Method, device and equipment for adjusting process parameters of chemical vapor deposition method

Abstract

The invention discloses a process parameter adjusting method, device and equipment for a chemical vapor deposition method, and relates to the technical field of chemical vapor deposition, the method comprises the following steps: executing at least one feature extraction operation to extract state features of a deposition body in a chemical vapor deposition process until the extracted state features exceed a preset range, the feature extraction operation comprises the steps that according to a target image, the state feature of the sedimentary body is automatically recognized, and the target image is an image obtained by automatically shooting the surface of the sedimentary body through an industrial camera system; if the state characteristics of the sedimentary body exceed the preset range, according to the target process parameters and the state characteristics of the sedimentary body exceeding the preset range, automatic process parameter adjustment operation is executed, the target process parameters are process parameters synchronously collected with the target image, and after the automatic process parameter adjustment operation is executed, the target image is obtained. And returning to execute at least one feature extraction operation and subsequent steps. According to the method, the quality of the prepared quartz glass is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of chemical vapor deposition, in particular to a process parameter adjustment method, device and equipment of a chemical vapor deposition method. BACKGROUND

[0002] Chemical vapor deposition (CVD) process is one of the important methods for preparing high-quality quartz glass. Through the chemical reaction of gaseous initial compounds on the surface of the substrate, solid quartz glass is deposited on the substrate. The process can accurately control the composition and structure of quartz glass, and prepare high-purity and low-defect quartz glass products. Therefore, it occupies an important position in the field of high-end quartz glass manufacturing.

[0003] However, in the process of preparing quartz glass by traditional chemical vapor deposition, process control mainly depends on the experience of operators. The operator needs to manually collect deposition process parameters, and also needs to observe the state of the deposition body by naked eye. When the quality of the deposition body changes, the operator needs to quickly judge and adjust the process parameters. However, in actual situations, manual operation has the following problems, which cannot be solved in time, seriously affecting the quality of the prepared quartz glass, resulting in problems such as bubbles, impurities, poor diameter of quartz glass, etc.: First, manual recording of process parameters and observation of state characteristics of the deposition body have serious hysteresis, resulting in hysteresis in judging whether the quality of the deposition body is abnormal and giving process parameter adjustment measures when abnormal; Second, once the quality of the deposition body changes, it is often difficult for the operator to quickly give effective process parameter adjustment measures in a short time. SUMMARY

[0004] The purpose of the present application is to provide a process parameter adjustment method, device and equipment of a chemical vapor deposition method, which can improve the quality of the prepared quartz glass.

[0005] To achieve the above purpose, the present application provides the following solutions: In a first aspect, the present application provides a process parameter adjustment method of a chemical vapor deposition method, comprising: performing at least one feature extraction operation to extract state characteristics of a deposition body of a chemical vapor deposition process until the extracted state characteristics exceed a preset range; If the state characteristics of the deposition body exceed the preset range, performing a process parameter automatic adjustment operation according to the target process parameters and the state characteristics of the deposition body exceeding the preset range; wherein the target process parameters are process parameters of the chemical vapor deposition process collected by a data collection device; After performing the automatic adjustment operation of the process parameters, return to perform at least one feature extraction operation and subsequent steps; The feature extraction operation comprises: According to the target image of the current time step, automatically identify the state features of the deposition body of the chemical vapor deposition process; wherein the target image is an image obtained by automatically photographing the surface of the deposition body by an industrial camera system and synchronously collecting the target process parameters.

[0006] Optionally, the automatic identification of whether the state features of the deposition body of the chemical vapor deposition process exceed the preset range according to the target image specifically comprises: Perform data processing on the target image of the current time step to obtain the real values of the state features of the deposition body; Compare the real values of the state features of the deposition body with the corresponding preset ranges, and if the real value of any state feature of the deposition body is not within the corresponding preset range, it is considered that the state features of the deposition body exceed the preset range.

[0007] Optionally, the data processing on the target image of the current time step to obtain the real values of the state features of the deposition body specifically comprises: Input the target image of the current time step into the trained deep learning network, extract the state feature vector of the deposition body from the input target image through the deep learning network, and then identify the real values of the state features of the deposition body according to the state feature vector.

[0008] Optionally, if the state features of the deposition body exceed the preset range, then according to the target process parameters and the state features of the deposition body that exceed the preset range, perform an automatic adjustment operation of the process parameters, specifically comprising: If any state feature of the deposition body exceeds the corresponding preset range, input the target process parameters of the current time step into the deep learning network, extract the process parameter time series feature vector through the deep learning network, and splice the process parameter time series feature vector and the state feature vector to obtain a fusion feature vector; According to the fusion feature vector, perform an automatic adjustment operation of the process parameters.

[0009] Optionally, the deep learning network comprises: A convolutional neural network is used to extract the state feature vector of the deposition body from the input target image, and identify the state features of the deposition body according to the state feature vector; A recurrent neural network is used to extract the process parameter time series feature vector according to the input target process parameters of the current time step combined with the target process parameters of the historical time steps. A fully connected layer is used to concatenate the state feature vector with the time series feature vector of the process parameters to obtain the fused feature vector.

[0010] Optionally, the convolutional neural network includes an input layer, a first convolutional block, a second convolutional block, a third convolutional block, a global average pooling layer, and an output layer, wherein: The first convolutional block is used to extract the first feature of the target image input by the input layer; The second convolutional block is used to extract a second feature from the target image; The third convolutional block is used to extract the third feature of the target image; The global average pooling layer is used to perform global average pooling on the feature maps output by the first convolutional block, the second convolutional block, and the third convolutional block, and output the state feature vector of the deposition body. The output layer is used to perform a nonlinear transformation on the state feature vector through an activation function, and to identify the state features of the sediment based on the vector output by the activation function.

[0011] Optionally, the automatic adjustment of process parameters includes: Based on the fused feature vector, the process parameter adjustment strategy for the current time step is generated by optimizing the model; The process parameters of the chemical vapor deposition process are adjusted according to the process parameter adjustment strategy at the current time step.

[0012] Optionally, the optimization model employs a reinforcement learning model; The process parameter adjustment strategy for the current time step, generated by optimizing the model based on the fused feature vector, specifically includes: The input fused feature vector is used as the current time step of the reinforcement learning module. t status S t The reinforcement learning module is based on S t Generate and output actions The action That is, the current time step. t The process parameter adjustment strategy.

[0013] Secondly, this application provides a process parameter adjustment device for chemical vapor deposition, comprising: Industrial camera systems are used to capture images of the surface of deposits during chemical vapor deposition processes to obtain target images; A data acquisition device is used to acquire process parameters of the chemical vapor deposition process to obtain target process parameters; wherein, the target process parameters are acquired synchronously with the target image; a data processing unit configured to perform at least one feature extraction operation to extract a state feature of a deposition body of a chemical vapor deposition process until the extracted state feature is out of a preset range; if the state feature of the deposition body is out of the preset range, performing a process parameter automatic adjustment operation according to a target process parameter and the state feature of the deposition body out of the preset range; wherein the target process parameter is a process parameter of the chemical vapor deposition process collected by a data collection device; after performing the process parameter automatic adjustment operation, returning to perform the at least one feature extraction operation and the subsequent steps; wherein the feature extraction operation comprises: automatically identifying the state feature of the deposition body of the chemical vapor deposition process according to a target image of a current time step; wherein the target image is an image obtained by automatically photographing a surface of the deposition body by an industrial camera system and is synchronously collected with the target process parameter.

[0014] In a third aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the process parameter adjustment method of the chemical vapor deposition method according to any one of the above.

[0015] According to the embodiments provided in the present application, the following technical effects are disclosed: The present application provides a process parameter adjustment method, device and equipment of a chemical vapor deposition method, which realizes automatic collection of process parameters of a chemical vapor deposition process by setting a data collection device and an industrial camera system, automatically photographs a surface of a deposition body by setting an industrial camera system to obtain a target image, and automatically identifies a state feature of the deposition body according to the target image, without manually observing and recording process parameters and state features of the deposition body, thereby avoiding the lag of manual observation and recording of process parameters and state features of the deposition body.

[0016] Since the state feature of the deposition body of the chemical vapor deposition process is out of the preset range, which means that the quality of the deposition body is abnormal, by performing at least one feature extraction operation (automatically identifying whether the state feature of the deposition body of the chemical vapor deposition process is out of the preset range according to a target image), until the state feature of the deposition body of the chemical vapor deposition process is out of the preset range, automatic monitoring of the quality abnormality of the deposition body is realized, so that the quality abnormality of the deposition body can be found in time, and the lag of manual judgment of the quality abnormality of the deposition body according to the state feature of the deposition body is avoided.

[0017] Since the state feature of the deposited body within the preset range indicates that the quality of the deposited body under the adjusted process parameters meets the requirements, i.e., the automatically adjusted process parameters overcome the problems existing in the deposition process under the original process parameters, when the state feature of the deposited body is out of the preset range, the automatic process parameter adjustment operation is performed, and after the automatic process parameter adjustment operation is performed, the at least one feature extraction operation and the subsequent steps are returned to be executed, so that the automatic adjustment of the process parameters is realized when the quality of the deposited body is abnormal, the process parameter adjustment is ensured in time when the quality of the deposited body is abnormal, the problem that manual observation and recording of the state feature of the process parameters and the deposited body have serious hysteresis, resulting in the problem that the quality of the deposited body is judged to be abnormal and the process parameter adjustment measures are given with hysteresis, are solved, and the quality of the prepared quartz glass is preliminarily improved.

[0018] Since the process parameter adjustment measures do not need to be given manually, and when the quality of the deposited body is abnormal, the time required for the automatic adjustment of the process parameters by the application to the state feature of the deposited body within the preset range is much less than the time required for the manual giving of the process parameter adjustment measures, the effective process parameter adjustment measures are quickly given when the quality of the deposited body is abnormal, the problem that once the quality of the deposited body changes, workers often have difficulty in quickly giving effective process parameter adjustment measures in a short time is solved, and the quality of the prepared quartz glass is further improved. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0020] Figure 1 A flowchart of a process parameter adjustment method of a chemical vapor deposition method provided by an embodiment of the application; Figure 2 A schematic diagram of a target image in the embodiment; Figure 1 Figure 3 A schematic diagram of a chemical vapor deposition process provided by an embodiment of the application; Figure 4 A functional module schematic diagram of a process parameter adjustment device of a chemical vapor deposition provided by an embodiment of the application; Figure 5 A structural schematic diagram of a computer device provided by an embodiment of the application. DETAILED DESCRIPTION

[0021] ​With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0022] The above-mentioned purposes, features and advantages of the present application will be more apparent and understandable. The present application will be described in further detail below with reference to the drawings and specific embodiments.

[0023] In one exemplary embodiment, a method for adjusting process parameters of a chemical vapor deposition method is provided, as shown in Figure 1 exemplarily comprising the following steps 101 to 103. Wherein: Step 101, performing at least one feature extraction operation to extract the state feature of the deposition body of the chemical vapor deposition process until the extracted state feature exceeds the preset range; wherein the feature extraction operation comprises: According to the target image of the current time step, automatically identifying the state feature of the deposition body of the chemical vapor deposition process, wherein the target image is an image obtained by automatically photographing the surface of the deposition body by an industrial camera system.

[0024] In the embodiments of the present application, the state feature at least includes the surface curvature, diameter, surface position, surface temperature and internal defect feature of the deposition body, and the internal defect feature at least includes the number of internal defects, and the internal defects include bubbles. The surface position refers to the distance between the top end of the semicircular arc shape of the surface of the deposition body and the preset reference plane in the target image. As Figure 2 shown is an exemplary target image, wherein the purple horizontal line is the preset reference plane, and the blue line is the top end of the semicircular arc shape of the surface of the deposition body. The surface position of the deposition body refers to the position difference between the top end of the semicircular arc shape of the surface of the deposition body and the reference plane. In the target image, if the distance between the top end of the semicircular arc shape of the surface of the deposition body and the reference plane is within the preset range (such as 2mm), it is considered that the surface position of the deposition body is within the preset range, otherwise, it is considered that the surface position of the deposition body exceeds the preset range.

[0025] The preset range of the state feature of the deposition body is the state feature data range that can guarantee the quality of the deposition body. If the state feature of the deposition body exceeds the preset range, it means that there is a problem in the deposition process, and the quality of the deposition body changes. The preset range of the state feature of the deposition body can be measured by simulation or actual test.

[0026] The stability of the surface curvature and the diameter directly affects the optical performance of the deposited quartz glass, the surface position affects the distance between the quartz glass surface and the burner port, affects the surface temperature, and the surface temperature directly affects the degree of chemical reaction. If the surface temperature is too high, the raw materials will volatilize, and if the temperature is too low, the reaction will be incomplete, resulting in serious defects such as bubbles. Therefore, by limiting the state characteristics to at least include the surface curvature, diameter, surface position, surface temperature and internal defect quantity of the deposition body, the optical performance and quality of the deposited quartz glass can be ensured.

[0027] In step 102, if the state characteristics of the deposition body exceed the preset range, a process parameter automatic adjustment operation is performed according to the target process parameters of the current time step and the state characteristics of the deposition body that exceed the preset range; wherein the target process parameters are process parameters of the chemical vapor deposition process collected by the data collection device, and the target image and the target process parameters are collected synchronously.

[0028] In the embodiments of the present application, by setting the data collection device and the industrial camera system, the process parameters of the chemical vapor deposition process are automatically collected, the industrial camera system is set to automatically photograph the surface of the deposition body to obtain the target image, and the state characteristics of the deposition body are automatically identified according to the target image to determine whether the state characteristics exceed the preset range, without the need for manual observation and recording of the process parameters and the state characteristics of the deposition body, thereby avoiding the influence of the timeliness of the process parameter adjustment caused by the lag of manual observation and recording of the process parameters and the state characteristics of the deposition body, and thus the quality of the prepared quartz glass is preliminarily ensured.

[0029] The chemical vapor deposition equipment for preparing quartz glass by chemical vapor deposition includes a deposition furnace body and a weighing machine. The weighing machine is a motor device for supporting and rotating the deposition substrate, and has rotating and lifting functions. The weighing machine rotates and lifts to drive the deposition substrate (substrate) and the quartz glass deposition body thereon to rotate and lift together.

[0030] The collected process parameters of the chemical vapor deposition process at least include the flow rates of the process gases, the feeding amount, the weighing machine withdrawal speed, the furnace temperature, the deposition body weight and the tail gas exhaust amount. The process gases include precursor gases, carrier gases and fuel gases. The data collection device includes gas flow sensors for detecting the flow rates of the precursor gases, the carrier gases and the fuel gases. The gas flow sensors can be installed on the delivery pipelines of the precursor gases, the carrier gases and the fuel gases.

[0031] When preparing the quartz glass, for example, Figure 3As shown, the precursor gas can adopt silicon tetrachloride, the carrier gas can adopt oxygen, the fuel gas can adopt hydrogen and oxygen, and the generated silicon dioxide particles are gathered on the deposition substrate. At this time, a gas flow sensor is installed on the conveying pipeline of hydrogen, oxygen as the fuel gas, oxygen as the carrier gas, and silicon tetrachloride to detect the flow of hydrogen, oxygen as the fuel gas, oxygen as the carrier gas, and silicon tetrachloride.

[0032] The feeding amount is the flow rate of the raw material (silicon tetrachloride), which is obtained by a mass flow meter installed on the feeding cabinet and is in units of g / min. The feeding amount directly affects the growth rate of the quartz glass.

[0033] The draw-off speed of the weighing machine refers to the speed at which the weighing machine drives the deposition body to descend at a uniform speed, and the weighing machine remains rotating during the draw-off process to make the deposition process uniform. The weighing machine draws off according to the speed at which the deposition body grows in height, and the draw-off speed is as equal as possible to the speed at which the deposition body increases in height. The data acquisition device includes a speed sensor, which can be installed on the rotor of the motor equipment of the weighing machine to monitor the draw-off speed of the weighing machine.

[0034] The furnace temperature refers to the internal temperature of the furnace body of the chemical vapor deposition equipment. The data acquisition device further includes a temperature sensor for detecting the furnace temperature. The structure of the temperature sensor is not limited and can be set as required. For example, the temperature sensor adopts a thermocouple or an infrared temperature sensor.

[0035] The data acquisition device further includes a second weight sensor, and the weight of the deposition body can be detected by the second weight sensor installed below the deposition substrate. The weight of the deposition body is the difference between the detection value of the second weight sensor and the weight of the deposition substrate.

[0036] The data acquisition device further includes an air volume sensor, and the exhaust air volume is collected by the air volume sensor installed on the exhaust air pipeline.

[0037] In step 103, after the automatic adjustment of the process parameters is performed, the at least one feature extraction operation and the subsequent steps are returned to be executed.

[0038] In the embodiments of the present application, after the automatic adjustment of the process parameters is performed, the at least one feature extraction operation and the subsequent steps are returned to be executed until all state features of the deposition body are within the preset range, which means that a process parameter adjustment process is completed. When all state features of the deposition body are within the preset range, it means that the quality of the deposition body under the adjusted process parameters meets the quality requirements of the deposition body, that is, the automatic adjustment of the process parameters timely solves the problems in the deposition process and ensures the quality of the deposition body.

[0039] The steps 101-103 are implemented to automatically collect the process parameters of the chemical vapor deposition process by setting the data collection device and the industrial camera system, automatically capture the surface of the deposition body by setting the industrial camera system to obtain the target image, and automatically identify the state characteristics of the deposition body according to the target image, without manually observing and recording the process parameters and the state characteristics of the deposition body, thereby avoiding the lag of manually observing and recording the process parameters and the state characteristics of the deposition body.

[0040] Since the state characteristics of the deposition body of the chemical vapor deposition process exceeding the preset range indicates that the quality of the deposition body is abnormal, the automatic monitoring of the quality of the deposition body is achieved by performing at least one feature extraction operation (automatically identifying whether the state characteristics of the deposition body of the chemical vapor deposition process exceed the preset range according to the target image) until the state characteristics of the deposition body of the chemical vapor deposition process exceed the preset range, so that the quality of the deposition body can be found in time, and the lag of manually judging the quality of the deposition body is avoided.

[0041] Since the state characteristics of the deposition body within the preset range indicates that the quality of the deposition body under the adjusted process parameters meets the requirements, that is, the automatically adjusted process parameters overcome the problems in the deposition process under the original process parameters, when the state characteristics of the deposition body exceed the preset range, the process parameter automatic adjustment operation is performed, and after the process parameter automatic adjustment operation is performed, the at least one feature extraction operation and the subsequent steps are returned to be executed, so that the automatic adjustment of the process parameters is realized when the quality of the deposition body is abnormal, the process parameter adjustment is ensured in time when the quality of the deposition body is abnormal, and the problem of serious lag of manually observing and recording the process parameters and the state characteristics of the deposition body is solved, so that the quality of the prepared quartz glass is preliminarily improved.

[0042] Since the process parameter adjustment measures do not need to be given manually, and the time required for the automatic adjustment of the process parameters by the application when the quality of the deposition body is abnormal until the state characteristics of the deposition body are within the preset range is much less than the time required for the manual process parameter adjustment measures, effective process parameter adjustment measures are given quickly when the quality of the deposition body is abnormal, the problem that workers often have difficulty in giving effective process parameter adjustment measures quickly in a short time when the quality of the deposition body changes is solved, and the quality of the prepared quartz glass is further improved.

[0043] In another exemplary embodiment of the application, the step 101 of automatically identifying whether any state characteristics of the deposition body of the chemical vapor deposition process exceed the preset range according to the target image of the current time step includes the following steps 201-202. Wherein: Step 201, data processing is performed on the target image of the current time step to obtain real values of the state features of the deposit.

[0044] Step 202, the real values of the state features of the deposit are compared with the corresponding preset ranges, and if the real value of any state feature of the deposit is not within the corresponding preset range, it is considered that the state feature of the deposit exceeds the preset range.

[0045] In an embodiment of the present application, for each state feature of the deposit, a preset range is set, and if the real value of any state feature is not within the corresponding preset range, the real value of any state feature is not within the preset range of the state feature.

[0046] In another exemplary embodiment of the present application, the above step 201 specifically includes: Step 301, inputting the target image of the current time step into the trained deep learning network, extracting the state feature vector of the deposit from the input target image through the deep learning network, and then identifying the real values of the state features of the deposit according to the state feature vector.

[0047] In an embodiment of the present application, the state feature vector contains state feature information of the deposit.

[0048] In another exemplary embodiment of the present application, the above step 102 includes steps 401 to 402. Wherein: Step 401, if any state feature of the deposit exceeds the corresponding preset range, inputting the target process parameter of the current time step into the trained deep learning network, extracting the process parameter time sequence feature vector through the deep learning network, and splicing the process parameter time sequence feature vector and the state feature vector extracted in step 301 to obtain a fusion feature vector.

[0049] In an embodiment of the present application, the process parameter time sequence feature vector represents the variation law of each process parameter with time and the dynamic correlation between each process parameter. If the dimension of the state feature vector is 256 and the dimension of the process parameter time sequence feature vector is 32, the dimension of the fusion feature vector obtained by splicing is 288.

[0050] Step 402, performing process parameter automatic adjustment operation according to the fusion feature vector.

[0051] In another exemplary embodiment of the present application, the deep learning network includes a convolutional neural network, a recurrent neural network and a fully connected layer. Wherein: The convolutional neural network is used to extract the state feature vector of the deposit from the input target image, and identify the state features of the deposit according to the state feature vector; The recurrent neural network is used to extract a process parameter time sequence feature vector according to the target process parameter of a current time step and in combination with target process parameters of historical time steps; The full connection layer is used to splice the state feature vector and the process parameter time sequence feature vector to obtain a fusion feature vector.

[0052] In the embodiment of the application, when the recurrent neural network extracts the process parameter time sequence feature vector according to the target process parameter of the current time step and in combination with the target process parameters of the historical time steps, the target process parameter of the current time step and the target process parameters of the historical time steps can be used to form a process parameter sequence containing continuous n time steps, and the last element of the process parameter sequence is the target process parameter of the current time step. Then, the recurrent neural network generates and outputs the process parameter time sequence feature vector according to the process parameter sequence.

[0053] In another exemplary embodiment of the application, the convolutional neural network comprises an input layer, a first convolutional block, a second convolutional block, a third convolutional block, a global average pooling layer and an output layer, wherein: The first convolutional block is used to extract first features of the target image input by the input layer; The second convolutional block is used to extract second features of the target image; The third convolutional block is used to extract third features of the target image; The global average pooling layer is used to perform global average pooling on feature maps output by the first convolutional block, the second convolutional block and the third convolutional block, and output a state feature vector of the deposit; The output layer is used to perform nonlinear transformation on the state feature vector output by the global average pooling layer through an activation function (such as a ReLU activation function), and identify state features of the deposit according to the vector output by the activation function.

[0054] In the embodiment of the application, the use of three convolutional blocks can achieve an optimal balance between the number of parameters and the ability of feature expression. The first features can reflect the surface position, surface curvature, diameter and the like of the deposit in the deposit image. The second features can reflect the surface temperature of the deposit in the deposit image, and the third features can reflect the internal defect features and defect change law of the deposit in the deposit image.

[0055] In another exemplary embodiment of the application, the first convolutional block comprises a first convolutional layer and a first max-pooling layer connected in sequence, the second convolutional block comprises a second convolutional layer and a second max-pooling layer connected in sequence, and the third convolutional block comprises a third convolutional layer.

[0056] In the embodiment of the application, the first convolutional layer adopts 64 3x3 convolutional kernels, the activation function adopts the ReLU activation function, the first and second maximum pooling layers adopt 2x2 maximum pooling, the second convolutional layer adopts 128 3x3 convolutional kernels, the third convolutional layer adopts 256 3x3 convolutional kernels, and the first, second and third features are accurately extracted.

[0057] The size of the target image is 224x224 pixels, the size of the feature map output by the first maximum pooling layer is 112x112x64, the size of the feature map output by the second maximum pooling layer is 56x56x128, and the size of the feature map output by the third convolutional layer is 56x56x256. The global average pooling layer outputs a one-dimensional vector of 256 dimensions.

[0058] In another exemplary embodiment of the application, the recurrent neural network adopts a long short-term memory network LSTM.

[0059] In another exemplary embodiment of the application, the input of the recurrent neural network is X t =[ Q t , T t , W t , S t , C t , P t ], Q t is the flow of the process gas at the time step t , T t is the furnace temperature at the time step t , W t is the weight of the deposited body at the time step t , t is the feeding amount at the time step t , C t is the unloading speed of the weigher at the time step t , P t is the exhaust air volume at the time step t .

[0060] In another exemplary embodiment of the application, the process parameter automatic adjustment operation of the above step 102 includes the following steps 501 to 502. Wherein: Step 501, generating a process parameter adjustment strategy for the current time step through an optimization model according to the fused feature vector.

[0061] In the embodiments of the present application, the structure of the optimization model is not specifically limited, and can be selected according to actual needs, as long as the process parameter adjustment strategy of the current time step can be generated according to the target process parameters and the state characteristics of the deposited body, so that an effective process parameter adjustment strategy is given in the shortest possible time, and then all the state characteristics of the deposited body are adjusted to the preset range, thereby improving the quality of the prepared quartz glass.

[0062] In step 502, the process parameters of the chemical vapor deposition process are adjusted according to the process parameter adjustment strategy of the current time step.

[0063] In the embodiments of the present application, the process parameters of the chemical vapor deposition process can be adjusted manually according to the optimal process parameter adjustment strategy of the current time step, or can be adjusted automatically according to the optimal process parameter adjustment strategy of the current time step. Specifically, when the process parameters of the chemical vapor deposition process are automatically adjusted according to the optimal process parameter adjustment strategy of the current time step, the opening degree of the electromagnetic valve on the delivery pipeline of each process gas, the rotating speed of the motor device of the weighing machine, and the opening degree of the electromagnetic valve on the exhaust pipeline are automatically controlled according to the process parameters fed back by the data acquisition device in real time.

[0064] In another exemplary embodiment of the present application, the optimization model of step 501 is a reinforcement learning model.

[0065] Correspondingly, step 501 includes: The input fusion feature vector is taken as the state of the reinforcement learning module at the current time step t S t The reinforcement learning module generates and outputs an action S t according to the state , and the action is the process parameter adjustment strategy of the current time step t .

[0066] For example, if it is found that the diameter of the deposited body exceeds the preset diameter, the reinforcement learning model may calculate that the process gas flow needs to be reduced or the weighing machine needs to be adjusted to reduce the speed, etc. (The reinforcement learning model finds the appropriate value of the process parameter adjustment in each data iteration, for example, the gas flow is adjusted by 10 in the old strategy, and the number of bubbles in the feature is found to increase, so the value of the gas flow adjustment in the new strategy will be reduced.) The calculated improved process parameters are transmitted to the actuators of the deposition equipment through the control interface, such as the gas flow regulating valve, the weighing machine speed controller, etc., to automatically adjust the process parameters, so that the deposition process returns to normal.

[0067] ​In another example embodiment of the present application, the reinforcement learning model is trained according to the following steps: using the following objective function L CLIP ( θ ) to optimize the value network V φ and the policy network π θ : ; ; A t =G t -V φ ( s t ) ; ; ; wherein θ denotes the parameters of the policy network of the reinforcement learning model, denotes the state generated by the old policy network s t and the action a t is calculated under the distribution of each set of state s t and action a t , and then the values of and are averaged for all sets of state s t and action a t ; denotes the ratio of the new policy probability to the old policy probability, denotes the advantage estimate value, denotes clipping to the interval [ ], which forces to be limited between and , preventing the deviation of the new policy from the old policy from being too large, and avoiding policy mutation leading to production out of control; denotes a hyperparameter for controlling the step size of policy update; denotes the probability distribution of selecting action in state in the new policy network , and ​representing the state in the old policy network probability distribution of the selected action G t representing the discounted cumulative reward starting from state V φ s t representing the value function of the state V φ output by the value network s t representing the total number of time steps of the preset chemical vapor deposition process, t representing the current time step, representing the discount factor, k representing the time offset starting from the current time step t representing the reward obtained at time step representing the weight coefficient, the current bubble number representing the total number of bubbles of the deposition body s t representing the ratio of the number of moles of produced silicon dioxide to the number of moles of used raw material silicon tetrachloride in the state s t s t is the fusion feature vector or the process state vector input to the reinforcement learning model at time step t is the L2 norm of the output action of the agent is the process parameter adjustment amount output by the PPO algorithm of the reinforcement learning model at time step t The larger the value of || is, the greater the "amplitude" and "intensity" of a single adjustment are. The weight coefficient can be adjusted according to the requirement for stability in actual production, the larger the value is, the more conservative and stable the system control strategy is.

[0068] In the early stage of training, the reinforcement learning model first receives the fusion feature vector as the state s t , and the policy network outputs different actions s t according to its current parameters θ ​​​​​​​​​​​​​​​​probability distribution Then, an action is selected from the probability distribution This action is an adjustment attempt on the process parameters, such as adjusting the amount of process gas or the amount of raw material. The selected action is applied to the actual deposition process environment, and then the feedback of the environment is observed, i.e. the new state s t+1 and reward are obtained. The reward is calculated according to a pre-designed reward function, which is closely related to the goals of the deposition process, such as improving the quality of the deposited body and improving production efficiency. For example, if the defect rate of the deposited body decreases after adjusting the process parameters, the reward is positive; otherwise, if the defect rate increases, the reward is negative. At the same time, the discounted cumulative reward s t is calculated according to the new state and reward , and the advantage estimate A t = - V φ ( ) is used to measure the advantage of the current action relative to the average level. Using a policy optimization algorithm, the parameters of the policy network are adjusted by maximizing the clipped objective function θ , and the above training process is iterated until the expected value no longer fluctuates or fluctuates within a preset fluctuation range.

[0069] The objective function is the mathematical expectation of and under the state-action pair distribution generated by the old policy . Specifically, during the reinforcement learning training process, a large number of state-action pair samples are generated from the old policy . For each sample , the values of and are calculated. After calculating the values corresponding to all samples, the average value is obtained by adding these values and dividing by the total number of samples. This average value is the objective function . This average value reflects the performance of the new policy under the sample distribution generated by the old policy, so that the new policy can improve the cumulative reward as much as possible while ensuring that it does not deviate too much from the old policy.

[0070] The clip function limits ​prevent the policy update from being too large, leading to instability. In the process of repeating the above process, the policy network gradually learns which states can obtain higher rewards, thereby continuously optimizing the policy.

[0071] Old policy network is the policy network that exists before the policy update. For a given state , the old policy network outputs a probability distribution of taking different actions . The old policy network processes the state through its network structure and outputs a probability value for each possible action i ( =1,2,3,...,n, n is the size of the action space). These probability values constitute the probability distribution of the old policy in state , that is: .

[0072] New policy network is the policy network to be optimized in the policy update process, and its parameters are θ . Similarly, for a given state , the new policy network processes the state through its network structure and outputs a probability distribution of taking different actions ( i =1,2,3,...,n, n is the size of the action space). Similar to the old policy, if the space is discrete, the new policy network outputs a probability value for each possible action , forming a probability distribution: .

[0073] Advantage estimation A t = G t -V φ ( s t ) depends on the accurate estimation of the state value V φ ( s t ) by the value network. Accurate value estimation is crucial for the optimization of the policy network. If the value network can accurately estimate the state value V φ ( s t ), the advantage estimation can more accurately reflect the advantage of a certain action relative to the average level, thereby guiding the policy network to make better decisions. The value network V φThe optimization objective is to accurately estimate the state s t The value. This is achieved by minimizing the mean squared error loss function L v (φ) = E st,Gt [ V φ ( s t )- G t ) 2 ]. The output V φ ( s t ) of the value network should be as close as possible to the actual discounted cumulative reward G t , so as to provide accurate value estimates for the policy network, helping the policy network to better evaluate the pros and cons of different states and actions.

[0074] The use weight of the raw material silicon tetrachloride in the deposition process is determined according to the number of moles of the raw material silicon tetrachloride. The data acquisition device further comprises a first weight sensor. The first weight sensor can be arranged at the bottom of the storage device of the raw material silicon tetrachloride, and the weight of the raw material silicon tetrachloride at the initial time is detected through the first weight sensor, and the residual weight of the raw material silicon tetrachloride is detected in real time. During the entire deposition process, if the raw material silicon tetrachloride is not added to the storage device, the use weight of the silicon tetrachloride in the deposition process is determined by the difference between the weight of the raw material silicon tetrachloride at the initial time and the residual weight. If the raw material silicon tetrachloride is supplemented to the storage device during the deposition process due to insufficient raw material silicon tetrachloride, then the use weight of the silicon tetrachloride is the cumulative use amount, and if the raw material silicon tetrachloride is supplemented to the storage device N times, then the cumulative use amount is the total use weight of the silicon tetrachloride for the N times.

[0075] In another exemplary embodiment of the present application, a large amount of historical production data (historical state characteristics of the deposited body and corresponding process parameters) is used to train the deep learning network. During the training process, the deep neural network continuously adjusts its weights and biases, learns the complex relationship between the image data of the deposited body and the corresponding process parameters, establishes a complex mapping relationship between the two, can predict the change of the state characteristics of the deposited body, and can also give the corresponding process adjustment strategy according to the change of the state characteristics of the deposited body, to realize the optimization control of the chemical vapor deposition quartz glass process. Through multiple iterative training, the deep learning network gradually converges. And according to the real-time collected target process parameters and target image of the new input, the parameters of the deep learning network are fine-tuned, and the mapping relationship between the state characteristics of the deposited body and the process parameters is further optimized.

[0076] As the deposition process progresses, the deep learning network is continuously reinforced, and the prediction of the growth trend of the deposition body becomes more and more accurate, and the adjustment of the process parameters becomes more and more accurate. During the deposition process, the growth rate of the deposition body is slow, and the peripheral process conditions will also change, and usually the process parameters will not change for half an hour before the deposition body characteristics reflect the change, which often causes serious defects in this period of time. For example, the change of hydrogen purity will cause the direct decrease of the furnace temperature and the weight surface temperature, and at this time, the other process parameters and the deposition body characteristics extracted usually do not change. Through prediction, the consequences of the decrease of the furnace temperature and the weight surface temperature can be known, and automatic adjustment of the process parameters is performed before the deposition body characteristics produce defects to avoid defects.

[0077] From the aspect of improving production efficiency, the method of the present application greatly reduces the human intervention of the operating workers through automatic data collection and intelligent process parameter adjustment, effectively reduces the production delay or product quality problems caused by human operation errors. In the past, manual parameter collection and process adjustment are not only prone to errors, but also have a complicated operation process and take a long time. Now the automatic system can collect data in real time and accurately and make adjustment decisions quickly, so that the whole production process is more smooth and efficient, and the production efficiency of the quartz glass is greatly improved. For example, the deposition process may deviate due to the error of manually recording the process gas amount, and re-adjustment or rework is required, but now the sensor accurately collects data and automatically transmits and analyzes, avoiding such errors and saving a lot of time and resources.

[0078] The present application has obvious advantages in overcoming the lag of manual adjustment. In the traditional way, the worker adjusts after observing the change of the deposition body characteristics or the generation of defects, which often misses the best opportunity and affects the product quality. The present method uses real-time data collection and rapid data analysis, and the system can react instantly and adjust the process parameters as soon as the deposition body appears abnormal. For example, when small bubbles begin to appear in the deposition body, the image recognition device immediately captures this defect information, and the data analysis model quickly calculates the process parameters that need to be adjusted, such as increasing the furnace temperature or changing the exhaust air volume, etc., to timely contain the further expansion of the defects and ensure the stability of the product quality.

[0079] The labor intensity of workers is greatly reduced. Originally, the workers need to pay attention to the equipment and manually adjust the process parameters during the two-month quartz glass deposition period, which is physically and mentally demanding. Now, with the help of the method of the present application, the workers only need to check the equipment operation status regularly, and most of the monitoring and adjustment work is automatically completed by the system. This not only liberates the labor of the workers and enables them to devote their energy to more valuable work, such as equipment maintenance and optimization, but also reduces the risk of operation errors caused by worker fatigue and improves the reliability of the overall production.

[0080] In terms of product quality improvement, the application finds the influence law of process parameters on the characteristics and defects of deposited quartz glass by establishing a data analysis model of accurate process parameters and deposited body characteristics and a database of the influence of process parameter changes on deposited body characteristics. Based on this, the process can be more accurately controlled in the production process to prevent and reduce the generation of product defects. At the same time, with the continuous accumulation of data in the deposition process, the model is continuously optimized, and the quality and performance parameters of the produced quartz glass are continuously improved. For example, after a period of production and model optimization, the number of bubbles in the produced quartz glass is significantly reduced, and the optical uniformity is significantly improved, which can better meet the strict requirements of high-end fields for the quality of quartz glass.

[0081] Based on the same inventive concept, the application also provides a chemical vapor deposition process parameter adjustment device for implementing the above-mentioned chemical vapor deposition process parameter adjustment method. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more chemical vapor deposition process parameter adjustment device embodiments provided below can refer to the limitations of the chemical vapor deposition process parameter adjustment method in the above text, which will not be repeated here.

[0082] In one exemplary embodiment, as shown in Figure 4 a chemical vapor deposition process parameter adjustment device 50 is provided, comprising: an industrial camera system 501 for shooting the surface of the deposited body in the chemical vapor deposition process to obtain a target image; a data acquisition device 502 for acquiring process parameters of the chemical vapor deposition process to obtain target process parameters; wherein the target process parameters are synchronously acquired with the target image; a data processing unit 503 for performing at least one feature extraction operation to extract the state characteristics of the deposited body in the chemical vapor deposition process until the extracted state characteristics exceed a preset range; If the state characteristics of the deposited body exceed the preset range, performing a process parameter automatic adjustment operation according to the target process parameters and the state characteristics of the deposited body exceeding the preset range; After performing the process parameter automatic adjustment operation, returning to perform the at least one feature extraction operation and the subsequent steps; wherein the feature extraction operation comprises: automatically identifying the state characteristics of the deposited body in the chemical vapor deposition process according to the target image of the current time step.

[0083] In the embodiments of the present application, the industrial camera system 501, the data acquisition device 502, the state features of the deposited body, and the process parameters of the chemical vapor deposition process are described in detail in the above method embodiments, and will not be described here.

[0084] In another example embodiment of the present application, the data processing unit 503 described above is further used for: performing data processing on the target image of the current time step to obtain real values of the state features of the deposited body; comparing the real values of the state features of the deposited body with the corresponding preset ranges, and if any of the state features of the deposited body is not within the corresponding preset range, it is considered that the state feature of the deposited body exceeds the preset range.

[0085] In another example embodiment of the present application, the data processing unit 503 described above is further used for: inputting the target image of the current time step into the trained deep learning network, extracting a state feature vector of the deposited body from the input target image through the deep learning network, and then identifying real values of the state features of the deposited body according to the state feature vector.

[0086] In the embodiments of the present application, the state feature vector contains state feature information of the deposited body.

[0087] In another example embodiment of the present application, the data processing unit 503 described above is further used for: if any of the state features of the deposited body exceeds the corresponding preset range, inputting the target process parameter of the current time step into the trained deep learning network, extracting a process parameter time sequence feature vector through the deep learning network, and splicing the process parameter time sequence feature vector and the state feature vector extracted in step 301 to obtain a fusion feature vector; performing a process parameter automatic adjustment operation according to the fusion feature vector.

[0088] In another example embodiment of the present application, the deep learning network includes a convolutional neural network, a recurrent neural network, and a fully connected layer. Among them: the convolutional neural network is used to extract a state feature vector of the deposited body from the input target image, and identify the state features of the deposited body according to the state feature vector; the recurrent neural network is used to extract a process parameter time sequence feature vector according to the input target process parameter of the current time step, combined with the target process parameters of the historical time steps; the fully connected layer is used to splice the state feature vector and the process parameter time sequence feature vector to obtain a fusion feature vector.

[0089] In another example embodiment of the present application, the convolutional neural network comprises an input layer, a first convolutional block, a second convolutional block, a third convolutional block, a global average pooling layer, and an output layer, wherein: The first convolutional block is configured to extract first features of the target image input by the input layer; The second convolutional block is configured to extract second features of the target image; The third convolutional block is configured to extract third features of the target image; The global average pooling layer is configured to perform global average pooling on feature maps output by the first convolutional block, the second convolutional block, and the third convolutional block, and output a state feature vector of the deposit body; The output layer is configured to perform nonlinear transformation on the state feature vector output by the global average pooling layer through an activation function (such as a ReLU activation function), and identify state features of the deposit body according to a vector output by the activation function.

[0090] In the embodiment of the present application, the use of three convolutional blocks can achieve an optimal balance between the number of parameters and the ability of feature expression. The first features can reflect the surface position, surface curvature, diameter, etc. of the deposit body in the deposit body image. The second features can reflect the surface temperature, etc. of the deposit body in the deposit body image, and the third features can reflect the internal defect features and defect change rules, etc. of the deposit body in the deposit body image.

[0091] In another example embodiment of the present application, the first convolutional block comprises a first convolutional layer and a first max-pooling layer connected in sequence, the second convolutional block comprises a second convolutional layer and a second max-pooling layer connected in sequence, and the third convolutional block comprises a third convolutional layer.

[0092] In the embodiment of the present application, the first convolutional layer uses 64 3x3 convolutional kernels, the activation function uses a ReLU activation function, the first max-pooling layer and the second max-pooling layer use 2x2 max-pooling, the second convolutional layer uses 128 3x3 convolutional kernels, and the third convolutional layer uses 256 3x3 convolutional kernels, thereby achieving accurate extraction of the first features, the second features, and the third features.

[0093] The size of the target image is 224x224 pixels, the size of the feature map output by the first max-pooling layer is 112x112x64, the size of the feature map output by the second max-pooling layer is 56x56x128, and the size of the feature map output by the third convolutional layer is 56x56x256. The global average pooling layer outputs a one-dimensional vector of 256 dimensions.

[0094] In another exemplary embodiment of this application, a deep learning network is trained using a large amount of historical production data (historical state characteristics of the sediment and corresponding process parameters) and newly acquired production data in real time. During training, the network continuously adjusts its weights and biases, learns the complex relationship between the image data of the sediment and the corresponding process parameters, and establishes a complex mapping relationship between the two. This allows the network to predict changes in the state characteristics of the sediment and provide corresponding process adjustment strategies based on these changes, thereby achieving optimized control of the chemical vapor deposition quartz glass process. Through multiple iterative training iterations, the deep learning network gradually converges.

[0095] In another exemplary embodiment of this application, the recurrent neural network employs a Long Short-Term Memory (LSTM) network.

[0096] In another exemplary embodiment of this application, the input to the recurrent neural network is X t =[ Q t , T t , W t , S t , C t , P t ], Q t For time steps t The flow rate of the process gas, T t For time steps t The furnace temperature W t For time steps t The weight of the sediment, S t For time steps t The amount of material fed, C t For time steps t The retraction speed of the grinding wheel, P t For time steps t The exhaust gas volume.

[0097] In another exemplary embodiment of this application, the data processing unit 503 described above is further configured to: Based on the fused feature vector, the process parameter adjustment strategy for the current time step is generated by optimizing the model; Adjust the process parameters of the chemical vapor deposition process according to the process parameter adjustment strategy at the current time step.

[0098] In the embodiments of the present application, the process parameters of the chemical vapor deposition process can be adjusted according to the optimal process parameter adjustment strategy of the current time step artificially, or automatically. Specifically, when the process parameters of the chemical vapor deposition process are automatically adjusted according to the optimal process parameter adjustment strategy of the current time step, the opening degree of the electromagnetic valve on the conveying pipeline of each process gas, the rotating speed of the motor equipment of the weighing machine, and the opening degree of the electromagnetic valve on the tail gas discharge pipeline are automatically controlled according to the process parameters fed back by the data acquisition device in real time.

[0099] In another exemplary embodiment of the present application, the optimization model adopts a reinforcement learning model.

[0100] Correspondingly, the data processing unit 503 described above is further configured to: take the input fusion feature vector as the state of the current time step of the reinforcement learning module t . S t The reinforcement learning module generates and outputs an action S t according to the state , and the action is the process parameter adjustment strategy of the current time step t .

[0101] In another exemplary embodiment of the present application, the data processing unit 503 described above is further configured to: train the reinforcement learning model according to the following steps: use the following objective function L CLIP ( θ ) to optimize the value network V φ and the policy network π θ : ; ; A t =G t -V φ ( s t ); ; ; wherein θ denotes the parameters of the policy network of the reinforcement learning model, denotes the parameters of the old policy network The generated state s t and actions a t Under the distribution, for each set of states s t and actions a t calculate and The value, and then for all group states s t and actions a t of Find the average; This represents the ratio of the probabilities of the new and old strategies. This represents the advantage estimate. Indicates will Clip to range [ ], forcibly Limited to and To prevent the new strategy from deviating too much from the old strategy and to avoid production runaway due to sudden strategy changes; This represents a hyperparameter used to control the step size of policy updates; Indicated in the new strategy network medium state Select action The probability distribution, In the old policy network medium state Select action The probability distribution; G t Indicates from state Initial discount cumulative rewards, V φ ( s t () represents the value network V φ Output status s t The value function; This indicates the total number of time steps in the preset chemical vapor deposition process. t Indicates the current time step. Indicates the discount factor. k Indicates the time step from the current time. t Start time offset express The reward obtained by the time step; , , The number of bubbles represents the weighting coefficient, and the current number of bubbles represents the state.s t The total number of bubbles in the lower sediment. Representing state s t The ratio of the number of moles of silicon dioxide produced to the number of moles of silicon tetrachloride used as raw material. State s t That is, time step t The fused feature vector or process state vector is input into the reinforcement learning model at any time. It is the action output by the intelligent agent. L2 norm ║2. The PPO algorithm for reinforcement learning models is at time step t The output process parameter adjustment amount. ║ The larger the ║2 value, the greater the "amplitude" and "severity" of a single adjustment. Weighting coefficient. It can be adjusted according to the stability requirements in actual production. The larger the value, the more conservative and robust the system control strategy.

[0102] In the early stages of training, the reinforcement learning model first receives the fused feature vector as its state. s t Policy networks are state-based s t According to its current parameters θ The output takes different actions probability distribution Then from the probability distribution Select an action from the middle sample This action This involves attempting to adjust process parameters, such as adjusting the process gas flow rate or the raw material feed rate. The selected action... This is applied to a real-world deposition process environment, and then the environmental feedback is observed to obtain a new state. s t+1 and rewards .award The reward function is calculated based on a pre-designed reward mechanism closely linked to the objectives of the deposition process, such as improving sediment quality and increasing production efficiency. For example, if adjusting process parameters reduces the defect rate of the sediment, a reward is given. A positive result indicates a positive result; conversely, a negative result indicates a higher defect rate. It is negative. Meanwhile, according to the new state... s t and rewards Calculate cumulative rewards for discounts And through advantage estimation A t = - V φ ( ) to measure the advantage of the current action over the average level. With the policy optimization algorithm, the parameters of the policy network are adjusted by maximizing the clipped target function θ , and the above training process is iterated until the expectation no longer fluctuates or fluctuates within a preset fluctuation range.

[0103] The target function is obtained by calculating the mathematical expectation of and under the state-action pair distribution generated by the old policy . Specifically, during the reinforcement learning training process, we generate a large number of state-action pair samples from the old policy . For each sample , we calculate and . After calculating the values corresponding to all samples, we add them up and divide by the total number of samples to obtain the average value, which is the target function . This average value reflects the performance of the new policy under the sample distribution generated by the old policy, allowing the new policy to improve the cumulative reward as much as possible while ensuring that it does not deviate too much from the old policy.

[0104] The clip function limits the range of to prevent excessive policy updates that can cause instability. In the repeated process, the policy network gradually learns which states can achieve higher rewards, thereby continuously optimizing the policy.

[0105] The old policy network exists before policy update. For a given state , the old policy network outputs a probability distribution of different actions . The old policy network processes the state through its network structure and outputs the probability value ( i =1,2,3,...,n, n is the size of the action space) of each possible action . These probability values constitute the probability distribution of the old policy under the state , i.e. .

[0106] The new policy network is the policy network to be optimized during policy update, with parameters θ . Similarly, for a given state , the new policy network processes the state through its network structure and outputs a probability distribution over actions ( i = 1, 2, 3,..., n, n is the size of the action space). Similar to the old policy, if the space is discrete, the new policy network outputs a probability value for each possible action , forming a probability distribution: .

[0107] Advantage estimation A t = G t -V φ ( s t ) relies on the accurate estimation of state value V φ ( s t ) by the value network. Accurate value estimation is crucial for the optimization of the policy network. If the value network can accurately estimate the state value V φ ( s t ), the advantage estimation can more accurately reflect the advantage of a certain action relative to the average level, thereby guiding the policy network to make better decisions. The optimization goal of the value network V φ is to accurately estimate the value of state s t . This is achieved by minimizing the mean square error loss function L v (φ) = E st,Gt [( V φ ( s t )- G t ) 2 ]. The output of the value network V φ ( s t ) should be as close as possible to the actual discounted cumulative reward G t , so as to provide accurate value estimation for the policy network and help the policy network better evaluate the pros and cons of different states and actions.

[0108] The silicon tetrachloride usage weight in the deposition process is determined according to the number of moles of the raw material silicon tetrachloride. The data acquisition device further comprises a first weight sensor. The first weight sensor can be arranged at the bottom of the storage device of the raw material silicon tetrachloride, and is used to detect the weight of the raw material silicon tetrachloride at the beginning and to detect the remaining weight of the raw material silicon tetrachloride in real time. During the entire deposition process, if the raw material silicon tetrachloride is not added to the storage device, the silicon tetrachloride usage weight in the deposition process is determined by the difference between the weight of the raw material silicon tetrachloride at the beginning and the remaining weight. If the raw material silicon tetrachloride is supplemented to the storage device due to insufficient raw material silicon tetrachloride during the deposition process, the silicon tetrachloride usage weight is the cumulative usage amount. If the raw material silicon tetrachloride is supplemented to the storage device N times, the cumulative usage amount is the total silicon tetrachloride usage weight of the N times.

[0109] In another exemplary embodiment of the present application, the data processing unit 503 described above is further configured to: The deep learning network is trained using a large amount of historical production data (historical state characteristics of the deposition body and corresponding process parameters). During the training process, the deep neural network continuously adjusts its weights and biases, learns the complex relationship between the image data of the deposition body and the corresponding process parameters, establishes a complex mapping relationship between the two, can predict the state characteristic changes of the deposition body, and can give the corresponding process adjustment strategy according to the state characteristic changes of the deposition body, to realize the optimization control of the chemical vapor deposition quartz glass process. Through multiple iterations of training, the deep learning network gradually converges. According to the newly input real-time collected target process parameters and target image, the parameters of the deep learning network are fine-tuned, and the mapping relationship between the state characteristics of the deposition body and the process parameters is further optimized.

[0110] In an exemplary embodiment, a computer device is provided, which can be a server or a terminal, and the internal structure diagram thereof can be as shown in Figure 5The computer device shown in the figure includes a processor, a memory, an Input / Output (I / O) interface, and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the process parameters of chemical vapor deposition, the state characteristics of the deposited body, and the process parameter adjustment data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement a process parameter adjustment method of a chemical vapor deposition method.

[0111] Those skilled in the art can understand that, Figure 5 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0112] In an exemplary embodiment, a computer device is also provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps in the above method embodiments.

[0113] In an exemplary embodiment, a computer readable storage medium is provided, storing a computer program, which is executed by a processor to implement the steps in the above method embodiments.

[0114] In an exemplary embodiment, a computer program product is provided, including a computer program, which is executed by a processor to implement the steps in the above method embodiments.

[0115] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use, and processing of related data need to comply with relevant regulations.

[0116] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, databases or other media used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0117] The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0118] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present application.

[0119] The principles and implementation modes of the present application are described by applying specific examples herein. The above description of the embodiments is only used to help understand the method and its core idea of the present application; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range can be changed. In conclusion, the content of the present application should not be understood as a limitation.

Claims

1. A method of adjusting process parameters of a chemical vapor deposition process, characterized by, The process parameter adjustment method of the chemical vapor deposition method comprises: performing at least one feature extraction operation to extract state features of a deposition body of a chemical vapor deposition process until the extracted state features exceed a preset range; if the state features of the deposition body exceed the preset range, performing a process parameter automatic adjustment operation according to target process parameters and the state features of the deposition body that exceed the preset range; wherein the target process parameters are process parameters of the chemical vapor deposition process collected by a data collection device; after performing the process parameter automatic adjustment operation, returning to perform the at least one feature extraction operation and subsequent steps; wherein the feature extraction operation comprises: automatically identifying the state features of the deposition body of the chemical vapor deposition process according to a target image of a current time step; wherein the target image is an image obtained by automatically photographing the surface of the deposition body by an industrial camera system, and the target image is synchronously collected with the target process parameters.

2. The method for adjusting process parameters of chemical vapor deposition according to claim 1, wherein The automatic identification of the state features of the deposition body of the chemical vapor deposition process according to the target image whether or not to exceed the preset range specifically comprises: performing data processing on the target image of the current time step to obtain real values of each state feature of the deposition body; comparing the real values of each state feature of the deposition body with the corresponding preset range, and if the real value of any state feature of the deposition body is not within the corresponding preset range, it is considered that the state features of the deposition body exceed the preset range.

3. The method for adjusting process parameters of chemical vapor deposition according to claim 2, wherein The data processing on the target image of the current time step to obtain the real values of each state feature of the deposition body specifically comprises: inputting the target image of the current time step into a trained deep learning network, extracting a state feature vector of the deposition body from the input target image through the deep learning network, and then identifying the real values of each state feature of the deposition body according to the state feature vector.

4. The method for adjusting process parameters of chemical vapor deposition according to claim 3, wherein If the state features of the deposition body exceed the preset range, the target process parameters of the current time step are input into the deep learning network to extract a process parameter time sequence feature vector, and the process parameter time sequence feature vector and the state feature vector are spliced to obtain a fusion feature vector. According to the fusion feature vector, a process parameter automatic adjustment operation is performed. The deep learning network comprises:

5. The method for adjusting process parameters of chemical vapor deposition according to claim 4, wherein a convolutional neural network for extracting a state feature vector of the deposition body from the input target image and identifying each state feature of the deposition body according to the state feature vector; a recurrent neural network for extracting the process parameter time sequence feature vector according to the target process parameters of the current time step combined with the target process parameters of the historical time steps; a fully connected layer for splicing the state feature vector and the process parameter time sequence feature vector to obtain the fusion feature vector. ​ 6. The method for adjusting process parameters of chemical vapor deposition according to claim 5, wherein The convolutional neural network comprises an input layer, a first convolutional block, a second convolutional block, a third convolutional block, a global average pooling layer, and an output layer, wherein: The first convolutional block is configured to extract first features of a target image input by the input layer; The second convolutional block is configured to extract second features of the target image; The third convolutional block is configured to extract third features of the target image; The global average pooling layer is configured to perform global average pooling on feature maps output by the first convolutional block, the second convolutional block, and the third convolutional block, and output a state feature vector of the deposit; The output layer is configured to perform nonlinear transformation on the state feature vector by an activation function, and identify state features of the deposit according to a vector output by the activation function.

7. The method for adjusting process parameters of chemical vapor deposition according to claim 4, wherein The process parameter automatic adjustment operation comprises: generating a process parameter adjustment strategy for a current time step by an optimization model according to the fused feature vector; and adjusting process parameters of the chemical vapor deposition process according to the process parameter adjustment strategy for the current time step.

8. The method for adjusting process parameters of a chemical vapor deposition method according to claim 7, wherein The optimization model adopts a reinforcement learning model. The process parameter automatic adjustment operation comprises: The input fusion feature vector is taken as a state of a reinforcement learning module at a current time step t S t The reinforcement learning module generates and outputs an action S t The action is a process parameter adjustment strategy at the current time step t .​​ 9. A process parameter adjustment device for a chemical vapor deposition method, characterized by The process parameter adjustment device for the chemical vapor deposition method comprises: an industrial camera system configured to capture a surface of a deposit of a chemical vapor deposition process to obtain a target image; a data acquisition device configured to acquire process parameters of the chemical vapor deposition process to obtain target process parameters; wherein the target process parameters are acquired synchronously with the target image; a data processing unit configured to perform at least one feature extraction operation to extract state features of the deposit of the chemical vapor deposition process until the extracted state features exceed a preset range; if the state features of the deposit exceed the preset range, performing a process parameter automatic adjustment operation according to the target process parameters and the state features of the deposit that exceed the preset range; wherein the target process parameters are process parameters of the chemical vapor deposition process acquired by the data acquisition device; after performing the process parameter automatic adjustment operation, returning to perform the at least one feature extraction operation and subsequent steps; The feature extraction operation comprises: automatically identifying state features of the deposit of the chemical vapor deposition process according to a target image of a current time step; wherein the target image is an image of the surface of the deposit obtained by automatically capturing the surface of the deposit by the industrial camera system, and is acquired synchronously with the target process parameters.

10. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the process parameter adjustment method for the chemical vapor deposition method according to any one of claims 1-8.

Images