Efficient calculation method and device for light beam quality time-varying sequence

Abstract

The invention provides a light beam quality time-varying sequence efficient calculation method and device, and the method comprises the steps: carrying out the processing of an optical load, design parameters and a working time mode of an optical-mechanical system according to a data set standard format, and creating a to-be-predicted optical-mechanical system data set; processing a to-be-predicted optical-mechanical system data set through the linear scaling mapping function and the nonlinear scaling mapping function to obtain a predicted data set; the prediction data set is input into a trained long short-term memory recurrent neural network based on the attention mechanism, a light beam quality time-varying sequence conforming to a working time mode is output, and the long short-term memory recurrent neural network based on the attention mechanism is an encoder and decoder architecture. The invention provides a method and a device for efficiently calculating a light beam quality time-varying sequence, which are used for solving the problem that the accurate transmission light beam quality time-varying sequence of a designed optical-mechanical system under a specified optical load is difficult to calculate within finite time in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical technology, in particular to a light beam quality time-varying sequence efficient calculation method and device. BACKGROUND

[0002] In the transmission process, laser is usually affected by the non-steady coupling thermal effect of the multi-physical field of light-heat-force-flow, so the efficient and accurate calculation of the transmission light beam quality derivation law is an important and difficult task, and directly determines the optical-mechanical system design and its final output performance, and the traditional multi-physical field simulation of light-mechanical-thermal-force-flow is difficult to calculate the accurate transmission light beam quality time-varying sequence of the designed optical-mechanical system under the specified optical load in a limited time. SUMMARY

[0003] The present application provides a light beam quality time-varying sequence efficient calculation method and device to solve the problem that it is difficult to calculate the accurate transmission light beam quality time-varying sequence of the designed optical-mechanical system under the specified optical load in a limited time in the prior art.

[0004] In a first aspect, the present application provides a light beam quality time-varying sequence efficient calculation method, comprising:

[0005] According to the data set standard format, the optical load and design parameters and working time mode of the optical-mechanical system are processed to create an optical-mechanical system data set to be predicted, the data set standard format is used to determine the ordering and data structure of each data in the optical-mechanical system data set, the data structure is an array, and the working time mode includes working time length and data frame rate;

[0006] The optical-mechanical system data set to be predicted is processed through a linear scaling mapping function and a nonlinear scaling mapping function to obtain a prediction data set;

[0007] The prediction data set is input into a trained long short-term memory recurrent neural network based on attention mechanism to output a light beam quality time-varying sequence conforming to the working time mode, the long short-term memory recurrent neural network based on attention mechanism is an encoder and decoder architecture, the input of the encoder is the prediction data set, and the output of the decoder is the light beam quality time-varying sequence.

[0008] Optionally, the optical load and design parameters of the optical-mechanical system include power, power density, relative humidity, air pressure, ambient temperature, incident angle, optical axis elevation angle, clear aperture, element type code, lens physical diameter, lens curvature, clear medium length and mirror film layer light-heat conversion rate, wherein:

[0009] The power and the power density belong to the optical load;

[0010] The relative humidity, the air pressure, the ambient temperature, the incidence angle, the optical axis elevation angle, the clear aperture diameter, the element type code, the lens physical diameter, the lens curvature, the clear medium length, and the mirror film layer photothermal conversion rate belong to the design parameters.

[0011] Optionally, the processing of the to-be-predicted optical-mechanical system dataset by the linear scaling mapping function and the nonlinear scaling mapping function comprises:

[0012] The power, the power density, the relative humidity, the air pressure, the ambient temperature, the incidence angle, the optical axis elevation angle, the clear aperture diameter, the element type code, the lens physical diameter, the lens curvature, the mirror film layer photothermal conversion rate, and the working time mode are processed by the linear scaling mapping function, and the expression of the linear scaling mapping function is as follows:

[0013]

[0014] wherein round() is an integer function, [x min ,x max ] is an original range, [x′ max ,x′ max ] is a scaled range;

[0015] The clear medium length is processed by the nonlinear scaling mapping function, and the expression of the nonlinear scaling mapping function is as follows:

[0016]

[0017] wherein p i , q i are adjustable parameters, which are set according to the optical-mechanical system.

[0018] Optionally, when there is data not belonging to the optical load and the design parameters in the data of the variable sequence for predicting the optical beam quality of a specific optical element in the optical-mechanical system, the data not belonging to the optical load and the design parameters are multiplexed into the same array with the optical load or the design parameter similar in physical meaning.

[0019] Optionally, the encoder and the decoder architecture are as follows:

[0020] The encoder comprises a first data embedding layer, a first stacked LSTM layer, a final state extraction layer, a context correlation layer, an attention concatenation layer, and a first nonlinear dense layer, and the decoder comprises a second data embedding layer, a second stacked LSTM layer, a correlation dot product layer, a second nonlinear dense layer, and a third nonlinear dense layer, wherein:

[0021] the first data embedding layer and the second data embedding layer are respectively configured to encode inputs of the encoder and the decoder into vectors;

[0022] the first stacked LSTM layer and the second stacked LSTM layer are respectively configured to increase dimensions of the encoder and the decoder;

[0023] the final state extraction layer is configured to extract a final state of a last layer in the first stacked LTSM layer and initialize attention weight of a last layer in the second stacked LTSM layer, the final state being data sliced along a second dimension in a three-dimensional tensor;

[0024] the correlation dot product layer is configured to perform a dot product on outputs of the last layer in the first stacked LSTM layer and the last layer in the second stacked LSTM layer;

[0025] the first nonlinear dense layer, the second nonlinear dense layer and the third nonlinear dense layer are respectively configured to process outputs of the attention concatenation layer, the correlation dot product layer and the first nonlinear dense layer according to a nonlinear function;

[0026] the context correlation layer is configured to perform a dot product on outputs of the last layer in the first stacked LSTM layer and the second nonlinear dense layer;

[0027] the attention concatenation layer is configured to concatenate outputs of the context correlation layer and the last layer in the second stacked LSTM layer along a third dimension in the three-dimensional tensor;

[0028] the output of the third nonlinear dense layer is added at the end of the input of the second data embedding layer.

[0029] Optionally, the attention mechanism based long short-term memory recurrent neural network is trained in different working time modes, and a training process in each working time mode comprises:

[0030] a training data set required in a current working time mode is created based on a training round, a number of training data sets required for each training round and a generator function, the number of training data sets required for each training round being set according to a running memory size of a host computer, and a pseudo code of the generator function is as follows:

[0031] Function single training data set generator function(){

[0032] a single unprocessed sample set is extracted from an original data set, the single unprocessed sample set including an optical load of an optical-mechanical system, design parameters and a measured time-varying sequence of beam quality when the optical path experiment is performed;

[0033] processing the optical load and design parameters of the optical path test time optical machine system and the current working time mode according to the data set standard format, creating an optical machine system data set to be trained;

[0034] processing the optical machine system data set to be trained through the linear scaling mapping function and the nonlinear scaling mapping function, obtaining an encoder input data set;

[0035] extracting an optical beam quality time-varying sequence to be trained from the measured optical beam quality time-varying sequence according to the current working time mode;

[0036] processing the optical beam quality time-varying sequence to be trained through the linear scaling mapping function, binarizing the processed optical beam quality time-varying sequence, and obtaining a decoder output data set;

[0037] right shifting the processed optical beam quality time-varying sequence to obtain a decoder input data set;

[0038] }

[0039] Function repeat call function(){

[0040] if (the training rounds have not been reached) {repeat call the single training data set generation function();}

[0041] }

[0042] Function training data set object construction function required for each round of training(){

[0043] divide the single training data set into the encoder input data set, the decoder output data set, and the decoder input data set;

[0044] assemble the training data set required for each round of training according to the number of training data sets required for each round of training;

[0045] call the repeat call function();

[0046] }

[0047] import the training data set required for each round of training into the attention mechanism-based long short-term memory recurrent neural network for training, and stop training when the training rounds are reached.

[0048] Optionally, before the training data set required for each round of training is imported into the attention mechanism-based long short-term memory recurrent neural network for training, one or two non-essential design parameters are randomly selected from the encoder input data set and set to zero.

[0049] In a second aspect, the present application provides a high-efficiency device for calculating time-varying sequence of beam quality, comprising a creating module, a processing module and a predicting module, wherein:

[0050] The creating module is configured to process the optical load and design parameters and working time mode of the optical-mechanical system according to a data set standard format, to create a data set of the optical-mechanical system to be predicted, and the data set standard format is used to determine the order and data structure of each data in the data set of the optical-mechanical system, the data structure is an array, and the working time mode comprises working time length and data frame rate.

[0051] The processing module is configured to process the data set of the optical-mechanical system to be predicted by using a linear scaling mapping function and a nonlinear scaling mapping function, to obtain a prediction data set.

[0052] The predicting module is configured to input the prediction data set into a long short-term memory recurrent neural network based on attention mechanism which has been trained, to output a time-varying sequence of beam quality conforming to the working time mode, and the long short-term memory recurrent neural network based on attention mechanism is an encoder-decoder architecture, the input of the encoder is the prediction data set, and the output of the decoder is the time-varying sequence of beam quality.

[0053] In a third aspect, the present application provides a computing device, comprising:

[0054] a memory configured to store program instructions;

[0055] a processor configured to invoke the program instructions stored in the memory, and execute any of the above methods according to the obtained program.

[0056] In a fourth aspect, the present application provides a computer-readable nonvolatile storage medium comprising computer-readable instructions, which, when read and executed by a computer, cause the computer to execute any of the above methods.

[0057] The above scheme realizes high-efficiency and accurate calculation of time-varying sequence of beam quality transmitted by an optical-mechanical link under the condition of a specified optical load and design parameters of the optical-mechanical system; can maximize the adaptation to the unexpected contraction and expansion characteristics of the optical-mechanical link module and the measured time-varying sequence of beam quality frequently encountered in real optical path tests, is suitable for full-size optical links, is compatible with the training and calculation process under the scene of missing part of the measured data, can achieve the expected convergence effect in a certain scale of data set, and can be expanded to tens to hundreds of optical and thermal control components to describe parameters including plane / curved transmissive and reflective lenses, plane / channel diaphragms, air charging and discharging ports, channel profiles in the optical-mechanical system and optical loads, and can achieve the expected convergence effect in a certain scale of data set, and meets the analysis requirements of optical-mechanical products of various platforms. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and all other drawings obtained by those skilled in the art without creative effort based on these drawings also belong to the protection scope of the present application.

[0059] Figure 1 A flowchart of a light beam quality time-varying sequence efficient calculation method provided by the embodiments of the present application is shown in the figure.

[0060] Figure 2 A schematic diagram of a light transmission medium length mapping curve provided by the embodiments of the present application is shown in the figure.

[0061] Figure 3 A schematic diagram of an encoder and a decoder provided by the embodiments of the present application is shown in the figure.

[0062] Figure 4 A structural schematic diagram of a light beam quality time-varying sequence efficient calculation device provided by the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0063] In order to make the objects, technical solutions and advantages of the present application more clear, the present application will be further described in detail below with reference to the drawings. Obviously, the described embodiments only constitute some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort also belong to the protection scope of the present application.

[0064] It should be noted that the terms "first", "second", etc. involved in the documents of the present application are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0065] Figure 1 The flow of a light beam quality time-varying sequence efficient calculation method provided by the embodiments of the present application is shown in detail, which includes:

[0066] S101, processing the optical load and the design parameters and the working time mode of the optical-mechanical system according to the data set standard format, and creating the optical-mechanical system data set to be predicted.

[0067] Specifically, the data set standard format is used to determine the order and data structure of each data in the optical-mechanical system data set, and the data structure is an array.

[0068] It should be noted that in the optical-mechanical system data set, the working time mode describes the entire optical-mechanical system, and the values of all arrays on the same column except the working time mode collectively describe an optical element in the optical-mechanical system.

[0069] The data set standard format provides a good normalization paradigm for subsequent continuous data updating and injection.

[0070] Specifically, the working time mode includes working time length and data frame rate. When the working time length is 10s and the data frame rate is 1Hz, this working time mode represents that the beam quality is extracted once every 1s in the 10s beam quality time-varying sequence.

[0071] In an example, the optical load and the design parameters of the optical-mechanical system include power, power density, relative humidity, air pressure, ambient temperature, incident angle, optical axis elevation angle, clear aperture, element type code, lens physical diameter, lens curvature, clear medium length, and mirror film layer photothermal conversion rate, wherein:

[0072] The power and the power density belong to the optical load;

[0073] The relative humidity, air pressure, ambient temperature, incident angle, optical axis elevation angle, clear aperture, element type code, lens physical diameter, lens curvature, clear medium length, and mirror film layer photothermal conversion rate belong to the design parameters.

[0074] In an example, when there is data in the data of the predicted beam quality time-varying sequence that does not belong to the optical load and the design parameters in a specific optical element in the optical-mechanical system, the data that does not belong to the optical load and the design parameters is multiplexed with the optical load or the design parameter similar in physical meaning as the same array.

[0075] For example, when the specific optical element is a gas filling device, its protection gas filling flow, protection gas filling flow rate, filled protection gas purity, gas filling positive pressure threshold, filled protection gas temperature and protection gas filling elevation angle do not belong to the optical load and design parameters. The protection gas filling flow and the protection gas filling flow rate belong to the pneumatic load, the power and the power density belong to the optical load, the protection gas filling flow has a similar physical meaning as the power, the protection gas filling flow rate has a similar physical meaning as the power density. The filled protection gas purity is used to reflect the proportion of impurity gas in the general gas, and the relative humidity is used to reflect the proportion of water vapor in the gas, both of which have similar physical meanings. The gas filling positive pressure threshold has a similar physical meaning as the gas pressure, the filled protection gas temperature has a similar physical meaning as the ambient temperature, and the protection gas filling elevation angle has a similar physical meaning as the optical axis elevation angle.

[0076] Therefore, for the gas filling device:

[0077] The protection gas filling flow and the power of the gas filling device are multiplexed in the same array, and the value of the array corresponding to the power on the column where the gas filling device is located actually represents the protection gas filling flow.

[0078] The protection gas filling flow rate and the power density of the gas filling device are multiplexed in the same array, and the value of the array corresponding to the power density on the column where the gas filling device is located actually represents the protection gas filling flow rate.

[0079] The filled protection gas purity and the relative humidity of the gas filling device are multiplexed in the same array, and the value of the array corresponding to the relative humidity on the column where the gas filling device is located actually represents the filled protection gas purity.

[0080] The gas filling positive pressure threshold and the gas pressure of the gas filling device are multiplexed in the same array, and the value of the array corresponding to the gas pressure on the column where the gas filling device is located actually represents the gas filling positive pressure threshold.

[0081] The filled protection gas temperature and the ambient temperature of the gas filling device are multiplexed in the same array, and the value of the array corresponding to the ambient temperature on the column where the gas filling device is located actually represents the filled protection gas temperature.

[0082] The protection gas filling elevation angle and the optical axis elevation angle of the gas filling device are multiplexed in the same array, and the value of the array corresponding to the optical axis elevation angle on the column where the gas filling device is located actually represents the protection gas filling elevation angle.

[0083] For example, when the specific optical element is an exhaust device, its exhaust port pressure relief threshold has a similar physical meaning as the gas pressure.

[0084] Therefore, for the exhaust device:

[0085] The exhaust port pressure relief threshold and the gas pressure of the exhaust device are multiplexed in the same array, and the value of the array corresponding to the gas pressure on the column where the exhaust device is located actually represents the exhaust port pressure relief threshold.

[0086] The optical-mechanical system dataset created according to the dataset standard format is shown in Table 1, which contains possible multiplexing conditions.

[0087] Table 1

[0088]

[0089] S102, processing the optical-mechanical system dataset to be predicted by a linear scaling mapping function and a nonlinear scaling mapping function to obtain a prediction dataset.

[0090] The threshold limiting mapping of the optical-mechanical system dataset to be predicted is realized by the linear scaling mapping function and the nonlinear scaling mapping function, that is, the optical-mechanical system dataset to be predicted is mapped to a data domain range recognizable by the attention mechanism-based long short-term memory recurrent neural network.

[0091] In an example, processing the optical-mechanical system dataset to be predicted by the linear scaling mapping function and the nonlinear scaling mapping function includes:

[0092] The power, power density, relative humidity, air pressure, ambient temperature, incident angle, optical axis elevation angle, clear aperture, element type code, lens physical diameter, lens curvature, mirror film layer light-heat conversion rate, and working time mode are processed by the linear scaling mapping function, and the expression of the linear scaling mapping function is as follows:

[0093]

[0094] wherein round() is an integer function, [x min ,x max ] is the original range, [x′ max ,x′ max ] is the scaled range;

[0095] The clear aperture length is processed by the nonlinear scaling mapping function, and the expression of the nonlinear scaling mapping function is as follows:

[0096]

[0097] wherein p i and q i are adjustable parameters, which are set according to the optical-mechanical system.

[0098] Since the etendue has a wide definition domain and the beam quality time-varying sequence is very sensitive to the change of the etendue when it describes the optical element of the curved mirror type, the etendue is not suitable for threshold limiting mapping using a linear scaling mapping function. The mapping function of the etendue is described in the form of a multinomial fraction, a nonlinear scaling mapping function is obtained, which can have a large slope in the numerical sensitive area of the etendue and a small slope in the numerical non-sensitive area of the etendue, and realize reasonable mapping distribution within a limited threshold.

[0099] The etendue mapping curve is shown in Figure 2 , wherein the horizontal axis is the etendue and the vertical axis is the mapped etendue.

[0100] S103, input the prediction data set into the trained attention mechanism based long short-term memory recurrent neural network, and output the beam quality time-varying sequence conforming to the working time mode.

[0101] Specifically, the attention mechanism based long short-term memory recurrent neural network is an encoder and decoder architecture, the input of the encoder is the prediction data set, and the output of the decoder is the beam quality time-varying sequence.

[0102] In an example, as shown in Figure 3 , the encoder and decoder architecture is:

[0103] The encoder includes a first data embedding layer, a first stacked LSTM layer, a last state extraction layer, a context correlation layer, an attention splicing layer and a first nonlinear dense layer, and the decoder includes a second data embedding layer, a second stacked LSTM layer, a correlation dot product layer, a second nonlinear dense layer and a third nonlinear dense layer, wherein:

[0104] The first data embedding layer and the second data embedding layer are respectively used to encode the input of the encoder and the decoder into a vector;

[0105] The first stacked LSTM layer and the second stacked LSTM layer are respectively used to increase the dimension of the encoder and the decoder to improve the ability to handle complex problems;

[0106] The last state extraction layer is used to extract the last state of the last layer in the first stacked LTSM layer and initialize the attention weight of the last layer in the second stacked LTSM layer, and the last state is the data obtained by slicing along the second dimension in the three-dimensional tensor;

[0107] The correlation dot product layer is used to perform dot product on the outputs of the last layer in the first stacked LSTM layer and the last layer in the second stacked LSTM layer;

[0108] The first nonlinear dense layer, the second nonlinear dense layer and the third nonlinear dense layer are respectively used for processing the outputs of the attention concatenation layer, the correlation dot product layer and the first nonlinear dense layer according to nonlinear functions;

[0109] The context correlation layer is used for performing a dot product on the outputs of the last layer in the first stacked LSTM layer and the second nonlinear dense layer;

[0110] The attention concatenation layer is used for concatenating the outputs of the context correlation layer and the last layer in the second stacked LSTM layer along the third dimension in a three-dimensional tensor;

[0111] The output of the third nonlinear dense layer is added at the end of the input of the second data embedding layer.

[0112] It should be noted that the output of the third nonlinear dense layer is the output of the decoder, and the input of the second data embedding layer is the input of the decoder.

[0113] Since the inputs of the encoder and the decoder and their dimensions are different, the dimensions of the first data embedding layer, the second data embedding layer, the first stacked LSTM layer, the second stacked LSTM layer, and the first nonlinear dense layer, the second nonlinear dense layer and the third nonlinear dense layer are different, although the same algorithm is used.

[0114] In an example, the attention mechanism based long short-term memory recurrent neural network is trained in different working time modes, and the training process in each working time mode includes:

[0115] According to the number of training data sets required for each training round and the number of training rounds, a sufficient number of unprocessed sample sets are extracted from the original data set, the number of training data sets required for each training round is set according to the running memory size of the upper computer, a single training data set includes an encoder input data set, a decoder output data set and a decoder input data set, and a single unprocessed sample set includes the optical load of the optical machine system during the optical path test, the design parameters and the measured optical beam quality time-varying sequence;

[0116] For each unprocessed sample set:

[0117] According to the data set standard format, the optical load of the optical machine system during the optical path test, the design parameters and the current working time mode are processed to create an optical machine system data set to be trained;

[0118] The optical machine system data set to be trained is processed by a linear scaling mapping function and a nonlinear scaling mapping function to obtain an encoder input data set;

[0119] According to the current working time mode, the optical beam quality time-varying sequence to be trained is extracted from the measured optical beam quality time-varying sequence;

[0120] The light beam quality time-varying sequence to be trained is processed by a linear scaling mapping function, and the processed light beam quality time-varying sequence is binarized to obtain a decoder output data set.

[0121] The processed light beam quality time-varying sequence is right-shifted to obtain a decoder input data set.

[0122] The training data set required for each round of training is imported into the long short-term memory recurrent neural network based on the attention mechanism for training, and the training is stopped when the training round is reached.

[0123] It should be noted that the light beam quality time-varying sequence actually measured in a single unprocessed sample set may not conform to the current working time mode, and therefore the light beam quality time-varying sequence to be trained needs to be extracted from the actually measured light beam quality time-varying sequence according to the current working time mode.

[0124] It should be noted that right shifting means discarding the latest data in the last position of the array to obtain historical data.

[0125] The number of training data sets required for each round of training is set according to the size of the host computer running memory, which realizes adaptive host computer running memory, can adapt to host computers with smaller running memory, shortens the light transmission simulation time required for originally calculating the light beam quality time-varying sequence from several weeks to within one hour, and can efficiently guide product design updates.

[0126] Training the long short-term memory recurrent neural network based on the attention mechanism under different working time modes not only helps to prevent overfitting of the long short-term memory recurrent neural network based on the attention mechanism during the training process, but also enables the trained long short-term memory recurrent neural network based on the attention mechanism to have better migration and application characteristics when processing different working time modes. Different working time modes are usually typical working time modes in optical path experiments.

[0127] During the training process, the decoder outputs the light beam quality at the current time each time, and the light beam quality at the current time is automatically added to the end of the decoder input to prompt the long short-term memory recurrent neural network based on the attention mechanism that the light beam quality at the current time generated by the long short-term memory recurrent neural network based on the attention mechanism is in the position of the overall light beam quality time-varying sequence, so that the long short-term memory recurrent neural network based on the attention mechanism can better understand the correlation between the light beam quality time-varying sequence at the current time and the light beam quality time-varying sequence at the previous time, and thus the light beam quality time-varying sequence calculated by the subsequent long short-term memory recurrent neural network based on the attention mechanism has better continuity. At the same time, the light beam quality at the current time is also used for comparison with the light beam quality at the current time in the decoder output data set to fit the weights of the long short-term memory recurrent neural network based on the attention mechanism.

[0128] Table 2 shows the predicted light beam quality time-varying sequence after training based on the measured light beam quality time-varying sequence of the light path experiment, wherein the working time in the working time mode is 10s, the data frame rate is 1Hz, the predicted light beam quality corresponding to each second is the same as the measured light beam quality of the light path experiment to the first decimal place, and the confidence is not less than 90%.

[0129] Table 2

[0130] Time / s 1 2 3 4 5 6 7 8 9 10 Predicted beam quality 1 1.1 1.3 1.2 1.2 1.1 1.2 1.3 1.5 1.4 Measured beam quality 1 1.1 1.3 1.2 1.2 1.1 1.2 1.3 1.5 1.4

[0131] In an example, the training data set required for each round of training is in JSON format.

[0132] In a possible implementation, the training process in each working time mode is as follows:

[0133] The training data set required in the current working time mode is created based on the training round, the number of training data sets required for each round of training and the generator function, and the number of training data sets required for each round of training is set according to the running memory size of the upper computer, and the pseudo code of the generator function is as follows:

[0134] Function single training data set generation function(){

[0135] A single unprocessed sample set is extracted from the original data set, and the single unprocessed sample set includes the optical load and design parameters of the optical-mechanical system during the light path experiment and the measured light beam quality time-varying sequence;

[0136] The optical load and design parameters of the optical-mechanical system during the light path experiment and the current working time mode are processed according to the data set standard format, and the optical-mechanical system data set to be trained is created;

[0137] The optical-mechanical system data set to be trained is processed through a linear scaling mapping function and a nonlinear scaling mapping function to obtain an encoder input data set;

[0138] The light beam quality time-varying sequence to be trained is extracted from the measured light beam quality time-varying sequence according to the current working time mode;

[0139] The light beam quality time-varying sequence to be trained is processed through a linear scaling mapping function, and the processed light beam quality time-varying sequence is binarized to obtain a decoder output data set;

[0140] The processed light beam quality time-varying sequence is right-shifted to obtain a decoder input data set;

[0141] }

[0142] Function repeated call function(){

[0143] if (not reached training round) { repeat calling single training dataset generation function () ;}

[0144] }

[0145] Function training dataset object construction function needed per training round (

[0146] Divide the single training dataset into an encoder input dataset, a decoder output dataset, and a decoder input dataset;

[0147] Assemble the training dataset needed per training round according to the number of training datasets needed per training round;

[0148] Call the repeat calling function () ;

[0149] }

[0150] Import the training dataset needed per training round into the attention mechanism-based long short-term memory recurrent neural network for training, and stop training when the training round is reached.

[0151] The generator function is one of the most flexible ways to create deep learning datasets, with both computing power resources (from small board-level host computers to large cluster servers) and data set preprocessing and data enhancement functions, and can continuously provide computer processor and memory capacity datasets for the training process of the attention mechanism-based long short-term memory recurrent neural network until the training is completed.

[0152] In one example, before importing the training dataset needed per training round into the attention mechanism-based long short-term memory recurrent neural network for training, one or two non-essential design parameters are randomly selected from the encoder input dataset and set to zero.

[0153] Excluding the essential optical load and design parameters in the encoder input dataset, one or two non-essential design parameters are randomly selected and set to zero to simulate network training under conditions where some design parameters are unknown, which not only helps to prevent overfitting of the attention mechanism-based long short-term memory recurrent neural network during training, but also makes the trained attention mechanism-based long short-term memory recurrent neural network have better migration and application characteristics when encountering new optical and mechanical systems.

[0154] In one example, the non-essential design parameters include environmental temperature, lens physical diameter, lens curvature, optical medium length, and mirror film layer light-heat conversion rate.

[0155] In one example, the light beam quality time-varying sequence efficient calculation method provided by the embodiment is implemented based on a host computer equipped with a general-purpose GPU.

[0156] The above scheme realizes efficient and accurate calculation of the time-varying sequence of the optical beam quality of the optical-mechanical link under the condition of the specified optical-mechanical system optical load and design parameters; can maximize the non-expected contraction and expansion characteristics of the time-varying sequence of the optical-mechanical link module and the measured transmission optical beam quality frequently encountered in real optical path tests, is suitable for full-size optical links, is compatible with the training and calculation process under the scene of the absence of part of the measured data, can achieve the expected convergence effect in a certain scale of data set, and can be expanded to tens to hundreds of optical, thermal control and other components with accuracy and adaptability, can describe parameters including plane / curved transmissive and reflective mirror, plane / channel diaphragm, air inlet / outlet, optical-mechanical internal channel profile and optical load, can achieve the expected convergence effect in a certain scale of data set, and meets the analysis requirements of optical-mechanical products of various platforms.

[0157] The above scheme has been verified in multiple optical-mechanical products, and the effect is good, which makes an important contribution to guiding the efficient optimization and iterative design of optical-mechanical products and saving project time and economic cost.

[0158] Based on the same technical concept, Figure 4 The structure of the optical beam quality time-varying sequence efficient calculation device provided by the embodiment of the application is shown, which comprises a creating module, a processing module and a prediction module, wherein:

[0159] The creating module is used for processing the optical load and design parameters and working time mode of the optical-mechanical system according to a data set standard format, creating an optical-mechanical system data set to be predicted, the data set standard format is used for determining the sorting and data structure of each data in the optical-mechanical system data set, the data structure is an array, and the working time mode comprises working time length and data frame rate;

[0160] The processing module is used for processing the optical-mechanical system data set to be predicted through a linear scaling mapping function and a nonlinear scaling mapping function, to obtain a prediction data set;

[0161] The prediction module is used for inputting the prediction data set into a long short-term memory recurrent neural network based on an attention mechanism which is trained, and outputting an optical beam quality time-varying sequence conforming to the working time mode, the long short-term memory recurrent neural network based on the attention mechanism is an encoder and decoder architecture, the input of the encoder is the prediction data set, and the output of the decoder is the optical beam quality time-varying sequence.

[0162] Based on the same technical concept, the embodiment of the application provides a computing device, which comprises:

[0163] A memory is used for storing program instructions;

[0164] A processor is used for calling the program instructions stored in the memory, and performing the above method according to the obtained program execution.

[0165] Based on the same technical concept, the embodiment of the present application provides a computer readable nonvolatile storage medium comprising computer readable instructions, which, when read and executed by a computer, enable the computer to perform the above method.

[0166] The present application is described in reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device that implements the flow Figure 1 one or more flows and / or blocks Figure 1 an apparatus that performs the functions specified in one or more blocks or flows.

[0167] These computer program instructions can also be stored in a computer readable memory that can direct the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including instruction apparatus, which implements the flow Figure 1 one or more flows and / or blocks Figure 1 an apparatus that performs the functions specified in one or more blocks or flows.

[0168] These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to produce a computer implemented process, so that the instructions executed on the computer or other programmable device provide a process for implementing the flow Figure 1 one or more flows and / or blocks Figure 1 an apparatus that performs the functions specified in one or more blocks or flows.

[0169] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications to the embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0170] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for high efficiency calculation of time-varying sequence of beam quality, characterized in that, The method comprises: processing the optical load and design parameters and working time mode of the optical mechanical system according to a data set standard format, the data set standard format being used to determine the order and data structure of each data in the optical mechanical system data set, the data structure being an array, and the working time mode including working time length and data frame rate; processing the optical mechanical system data set to be predicted through a linear scaling mapping function and a nonlinear scaling mapping function to obtain a prediction data set; inputting the prediction data set into a trained long short-term memory recurrent neural network based on an attention mechanism to output an optical beam quality time-varying sequence conforming to the working time mode, the long short-term memory recurrent neural network based on the attention mechanism being an encoder-decoder architecture, the input of the encoder being the prediction data set, and the output of the decoder being the optical beam quality time-varying sequence.

2. The method of claim 1, wherein, The optical load and design parameters of the optical mechanical system include power, power density, relative humidity, air pressure, ambient temperature, incident angle, optical axis elevation angle, clear aperture, element type code, lens physical diameter, lens curvature, clear medium length, and mirror film layer light-heat conversion rate, wherein: The power and the power density belong to the optical load; The relative humidity, the air pressure, the ambient temperature, the incident angle, the optical axis elevation angle, the clear aperture, the element type code, the lens physical diameter, the lens curvature, the clear medium length, and the mirror film layer light-heat conversion rate belong to the design parameters.

3. The method of claim 1, wherein, When there is data not belonging to the optical load and the design parameters in the data of the optical beam quality time-varying sequence predicted by a specific optical element in the optical mechanical system, the data not belonging to the optical load and the design parameters are multiplexed with an optical load or a design parameter similar in physical meaning in the same array.

4. The method of claim 2, wherein, The processing of the optical mechanical system data set to be predicted through the linear scaling mapping function and the nonlinear scaling mapping function comprises: processing the power, the power density, the relative humidity, the air pressure, the ambient temperature, the incident angle, the optical axis elevation angle, the clear aperture, the element type code, the lens physical diameter, the lens curvature, the mirror film layer light-heat conversion rate, and the working time mode through the linear scaling mapping function, the expression of the linear scaling mapping function being as follows: wherein round() is a rounding function, [x min ,x max ] is the original range, and [x′ max ,x′ max ] is the scaled range; processing the clear medium length through the nonlinear scaling mapping function, the expression of the nonlinear scaling mapping function being as follows: wherein p i , q i are adjustable parameters, set according to the optical-mechanical system.

5. The method of claim 1, wherein, The encoder-decoder architecture comprises: The encoder comprises a first data embedding layer, a first stacked LSTM layer, a final state extraction layer, a context correlation layer, an attention concatenation layer, and a first nonlinear dense layer, and the decoder comprises a second data embedding layer, a second stacked LSTM layer, a correlation dot product layer, a second nonlinear dense layer, and a third nonlinear dense layer, wherein: The first data embedding layer and the second data embedding layer are respectively used to encode the input of the encoder and the decoder into a vector. The first stacked LSTM layer and the second stacked LSTM layer are respectively used to increase dimensions of the encoder and the decoder; The last state extraction layer is used to extract a last state of a last layer in the first stacked LTSM layer and to initialize an attention weight of a last layer in the second stacked LTSM layer, the last state being data obtained by slicing along a second dimension in a three-dimensional tensor; The correlation dot product layer is used to perform a dot product on outputs of the last layer in the first stacked LSTM layer and the last layer in the second stacked LSTM layer; The first nonlinear dense layer, the second nonlinear dense layer and the third nonlinear dense layer are respectively used to process outputs of the attention concatenation layer, the correlation dot product layer and the first nonlinear dense layer according to a nonlinear function; The context correlation layer is used to perform a dot product on the output of the last layer in the first stacked LSTM layer and the output of the second nonlinear dense layer; The attention concatenation layer is used to concatenate the output of the context correlation layer and the output of the last layer in the second stacked LSTM layer along a third dimension in the three-dimensional tensor; The output of the third nonlinear dense layer is added at the end of the input of the second data embedding layer.

6. The method of claim 1, wherein, The attention mechanism-based long short-term memory recurrent neural network is trained in different working time modes, and the training process in each working time mode includes: Based on a training round, a number of training data sets required for each round of training and a generator function, a training data set required in the current working time mode is created, the number of training data sets required for each round of training is set according to the running memory size of the host computer, and the pseudo code of the generator function is: Function single training data set generation function(){ A single unprocessed sample set is extracted from an original data set, the single unprocessed sample set including an optical load of an optical-mechanical system during an optical path test, design parameters and a measured time-varying sequence of beam quality; The optical load of the optical-mechanical system during the optical path test, the design parameters and the current working time mode are processed according to the data set standard format, and an optical-mechanical system data set to be trained is created; The optical-mechanical system data set to be trained is processed through the linear scaling mapping function and the nonlinear scaling mapping function to obtain an encoder input data set; A time-varying sequence of beam quality to be trained is extracted from the measured time-varying sequence of beam quality according to the current working time mode; The time-varying sequence of beam quality to be trained is processed through the linear scaling mapping function, and the processed time-varying sequence of beam quality is binarized to obtain a decoder output data set; The processed time-varying sequence of beam quality is right-shifted to obtain a decoder input data set; } Function repeat call function(){ if (the training round has not been reached), the single training data set generation function is repeatedly called; } Function each round of training training data set object construction function() dividing the single training dataset into the encoder input dataset, the decoder output dataset, and the decoder input dataset; grouping the training dataset required for each training round according to the number of training datasets required for each training round; calling the repeated calling function (). } training the attention mechanism based long short-term memory recurrent neural network with the training dataset required for each training round, and stopping the training when the training round is reached.

7. The method of claim 6, wherein, Before the training of the attention mechanism based long short-term memory recurrent neural network with the training dataset required for each training round, randomly selecting one or two non-essential design parameters from the encoder input dataset to be zeroed.

8. A high-efficiency computing device for time-varying sequence of beam quality, characterized in that, comprising a creating module, a processing module and a predicting module, wherein: the creating module is configured to process the optical load and the design parameters and the working time mode of the optical-mechanical system according to a dataset standard format to create an optical-mechanical system dataset to be predicted, the dataset standard format is used to determine the order and data structure of each data in the optical-mechanical system dataset, the data structure is an array, and the working time mode comprises working time length and data frame rate; the processing module is configured to process the optical-mechanical system dataset to be predicted through a linear scaling mapping function and a nonlinear scaling mapping function to obtain a prediction dataset; the predicting module is configured to input the prediction dataset into a trained attention mechanism based long short-term memory recurrent neural network to output an optical beam quality time-varying sequence conforming to the working time mode, the attention mechanism based long short-term memory recurrent neural network is an encoder and decoder architecture, the input of the encoder is the prediction dataset, and the output of the decoder is the optical beam quality time-varying sequence.

9. A computing device, comprising: comprise: a memory for storing program instructions; a processor for calling the program instructions stored in the memory, and executing the method according to any one of claims 1-7 according to the obtained program execution right.

10. A computer-readable non-transitory storage medium, characterized in that, comprise computer readable instructions, when the computer reads and executes the computer readable instructions, the computer executes the method according to any one of claims 1-7. comprise computer readable instructions, when the computer reads and executes the computer readable instructions, the computer executes the method according to any one of claims 1-7.

Images