Training method, kiln molten glass viscosity control method and device, and terminal

Abstract

The invention provides a training method, a kiln molten glass viscosity control method, a kiln molten glass viscosity control device and a terminal. The kiln molten glass viscosity control method comprises the following steps: acquiring real-time operation state data of a kiln; the real-time operation state data of the kiln comprises the liquid level of molten glass in the kiln, the feeding speed of a batch feeder, the temperature in the kiln and the pressure in the kiln, which are collected at the current moment; inputting the real-time operation state data of the kiln into a pre-trained neural network model, and outputting real-time molten glass viscosity; generating a control signal based on the real-time viscosity of the molten glass; and based on the control signal, automatically adjusting the real-time viscosity of the molten glass to a target viscosity range. According to the method, the excellent self-learning ability and the mapping ability of the neural network model to the complex nonlinear relationship are fully utilized, the viscosity of the molten glass can be predicted in real time, and related process parameters are automatically adjusted according to the prediction result, so that the viscosity of the molten glass is accurately controlled, and the viscosity of the molten glass is maintained in a specific target range.

Description

Technical Field

[0001] This application belongs to the technical field of glass production, and relates to a training method, a method and device for controlling the viscosity of glass liquid in a kiln, and a terminal. Background Art

[0002] During the glass production process, the viscosity of the glass liquid in the kiln is a key parameter, which directly affects the production stability and product quality. The change in the viscosity of the glass liquid will significantly change its flow state, and thus have a complex impact on the production process. For example, when the viscosity of the glass liquid is too low, the flow rate of the glass liquid at the kiln outlet will become unstable, resulting in quality problems such as uneven thickness during the forming process of glass products. On the contrary, if the viscosity of the glass liquid is too high, the pressure distribution of the glass liquid in the kiln will change, increasing the impact on the kiln wall, accelerating the damage of the kiln, and even possibly triggering safety accidents, seriously interfering with the continuity of production. Therefore, maintaining an appropriate viscosity of the glass liquid not only helps to ensure product quality, but also optimizes energy utilization, reduces heat loss, and maintains a stable thermal environment.

[0003] Currently, the viscosity of the glass liquid in the kiln mainly depends on empirical operation and manual adjustment, lacking real-time monitoring and precise control means. This method is not only inefficient, but also difficult to cope with complex production environment changes, resulting in insufficient control accuracy of the glass liquid viscosity and affecting production efficiency and product quality. Summary of the Invention

[0004] The purpose of this application is to provide a training method, a method and device for controlling the viscosity of glass liquid in a kiln, and a terminal, which are used to solve the technical problem that the prior art cannot intelligently control the viscosity of glass liquid, resulting in too high or too low viscosity of glass liquid.

[0005] In a first aspect, this application provides a training method, including: obtaining historical operation state data of a kiln as training samples; the historical operation state data includes the liquid level of the glass liquid in the kiln, the feeding speed of the feeder, the temperature in the kiln, and the pressure in the kiln collected at historical moments; initializing the parameters of the neural network model to be trained to obtain an initialized neural network model; based on the initialized neural network model, performing forward propagation calculation on the training samples to obtain a predicted value of the glass liquid viscosity, and calculating the prediction error of the predicted value relative to the true value; taking minimizing the prediction error as the goal, using the gradient descent method to perform iterative optimization processing on the parameters of the initialized neural network model to obtain a trained neural network model.

[0006] In an implementation of the first aspect, initializing the parameters of the neural network model to be trained, the initialized neural network model includes: obtaining the numbers of the input layer neurons, the hidden layer neurons, and the output layer neurons of the neural network model to be trained; using a neural network weight initialization method to initialize the connection weights between the input layer neurons and the hidden layer neurons, and the connection weights between the hidden layer neurons and the output layer neurons; using a random initialization method to initialize the bias terms of the hidden layer neurons and the bias terms of the output layer neurons.

[0007] In an implementation of the first aspect, the number of the hidden layer neurons is determined by the numbers of the input layer neurons and the output layer neurons; the calculation formula for the number of the hidden layer neurons is:

[0008]

[0009] where h is the number of the hidden layer neurons, m is the number of the input layer neurons, n is the number of the output layer neurons, and a is any constant within the value range of [1, 10].

[0010] In an implementation of the first aspect, based on the initialized neural network model, performing forward propagation calculation on the training samples to obtain the predicted value of the glass melt viscosity includes: based on the training samples, the connection weights between the input layer neurons and the hidden layer neurons, and the bias terms of the hidden layer neurons, calculating the outputs of different hidden layer neurons in the initialized neural network model, and combining the outputs of the hidden layer neurons to obtain a hidden layer output vector; based on the hidden layer output vector, the connection weights between the hidden layer neurons and the output layer neurons, and the bias terms of the output layer neurons, calculating the output of the output layer neurons to obtain the predicted value of the glass melt viscosity.

[0011] In an implementation of the first aspect, aiming to minimize the prediction error, the gradient descent method is used to iteratively optimize the parameters of the initialized neural network model. The trained neural network model is obtained as follows: Determine whether the prediction error converges; if so, end the training of the neural network model; otherwise, calculate the gradient of the prediction error with respect to the connection weights between the hidden layer neurons and the output layer neurons to obtain a first gradient; calculate the gradient of the prediction error with respect to the bias term of the output layer neurons to obtain a second gradient; based on the first gradient and a preset learning rate, update the connection weights between the hidden layer neurons and the output layer neurons; based on the second gradient and the preset learning rate, update the bias term of the output layer neurons; calculate the gradient of the prediction error with respect to the connection weights between the input layer neurons and the hidden layer neurons to obtain a third gradient; calculate the gradient of the prediction error with respect to the bias term of the hidden layer neurons to obtain a fourth gradient; based on the third gradient and the preset learning rate, update the connection weights between the input layer neurons and the hidden layer neurons; based on the fourth gradient and the preset learning rate, update the bias term of the hidden layer neurons to obtain an updated neural network model; based on the updated neural network model, calculate a new prediction error; repeat the above process of convergence determination of the prediction error and update of the neural network model until the prediction error converges.

[0012] In an implementation of the first aspect, before performing the forward propagation calculation on the training samples, preprocessing of the training samples is further included, where the steps of the preprocessing at least include: using a normalization method to uniformly map the training samples to a preset interval; using a moving average filtering method to denoise the training samples.

[0013] In the second aspect, the present application provides a method for controlling the viscosity of glass melt in a kiln, including: obtaining real-time operating state data of the kiln; the real-time operating state data of the kiln includes the liquid level of the glass melt in the kiln, the feeding speed of the feeder, the temperature in the kiln, and the pressure in the kiln collected at the current moment; inputting the real-time operating state data of the kiln into a pre-trained neural network model to output the real-time glass melt viscosity; generating a control signal based on the real-time glass melt viscosity; automatically adjusting the real-time glass melt viscosity to a target viscosity range based on the control signal.

[0014] In an implementation of the second aspect, the training method of the neural network model includes: obtaining historical operation state data of the kiln as training samples; the historical operation state data includes the liquid level of the molten glass in the kiln, the feeding speed of the feeder, the temperature in the kiln, and the pressure in the kiln collected at historical moments; initializing the parameters of the neural network model to be trained to obtain the initialized neural network model; based on the initialized neural network model, performing forward propagation calculation on the training samples to obtain the predicted value of the molten glass viscosity, and calculating the prediction error of the predicted value relative to the true value; aiming at minimizing the prediction error, using the gradient descent method to perform iterative optimization processing on the parameters of the initialized neural network model to obtain the trained neural network model.

[0015] In a third aspect, the present application provides a device for controlling the viscosity of molten glass in a kiln, including: a real-time data acquisition module for acquiring the real-time operation state data of the kiln; the real-time operation state data of the kiln includes the liquid level of the molten glass in the kiln, the feeding speed of the feeder, the temperature in the kiln, and the pressure in the kiln collected at the current moment; a molten glass viscosity prediction module for inputting the real-time operation state data of the kiln into a pre-trained neural network model and outputting the real-time molten glass viscosity; a control signal generation module for generating a control signal based on the real-time molten glass viscosity; a molten glass viscosity adjustment module for automatically adjusting the real-time molten glass viscosity to the target viscosity range based on the control signal.

[0016] In a fourth aspect, the present application provides a terminal, including: a processor and a memory; the memory is used for storing a computer program; the processor is used for executing the computer program stored in the memory so that the terminal executes the training method described in any one of the above and / or the method for controlling the viscosity of molten glass in a kiln described in any one of the above.

[0017] As described above, the training method, the method and device for controlling the viscosity of molten glass in a kiln, and the terminal described in the present application have the following

[0018] Beneficial effects:

[0019] (1) Make full use of the excellent self-learning ability of the neural network model and the mapping ability for complex non-linear relationships, can predict the viscosity of molten glass in real time, and automatically adjust relevant process parameters according to the prediction results to achieve precise control of the viscosity of molten glass;

[0020] (2) By designing a reasonable intelligent control strategy, the viscosity of the glass liquid can be maintained within a specific target range, ensuring stable and consistent flow characteristics of the glass liquid inside the furnace, achieving stable production of high-quality glass products, ensuring the working efficiency of the feeder, the rational utilization of raw materials, and the safety of the production process, effectively reducing raw material waste, optimizing the feeding rhythm, improving production efficiency, and promoting the intelligent development process of the glass production industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It shows a hardware structure block diagram of the mobile terminal described in this application in an embodiment.

[0022] Figure 2 It shows a flowchart of the method for controlling the viscosity of the furnace glass liquid described in this application in an embodiment.

[0023] Figure 3 It shows a schematic diagram of the positions of different types of sensors described in this application in an embodiment.

[0024] Figure 4 It shows a schematic diagram of the structure of the pre-trained neural network model described in this application in an embodiment.

[0025] Figure 5 It shows a flowchart of the training method of the neural network model described in this application in an embodiment.

[0026] Figure 6 It shows an initialization flowchart of the method for controlling the viscosity of the furnace glass liquid described in this application in an embodiment.

[0027] Figure 7 It shows a forward propagation calculation flowchart of the method for controlling the viscosity of the furnace glass liquid described in this application in an embodiment.

[0028] Figure 8 It shows a flowchart of the gradient descent method described in this application in an embodiment.

[0029] Figure 9 It shows a schematic diagram of the structure of the device for controlling the viscosity of the furnace glass liquid described in this application in an embodiment.

[0030] Figure 10 It shows a schematic diagram of the structure of the terminal described in this application in an embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The following describes the implementation manners of the present application through specific examples. Those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in this specification. The present application can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0032] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner. Therefore, only the components related to the present application are shown in the diagrams, rather than being drawn according to the number, shape, and size of the components in actual implementation. The type, quantity, and proportion of each component in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0033] In addition, in the present application, descriptions such as "first" and "second" are only for descriptive purposes, and cannot be understood as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" can explicitly or implicitly include at least one of such features. In addition, the technical solutions between various embodiments can be combined with each other, but it must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present application.

[0034] The following embodiments of the present application provide a training method, a method and device for controlling the viscosity of furnace glass liquid, and a terminal. The present application makes full use of the excellent self-learning ability of the neural network model and the mapping ability for complex non-linear relationships, can predict the viscosity of the glass liquid in real time, and automatically adjust relevant process parameters according to the prediction results to achieve precise control of the viscosity of the glass liquid.

[0035] Both the method for controlling the viscosity of furnace glass liquid and the training method provided by the embodiments of the present application can run in similar devices such as mobile terminals and computer terminals. Taking running on the mobile terminal as an example, Figure 1 is the hardware structure block diagram of the mobile terminal, as Figure 1 taking one mobile terminal as an example, the mobile terminal may include: a processor and a memory. The processor may be a central processing unit, and the memory is used to store data. Figure 1 The mobile terminal in is only for illustration and does not limit the specific structure of the mobile terminal.

[0036] Optionally, the mobile terminal may further include: a communication transmission device and an input / output device.

[0037] Optionally, the memory may be used to store computer programs, such as software programs and modules of application software, and the memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the mobile terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0038] Optionally, the communication transmission device can be used to receive or send data via a network, which may include a wireless network provided by a communication provider of the mobile terminal. The communication transmission device may include a NIC (Network Interface Controller) that can be connected to other network devices through a base station so as to communicate with the Internet.

[0039] The technical solutions in the embodiments of the present application will be described in detail below in conjunction with the drawings in the embodiments of the present application.

[0040] See also Figure 2 , which is a flow chart of a method for controlling the viscosity of glass liquid in a furnace according to the present application in one embodiment. Figure 2 As shown, the method for controlling the viscosity of the furnace glass liquid provided in the embodiment of the present application includes the following steps S10 to S40.

[0041] In step S10, real-time operation status data of the kiln is obtained; the real-time operation status data of the kiln includes the liquid level of the glass liquid in the kiln, the feeding speed of the feeder, the temperature in the kiln and the pressure in the kiln collected at the current moment.

[0042] In this embodiment, the liquid level of the molten glass in the kiln, the feeding speed of the feeder, the temperature in the kiln and the pressure in the kiln are all important factors affecting the viscosity of the molten glass.

[0043] Specifically, the level of the glass liquid in the kiln reflects the amount of glass liquid in the kiln, and the feed rate of the feeder determines the supply of raw materials used to generate glass. In addition, the viscosity of the glass liquid will change significantly with the change of the temperature in the kiln. Generally, the higher the temperature in the kiln, the lower the viscosity of the glass liquid. The pressure in the kiln affects the viscosity of the glass liquid by affecting the bubble content and fluidity in the glass liquid.

[0044] During the data collection stage, accurate and reliable real-time operating status data of the kiln can be obtained by deploying different types of sensors inside the kiln.

[0045] See also Figure 3, which is a schematic diagram showing the positions of different types of sensors described in the present application in one embodiment.

[0046] For example, to measure the feed rate v of a feeder, a high-precision electromagnetic flowmeter can be used and installed at the key position of raw material feeding. The measurement accuracy of the electromagnetic flowmeter is as high as ±0.5%, which can accurately capture the instantaneous and cumulative flow information of the feed rate, thereby providing solid data support for the neural network model.

[0047] For the measurement of the temperature T in the kiln, thermocouple temperature sensors can be used and widely distributed in multiple key parts of the kiln, such as the kiln vault, cooling part, clarification part pool bottom and cooling part pool bottom. Thermocouple temperature sensors have a temperature measurement range of 0 to 1600°C and a resolution of 0.1°C, which can fully and accurately obtain the temperature distribution details in the kiln.

[0048] For the measurement of the kiln pressure P, a piezoresistive pressure sensor can be selected and installed at the pressure measurement interface of the kiln. The measurement accuracy of the piezoresistive pressure sensor is ±0.05MPa, which can keenly monitor the slight changes in the pressure inside the kiln.

[0049] It should be noted that the real-time operating status data of the kiln may also include other factors that affect the viscosity of the glass liquid, so as to achieve comprehensive monitoring of the operating status of the kiln. To save space, they are not listed here one by one. The present application provides strong support for the subsequent control and adjustment of the viscosity of the glass liquid by real-time and comprehensive monitoring of the operating status data of the kiln.

[0050] In step S20, the real-time operating status data of the kiln is input into a pre-trained neural network model, and the real-time glass liquid viscosity is output.

[0051] See also Figure 4 , which is a schematic diagram of the structure of the pre-trained neural network model described in the present application in one embodiment.

[0052] In one embodiment of the present application, the pre-trained neural network model may be a back propagation (BP) neural network model.

[0053] The BP neural network model provides a solution for accurately and quickly predicting the viscosity of molten glass with its powerful self-learning and adaptive capabilities and good mapping capabilities for complex nonlinear relationships. Specifically, the BP neural network model can effectively capture the complex nonlinear relationship between the operating status data of the kiln and the viscosity of the molten glass through the connection and weight adjustment of multiple layers of neurons. In practical applications, when real-time data is input into the trained BP neural network model, the model can quickly calculate the current viscosity value of the molten glass, providing a reliable basis for the generation of subsequent control signals.

[0054] In step S30, a control signal is generated based on the real-time glass liquid viscosity.

[0055] In one embodiment of the present application, based on the real-time glass liquid viscosity, generating a control signal includes: setting a target viscosity range; the target viscosity range is related to the production process, product quality requirements or energy efficiency; comparing the output real-time glass liquid viscosity with the set target viscosity range, and calculating the viscosity deviation; generating the control signal based on the viscosity deviation.

[0056] Under complex working conditions, the BP neural network model can demonstrate adaptive adjustment capabilities, thereby achieving dynamic updating of control signals.

[0057] For example, when the temperature in the kiln rises or falls rapidly, causing the viscosity of the glass liquid to change significantly, the BP neural network model can sense the change in the viscosity of the glass liquid in real time and adjust the control signal accordingly, thereby quickly offsetting the impact of temperature fluctuations on the viscosity of the glass liquid. Similarly, when there is a sudden change in the feed speed of the feeder, the BP neural network model can adjust the control signal of the feeder in time to ensure that the feed amount is consistent with production needs, avoiding the problem of instability of the glass liquid viscosity due to excessive or insufficient feed.

[0058] In this implementation, the intelligent control of the glass liquid viscosity shows high robustness and real-time performance, and can dynamically adapt to the complex and changeable working conditions in the glass production process, thereby comprehensively improving the automation level and production quality of glass production.

[0059] In step S40, based on the control signal, the real-time glass liquid viscosity is automatically adjusted to a target viscosity range.

[0060] Specifically, based on the control signal, the operating behavior of the actuator is dynamically adjusted to achieve the goal of intelligent control of the viscosity of the glass liquid.

[0061] For example, when it is necessary to reduce the viscosity of the molten glass, the liquid level of the molten glass in the kiln can be lowered by reducing the input of raw materials by the feeder, so as to increase the viscosity of the molten glass; conversely, the viscosity of the molten glass can be reduced by increasing the input of raw materials. In addition, the viscosity of the molten glass can be reduced by increasing the heating power of the furnace heating device to raise the temperature in the kiln; or the viscosity of the molten glass can be increased by reducing the heating power of the furnace heating device.

[0062] In this implementation, by designing a reasonable intelligent control strategy, the viscosity of the molten glass can be maintained within a specific target range, ensuring that the flow characteristics of the molten glass in the kiln are stable and consistent, realizing the stable production of high-quality glass products, ensuring the working efficiency of the feeder, the reasonable utilization of raw materials, and the safety of the production process, effectively reducing raw material waste, optimizing the feeding rhythm, improving production efficiency, and promoting the intelligent development process of the glass production industry.

[0063] Please refer to Figure 5 , which shows the flow chart of the training method of the neural network model described in this application in an embodiment. As Figure 5 shown, the training method of the neural network model may include the following steps S21 to step S24.

[0064] In step S21, obtain the historical operating state data of the kiln as training samples; the historical operating state data includes the liquid level of the molten glass in the kiln, the feeding speed of the feeder, the temperature in the kiln, and the pressure in the kiln collected at historical moments.

[0065] In this embodiment, the liquid level height of the molten glass in the kiln is represented as h, the feeding speed of the feeder is represented as v, the temperature in the kiln is represented as T, and the pressure in the kiln is represented as P.

[0066] In this embodiment, the liquid level of the molten glass in the kiln, the feeding speed of the feeder, the temperature in the kiln, and the pressure in the kiln are all important factors affecting the viscosity of the molten glass.

[0067] The target viscosity range of the molten glass viscosity is determined by comprehensively considering various factors. From the perspective of the process requirements of glass products, different types and specifications of glass products, such as flat glass and glass bottles, etc., require specific molten glass flow conditions and pressure conditions, and these conditions are closely related to the viscosity of the molten glass. At the same time, it is also necessary to combine the structure and size of the kiln to ensure that the molten glass has a suitable residence time and heat exchange efficiency in the kiln. In addition, the energy consumption factor cannot be ignored. A suitable viscosity of the molten glass helps to maintain a stable thermal environment, reduce heat loss, and thus optimize energy utilization.

[0068] In the data acquisition stage, accurate and reliable historical operation state data of the kiln can also be obtained by arranging different types of sensors inside the kiln. The types and arrangement methods of the sensors have been described in detail in the embodiments of step S10, and will not be elaborated here.

[0069] In step S22, initialize the parameters of the neural network model to be trained to obtain the initialized neural network model.

[0070] Please refer to Figure 6 , which shows the initialization flowchart of the method for controlling the viscosity of the glass melt in the kiln according to an embodiment of the present application. As Figure 6 shown, initializing the parameters of the neural network model to be trained to obtain the initialized neural network model includes steps S220 to S222.

[0071] In step S220, obtain the numbers of neurons in the input layer, hidden layer, and output layer of the neural network model to be trained.

[0072] In an embodiment of the present application, the number of neurons in the input layer is determined by the historical operation state data of the kiln, and the number of neurons in the output layer is determined by the viscosity of the glass melt.

[0073] Specifically, since the historical operation state data of the kiln includes the liquid level of the glass melt in the kiln, the feeding speed of the feeder, the temperature in the kiln, and the pressure in the kiln, these four key parameters together constitute the basic information reflecting the flow and heat exchange of the glass melt in the kiln. In order to comprehensively capture this information, 4 neurons are set in the input layer, and each neuron corresponds to a key parameter respectively, ensuring that the neural network model can receive complete and accurate input signals. Since the viscosity of the glass melt is a scalar value, only 1 neuron is set in the output layer to output the predicted viscosity of the glass melt.

[0074] In an embodiment of the present application, a strategy combining the empirical formula method and the experimental comparison method can be adopted to determine the number of neurons in the hidden layer.

[0075] Taking the empirical formula method as an example, the number of neurons in the hidden layer is determined by the number of neurons in the input layer and the number of neurons in the output layer. The calculation formula for the number of neurons in the hidden layer is:

[0076]

[0077] where h is the number of neurons in the hidden layer, m is the number of neurons in the input layer, n is the number of neurons in the output layer, and a is an arbitrary constant in the value range of [1, 10].

[0078] For example, when the number of neurons in the input layer m = 5, the number of neurons in the output layer n = 1, and a = 5, the number of neurons in the hidden layer h can be estimated to be approximately 8.

[0079] It should be noted that although the empirical formula method can provide a preliminary reference, to obtain a more accurate neural network structure, it is necessary to further carry out the experimental comparison method.

[0080] Specifically, multiple neural network architectures with different numbers of neurons in the hidden layer can be constructed, and each group of neural network architectures can be strictly trained and tested. During the training process, professional evaluation indicators such as Mean Squared Error (MSE) or Mean Absolute Error (MAE) can be used to evaluate the model performance, and the model with the best performance can be selected for application.

[0081] In this implementation, first, an initial estimated value of the number of neurons in the hidden layer is obtained using the empirical formula method, and then further optimization is carried out through the experimental comparison method on this basis. This can not only ensure that there is a certain theoretical basis as a guide but also find the most suitable network structure for the current problem through actual operations. This strategy of combining the empirical formula method and the experimental comparison method helps to construct a more accurate and efficient neural network model.

[0082] In step S221, the neural network weight initialization method is used to initialize the connection weights between the input layer neurons and the hidden layer neurons and the connection weights between the hidden layer neurons and the output layer neurons.

[0083] In an embodiment of the present application, the neural network weight initialization method can be the Xavier initialization method. Xavier initialization (also known as Gloroti initialization) can be used to solve the problem of gradient disappearance or explosion in deep neural networks. The core idea of the Xavier initialization method is to make the signal stable during the forward and backward propagation processes of the network by setting reasonable initial weights.

[0084] Specifically, the formula used in Xavier initialization is:

[0085]

[0086] where w is the connection weight between the neurons in the current layer and the neurons in the next layer, n j is the number of neurons in the current layer, n j+1 is the number of neurons in the next layer, and r is a random number in the range of [−1, 1].

[0087] It should be noted that the above formula is applicable to the initialization of the connection weights between the input layer neurons and the hidden layer neurons, as well as the connection weights between the hidden layer neurons and the output layer neurons, and will not be elaborated here.

[0088] In step S222, a random initialization method is adopted to initialize the bias terms of the hidden layer neurons and the bias terms of the output layer neurons.

[0089] In an embodiment of the present application, the bias terms of the hidden layer neurons and the bias terms of the output layer neurons are both random numbers within the value range of [0,

[0090] 0.1].

[0091] The selection of the initialization strategy plays a crucial role in the training effect and convergence speed of the BP neural network model. By adopting a strategy that combines Xavier initialization and random initialization, the present application can make full use of the advantages of both initializations, ensure that the activation value distributions of each layer are reasonable and orderly in the initial stage of neural network model training, effectively avoid thorny problems such as gradient disappearance or gradient explosion, and also endow the neural network model with a certain degree of initial flexibility and adaptability.

[0092] In step S23, based on the initialized neural network model, forward propagation calculation is performed on the training samples to obtain the predicted value of the glass liquid viscosity, and the prediction error of the predicted value relative to the true value is calculated.

[0093] Please refer to Figure 7 , which shows the forward propagation calculation flowchart in an embodiment of the method for controlling the viscosity of furnace glass liquid according to the present application. As Figure 7 shown, in an embodiment of the present application, based on the initialized neural network model, forward propagation calculation is performed on the training samples, and the steps for obtaining the predicted value of the glass liquid viscosity may include the following steps S230 and S231.

[0094] In step S230, based on the training samples, the connection weights between the input layer neurons and the hidden layer neurons, and the bias terms of the hidden layer neurons, the outputs of different hidden layer neurons in the initialized neural network model are calculated, and the outputs of each hidden layer neuron are combined to obtain a hidden layer output vector.

[0095] Specifically, according to the formula calculate the outputs of each hidden layer neuron in the neural network model, and combine them to obtain the hidden layer output vector H=(h1, h2, …, h h ) T ; where is the activation function, m is the number of input layer neurons, is the connection weight between the i-th input layer neuron and the j-th hidden layer neuron, and x i is the training sample input by the i-th input layer neuron, is the bias term of the j-th hidden layer neuron, and h h is the output of the h-th hidden layer neuron.

[0096] In step S231, based on the hidden layer output vector, the connection weight between the hidden layer neuron and the output layer neuron, and the bias term of the output layer neuron, calculate the output of the output layer neuron to obtain the predicted value of the glass liquid viscosity.

[0097] Specifically, according to the formula calculate the output of the output layer neuron to obtain the predicted value of the glass liquid viscosity, where is the activation function, h is the number of hidden layer neurons, is the connection weight between the j-th hidden layer neuron and the output layer neuron, and h j is the output of the j-th hidden layer neuron, and b 2 is the bias term of the output layer neuron.

[0098] In an embodiment of the present application, the mean square error formula or the mean absolute error formula is used to calculate the prediction error of the predicted value relative to the true value.

[0099] For example, the total number of training samples is N. For the k-th training sample, the calculation method of the prediction error of the predicted value of the glass liquid viscosity relative to the true value is:

[0100]

[0101] where y k is the predicted value of the glass liquid viscosity obtained by the neural network model through forward propagation calculation for the k-th training sample, and t k is the corresponding true value.

[0102] The sum of the prediction errors of the N training samples can be expressed as:

[0103]

[0104] In other embodiments, the sum of the prediction errors of the N training samples can also be expressed as:

[0105]

[0106] where e pred,k is the prediction error corresponding to the k-th training sample.

[0107] In an embodiment of the present application, before performing forward propagation calculation on the training samples, preprocessing of the training samples is further included, where the steps of the preprocessing at least include: uniformly mapping the training samples to a preset interval by using a normalization method; performing denoising processing on the training samples by using a moving average filtering method.

[0108] Since there are significant differences in the dimensions and numerical ranges of the liquid level of the molten glass in the kiln, the feeding speed of the feeder, the temperature in the kiln, and the pressure in the kiln collected, if directly used for the training of the neural network model, it will seriously affect the training efficiency and accuracy. The present application uniformly maps the training samples to a preset interval by using a normalization method, which can eliminate the weight differences caused by different dimensions between different features.

[0109] Taking the feeding speed of the feeder as an example, the following formula can be used to perform normalization processing on the feeding speed of the feeder:

[0110]

[0111] where v is the feeding speed of the feeder, v min is the minimum value in the feeding speed of the feeder, v max is the maximum value in the feeding speed of the feeder, v norm is the feeding speed of the feeder after normalization.

[0112] Since the glass production site environment is complex and changeable, the collected data is extremely vulnerable to noise interference. To solve this problem, the present application performs denoising processing on the data by using a moving average filtering method, effectively smoothing the noise fluctuations in the data and improving the data quality.

[0113] Specifically, the following formula can be used to perform denoising processing on the collected data:

[0114]

[0115] where {x j} is the original data sequence (i.e., the data sequence before filtering), {y i} is the data sequence after filtering, and k is the moving average window size.

[0116] In this implementation manner, the normalization method enables different features to participate in training on the same scale, avoiding the problem of weight imbalance caused by dimensional differences. The moving average filtering method improves the quality and reliability of the data and reduces the interference of noise on model learning. After being processed by these preprocessing steps, the training samples become more suitable for neural network training, thus laying a solid foundation for constructing a high-performance and high-accuracy kiln furnace molten glass viscosity control model.

[0117] In step S24, with the goal of minimizing the prediction error, the gradient descent method is used to iteratively optimize the parameters of the initialized neural network model to obtain a trained neural network model.

[0118] In this application, the weights and biases of the network are continuously optimized through error backpropagation. In the error backpropagation stage, the error starts from the output layer and propagates through the hidden layers to the input layer in the opposite direction of the forward propagation. During the propagation process, the weights and bias terms of the connections between neurons in each layer are adjusted according to the error, so that the output value of the network gradually approaches the target value.

[0119] Please refer to Figure 8 , which shows the flowchart of the gradient descent method described in this application in an embodiment.

[0120] As Figure 8 shown, with the goal of minimizing the prediction error, using the gradient descent method to iteratively optimize the parameters of the initialized neural network model to obtain a trained neural network model includes: determining whether the prediction error converges; if so, ending the training of the neural network model; otherwise, calculating the gradient of the prediction error with respect to the connection weight between the hidden layer neurons and the output layer neurons to obtain the first gradient; calculating the gradient of the prediction error with respect to the bias term of the output layer neurons to obtain the second gradient; based on the first gradient and the preset learning rate, updating the connection weight between the hidden layer neurons and the output layer neurons; based on the second gradient and the preset learning rate, updating the bias term of the output layer neurons; calculating the gradient of the prediction error with respect to the connection weight between the input layer neurons and the hidden layer neurons to obtain the third gradient; calculating the gradient of the prediction error with respect to the bias term of the hidden layer neurons to obtain the fourth gradient; based on the third gradient and the preset learning rate, updating the connection weight between the input layer neurons and the hidden layer neurons; based on the fourth gradient and the preset learning rate, updating the bias term of the hidden layer neurons to obtain an updated neural network model; based on the updated neural network model, calculating a new prediction error; repeating the above prediction error convergence judgment and neural network model update process until the prediction error converges.

[0121] Specifically, the calculation formula for the first gradient is:

[0122]

[0123] where N is the total number of training samples, y k is the predicted value of the glass melt viscosity obtained by the neural network model through forward propagation calculation for the k-th training sample, t k is the corresponding true value, is the derivative of the activation function, is the connection weight between the j-th hidden layer neuron and the output layer neuron, h is the number of hidden layer neurons, h j is the output of the j-th hidden layer neuron, b 2 is the bias term of the output layer neuron.

[0124] The update formula for the connection weight between the hidden layer neuron and the output layer neuron is:

[0125]

[0126] where η is the preset learning rate, is the first gradient, is the connection weight between the hidden layer neuron and the output layer neuron before update, is the connection weight between the hidden layer neuron and the output layer neuron after update.

[0127] The update formula for the bias term of the output layer neuron is:

[0128]

[0129] where η is the preset learning rate, is the second gradient, b 2(old) is the bias term of the output layer neuron before update, b 2 (new) is the bias term of the output layer neuron after update.

[0130] The update formula for the connection weight between the input layer neuron and the hidden layer neuron is:

[0131]

[0132] where η is the preset learning rate, is the third gradient, is the connection weight between the input layer neuron and the hidden layer neuron before update, is the connection weight between the input layer neuron and the hidden layer neuron after update.

[0133] The update formula for the bias term of the hidden layer neuron is:

[0134]

[0135] where η is the preset learning rate, is the fourth gradient, is the bias term of the hidden layer neuron before update, is the bias term of the hidden layer neuron after update.

[0136] It should be noted that the learning rate described in the embodiments of the present application controls the step size of each weight update. If the learning rate is too large, it may cause the neural network model to fail to converge or even diverge; if the learning rate is too small, the training speed of the neural network model will be too slow. To balance the relationship between the model training speed and stability, the present application can select an appropriate learning rate value according to specific requirements.

[0137] The protection scope of the furnace glass liquid viscosity control method described in the embodiments of the present application is not limited to the execution order of the steps listed in this embodiment. Any solution achieved by adding or reducing steps of the prior art and replacing steps according to the principle of the present application is included in the protection scope of the present application.

[0138] Please refer to Figure 9 , which shows the structural schematic diagram of the furnace glass liquid viscosity control device described in the present application in an embodiment.

[0139] As Figure 9 shown, the embodiments of the present application provide a furnace glass liquid viscosity control device, including a real-time data acquisition module, a glass liquid viscosity prediction module, a control signal generation module, and a glass liquid viscosity adjustment module.

[0140] The real-time data acquisition module is used to acquire the real-time operation state data of the furnace; the real-time operation state data of the furnace includes the liquid level of the glass liquid in the furnace, the feeding speed of the feeder, the temperature in the furnace, and the pressure in the furnace collected at the current moment.

[0141] The glass liquid viscosity prediction module is used to input the real-time operation state data of the furnace into a pre-trained neural network model and output the real-time glass liquid viscosity.

[0142] The control signal generation module is used to generate a control signal based on the real-time glass liquid viscosity.

[0143] The glass liquid viscosity adjustment module is used to automatically adjust the real-time glass liquid viscosity to the target viscosity range based on the control signal.

[0144] It should be noted that the structures and principles of the real-time data acquisition module, the glass liquid viscosity prediction module, the control signal generation module, and the glass liquid viscosity adjustment module correspond one by one to the steps in the above-mentioned furnace glass liquid viscosity control method, so they will not be elaborated here.

[0145] The furnace glass liquid viscosity control device provided by the embodiments of the present application can implement the furnace glass liquid viscosity control method described in the present application. However, the implementation devices of the furnace glass liquid viscosity control method described in the present application include but are not limited to the structures of the furnace glass liquid viscosity control devices listed in this embodiment. Any structural deformation and replacement of the prior art made according to the principle of the present application are included in the protection scope of the present application.

[0146] An embodiment of the present application provides a training method, including the following steps S100 to S400.

[0147] In step S100, obtain the historical operating state data of the kiln as training samples; the historical operating state data includes the liquid level of the molten glass in the kiln, the feeding speed of the feeder, the temperature in the kiln, and the pressure in the kiln collected at historical moments.

[0148] In step S200, initialize the parameters of the neural network model to be trained to obtain the initialized neural network model.

[0149] In step S300, based on the initialized neural network model, perform forward propagation calculation on the training samples to obtain the predicted value of the molten glass viscosity, and calculate the prediction error of the predicted value relative to the true value.

[0150] In step S400, with the goal of minimizing the prediction error, use the gradient descent method to iteratively optimize the parameters of the initialized neural network model to obtain the trained neural network model.

[0151] It should be noted that the embodiments of steps S100 to S400 are the same as those described in steps S21 to S24, and will not be repeated here.

[0152] The protection scope of the training method described in the embodiments of the present application is not limited to the execution order of the steps listed in this embodiment. Any solution achieved by adding or subtracting steps of the prior art or replacing steps according to the principles of the present application is included in the protection scope of the present application.

[0153] Please refer to Figure 10 , which shows the structural schematic diagram of the terminal in an embodiment of the present application.

[0154] As Figure 10 shown, an embodiment of the present application provides a terminal, including: a processor and a memory.

[0155] The memory is used to store computer programs.

[0156] The processor is used to execute the computer programs stored in the memory so that the terminal executes the above-mentioned molten glass viscosity control method and / or training method of the kiln.

[0157] Preferably, the processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The memory may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, a magnetic disk, or an optical disc.

[0158] In several embodiments provided in the present application, it should be understood that the disclosed system, apparatus, or method may be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of modules / units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple modules or units may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces, and the indirect coupling or communication connection of devices, modules, or units may be in an electrical, mechanical, or other form.

[0159] The modules / units described as separate components may or may not be physically separated. The components shown as modules / units may or may not be physical modules, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules / units can be selected according to actual needs to achieve the objectives of the embodiments of the present application. For example, in various embodiments of the present application, each functional module / unit can be integrated in a processing module, or each module / unit can exist physically alone, or two or more modules / units can be integrated in one module / unit.

[0160] Those of ordinary skill in the art should further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. A professional technician can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present application.

[0161] The embodiments of the present application also provide a computer-readable storage medium. Those of ordinary skill in the art can understand that all or part of the steps in the methods of the above embodiments can be completed by instructing a processor through a program. The program can be stored in a computer-readable storage medium. The storage medium is a non-transitory medium, such as random access memory, read-only memory, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disc, and any combination thereof. The above storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a digital video disc (DVD)), or a semiconductor medium (such as a solid-state disk (SSD)).

[0162] An embodiment of the present application may further provide a computer program product, which includes one or more computer instructions. When the computer instructions are loaded and executed on a computing device, they generate, in whole or in part, the processes or functions described in the embodiments of the present application. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another website, computer, or data center via wired (such as coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) means.

[0163] When the computer program product is executed by a computer, the computer executes the method described in the foregoing method embodiments. The computer program product may be a software installation package. In the case where the foregoing method is needed, the computer program product may be downloaded and executed on the computer.

[0164] The descriptions of the processes or structures corresponding to the above respective drawings each have their own focuses. For parts not detailed in a certain process or structure, reference may be made to the relevant descriptions of other processes or structures.

[0165] The above embodiments are only illustrative of the principles and effects of the present application and are not intended to limit the present application. Any person familiar with this technology may make modifications or changes to the above embodiments without departing from the spirit and scope of the present application. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed in the present application should still be covered by the claims of the present application.

Claims

1. A training method, characterized in that: include: Obtain the historical operating status data of the kiln as training samples; The historical operation status data includes the liquid level of the glass liquid in the kiln, the feeding speed of the feeder, the temperature in the kiln and the pressure in the kiln collected at the historical moment; Initialize the parameters of the neural network model to be trained to obtain the initialized neural network model; Based on the initialized neural network model, forward propagation calculation is performed on the training sample to obtain a predicted value of the glass liquid viscosity, and a prediction error of the predicted value relative to a true value is calculated; With the goal of minimizing the prediction error, the parameters of the initialized neural network model are iteratively optimized using the gradient descent method to obtain a trained neural network model.

2. The method according to claim 1, characterized in that Initialize the parameters of the neural network model to be trained, and obtain the initialized neural network model including: Obtaining the number of input layer neurons, hidden layer neurons and output layer neurons of the neural network model to be trained; Using a neural network weight initialization method, the connection weights between the input layer neurons and the hidden layer neurons, and the connection weights between the hidden layer neurons and the output layer neurons are initialized; A random initialization method is adopted to initialize the bias items of the hidden layer neurons and the bias items of the output layer neurons.

3. The method according to claim 2, characterized in that The number of neurons in the hidden layer is determined by the number of neurons in the input layer and the number of neurons in the output layer; the calculation formula for the number of neurons in the hidden layer is: Wherein h is the number of neurons in the hidden layer, m is the number of neurons in the input layer, n is the number of neurons in the output layer, and a is an arbitrary constant in the value range of [1,10].

4. The method according to claim 2, characterized in that: Based on the initialized neural network model, forward propagation calculation is performed on the training sample to obtain the predicted value of the glass liquid viscosity, including: Based on the training samples, the connection weights between the input layer neurons and the hidden layer neurons, and the bias terms of the hidden layer neurons, the outputs of different hidden layer neurons in the initialized neural network model are calculated, and the outputs of the hidden layer neurons are combined to obtain a hidden layer output vector; Based on the hidden layer output vector, the connection weights between the hidden layer neurons and the output layer neurons, and the bias items of the output layer neurons, the output of the output layer neurons is calculated to obtain a predicted value of the glass liquid viscosity.

5. The method according to claim 2, characterized in that: With the goal of minimizing the prediction error, the parameters of the initialized neural network model are iteratively optimized using the gradient descent method to obtain a trained neural network model including: Determining whether the prediction error converges; If yes, end the training of the neural network model; Otherwise, the gradient of the prediction error with respect to the connection weight between the hidden layer neurons and the output layer neurons is calculated to obtain a first gradient; the gradient of the prediction error with respect to the bias term of the output layer neurons is calculated to obtain a second gradient; based on the first gradient and a preset learning rate, the connection weight between the hidden layer neurons and the output layer neurons is updated; based on the second gradient and a preset learning rate, the bias term of the output layer neurons is updated; Calculate the gradient of the prediction error with respect to the connection weight between the input layer neurons and the hidden layer neurons to obtain a third gradient; calculate the gradient of the prediction error with respect to the bias term of the hidden layer neurons to obtain a fourth gradient; based on the third gradient and a preset learning rate, update the connection weight between the input layer neurons and the hidden layer neurons; based on the fourth gradient and a preset learning rate, update the bias term of the hidden layer neurons to obtain an updated neural network model; Based on the updated neural network model, a new prediction error is calculated; Repeat the above-mentioned prediction error convergence judgment and neural network model update process until the prediction error converges.

6. The method according to claim 1, characterized in that Before performing forward propagation calculation on the training samples, the training samples are preprocessed, wherein the preprocessing step at least includes: Using a normalization method to uniformly map the training samples to a preset interval; The training samples are denoised using a moving average filtering method.

7. A method for controlling the viscosity of furnace glass liquid, characterized in that: include: Acquire real-time operation status data of the kiln; the real-time operation status data of the kiln includes the liquid level of the glass liquid in the kiln, the feeding speed of the feeder, the temperature in the kiln and the pressure in the kiln collected at the current moment; Inputting the real-time operating status data of the kiln into a pre-trained neural network model to output the real-time glass liquid viscosity; Based on the real-time glass liquid viscosity, generating a control signal; Based on the control signal, the real-time glass liquid viscosity is automatically adjusted to a target viscosity range.

8. The method according to claim 7, characterized in that The training method of the neural network model includes: Acquire historical operation status data of the kiln as training samples; the historical operation status data includes the liquid level of the glass liquid in the kiln, the feeding speed of the feeder, the temperature in the kiln and the pressure in the kiln collected at historical moments; Initialize the parameters of the neural network model to be trained to obtain the initialized neural network model; Based on the initialized neural network model, forward propagation calculation is performed on the training sample to obtain a predicted value of the glass liquid viscosity, and a prediction error of the predicted value relative to a true value is calculated; With the goal of minimizing the prediction error, the parameters of the initialized neural network model are iteratively optimized using the gradient descent method to obtain a trained neural network model.

9. A furnace glass liquid viscosity control device, characterized in that: include: A real-time data acquisition module is used to acquire real-time operation status data of the kiln; the real-time operation status data of the kiln includes the liquid level of the glass liquid in the kiln, the feeding speed of the feeder, the temperature in the kiln and the pressure in the kiln collected at the current moment; A glass liquid viscosity prediction module, used for inputting the real-time operating status data of the kiln into a pre-trained neural network model and outputting the real-time glass liquid viscosity; A control signal generating module, used for generating a control signal based on the real-time glass liquid viscosity; The glass liquid viscosity adjustment module is used to automatically adjust the real-time glass liquid viscosity to a target viscosity range based on the control signal.

10. A terminal, characterized in that: include: Processor and memory; The memory is used to store computer programs; The processor is used to execute the computer program stored in the memory so that the terminal executes the training method described in any one of claims 1 to 6 and / or the furnace glass liquid viscosity control method described in any one of claims 7 to 8.

Images