CN-122024882-A - Nonferrous metal smelting monitoring optimization method based on multi-mode sensing and trans-scale physical neural network

CN122024882ACN 122024882 ACN122024882 ACN 122024882ACN-122024882-A

Abstract

The invention discloses a nonferrous metal smelting monitoring optimization method based on multi-mode sensing and a cross-scale physical neural network, which comprises the steps of constructing a dynamic graph structure according to multi-source sensor data, fusing multi-mode sensing information, extracting space-time characteristics related to carbon content and temperature, inputting the space-time characteristics into the neural network embedded with thermodynamic and reaction dynamics constraints, introducing the cross-scale physical, establishing a multi-scale physical consistency prediction model, constructing a contrast learning mechanism based on the physical coupling relation between the carbon content and the temperature, enhancing the model, adopting a multi-objective optimization algorithm to carry out collaborative iterative optimization on the prediction precision and the physical consistency, and utilizing the optimized model to realize continuous online estimation and abnormal early warning of the carbon content and the temperature. According to the invention, by fusing multi-mode sensing information and embedding cross-scale physical constraints, an interpretable intelligent neural network model is constructed, and high-precision continuous monitoring and process optimization of carbon content and temperature in the nonferrous metal smelting process are realized.

Inventors

YANG KAI
LIU CHUNMEI
Xiao qingtai
XU JIANXIN
WANG HUA

Assignees

昆明理工大学

Dates

Publication Date: 20260512
Application Date: 20260204

Claims (10)

1. The nonferrous metal smelting monitoring and optimizing method based on the multi-mode sensing and the trans-scale physical neural network is characterized by comprising the following steps of: constructing a dynamic graph structure of the space-time correlation and physical coupling relation of the characterization process according to the multi-source sensor data; Based on the dynamic graph structure, multi-mode sensing information is fused, and space-time characteristics related to carbon content and temperature are extracted; Inputting the extracted space-time characteristics into a neural network embedded with thermodynamic and reaction kinetic constraints, and introducing a trans-scale physical embedded neural network to establish a multi-scale physical consistency prediction model from a microscopic reaction mechanism to macroscopic heat and mass transfer; Constructing a contrast learning mechanism based on a physical coupling relation between carbon content and temperature for enhancing the multi-scale physical consistency prediction model, and adopting a multi-objective optimization algorithm to carry out collaborative iterative optimization on prediction precision and physical consistency; And (3) utilizing the optimized multiscale physical consistency prediction model to realize continuous online estimation and abnormal early warning of carbon content and temperature, and outputting process adjustment suggestions through interpretability analysis.
2. The method of claim 1, wherein constructing a dynamic graph structure characterizing a process space-time correlation and physical coupling relationship from the multi-source sensor data comprises: According to spectrum, infrared thermal imaging and acoustic emission multi-mode data acquired by the sensors, taking data feature vectors corresponding to the sensors as graph nodes; Constructing an adjacency matrix among nodes according to the process time sequence, the equipment space layout and the flow field similarity obtained by computational fluid dynamics simulation; and according to the constructed adjacency matrix, adopting a graph attention network to dynamically fuse node characteristics and allocate weights.
3. The method according to claim 1, wherein based on the dynamic graph structure, the multi-modal sensing information is fused and the space-time characteristics related to carbon content and temperature are extracted, and the specific process comprises: according to the spectrum data, after baseline correction and noise reduction treatment, extracting the characteristic peak intensity and peak area of the specific element; According to the infrared thermal imaging data, acquiring a two-dimensional temperature distribution matrix of the surface of the melt after emissivity correction; extracting time-frequency characteristics of energy, frequency and ringing count after wavelet transformation according to the acoustic emission signal; according to the infrared temperature distribution matrix, a convolutional neural network is adopted to extract the spatial distribution characteristics of the infrared temperature distribution matrix; Extracting the time evolution characteristics of the acoustic emission time-frequency characteristic sequence by adopting a long-short-time memory network according to the spectral characteristic peak sequence and the acoustic emission time-frequency characteristic sequence; And according to all the features extracted from the graph nodes, the infrared images and the time sequence, splicing and fusing to form a unified space-time fusion feature vector.
4. The method of claim 1, wherein inputting the extracted spatiotemporal features into a neural network embedded with thermodynamic and reaction kinetic constraints comprises: Constructing physical residual constraint according to a first law of thermodynamics and a carbon-oxygen reaction kinetic equation; And constructing a weighted sum as a total loss function of the neural network according to the data fitting loss and the constructed physical residual constraint.
5. The method according to claim 4, wherein constructing the total loss function, in particular further comprises: And dynamically adjusting the weights of the data fitting loss term and the physical residual constraint term in the total loss function according to different stages of the smelting process.
6. The method of claim 1, wherein introducing a cross-scale physical embedded neural network to build a multi-scale physical consistency prediction model comprises: Respectively constructing corresponding special physical sub-networks according to microscopic reaction dynamics, mesoscopic phase interface evolution and macroscopic heat and mass transfer processes; embedding a control equation describing a physical process of a corresponding scale into each scale sub-network as a training constraint; information interaction and feature fusion among sub-networks with different scales are realized through a cross-scale attention mechanism; constraining collaborative optimization of microscopic and macroscopic physical fields in a training process by cross-scale gradient alignment loss; And constructing a multi-scale physical consistency loss function, and ensuring the compatibility and continuity of different scale prediction results in a physical sense.
7. The method of claim 1, wherein for the multi-scale physical consistency prediction model, a contrast learning mechanism is constructed based on a physical coupling relation between carbon content and temperature to be enhanced, and a multi-objective optimization algorithm is adopted to perform collaborative iterative optimization, and the specific process comprises: Constructing a theoretical coupling relation between carbon content change and temperature change according to a mass conservation and energy conservation equation; constructing a cross-modal positive sample pair according to the theoretical coupling relation and the data space-time proximity; designing a contrast loss function to enable the similarity of the positive sample pair in the characteristic space to be related to the theoretical physical coupling strength of the positive sample pair; constructing a multi-objective optimization function set according to the carbon content prediction error, the temperature inversion error and the physical constraint total residual error; Under the Newton-Laportson iteration frame, calculating a joint Jacobian matrix of model parameters relative to each optimization target; Processing the candidate solution set generated by iteration and generating a pareto front by adopting a non-dominant sorting and crowding distance sorting technology; According to the real-time process demand, dynamically selecting an optimal solution from the pareto front to update the model parameters.
8. The method of claim 1, wherein the continuous on-line estimation and anomaly pre-warning of carbon content and temperature is achieved using an optimized multi-scale physical consistency prediction model, and process adjustment suggestions are output through an interpretability analysis, the specific process comprising: Initializing prototype vectors representing carbon content and temperature characteristics of stable working conditions; Dynamically adjusting a prototype vector by adopting a momentum update mechanism according to the characteristic representation obtained by online prediction so as to track slow drift of the process; calculating the distance between the current feature representation and the corresponding prototype vector, and setting a dynamic early warning threshold according to statistics of the historical distance; Triggering abnormal state early warning according to the condition that the current distance exceeds a dynamic threshold value; analyzing the contribution degree of the prediction output to the final prediction result by calculating the gradient of the prediction output to the input sensor data and each physical constraint item; Generating a thermodynamic diagram, and identifying a key process stage and a molten pool sensitive area; and generating a structured report containing weak link positioning and parameter adjustment suggestions according to the contribution analysis result.
9. An electronic device comprising a memory, a processor and a computing program stored in the memory and executable on the processor, characterized in that the processor implements the method of any of claims 1-8 when executing the computing program.
10. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the method of any of claims 1-8.

Description

Nonferrous metal smelting monitoring optimization method based on multi-mode sensing and trans-scale physical neural network Technical Field The invention belongs to the technical field of intelligent monitoring and process optimization of nonferrous metal smelting processes, and particularly relates to a nonferrous metal smelting monitoring and optimizing method based on a multi-mode sensing and trans-scale physical neural network. Background In nonferrous metal smelting, real-time and accurate measurement of the carbon content and the temperature field in the melt is a core precondition for guaranteeing the product quality, optimizing the energy consumption and realizing intelligent control. However, current technical applications in this field still face a number of challenges. Traditional measurement methods rely mainly on manual off-line sampling and laboratory analysis. The method has serious response delay, cannot capture continuous dynamic change of the melt state, is easy to introduce interference and safety risks in the sampling process, and is difficult to meet the requirement of real-time closed-loop control in the modern smelting process. In order to overcome the hysteresis of the off-line analysis, an on-line sensing technology based on infrared temperature measurement, spectral analysis or acoustic emission monitoring is applied. However, the smelting environment has the characteristics of high temperature, multiple phases and strong interference, a single-mode sensing signal is easy to be polluted by noise, the information dimension is limited, complex multi-physical field coupling states inside a melt are difficult to comprehensively and robustly reflect, and the problems of incomplete information and insufficient reliability exist. With the development of artificial intelligence technology, a machine learning model based on pure data driving is introduced to conduct predictive analysis. Although such models can mine data correlation, the embedding of underlying physical laws such as thermodynamics, reaction kinetics and the like is generally lacking. This results in a sudden drop in generalization ability when the model changes in operating conditions, data is sparse, or abnormal conditions not covered by the training set occur, the prediction result may deviate from the physical common sense, and the reliability is low in practical industrial application. The smelting process is essentially a multi-scale process with close coupling of microscopic reaction, mesoscopic phase change and macroscopic flow heat and mass transfer. The existing modeling method is often limited to a single scale or is used for simply splicing models with different scales, so that consistency modeling and collaborative solving of a trans-scale physical rule are difficult to realize, the prediction results among different scales are often contradictory, and the physical quantity is discontinuous at the scale juncture, so that the prediction precision and the physical credibility of the model are limited. In addition, industrial sites have a judicious attitude towards "black box" models. The existing prediction model is often poor in interpretation, the influence of key process parameters cannot be clearly disclosed, and a fault source is positioned, so that a craftsman is difficult to formulate a reliable operation adjustment strategy according to model output, and the deep application of the prediction model in actual production is hindered. In recent years, a Physical Information Neural Network (PINN) provides a new paradigm for fusing data and physical laws. The control equation is used as a regularization term of a loss function, and the constraint neural network learns solutions conforming to the physical laws. However, the conventional PINN method still has obvious limitations when facing to a complex industrial process, namely, the method lacks a system fusion mechanism for multi-source heterogeneous sensing information, secondly, the method can not effectively solve the problems of microscale to macroscopic multiscale physical embedding and collaborative optimization, thirdly, when facing to a continuously operated industrial system, the self-adaptive optimization capability and the online abnormality early warning function are weak, fourthly, the interpretability form of the model output is single, and the process decision is difficult to directly support. Therefore, the prior art fails to provide a carbon content and temperature continuous measurement and analysis method which can systematically fuse multi-mode sensing information, deeply embed and cooperatively optimize a multi-scale physical mechanism, has online self-adaption and abnormality early warning capability, and can output high-interpretation guidance. Disclosure of Invention In order to solve the technical problems, the invention provides a nonferrous metal smelting monitoring and optimizing method based on multi-mode sensing and a cross-scale