Method and system for improving collimator optical fiber diameter precision based on RNN-PI deep learning

Abstract

The invention provides a method and a system for improving collimator optical fiber diameter precision based on RNN-PI deep learning, the method overcomes the defect that the furnace temperature is neglected due to the fact that the drawing speed is adjusted in a traditional optical fiber drawing control process, the drawing diameter is predicted by an RNN network, and the precision of the collimator optical fiber diameter is improved on the basis of the deviation between a predicted value and an ideal value. On one hand, the wire drawing speed and the wire drawing temperature are subjected to self-adaptive fine adjustment based on PI-PSO, the PI optimal solution retrieval process is optimized, optimal control parameters are found more quickly, local optimum is avoided, and the deviation between the expected diameter and the expected diameter is responded; in the fine adjustment process, based on the relevancy of the wire drawing speed, the wire drawing temperature and the diameter deviation, the weight of the wire drawing speed fine adjustment link and the weight of the wire drawing temperature fine adjustment link are dynamically controlled according to the diameter deviation saliency so as to restrain overshoot. According to the method, the obvious influence of the temperature and the speed on the wire drawing diameter is considered at the same time, the optical fiber precision of the capillary tube collimator is improved, and then the micropore array structure of the micro-channel X-ray collimator is optimized.

Description

Technical Field

[0001] The present invention relates to the field of optoelectronic technology, and in particular to a method and system for improving the diameter accuracy of a collimator optical fiber based on RNN-PI deep learning, so as to improve the quality of a capillary collimator optical fiber. Background Art

[0002] Space X-rays originate from high-energy physical processes on celestial bodies and are related to extreme physical conditions such as high temperature, high density, strong magnetic field, and strong gravitational field. They are an important way to study the properties of celestial bodies such as black holes and neutron stars. X-ray space observation equipment has also gradually developed from early survey telescopes to more complex and larger astronomical observation platforms, and the observation targets and scientific topics are becoming increasingly rich. my country's space X-ray observation technology has developed very rapidly. The new X-ray microchannel optical device based on capillary collimators has the advantages of lightweight, high pointing accuracy, large collection area, wide band, simple structure, easy mass production and high reliability, and has achieved remarkable application results in the field of space X-ray astronomy.

[0003] In 2016, the capillary X-ray collimator carried by the Pulsar Experiment 01 (XPNAV-1) verified the pulsar navigation technology. In 2017, NASA proposed the medium-sized exploration mission STROBE-X, in which the High Energy Modular Array HEMA (High Energy Modular Array) adopted the structural design of the microchannel X-ray collimator. In 2018, the Mercury planetary orbiter Bepicolombo launched by ESA used four square hole capillary collimators as the optical system on the collimating telescope MIXS-C payload to monitor the solar X-ray flux. In 2018, the "Aurora Detector" for space soft X-ray polarization detection launched by Tsinghua University used a capillary collimator at the front end of its gas detector to meet the collimator requirements of the cube satellite under extremely lightweight conditions.

[0004] The capillary collimator is mainly developed by drawing on the manufacturing technology of the microchannel plate (MCP) in the field of military night vision. MCP is an advanced microporous array optical element. As an electron multiplier, it can transmit and enhance electronic images. It is widely used in image intensifiers, display technology, high-speed oscilloscope tubes, environmental monitoring, high-energy physics experiments and other fields. Each X-ray collimator contains millions of micron-level circular channel arrays, and each microchannel is drawn by optical fiber. Therefore, the quality of the optical fiber directly determines the performance of the capillary collimator. Compared with the traditional microchannel plate MCP, the capillary X-ray collimator needs to meet more stringent technical requirements in terms of spatial array structure and X-ray transmittance.

[0005] Therefore, improving the uniformity and accuracy of X-ray collimator fiber diameter is a major challenge in research and manufacturing. Traditional MCP fiber drawing control methods and traditional methods focus on adjusting the drawing speed, ignoring the important influence of furnace temperature on fiber diameter. Higher performance optical devices place higher demands on the fiber drawing process, requiring higher diameter uniformity and accuracy. Summary of the invention

[0006] According to the first aspect of the present invention, a method for improving the diameter accuracy of the optical fiber collimator based on RNN-PI deep learning is provided to overcome the defect that the traditional optical fiber drawing control process focuses on adjusting the drawing speed while ignoring the furnace temperature. Based on the prediction of the drawing diameter by the RNN neural network proposed in the present invention, based on the deviation between the predicted value and the ideal value, based on the PI-PSO, the drawing speed and the drawing temperature control links are dynamically adjusted simultaneously, and the significant influence of temperature and speed on the drawing diameter is considered. The optical fiber accuracy of the capillary collimator is improved, and the microhole array structure of the microchannel X-ray collimator is optimized, and a high-performance capillary collimator that meets the needs of space applications is developed.

[0007] According to another object of the present invention, a method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning is used to optimize the optimal solution retrieval process of PI (PI-PSO) based on the particle swarm algorithm PSO, and its global search capability is used to assist the PI algorithm to find the optimal control parameters faster and avoid falling into the local optimum.

[0008] According to another object of the present invention, a method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning, in view of the fact that the dynamic adjustment of temperature and speed are independent links, in order to avoid overshoot, the present invention is designed to establish a dynamic adjustment mechanism of weighted coefficients. On the one hand, based on the proportional integral controller PI-PSO, the drawing speed and drawing temperature are adaptively fine-tuned respectively to respond to the deviation between the expected and desired diameters; on the other hand, in the fine-tuning process, based on the correlation between the drawing speed, drawing temperature and diameter deviation and the significance of the diameter deviation, the weights of the drawing speed fine-tuning link and the drawing temperature fine-tuning link are dynamically controlled to suppress overshoot.

[0009] As an implementation method, a method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning is used. Through an adaptive weighted control strategy, the error of optical fiber diameter is used as the main evaluation index. At the same time, the rate of change of diameter error is considered. The correlation between drawing speed, drawing furnace temperature and optical fiber diameter error is analyzed to determine their contribution to the diameter error and the significance of the diameter error change, and establish a dynamic adjustment mechanism for weighted coefficients. If the optical fiber diameter error is mainly caused by furnace temperature fluctuations, that is, the correlation between furnace temperature and diameter error is high, and the diameter error changes significantly with furnace temperature changes, then the weighted coefficient of furnace temperature adjustment is increased; conversely, if the drawing speed has a greater impact on the diameter error, then the weighted coefficient of drawing speed adjustment is increased.

[0010] According to the above-mentioned object of the present invention, the method for improving the diameter accuracy of the collimator optical fiber based on RNN-PI deep learning comprises the following steps:

[0011] Step 1: Input the optical fiber drawing process data in the current drawing control cycle into the RNN recurrent neural network to predict the drawing diameter of the next drawing control cycle and obtain the diameter prediction value D pre The optical fiber drawing process data includes drawing temperature, drawing speed, pulling force value and circular single-filament rod diameter value;

[0012] Step 2: Based on the predicted diameter value D of the next wire drawing control cycle pre The collimator fiber drawing diameter D def Based on the diameter deviation, the wire drawing speed and wire drawing temperature are adaptively adjusted based on the proportional integral controller to respond to the deviation between the expected and desired diameters.

[0013] Step 3: During the fine-tuning process, based on the correlation between the drawing speed, drawing temperature and diameter deviation and the significance of the diameter deviation, the weights of the drawing speed fine-tuning link and the drawing temperature fine-tuning link are dynamically controlled to suppress overshoot.

[0014] As an optional implementation, the proportional-integral controller determines the optimal gain coefficient for adaptive fine-tuning based on particle swarm optimization according to the diameter deviation, so as to respectively determine the adjustment values ​​of the wire drawing speed fine-tuning link and the wire drawing temperature fine-tuning link, as shown below:

[0015]

[0016] Among them, u t Indicates the adjustment value of drawing speed or drawing temperature, e t Indicates the diameter deviation, K P is the proportional gain coefficient, K I is the integral gain coefficient, K D is the derivative gain coefficient.

[0017] As an optional implementation, the proportional-integral controller determines the optimal gain coefficient for adaptive fine-tuning based on the diameter deviation and particle swarm optimization, including:

[0018] Step 2.1, set the number of particle groups N, each particle represents a set of control parameters of PI control, including the proportional gain coefficient K P , integral gain coefficient K P and the derivative gain coefficient K D , each particle is represented in vector form as:

[0019] Step 2.2, randomly initialize the position and velocity of each particle;

[0020] Step 2.3, iteratively update the velocity and position of the i-th particle, i = 1, 2, 3, ..., N:

[0021] v i，d (t+1)=wv i，d (t)+c 1 r 1 (t)[p i，d (t)-x i，d (t)]+c 2 r 2 (t)[g d (t)-x i，d (t)];

[0022] Among them, v i，d (t+1) represents the velocity of the ith particle in dimension d at time t+1, d = 1, 2, 3, corresponding to K P , K I , K D Three dimensions; w represents inertia weight; c 1 With c 2 Respectively represent learning factors; r 1 (t) and r 2 (t) represent randomly generated numbers between [0,1]; p i，d (t) represents the individual optimal position of the i-th particle in dimension d; g d (t) is the global optimal position of the group in dimension d; x i，d (t) is the current position of the ith particle in dimension d;

[0023] Step 2.4, update the individual optimal position of each particle: For the i-th particle, if its current fitness value is less than the recorded individual optimal fitness value, update the individual optimal position to the current position, that is: if f(x i(t+1))<f(p i (t)), then p i (t+1)=x i (t+1); where f(·) represents the particle fitness function, and the diameter error is used as the evaluation criterion;

[0024] Step 2.4: Update the global optimal solution, that is:

[0025]

[0026] Step 2.5: According to the preset iteration termination condition, stop the iteration update and output a set of gain coefficients K of the global optimal solution. P , integral gain coefficient K P and the derivative gain coefficient K D .

[0027] As an optional implementation, the proportional gain coefficient K P , integral gain coefficient K P and the derivative gain coefficient K D The initialization ranges of the positions are: [0,10], [0,1], [0,5]; the speed range of each dimension of the particle is [-1,1].

[0028] As an optional implementation, the inertia weight w is used to balance the global search and local search capabilities of the particle, and its value range is [0.4, 0.8].

[0029] As an optional implementation, the learning factor c 1 With c 2 They are used to balance the influence of the particle's own experience and the group's experience, and their value range is [1,2].

[0030] As an optional implementation, the preset iteration termination condition is set to a maximum number of iterations Tmax, which is 100-200.

[0031] As an optional implementation, the preset iteration termination condition is set to a target fitness value, and the iteration is stopped when the fitness value of the found optimal particle is less than the set target fitness value.

[0032] As an optional implementation, in the aforementioned step 3, during the fine-tuning process, based on the correlation between the drawing speed, the drawing temperature and the diameter deviation and the significance of the diameter deviation, the weights of the drawing speed fine-tuning link and the drawing temperature fine-tuning link are dynamically controlled to suppress overshoot, including:

[0033] Step 3.1, constructing sequences according to the drawing speed, drawing temperature, and diameter error in P drawing control cycles including the current drawing control cycle, respectively, and recording them as sequences A, B, and C;

[0034] Step 3.2: Calculate the first correlation coefficient r of sequence A and sequence C respectively A,C , and the second correlation coefficient r between the B sequence and the C sequence B,C ;

[0035] Step 3.3: According to the first correlation coefficient r A,C , the second correlation coefficient r B,C The correlation between the drawing speed, drawing temperature and diameter error in the current drawing control cycle is determined, and if the change rate of the diameter error in the current drawing control cycle is greater than or equal to a preset threshold, the dynamic weighting coefficient corresponding to the one with a greater correlation is controlled and adjusted.

[0036] In a further embodiment, in step 3.1, the following steps are also included:

[0037] Calculate the mean and standard deviation of diameter errors in P wire drawing control cycles including the current wire drawing control cycle;

[0038] If the standard deviation exceeds the mean by a predetermined proportion, the number of control cycles P is dynamically adjusted and expanded.

[0039] According to a second aspect of the present invention, a computer system is provided, comprising:

[0040] one or more processors;

[0041] A memory storing operable instructions, wherein when the instructions are executed by the one or more processors, the one or more processors perform operations, wherein the operations include executing the aforementioned method for improving the accuracy of collimator optical fiber diameter based on RNN-PI deep learning.

[0042] It should be understood that all combinations of the aforementioned concepts and the additional concepts described in more detail below can be considered as part of the inventive subject matter of the present disclosure as long as such concepts are not mutually inconsistent. In addition, all combinations of the claimed subject matter are considered as part of the inventive subject matter of the present disclosure.

[0043] The foregoing and other aspects, embodiments and features of the present invention can be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the present invention, such as the features and / or beneficial effects of the exemplary embodiments, will be apparent from the following description or learned from the practice of the specific embodiments according to the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The accompanying drawings are not intended to be drawn to scale. In the accompanying drawings, each identical or nearly identical component shown in each figure may be represented by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the present invention will now be described by way of example and with reference to the accompanying drawings.

[0045] Figure 1 It is a flow chart of a method for improving the diameter accuracy of a collimator optical fiber based on RNN-PI deep learning according to an embodiment of the present invention.

[0046] Figure 2 It is a schematic diagram of PI control in a method for improving the diameter accuracy of a collimator optical fiber based on RNN-PI deep learning according to an embodiment of the present invention.

[0047] Figure 3 It is a schematic diagram of a process of optimizing PI optimization based on PSO particle swarm algorithm to obtain a global optimal solution for control coefficients according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to better understand the technical content of the present invention, specific embodiments are given and described as follows in conjunction with the accompanying drawings.

[0049] Various aspects of the present invention are described in this disclosure with reference to the accompanying drawings, in which many illustrative embodiments are shown. The embodiments of the present disclosure are not necessarily intended to include all aspects of the present invention. It should be understood that the various concepts and embodiments introduced above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed by the present invention are not limited to any implementation. In addition, some aspects disclosed by the present invention can be used alone or in any appropriate combination with other aspects disclosed by the present invention.

[0050] {Example 1}

[0051] As shown in the accompanying drawings, a method for improving the diameter accuracy of a collimator optical fiber based on RNN-PI deep learning according to an embodiment of the present invention includes the following steps:

[0052] Step 1: Input the optical fiber drawing process data in the current drawing control cycle into the RNN recurrent neural network to predict the drawing diameter of the next drawing control cycle and obtain the diameter prediction value D pre The optical fiber drawing process data includes drawing temperature, drawing speed, pulling force value and circular single-filament rod diameter value;

[0053] Step 2: Based on the predicted diameter value D of the next wire drawing control cycle pre The collimator fiber drawing diameter D defBased on the diameter deviation, the wire drawing speed and wire drawing temperature are adaptively adjusted based on the proportional integral controller to respond to the deviation between the expected and desired diameters.

[0054] Step 3: During the fine-tuning process, based on the correlation between the drawing speed, drawing temperature and diameter deviation and the significance of the diameter deviation, the weights of the drawing speed fine-tuning link and the drawing temperature fine-tuning link are dynamically controlled to suppress overshoot.

[0055] In an embodiment of the present invention, an RNN neural network is used as a prediction model. Its key advantage in processing long-term time series data is that it allows multiple gates to learn and retain important information stored in long-term memory, while filtering out irrelevant data and having memory.

[0056] In an embodiment of the present invention, the RNN network receives input X at time t t After that, the value of the hidden layer is S t , the output value is O t , the implementation of RNN is as follows,

[0057] S t =f(U×X t +W×S t-1 +b)

[0058] O t =g(V×S t )

[0059] W is the weight matrix when the value of the hidden layer at the previous moment t-1 is transferred to the hidden layer at the next moment, V is the weight matrix from the hidden layer to the output layer, and b is the bias term. f() is the activation function of the hidden layer, and g() is the activation function of the output layer.

[0060] It should be understood that the RNN neural network can be obtained through pre-training and used as a prediction model in the capillary collimator optical fiber quality preparation process.

[0061] As an optional implementation, the proportional-integral controller determines the optimal gain coefficient for adaptive fine-tuning based on particle swarm optimization according to the diameter deviation, so as to respectively determine the adjustment values ​​of the wire drawing speed fine-tuning link and the wire drawing temperature fine-tuning link, as shown below:

[0062]

[0063] Among them, u t Indicates the adjustment value of drawing speed or drawing temperature, e t Indicates the diameter deviation, K P is the proportional gain coefficient, K I is the integral gain coefficient, K D is the derivative gain coefficient.

[0064] As an optional implementation, the proportional-integral controller determines the optimal gain coefficient for adaptive fine-tuning based on the diameter deviation based on particle swarm optimization (PSO), including:

[0065] Step 2.1, set the number of particle groups N, each particle represents a set of control parameters of PI control, including the proportional gain coefficient K P , integral gain coefficient K P and the derivative gain coefficient K D , each particle is represented in vector form as:

[0066] Step 2.2, randomly initialize the position and velocity of each particle;

[0067] Step 2.3, iteratively update the velocity and position of the i-th particle, i = 1, 2, 3, ..., N:

[0068] v i，d (t+1)=wv i，d (t)+c 1 r 1 (t)[p i，d (t)-x i，d (t)]+c 2 r 2 (t)[g d (t)-x i，d (t)];

[0069] Among them, v i，d (t+1) represents the velocity of the ith particle in dimension d at time t+1, d = 1, 2, 3, corresponding to K P , K I , K D Three dimensions; w represents inertia weight; c 1 With c 2 Respectively represent learning factors; r 1 (t) and r 2 (t) represent randomly generated numbers between [0,1]; p i，d (t) represents the individual optimal position of the i-th particle in dimension d; g d (t) is the global optimal position of the group in dimension d; x i，d (t) is the current position of the ith particle in dimension d;

[0070] Step 2.4, update the individual optimal position of each particle: For the i-th particle, if its current fitness value is less than the recorded individual optimal fitness value, update the individual optimal position to the current position, that is: if f(xi (t+1))<f(p i (t)), then p i (t+1)=x i (t+1); where f(·) represents the particle fitness function, and the diameter error is used as the evaluation criterion;

[0071] Step 2.4: Update the global optimal solution, that is:

[0072]

[0073] Step 2.5: According to the preset iteration termination condition, stop the iteration update and output a set of gain coefficients K of the global optimal solution. P , integral gain coefficient K P and the derivative gain coefficient K D .

[0074] As an optional implementation, the proportional gain coefficient K P , integral gain coefficient K P and the derivative gain coefficient K D The initialization ranges of the positions are: [0,10], [0,1], [0,5]; the speed range of each dimension of the particle is [-1,1].

[0075] As an optional implementation, the inertia weight w is used to balance the global search and local search capabilities of the particle, and its value range is [0.4, 0.8].

[0076] As an optional implementation, the learning factor c 1 With c 2 They are used to balance the influence of the particle's own experience and the group's experience, and their value range is [1,2].

[0077] As an optional implementation, the preset iteration termination condition is set to a maximum number of iterations Tmax, which is 100-200.

[0078] As an optional implementation, the preset iteration termination condition is set to a target fitness value, and the iteration is stopped when the fitness value of the found optimal particle is less than the set target fitness value.

[0079] Therefore, combined with the aforementioned drawing speed or drawing temperature adjustment value u t The calculation formula can be used to obtain the adjustment values ​​of the wire drawing speed fine-tuning link and the wire drawing temperature fine-tuning link.

[0080] In this embodiment, the adjustment values ​​of the drawing speed and the drawing temperature are generated independently. If they are directly executed on the X-ray collimator optical fiber drawing machine, overcontrol will be caused. For this reason, in an embodiment of the present invention, the overshoot effect of the two independent adjustments is reduced by dynamically configuring the fine-tuning weighting coefficients α and β. The fine-tuning weighting coefficient α corresponds to the drawing speed adjustment value, and the fine-tuning weighting coefficient β corresponds to the drawing temperature adjustment value, α+β=1.

[0081] Therefore, the drawing speed adjustment value output by PI-PSO is multiplied by α, and the drawing temperature adjustment value is multiplied by β. These are used as the final adjustment values ​​and applied to the X-ray collimator optical fiber drawing machine for feedback control to achieve precise control of the drawing diameter.

[0082] As an optional implementation, in the aforementioned step 3, during the fine-tuning process, based on the correlation between the drawing speed, the drawing temperature and the diameter deviation and the significance of the diameter deviation, the weights of the drawing speed fine-tuning link and the drawing temperature fine-tuning link are dynamically controlled to suppress overshoot, including:

[0083] Step 3.1, construct sequences according to the drawing speed, drawing temperature and diameter error in N drawing control cycles including the current drawing control cycle, and record them as sequences A, B and C; for the sake of convenience, n cycles are taken as an example for explanation; sequence A = [a1, a2, a3, ..., an], sequence B = [b1, b2, b3, ..., bn], sequence C = [c1, c2, c3, ..., cn];

[0084] Step 3.2: Calculate the first correlation coefficient r of sequence A and sequence C respectively A,C , and the second correlation coefficient r between the B sequence and the C sequence B,C ;

[0085] Step 3.3: According to the first correlation coefficient r A,C , the second correlation coefficient r B,C The correlation between the drawing speed, drawing temperature and diameter error in the current drawing control cycle is determined, and if the change rate of the diameter error in the current drawing control cycle is greater than or equal to a preset threshold, the dynamic weighting coefficient corresponding to the one with a greater correlation is controlled and adjusted.

[0086] In an optional embodiment, the adjustment range of the dynamic weighting coefficient is generally set to be adjusted according to a preset gradient percentage.

[0087] As an optional embodiment, in the aforementioned step 3.2, the correlation coefficient may be evaluated using a Pearson correlation coefficient, a Spearman correlation coefficient, or the like.

[0088] In this example, the Pearson correlation coefficient is used as an example to describe the operation process of step 3.2 in detail as follows:

[0089] First, calculate the series mean:

[0090] For sequence A, sequence B, and sequence C, calculate their means respectively.

[0091]

[0092]

[0093] Then, calculate the covariance Cov(A, C) between sequence A and sequence C and the covariance Cov(B, C) between sequence B and sequence C:

[0094]

[0095]

[0096] Then, calculate the standard deviation S of the three series A , S B , S C :

[0097] Standard deviation of sequence A:

[0098] Standard deviation of sequence B:

[0099] Standard deviation of sequence C:

[0100] Finally, the correlation coefficient between the sequences is calculated as follows:

[0101] Pearson correlation coefficient r between sequence A and sequence C A,C :

[0102] Pearson correlation coefficient r between sequence B and sequence C B,C :

[0103] On this basis, we further compared the Pearson correlation coefficient r A,C Pearson correlation coefficient r B,C The size of the wire drawing speed, wire drawing temperature and diameter deviation is determined by:

[0104] -If |r A,C |>|r B,C |, it means that sequence A is more related to sequence C;

[0105] -If |r A,C |<|r B,C |, it means that sequence B is more correlated with sequence C.

[0106] In a further embodiment, in step 3.1, the following steps are also included:

[0107] Calculate the mean and standard deviation of diameter errors in P wire drawing control cycles including the current wire drawing control cycle;

[0108] If the standard deviation exceeds the mean by a predetermined proportion, such as a preset value in the range of 5%-10%, the number of control cycles P is dynamically adjusted and expanded.

[0109] In this embodiment, if the fluctuation of the diameter error is large, the cycle size within the control window range is dynamically expanded to obtain a more accurate result.

[0110] In an optional embodiment, when the fluctuation is stable, the number of control cycles is gradually restored to P. Under specific process precision control requirements, the number of control cycles P can be dynamically configured, and in this example, the value is selected within the range of 5-10.

[0111] {Example 2}

[0112] In combination with the method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning in the above embodiment, a computer system is also proposed, including: one or more processors, and a memory.

[0113] The aforementioned memory is used to store operable instructions, and when the instructions are executed by the one or more processors, the one or more processors perform operations. The operations include the process of executing the method for improving the diameter accuracy of the collimator optical fiber based on RNN-PI deep learning of the aforementioned embodiment.

[0114] Although the present invention has been disclosed as above with preferred embodiments, it is not intended to limit the present invention. A person with ordinary knowledge in the technical field to which the present invention belongs may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be determined by the definition of the claims.

Claims

1. A method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning, characterized in that: The following steps are involved: Step 1: Input the optical fiber drawing process data in the current drawing control cycle into the RNN recurrent neural network to predict the drawing diameter of the next drawing control cycle and obtain the diameter prediction value D pre The optical fiber drawing process data includes drawing temperature, drawing speed, pulling force value and circular single-filament rod diameter value; Step 2: Based on the predicted diameter value D of the next wire drawing control cycle pre The collimator fiber drawing diameter D is preset def Based on the diameter deviation, the wire drawing speed and wire drawing temperature are adaptively adjusted based on the proportional integral controller to respond to the deviation between the expected and desired diameters. Step 3: During the fine-tuning process, based on the correlation between the drawing speed, drawing temperature and diameter deviation and the significance of the diameter deviation, the weights of the drawing speed fine-tuning link and the drawing temperature fine-tuning link are dynamically controlled to suppress overshoot.

2. The method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning according to claim 1, characterized in that: The proportional-integral controller determines the optimal gain coefficient for adaptive fine-tuning based on the particle swarm optimization according to the diameter deviation, so as to respectively determine the adjustment values ​​of the wire drawing speed fine-tuning link and the wire drawing temperature fine-tuning link, which are expressed as follows: Among them, u t Indicates the adjustment value of drawing speed or drawing temperature, e t Indicates the diameter deviation, K P is the proportional gain coefficient, K I is the integral gain coefficient, K D is the derivative gain coefficient.

3. The method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning according to claim 2 is characterized in that: The proportional-integral controller determines the optimal gain coefficient for adaptive fine-tuning based on particle swarm optimization according to the diameter deviation, including: Step 2.1, set the number of particle groups N, each particle represents a set of control parameters of PI control, including the proportional gain coefficient K P , integral gain coefficient K P and the derivative gain coefficient K D , each particle is represented in vector form as: Step 2.2, randomly initialize the position and velocity of each particle; Step 2.3, iteratively update the velocity and position of the i-th particle, i = 1, 2, 3, ..., N: υ i，d (t+1)=wυ i，d (t)+c1r1(t)[p i，d (t))—x i，d (t)]+c2r2(t)[g d (t))-x i，d (t)]; Among them, i，d (t+1) represents the velocity of the ith particle in dimension d at time t+1, d = 1, 2, 3, corresponding to K P , K I , K D Three dimensions; w represents the inertia weight; c1 and c2 represent the learning factors; r1(t) and r2(t) represent randomly generated numbers between [0,1]; p i，d (t) represents the individual optimal position of the i-th particle in dimension d; g d (t) is the global optimal position of the group in dimension d; x i，d (t) is the current position of the ith particle in dimension d; Step 2.4, update the individual optimal position of each particle: For the i-th particle, if its current fitness value is less than the recorded individual optimal fitness value, update the individual optimal position to the current position, that is: if f(x i (t+1))<f(p i (t)), then p i (t+1)=x i (t+1); where f(·) represents the particle fitness function, and the diameter error is used as the evaluation criterion; Step 2.4: Update the global optimal solution, that is: Step 2.5: According to the preset iteration termination condition, stop the iteration update and output a set of gain coefficients K of the global optimal solution. P , integral gain coefficient K P and the derivative gain coefficient K D .

4. The method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning according to claim 3 is characterized in that: The proportional gain coefficient K P , integral gain coefficient K P and the derivative gain coefficient K D The initialization ranges of the positions are: [0,10], [0,1], [0,5]; the speed range of each dimension of the particle is [-1,1].

5. The method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning according to claim 3 is characterized in that: The inertia weight w is used to balance the global search and local search capabilities of the particle, and its value range is [0.4, 0.8].

6. The method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning according to claim 3, characterized in that: The learning factors c1 and c2 are used to balance the influence of the particle's own experience and the group's experience, and their value range is [1,2].

7. The method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning according to claim 3, characterized in that: The preset iteration termination condition is set to the maximum number of iterations Tmax, which is 100 to 200.

8. The method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning according to claim 3, characterized in that: The preset iteration termination condition is set as the target fitness value, and the iteration is stopped when the fitness value of the found optimal particle is less than the set target fitness value.

9. The method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning according to any one of claims 1 to 8, characterized in that: In step 3, during the fine-tuning process, based on the correlation between the drawing speed, the drawing temperature and the diameter deviation and the significance of the diameter deviation, the weights of the drawing speed fine-tuning link and the drawing temperature fine-tuning link are dynamically controlled to suppress overshoot, including: Step 3.1, constructing sequences according to the drawing speed, drawing temperature, and diameter error in P drawing control cycles including the current drawing control cycle, respectively, and recording them as sequences A, B, and C; Step 3.2: Calculate the first correlation coefficient r of sequence A and sequence C respectively A,C , and the second correlation coefficient r between the B sequence and the C sequence B,C ; Step 3.3: According to the first correlation coefficient r A,C , the second correlation coefficient r B,C The correlation between the drawing speed, drawing temperature and diameter error in the current drawing control cycle is determined, and if the change rate of the diameter error in the current drawing control cycle is greater than or equal to a preset threshold, the dynamic weighting coefficient corresponding to the one with a greater correlation is controlled and adjusted.

10. The method for improving the diameter accuracy of collimator optical fiber based on RNN-PI deep learning according to claim 9, characterized in that: The step 3.1 also includes the following steps: Calculate the mean and standard deviation of diameter errors in P wire drawing control cycles including the current wire drawing control cycle; If the standard deviation exceeds the mean by a predetermined proportion, the number of control cycles P is dynamically adjusted and expanded.

11. A computer system, characterized in that: include: one or more processors; A memory storing operable instructions, wherein when the instructions are executed by the one or more processors, the one or more processors perform operations, wherein the operations include the process of executing the method for improving the diameter accuracy of the collimator optical fiber based on RNN-PI deep learning as described in any one of claims 1 to 10.

Images