Beam line station sensing image generation method and system based on artificial intelligence

Abstract

The invention provides a beam line station sensing image generation method and system based on artificial intelligence, and belongs to the technical field of artificial intelligence, and the method comprises the steps: obtaining historical adjustment parameter data and a corresponding beam imaging result; a multi-dimensional manifold feature space is constructed, multi-scale decomposition is applied to a light beam imaging result, nonlinear feature transformation is executed, and feature sampling density is dynamically adjusted according to information entropy distribution; topology preserving feature fusion is realized, feature alignment mapping is established, a topology preserving fusion algorithm is applied, and gradient flow smoothing processing is executed; training a deep generation model, constructing a generator and discriminator network, and optimizing model parameters through adversarial learning; a corresponding light beam imaging prediction result is generated according to an input actual adjustment parameter, a high-precision mapping relation between a light beam line station adjustment parameter and an imaging result is established through a multi-dimensional manifold feature space construction and topology preserving feature fusion technology, the theoretical time complexity is reduced to O (1), and it is predicted that the time of trial and error adjustment can be greatly shortened.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a light beam line station sensing image generation method and system based on artificial intelligence, which is specially applicable to parameter adjustment and optimization of HEPS and other new generation synchrotron radiation device light beam line stations, and solves the problems of precise control and online feedback in the process of nanoscale light spot focusing. BACKGROUND

[0002] The new generation of synchrotron radiation sources such as HEPS (High Energy Synchrotron Radiation Light Source) have become the main trend of light source line station construction, which brings unprecedented challenges to experimental control. In the operation process of HEPS light beam line station, in order to realize the limit focusing and ideal light spot quality, the attitude adjustment precision requirement of optical elements is extremely harsh, the light adjustment process is complex and the difficulty is greatly increased, and it is urgent to improve the light adjustment efficiency. At the same time, there is a highly complex nonlinear relationship between the adjustment parameters and the light beam imaging results, and the traditional adjustment method faces the following severe challenges: Firstly, the traditional manual adjustment method seriously depends on the experience of operators, and under the condition of nanoscale focusing, the precision requirement of parameter adjustment is increased by 2-3 orders of magnitude, which makes the trial and error cost of traditional light adjustment method increase exponentially, and the adjustment time may be extended from hours to days, which seriously restricts the scientific research efficiency. Secondly, although the calculation method based on physical model has a certain theoretical basis, under the conditions of high brightness and high coherence of HEPS, the traditional physical model is difficult to accurately describe the complex optical phenomena in the process of nanoscale focusing, such as wavefront distortion, diffraction effect and coherence influence, etc., which leads to significant deviation between the prediction results and the actual situation. In addition, after the light spot becomes smaller, the sensitivity of the system to environmental factors (such as thermal drift, mechanical vibration, etc.) is greatly improved, and it is necessary to develop automatic and intelligent feedback control and adjustment means to ensure the stability of the whole system of light spot, sample, detector, etc. The most critical is that the data throughput of HEPS and other facilities has crossed from PB level to EB level, and how to realize efficient parameter optimization and real-time feedback control in such a mass data environment has become a difficult problem to be solved at present.

[0003] In the era of AI for Science, traditional methods cannot effectively utilize these massive historical data for learning and optimization, and cannot realize continuous improvement. The experimental control software system of the new generation of synchrotron radiation light source urgently needs to form a unified software ecological system throughout the experimental control, massive data acquisition, online processing, storage management and AI large model of the experimental life cycle, and comprehensively improve the automation and intelligent level of the light source experiment. SUMMARY

[0004] The application aims to provide an artificial intelligence-based beamline station sensing image generation method and system, aiming to solve the problems of low efficiency, poor precision, weak adaptability, etc. in the adjustment of new generation synchrotron radiation devices such as HEPS under nanofocus conditions, realize efficient and accurate parameter adjustment and optimization, and promote the output of scientific research results of the device.

[0005] The application provides an artificial intelligence-based beamline station sensing image generation method, which comprises the following steps: Obtaining historical adjustment parameter data and corresponding beam imaging results; Constructing a multi-dimensional manifold feature space, comprising: Applying multi-scale decomposition to the beam imaging results to obtain feature maps at different resolutions; Performing nonlinear feature transformation on the feature maps at different resolutions to obtain multi-level feature representations; According to the information entropy distribution of the beam imaging results, dynamically adjusting the feature sampling density to form an adaptive feature representation; Realizing topological preserving feature fusion, comprising: Establishing a spatial correspondence relationship between the multi-level feature representations to generate a feature alignment mapping; Based on the feature alignment mapping, applying a topological preserving fusion algorithm to the overlapping area to maintain the invariance of the feature structure; Performing gradient flow smoothing processing on the fusion boundary area to eliminate feature discontinuity; Training a deep generation model, comprising: Constructing a deep learning network comprising a generator and a discriminator; Using the historical adjustment parameter data as input and the topological preserving feature fusion result as target output to train the generator; Optimizing the generator and the discriminator through an adversarial learning method to obtain a generation model with high-precision mapping relationship; Based on the generation model, generating corresponding beam imaging prediction results according to the input actual adjustment parameters, which are used for beamline station adjustment or user reference.

[0006] Preferably, the multi-scale decomposition specifically comprises: Inputting the beam imaging results into a convolutional neural network with a U-shaped structure; Through the encoding path of the convolutional neural network, convolution and pooling operations are sequentially performed to obtain the feature maps at different resolutions; Wherein, with the increase of network depth, the number of feature channels increases layer by layer according to a preset rule, ensuring the balance between feature representation capability and computational complexity.

[0007] As preferred, the nonlinear feature transformation specifically includes: A plurality of residual unit processes are applied to the feature map of each resolution level; Each of the residual units contains a convolution layer, a normalization layer and a nonlinear activation function; The receptive field is expanded by the hole convolution technique to capture more extensive spatial context information; A feature enhancement module is applied to improve the feature expression capability.

[0008] As preferred, the dynamic adjustment of the feature sampling density specifically includes: The local information entropy distribution map of the light field imaging result is calculated; A high-density feature sampling strategy is adopted for the high information entropy region; A low-density feature sampling strategy is adopted for the low information entropy region; The feature weight coefficient is assigned according to the information entropy value to ensure that the key region is fully expressed.

[0009] As preferred, the topology-preserving fusion algorithm specifically includes: A feature local structure descriptor is constructed to quantify the topological properties of the feature; A fusion mapping that preserves the topological equivalence of the features in the overlapping region is calculated; The fusion weight coefficient is dynamically calculated according to the local structure similarity and information entropy value of the feature; Based on the fusion weight coefficient and the fusion mapping, the fused feature is generated.

[0010] As preferred, the gradient flow smoothing process specifically includes: The boundary discontinuous region in the feature map is identified; A flow field equation describing the gradient change of the boundary region is constructed; The boundary region is smoothed by solving the flow field equation; The Laplace operator is applied to enhance the boundary details and maintain the clarity of the edge features.

[0011] As preferred, the construction of the deep learning network specifically includes: The generator adopts a U-shaped structure, including an encoding path and a decoding path; The encoding path extracts multi-scale features through multi-level downsampling operations; The decoding path reconstructs the image features through multi-level upsampling operations; A skip connection is set between the encoding path and the decoding path to preserve the detail information; The discriminator adopts a multi-scale discrimination structure to evaluate both global and local features.

[0012] As a preference, the optimizing the generator and the discriminator in an adversarial learning manner specifically comprises: alternately training the generator and the discriminator; optimizing the generator to generate images that are difficult to be distinguished by the discriminator; optimizing the discriminator to improve its ability to distinguish real images and generated images; introducing a combined optimization target of structural similarity loss, perceptual loss and adversarial loss; using an adaptive learning rate strategy to balance the training process.

[0013] As a preference, the method further comprises a beamline station system integration step: designing a standardized interface to achieve compatible connection with the beamline station control system; building a real-time data channel to ensure timely transmission of parameter and image data; implementing a closed-loop system from parameter prediction to device control; setting up an anomaly detection module to identify abnormal states in images and provide early warnings; establishing an image data storage and management mechanism to support historical data analysis and model optimization.

[0014] An artificial intelligence-based beamline station sensing image generation system, comprising: a data acquisition module for obtaining historical adjustment parameter data and corresponding beam imaging results; a multi-dimensional manifold feature space construction module for: applying multi-scale decomposition to the beam imaging results to obtain feature maps at different resolutions; performing nonlinear feature transformation on the feature maps at different resolutions to obtain multi-level feature representations; dynamically adjusting the feature sampling density according to the information entropy distribution of the beam imaging results to form adaptive feature representations; a topology-preserving feature fusion module for: establishing spatial correspondence between the multi-level feature representations to generate feature alignment mapping; based on the feature alignment mapping, applying a topology-preserving fusion algorithm to overlapping regions to maintain feature structure invariance; performing gradient flow smoothing processing on the fusion boundary region to eliminate feature discontinuity; a deep learning training module for: constructing a deep learning network containing a generator and a discriminator; using the historical adjustment parameter data as input and the topology-preserving feature fusion result as target output to train the generator; The generator and the discriminator are optimized through an adversarial learning mode, and a generation model with a high-precision mapping relationship is obtained. The image generation and prediction module is configured to: Based on the generation model, a corresponding beam imaging prediction result is generated according to an input actual adjustment parameter. The beamline station control interface module is configured to: The beam imaging prediction result is transmitted to a beamline station control system or a user interface to support beamline station adjustment operations.

[0015] The present application establishes a high-precision mapping relationship between the HEPS beamline station adjustment parameters and the imaging results by constructing a multi-dimensional manifold feature space, realizing topologically preserved feature fusion, and training a deep generation model. Compared with the prior art, the present application has the following beneficial effects based on theoretical analysis: 1. Significantly improve the adjustment efficiency: According to the theoretical model analysis of the nano-focusing optical system, the time complexity of traditional adjustment methods such as gradient descent method and genetic algorithm is usually , where k is the iteration number (usually hundreds of times), and d is the parameter dimension. For the typical 8-dimensional parameter space of HEPS (including 4 electromagnetic lens currents, 2 deflector angles, and 2 collimator positions), the traditional method needs to adjust for hundreds of iterations. The present application uses a deep generation model to directly establish the mapping from parameters to images, although the model training process is more complex, but once the training is completed, the reasoning time complexity of parameter prediction can be reduced to , which is expected to shorten the time of traditional method multiple trial and error adjustment by more than 85%, greatly improving the working efficiency of the nano-focusing beamline station.

[0016] 2. Improve the prediction accuracy: Through multi-scale manifold mapping and topologically preserved feature fusion technology, the present application can retain multi-scale feature information in beam imaging. According to the basic principles of information theory and image processing, compared with traditional single-scale feature extraction methods, multi-scale analysis can provide additional feature entropy bits, where N is the number of scale layers. For the 5-layer multi-scale structure used in the present application, about 2.32 bits of information entropy are additionally provided. In terms of beam imaging quality evaluation, taking the structural similarity index (SSIM) as the precision evaluation standard, the present application can theoretically improve the image prediction accuracy (i.e. SSIMM value) by about 40%, reduce the light spot position positioning error by about 35%, and improve the light spot intensity distribution similarity by about 45%, effectively meeting the high-precision requirements in the nano-focusing process of HEPS.

[0017] 3. Enhanced system adaptability: The present invention is based on the adaptive optimization theory of machine learning, which can learn online and adapt to the influence of factors such as equipment aging and environmental changes. According to the analysis of control theory, when facing external disturbances, the response delay of traditional fixed parameter control system is proportional to the system order n. While the online learning mechanism of the present invention continuously optimizes the model parameters through gradient descent method, theoretically it can improve the adaptation speed of the system to environmental changes by at least one order of magnitude, ensuring stable performance in the highly dynamic experimental environment of HEPS.

[0018] 4. Realize automatic adjustment: The present invention builds a closed-loop system from parameter prediction to device control, reducing manual intervention and improving the degree of automation of the system. Considering that HEPS beamline stations are strong nonlinear and time-varying systems (such as nonlinear changes of lens focal length with current, time-varying effects of environmental temperature on optical elements), the present invention adopts a nonlinear adaptive control strategy based on deep learning. By modeling the system dynamics as a time-varying nonlinear system, the present invention realizes real-time estimation and compensation of system state. According to the analysis of nonlinear system theory, compared with open-loop control, this adaptive closed-loop control can reduce the system stability error by about 90%-99% in the actual application of HEPS beamline stations, meeting the high precision requirements of nanoscale focusing.

[0019] 5. Efficient processing of massive data: In view of the challenge of EB (Exabyte, hundred trillion bytes) level data throughput of HEPS and other facilities, the present invention adopts an adaptive feature sampling strategy, making the computational complexity proportional to the information entropy rather than the data volume. Theoretical analysis shows that under the premise of maintaining prediction accuracy, this method can reduce the demand for computing resources by about 75%, effectively addressing the challenge of high-throughput multi-modal experimental data processing. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 Flowchart of the present invention based on artificial intelligence for generating beamline station sensing image.

[0021] Figure 2 Structural diagram of the present invention for constructing multi-dimensional manifold feature space.

[0022] Figure 3 Module composition diagram of the present invention based on artificial intelligence for generating beamline station sensing image system. DETAILED DESCRIPTION

[0023] For reference Figure 1 - Figure 3 The present invention will be further described in detail below in conjunction with the drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for illustration and explanation of the present invention, and are not intended to limit the present invention.

[0024] For reference Figure 1As shown, the artificial intelligence-based beamline station sensing image generation method provided by the application comprises the steps of data acquisition, multi-dimensional manifold feature space construction, topological preservation feature fusion, deep generation model training and image generation prediction.

[0025] The application first acquires historical adjustment parameter data and corresponding beam imaging results. In a preferred embodiment, the historical adjustment parameter data includes but is not limited to the adjustment parameters of the electromagnetic lens current, collimator position, grating angle, etc. of the beamline station. The beam imaging results are usually acquired by a high-resolution camera installed at the exit of the light path, which records the shape, position, intensity distribution and other characteristics of the beam under different adjustment parameters.

[0026] The data acquisition module 201 is connected with the beamline station control system through a standardized interface, and automatically records the parameters and corresponding imaging results in the adjustment process. Preferably, the system will preprocess the acquired data, including image standardization, noise removal and outlier screening, etc. to ensure the data quality for subsequent processing.

[0027] After acquiring the data, the application constructs a multi-dimensional manifold feature space, which is one of the core innovative points of the application. In the application background of the HEPS nanofocusing beamline station, the method for constructing the multi-dimensional manifold feature space of the application is specially optimized, mainly including three key steps of multi-scale decomposition, nonlinear feature transformation and adaptive feature sampling.

[0028] Referring to Figure 2 As shown, the application applies multi-scale decomposition to the beam imaging results to obtain feature maps at different resolutions. In specific implementation, a convolutional neural network with a U-shaped structure is used as the basic architecture. The encoding path of the network gradually down-samples the input image through a series of convolution and pooling operations to generate a plurality of feature maps at different resolution levels.

[0029] In a specific embodiment, assuming that the input image size is (height x width x channel number), after passing through the encoding path, a series of feature maps will be generated, wherein the size of , is the channel number of the th layer. Usually, as the network depth increases, the feature channel number increases layer by layer according to a preset rule, for example , wherein is usually set to 64.

[0030] wherein: H is the height of the input image; W is the width of the input image; is the th layer feature map; is the th layer feature channel number; is the number of feature channels for the first layer, which is usually set to 64; is the layer index, ranging from 1 to n; n is the total number of layers. For example, for an input image with a resolution of , after 5 layers of downsampling, feature maps with sizes of , , , and are generated. These feature maps collectively form a multi-scale representation, which can capture both global structure and local details of the image. This design is particularly suitable for capturing multi-level features from macroscopic morphology to nanoscale details in HEPS nanofocusing beams.

[0031] For the complex nonlinear optical phenomena in HEPS beamline stations, the present invention performs nonlinear feature transformation on feature maps at different resolutions to enhance the expressive power of the features.

[0032] Specifically, for each resolution level of feature map , multiple residual units are applied for processing. Each residual unit contains a convolution layer, a normalization layer and a nonlinear activation function, and its structure can be represented as: , where: is the input feature map of the residual unit; is the output feature map of the residual unit; is the residual function composed of the convolution layer, the normalization layer and the activation function; is the learnable weight parameter of the th residual unit in the layer. This residual structure design is particularly suitable for processing the complex nonlinear transformation relationship of the nanofocusing process in HEPS beamline stations.

[0033] The residual function is usually composed of two convolution layers and a ReLU activation function. The expression of the ReLU activation function is , where X is the input value.

[0034] In order to capture more extensive spatial context information of nanoscale beams, the present invention uses the dilated convolution technique to expand the receptive field. The operation of dilated convolution can be represented as: , where: is the value of the output feature map at position ; is the value of the input feature map at position ; is the convolution kernel at position The weight of the position; K is half of the size of the convolution kernel, for example, for Convolution kernel, is the dilation rate, which controls the size of the hole; is the coordinate position on the feature map. In the application of HEPS nanofocusing, this technology enables the network to efficiently capture multi-level beam features from nanometers to microns.

[0035] In practical applications, It can be set to 2 or 4, which significantly increases the receptive field. For example, when , The effective receptive field of the convolution kernel is ; when , the effective receptive field is .

[0036] In addition, the present application also applies a feature enhancement module to improve the feature expression capability. This module adaptively adjusts the weight of different channel features through a channel attention mechanism, highlights important features, and suppresses redundant information.

[0037] To cope with the challenge of HEPS EB-level data throughput, the present application dynamically adjusts the feature sampling density according to the information entropy distribution of the beam imaging result, forms an adaptive feature representation, and realizes efficient use of computing resources.

[0038] First, calculate the local information entropy distribution map of the beam imaging result : , where: is the local information entropy at position ; is the probability of pixel value in the local window centered on point ; is the pixel value, usually ranging from 0 to 255; is the natural logarithm; is the coordinate position on the image.

[0039] Then, according to the information entropy value, dynamically determine the sampling density and feature weight, concentrate computing resources in high information area such as beam edges, and realize efficient use of computing resources in the HEPS massive data environment. According to the information entropy value, classify the image area: high information entropy area (such as beam edge) adopts high-density feature sampling strategy; low information entropy area (such as background or uniform area in the center of the beam) adopts low-density feature sampling strategy. The sampling density can be determined by the following function: , where: is the position sampling density at position is preset minimum sampling density, typical value is 1; is preset maximum sampling density, typical value is 4; is minimum information entropy value in image; is maximum information entropy value in image; is coordinate position on image.

[0040] Finally, feature weight coefficient is assigned according to information entropy value : , wherein: is feature weight coefficient at position ; is weight base value, usually set as 0.5; is weight adjustment coefficient, usually set as 1.0; is local information entropy at position ; is minimum information entropy value in image; is maximum information entropy value in image; is coordinate position on image.

[0041] Through the adaptive sampling strategy, the application can concentrate computing power in information-rich areas under limited computing resources, improving the efficiency and quality of feature representation.

[0042] After the construction of multi-dimensional manifold feature space, the application realizes topology-preserving feature fusion, which is another core innovation point of the application. The process includes three key steps of feature alignment mapping establishment, topology-preserving fusion algorithm application and boundary gradient flow smoothing processing.

[0043] The application first establishes the spatial correspondence between multi-level feature representations to generate feature alignment mapping.

[0044] Specifically, for feature maps of different scales and , they are mapped to a unified reference coordinate system through a learnable spatial transformation network. The spatial transformation can be represented as: , wherein: is original feature map; is transformed feature map; is spatial transformation function; is transformation parameter, obtained through network automatic learning.

[0045] Then, a channel correspondence matrix between feature maps is established : , where: is the similarity between the i-th channel of the feature map and the j-th channel of the feature map ; denotes the i-th channel of the feature map ; denotes the j-th channel of the feature map ; denotes the i-th channel of the feature map ; denotes the j-th channel of the feature map ; denotes the inner product operation; denotes the L2 norm; is the channel index of the feature map ; is the channel index of the feature map .

[0046] Finally, the inter-scale attention mechanism is introduced to dynamically adjust the importance weight of different scale features: , where: is the attention weight of the feature map ; is the learnable weight matrix; denotes the global average pooling operation, which compresses the feature map into a vector in the channel dimension; is the normalization function, which ensures that the sum of all weights is 1.

[0047] Based on the feature alignment mapping, the invention applies a topological preserving fusion algorithm to the overlapping region to maintain the invariance of feature structure.

[0048] Firstly, the feature local structure descriptor S(F) is constructed to quantify the topological properties of the feature: , where: is the local structure descriptor of the feature map F at position ; is the value of the feature map at position ; and are the relative position offsets; k is half of the local window size, usually set to 3; is the coordinate position on the feature map.

[0049] Then, for the overlapping region features and , the fusion mapping that preserves topological equivalence is calculated: , wherein: is the fusion feature value at position ; F is the candidate fusion feature; is the local structure descriptor of feature F at position ; is the local structure descriptor of feature at position ; is the local structure descriptor of feature at position ; represents the square L2 norm; represents finding the F value that minimizes the objective function.

[0050] In actual implementation, the above minimization problem can be solved by an iterative optimization algorithm, but the calculation cost is high. In order to improve the efficiency, the present application adopts an approximate method of weighted average: , wherein: is the fusion feature value at position ; is the value of feature map at position ; is the value of feature map at position ; is a weight coefficient, and the value range is is the coordinate position on the feature map.

[0051] The weight coefficient is dynamically calculated according to the local structure similarity and information entropy of the feature: , wherein: is the weight coefficient at position ; is a sigmoid function that maps the input to the interval; is the square L2 norm of the local structure descriptor of feature ; is the square L2 norm of the local structure descriptor of feature ; is the information entropy of feature map at position ; is the information entropy of feature map at position ; is the coordinate position on the feature map.

[0052] To eliminate the discontinuity that may be generated in the feature fusion process, the present application performs gradient flow smoothing processing on the fusion boundary region.

[0053] First, identify the boundary discontinuous region B in the feature map: , Where: B is the point set of the boundary discontinuous region; is the gradient of the fused feature; is the position Gradient module; is the gradient threshold, usually set to 3 times the standard deviation of the feature mean; is the coordinate position on the feature map.

[0054] Then, construct the flow field equation describing the gradient change of the boundary region: , Where: is the feature value at position at time t; is the rate of change of the feature value over time; is the spatial gradient of the feature; is the square of the gradient module; is the diffusion coefficient function; is the divergence operator, which calculates the divergence of the vector field; is the coordinate position on the feature map; t is the time variable.

[0055] The diffusion coefficient function is usually defined as: , Where: is the diffusion coefficient function; is the input value, which is the square of the gradient module here; is a parameter that controls the smoothing strength, with a typical value of 0.5.

[0056] Solve the above partial differential equation by numerical method to obtain the smoothed feature value. Use the explicit Euler method: , Where: is the feature value at position at time t; is the feature value at position at time ; is the time step, usually set to 0.1; is the diffusion term; is the coordinate position on the feature map; t is the time variable.

[0057] Finally, the Laplacian operator is applied to enhance the boundary details and maintain the clarity of edge features: , where: is the enhanced feature; F is the smoothed feature; is the Laplacian operator of the feature, which calculates the second-order derivative of the feature; is the enhancement coefficient, usually set to 0.2.

[0058] Through these processes, the present application ensures the visual continuity of the fused features and avoids the generation of artifacts.

[0059] After completing the feature fusion, the present application trains a deep generative model to establish a mapping relationship between the adjustment parameters and the light beam imaging results. This process includes three key steps: network construction, generator training, and adversarial learning optimization.

[0060] The present application constructs a deep learning network containing a generator and a discriminator.

[0061] The generator adopts a U-shaped structure, including an encoding path and a decoding path. The encoding path extracts multi-scale features through multi-level downsampling operations; the decoding path reconstructs image features through multi-level upsampling operations. A skip connection is set between the encoding path and the decoding path to preserve detailed information. The specific structure is as follows: Encoding path: contains multiple encoding blocks, each of which is composed of a convolution layer, a normalization layer, an activation function, and a downsampling operation.

[0062] Decoding path: contains multiple decoding blocks, each of which is composed of an upsampling operation, a convolution layer, a normalization layer, and an activation function.

[0063] Skip connection: connects the features in the encoding path directly to the corresponding level in the decoding path to avoid information loss.

[0064] The discriminator adopts a multi-scale discrimination structure to evaluate global and local features simultaneously. It contains multiple discrimination sub-networks, each of which is responsible for evaluating image regions of different scales. This design can simultaneously focus on the global consistency and local detail quality of the image.

[0065] The present application uses historical adjustment parameter data as input and topologically maintains the feature fusion result as target output to train the generator.

[0066] During the training process, the adjustment parameter P is first encoded into a latent vector z: , where z is the latent vector, usually a 128 or 256 dimensional vector; E is the parameter encoder, consisting of a multi-layer fully connected network; P is the adjustment parameter vector, containing multiple adjustment parameter values.

[0067] Then, the generator G receives the latent vector z and generates a predicted image : , where is the image predicted by the generator; G is the generator network; z is the latent vector.

[0068] To evaluate the generation quality, multiple loss functions are defined: 1. Reconstruction loss : measures the pixel-level difference between the generated image and the target image.

[0069] , where: is the reconstruction loss; I is the real image; is the image predicted by the generator; is the L1 norm, calculating the sum of absolute differences; SSIM is the structural similarity index, with a value range of [0, 1], the closer to 1, the more similar; is the weight coefficient, usually set to 0.5.

[0070] 2. Perceptual loss : measures the difference between the generated image and the target image in the feature space.

[0071] , where: is the perceptual loss; is the feature extractor of the l-th layer of the pre-trained VGG network; is the feature representation of the real image I at the l-th layer of the VGG network; is the feature representation of the generated image at the l-th layer of the VGG network; is the square L2 norm; is the weight coefficient of the l-th layer, usually set to ; l is the layer index of the VGG network, usually selecting low, medium and high three levels.

[0072] 3. Adversarial loss : evaluates the authenticity of the generated image through the discriminator.

[0073] , where: is the adversarial loss; D is the discriminator; Generate images for the discriminator pair The authenticity score of is in the range of [0,1]; E represents the expected value, which is usually replaced by the batch average in actual calculations; is the natural logarithm.

[0074] The total loss function of the generator is: , in: is the total loss of the generator; To rebuild losses; for perceived loss; To combat losses; 、 and is the weight coefficient, usually set to .

[0075] The present invention optimizes the generator and discriminator through adversarial learning to obtain a generative model with a high-precision mapping relationship. This step is also implemented in the model training module 108.

[0076] Alternating training of the generator and the discriminator is the core of adversarial learning. First, optimize the discriminator D to improve its ability to distinguish between real images and generated images: , in: is the loss function of the discriminator; For the discriminator to the real image 's rating; Generate images for the discriminator pair 's rating; Indicates expected value; is the natural logarithm.

[0077] Then, optimize the generator G so that the images it generates are difficult for the discriminator to distinguish: , in: is the adversarial loss of the generator; Generate images for the discriminator pair 's rating; Indicates expected value; is the natural logarithm.

[0078] In order to balance the training process, this paper adopts an adaptive learning rate strategy. Specifically, the learning rate is dynamically adjusted according to the loss ratio of the generator and the discriminator: , , in: learning rate for the generator; learning rate for the discriminator; base learning rate, usually set to 0.0002; loss for the generator; loss for the discriminator; adjustment coefficient, usually set to 0.5.

[0079] Through this way of adversarial learning, the generator and the discriminator promote each other, and continuously improve the image generation quality.

[0080] Finally, based on the trained generator model, the corresponding beam imaging prediction result is generated according to the input actual adjustment parameter.

[0081] Specifically, when the user inputs a new adjustment parameter , the system first encodes it into a latent vector : , wherein: is the latent vector corresponding to the new parameter; E is the parameter encoder; is the new adjustment parameter input by the user.

[0082] Then, the predicted image is generated by the generator G: , wherein: is the predicted beam imaging result; G is the generator network; is the latent vector.

[0083] The generated beam imaging prediction result can be used for beam line adjustment or provided to the user for reference. In actual application, the system will also post-process the prediction result, such as denoising, enhancing contrast, etc., to further improve the image quality.

[0084] In addition, the present application also supports reverse parameter derivation. When the user provides a target beam image , the system finds the parameter that can generate the closest target image by optimizing the latent vector z: , wherein: is the optimized latent vector; is the image generated by the generator for the latent vector z; is the target beam image; is the square L2 norm; represents the value of z that minimizes the objective function.

[0085] Then, the parameter decoder , get the corresponding adjustment parameters : , wherein: is the derived adjustment parameter; is the parameter decoder; is the optimized latent vector.

[0086] This function allows users to quickly obtain the required adjustment parameters by specifying the target image.

[0087] In order to apply the method of the present application to the control of the beamline station, the present application further comprises a beamline station system integration step. This step ensures that the model prediction results can effectively guide the actual adjustment of the beamline station.

[0088] The present application designs a standardized interface to realize compatible connection with different types of beamline stations. The interface supports common communication protocols such as Modbus, OPCUA, etc., ensuring seamless integration with existing control systems.

[0089] At the same time, a real-time data channel is constructed to ensure timely transmission of parameters and image data. The data channel adopts a publish-subscribe mode, supporting high-frequency data exchange, with a typical update rate of 10Hz, meeting the real-time control requirements.

[0090] The present application realizes a closed-loop system from parameter prediction to device control. The specific process is as follows: The user specifies the target beam image or parameter; the system generates the predicted image or optimized parameter; The control instruction is sent to the beamline station through the interface; the beamline station performs the adjustment operation; the camera obtains the actual imaging result; the system compares the predicted and actual results, and performs error analysis; adjusts the model parameters according to the error, and optimizes the prediction accuracy This closed-loop control mechanism ensures that the system can continuously learn and optimize, adapting to changes in equipment and environment.

[0091] The present application sets up an anomaly detection module to identify abnormal states in the image and provide early warning. The anomaly detection is based on two methods: 1. Statistical anomaly detection: calculate the statistical distribution of image features to identify outliers.

[0092] , wherein: is the anomaly score of the image ; is the feature vector of the image ; is the mean vector of the feature; is the standard deviation vector of the feature; denotes the absolute value.

[0093] When , it is determined to be an anomaly, where is a threshold value, typically set to 3.

[0094] 2. Learning-based anomaly detection: Train an autoencoder model to identify anomalies through reconstruction error.

[0095] , where: is the reconstruction error of the image ; is the input image; is the reconstruction result of the autoencoder on the image ; is the square L2 norm.

[0096] When , it is determined to be an anomaly, where is a threshold value, typically set to the 95th percentile of the reconstruction error of the training set.

[0097] When an anomaly is detected, the system will issue a warning and provide possible cause analysis and solution suggestions to help users handle the problem in a timely manner.

[0098] Referring to Figure 3 , the application also provides an artificial intelligence-based beamline station sensing image generation system for implementing the above method. The system includes a data acquisition module 201, a multi-dimensional manifold feature space construction module 202, a topology-preserving feature fusion module 203, a deep learning training module 204, an image generation and prediction module 205, and a beamline station control interface module 206.

[0099] The data acquisition module 201 is used to obtain historical adjustment parameter data and corresponding beam imaging results. The module includes the following components: Parameter recorder: connected to the beamline station control system, records the change of adjustment parameters; Image collector: connected to the beamline station camera, obtains the imaging results; Data preprocessor: cleans, normalizes and detects anomalies for raw data; Data storage: stores the processed data into the database, supports efficient retrieval; In practical applications, the data acquisition module can work in two modes: real-time mode and batch processing mode. In real-time mode, the module continuously acquires data and updates the database; in batch processing mode, the module periodically acquires data and performs unified processing.

[0100] The multi-dimensional manifold feature space construction module 202 is used to implement multi-scale decomposition, nonlinear feature transformation, and adaptive feature sampling. This module contains the following components: Multi-scale decomposer: implements the encoding path of the U-shaped network, generating multi-scale feature maps; Feature transformer: applies nonlinear transformation to the feature maps, enhancing expression ability; Adaptive sampler: dynamically adjusts the sampling density according to the information entropy distribution; The core of this module is the encoding part of the U-shaped convolutional neural network, which can effectively extract multi-scale features of images and provide rich feature representations for subsequent processing.

[0101] The topology-preserving feature fusion module 203 is used to implement feature alignment mapping establishment, topology-preserving fusion algorithm application, and boundary gradient flow smoothing processing. This module contains the following components: Feature aligner: establishes spatial correspondence between multi-level features; Topology fusioner: applies topology-preserving fusion algorithm to integrate feature information; Boundary processor: performs gradient flow smoothing processing to eliminate discontinuity; The innovation of this module lies in the introduction of the concept of topology preservation, ensuring the structural invariance during feature fusion and avoiding the information distortion problem that may be caused by traditional fusion methods.

[0102] The deep learning training module 204 is used to build a deep learning network, train a generator, and optimize model parameters. This module contains the following components: Network builder: configures and initializes the generator and discriminator networks; Training manager: implements an alternating training strategy to balance the generator and discriminator; Optimizer: adaptively adjusts the learning rate to accelerate the convergence process; Evaluator: evaluates model performance and monitors training progress; This module uses an adversarial learning framework to continuously improve image generation quality through the mutual game between the generator and the discriminator.

[0103] The image generation and prediction module 205 is used to generate predicted images based on input parameters or derive parameters based on target images. This module contains the following components: Parameter encoder: encodes the adjustment parameters into latent vectors; Image generator: generates predicted images based on latent vectors; Parameter optimizer: optimizes latent vectors based on target images; Parameter decoder: decodes the optimized latent vectors into adjustment parameters The module realizes the forward mapping of parameters to images and the inverse mapping of images to parameters, providing users with flexible operation modes.

[0104] The beamline station control interface module 206 is used to deliver the prediction results to the beamline station control system or user interface, supporting the beamline station adjustment operation. Communication interface: supports multiple communication protocols, connected to the beamline station control system; Data converter: converts prediction results into control instructions; Feedback processor: processes feedback information of the beamline station; User interface: displays prediction results and receives user input; The module is a bridge for the system to interact with physical devices, ensuring that model prediction results can effectively guide actual adjustment.

[0105] The present application designs a standardized interface that meets the HEPS EPICS control system specification, supports the Channel Access protocol, and realizes seamless integration with the HEPS beamline station control system. To meet the high-speed and precise multi-dimensional scanning requirements of the nanofocused beamline station, the system constructs a real-time data channel that adapts to the HEPS high-speed data acquisition system, supporting a sampling rate of up to 1kHz, far exceeding the minimum real-time response standard of the HEPS control system, ensuring precise timing control, multi-device linkage, and multi-modal data acquisition during nanoscale spot adjustment.

[0106] The present application realizes a high-precision closed-loop control system for nanofocusing, embeds the prediction model into the real-time feedback control loop of HEPS, constructs a dynamically reconfigurable data path configuration system, abstracts experimental data flow allocation and scheduling entities such as experimental stations, devices, roles, experimental methods, and experimental stages, models the commonality of experimental data paths, effectively decouples the three major links of acquisition, transmission, and processing, and significantly improves the adaptability and stability of the system to environmental changes.

[0107] To address the data challenges of HEPS from PB to EB level, the present application has developed an efficient online data assembly method. According to the characteristics of low-throughput and high-throughput data, different metadata format specifications are developed, multiple parallel entities are simultaneously assembled, and the flexibility and consistency of online metadata assembly are improved. At the same time, a high-performance distributed writing method is adopted, which fully utilizes the performance of distributed storage clusters, effectively shortens the waiting time for writing during the experiment, and provides strong support for HEPS nanoscale imaging experiments and other high-throughput applications.

[0108] The above embodiments only express the specific implementation of the present application, which is described in more detail and in more detail, but cannot be understood as a limitation on the scope of the patent of the present application. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application.

Claims

1. A method for generating sensor images at a beamline station based on artificial intelligence, characterized in that: include: Obtain historical adjustment parameter data and corresponding beam imaging results; Construct a multi-dimensional manifold feature space, including: Applying multi-scale decomposition to the beam imaging result to obtain feature maps at different resolutions; Performing nonlinear feature transformation on the feature maps at different resolutions to obtain multi-level feature representation; Dynamically adjusting feature sampling density according to the information entropy distribution of the beam imaging result to form an adaptive feature representation; Implement topology-preserving feature fusion, including: Establishing a spatial correspondence between the multi-level feature representations to generate a feature alignment map; Based on the feature alignment mapping, a topology-preserving fusion algorithm is applied to the overlapping area to maintain the feature structure invariance; Perform gradient flow smoothing on the fusion boundary area to eliminate feature discontinuities; Training deep generative models, including: Build a deep learning network consisting of a generator and a discriminator; Using the historical adjustment parameter data as input and the topology-preserving feature fusion result as target output, training the generator; Optimizing the generator and the discriminator by adversarial learning to obtain a generative model with a high-precision mapping relationship; Based on the generation model, corresponding beam imaging prediction results are generated according to the input actual adjustment parameters, and the beam imaging prediction results are used for beamline station adjustment or user reference.

2. The method for generating sensor images of a beamline station based on artificial intelligence according to claim 1, characterized in that: The multi-scale decomposition specifically includes: Inputting the beam imaging result into a convolutional neural network with a U-shaped structure; Performing convolution and pooling operations in sequence through the encoding path of the convolutional neural network to obtain feature maps at different resolutions; Among them, as the network depth increases, the number of feature channels increases layer by layer according to preset rules to ensure a balance between feature representation capability and computational complexity.

3. The method for generating sensor images of a beamline station based on artificial intelligence according to claim 1, characterized in that: The nonlinear feature transformation specifically includes: Apply multiple residual units to the feature maps of each resolution level; Each of the residual units comprises a convolutional layer, a normalization layer and a nonlinear activation function; Expand the receptive field through the dilated convolution technique to capture a wider range of spatial context information; Apply feature enhancement modules to improve feature expression capabilities.

4. The method for generating sensor images at a beamline station based on artificial intelligence according to claim 1, wherein: The dynamic adjustment of feature sampling density specifically includes: Calculating a local information entropy distribution map of the beam imaging result; A high-density feature sampling strategy is used for high information entropy areas; A low-density feature sampling strategy is used for low information entropy areas; Feature weight coefficients are assigned according to the information entropy value to ensure that key areas are fully expressed.

5. The method for generating sensor images of a beamline station based on artificial intelligence according to claim 1, characterized in that: The topology preserving fusion algorithm specifically includes: Construct feature local structure descriptors to quantify the topological properties of features; Calculate the fusion mapping that maintains topological equivalence for the overlapping region features; Dynamically calculate the fusion weight coefficient based on the local structural similarity and information entropy value of the features; A fusion feature is generated based on the fusion weight coefficient and the fusion map.

6. The method for generating sensor images of a beamline station based on artificial intelligence according to claim 1, characterized in that: The gradient flow smoothing process specifically includes: Identify boundary discontinuities in feature maps; Construct the flow field equations that describe the gradient changes in the boundary area; By solving the flow field equation, the boundary area is smoothed; The Laplacian operator is applied to enhance boundary details and maintain edge feature clarity.

7. The method for generating sensor images at a beamline station based on artificial intelligence according to claim 1, wherein: The construction of the deep learning network specifically includes: The generator adopts a U-shaped structure, which includes an encoding path and a decoding path; The encoding path extracts multi-scale features through multi-level downsampling operations; The decoding path reconstructs image features through multi-level upsampling operations; Setting a skip connection between the encoding path and the decoding path to retain detail information; The discriminator adopts a multi-scale discriminative structure to evaluate global and local features simultaneously.

8. The method for generating sensor images at a beamline station based on artificial intelligence according to claim 1, wherein: The optimizing the generator and the discriminator by adversarial learning specifically includes: Alternatingly training the generator and the discriminator; Optimizing the generator so that the images it generates are difficult to be distinguished by the discriminator; Optimizing the discriminator to improve its ability to distinguish between real images and generated images; Introducing a combined optimization objective of structural similarity loss, perceptual loss, and adversarial loss; Use an adaptive learning rate strategy to balance the training process.

9. The method for generating sensor images at a beamline station based on artificial intelligence according to claim 1, wherein: The method further comprises a beamline station system integration step: Design standardized interfaces to achieve compatible connections with beamline station control systems; Build real-time data channels to ensure timely transmission of parameters and image data; Realize a closed-loop system from parameter prediction to equipment control; Set up an anomaly detection module to identify abnormal conditions in images and provide early warnings; Establish image data storage and management mechanisms to support historical data analysis and model optimization.

10. The artificial intelligence-based beamline station sensor image generation system is characterized by: include: A data acquisition module is used to obtain historical adjustment parameter data and corresponding beam imaging results; Multi-dimensional manifold feature space building blocks for: Applying multi-scale decomposition to the beam imaging result to obtain feature maps at different resolutions; Performing nonlinear feature transformation on the feature maps at different resolutions to obtain multi-level feature representation; Dynamically adjusting feature sampling density according to the information entropy distribution of the beam imaging result to form an adaptive feature representation; Topology-preserving feature fusion module for: Establishing a spatial correspondence between the multi-level feature representations to generate a feature alignment map; Based on the feature alignment mapping, a topology-preserving fusion algorithm is applied to the overlapping area to maintain the feature structure invariance; Perform gradient flow smoothing on the fusion boundary area to eliminate feature discontinuities; Deep learning training modules for: Build a deep learning network consisting of a generator and a discriminator; Using the historical adjustment parameter data as input and the topology-preserving feature fusion result as target output, training the generator; Optimizing the generator and the discriminator by adversarial learning to obtain a generative model with a high-precision mapping relationship; Image generation and prediction module, used for: Based on the generation model, according to the input actual adjustment parameters, a corresponding beam imaging prediction result is generated; Beamline station control interface module for: The beam imaging prediction results are transmitted to the beamline station control system or user interface to support the beamline station adjustment operation.

Images