Quartz glass non-contact tube drawing process parameter optimization method based on deep learning

Abstract

The invention discloses a quartz glass non-contact tube drawing process parameter optimization method based on deep learning, which comprises the following steps: 1) in the continuous forming process of a glass tube, collecting process parameters and forming quality indexes in real time; 2) performing real-time imaging on the surface and the cross section of the glass tube; 3) taking the image obtained in the step 2) and the corresponding process parameter combination as input, and performing defect detection and classification based on a deep learning detection model; 4) obtaining a forming quality cost function according to the geometric accuracy index and the defect risk index, and evaluating the forming quality under different process parameters; 5) inputting the comprehensive quality score and the process parameters of the corresponding batch into a Bayesian optimization model to obtain an optimal process parameter combination, and 6) adjusting the key control variables of the non-contact tube drawing device according to the optimal process parameter combination output by Bayesian optimization.

Description

TECHNICAL FIELD

[0001] The application relates to an optical material processing technology, in particular to a quartz glass non-contact tube drawing process parameter optimization method based on deep learning. BACKGROUND

[0002] Quartz glass is widely used in the fields of semiconductor manufacturing, optical fiber communication, optical devices, aerospace, special lighting and the like due to its excellent optical transmittance, high temperature stability, chemical corrosion resistance and low thermal expansion coefficient. The production quality of quartz glass tube directly affects the performance and reliability of the tube in downstream high-end applications. As a forming method for reducing mechanical contact and surface defect risk, the non-contact tube drawing process can realize continuous and stable glass tube production through induction heating, rotation traction and internal pressure control.

[0003] In the existing production process, the process parameters mainly include heating temperature distribution, feeding speed, traction rotation speed, tube internal pressure and the like. Small fluctuations of these parameters will cause problems such as tube outer diameter deviation, roundness reduction and uneven wall thickness. In addition, defects such as bubbles, inclusions, surface scratches, corrugated deformation and micro-cracks are prone to occur in the production process, which not only affect the appearance quality of the glass tube, but also may cause failure in subsequent processing and application.

[0004] At present, the quality detection in the industry is mainly offline sampling inspection by manual work and partial single-point sensor monitoring, and there are problems such as detection lag, low coverage rate and insufficient defect identification precision, which makes it difficult to reflect the quality fluctuation in the production process in time. At the same time, the existing process optimization often relies on the experience trial and error method, which has a long adjustment period, high cost and lacks a multi-parameter collaborative optimization mechanism based on full data. Especially under the condition of coexistence of multiple indexes (geometric precision, defect rate and production efficiency) and mutual association, the traditional optimization method is difficult to effectively reduce the defect rate while ensuring the size precision.

[0005] Therefore, a closed-loop control method combining online full-process data acquisition, automatic defect analysis and intelligent process parameter optimization is needed to realize real-time optimization and stable control of the quartz glass non-contact tube drawing process, so as to improve the size consistency of the tube and reduce the defect rate, and shorten the process groping period. SUMMARY

[0006] The technical problem to be solved by the application is to provide a quartz glass non-contact tube drawing process parameter optimization method based on deep learning in view of the defects in the prior art.

[0007] The technical scheme adopted by the application to solve the technical problem is: a quartz glass non-contact tube drawing process parameter optimization method based on deep learning, comprising the following steps: 1) In the continuous forming process of glass tube, the process parameters and forming quality indicators are collected in real time to obtain the process parameter combination X = (T(y), P(t), v feed (t), v draw (t), ω(t), p in (t), q cool (t)) varying with time in the continuous forming process of glass tube. Wherein, T(y) is the heating temperature distribution, P(t) is the heating zone power, v feed (t) is the feeding speed, v draw (t) is the pulling speed, ω(t) is the rotation speed, p in (t) is the tube internal gas pressure, q cool (t) is the cooling rate; The forming geometric precision indicators include the roundness e of the glass tube, the wall thickness inconsistency SID, the cross-sectional area CSA, the wall thickness range and the length direction uniformity. 2) Real-time imaging of the glass tube surface and cross section to obtain the geometric and surface information of the formed tube at the corresponding time; 3) Taking the image obtained in step 2) and the process parameter combination at the corresponding time as input, the deep learning detection model based on convolutional neural network CNN detects and classifies the image; The detection model outputs the defect category probability , uncertainty evaluation and predicted wall thickness distribution and predicted ovality , and further generates a geometric defect intensity risk indicator; The amplitude of the axial corrugation main frequency is extracted from the obtained image as a surface periodic geometric defect intensity risk indicator; 4) Obtain the forming quality cost function according to the geometric precision indicators and defect risk indicators to evaluate the forming quality under different process parameters; Geometric accuracy:

[0008] Defect risk:

[0009] Forming quality cost function : ; Wherein, , , are weight coefficients; and are weight coefficients; is the defect risk weight coefficient; , , and are normalized values of average ellipticity , average wall thickness inconsistency , average waviness amplitude and average category defect intensity respectively; wherein average waviness amplitude is obtained by defect intensity risk indicator; average ellipticity , average wall thickness inconsistency , average waviness amplitude are obtained by averaging model output values; is a confidence function; 5) input the comprehensive quality score and corresponding batch of process parameters into a physical prior driven Bayesian optimization model, the Bayesian optimization constructs an optimization function by minimizing the forming quality cost function and setting constraints, and obtains the optimal process parameter combination; 6) according to the optimal process parameter combination output by the Bayesian optimization, adjust the key control variables of the non-contact pipe drawing device (adjust the key control variables of the non-contact pipe drawing device such as feeding speed, induction heating temperature and power distribution, traction rotation speed, pipe internal gas pressure and cooling rate), and execute the optimized parameters in the next batch of production.

[0010] According to the above scheme, in the step 1), the collected process parameter data is denoised to remove the noise of short-term disturbance in the process parameter combination.

[0011] According to the above scheme, in the step 4), when average ellipticity and average wall thickness inconsistency values are is the average defect risk, which is represented as follows: ; wherein is the average uncertainty entropy, which is obtained by historical data; is an amplification coefficient of uncertainty penalty on defects.

[0012] According to the above scheme, in the step 4), is the average risk of the worst alpha proportion sample, which is represented as follows:

[0013] wherein is a threshold variable, which corresponds to the optimal proxy of value at risk of quantile ; the optimal is the tail expectation of the part , represents the excess loss of the part over the threshold .

[0014] According to the above scheme, in the Bayesian optimization in step 5), the forming quality cost function further considers the production efficiency, =

[0015] wherein, is a weight coefficient, represents the production efficiency, which is obtained by weighting the unit energy consumption and the qualified rate.

[0016] According to the above scheme, the constraint conditions in step 5) include size error limit value constraints and defect probability threshold constraints.

[0017] According to the above scheme, the defect probability threshold constraint is:

[0018] ; ; ; ; ; wherein, represents that the temperature gradient does not exceed the maximum value; represents that the equivalent stress is less than the safe yield stress of the material; represents that the ovality should be controlled within the allowable range to ensure geometric stability; represents that the wall thickness inconsistency is less than the maximum allowable value; represents that the defect strength should be controlled within the safe range.

[0019] The beneficial effects of the present application are: The present application realizes the self-adaptive adjustment of process parameters and the optimization of product quality by establishing an intelligent closed loop of "process data acquisition-defect image analysis-quality scoring and optimization-parameter iteration and closed loop control". BRIEF DESCRIPTION OF DRAWINGS

[0020] The present application will be further described below in conjunction with the drawings and examples, wherein: Figure 1 is a method flowchart of an embodiment of the present application. DETAILED DESCRIPTION ​

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] like Figure 1 As shown, a method for optimizing the process parameters of non-contact quartz glass tube drawing based on deep learning includes the following steps: 1) During the continuous forming process of glass tubes, process parameters and forming quality indicators are collected in real time to obtain the combination of process parameters X = (T(y), P(t), v) as a function of time during the continuous forming process of glass tubes. feed (t), v draw (t), ω(t), p in (t), q cool (t) Where T(y) is the heating temperature distribution, P(t) is the heating zone power, and v feed (t) represents the feed rate, v draw (t) is the traction speed, ω(t) is the rotational speed, and p in (t) represents the air pressure inside the pipe, q cool (t) represents the cooling rate; The temperature distribution T(y) of the heating material is obtained by deploying fiber optic temperature sensors and infrared thermometers to show the temperature variation law of the molten zone along the height direction y. The heating zone power P(t) represents the change of input power in each power segment of the induction heating furnace over time. Feed speed v feed (t) represents the feeding rate of the quartz billet in the feeding mechanism; Traction speed v draw (t) represents the speed at which the glass tube is pulled out; The rotational speed ω(t) is the rotational angular velocity applied by the traction device; The air pressure inside the pipe, p in (t) represents the internal cavity pressure obtained through the barometric pressure sensor; Cooling rate q cool (t) represents the rate at which the cooling airflow or coolant cools the glass tube; Ambient temperature and humidity are external disturbance conditions in the production environment; The forming geometric accuracy indicators include the roundness e(t) of the glass tube, the wall thickness inconsistency SID, the cross-sectional area CSA, the wall thickness range, and the uniformity in the length direction. 2) Denoise the process data; By removing the noise of short-term disturbance in the process parameter combination, it ensures the efficient support of data to the model training and optimization process, and enhances the perception ability of the model to long-term trend and periodic change; In order to distinguish the influence of short-term disturbance and long-term drift on process stability, a delay autocorrelation function with exponential decay is defined. This method not only considers the correlation of signals at different time delays, but also introduces a decay time constant The influence of long-distance delay is suppressed, so that short-period fluctuations and long-period drifts can be separated and analyzed.

[0023]

[0024] Among them: The time series of a certain process parameter at time t; : The mean value of the sequence.

[0025] : The variance of the sequence.

[0026] : Time delay step.

[0027] : Mathematical expectation operation, used to calculate the autocorrelation function.

[0028] : Decay time constant, determines the weight of long-distance correlation. Small Emphasize short-term disturbance, large Emphasize long-term drift.

[0029] ): Multiscale delay autocorrelation function, used to separate and analyze "short-period disturbance" and "long-period drift".

[0030] By adjusting , long-term drift (large ) can be emphasized in analysis; 3) Real-time imaging of glass tube surface and cross section to obtain geometric and surface information of the formed pipe at the corresponding time; 4) Defect detection and classification of images based on deep learning model of convolutional neural network CNN; 3) The images obtained in step 2) and the process parameter combination at the corresponding time are input, and the deep learning detection model based on convolutional neural network CNN is used for defect detection and classification of images; The detection model is a pre-trained model. Through deep feature extraction and multi-task learning, the detection model outputs defect category probability , uncertainty evaluation and predicted wall thickness distribution and predict the ellipticity and further generate a geometric defect intensity risk indicator; extract the amplitude of the axial waviness dominant frequency from the acquired image as the surface periodic geometric defect intensity risk indicator; unfold the glass surface image into a two-dimensional grayscale image; perform a Fourier transform (FFT) along the axial (longitudinal) direction; the dominant frequency amplitude is the amplitude corresponding to the dominant frequency in the energy spectrum, reflecting the intensity of the surface periodic defect.

[0031] extract the axial waviness dominant frequency amplitude from the image through Fourier spectrum analysis to represent the intensity of the surface periodic defect; 5) obtain a forming quality cost function according to the geometric accuracy indicator and the defect risk indicator to evaluate the forming quality under different process parameters; Geometric accuracy:

[0032] Defect risk:

[0033] Forming quality cost function : ; wherein, , , are weight coefficients; and are weight coefficients; is a defect risk weight coefficient; , , and are the normalized values of the average ellipticity , the average wall thickness inconsistency , the average waviness amplitude and the average category defect intensity ; the average ellipticity , the average wall thickness inconsistency , the average waviness amplitude are obtained by averaging the model output values; wherein, the average waviness amplitude is obtained through the defect intensity risk indicator; ; ; ; ; is:

[0034] wherein, is the average uncertainty entropy, obtained from historical data; is the amplification coefficient of uncertainty penalty on defects; is the average risk of the worst alpha proportion sample, =0.9~0.99 is commonly used to emphasize the "worst 10%~1%" quality section; is the tail risk

[0035]

[0036] =0.9~0.99 is commonly used to emphasize the "worst 10%~1%" quality section; is the confidence function; 6) The comprehensive quality score and the corresponding batch process parameters are input into the physical prior driven Bayesian optimization model. Bayesian optimization can efficiently search for the optimal process parameter combination in the global range by constructing an optimization function under multi-objective (geometric accuracy, defect rate, production efficiency) and multi-constraint conditions (size error limit value, defect probability threshold, energy consumption limit, etc.); The optimization function is min ; In this embodiment, the forming quality cost function also considers the production efficiency, =

[0037] wherein, is the weight coefficient, represents the production efficiency, which is obtained by weighting the unit energy consumption and the qualified rate.

[0038] The defect probability threshold constraint is:

[0039] ; ; ; ; ; wherein, represents that the temperature gradient should not exceed the maximum value, avoiding problems caused by uneven temperature; Indicates that the equivalent stress should be less than the safe yield stress of the material to prevent material failure. Indicates that the ovality should be controlled within the allowable range to ensure geometric stability. Indicates that the wall thickness inconsistency should be less than the maximum allowable value to ensure product uniformity. Indicates that the defect intensity should be controlled within the safe range. Indicates the energy consumption limit. is the energy consumption per unit length, is the sampled maximum value of energy consumption per unit length.

[0040] 7) Adjust the key control variables of the non-contact pipe drawing device according to the optimal combination of process parameters output by the Bayesian optimization, and execute the optimized parameters in the next batch of production. It should be understood that those of ordinary skill in the art can make improvements or changes according to the above description, and all these improvements and changes shall belong to the protection scope of the appended claims of the present application.

Claims

1. A deep learning-based method for optimizing parameters of a contactless tube-pulling process of quartz glass, characterized in that, The method comprises the following steps: 1) During the continuous forming process of glass tubes, process parameters and forming quality indicators are collected in real time to obtain the combination of process parameters X = (T(y), P(t), v) as a function of time during the continuous forming process of glass tubes. feed (t), v draw (t), ω(t), p in (t), q cool (t) where T(y) is the heating temperature profile, P(t) is the heating zone power, v feed (t) is the feed speed, v draw (t) is the pull speed, ω(t) is the rotational speed, p in (t) is the gas pressure inside the tube, q cool (t) is the cooling rate; The forming geometric accuracy index comprises roundness e, wall thickness inconsistency SID, cross-sectional area CSA, wall thickness range, and length direction uniformity of the glass tube; 2) Real-time imaging is performed on the surface and cross section of the glass tube to obtain geometric and surface information of the forming tube material at the corresponding moment; 3) The image obtained in step 2) and the corresponding process parameter combination are taken as inputs, and a defect detection and classification is performed based on a deep learning detection model; Detecting model output defect class probabilities , uncertainty assessment and predicted wall thickness distribution and predicted ovality and further generating a geometric defect strength risk indicator; The amplitude of the axial corrugation main frequency is extracted from the obtained image as a surface periodic geometric defect intensity risk index; 4) A forming quality cost function is obtained according to the geometric accuracy index and the defect risk index to evaluate the forming quality under different process parameters; Geometric accuracy: Defect risk: molding quality cost function : ； wherein, , , is a weight coefficient; and is a weight coefficient; is a defect risk weight coefficient; , , and are normalized values of the average ellipticity , average wall thickness inconsistency , average waviness amplitude and average class defect intensity , respectively; wherein the average waviness amplitude obtained by the defect intensity risk indicator; average ovality average wall thickness inconsistency average waviness amplitude obtained by averaging the model output values; is a confidence function; 5) input the comprehensive quality score and the corresponding batch process parameters into the Bayesian optimization model, and the Bayesian optimization model minimizes the molding quality cost function and sets constraints to build an optimization function to obtain the optimal process parameter combination; 6) According to the optimal process parameter combination output by the Bayesian optimization, the key control variables of the non-contact pipe drawing device are adjusted, and the optimized parameters are executed in the next batch of production.

2. The deep learning based quartz glass contactless tube-pulling process parameter optimization method according to claim 1, characterized in that, In step 1), the collected process parameter data is also denoised to remove the noise of short-term disturbance in the process parameter combination.

3. The deep learning based quartz glass contactless tube-pulling process parameter optimization method according to claim 1, characterized in that, In step 4), the forming quality cost function is as follows: Geometric accuracy: Defect risk: molding quality cost function : ； wherein, , , is a weight coefficient; and is a weight coefficient; is a defect risk weight coefficient; , , and are normalized values of the average ellipticity , average wall thickness inconsistency , average waviness amplitude and average class defect intensity , respectively; wherein the average waviness amplitude obtained by the defect intensity risk indicator; average ovality average wall thickness inconsistency average waviness amplitude obtained by averaging the model output values; is a confidence function; is the average risk for the worst alpha proportion sample; The average defect risk is 0.

4. The deep learning-based quartz glass contactless tube-pulling process parameter optimization method according to claim 3, characterized in that, In the step 4), the average ellipticity and the average wall thickness inconsistency when the value is For the average defect risk, this is expressed as follows: ； wherein, is the average uncertainty entropy, obtained from historical data; is the amplification factor of the uncertainty penalty on defects.

5. The deep learning based quartz glass contactless tube-pulling process parameter optimization method according to claim 3, characterized in that, In step 4) above, The average risk for the worst alpha proportion sample is denoted as follows: wherein, is a threshold variable corresponding to the quantile value at risk of the optimized agent; the optimal is the value at , represents the tail expectation beyond the threshold part, i.e. > the average of the excess loss.

6. The deep learning based quartz glass contactless tube-pulling process parameter optimization method according to claim 3, characterized in that, In the Bayesian optimization in step 5), the forming quality cost function also considers the production efficiency, = wherein, is a weight coefficient, represents production efficiency, obtained by weighting unit energy consumption and pass rate.

7. The deep learning-based quartz glass contactless tube-pulling process parameter optimization method according to claim 1, wherein The constraint conditions in step 5) include size error limit value constraints and defect probability threshold constraints.

8. The deep learning based quartz glass contactless tube-pulling process parameter optimization method according to claim 1, wherein, The defect probability threshold constraint is: The temperature gradient does not exceed the maximum value; the equivalent stress is less than the safe yield stress of the material; the ovality is controlled within the allowable range; the wall thickness inconsistency is less than the maximum allowable value; and the defect intensity should be controlled within the safe range. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the method of any one of claims 1 to 8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the method of any one of claims 1 to 8.

Images