A method, system and apparatus for producing architectural glass

Abstract

The application relates to the technical field of glass production, and provides a building glass production method, a system and equipment, which comprise the following steps: obtaining historical production information data sets of multiple different batches of building glass of the same type; constructing a multi-target dynamic loss function, and constructing a neural network model based on the loss function; training the neural network model according to the historical production information data sets; obtaining basic information of building glass to be produced, inputting the basic information into the neural network model to obtain optimal process control parameters, and producing the building glass according to the optimal process control parameters. The application improves the robustness of the neural network model, thereby obtaining optimal process control parameters matched with current raw material information and target performance information of the building glass to be produced, and improves the intelligent degree and production benefit.

Description

Technical Field

[0001] This invention relates to the field of glass production technology, and more specifically, to a method, system, and equipment for producing architectural glass. Background Technology

[0002] Architectural glass refers to transparent or translucent glass materials used in building components. It has properties such as light transmission, heat insulation, and sound insulation. It can be divided into many types, including ordinary flat glass, tempered glass, laminated glass, LOW-E glass, patterned glass, and frosted glass.

[0003] A search revealed several typical existing technologies. For example, application number CN2021110133635 discloses an intelligent production process and system for architectural glass, including processes such as raw material cutting, edging, cleaning, hot dipping, and composite glass production. Another example is application number CN2023116763280, which discloses a production control system and method for tempered glass. Based on deep neural network model artificial intelligence technology, it analyzes and classifies the preheating process of tempered glass to determine whether to stop preheating, thereby improving the efficiency and quality of the tempered glass production process. Yet another example is application number CN2023102954723, which discloses an intelligent production control method for tempered glass. This method characterizes the crystal structure of nickel sulfide grains in tempered glass using photoluminescence principles. Based on changes in its crystal phase, it determines the timing of spontaneous breakage of the tempered glass, allowing it to be removed from the homogenizing furnace in advance to prevent contamination and damage to the furnace.

[0004] The three existing technologies mentioned above all involve glass construction production methods, but they do not take into account multiple factors that affect the production efficiency of architectural glass, making it difficult to obtain more ideal process control parameters. Their level of intelligence and production efficiency still have room for further improvement. Summary of the Invention

[0005] Based on this, in order to solve the problems existing in the prior art, the present invention provides a method, system and equipment for producing architectural glass. It constructs a neural network model based on multi-scale data information including defect information, energy consumption information and environmental factor information, and trains the neural network model with historical production information datasets of multiple different batches of architectural glass of the same type. By taking into account multiple factors affecting the efficiency of architectural glass production, the robustness of the neural network model can be improved, thereby obtaining the optimal process control parameters that match the current raw material information and target performance information of the architectural glass to be produced, improving the level of intelligence and production efficiency. The specific technical solution is as follows:

[0006] A method for producing architectural glass includes the following steps:

[0007] Obtain the type of architectural glass to be produced and obtain a dataset of historical production information for multiple different batches of the same type of architectural glass.

[0008] A multi-objective dynamic loss function is constructed based on defect information, energy consumption information, performance information, and environmental factor information, and a neural network model is built based on the loss function.

[0009] The neural network model is trained based on the historical production information dataset.

[0010] The basic information of the architectural glass to be produced is obtained, the basic information is input into the neural network model to obtain the optimal process control parameters, and the architectural glass is produced according to the optimal process control parameters.

[0011] The historical production information dataset includes historical raw material information, historical process control parameters, historical defect information, historical performance information, historical energy consumption information, and historical environmental factor information. The basic information includes current raw material information, target performance information, and environmental factor information of the production process.

[0012] The described architectural glass production method constructs a neural network model and trains the model with historical production information datasets of multiple different batches of architectural glass of the same type. By taking into account multiple factors affecting the efficiency of architectural glass production, the robustness of the neural network model can be improved. This allows the acquisition of optimal process control parameters that match the current raw material information and target performance information of the architectural glass to be produced, thereby improving the level of intelligence and production efficiency.

[0013] Preferably, the specific method for training the neural network model based on the historical production information dataset includes the following steps:

[0014] Obtain the correlation between multiple different batches of architectural glass of the same type and the architectural glass to be produced;

[0015] The historical production information dataset is filtered based on a preset correlation threshold and the correlation degree.

[0016] The neural network model is trained based on the selected historical production information dataset.

[0017] Preferably, the multi-objective dynamic loss function L = L1 + L2 + L3 + β × ||W defect W energy || 2 ;

[0018] Where L1=α defect ||w defect || 2 +α energy ||wenergy || 2 Let α represent the multi-objective regularization term. defect α energy w represents the defect loss weighting coefficient and the energy consumption loss weighting coefficient, respectively. defect w energy Let f(env) and f(env) represent the feature vector of defect information and the feature vector of energy consumption loss, respectively. Let |||| denote the norm, and L2 = Relu(f(env) - threshold) represent the environmental sensitivity penalty term. Represents the comprehensive environmental sensitivity evaluation function, Relu() represents the activation function, threshold represents the safety threshold for environmental factors, M1 represents the number of types of environmental factors, and λ represents the overall environmental sensitivity evaluation function. i This represents the weight of the evaluation value of the i-th type of environmental factor, and e represents the natural constant. E represents an intermediate variable. i E represents the actual value of the i-th environmental factor. i ' represents the standard value of the i-th environmental factor, and || represents the absolute value function. α represents the performance loss term. performance This represents the performance loss weighting coefficient, M2 represents the number of types of performance information, and p i p represents the actual value of the performance information for the i-th type. i 'Represents the predicted value β×||W for the i-th type of performance information. defect W energy || 2 W represents the weighted regularization term, β represents the regularization strength, and W represents the weighted regularization term. defect W represents the defect loss weight matrix. energy This represents the energy loss weight matrix.

[0019] Preferably, the correlation degree R = R' + R” + 1 / R”', where R', R”, and R”' represent the correlation coefficient of raw materials, the correlation coefficient of performance parameters, and the correlation coefficient of defects, respectively.

[0020] R' is obtained by calculating the weighted value of the Pearson correlation coefficient between the characteristic information values ​​of current raw materials and historical raw materials of multiple types;

[0021] R” is obtained by calculating the weighted value of the Pearson correlation coefficient between multiple types of target performance parameters and historical performance parameter feature information values;

[0022] R”' is obtained by calculating the weighted average error between the historical defect feature values ​​and the corresponding defect feature standard values ​​of multiple types of historical defect information.

[0023] Preferably, the specific method for filtering the historical production information dataset includes: removing and filtering historical production information of batches of architectural glass with a correlation degree less than a preset correlation degree threshold;

[0024] The specific method for training the neural network model based on the filtered historical production information dataset includes: preprocessing the historical production information of multiple batches of architectural glass after filtering to generate a historical production information dataset, and training the neural network model based on the historical production information dataset to obtain the trained neural network model.

[0025] An architectural glass production system for implementing the architectural glass production method, comprising:

[0026] The production information acquisition module is used to acquire the type of architectural glass to be produced and to acquire historical production information datasets of multiple different batches of architectural glass of the same type.

[0027] The neural network construction module is used to construct a multi-objective dynamic loss function based on defect information, energy consumption information, performance information and environmental factor information, and to construct a neural network model based on the loss function;

[0028] The training module is used to train the neural network model based on the historical production information dataset;

[0029] The process parameter acquisition module is used to acquire basic information about the architectural glass to be produced, input the basic information into the neural network model to obtain optimal process control parameters, and produce the architectural glass according to the optimal process control parameters.

[0030] The historical production information dataset includes historical raw material information, historical process control parameters, historical defect information, historical performance information, historical energy consumption information, and historical environmental factor information. The basic information includes current raw material information, target performance information, and environmental factor information of the production process.

[0031] Preferably, the training module includes:

[0032] The correlation degree acquisition unit is used to acquire the correlation degree between multiple different batches of architectural glass of the same type and the architectural glass to be produced.

[0033] A filtering unit is used to filter historical production information datasets based on a preset correlation threshold and the correlation degree.

[0034] The training unit is used to train the neural network model based on the filtered historical production information dataset.

[0035] Preferably, the neural network building module includes:

[0036] The loss function construction unit is used to construct a multi-objective dynamic loss function L=L1+L2+β×||W based on defect information, energy consumption information, and environmental factor information. defect W energy || 2 ;

[0037] A neural network model building unit is used to build a neural network model based on the multi-objective dynamic loss function.

[0038] Where L1=α defect ||w defect || 2 +α energy ||w energy || 2 Let α represent the multi-objective regularization term. defect α energy w represents the defect loss weighting coefficient and the energy consumption loss weighting coefficient, respectively. defect w energy Let f(env) and f(env) represent the feature vector of defect information and the feature vector of energy consumption loss, respectively. Let |||| denote the norm, and L2 = Relu(f(env) - threshold) represent the environmental sensitivity penalty term. Represents the comprehensive environmental sensitivity evaluation function, Relu() represents the activation function, threshold represents the safety threshold for environmental factors, M1 represents the number of types of environmental factors, and λ represents the overall environmental sensitivity evaluation function. i This represents the weight of the evaluation value of the i-th type of environmental factor, and e represents the natural constant. E represents an intermediate variable. i E represents the actual value of the i-th environmental factor. i ' represents the standard value of the i-th environmental factor, and || represents the absolute value function. α represents the performance loss term. performance This represents the performance loss weighting coefficient, M2 represents the number of types of performance information, and p i p represents the actual value of the performance information for the i-th type. i 'Represents the predicted value β×||W for the i-th type of performance information. defect W energy || 2 W represents the weighted regularization term, β represents the regularization strength, and W represents the weighted regularization term. defect W represents the defect loss weight matrix. energy This represents the energy loss weight matrix.

[0039] Preferably, the correlation degree acquisition unit includes:

[0040] The first subunit is used to obtain the raw material correlation coefficient R' by calculating the weighted value of the Pearson correlation coefficient between the characteristic information values ​​of multiple types of current raw materials and historical raw materials;

[0041] The second subunit is used to obtain the performance parameter correlation coefficient R” by calculating the weighted value of the Pearson correlation coefficient between multiple types of target performance parameters and historical performance parameter feature information values;

[0042] The third subunit is used to obtain the defect correlation coefficient R”' by calculating the weighted average error between the historical defect feature values ​​and the corresponding defect feature standard values ​​of multiple types of historical defect information;

[0043] The correlation degree R = R' + R” + 1 / R”'.

[0044] An architectural glass production equipment, comprising:

[0045] Controller;

[0046] Memory, which stores executable instructions;

[0047] The executable instructions can be run on the controller to implement the architectural glass production method. Attached Figure Description

[0048] The invention will be further understood from the following description taken in conjunction with the accompanying drawings. The components in the drawings are not necessarily drawn to scale, but rather the emphasis is on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0049] Figure 1 This is a schematic diagram of the overall process of a method for producing architectural glass according to an embodiment of the present invention;

[0050] Figure 2 This is a flowchart illustrating a specific method for training a neural network model in one embodiment of the present invention;

[0051] Figure 3 This is a flowchart illustrating the specific method for calculating the correlation coefficient of raw materials, the correlation coefficient of performance parameters, and the correlation coefficient of defects in one embodiment of the present invention.

[0052] Figure 4 This is a schematic diagram of the overall structure of an architectural glass production system according to one embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to its embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0054] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0056] In this invention, "first" and "second" do not represent a specific quantity or order, but are merely used to distinguish names.

[0057] Before describing the embodiments of the present invention, a brief introduction to the prior art will be given.

[0058] Based on application scenarios and functions, architectural glass can be classified into safety and protective glass (such as tempered glass, laminated glass, and explosion-proof glass), energy-saving and environmentally friendly glass (such as insulated glass, low-emissivity glass, and heat-absorbing glass), decorative art glass (such as painted / patterned glass and embossed glass), and functional glass (such as fire-resistant glass, soundproof glass, and self-cleaning glass). Architectural glass production generally includes raw material preparation, cleaning and drying, melting and forming, heat treatment and strengthening (such as degradation and tempering), deep processing, and quality inspection and packaging. The melting and forming and heat treatment strengthening steps involve numerous process control parameters. These parameters often need to be adjusted based on raw materials, preset target performance parameters (such as light transmittance, impact resistance, flatness, hardness, and toughness), and quality requirements for surface and internal defects to meet production needs and improve production efficiency.

[0059] The glass production methods disclosed in applications CN2021110133635, CN2023116763280, and CN2023102954723 do not take into account multiple factors affecting the production efficiency of architectural glass. That is, they do not comprehensively optimize process control parameters based on basic information such as raw material information, preset target performance information, and quality requirements for internal surface defects during the glass production process. Their level of intelligence and production efficiency have room for further optimization.

[0060] To optimize the intelligence and efficiency of glass production, this invention provides a method for producing architectural glass. This method constructs a neural network model using multi-scale data, including defect information, energy consumption information, and environmental factor information. It trains the neural network model by acquiring historical production information datasets of multiple batches of architectural glass of the same type. By considering multiple factors affecting the efficiency of architectural glass production, the robustness of the neural network model is improved. This allows for the acquisition of optimal process control parameters, including current raw material information and target performance information for the architectural glass to be produced, thereby enhancing the intelligence and efficiency of production.

[0061] like Figure 1 As shown, the architectural glass production method includes the following steps:

[0062] S1, obtain the type of architectural glass to be produced and obtain a dataset of historical production information for multiple different batches of architectural glass of the same type.

[0063] The historical production information dataset includes historical raw material information, historical process control parameters, historical defect information, historical performance information, historical energy consumption information, and historical environmental factor information. Preferably, a virtual digital model can be constructed based on digital twin technology, and simulations can be performed based on this digital model to obtain a richer and more meaningful historical production information dataset. Of course, the historical production information dataset of multiple different batches of architectural glass of the same type can be a dataset of multiple different batches of architectural glass of the same type from the production service database of the same manufacturer, or it can be a dataset of multiple different batches of architectural glass of the same type from the production server databases of multiple manufacturers.

[0064] The raw material information includes, but is not limited to, parameter information (including type and proportion) of materials used in the production of architectural glass, such as quartz sand, limestone, feldspar, and soda ash; the process control parameters include, but are not limited to, parameter information including melting temperature and time, and temperature and time at each stage of annealing; the defect information includes, but is not limited to, parameter information of surface and internal defects of architectural glass (such as scratches, bubbles, dents, distortion, cracks, and defects); the performance information includes, but is not limited to, parameter information including light transmittance, impact resistance, pressure resistance, hardness, and toughness; the energy consumption information includes, but is not limited to, parameter information including gas consumption, cooling water consumption, and electrical energy; and the environmental factor information includes, but is not limited to, parameter information including ambient temperature, humidity, and vibration values.

[0065] One of the purposes of obtaining the type of architectural glass to be produced and the historical production information dataset of multiple different batches of architectural glass of the same type is to improve the reference value of the historical production information dataset, so as to better train the neural network model.

[0066] S2. Construct a multi-objective dynamic loss function based on defect information, energy consumption information, performance information and environmental factor information, and build a neural network model based on the loss function.

[0067] The neural network models include, but are not limited to, CNN convolutional neural network models, BP neural network models, and fusion neural network models of Transformer and CNN.

[0068] For fusion neural network models that combine Transformer and CNN, aiming to combine the advantages of both to improve model performance, the fusion methods include:

[0069] 1. Early Layer Fusion: In the early stages of the network, feature maps from the CNN are fused with certain layers of the Transformer. This typically involves converting the output of an intermediate layer of the CNN into a sequence format that the Transformer can process, and then feeding it into the Transformer.

[0070] 2. Lateral Layer Fusion: This involves fusing data across multiple layers of the network. The outputs of the CNN and Transformer at each layer can be combined to create a richer representation that incorporates both local and global information.

[0071] 3. Sequential Fusion: First, a CNN is used to process the input data to extract local features, and then these features are passed as input to a Transformer. This method allows the Transformer to model long-range dependencies based on the already extracted local features.

[0072] 4. Parallel Fusion: CNN and Transformer run in parallel, processing the input data separately. Then, the outputs of the two are combined in some form (such as concatenation, addition, or multiplication) to create a representation that contains both local and global information.

[0073] 5. Encoder-Decoder Architecture: In this architecture, a CNN acts as the encoder to extract local features from the input data, while a Transformer acts as the decoder to utilize these features and generate the output. The decoder can leverage the Transformer's self-attention mechanism to model long-range dependencies and produce task-relevant output.

[0074] 6. Cross-Teaching Approach: In semi-supervised learning scenarios, labeled and unlabeled images are used as input. For labeled data, the CNN and Transformer are supervised by the real labels respectively. For unlabeled data, the parameters of the Transformer / CNN are updated using predictions generated by the CNN and Transformer respectively. This approach can encourage consistency between different networks and improve model performance.

[0075] Of course, the neural network model can be a single type of neural network for deep learning, or it can be a fusion neural network that combines two different types of neural networks, such as a fusion neural network model of Transformer and CNN, to improve accuracy and robustness and / or shorten training time and improve training effect.

[0076] By combining two neural networks and utilizing information from both forward and backward propagation, the model can make fuller use of the contextual information in the input data, thereby improving its accuracy and robustness. For example, a two-layer bidirectional neural network structure can consider both preceding and following contextual information simultaneously, enhancing the model's robustness and effectiveness. Alternatively, by combining two neural networks, contextual information can be used in both forward and backward propagation, reducing the problems of vanishing and exploding gradients, thus shortening training time and improving training effectiveness. This structure can also improve the model's generalization ability.

[0077] When combining two neural networks, different feature fusion methods can be used, including:

[0078] 1. Simple addition or concatenation: Simply add or concatenate the features extracted by different neural networks. For example, for feature vectors extracted by two neural networks, they can be simply added together or concatenated along the feature dimension to form a longer feature vector.

[0079] 2. Weighted Addition or Concatenation: Weighted addition or concatenation builds upon simple addition or concatenation by weighting the features extracted by different neural networks to better fuse them. For example, the features extracted by each neural network can be weighted based on their performance. This method can reduce feature dimensionality and improve the model's generalization ability, but it requires accurate evaluation of the performance of different neural networks to determine the weighting coefficients.

[0080] 3. Feature Crossing: Feature crossing involves combining features extracted by different neural networks to generate new feature representations. For example, features extracted by two neural networks can be multiplied element-wise to generate a new feature vector. This method can increase feature diversity and improve the model's expressive power, but it may increase feature dimensionality and model complexity.

[0081] 4. Feature Selection: Feature selection involves choosing the most representative features from those extracted by different neural networks to reduce feature dimensionality and improve the model's generalization ability. For example, correlation analysis and PCA can be used to filter features and select the most representative ones. This method can reduce feature dimensionality and improve the model's generalization ability, but it may lose some important feature information.

[0082] S3, Train the neural network model based on the historical production information dataset.

[0083] Preferably, such as Figure 2 As shown, in step S3, the specific method for training the neural network model based on the historical production information dataset includes the following steps:

[0084] S31, obtain the correlation between multiple different batches of architectural glass of the same type and the architectural glass to be produced.

[0085] Specifically, the correlation between each batch of architectural glass and the architectural glass to be produced can be calculated using the formula R = R' + R” + 1 / R”', where R', R”, and R”' represent the correlation coefficient of raw materials, the correlation coefficient of performance parameters, and the correlation coefficient of defects, respectively.

[0086] like Figure 3 As shown, the specific methods for calculating the correlation coefficients of raw materials, performance parameters, and defects include the following steps:

[0087] S311, the raw material correlation coefficient is obtained by calculating the weighted value of the Pearson correlation coefficient between the characteristic information values ​​of multiple types of current raw materials and historical raw materials;

[0088] S312, the correlation coefficient of the performance parameter is obtained by calculating the weighted value of the Pearson correlation coefficient between multiple types of target performance parameters and historical performance parameter feature information values;

[0089] S313, the defect correlation coefficient is obtained by calculating the weighted average error between the historical defect feature values ​​and the corresponding defect feature standard values ​​of multiple types of historical defect information.

[0090] More specifically, raw material correlation coefficient

[0091] in, RMCx represents the correlation coefficient value of the i-th type of raw material, N1 represents the number of types of raw materials, and n1 represents the number of dimensions of the feature information of the i-th type of raw material, where n1 > 1. jRMCy represents the j-th feature information value of the i-th type of current raw material under a certain dimension. j λ represents the j-th feature information value of the i-th type of historical raw materials under a certain dimension. 1i The adjustment weighting factor representing the correlation coefficient value of the i-th type of raw material can be set by technical personnel based on experience.

[0092] The number of dimensions can be understood as the number of types of characteristic information of a certain type of raw material; the characteristic information includes, but is not limited to, the chemical composition content of the raw material (such as the content of silicon dioxide, iron oxide and aluminum oxide in quartz sand), particle size distribution, purity and impurity content (such as the content of elements such as iron, aluminum, calcium, magnesium, potassium and sodium), and physical properties (such as specific gravity, moisture content, hardness, etc.).

[0093] Correlation coefficient of performance parameters

[0094] in, PPx represents the correlation coefficient value of the i-th type of performance parameter, M2 represents the number of types of performance parameters, n2 represents the number of feature information of the i-th type of performance parameter and n2>1, PPx j PPy represents the j-th feature information value under the i-th type of target performance parameter. j λ represents the j-th feature information value under the historical performance parameter of the i-th category. 2i The adjustment weighting factor representing the correlation coefficient value of the i-th type of performance parameter can be set by technical personnel based on experience.

[0095] The performance parameters refer to the performance information of the corresponding type of architectural glass, including but not limited to light transmittance, impact resistance, pressure resistance, hardness, and toughness. The number of characteristic information points for a certain type of historical performance parameter can be understood as the number of batches of architectural glass whose performance parameters were collected. For example, when collecting performance parameters of architectural glass, it is pre-set that performance parameters are collected from n² pieces of architectural glass in each batch, and the number of types of performance parameters collected from each piece of architectural glass is M². More specifically, assuming that the types of performance parameters include light transmittance, impact resistance, pressure resistance, hardness, and toughness, then M² = 5. If 100 pieces of architectural glass in each batch are collected for the detection of performance parameter characteristic information values, then n² = 100.

[0096] The target performance information is a preset value, which can be set by technicians according to production needs or product order requirements. For the n2 characteristic information values ​​under the i-th type of target performance parameter, all are uniformly set values, i.e., PPx1, PPx2, ..., PPx n2 All of them have the same preset value.

[0097] Defect correlation coefficient

[0098] in, D represents the mean square error value of the defect information of the i-th type, N3 represents the number of defect information types, n3 represents the number of feature information of the historical defect information of the i-th type and n3>1, and D j D represents the feature value of the j-th historical defect under the historical defect information of the i-th type. j 'Represents the standard value of the j-th defect characteristic under the historical defect information of the i-th type, which can be set by technical personnel, λ 3i The adjustment weighting factor representing the mean square error value of the historical defect information of the i-th type can be set by technicians based on experience.

[0099] The defect information includes, but is not limited to, defect parameter information of surface and internal defects of architectural glass (such as scratches, bubbles, dents, distortions, cracks, and defects). The standard value of the j-th defect feature under the i-th type of historical defect information is a preset value, which can be set by technical personnel according to production needs, product order requirements, or production experience.

[0100] The number of feature information items for a certain type of historical defect information can be understood as the number of batches of architectural glass for which defect information is collected. For example, when collecting defect information on architectural glass, it is pre-set that defect information is collected from n3 pieces of architectural glass in each batch, and the number of types of defect information collected from each piece of architectural glass is N3. More specifically, assuming that the types of defect information include scratches, bubbles, dents, distortions, cracks, and defects, then N3 = 6. If defect feature values ​​are collected from 50 pieces of architectural glass in each batch, then n3 = 50.

[0101] S32, the historical production information dataset is filtered according to the preset correlation threshold and the correlation.

[0102] Specific methods for filtering historical production information datasets include: removing and filtering historical production information of batches of architectural glass with a correlation degree less than a preset correlation degree threshold;

[0103] S33, Train the neural network model based on the filtered historical production information dataset.

[0104] The specific method for training the neural network model based on the filtered historical production information dataset includes: preprocessing the historical production information of multiple batches of architectural glass after filtering to generate a historical production information dataset, and training the neural network model based on the historical production information dataset to obtain the trained neural network model.

[0105] By observing the formula R = R' + R” + 1 / R”’, it can be seen that the correlation degree combines the correlation coefficient of raw materials, the correlation coefficient of performance parameters, and the correlation coefficient of defects. It considers the correlation coefficient between historical performance information and target performance information, the correlation coefficient between current raw material information and historical raw material information, and the mean square error between historical defect feature values ​​and preset defect feature standard values. By filtering the historical production information dataset according to the preset correlation degree threshold and the correlation degree itself, batches of architectural glass with significant differences between historical raw material information and current raw material information, historical performance information and target performance information, and those with large overall defects can be removed. This means that the historical production information of unqualified batches of architectural glass is filtered out to ensure the value of the historical production information dataset, improve its quality and reliability, and thus improve the effect and accuracy of training the neural network model using the historical production information dataset.

[0106] S4. Obtain basic information about the architectural glass to be produced, input the basic information into the neural network model to obtain optimal process control parameters, and produce the architectural glass according to the optimal process control parameters.

[0107] Here, the basic information includes current raw material information, target performance information, and environmental factor information of the production process. For the neural network model, process control parameters are used as output, and raw material information and environmental factor information are used as input. A loss function is constructed based on defect information, performance information, environmental factor information, and energy consumption information. Through iterative training, the loss function is minimized to obtain the optimal process control parameters.

[0108] In the current technology of architectural glass production, the loss function of obtaining optimal process parameters based on deep learning neural networks often only considers one of the defects, performance or energy loss in glass production. This method optimizes for a single production goal and cannot comprehensively improve the overall production efficiency of architectural glass production.

[0109] Compared to existing technologies, the architectural glass production method of this invention constructs a multi-objective dynamic loss function based on defect information, energy consumption information, performance information, and environmental factor information. This not only overcomes the problem that existing architectural glass production methods, which optimize for a single production objective, cannot comprehensively improve the overall production efficiency of architectural glass production, but also helps to improve the overall production efficiency of architectural glass according to actual production needs. Furthermore, by incorporating environmental factor information into the loss function, it can take into account the potential impact of environmental factors on the quality and efficiency of architectural glass, such as defects, energy consumption, and performance, thus ensuring the robustness and accuracy of the neural network model.

[0110] In summary, the architectural glass production method described above, by constructing a neural network model and training the model with historical production information datasets of multiple different batches of architectural glass of the same type, takes into account multiple factors affecting the efficiency of architectural glass production, thereby improving the robustness of the neural network model. This allows for the acquisition of optimal process control parameters that match the current raw material information and target performance information of the architectural glass to be produced, thus improving the level of intelligence and production efficiency.

[0111] In addition, by filtering the historical production information dataset according to the preset correlation threshold and the correlation, the historical production information of batches of architectural glass with large differences between historical raw material information and current raw material information, historical performance information and target performance information, and large overall defects can be removed. That is, the historical production information of batches of architectural glass that do not meet the requirements is removed and filtered to ensure the value of the historical production information dataset, improve its quality and reliability, and thus improve the effect and accuracy of training the neural network model through the historical production information dataset.

[0112] As a preferred technical solution, in step S2, the multi-objective dynamic loss function L=L1+L2+L3+β×||W defect W energy || 2 ;

[0113] Where L1=α defect ||w defect || 2 +α energy ||w energy || 2 Let α represent the multi-objective regularization term. defect α energy w represents the defect loss weighting coefficient and the energy consumption loss weighting coefficient, respectively. defect w energy Let f(env) and f(env) represent the feature vector of defect information and the feature vector of energy consumption loss, respectively. Let |||| denote the norm, and L2 = Relu(f(env) - threshold) represent the environmental sensitivity penalty term. Represents the comprehensive environmental sensitivity evaluation function, Relu() represents the activation function, threshold represents the safety threshold for environmental factors, M1 represents the number of types of environmental factors, and λ represents the overall environmental sensitivity evaluation function. i The weight of the evaluation value for the i-th type of environmental factor can be set by technical personnel, and e represents a natural constant. E represents an intermediate variable. i E represents the actual value of the i-th environmental factor. i ' represents the standard value of the i-th environmental factor, and || represents the absolute value function. α represents the performance loss term. performanceThis represents the performance loss weighting coefficient, M2 represents the number of types of performance information, and p i p represents the actual value of the performance information for the i-th type. i ' represents the predicted value of the performance information for the i-th type, β×||W defect W energy || 2 W represents the weighted regularization term, β represents the regularization strength, which can be set by technical personnel, with an initial value between 0.01 and 0.1. defect W represents the defect loss weight matrix. energy This represents the energy loss weight matrix.

[0114] Weight regularization term β×||W defect W energy || 2 It is mainly used to apply Frobenius norm constraints to the defect loss weight matrix and the energy consumption loss weight matrix to prevent overfitting of the neural network model during training.

[0115] The environmental sensitivity penalty term L2 = ReLU(f(env) - threshold) integrates environmental factor features from multiple dimensions. By setting an activation function, a non-linear threshold response mechanism is created. When the overall environmental conditions deteriorate and exceed the environmental factor safety threshold, a gradient penalty is triggered, increasing the constraint strength. One of the main functions of setting the environmental sensitivity penalty term is to construct a dynamically updated loss function based on changes in the actual values ​​of environmental factors and to increase the model's constraint strength, enabling the neural network model to adapt to the diversity of input sample data. Specifically,

[0116] For the energy loss weighting coefficient α energy Defect loss weighting coefficient α defect The performance loss weighting coefficient can be set by technical personnel based on their experience.

[0117] In data analysis and machine learning, feature vectors are typically vectors composed of multiple features. Each feature may have different units of measurement or numerical ranges, which can lead to uneven weight distribution among features during model training, affecting model performance and accuracy. Feature vector normalization eliminates the influence of units of measurement between features by mapping the numerical range of each feature to the same interval, making it easier for the model to capture the relationships and patterns between features. Therefore, according to the formula L=L1+L2+β×||W defect W energy || 2Before calculating the loss function, preferably, the defect information feature vector and the energy loss feature vector can be normalized separately by normalization processing to unify the numerical range between different features and avoid the problem of unstable model training or slow convergence speed due to excessive differences between features.

[0118] Similarly, when training a neural network model, the feature vectors of sample data in the historical production information dataset can be normalized to improve the model's training speed and stability, as well as its generalization ability and accuracy. Feature vector normalization effectively avoids the impact of differences in the dimensions of features on the model, making it easier for the model to capture patterns and regularities in the data.

[0119] When performing feature vector normalization, the most commonly used methods include min-max normalization and Z-score standardization. Min-max normalization maps the numerical range of each feature to [0,1], preserving the distribution information of the original data; while Z-score standardization maps the data to a normal distribution with a mean of 0 and a standard deviation of 1, which is suitable for data that conforms to a normal distribution.

[0120] When inputting the basic information into the neural network model to obtain the optimal process control parameters, the performance loss term in the loss function... Preferably, p i ' represents the predicted value of the performance information for the i-th category, p i This represents the actual value of the performance information of the i-th type, which can be replaced by the feature information value under the target performance parameter of the i-th type.

[0121] By constructing the multi-objective dynamic loss function and incorporating environmental factor information, the potential impact of environmental factors on the quality and efficiency of architectural glass, such as defects, energy consumption, and performance, during the production process can be taken into account, ensuring the robustness and accuracy of the neural network model. Simultaneously, it overcomes the problem that existing architectural glass production methods, which optimize for a single production objective, cannot comprehensively improve the overall production efficiency of architectural glass. This facilitates improving the overall production efficiency of architectural glass based on actual production needs. Based on the constructed multi-objective dynamic loss function and neural network model, by adjusting the weight coefficients for defect loss, energy consumption loss, and performance loss, the optimal process parameters that meet the actual needs of architectural glass manufacturers for comprehensive production efficiency can be obtained, thereby improving the overall intelligence of the system.

[0122] An embodiment of the present invention also provides an architectural glass production system for implementing the aforementioned architectural glass production method, such as... Figure 4As shown, it includes a production information acquisition module, a neural network construction module, a training module, and a process parameter acquisition module.

[0123] The production information acquisition module is used to acquire the type of architectural glass to be produced and to acquire historical production information datasets of multiple different batches of architectural glass of the same type; the neural network construction module is used to construct a multi-objective dynamic loss function based on defect information, energy consumption information, performance information and environmental factor information, and to construct a neural network model based on the loss function.

[0124] The training module is used to train the neural network model based on the historical production information dataset; the process parameter acquisition module is used to acquire the basic information of the architectural glass to be produced, input the basic information into the neural network model to obtain the optimal process control parameters, and produce the architectural glass according to the optimal process control parameters.

[0125] The historical production information dataset includes historical raw material information, historical process control parameters, historical defect information, historical performance information, historical energy consumption information, and historical environmental factor information. The basic information includes current raw material information, target performance information, and environmental factor information of the production process.

[0126] Specifically, the process control parameters can be parameters related to several stages of annealing during glass production, such as the temperature range, heating rate, and heating time for primary and secondary annealing in the heating stage; the temperature range and holding time in the holding stage; the cooling rate and temperature range in the slow cooling stage; and the cooling rate and cooling method in the rapid cooling stage. Correspondingly, the performance information includes, but is not limited to, residual stress parameters, structural relaxation, grain size, grain boundary clarity, elastic modulus, coefficient of thermal expansion, light transmittance, and refractive index.

[0127] Preferably, the training module includes a correlation acquisition unit, a filtering unit, and a training unit.

[0128] The correlation acquisition unit is used to acquire the correlation between multiple different batches of architectural glass of the same type and the architectural glass to be produced; the filtering unit is used to filter the historical production information dataset according to the preset correlation threshold and the correlation; the training unit is used to train the neural network model according to the filtered historical production information dataset.

[0129] Specifically, the correlation acquisition unit includes a first sub-unit, a second sub-unit, and a third sub-unit.

[0130] The first subunit is used to obtain the raw material correlation coefficient R' by calculating the weighted value of the Pearson correlation coefficient between the current raw material and historical raw material feature information values ​​of multiple types; the second subunit is used to obtain the performance parameter correlation coefficient R” by calculating the weighted value of the Pearson correlation coefficient between the target performance parameter and historical performance parameter feature information values ​​of multiple types; the third subunit is used to obtain the defect correlation coefficient R”' by calculating the weighted value of the mean square error between the historical defect feature value and the corresponding defect feature standard value of multiple types of historical defect information.

[0131] After calculating the correlation degree R = R' + R” + 1 / R”', the historical production information of batches of architectural glass with a correlation degree less than the preset correlation degree threshold is filtered out. The historical production information of multiple batches of architectural glass after filtering is preprocessed to generate a historical production information dataset.

[0132] The preprocessing methods include, but are not limited to, data cleaning, partitioning, normalization, and augmentation. Data partitioning refers to dividing the data into training, validation, and test sets. Since data cleaning, partitioning, normalization, and augmentation are conventional techniques in this field, they will not be elaborated upon here.

[0133] By filtering the historical production information dataset according to the preset correlation threshold and the correlation, the historical production information of batches of architectural glass with large differences between historical raw material information and current raw material information, historical performance information and target performance information, and large overall defects can be removed. In other words, the historical production information of batches of architectural glass that do not meet the requirements is removed to ensure the value of the historical production information dataset, improve its quality and reliability, and thus improve the effect and accuracy of training the neural network model through the historical production information dataset.

[0134] For environmental factor information in the production process, preferably, it can be obtained through prediction. The specific steps include: first, preprocessing the historical production information of multiple batches of architectural glass after screening and elimination to generate a historical production information dataset, extracting historical environmental factor information from it, and performing fitting processing on the historical environmental information of each type to obtain a historical environmental factor information fitting curve function. This fitting curve function is a curve function of the feature value of historical environmental factor information with respect to time. Then, based on the fitting curve function corresponding to each type of historical environmental factor information, the predicted value of historical environmental factor information is obtained and used as the environmental factor information in the production process.

[0135] Of course, to simplify calculations, the current environmental factor information can also be directly used as the environmental factor information during the production process of the architectural glass to be produced. More specifically, for example, if the architectural glass production process specifically refers to the melting or forming annealing stage, the environmental factor information at a certain moment before the start of melting or forming annealing can be used as the environmental factor information during the production process of the architectural glass to be produced.

[0136] Preferably, the neural network building module includes a loss function building unit and a neural network model building unit.

[0137] The loss function construction unit is used to construct a multi-objective dynamic loss function L=L1+L2+β×||W based on defect information, energy consumption information, and environmental factor information. defect W energy || 2 The neural network model building unit is used to build a neural network model based on the multi-objective dynamic loss function.

[0138] Where L1=α defect ||w defect || 2 +α energy ||w energy || 2 Let α represent the multi-objective regularization term. defect α energy w represents the defect loss weighting coefficient and the energy consumption loss weighting coefficient, respectively. defect w energy Let f(env) and f(env) represent the feature vector of defect information and the feature vector of energy consumption loss, respectively. Let |||| denote the norm, and L2 = Relu(f(env) - threshold) represent the environmental sensitivity penalty term. Represents the comprehensive environmental sensitivity evaluation function, Relu() represents the activation function, threshold represents the safety threshold for environmental factors, M1 represents the number of types of environmental factors, and λ represents the overall environmental sensitivity evaluation function. i This represents the weight of the evaluation value of the i-th type of environmental factor, and e represents the natural constant. E represents an intermediate variable. i E represents the actual value of the i-th environmental factor. i ' represents the standard value of the i-th environmental factor, and || represents the absolute value function. α represents the performance loss term. performance This represents the performance loss weighting coefficient, M2 represents the number of types of performance information, and p i p represents the actual value of the performance information for the i-th type. i 'Represents the predicted value β×||W for the i-th type of performance information. defect W energy || 2W represents the weighted regularization term, β represents the regularization strength, and the initial value is between 0.01 and 0.1. defect W represents the defect loss weight matrix. energy This represents the energy loss weight matrix.

[0139] Weight regularization term β×||W defect W energy || 2 It is mainly used to apply Frobenius norm constraints to the defect loss weight matrix and the energy consumption loss weight matrix to prevent overfitting of the neural network model during training.

[0140] The environmental sensitivity penalty term L2 = ReLU(f(env) - threshold) integrates environmental factor features from multiple dimensions. By setting an activation function, a non-linear threshold response mechanism is created. When the overall environmental conditions deteriorate and exceed the environmental factor safety threshold, a gradient penalty is triggered, increasing the constraint strength. One of the main functions of setting the environmental sensitivity penalty term is to construct a dynamically updated loss function based on changes in the actual values ​​of environmental factors and to increase the model's constraint strength, enabling the neural network model to adapt to the diversity of input sample data. Specifically,

[0141] For the energy loss weighting coefficient α energy Defect loss weighting coefficient α defect and the performance loss weighting coefficient α performance All of these can be set by technicians based on their experience. Preferably, α energy :α defect :α performance =0.1:0.4:05 or α energy :α defect :α performance = 0.1:0.6:03. The initial value of the regularization strength is between 0.01 and 0.1, and the regularization strength gradually decreases with each training epoch. Where μ0 represents the initial value of the regularization strength, and Epoch represents the training epoch.

[0142] Preferably, Where μ0 represents the initial value of the regularization strength, Epoch represents the training epoch, and α' represents the regularization strength adjustment coefficient. Both α' and μ0 can be set by technicians based on experience. (The formula is used to...) This allows the regularization strength to gradually decrease with each training epoch, and its integration with the comprehensive environment sensitivity evaluation function allows for dynamic adjustment of the regularization strength based on environmental changes. When the overall environment is harsh, weakening the regularization strength avoids underfitting of the model due to adverse environmental conditions, thereby improving the model's accuracy and generalization ability. In summary, through the formula... Adjusting the regularization strength, which integrates training rounds and environmental sensitivity, improves the robustness and generalization ability of the neural network model, reducing the probability of underfitting and overfitting during training. Based on this, the trained neural network model, using the multi-objective dynamic loss function, can obtain the optimal process parameters that meet the actual production efficiency needs of architectural glass manufacturers, thus improving the overall intelligence of the system and the overall production efficiency of architectural glass production.

[0143] In summary, the architectural glass production system, by constructing a multi-objective dynamic loss function and incorporating environmental factor information into the loss function, can take into account the potential impact of environmental factors on the quality and efficiency of architectural glass, such as defects, energy consumption, and performance, thus ensuring the robustness and accuracy of the neural network model. Simultaneously, it overcomes the problem of existing architectural glass production methods that optimize for a single production objective, failing to comprehensively improve the overall production efficiency of architectural glass production. This facilitates improving the overall production efficiency of architectural glass based on actual production needs. Based on the constructed multi-objective dynamic loss function and neural network model, by adjusting the weight coefficients of defect loss, energy consumption loss, and performance loss, the optimal process parameters that meet the actual needs of architectural glass manufacturers can be obtained, thereby improving the overall intelligence of the system.

[0144] This invention also provides an architectural glass production device, comprising: a controller; and a memory storing executable instructions; wherein the executable instructions can run on the controller to implement the architectural glass production method.

[0145] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0146] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for producing architectural glass, characterized in that, The architectural glass production method includes the following steps: Obtain the type of architectural glass to be produced and obtain a dataset of historical production information for multiple different batches of the same type of architectural glass. A multi-objective dynamic loss function is constructed based on defect information, energy consumption information, performance information, and environmental factor information, and a neural network model is built based on the loss function. The neural network model is trained based on the historical production information dataset. The basic information of the architectural glass to be produced is obtained, the basic information is input into the neural network model to obtain the optimal process control parameters, and the architectural glass is produced according to the optimal process control parameters. The historical production information dataset includes historical raw material information, historical process control parameters, historical defect information, historical performance information, historical energy consumption information, and historical environmental factor information. The basic information includes current raw material information, target performance information, and environmental factor information of the production process. The multi-objective dynamic loss function ; in, Represents a multi-objective regularization term. These represent the defect loss weighting coefficient and the energy consumption loss weighting coefficient, respectively. These represent the defect information feature vector and the energy loss feature vector, respectively. Represents the norm, Indicates environmentally sensitive penalty items. This represents the comprehensive environmental sensitivity evaluation function. This represents the activation function. Indicates the safety threshold of environmental factors. Indicates the number and types of environmental factors. Indicates the first The weight of the evaluation value for each type of environmental factor. Represents the natural constant. Indicates intermediate variables. Indicates the first Actual values ​​of each environmental factor Indicates the first Standard values ​​for each environmental factor Represents the absolute value function. Represents the performance loss term. This represents the performance loss weighting coefficient. Indicates the number of types of performance information. Indicates the first Actual values ​​of performance information for each category Indicates the first Predicted values ​​of performance information for each category This represents the weight regularization term. Indicates the regularization strength. This represents the defect loss weight matrix. This represents the energy loss weight matrix.

2. The method for producing architectural glass as described in claim 1, characterized in that, The specific method for training the neural network model based on the historical production information dataset includes the following steps: Obtain the correlation between multiple different batches of architectural glass of the same type and the architectural glass to be produced; The historical production information dataset is filtered based on a preset correlation threshold and the correlation degree. The neural network model is trained based on the selected historical production information dataset.

3. The method for producing architectural glass as described in claim 2, characterized in that, correlation These represent the correlation coefficients of raw materials, performance parameters, and defects, respectively. It is obtained by calculating the weighted value of the Pearson correlation coefficient between the characteristic information values ​​of multiple types of current raw materials and historical raw materials; It is obtained by calculating the weighted value of the Pearson correlation coefficient between multiple types of target performance parameters and historical performance parameter feature information values; It is obtained by calculating the weighted average error between the historical defect feature values ​​and the corresponding defect feature standard values ​​of multiple types of historical defect information.

4. The method for producing architectural glass as described in claim 3, characterized in that, Specific methods for filtering historical production information datasets include: removing and filtering historical production information of batches of architectural glass with a correlation degree less than a preset correlation degree threshold; The specific method for training the neural network model based on the filtered historical production information dataset includes: preprocessing the historical production information of multiple batches of architectural glass after filtering to generate a historical production information dataset, and training the neural network model based on the historical production information dataset to obtain the trained neural network model.

5. A building glass production system for implementing the building glass production method as described in any one of claims 1-4, characterized in that, The architectural glass production system includes: The production information acquisition module is used to acquire the type of architectural glass to be produced and to acquire historical production information datasets of multiple different batches of architectural glass of the same type. The neural network construction module is used to construct a multi-objective dynamic loss function based on defect information, energy consumption information, performance information and environmental factor information, and to construct a neural network model based on the loss function; The training module is used to train the neural network model based on the historical production information dataset; The process parameter acquisition module is used to acquire basic information about the architectural glass to be produced, input the basic information into the neural network model to obtain optimal process control parameters, and produce the architectural glass according to the optimal process control parameters. The historical production information dataset includes historical raw material information, historical process control parameters, historical defect information, historical performance information, historical energy consumption information, and historical environmental factor information. The basic information includes current raw material information, target performance information, and environmental factor information of the production process. The neural network building blocks include: The loss function construction unit is used to construct a multi-objective dynamic loss function based on defect information, energy consumption information, and environmental factor information. ; A neural network model building unit is used to build a neural network model based on the multi-objective dynamic loss function. in, Represents a multi-objective regularization term. These represent the defect loss weighting coefficient and the energy consumption loss weighting coefficient, respectively. These represent the defect information feature vector and the energy loss feature vector, respectively. Represents the norm, Indicates environmentally sensitive penalty items. This represents the comprehensive environmental sensitivity evaluation function. This represents the activation function. Indicates the safety threshold of environmental factors. Indicates the number and types of environmental factors. Indicates the first The weight of the evaluation value for each type of environmental factor. Represents the natural constant. Indicates intermediate variables. Indicates the first Actual values ​​of each environmental factor Indicates the first Standard values ​​for each environmental factor Represents the absolute value function. Represents the performance loss term. This represents the performance loss weighting coefficient. Indicates the number of types of performance information. Indicates the first Actual values ​​of performance information for each category Indicates the first Predicted values ​​of performance information for each category This represents the weight regularization term. Indicates the regularization strength. This represents the defect loss weight matrix. This represents the energy loss weight matrix.

6. The architectural glass production system as described in claim 5, characterized in that, The training module includes: The correlation degree acquisition unit is used to acquire the correlation degree between multiple different batches of architectural glass of the same type and the architectural glass to be produced. A filtering unit is used to filter historical production information datasets based on a preset correlation threshold and the correlation degree. The training unit is used to train the neural network model based on the filtered historical production information dataset.

7. The architectural glass production system as described in claim 6, characterized in that, The correlation acquisition unit includes: The first subunit is used to obtain the raw material correlation coefficient by calculating the weighted value of the Pearson correlation coefficient between the characteristic information values ​​of multiple types of current raw materials and historical raw materials. ; The second subunit is used to obtain the performance parameter correlation coefficient by calculating the weighted value of the Pearson correlation coefficient between multiple types of target performance parameters and historical performance parameter feature information values. ; The third subunit is used to obtain the defect correlation coefficient by calculating the weighted average error between the historical defect feature values ​​and the corresponding defect feature standard values ​​of multiple types of historical defect information. ; Among them, correlation .

8. A type of architectural glass production equipment, characterized in that, The architectural glass production equipment includes: Controller; Memory, which stores executable instructions; The executable instructions can run on the controller and implement the architectural glass production method as described in any one of claims 1 to 4.

Images