Inorganic mineral fiber production system and method based on intelligent control

Abstract

The invention discloses an inorganic mineral fiber production system and method based on intelligent control, and relates to the technical field of artificial intelligence. The temperature distribution diagram, the flow velocity distribution diagram and the conductivity distribution diagram are fused into a multi-physical-field distribution diagram, the three-dimensional shape model and the multi-physical-field distribution diagram are coupled, and melt state parameters are extracted; drawing process parameters and fiber quality indexes are collected, the causal relationship and causal strength among the melt state parameters, the drawing process parameters and the fiber quality indexes are analyzed, a weighted Bayesian network is constructed, and a key reason set is searched; training a source domain quality prediction model, and migrating the source domain quality prediction model to a target domain to obtain a target domain quality prediction model; and obtaining candidate wire drawing process parameters according to the key reason set, predicting a fiber quality index prediction value, calculating a quality error as feedback, and designing an optimization objective function to search for the optimal wire drawing process parameters.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to an inorganic mineral fiber production system and method based on intelligent control. Background Art

[0002] Using industrial solid wastewater-quenched slag as raw material is a new approach to developing high-strength, high-modulus inorganic mineral fibers. The melt-drawing process directly determines the fiber's microstructure and ultimate performance. However, existing production processes still have some deficiencies in intelligent control. Traditional process parameter settings rely primarily on operator experience and historical data. As raw material composition and production environment change, process parameters often cannot be dynamically adjusted based on real-time data, resulting in quality fluctuations and low production efficiency.

[0003] Traditional production control usually relies on experience and trial and error in setting process parameters in the melt drawing process, which is often difficult to cope with the dynamic changes in raw material characteristics and working conditions, and the adjustment is delayed and the quality fluctuates greatly. Summary of the Invention

[0004] The present application aims to at least partially address one of the technical problems in the related art. To this end, one purpose of the present application is to propose an inorganic mineral fiber production system and method based on intelligent control, thereby improving the production quality of the melt drawing process of inorganic mineral fiber production, thereby improving the final quality of the inorganic mineral fiber production.

[0005] One aspect of the present application provides a method for producing inorganic mineral fibers based on intelligent control, comprising:

[0006] Step S100: acquiring an image sequence of the melt surface at different viewing angles, performing feature extraction and semantic segmentation on the image sequence to obtain image features and semantic masks, and constructing a three-dimensional morphology model of the melt surface;

[0007] Step S200: obtaining a temperature distribution map, a flow velocity distribution map, and a conductivity distribution map of the melt surface, fusing the temperature distribution map, the flow velocity distribution map, and the conductivity distribution map into a multi-physics field distribution map, coupling the three-dimensional morphology model with the multi-physics field distribution map, and extracting melt state parameters;

[0008] Step S300: collecting drawing process parameters and fiber quality indicators, analyzing the causal relationship and causal strength among melt state parameters, drawing process parameters and fiber quality indicators, constructing a weighted Bayesian network, and searching for a key cause set for each fiber quality indicator;

[0009] Step S400: using historical melt state parameters, drawing process parameters, and fiber quality indicators to train a source domain quality prediction model, and migrating the source domain quality prediction model to a target domain to obtain a target domain quality prediction model;

[0010] Step S500: Obtain candidate drawing process parameters based on the key cause set, take the candidate drawing process parameters and melt state parameters as input, predict the fiber quality index prediction value based on the target domain quality prediction model, calculate the quality error as feedback, and design an optimization objective function to search for the optimal drawing process parameters.

[0011] The specific method of obtaining an image sequence of the melt surface at different viewing angles, performing feature extraction and semantic segmentation on the image sequence to obtain image features and semantic masks, and constructing a three-dimensional morphology model of the melt surface is as follows:

[0012] Step S110: Setting the melt surface imaging perspectives, and acquires a sequence of continuous frames at each imaging perspective ;in, Indicates the imaging angle number, , t represents the time step, , represents the total number of time steps;

[0013] Step S120: Extract features from the image sequence at each imaging perspective to obtain the image features of each image at each imaging perspective and each time step. , perform semantic segmentation on the images at each imaging perspective and each time step in the image sequence to obtain the semantic mask of each image ;

[0014] Step S130: For time step t, map the image features and semantic mask to three-dimensional space to obtain a three-dimensional feature point cloud and a three-dimensional semantic mask point cloud, and construct a voxel grid V in the three-dimensional space. For each voxel , calculate voxel The eigenvalues ​​between the 3D feature point cloud at each imaging perspective , calculate voxel The semantic value between the 3D semantic mask point cloud at each imaging perspective ;

[0015] Step S140: Setting a threshold for semantic value binarization , binarize the semantic value of the voxel to obtain the three-dimensional semantic mask voxel ;

[0016] Step S150: Extract the center coordinates of the voxels with a median value of 1 in the 3D semantic mask voxels to form the 3D point cloud coordinates at time step t ; The three-dimensional point cloud coordinates include M three-dimensional points;

[0017] Step S160: Continuous The three-dimensional point cloud coordinates of each time step are spliced ​​to obtain a complete three-dimensional morphology model of the melt surface.

[0018] The specific method of obtaining the temperature distribution map, flow velocity distribution map and conductivity distribution map of the melt surface, fusing the temperature distribution map, flow velocity distribution map and conductivity distribution map into a multi-physics field distribution map, coupling the three-dimensional morphology model and the multi-physics field distribution map, and extracting the melt state parameters is as follows:

[0019] Step S210: Obtaining the temperature distribution diagram of the melt surface , velocity distribution diagram and conductivity distribution map , the temperature distribution map, velocity distribution map and conductivity distribution map are weighted averaged to obtain the multi-physics field distribution map ;

[0020] Step S220: Couple the 3D topography model and the multi-physics field distribution map. For each time step t and each 3D point , according to the three-dimensional point At the corresponding position in the multi-physics field distribution map, extract the multi-physics field value at that position , combining the 3D points with the multi-physics field values ​​to obtain the coupled characteristics of the melt state , and obtain the final multi-physics field coupling characterization model ;

[0021] Step S230: For each time step, the coupling characteristics of the melt state are extracted from the multi-physics field coupling characterization model as the melt state parameters of the current time step.

[0022] The specific method of collecting the drawing process parameters and fiber quality indicators, analyzing the causal relationship and causal strength among the melt state parameters, the drawing process parameters and the fiber quality indicators, constructing a weighted Bayesian network, and searching for the key cause set of each fiber quality indicator is as follows:

[0023] Step S310: collecting melt state parameters S and drawing process parameters P during the historical drawing process, performing quality inspection on the produced inorganic mineral fiber, and obtaining its fiber quality index Q;

[0024] Step S320: Divide the acquired melt state parameters, drawing process parameters and fiber quality indicators into Nd groups of sample data, and learn the structure G and parameters θ of the Bayesian network based on the sample data;

[0025] Step S330: In the Bayesian network In the equation, node v represents each parameter among melt state parameters, drawing process parameters and fiber quality indexes, and edge u represents the causal relationship between each parameter. For parameter Xi and parameter Xj, if there is a directed edge between vi and vj, it means that Xi is the direct cause of Xj.

[0026] Step S340: Calculate the mutual information between each parameter Xi and parameter Xj , average causal effect and causal certainty ;

[0027] Step S350: define causal strength, for each pair of directed edges in the directed acyclic graph , calculate the causal strength between parameter Xi and parameter Xj based on mutual information, average causal effect and causal confidence , assign the calculated causal strength to the corresponding edge in the Bayesian network to obtain a weighted Bayesian network;

[0028] Step S360: For each fiber quality index , find its key cause set in the weighted Bayesian network ,in, For the A key cause node.

[0029] The method for obtaining the key cause set is:

[0030] Step S361: For each fiber quality index , find out the set of all direct cause nodes in the weighted Bayesian network ;

[0031] Step S362: For each direct cause node in the direct cause node set, calculate the direct cause node and the fiber quality index. The causal strength between nodes;

[0032] Step S363: Sort the causal strength from large to small, and select the Direct cause nodes are used as key cause nodes to form a key cause set .

[0033] The specific method of using the historical melt state parameters, drawing process parameters, and fiber quality indicators to train the source domain quality prediction model and migrating the source domain quality prediction model to the target domain to obtain the target domain quality prediction model is as follows:

[0034] Step S410: Obtain historical samples based on the melt state parameters, drawing process parameters and fiber quality indicators at historical times , train the source domain quality prediction model and obtain the source domain quality prediction model parameters ,in, Represents the characteristic vector composed of the melt state parameters and drawing process parameters of the nth historical sample, is the corresponding fiber quality index, is the number of historical samples;

[0035] Step S420: For a new batch of water-quenched slag raw materials, collect target domain data ,in, Indicates the The characteristic vector composed of the melt state parameters and drawing process parameters of the target domain data, is the corresponding fiber quality index, is the number of samples of target domain data;

[0036] Step S430: Use the migration algorithm to migrate the source domain quality prediction model to the target domain, using the source domain quality prediction model parameters as the initial values ​​and using the target domain data. Optimize the source domain quality prediction model to obtain the target domain quality prediction model, design the domain adaptation loss function and the prediction loss function, and construct the loss function by weighted summation;

[0037] Step S440: Taking minimizing the loss function as the training objective, find the optimal quality prediction model parameters ,When the loss function reaches convergence, the training is completed and the trained target domain quality prediction model is obtained;

[0038] Step S450: Design a sliding window W and preset the window length , stores the most recent window length When the melt state parameters and drawing process parameters in production change, the target domain data in the sliding window is used to update the target domain quality prediction model parameters, and the updated target domain quality prediction model parameters are used to predict the fiber quality indicators of the next batch.

[0039] The specific method for obtaining candidate wire drawing process parameters according to the key cause set, taking the candidate wire drawing process parameters and melt state parameters as input, predicting the fiber quality index prediction value based on the target domain quality prediction model, calculating the quality error as feedback, and designing an optimization objective function to search for the optimal wire drawing process parameters is as follows:

[0040] Step S510: obtaining melt state parameters of the current wire drawing process, taking the wire drawing process parameters in the key cause concentration as candidate wire drawing process parameters, and initializing the values ​​of the candidate wire drawing process parameters;

[0041] Step S520: using the current melt state parameters and the candidate drawing process parameters as input, and using the trained target domain quality prediction model to predict the fiber quality index prediction value;

[0042] Step S530: Fiber quality index prediction value and fiber quality index target value Compare and calculate the mass error ;

[0043] Step S540: Using the quality error as feedback, design an optimization objective function to search for the optimal drawing process parameters ;

[0044] Step S550: using a genetic algorithm to solve the optimization objective function and obtain the optimal drawing process parameters, taking the current melt state parameters and the optimal drawing process parameters as input data, and using the target domain quality prediction model to output the optimized fiber quality index prediction value;

[0045] Step S560: Calculate the quality error between the optimized fiber quality index prediction value and the fiber quality index target value. When the quality error is less than the error threshold, apply the current optimal drawing process parameters to production control. Otherwise, continue to optimize until the quality error is less than the error threshold.

[0046] One aspect of the present application provides an inorganic mineral fiber production system based on intelligent control, comprising:

[0047] The melt surface modeling module is used to obtain image sequences of the melt surface at different viewing angles, perform feature extraction and semantic segmentation on the image sequences, obtain image features and semantic masks, and construct a three-dimensional morphology model of the melt surface;

[0048] A state parameter acquisition module is used to obtain the temperature distribution map, flow velocity distribution map and conductivity distribution map of the melt surface, fuse the temperature distribution map, flow velocity distribution map and conductivity distribution map into a multi-physics field distribution map, couple the three-dimensional morphology model and the multi-physics field distribution map, and extract the melt state parameters;

[0049] The key cause search module is used to collect drawing process parameters and fiber quality indicators, analyze the causal relationship and causal strength between melt state parameters, drawing process parameters and fiber quality indicators, construct a weighted Bayesian network, and search for the key cause set of each fiber quality indicator;

[0050] The fiber quality prediction module is used to train the source domain quality prediction model using historical melt state parameters, drawing process parameters, and fiber quality indicators, and migrate the source domain quality prediction model to the target domain to obtain the target domain quality prediction model;

[0051] The optimal process optimization module is used to obtain candidate drawing process parameters based on the key cause set, take the candidate drawing process parameters and melt state parameters as input, predict the fiber quality index prediction value based on the target domain quality prediction model, calculate the quality error as feedback, and design the optimization objective function to search for the optimal drawing process parameters.

[0052] One aspect of the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in an inorganic mineral fiber production method based on intelligent control.

[0053] One aspect of the present application provides a readable storage medium storing a computer program, wherein the computer program is suitable for being loaded by a processor to execute steps in a method for producing inorganic mineral fibers based on intelligent control.

[0054] The inorganic mineral fiber production system and method based on intelligent control proposed in this application have the following advantages over the existing technology:

[0055] This application uses computer vision technology to achieve refined three-dimensional reconstruction of the melt surface morphology. By coupling the multi-physical field distribution map with the three-dimensional morphology model, the melt state parameters that comprehensively reflect the geometric morphology and physical properties of the melt are obtained, and the spatiotemporal evolution law of the melt is more comprehensively portrayed.

[0056] This application constructs a weighted Bayesian network model to characterize the probabilistic dependency between variables in a complex system. By searching for the key cause set of each fiber quality indicator, the key factors affecting fiber quality are discovered, and the most effective control measures for quality improvement are determined.

[0057] This application uses a transfer learning strategy to migrate the source domain quality prediction model, trained with historical data, to the target domain. The model is then optimized using target domain data, resulting in a target domain quality prediction model for new batches of raw materials. Furthermore, by designing a sliding window and continuously updating the prediction model using real-time data acquired during production, online adaptive optimization of the model is achieved.

[0058] This application uses the error between the predicted value and the target value of the quality indicator output by the model as feedback, searches for the optimal process parameters by designing an optimization objective function, and realizes intelligent closed-loop optimization control of the production process. The wire drawing process parameters where key causes are concentrated are taken as the main objects of optimization, which reduces the complexity of optimization while ensuring the effectiveness of optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A flow chart of the method for producing inorganic mineral fibers based on intelligent control provided in this application;

[0060] Figure 2 Flowchart of the method for obtaining the target domain quality prediction model provided in this application;

[0061] Figure 3 Flowchart of the optimal wire drawing process parameter optimization method provided for this application;

[0062] Figure 4 This is a functional module diagram of the inorganic mineral fiber production system based on intelligent control provided in this application. DETAILED DESCRIPTION

[0063] To better understand the present application, various aspects of the present application will be described in more detail with reference to the accompanying drawings. It should be understood that these detailed descriptions are merely descriptions of exemplary embodiments of the present application and are not intended to limit the scope of the present application in any way. Throughout the specification, the same reference numerals refer to the same elements. The expression "and / or" includes any and all combinations of one or more of the associated listed items.

[0064] In the accompanying drawings, the size, dimensions, and shapes of the elements have been slightly adjusted for ease of illustration. The accompanying drawings are for illustration only and are not drawn strictly to scale. As used herein, the terms "substantially," "approximately," and similar terms are used to indicate approximate values, not degrees, and are intended to illustrate inherent deviations in measurements or calculations that would be recognized by a person of ordinary skill in the art. In addition, in this application, the order in which the steps are described does not necessarily represent the order in which these steps would occur in actual operation, unless otherwise specified or inferred from the context.

[0065] It should also be understood that expressions such as "including", "comprising", "having", "containing" and / or "comprising" are open rather than closed expressions in this specification, which indicate the presence of the stated features, elements and / or components, but do not exclude the presence of one or more other features, elements, components and / or combinations thereof. In addition, when expressions such as "at least one of..." appear after a list of listed features, they modify the entire list of features rather than just the individual elements in the list. In addition, when describing embodiments of the present application, "may" is used to mean "one or more embodiments of the present application". And, the term "exemplary" is intended to refer to an example or illustration.

[0066] Unless otherwise defined, all terms used herein (including engineering terms and scientific and technical terms) have the same meaning as commonly understood by those skilled in the art to which this application belongs. It should also be understood that, unless otherwise specified in this application, words defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an idealized or overly formal sense.

[0067] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0068] Example 1

[0069] like Figure 1 As shown, the inorganic mineral fiber production method based on intelligent control provided by this application includes:

[0070] Step S100: acquiring an image sequence of the melt surface at different viewing angles, performing feature extraction and semantic segmentation on the image sequence to obtain image features and semantic masks, and constructing a three-dimensional morphology model of the melt surface;

[0071] The specific method of obtaining an image sequence of the melt surface at different viewing angles, performing feature extraction and semantic segmentation on the image sequence to obtain image features and semantic masks, and constructing a three-dimensional morphology model of the melt surface is as follows:

[0072] Step S110: Setting the melt surface imaging perspectives, and acquires a sequence of continuous frames at each imaging perspective ;in, Indicates the imaging angle number, , t represents the time step, , represents the total number of time steps;

[0073] Step S120: Extract features from the image sequence at each imaging perspective to obtain the image features of each image at each imaging perspective and each time step. , perform semantic segmentation on the images at each imaging perspective and each time step in the image sequence to obtain the semantic mask of each image ;

[0074] The semantic mask refers to a binary image containing information on whether each pixel in each image belongs to the melt surface area. If the pixel belongs to the melt surface area, the value of the semantic mask is 1; if the pixel belongs to the background area, the value of the semantic mask is 0.

[0075] Step S130: For time step t, map the image features and semantic mask to three-dimensional space to obtain a three-dimensional feature point cloud and a three-dimensional semantic mask point cloud, and construct a voxel grid V in the three-dimensional space. For each voxel , calculate voxel The eigenvalues ​​between the 3D feature point cloud at each imaging perspective , calculate voxel The semantic value between the 3D semantic mask point cloud at each imaging perspective ;

[0076] The calculation formula of the characteristic value is: ,in, Voxel Hedi The weight between the b-th three-dimensional feature point clouds under the imaging perspective, For the The image features of the b-th 3D feature point cloud under the imaging perspective;

[0077] The voxel Hedi The calculation formula of the weight between the b-th three-dimensional feature point clouds under the imaging perspective is: ,in, Voxel The center and The spatial distance between the b-th three-dimensional feature point clouds under the imaging perspective, is a parameter that controls the speed of weight decay. represents the exponential function with base e;

[0078] The value of the parameter controlling the weight decay rate is set by those skilled in the art based on experience.

[0079] The calculation formula of the semantic value is: ,in, Indicates the The semantic mask of the b'th 3D semantic mask point cloud under the imaging perspective, Representing voxels Hedi The weight between the b'th 3D semantic mask point clouds under the b'th imaging perspective;

[0080] The voxel Hedi The calculation formula of the weight between the b'th 3D semantic mask point clouds under the imaging perspective is: ,in, Voxel The center and The spatial distance between the b'th 3D semantic mask point clouds under the b'th imaging perspective;

[0081] Step S140: Setting a threshold for semantic value binarization , binarize the semantic value of the voxel to obtain the three-dimensional semantic mask voxel ;

[0082] The expression of the three-dimensional semantic mask voxel is: ;

[0083] The threshold value for the semantic value binarization is set by those skilled in the art based on experience.

[0084] Step S150: Extract the center coordinates of the voxels with a median value of 1 in the 3D semantic mask voxels to form the 3D point cloud coordinates at time step t ; The three-dimensional point cloud coordinates include M three-dimensional points;

[0085] M is the number of points in the 3D point cloud;

[0086] The expression of the three-dimensional point cloud coordinates at the time step t is: ,in, Voxel The center coordinates of

[0087] Step S160: Continuous The three-dimensional point cloud coordinates of each time step are spliced ​​to obtain a complete three-dimensional morphology model of the melt surface.

[0088] Step S200: obtaining a temperature distribution map, a flow velocity distribution map, and a conductivity distribution map of the melt surface, fusing the temperature distribution map, the flow velocity distribution map, and the conductivity distribution map into a multi-physics field distribution map, coupling the three-dimensional morphology model with the multi-physics field distribution map, and extracting melt state parameters;

[0089] The specific method of obtaining the temperature distribution map, flow velocity distribution map and conductivity distribution map of the melt surface, fusing the temperature distribution map, flow velocity distribution map and conductivity distribution map into a multi-physics field distribution map, coupling the three-dimensional morphology model and the multi-physics field distribution map, and extracting the melt state parameters is as follows:

[0090] Step S210: Obtaining the temperature distribution diagram of the melt surface , velocity distribution diagram and conductivity distribution map , the temperature distribution map, velocity distribution map and conductivity distribution map are weighted averaged to obtain the multi-physics field distribution map ;

[0091] The calculation formula of the multi-physics field distribution map is: ,in, 、 、 are the weight coefficients of temperature distribution map, velocity distribution map and conductivity distribution map respectively;

[0092] The sum of the weight coefficients of the temperature distribution map, the flow velocity distribution map, and the conductivity distribution map is 1, and the specific values ​​are set by those skilled in the art according to actual needs.

[0093] Step S220: Couple the 3D topography model and the multi-physics field distribution map. For each time step t and each 3D point , according to the three-dimensional point At the corresponding position in the multi-physics field distribution map, extract the multi-physics field value at that position , combining the 3D points with the multi-physics field values ​​to obtain the coupled characteristics of the melt state , and obtain the final multi-physics field coupling characterization model ;

[0094] Step S230: For each time step, the coupling characteristics of the melt state are extracted from the multi-physics field coupling characterization model as the melt state parameters of the current time step.

[0095] The multi-physics field coupling characterization model in the above steps integrates the geometric characteristics of the melt surface morphology and the physical characteristics of the multi-physics field distribution. It can comprehensively characterize the spatiotemporal evolution of the melt state and provide an important data basis for subsequent quality influencing mechanism analysis and process parameter optimization.

[0096] Step S300: collecting drawing process parameters and fiber quality indicators, analyzing the causal relationship and causal strength among melt state parameters, drawing process parameters and fiber quality indicators, constructing a weighted Bayesian network, and searching for a key cause set for each fiber quality indicator;

[0097] The specific method of collecting the drawing process parameters and fiber quality indicators, analyzing the causal relationship and causal strength among the melt state parameters, the drawing process parameters and the fiber quality indicators, constructing a weighted Bayesian network, and searching for the key cause set of each fiber quality indicator is as follows:

[0098] Step S310: collecting melt state parameters S and drawing process parameters P during the historical drawing process, performing quality inspection on the produced inorganic mineral fiber, and obtaining its fiber quality index Q;

[0099] Among them, the melt state parameters include temperature, flow rate and conductivity value, the drawing process parameters include drawing speed, traction force and cooling conditions, and the fiber quality indicators include diameter, strength and modulus.

[0100] Step S320: Divide the acquired melt state parameters, drawing process parameters and fiber quality indicators into Nd groups of sample data, and learn the structure G and parameters θ of the Bayesian network based on the sample data;

[0101] The Bayesian network learning method includes structure learning and parameter learning, and the specific method is:

[0102] Step S321: the structure learning refers to learning the topological structure of the Bayesian network from the sample data, using the K2 algorithm to initialize the parent node set of all nodes to an empty set;

[0103] Step S322: For each node vi, enumerate all its parent node sets and calculate the scoring function value of the node after adding different parent node sets;

[0104] The calculation formula of the scoring function value is: ,in, Represents the parent node set of node vi, represents the number of samples, k1 is the number of nodes in the parent node set of node vi, log represents the logarithmic function, Represents a set of known parent nodes The conditional probability distribution of node vi under ;

[0105] Step S323: Select the parent node set that maximizes the scoring function value of the node, and update the parent node set of node vi;

[0106] Step S324: Repeat steps S322 to S323 until the maximum number of iterations is reached;

[0107] The maximum number of iterations is set by those skilled in the art according to actual needs;

[0108] Step S325: The parameter learning adopts Bayesian estimation. For node vi and a given parent node set, calculate the node vi in ​​the parent node set. The value is The conditional probability under , get all the conditional probability values ​​in the conditional probability table of node vi, calculate the conditional probability table of all nodes to get the parameters of the Bayesian network.

[0109] Step S330: In the Bayesian network In the equation, node v represents each parameter among melt state parameters, drawing process parameters and fiber quality indexes, and edge u represents the causal relationship between each parameter. For parameter Xi and parameter Xj, if there is a directed edge between vi and vj, it means that Xi is the direct cause of Xj.

[0110] The parameter Xi and the parameter Xj correspond to the node vi and the node vj respectively.

[0111] Step S340: Calculate the mutual information between each parameter Xi and parameter Xj , average causal effect and causal certainty ;

[0112] The calculation formula of the mutual information is: ,in, represents the joint probability distribution of nodes vi and vj, and are the edge probability distributions of nodes vi and vj respectively.

[0113] The average causal effect represents the average change in vj when node vi changes by one unit;

[0114] The average causal effect is calculated as follows: ,in, Indicates that the node vi is fixed to Xi+1, Indicates that the node vi is fixed to Xi, Indicates Under the intervention of expected value;

[0115] The causal confidence level indicates the degree of confidence in the causal relationship between parameter Xi and parameter Xj;

[0116] The calculation method of the causal confidence is: ,in, is a hyperparameter, represents the partial correlation coefficient between nodes vi and vj;

[0117] The partial correlation coefficient measures the correlation between nodes vi and vj while controlling other variables. The hyperparameter is used to control the growth rate of causal confidence and is set by those skilled in the art based on experience.

[0118] Step S350: define causal strength, for each pair of directed edges in the directed acyclic graph , calculate the causal strength between parameter Xi and parameter Xj based on mutual information, average causal effect and causal confidence , assign the calculated causal strength to the corresponding edge in the Bayesian network to obtain a weighted Bayesian network;

[0119] The calculation formula of the causal strength is: ,in, is the average causal effect of parameter Xi on parameter Xj, is the causal confidence of parameter Xi on parameter Xj;

[0120] Step S360: For each fiber quality index , find its key cause set in the weighted Bayesian network ,in, For the A key cause node.

[0121] The method for obtaining the key cause set is:

[0122] Step S361: For each fiber quality index , find out the set of all direct cause nodes in the weighted Bayesian network ;

[0123] Step S362: For each direct cause node in the direct cause node set, calculate the direct cause node and the fiber quality index. The causal strength between nodes;

[0124] Step S363: Sort the causal strength from large to small, and select the Direct cause nodes are used as key cause nodes to form a key cause set .

[0125] The parameters in the key cause concentration are composed of melt state parameters and drawing process parameters, that is, the melt state parameters and drawing process parameters are causes, and the fiber quality indicators are the results.

[0126] Step S400: using historical melt state parameters, drawing process parameters, and fiber quality indicators to train a source domain quality prediction model, and migrating the source domain quality prediction model to a target domain to obtain a target domain quality prediction model;

[0127] The specific method of using the historical melt state parameters, drawing process parameters, and fiber quality indicators to train the source domain quality prediction model and migrating the source domain quality prediction model to the target domain to obtain the target domain quality prediction model is as follows:

[0128] Step S410: Obtain historical samples based on the melt state parameters, drawing process parameters and fiber quality indicators at historical times , train the source domain quality prediction model and obtain the source domain quality prediction model parameters ,in, Represents the characteristic vector composed of the melt state parameters and drawing process parameters of the nth historical sample, is the corresponding fiber quality index, is the number of historical samples;

[0129] Step S420: For a new batch of water-quenched slag raw materials, collect target domain data ,in, Indicates the The characteristic vector composed of the melt state parameters and drawing process parameters of the target domain data, is the corresponding fiber quality index, is the number of samples of target domain data;

[0130] The number of samples of the target domain data is much smaller than the number of historical samples.

[0131] Step S430: Use the migration algorithm to migrate the source domain quality prediction model to the target domain, using the source domain quality prediction model parameters as the initial values ​​and using the target domain data. Optimize the source domain quality prediction model to obtain the target domain quality prediction model, design the domain adaptation loss function and the prediction loss function, and construct the loss function by weighted summation;

[0132] The calculation formula of the loss function is: ,in, is the prediction loss function, is the domain adaptation loss function, is the weight coefficient, Predict model parameters for target domain quality;

[0133] described The value of is set by those skilled in the art based on experience.

[0134] The calculation formula of the domain adaptation loss function is: ,in, is the feature mapping function, which is used to map the input feature vector to the reproducing kernel Hilbert space middle, represents the reproducing kernel Hilbert space The squared norm on ;

[0135] The domain adaptation loss function measures the distance between the mean vectors of the source domain and target domain data in the reproducing kernel Hilbert space. The smaller the distance, the closer the data distribution of the two domains is.

[0136] The calculation formula of the prediction loss function is: ,in, is the predicted value of the target domain quality prediction model.

[0137] Step S440: Taking minimizing the loss function as the training objective, find the optimal quality prediction model parameters ,When the loss function reaches convergence, the training is completed and the trained target domain quality prediction model is obtained;

[0138] Step S450: Design a sliding window W and preset the window length , stores the most recent window length When the melt state parameters and drawing process parameters in production change, the target domain data in the sliding window is used to update the target domain quality prediction model parameters, and the updated target domain quality prediction model parameters are used to predict the fiber quality indicators of the next batch;

[0139] The update formula of the updated target domain quality prediction model parameters is: ,in, Respectively represent the target domain quality prediction model parameters updated at the t'th and t'+1th times, is the learning rate, is the loss function of the target domain data in the sliding window, is the gradient operator, Indicates the partial derivative of the target domain quality prediction model parameters, Represents the calculation loss function Gradient at the parameters of the quality prediction model of the current target domain;

[0140] The learning rate is set by those skilled in the art based on experience.

[0141] The calculation formula of the loss function of the target domain data in the sliding window is: ,in, is the balance coefficient of the domain adaptation loss function;

[0142] described The value of is set by those skilled in the art based on experience;

[0143] Figure 2 Flowchart of the method for obtaining the target domain quality prediction model provided in this application.

[0144] Step S500: obtaining candidate drawing process parameters according to the key cause set, taking the candidate drawing process parameters and melt state parameters as input, predicting the fiber quality index prediction value based on the target domain quality prediction model, calculating the quality error as feedback, and designing an optimization objective function to search for the optimal drawing process parameters;

[0145] The specific method for obtaining candidate wire drawing process parameters according to the key cause set, taking the candidate wire drawing process parameters and melt state parameters as input, predicting the fiber quality index prediction value based on the target domain quality prediction model, calculating the quality error as feedback, and designing an optimization objective function to search for the optimal wire drawing process parameters is as follows:

[0146] Step S510: obtaining melt state parameters of the current wire drawing process, taking the wire drawing process parameters in the key cause concentration as candidate wire drawing process parameters, and initializing the values ​​of the candidate wire drawing process parameters;

[0147] Step S520: using the current melt state parameters and the candidate drawing process parameters as input, and using the trained target domain quality prediction model to predict the fiber quality index prediction value;

[0148] Step S530: Fiber quality index prediction value and fiber quality index target value Compare and calculate the mass error ;

[0149] The calculation formula of the mass error is: ;

[0150] Step S540: Using the quality error as feedback, design an optimization objective function to search for the optimal drawing process parameters ;

[0151] The calculation formula of the optimization objective function is: ,in, is the current candidate drawing process parameter, is the last candidate drawing process parameter, is the regularization coefficient, is the module length, represents the candidate wire drawing process parameters that minimize the optimization objective function;

[0152] The regularization coefficient is used to balance the smoothness between the target value of the fiber quality index and the change of the drawing process parameters, and is set by those skilled in the art based on experience.

[0153] The constraints of the optimization objective function include physical and chemical constraints, process experience constraints and equipment capacity constraints. According to the physical and chemical constraints, the drawing speed is set not to exceed the maximum rheological rate of the melt material. According to the equipment capacity constraints, the drawing speed, traction force and cooling conditions are set not to exceed the maximum production capacity of the equipment. According to the process experience constraints, the drawing speed, traction force and cooling conditions are set not to exceed the empirical value range.

[0154] Step S550: using a genetic algorithm to solve the optimization objective function and obtain the optimal drawing process parameters, taking the current melt state parameters and the optimal drawing process parameters as input data, and using the target domain quality prediction model to output the optimized fiber quality index prediction value;

[0155] Step S560: Calculate the quality error between the optimized fiber quality index prediction value and the fiber quality index target value. When the quality error is less than the error threshold, apply the current optimal drawing process parameters to production control. Otherwise, continue to optimize until the quality error is less than the error threshold.

[0156] The value of the error threshold is set by those skilled in the art based on experience.

[0157] Figure 3 Flowchart of the optimal wire drawing process parameter optimization method provided in this application.

[0158] The above steps use a key cause set to identify candidate drawing process parameters with the greatest impact on fiber quality indicators, which are then used as key optimization parameters. This approach, based on transfer learning, adaptively optimizes the target domain quality prediction model and quality-oriented self-optimizing drawing process parameter control, forming a closed-loop quality optimization and control method. This method adaptively tracks changes in the production process and achieves target fiber quality indicators by optimizing drawing process parameters.

[0159] Example 2

[0160] like Figure 4 As shown, the inorganic mineral fiber production system based on intelligent control provided by this application includes:

[0161] The melt surface modeling module is used to obtain image sequences of the melt surface at different viewing angles, perform feature extraction and semantic segmentation on the image sequences, obtain image features and semantic masks, and construct a three-dimensional morphology model of the melt surface;

[0162] A state parameter acquisition module is used to obtain the temperature distribution map, flow velocity distribution map and conductivity distribution map of the melt surface, fuse the temperature distribution map, flow velocity distribution map and conductivity distribution map into a multi-physics field distribution map, couple the three-dimensional morphology model and the multi-physics field distribution map, and extract the melt state parameters;

[0163] The key cause search module is used to collect drawing process parameters and fiber quality indicators, analyze the causal relationship and causal strength between melt state parameters, drawing process parameters and fiber quality indicators, construct a weighted Bayesian network, and search for the key cause set of each fiber quality indicator;

[0164] The fiber quality prediction module is used to train the source domain quality prediction model using historical melt state parameters, drawing process parameters, and fiber quality indicators, and migrate the source domain quality prediction model to the target domain to obtain the target domain quality prediction model;

[0165] The optimal process optimization module is used to obtain candidate drawing process parameters based on the key cause set, take the candidate drawing process parameters and melt state parameters as input, predict the fiber quality index prediction value based on the target domain quality prediction model, calculate the quality error as feedback, and design the optimization objective function to search for the optimal drawing process parameters.

[0166] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0167] The above-described specific embodiments further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. An inorganic mineral fiber production method based on intelligent control, characterized in that: include: Obtain image sequences of the melt surface at different viewing angles, perform feature extraction and semantic segmentation on the image sequences, obtain image features and semantic masks, and construct a three-dimensional morphology model of the melt surface; Obtaining a temperature distribution map, a velocity distribution map, and a conductivity distribution map on the melt surface, fusing the temperature distribution map, the velocity distribution map, and the conductivity distribution map into a multi-physics field distribution map, coupling the three-dimensional morphology model and the multi-physics field distribution map, and extracting melt state parameters; Collect wire drawing process parameters and fiber quality indicators, analyze the causal relationship and causal strength between melt state parameters, wire drawing process parameters and fiber quality indicators, construct a weighted Bayesian network, and search for the key cause set of each fiber quality indicator; The source domain quality prediction model is trained using the melt state parameters, wire drawing process parameters, and fiber quality indicators of historical time, and the source domain quality prediction model is migrated to the target domain to obtain the target domain quality prediction model; The candidate wire drawing process parameters are obtained according to the key cause set. The candidate wire drawing process parameters and melt state parameters are taken as input. The predicted values ​​of fiber quality indicators are predicted based on the target domain quality prediction model. The quality error is calculated as feedback, and the optimization objective function is designed to search for the optimal wire drawing process parameters.

2. The method for producing inorganic mineral fibers based on intelligent control according to claim 1, characterized in that: The specific method of obtaining an image sequence of the melt surface at different viewing angles, performing feature extraction and semantic segmentation on the image sequence, obtaining image features and semantic masks, and constructing a three-dimensional morphology model of the melt surface is: Setting the melt surface imaging angles, and obtain a sequence of continuous frames at each imaging angle ;in, Indicates the imaging angle number, , t represents the time step, , represents the total number of time steps; Perform feature extraction on the image sequence at each imaging perspective to obtain the image features of each image at each imaging perspective and each time step , perform semantic segmentation on the images at each imaging perspective and each time step in the image sequence to obtain the semantic mask of each image ; For time step t, the image features and semantic masks are mapped to three-dimensional space to obtain three-dimensional feature point cloud and three-dimensional semantic mask point cloud, and a voxel grid V is constructed in the three-dimensional space. For each voxel , calculate voxel The eigenvalues ​​between the 3D feature point cloud at each imaging perspective , calculate voxel The semantic value between the 3D semantic mask point cloud at each imaging perspective ; Set the threshold for semantic value binarization , binarize the semantic value of the voxel to obtain the 3D semantic mask voxel ; The center coordinates of the voxels with a value of 1 in the 3D semantic mask voxels are extracted to form the 3D point cloud coordinates at time step t ; The three-dimensional point cloud coordinates include M three-dimensional points; For continuous The three-dimensional point cloud coordinates of each time step are spliced ​​to obtain a complete three-dimensional morphology model of the melt surface.

3. The method for producing inorganic mineral fibers based on intelligent control according to claim 2, characterized in that: The specific method of obtaining the temperature distribution map, flow velocity distribution map and conductivity distribution map of the melt surface, fusing the temperature distribution map, flow velocity distribution map and conductivity distribution map into a multi-physical field distribution map, coupling the three-dimensional morphology model and the multi-physical field distribution map, and extracting the melt state parameters is as follows: Obtaining the temperature distribution of the melt surface , velocity distribution diagram And conductivity distribution map , the temperature distribution map, velocity distribution map and conductivity distribution map are weighted averaged to obtain the multi-physics field distribution map ; The 3D topography model and the multi-physics field distribution map are coupled. For each time step t and each 3D point , according to the three-dimensional point At the corresponding position in the multi-physics field distribution map, extract the multi-physics field value at that position , combining the 3D points with the multi-physics values ​​to obtain the coupled characteristics of the melt state , and obtain the final multi-physics field coupling characterization model ; For each time step, the coupling characteristics of the melt state are extracted from the multi-physics field coupling characterization model as the melt state parameters of the current time step.

4. The method for producing inorganic mineral fibers based on intelligent control according to claim 3, characterized in that: The specific method of collecting the drawing process parameters and fiber quality indicators, analyzing the causal relationship and causal strength among the melt state parameters, the drawing process parameters and the fiber quality indicators, constructing a weighted Bayesian network, and searching for the key cause set of each fiber quality indicator is as follows: Collect the melt state parameters S and drawing process parameters P in the historical drawing process, conduct quality inspection on the produced inorganic mineral fibers, and obtain the fiber quality index Q; The melt state parameters, wire drawing process parameters and fiber quality indicators obtained are divided into Nd groups of sample data, and the structure G and parameter θ of the Bayesian network are learned based on the sample data; In Bayesian Networks In the equation, node v represents each parameter of melt state parameter, wire drawing process parameter and fiber quality index, and edge u represents the causal relationship between each parameter; for parameter Xi and parameter Xj, if there is a directed edge between vi and vj, it means that Xi is the direct cause of Xj; Calculate the mutual information between each parameter Xi and parameter Xj , average causal effect and causal certainty ; Define causal strength, for each pair of directed edges in a directed acyclic graph , the causal strength between parameter Xi and parameter Xj is calculated based on mutual information, average causal effect and causal confidence , assign the calculated causal strength to the corresponding edge in the Bayesian network to obtain a weighted Bayesian network; For each fiber quality index , find out its key cause set in the weighted Bayesian network ,in, For the A key cause node.

5. The method for producing inorganic mineral fibers based on intelligent control according to claim 4, characterized in that: The method for obtaining the key cause set is: For each fiber quality index , find out the set of all direct cause nodes in the weighted Bayesian network ; For each direct cause node in the direct cause node set, calculate the direct cause node and the fiber quality index The causal strength between nodes; Sort the causal strength from large to small and select the first Direct cause nodes are used as key cause nodes to form a key cause set .

6. The method for producing inorganic mineral fibers based on intelligent control according to claim 5, characterized in that: The specific method of using the melt state parameters, drawing process parameters, and fiber quality indicators of the historical time to train the source domain quality prediction model and migrating the source domain quality prediction model to the target domain to obtain the target domain quality prediction model is: Get historical samples based on melt state parameters, drawing process parameters and fiber quality indicators at historical time , train the source domain quality prediction model and obtain the source domain quality prediction model parameters ,in, Represents the characteristic vector composed of the melt state parameters and wire drawing process parameters of the nth historical sample, is the corresponding fiber quality index, is the number of historical samples; For a new batch of water-quenched slag raw materials, collect target domain data ,in, Indicates The characteristic vector composed of melt state parameters and wire drawing process parameters of target domain data, is the corresponding fiber quality index, is the number of samples of target domain data; The migration algorithm is used to migrate the source domain quality prediction model to the target domain, using the source domain quality prediction model parameters as the initial values ​​and the target domain data Optimize the source domain quality prediction model to obtain the target domain quality prediction model, design the domain adaptation loss function and the prediction loss function, and construct the loss function by weighted summation; Taking minimizing the loss function as the training objective, we can find the optimal quality prediction model parameters. , when the loss function reaches convergence, the training is completed and the trained target domain quality prediction model is obtained; Design sliding window W and preset window length , stores the most recent window length When the melt state parameters and drawing process parameters in production change, the target domain data in the sliding window is used to update the target domain quality prediction model parameters, and the updated target domain quality prediction model parameters are used to predict the fiber quality indicators of the next batch.

7. The method for producing inorganic mineral fibers based on intelligent control according to claim 6, characterized in that: The specific method of obtaining candidate wire drawing process parameters according to the key cause set, taking the candidate wire drawing process parameters and melt state parameters as input, predicting the predicted value of fiber quality index based on the target domain quality prediction model, calculating the quality error as feedback, and designing the optimization objective function to search for the optimal wire drawing process parameters is as follows: Obtain the melt state parameters of the current wire drawing, take the wire drawing process parameters in the key cause concentration as candidate wire drawing process parameters, and initialize the values ​​of the candidate wire drawing process parameters; Taking the current melt state parameters and candidate drawing process parameters as input, the trained target domain quality prediction model is used to predict the fiber quality index prediction value; The fiber quality index prediction value Fiber quality index target value Compare and calculate the mass error ; Using quality error as feedback, design optimization objective function to search for optimal wire drawing process parameters ; Genetic algorithm is used to solve the optimization objective function and obtain the optimal drawing process parameters. The current melt state parameters and the optimal drawing process parameters are used as input data, and the target domain quality prediction model is used to output the predicted value of the optimized fiber quality index. The quality error between the predicted value of the optimized fiber quality index and the target value of the fiber quality index is calculated. When the quality error is less than the error threshold, the current optimal drawing process parameters are applied to production control, otherwise the optimization is continued until the quality error is less than the error threshold.

8. An inorganic mineral fiber production system based on intelligent control, which is used to implement the inorganic mineral fiber production method based on intelligent control according to any one of claims 1 to 7, characterized in that: include: The melt surface modeling module is used to obtain image sequences of the melt surface at different viewing angles, perform feature extraction and semantic segmentation on the image sequences, obtain image features and semantic masks, and construct a three-dimensional morphology model of the melt surface; A state parameter acquisition module is used to obtain a temperature distribution map, a flow velocity distribution map and a conductivity distribution map on the melt surface, fuse the temperature distribution map, the flow velocity distribution map and the conductivity distribution map into a multi-physical field distribution map, couple the three-dimensional morphology model and the multi-physical field distribution map, and extract the melt state parameters; The key cause search module is used to collect wire drawing process parameters and fiber quality indicators, analyze the causal relationship and causal strength between melt state parameters, wire drawing process parameters and fiber quality indicators, construct a weighted Bayesian network, and search for the key cause set of each fiber quality indicator; The fiber quality prediction module is used to train the source domain quality prediction model by using the melt state parameters, drawing process parameters, and fiber quality indicators of historical time, and migrate the source domain quality prediction model to the target domain to obtain the target domain quality prediction model; The optimal process optimization module is used to obtain candidate wire drawing process parameters according to the key cause set, take the candidate wire drawing process parameters and melt state parameters as input, predict the fiber quality index prediction value based on the target domain quality prediction model, calculate the quality error as feedback, and design the optimization objective function to search for the optimal wire drawing process parameters.

9. An electronic device, characterized in that: It comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for producing inorganic mineral fibers based on intelligent control as described in any one of claims 1 to 7 are implemented.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the steps in the inorganic mineral fiber production method based on intelligent control as described in any one of claims 1 to 7.

Images