detection. The core components include:

[0281] (1) Online detection layer, which collects data through sensors.

[0282] For example, data is collected using sensors covering the entire production line, such as an X-ray fluorescence analyzer (composition), a laser particle size analyzer (particle size), and a thermal imager (temperature), with the frequency of collecting key parameters being 10 times per second.

[0283] (2) Dynamic correction engine, integrating LSTM time series prediction model and adaptive PID algorithm, automatically generates parameter adjustment suggestions every hour.

[0284] (3) Execution control layer, including PLC linkage adjustment of belt scale, feeding valve and other actuators; in one embodiment of the present invention, the control signal transmission delay is <50ms.

[0285] S2: Hourly closed-loop control implementation

[0286] The following examples illustrate the core logic of the dynamic correction of the present invention:

[0287] Within one hour, the operation mode of the time-sharing control cycle is:

[0288] (1) 00:00-00:15: Data acquisition, obtaining more than 200 parameters such as composition, particle size, temperature, etc.

[0289] (2) 00:15-00:30: Anomaly detection, screening abnormal indicators based on the 3σ principle;

[0290] (3) 00:30-00:45: Model calculation, generating three sets of candidate solutions for ratio adjustment;

[0291] (4) 00:45-01:00: Execute verification, issue instructions and verify the results.

[0292] This operating mode also includes the following dynamic adjustment strategies:

[0293] (1) Normal mode: fine-tune parameters every hour (flux ±0.5kg / t, belt speed ±0.1m / min)

[0294] (2) Emergency mode: When the component fluctuation is detected > 2%, the adjustment is completed within 5 minutes

[0295] The present invention proposes an intelligent ironmaking optimization system, device, storage medium and program product based on industrial big data, which integrates deep neural networks, multi-objective optimization algorithms and metallurgical process mechanism models to achieve coordinated optimization of raw material cost savings and stable blast furnace operation. It is suitable for ore proportioning decisions at the front end of sintered ore production.

[0296] The following describes the technical effects of the present invention by combing through the test data:

[0297] 1. Dynamic raw material knowledge graph

[0298]

[0299] 2. Large model for metallurgical property prediction

[0300]

[0301]

[0302] 3. Robust optimization decision engine

[0303]

[0304] 4. Online closed-loop control system

[0305]

Claims

1. An intelligent ore blending system before iron production, characterized in that: Includes the following modules: Raw material knowledge graph module: Constructs a three-dimensional feature matrix to describe ore properties, builds a dynamic raw material knowledge graph, and constructs an ore-origin-supplier association database based on knowledge extraction technology to achieve multi-dimensional encoding of raw material properties; Metallurgical prediction module: Build a sintering performance prediction network to adjust the attention mechanism; Within the framework of a multimodal fusion method of large models and material properties, a spatiotemporal graph convolutional network (ST-GCN) is used to model the complex mapping relationship between ore proportions and sinter properties. By integrating the ore proportioning multi-objective optimization model with the large model multimodal model, joint optimization of the model is achieved. Robust optimization module: Integrates robust optimization theory with transfer learning algorithms to build an interference-resistant ore allocation optimization decision model; Closed-loop control module: Integrates online detection instruments and dynamic parameter correction models to achieve hourly ratio closed-loop control; The cyclic operation steps of the closed-loop control include: online detection, dynamic correction, execution control, production feedback, and online detection.

2. The system according to claim 1, wherein: The raw material knowledge graph module includes: Construct a three-dimensional feature matrix describing ore properties; By defining quintuples to describe ore entities, a dynamic ore knowledge graph is constructed.

3. The system according to claim 2, characterized in that The quintuple includes: Use hash values as unique ore identifiers, ore chemical composition vectors, ore physical property sets, ore supply chain relationship matrices, and dynamic feature tensors.

4. The system according to claim 1, wherein: The metallurgical prediction module includes: Using a dual-path attention mechanism to build a deep sintering performance prediction network; Large-scale multimodal models, including the optimization design of spatiotemporal convolution operators within the framework of multimodal fusion methods of large models and material characteristics, process knowledge distillation, and key constraints of sintering processes; Fusion of multi-objective optimization model for ore distribution and multi-modal model of large model.

5. The system according to claim 4, characterized in that The dual-path attention mechanism includes: ore parameter attention and temporal feature attention.

6. The system according to claim 4, characterized in that The multimodal fusion method framework of the large model and material characteristics includes: Obtain raw material modalities based on material testing data; Characterize the raw material modality; Cross-modal alignment of feature encodings via hypersphere embedding; Knowledge-enhanced reasoning, including fusing cross-modal aligned features with a knowledge base; Decision optimization via physics-guided attention weight modification techniques.

7. The system according to claim 4, wherein: The spatiotemporal convolution operator optimization design includes: The spatial domain convolution operator is optimized through an improved residual graph convolution layer; The time domain convolution operator is optimized through multi-scale dilated causal convolution.

8. The system according to claim 4, wherein: The process knowledge distillation is specifically to incorporate the sintering operation expert experience into the pre-training stage through the physics-guided attention weight correction technology.

9. The system according to claim 4, wherein: The key constraints of the sintering process include: Thermodynamic constraints: Add a residual calculation module after the model output layer to compare the predicted temperature with the solution of the thermodynamic equation in real time, and regularly update the Q_loss value through an online thermal imager; Chemical reaction constraints: by calculating the mass ratio of basic oxides to acidic oxides, the amount of liquid phase generated is limited; Equipment operation constraints: The total mass of materials passing through the sintering machine per unit time is used as a trigger parameter for dynamic adjustment.

10. The system according to claim 4, wherein: The fusion of the multi-objective optimization model for ore blending and the large-scale multi-modal model includes: Define a multi-objective optimization function that comprehensively considers economy, metallurgical performance balance, and process constraints; The improved NSGA-III algorithm is used to solve the optimization function and define a reference point adaptive adjustment strategy.

11. The system according to claim 1, wherein: The robust optimization module includes: Robust dual transformation; Explicitly transform the dual problem, including obtaining strong duality conditions via Legendre-Fenchel transformation; Robust duality gap analysis, including: defining approximation error bounds and uncertainty modeling; Transfer learning, including: feature alignment loss and rule extraction.

12. The system according to claim 1, wherein: The closed-loop control module includes the following operating steps: Online detection: data collection through sensors; Dynamic correction: Integrates the LSTM time series prediction model with the adaptive PID algorithm to automatically generate parameter adjustment suggestions every hour; Execution control: control the executive agencies in the system; Production feedback; Return to online testing.

13. The system according to claim 1, wherein: The closed-loop control module includes an hourly closed-loop control mode, and its operation steps include: Data collection: obtain 200+ parameters such as composition, particle size, temperature, etc. Anomaly detection: Screening abnormal indicators based on the 3σ principle; Model calculation: generate three sets of candidate solutions for ratio adjustment; Execution verification: issue instructions and verify the effects; The following dynamic adjustment strategies are also included: Normal mode: fine-tune parameters every hour; Emergency mode: When a component fluctuation >2% is detected, adjustments will be made within 5 minutes.

14. A multimodal fusion method for a large model of a metallurgical prediction module and material properties in a system as claimed in claim 1, characterized in that: include: Obtain raw material modalities based on material testing data; Characterize the raw material modality; Cross-modal alignment of feature encodings via hypersphere embedding; Knowledge-enhanced reasoning, including fusing cross-modal aligned features with a knowledge base; Decision optimization via physics-guided attention weight modification techniques.

15. The method according to claim 14, wherein Also includes: The spatiotemporal convolution operators are optimized by using improved residual graph convolution layers and multi-scale dilated causal convolutions respectively; Through the physics-guided attention weight correction technology, the experience of sintering operation experts is incorporated into the pre-training stage to perform process knowledge distillation; The key constraints of the sintering process include comprehensive thermodynamic constraints, chemical reaction constraints and equipment operation constraints.

16. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the method flow in the system according to any one of claims 1 to 13 or the method according to any one of claims 14 to 15.

17. A device system, comprising a computer device, and also comprising a raw material storage device, a feeding device, a crushing and screening device, a grinding device, a beneficiation device, a batching device, and a conveying device involved in a pre-iron ore blending system, characterized in that: The device system operates in coordination to implement the method flow in the system according to any one of claims 1 to 13 or the method according to any one of claims 14 to 15.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method flow in the system of any one of claims 1 to 13 or the method of any one of claims 14 to 15 is implemented.

19. A computer program product, characterized in that When the computer program is executed by a processor, the method flow in the system of any one of claims 1 to 13 or the method of any one of claims 14 to 15 is implemented.

Images