Deformable mirror control method and device capable of quickly adapting to actual application environment

Abstract

The invention discloses a deformable mirror control method and device capable of quickly adapting to an actual application environment, which can enhance the field self-adaption and field generalization ability of an existing control model under the condition of low training cost, flexibly adapt to the actual application environment and realize high-precision control and wavefront compensation of a deformable mirror. The method comprises the following steps: (1) complementarily collecting a small amount of deformable mirror voltage and surface shape data to construct a migration data set; (2) evaluating the data set similarity degree of the initial modeling environment of the source domain and the actual application environment of the target domain; (3) determining a deformable mirror control model migration scheme; (4) training the network to realize the migration of the deformable mirror control model; and (5) evaluating and verifying the migration effect of the deformable mirror control model.

Description

Technical Field

[0001] The present invention relates to the technical field of photoelectric detection, and in particular to a deformable mirror control method that can quickly adapt to an actual application environment, and a deformable mirror control device that can quickly adapt to an actual application environment. Background Art

[0002] Deformable mirror (DM), also known as wavefront corrector, is one of the core components in adaptive optical systems and has important applications in astronomical observation, biomedicine, laser beam shaping, long-distance communication, free-form surface detection and other fields.

[0003] The deformable mirror mainly includes a surface composed of a voltage-controlled driver and a flexible metal layer. By setting the control driver voltage according to the surface shape information to be compensated, the DM surface shape can be controlled to change the optical path of wavefront transmission to dynamically correct the wavefront of the optical system. It can be seen that in the above process, the setting of the control voltage directly affects the wavefront correction accuracy of the deformable mirror. The role of the automatic control method of the deformable mirror is to automatically output the control voltage according to the characteristic vector of the target surface shape. Therefore, the accuracy and efficiency of the control method determine the accuracy and efficiency of the deformable mirror wavefront compensation. In the application of free-form surface detection, the free-form surface to be measured can be used as the target surface shape input control system to drive the deformable mirror, so that the deformable mirror can be used as a dynamic compensation mirror to achieve zero-difference interference with the surface to be measured, and the dynamic interference measurement of the free-form surface shape can be achieved without replacing the compensation mirror multiple times. At this time, the accuracy and efficiency of the deformable mirror control method also directly determine the accuracy and efficiency of free-form surface detection.

[0004] According to the above analysis, the high rate and high precision of the deformable mirror control method have an important impact on the development and application requirements of the deformable mirror. However, it is difficult for the existing deformable mirror control method to meet the requirements of control accuracy and efficiency at the same time. Commonly used deformable mirror control methods include optimization control methods and model-based control methods. When there is no pre-collected data, a search optimization control method can be used, such as a hill climbing algorithm, a random parallel gradient descent algorithm, etc. This method only optimizes the control voltage according to the compensation effect of the deformable mirror. Although it does not require advance modeling, the convergence speed is slow, so the control efficiency is very low. In order to achieve higher control efficiency, a model-based control method is often used. However, it is very difficult to accurately obtain the control model using traditional methods. For example, the commonly used physical modeling solution method contains many necessary approximate conditions that limit the accuracy. The method of solving the control model by optimizing the influence function matrix cannot fully characterize the problems of driver coupling, nonlinear response, etc., and each control needs to be re-optimized, which is time-consuming and has low control efficiency. Therefore, a deformable mirror control method that can obtain the control model without approximation and optimization is necessary for efficient and high-precision control of the deformable mirror. Neural networks have a powerful ability to fit complex multi-dimensional nonlinear functions and can directly build high-precision control models. The model builds the data set based on the fitting learning of the data set without any approximate conditions. And the trained neural network control model can directly output the control voltage as long as the surface data is input, without the need for optimization process. Therefore, the control method based on neural networks has advantages in efficiency and accuracy.

[0005] However, the existing control methods based on any neural network have very strict requirements on the stability of the deformable mirror application environment, that is, all interference factors such as temperature, humidity, disturbance, etc. in the actual application environment must be consistent with the environmental factors when the data set was collected. This is because the deformable mirror driver is very sensitive to temperature and humidity. Under different environmental conditions, the deformable mirror shape produced by the same driving voltage is significantly different, which also causes the wavefront compensation amount of the deformable mirror to change significantly. The general actual application environment is often different from the existing model data set environment. At this time, because the trained existing model cannot adapt to the target environment, the wavefront compensation effect of the deformable mirror is often greatly reduced.

[0006] In order to adapt to the actual application environment, it is a feasible method to collect data and retrain the model in the current environment. However, in order to ensure the model effect, model training requires a large amount of newly collected data and a long and complex training optimization process, which is time-consuming and labor-intensive. Moreover, completely abandoning the existing models trained in other environments is a huge waste of time and resources. Therefore, there is an urgent need for a control method to improve the adaptability of the existing model at a very low time and resource cost, so that it can adapt to the actual application environment and achieve high-precision deformable mirror control. Summary of the invention

[0007] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to provide a deformable mirror control method that can quickly adapt to the actual application environment, which can enhance the domain adaptation and domain generalization capabilities of the existing control model at a low training cost, flexibly adapt to the actual application environment, and realize high-precision control and wavefront compensation of the deformable mirror.

[0008] The technical solution of the present invention is: the deformable mirror control method that can quickly adapt to the actual application environment comprises the following steps:

[0009] (1) Sampling a small amount of deformable mirror voltage and surface shape data to build a migration data set: In actual application scenarios, select several sets of typical control voltages to drive the deformable mirror and measure the surface shape. These small amounts of voltage and surface shape data collected in the application scenarios serve as the target data set of the deformable mirror.

[0010] The target domain dataset is divided into a support set and a query set;

[0011] (2) Evaluate the similarity between the datasets of the initial modeling environment in the source domain and the actual application environment in the target domain.

[0012] Degree: First, determine the similarity metric, and then directly calculate the metric using the target domain data of the actual application environment and the source domain data of the initial modeling environment;

[0013] (3) Determine the type of transfer learning scheme for the deformable mirror control model: The basis for selecting the control is the amount of data in the migration data set and the similarity between the target environment and the source environment of the actual application of the deformable mirror; select a fine-tuning scheme for the control of the deformable mirror under general environmental conditions, and combine the migration channel clipping method to improve the effect of transfer learning, ensure control accuracy and minimize the increase in training costs;

[0014] (4) Training the network to realize the migration of the deformable mirror control model: According to the fine-tuning scheme selected in step (3), the first x layers of the source domain model trained in the source domain environment are frozen, and only a small amount of supplementary data is used to train the remaining y layers of the network; the nonlinear relationship between the surface shape and voltage that is common to the deformable mirror control model in any environment is retained by using the first x layers.

[0015] Linear fitting features, and use the unfrozen back y layers of the network to extract individual features related to the actual application environment; Combined with the migration channel pruning scheme, fine-tune the source domain model, and then use the migration channel evaluation to evaluate the importance of different channels in all network layers, prune the less important channels, and iterate until a balance between accuracy and computational overhead is achieved;

[0016] (5) Evaluation and verification of the migration effect of the deformable mirror control model: First, determine the two control accuracy evaluation indicators of the deformable mirror control voltage and surface compensation accuracy, and conduct a comprehensive accuracy analysis of the migration model and the source domain model in the target domain dataset, that is, the actual application environment of the deformable mirror. Based on the accuracy comparison results, evaluate the model migration effect and adjust the migration plan.

[0017] The present invention establishes a deformable mirror control model that can quickly adapt to the actual application environment through the steps of target domain data set establishment, similarity evaluation, migration scheme selection and implementation, and model migration effect evaluation. The present invention enhances the domain adaptation and domain generalization capabilities of the existing control model at a low training cost through model migration schemes such as fine-tuning, and flexibly adapts to the actual application environment. The migration channel clipping scheme further improves the control model migration effect when the environment difference is large, making the applicable application environment wider and the control accuracy higher. Therefore, the present invention can enhance the domain adaptation and domain generalization capabilities of the existing control model at a low training cost, flexibly adapt to the actual application environment, and realize high-precision control and wavefront compensation of the deformable mirror.

[0018] A deformable mirror control device that can quickly adapt to actual application environments is also provided, comprising:

[0019] The supplementary sampling module is configured to supplement a small amount of deformable mirror voltage and surface shape data to build a migration data set: in the actual application scenario, several sets of typical control voltages are selected to drive the deformable mirror and measure the surface shape. These small amounts of voltage and surface shape data supplemented in the application scenario are used as the target domain data set of the deformable mirror, and the target domain data set is divided into a support set and a query set;

[0020] The evaluation module is configured to evaluate the similarity between the datasets of the source domain initial modeling environment and the target domain actual application environment: first, a similarity metric is determined, and then the metric is directly calculated using the target domain data of the actual application environment and the source domain data of the initial modeling environment;

[0021] Determine the module whose configuration determines the type of scheme for transfer learning of the deformable mirror control model:

[0022] The basis for selecting control is the amount of data in the migration data set and the similarity between the target environment and the source environment of the actual application of the deformable mirror; a fine-tuning scheme is selected for the control of the deformable mirror under general environmental conditions, and the migration channel clipping method is combined to improve the effect of transfer learning, ensure control accuracy and minimize the increased training cost;

[0023] A migration module is configured to train the network to realize the migration of the deformable mirror control model: according to the selected fine-tuning scheme, the first x layers of the source domain model trained in the source domain environment are frozen, and only a small amount of supplementary data is used to train the remaining y layers of the network; the first x layers are used to retain the nonlinear fitting characteristics between the common surface shape and voltage of the deformable mirror control model in any environment, and the unfrozen rear y layers of the network are used to extract the individual characteristics related to the actual application environment; combined with the migration channel pruning scheme, the source domain model is fine-tuned, and then the migration channel evaluation is used to evaluate the importance of different channels of all network layers, and the less important channels are pruned, and iterative adjustments are made until a balance between accuracy and computational overhead is achieved;

[0024] The verification module is configured to evaluate and verify the migration effect of the deformable mirror control model: first determine the two control accuracy evaluation indicators of the deformable mirror control voltage and surface compensation accuracy, and conduct a comprehensive accuracy analysis of the model testing and experimental verification of the migration model and the source domain model in the target domain dataset, that is, the actual application environment of the deformable mirror. According to the accuracy comparison results, the model migration effect is evaluated and the migration plan is adjusted. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of a deformable mirror control method that can quickly adapt to actual application environment according to the present invention.

[0026] Figure 2 It is a flow chart of the migration channel pruning method. DETAILED DESCRIPTION

[0027] like Figure 1 As shown, the deformable mirror control method that can quickly adapt to the actual application environment includes the following steps:

[0028] (1) Sampling a small amount of deformable mirror voltage and surface shape data to build a migration data set: In actual application scenarios, select several sets of typical control voltages to drive the deformable mirror and measure the surface shape. These small amounts of voltage and surface shape data collected in the application scenarios serve as the target data set of the deformable mirror.

[0029] The target domain dataset is divided into a support set and a query set;

[0030] (2) Evaluate the similarity between the datasets of the initial modeling environment in the source domain and the actual application environment in the target domain.

[0031] Degree: First, determine the similarity metric, and then directly calculate the metric using the target domain data of the actual application environment and the source domain data of the initial modeling environment;

[0032] (3) Determine the type of transfer learning scheme for the deformable mirror control model: The basis for selecting the control is the amount of data in the migration data set and the similarity between the target environment and the source environment of the actual application of the deformable mirror; select a fine-tuning scheme for the control of the deformable mirror under general environmental conditions, and combine the migration channel clipping method to improve the effect of transfer learning, ensure control accuracy and minimize the increase in training costs;

[0033] (4) Training the network to realize the migration of the deformable mirror control model: According to the fine-tuning scheme selected in step (3), the first x layers of the source domain model trained in the source domain environment are frozen, and only a small amount of supplementary data is used to train the remaining y layers of the network; the nonlinear relationship between the surface shape and voltage that is common to the deformable mirror control model in any environment is retained by using the first x layers.

[0034] Linear fitting features, and use the unfrozen back y layers of the network to extract individual features related to the actual application environment; Combined with the migration channel pruning scheme, fine-tune the source domain model, and then use the migration channel evaluation to evaluate the importance of different channels in all network layers, prune the less important channels, and iterate until a balance between accuracy and computational overhead is achieved;

[0035] (5) Evaluation and verification of the migration effect of the deformable mirror control model: First, determine the two control accuracy evaluation indicators of the deformable mirror control voltage and surface compensation accuracy, and conduct a comprehensive accuracy analysis of the migration model and the source domain model in the target domain dataset, that is, the actual application environment of the deformable mirror. Based on the accuracy comparison results, evaluate the model migration effect and adjust the migration plan.

[0036] The present invention establishes a deformable mirror control model that can quickly adapt to the actual application environment through the steps of target domain data set establishment, similarity evaluation, migration scheme selection and implementation, and model migration effect evaluation. The present invention enhances the domain adaptation and domain generalization capabilities of the existing control model at a low training cost through model migration schemes such as fine-tuning, and flexibly adapts to the actual application environment. The migration channel clipping scheme further improves the control model migration effect when the environment difference is large, making the applicable application environment wider and the control accuracy higher. Therefore, the present invention can enhance the domain adaptation and domain generalization capabilities of the existing control model at a low training cost, flexibly adapt to the actual application environment, and realize high-precision control and wavefront compensation of the deformable mirror.

[0037] Preferably, before the step (1), the method further includes obtaining a source domain model for deformable mirror control: driving the deformable mirror and measuring the deformable mirror surface shape under environmental conditions where a large amount of data can be collected, recording the control voltage and the corresponding surface shape data, and establishing a sufficient source domain data set; selecting a neural network structure to establish a deformable mirror control model, and performing optimization training; the deformable mirror control model trained with a sufficient source domain data set is the source domain model; the source domain refers to a field that has been completed and contains knowledge in transfer learning, and the source domain model has good control accuracy and sufficient knowledge for reuse and transfer.

[0038] Preferably, in step (1), the target domain refers to the field to be migrated, which is the environmental conditions in which the deformable mirror is actually used, and is different from the environmental conditions of the source domain; the support set refers to the samples in the small amount of supplementary data used to train the target domain model; and the query set refers to the samples in the small amount of supplementary data used to test the target domain model.

[0039] Preferably, in step (2), the measurement indicator of similarity is the maximum mean difference (MMD), which is the distance between two distributions in the regenerated Hilbert space; the basic characteristics of the deformable mirror refer to the stability and repeatability of the deformable mirror with changes in environmental factors such as temperature and humidity.

[0040] Preferably, in step (3), the model migration scheme includes fine-tuning and retaining initialization weights; fine-tuning is to directly retain some features learned by the network from the source domain model by freezing the parameters of the first several layers; retaining initialization weights is to retain the network's feature expression ability for the basic control model of the deformable mirror by using the source domain model parameters as initialization parameters; when the similarity gap between the source domain environmental conditions and the target domain is small, or the target domain data set is small, the fine-tuning migration method is selected; when the scale of the supplementary data set is large and the similarity gap between the target domain and the source domain environmental conditions is large, the migration method of retaining initialization weights is selected.

[0041] Preferably, in step (4), the number of network layers y retrained during fine-tuning depends on two parameters: the amount of supplementary data and the similarity between the source domain and the target domain; the larger the amount of supplementary data, the more network layers can be trained, the larger the y value, and the better the performance of the network in the target environment; the lower the similarity between the source domain and the target domain, the more individual features are possessed, the more network layers need to be trained, and the larger the y value; the migration channel pruning in step (4) is to reduce the difference between the characteristics of the target data set and the characteristics of the source domain data while pruning those unimportant channels.

[0042] Preferably, in the step (5), the voltage accuracy is represented by the root mean square, mean absolute deviation and relative deviation of the model output value and the true value of the data set, and is determined by analyzing the difference between the model prediction value and the true value through an algorithm. The surface shape control accuracy index is the peak-to-peak value pv and root mean square rms of the difference surface shape, the voltage accuracy of the model test, and is given by the surface shape difference after the deformable mirror is driven by the true voltage and the predicted voltage.

[0043] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes the steps of the above-mentioned embodiment method, and the storage medium can be: ROM / RAM, disk, CD, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a deformable mirror control device that can quickly adapt to the actual application environment. The device is usually represented in the form of functional modules corresponding to the steps of the method. The device includes:

[0044] The supplementary sampling module is configured to supplement a small amount of deformable mirror voltage and surface shape data to build a migration data set: in the actual application scenario, several sets of typical control voltages are selected to drive the deformable mirror and measure the surface shape. These small amounts of voltage and surface shape data supplemented in the application scenario are used as the target domain data set of the deformable mirror, and the target domain data set is divided into a support set and a query set;

[0045] The evaluation module is configured to evaluate the similarity between the datasets of the source domain initial modeling environment and the target domain actual application environment: first, a similarity metric is determined, and then the metric is directly calculated using the target domain data of the actual application environment and the source domain data of the initial modeling environment;

[0046] Determine the module whose configuration determines the type of scheme for transfer learning of the deformable mirror control model:

[0047] The basis for selecting control is the amount of data in the migration data set and the similarity between the target environment and the source environment of the actual application of the deformable mirror; a fine-tuning scheme is selected for the control of the deformable mirror under general environmental conditions, and the migration channel clipping method is combined to improve the effect of transfer learning, ensure control accuracy and minimize the increased training cost;

[0048] A migration module is configured to train the network to realize the migration of the deformable mirror control model: according to the selected fine-tuning scheme, the first x layers of the source domain model trained in the source domain environment are frozen, and only a small amount of supplementary data is used to train the remaining y layers of the network; the first x layers are used to retain the nonlinear fitting characteristics between the common surface shape and voltage of the deformable mirror control model in any environment, and the unfrozen rear y layers of the network are used to extract the individual characteristics related to the actual application environment; combined with the migration channel pruning scheme, the source domain model is fine-tuned, and then the migration channel evaluation is used to evaluate the importance of different channels of all network layers, and the less important channels are pruned, and iterative adjustments are made until a balance between accuracy and computational overhead is achieved;

[0049] The verification module is configured to evaluate and verify the migration effect of the deformable mirror control model: first determine the two control accuracy evaluation indicators of the deformable mirror control voltage and surface compensation accuracy, and conduct a comprehensive accuracy analysis of the model testing and experimental verification of the migration model and the source domain model in the target domain dataset, that is, the actual application environment of the deformable mirror. According to the accuracy comparison results, the model migration effect is evaluated and the migration plan is adjusted.

[0050] Preferably, the device also includes an acquisition module, which is configured to obtain a source domain model for deformable mirror control: under environmental conditions where a large amount of data can be collected, the deformable mirror is driven and the deformable mirror surface shape is measured, the control voltage and the corresponding surface shape data are recorded, and a sufficient source domain data set is established; a neural network structure is selected to establish a deformable mirror control model, and optimization training is performed; the deformable mirror control model trained with a sufficient source domain data set is the source domain model; the source domain refers to a field that has been completed and contains knowledge in transfer learning, and the source domain model has good control accuracy capabilities and sufficient knowledge for reuse and migration.

[0051] Preferably, in the migration module, the number of network layers y retrained during fine-tuning depends on two parameters: the amount of supplementary data and the similarity between the source domain and the target domain; the larger the amount of supplementary data, the more network layers can be trained, the larger the y value, and the better the performance of the network in the target environment; the lower the similarity between the source domain and the target domain, the more individual features it has, the more network layers need to be trained, and the larger the y value; the migration channel pruning in step (4) is to reduce the difference between the characteristics of the target data set and the characteristics of the source domain data while pruning those unimportant channels;

[0052] In the verification module, the voltage accuracy is expressed by the root mean square, mean absolute deviation, and relative deviation of the model output value and the true value of the data set, and is determined by the difference between the model prediction value and the true value through algorithm analysis. The surface shape control accuracy index is the peak-to-peak value pv and root mean square rms of the difference surface shape, the voltage accuracy of the model test, and is given by the difference in surface shape after the deformable mirror is driven by the true voltage and the predicted voltage.

[0053] A specific implementation example of the present invention is described in detail below. This deformable mirror control method that quickly adapts to the actual application environment includes the following steps:

[0054] (I) Obtain a source domain model for deformable mirror control. Drive the deformable mirror and measure the deformable mirror surface shape under environmental conditions where a large amount of data can be collected, record the control voltage and the corresponding surface shape data, and establish a sufficient source domain data set. Select any effective neural network structure to establish a deformable mirror control model and perform optimization training. The deformable mirror control model trained with a sufficient source domain data set is the source domain model. If there is already a control model trained in any environment, there is no need to repeat the training in this step, and it can be directly used as the initial model.

[0055] (II) A small amount of supplementary data is collected to construct a migration dataset. In the actual application scenario, several sets of typical control voltages are selected to drive the deformable mirror and measure the surface shape. These small amounts of supplementary voltage and surface shape data collected in the application scenario are used as the target domain dataset of the deformable mirror. The target domain dataset is divided into a support set and a query set.

[0056] (III) Evaluate the similarity between the source domain dataset and the migration dataset. First, determine the similarity metric, and then directly calculate the metric using the target domain data and the source domain data. In addition to accurate quantitative calculations, qualitative evaluation of the similarity between the source domain and the target domain is also important. Specifically, it can be estimated by detecting the basic characteristics of the deformable mirror, or by using the source domain model to test the control accuracy of the deformable mirror in the target environment of actual application.

[0057] (IV) Determine the type of transfer learning scheme. The basis for selecting the control is the amount of data in the migration dataset and the similarity between the target environment and the source environment of the actual application of the deformable mirror. For the control of the deformable mirror in general environmental conditions, a fine-tuning scheme is usually selected, and the transfer channel clipping method is combined to improve the effect of transfer learning, ensure control accuracy and minimize the increase in training costs.

[0058] (V) Train the network to achieve model migration. According to the fine-tuning scheme selected in step (IV), the first x layers of the source domain model trained in the source domain environment are frozen, and only a small amount of supplementary data is used to train the remaining y layers of the network. The first x layers are used to retain the nonlinear fitting characteristics between the common surface shape and voltage of the deformable mirror control model in any environment, and the unfrozen rear y layers of the network are used to extract individual characteristics related to the actual application environment. Combined with the migration channel pruning scheme, the source domain model needs to be fine-tuned, and then the migration channel evaluation is used to evaluate the importance of different channels of all network layers, prune the less important channels, and iterate until a balance between accuracy and computational overhead is achieved.

[0059] (VI) Evaluation and verification of model migration effect. First, determine the two control accuracy evaluation indicators of the deformable mirror control voltage and surface compensation accuracy, and conduct a comprehensive accuracy analysis of the model testing and experimental verification of the migration model and the source domain model in the target domain dataset, i.e., the actual application environment of the deformable mirror. Based on the accuracy comparison results, evaluate the model migration effect and adjust the migration plan.

[0060] The beneficial effects of the present invention are as follows:

[0061] 1. The deformable mirror control method and device disclosed in the present invention can quickly adapt to the actual application environment. The method and device adopt a fine-tuning model migration scheme, retain the common characteristics of accurate neural network model parameter learning trained by existing data sets, and only add a small amount of data for reuse and migration, so that learning from scratch can be transformed into accumulative learning, which greatly reduces the cost of adapting the deformable mirror control model to the actual application environment.

[0062] 2. The deformable mirror control method and device disclosed in the present invention can quickly adapt to the actual application environment. The migration channel clipping scheme is used to assist the transfer learning process, which further enhances the adaptability and control accuracy of the deformable mirror control model to different application fields, so that the model can achieve high-precision control while having domain adaptation and domain generalization capabilities.

[0063] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the protection scope of the technical solution of the present invention.

Claims

1. A deformable mirror control method that can quickly adapt to actual application environments, Features: It includes the following steps: (1) A small amount of deformable mirror voltage and surface shape data is collected to build a migration data set: In actual application scenarios, several sets of typical control voltages are selected to drive the deformable mirror and measure the surface shape. These small amounts of voltage and surface shape data collected in the application scenarios are used as the target domain data set of the deformable mirror, and the target domain data set is divided into a support set and a query set. (2) Evaluate the similarity between the datasets of the source domain initial modeling environment and the target domain actual application environment: First, determine the similarity metric, and then directly calculate the metric using the target domain data of the actual application environment and the source domain data of the initial modeling environment; (3) Determine the type of transfer learning scheme for the deformable mirror control model: The basis for selecting the control is the amount of data in the migration data set and the similarity between the target environment and the source environment of the actual application of the deformable mirror; select a fine-tuning scheme for the control of the deformable mirror under general environmental conditions, and combine the migration channel clipping method to improve the effect of transfer learning, ensure control accuracy and minimize the increase in training costs; (4) Training the network to realize the transfer of the deformable mirror control model: According to the fine-tuning scheme selected in step (3), the first x layers of the source domain model trained in the source domain environment are frozen. Only a small amount of supplementary data is used to train the remaining y-layer network; the nonlinear fitting characteristics between the common surface shape and voltage of the deformable mirror control model in any environment are retained by the first x layers, and the individual characteristics related to the actual application environment are extracted by the unfrozen rear y-layer network; Combined with the migration channel pruning solution, the source domain model is fine-tuned. Then, the migration channel evaluation is used to evaluate the importance of different channels in all network layers, prune the less important channels, and iteratively adjust until a balance between accuracy and computational overhead is achieved; (5) Evaluation and verification of the migration effect of the deformable mirror control model: First, determine the two control accuracy evaluation indicators of the deformable mirror control voltage and surface compensation accuracy, and conduct a comprehensive accuracy analysis of the migration model and the source domain model in the target domain dataset, that is, the actual application environment of the deformable mirror. Based on the accuracy comparison results, evaluate the model migration effect and adjust the migration plan.

2. The method for controlling a deformable mirror that can quickly adapt to actual application environments according to claim 1, Features: Prior to step (1), the method further includes obtaining a source domain model for deformable mirror control: driving the deformable mirror and measuring the deformable mirror surface shape under environmental conditions where a large amount of data can be collected, recording the control voltage and the corresponding surface shape data, and establishing a sufficient source domain data set; selecting a neural network structure to establish a deformable mirror control model, and performing optimization training; the deformable mirror control model trained with a sufficient source domain data set is the source domain model; the source domain refers to a field that has been learned and contains knowledge in transfer learning, and the source domain model has good control accuracy and sufficient knowledge for reuse and transfer.

3. The method for controlling a deformable mirror that can quickly adapt to actual application environments according to claim 2, Features: In the step (1), the target domain refers to the field to be migrated, which is the environmental conditions in which the deformable mirror is actually used, and is different from the environmental conditions of the source domain; the support set refers to the samples in the small amount of supplementary data used to train the target domain model; and the query set refers to the samples in the small amount of supplementary data used to test the target domain model.

4. The method for controlling a deformable mirror that can quickly adapt to actual application environments according to claim 3, Features: In the step (2), the measurement index of similarity is the maximum mean difference (MMD), which is the distance between two distributions in the regenerated Hilbert space; the basic characteristics of the deformable mirror refer to the stability and repeatability of the deformable mirror with changes in environmental factors such as temperature and humidity.

5. The method for controlling a deformable mirror that can quickly adapt to actual application environments according to claim 4, Features: In step (3), the model migration scheme includes fine-tuning and retaining the initialization weights; the fine-tuning directly retains some features learned by the network from the source domain model by freezing the parameters of the first several layers; Retaining the initialization weights is to retain the network's ability to express the characteristics of the basic control model of the deformable mirror by using the source domain model parameters as initialization parameters; when the similarity gap between the source domain environmental conditions and the target domain is small, or the target domain data set is small, the fine-tuning migration method is selected; when the scale of the supplementary data set is large and the similarity gap between the target domain and the source domain environmental conditions is large, the migration method of retaining the initialization weights is selected.

6. The method for controlling a deformable mirror that can quickly adapt to actual application environments according to claim 5, Features: In the step (4), the number of network layers y retrained during fine-tuning depends on two parameters: the amount of supplementary data and the similarity between the source domain and the target domain; the larger the amount of supplementary data, the more network layers can be trained, the larger the y value, and the better the performance of the network in the target environment; the lower the similarity between the source domain and the target domain, the more individual features it has, the more network layers need to be trained, and the larger the y value; the migration channel pruning in step (4) is to reduce the difference between the characteristics of the target data set and the characteristics of the source domain data while pruning those unimportant channels.

7. The method for controlling a deformable mirror that can quickly adapt to actual application environments according to claim 6, Features: In the step (5), the voltage accuracy is represented by the root mean square, mean absolute deviation and relative deviation of the model output value and the true value of the data set, and is determined by the difference between the model prediction value and the true value through algorithm analysis. The surface shape control accuracy index is the peak-to-peak value pv and root mean square rms of the difference surface shape, the voltage accuracy of the model test, and is given by the difference in surface shape after the deformable mirror is driven by the true voltage and the predicted voltage.

8. A deformable mirror control device that can quickly adapt to the actual application environment. Features: It includes: A supplementary sampling module is configured to supplement a small amount of deformable mirror voltage and surface shape data to construct a migration data set: in the actual application scenario, several sets of typical control voltages are selected to drive the deformable mirror and measure the surface shape. These small amounts of voltage and surface shape data supplemented in the application scenario are used as the target domain data set of the deformable mirror, and the target domain data set is divided into a support set and a query set; an evaluation module is configured to evaluate the similarity between the source domain initial modeling environment and the target domain actual application environment data set: first, determine the similarity metric, and then use the target domain data of the actual application environment and the source domain data of the initial modeling environment to directly calculate the metric; A determination module is configured to determine the type of scheme for transfer learning of the deformable mirror control model: the basis for selecting the control determination is the amount of data in the migration data set and the similarity between the target environment and the source environment of the actual application of the deformable mirror; a fine-tuning scheme is selected for the control of the deformable mirror under general environmental conditions, and the method of migration channel clipping is combined to improve the effect of transfer learning, ensure control accuracy and minimize the increased training cost; The migration module is configured to train the network to realize the migration of the deformable mirror control model: according to the selected fine-tuning scheme, the first x layers of the source domain model trained in the source domain environment are frozen, and only a small amount of supplementary data is used to train the remaining y layers of the network; the first x layers are used to retain the nonlinear fitting characteristics between the common surface shape and voltage of the deformable mirror control model in any environment, and the unfrozen rear y layers of the network are used to extract the individual characteristics related to the actual application environment; Combined with the migration channel pruning scheme, the source domain model is fine-tuned, and then the migration channel evaluation is used to evaluate the importance of different channels of all network layers, and the less important channels are pruned, and iterative adjustments are made until a balance between accuracy and computational overhead is achieved; a verification module is configured to evaluate and verify the migration effect of the deformable mirror control model: first, the two control accuracy evaluation indicators of the deformable mirror control voltage and surface compensation accuracy are determined, and the migration model and the source domain model are tested and experimentally verified in the target domain dataset, i.e., the actual application environment of the deformable mirror. A comprehensive analysis of the accuracy is performed, and based on the accuracy comparison results, the model migration effect is evaluated and the migration scheme is adjusted.

9. The deformable mirror control device capable of quickly adapting to actual application environment according to claim 8, Features: It also includes an acquisition module, which is configured to obtain a source domain model for deformable mirror control: under environmental conditions where a large amount of data can be collected, the deformable mirror is driven and the deformable mirror surface shape is measured, the control voltage and the corresponding surface shape data are recorded, and a sufficient source domain data set is established; a neural network structure is selected to establish a deformable mirror control model, and optimization training is performed; the deformable mirror control model trained with a sufficient source domain data set is the source domain model; the source domain refers to a field that has been completed and contains knowledge in transfer learning, and the source domain model has good control accuracy capabilities and sufficient knowledge for reuse and migration.

10. The deformable mirror control device capable of quickly adapting to actual application environment according to claim 9, Features: In the migration module, the number of network layers y retrained during fine-tuning depends on two parameters: the amount of supplementary data and the similarity between the source domain and the target domain; the larger the amount of supplementary data, the more network layers can be trained, the larger the y value, and the better the performance of the network in the target environment; the lower the similarity between the source domain and the target domain, the more individual features it has, the more network layers need to be trained, and the larger the y value; the migration channel pruning in step (4) is to reduce the difference between the characteristics of the target data set and the characteristics of the source domain data while pruning those unimportant channels; In the verification module, the voltage accuracy is expressed by the root mean square, mean absolute deviation, and relative deviation of the model output value and the true value of the data set, and is determined by the difference between the model prediction value and the true value through algorithm analysis. The surface shape control accuracy index is the peak-to-peak value pv and root mean square rms of the difference surface shape, the voltage accuracy of the model test, and is given by the difference in surface shape after the deformable mirror is driven by the true voltage and the predicted voltage.

Images