Photovoltaic cell film quality detection method and system based on synchronous stretching

Abstract

The invention provides a synchronous stretching-based photovoltaic cell film quality detection method and system, and relates to the technical field of quality detection, and the method comprises the steps: obtaining a surface topography image sequence of a photovoltaic cell film, constructing a surface topography dynamic evolution manifold model, generating time-varying surface topography point cloud data, and dividing the time-varying surface topography point cloud data into adaptive voxel grids; extracting multi-scale dynamic features, extracting space-time coupling features of the surface topography through time causal convolution and space graph convolution, and dynamically mapping the space-time coupling features into a defect probability sequence; the method comprises the following steps: obtaining a multi-physics field instability indication through dynamic numerical simulation based on multi-physics field coupling data, carrying out cross-scale dynamic association on the multi-physics field instability indication and a defect probability sequence, constructing a defect evolution prediction model, and carrying out real-time training and updating on the defect evolution prediction model through an adaptive dynamic Bayesian network reasoning algorithm. And generating a photovoltaic cell film quality early warning signal.

Description

Technical Field

[0001] The present invention relates to the technical field of quality detection, and in particular to a photovoltaic cell film quality detection method and system based on synchronous stretching. Background Art

[0002] Photovoltaic cell thin film is the core component of solar photovoltaic power generation system, and its quality and performance directly affect the photoelectric conversion efficiency and service life. With the rapid development of flexible photovoltaic cells, roll-to-roll preparation process has become the mainstream method for the production of photovoltaic cell thin film. During the roll-to-roll preparation process, the thin film material undergoes continuous mechanical stretching deformation, accompanied by the action of multiple physical fields such as heat, electricity, and light. Defects are easily generated inside the material, resulting in a decrease in the quality of the film. Timely detection and warning of the evolution trend of thin film defects are of great significance for controlling production quality and improving yield rate.

[0003] At present, the quality inspection of photovoltaic cell films mainly adopts optical imaging, electrical testing and other methods. For example, photoluminescence imaging and electroluminescence imaging can detect the morphological defects on the surface of the film; based on electrical parameters, the charge transfer performance of the film can be characterized. However, most of these methods are limited to offline detection and sampling analysis, which makes it difficult to monitor the dynamic changes of film quality in real time, and lacks in-depth understanding of the defect generation mechanism and evolution law. In addition, the existing methods mainly focus on the material response under a single physical field, ignoring the influence of multi-physical field coupling effects such as mechanical-thermal-electrical on defect evolution, and the prediction accuracy and reliability need to be improved.

[0004] In summary, in order to overcome the shortcomings, a method is urgently needed that can detect the quality of photovoltaic cell thin films in real time during the roll-to-roll preparation process, reveal the law of defect evolution under the coupling of multiple physical fields, and warn of quality risks. Based on the real-time collection and fusion analysis of multi-physical field parameters during the synchronous stretching process, a prediction model for defect evolution is constructed to achieve online warning of film quality, which can provide new ideas for the quality control of photovoltaic cell thin films. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a photovoltaic cell film quality detection method and system based on synchronous stretching, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for detecting quality of a photovoltaic cell film based on synchronous stretching is provided, comprising:

[0008] Based on progressive synchronous stretching, a surface morphology image sequence of a photovoltaic cell film is obtained; based on an adaptive fractal dimension analysis algorithm, a multi-scale analysis is performed on the surface morphology image sequence to extract surface features; the surface morphology features are mapped to a low-dimensional manifold space through a feature mapping method based on manifold learning, and a surface morphology dynamic evolution manifold model is constructed; based on the surface morphology dynamic evolution manifold model, the corresponding surface morphology is reconstructed through a deep learning three-dimensional space reconstruction algorithm to generate time-varying surface morphology point cloud data; and the time-varying surface morphology point cloud data is divided into an adaptive voxel grid through an adaptive voxelization strategy;

[0009] For each adaptive voxel grid, extract the multi-scale dynamic features in the adaptive voxel grid, input the multi-scale dynamic features into the pre-trained spatiotemporal convolutional neural network, extract the spatiotemporal coupling features of the surface morphology through temporal causal convolution and spatial graph convolution, and dynamically map the spatiotemporal coupling features into the defect probability sequence corresponding to the adaptive voxel grid through the time recursive fully connected layer;

[0010] Based on the multi-physical field coupling data in the progressive synchronous stretching process, dynamic numerical simulation is carried out through the adaptive finite element method to obtain the multi-physical field instability indicators, which are dynamically correlated with the defect probability sequence across scales, and a defect evolution prediction model based on multivariate time series is constructed. The defect evolution prediction model is trained and updated in real time through the adaptive dynamic Bayesian network inference algorithm to generate early warning signals for photovoltaic cell film quality.

[0011] In an optional embodiment,

[0012] Based on progressive synchronous stretching, the surface morphology image sequence of photovoltaic cell film is obtained. Based on the adaptive fractal dimension analysis algorithm, the surface morphology image sequence is analyzed at multiple scales to extract surface features. The surface morphology features are mapped to low-dimensional manifold space through the feature mapping method based on manifold learning. The surface morphology dynamic evolution manifold model is constructed, including:

[0013] Based on progressive synchronous stretching, the high-speed camera device and the stretching device are controlled to synchronously collect data through a synchronous trigger control circuit to obtain continuous image frames. Combined with the optical flow motion compensation algorithm, the continuous image frames are aligned with motion vectors to obtain the surface morphology image sequence of the photovoltaic cell film.

[0014] Through non-subsampled contourlet transform, each image in the surface morphology image sequence is decomposed in a multi-scale and multi-directional manner to obtain transformation subbands, and the transformation coefficients corresponding to each of the transformation subbands are determined. The transformation coefficient energy spectrum is determined by calculating the energy of the transformation coefficient, and the fractal dimension of each transformation subband is calculated by using a box counting method to construct a multi-scale fractal dimension spectrum. Through multi-scale gradient analysis of the transformation coefficient energy spectrum, the optimal scale range is adaptively determined, and the statistical features of the multi-scale fractal dimension spectrum are extracted to generate surface morphology features.

[0015] A high-dimensional data dimensionality reduction method is adopted to maintain local and global structures by adjusting the KL divergence, realize nonlinear mapping of surface morphology features to low-dimensional manifolds, and generate a low-dimensional manifold space. In the low-dimensional manifold space, the similarity between different surface morphologies is measured by geodesic distance, and a time series transition probability matrix is ​​generated. A manifold model of dynamic evolution of surface morphology is constructed to determine the dynamic evolution law of surface morphology.

[0016] In an optional embodiment,

[0017] The high-dimensional data dimensionality reduction method comprises:

[0018] Input the surface morphology features into the high-dimensional feature space, determine the high-dimensional sample points, calculate the Euclidean distance between each high-dimensional sample point, and construct the Euclidean distance matrix;

[0019] Based on the low-dimensional manifold space, randomly select position coordinates, determine low-dimensional sample points, and set the t-distribution degree of freedom parameters to determine the initial low-dimensional manifold space;

[0020] In the high-dimensional feature space, based on the Euclidean distance matrix, the conditional probability between the high-dimensional sample points is calculated, the corresponding distance values ​​are converted into similarity metrics using a Gaussian kernel function, and the probability distribution of the high-dimensional sample points is constructed;

[0021] In the initial low-dimensional manifold space, according to the position coordinates of the current low-dimensional sample points, the conditional probabilities between the low-dimensional sample points are calculated, and the probability distribution of the low-dimensional sample points is constructed using t-distribution as a similarity measure;

[0022] Calculate the KL divergence between the probability distribution of high-dimensional sample points and the probability distribution of low-dimensional sample points, construct the optimization objective function, iteratively adjust the coordinates of the low-dimensional sample points through the gradient descent method, combine with the dynamic adjustment of the t distribution degree of freedom parameter, control the balance between the local structure and the global structure, minimize the optimization objective function, obtain the probability distribution of the low-dimensional sample points that best conforms to the probability distribution of the high-dimensional sample points, and determine the optimal low-dimensional manifold space.

[0023] In an optional embodiment,

[0024] Also includes:

[0025] The probability distribution of low-dimensional sample points is as follows:

[0026]

[0027] Among them, q ij Represents the similarity probability between low-dimensional sample points, y i Indicates the coordinate position of the i-th sample point, y j represents the coordinate position of the jth sample point, ν represents the degree of freedom parameter of the t distribution, ∑k≠l represents the sum of the relationships between any different sample points among all sample points, y k Indicates the coordinate position of the kth sample point, y l Represents the coordinate position of the lth sample point;

[0028] Dynamically adjust the t distribution degree of freedom parameter, the formula is as follows:

[0029]

[0030] Where ν(c) represents the degree of freedom parameter of the t distribution at the cth iteration, c represents the cth iteration, and v min Indicates the preset minimum value of the degree of freedom parameter, v max Represents the maximum value of the preset degree of freedom parameter, C represents the total number of iterations, α represents the steepness of the change, and β represents the starting point of dynamic adjustment.

[0031] In an optional embodiment,

[0032] Based on the surface topography dynamic evolution manifold model, the corresponding surface topography is reconstructed through a deep learning three-dimensional space reconstruction algorithm to generate time-varying surface topography point cloud data, and the time-varying surface topography point cloud data is divided into adaptive voxel grids through an adaptive voxelization strategy, including:

[0033] A 3D reconstruction model is constructed based on a conditional generative adversarial network, wherein the conditional generative adversarial network includes a generator and a discriminator. The generator takes low-dimensional manifold coordinates and timestamps as conditional inputs, and generates 3D surface morphology samples through multi-layer deconvolution and jump connections. The discriminator takes the 3D surface morphology samples and the real surface morphology as inputs, and determines the authenticity measure of the 3D surface morphology samples through multi-layer convolution and downsampling. The final 3D surface morphology is generated through adversarial loss function combined with temporal consistency loss function and adversarial iteration of the generator and the discriminator. Based on the final 3D surface morphology corresponding to each timestamp, the 3D coordinate data is obtained according to the predetermined surface morphology reference point, and the time-varying surface morphology point cloud data is generated.

[0034] Based on the time-varying surface morphology point cloud data, the global three-dimensional surface morphology corresponding to the time-varying surface morphology point cloud data is recursively divided into eight sub-regions through octree space partitioning. If the granularity of the sub-region is greater than a preset coarse-grained threshold, the recursive partitioning is continued until all sub-regions reach the preset coarse-grained threshold, thereby generating coarse-grained voxels.

[0035] Performing KD tree partitioning on the local three-dimensional surface morphology corresponding to the coarse-grained voxels, selecting the coordinate axis with the maximum variance as the segmentation axis, and recursively dividing the local three-dimensional surface morphology into two sub-areas until a preset fine-grained threshold is met, thereby generating an adaptive voxelization result;

[0036] For each voxel in the adaptive voxelization result, the internal point cloud density and curvature change are calculated, the voxel size is adaptively adjusted, the voxel merging criterion based on entropy is combined, redundant voxels are merged, a time-varying adaptive voxel grid index is generated, and the spatial position, size and timestamp of each voxel are determined.

[0037] In an optional embodiment,

[0038] For each adaptive voxel grid, the multi-scale dynamic features in the adaptive voxel grid are extracted, and the multi-scale dynamic features are input into the pre-trained spatiotemporal convolutional neural network. The spatiotemporal coupling features of the surface morphology are extracted through temporal causal convolution and spatial graph convolution. The spatiotemporal coupling features are dynamically mapped to the defect probability sequence corresponding to the adaptive voxel grid through the time recursive fully connected layer, including:

[0039] Performing multi-scale decomposition on the adaptive voxel grid by three-dimensional wavelet transform to obtain voxel grid representations at different scales;

[0040] Based on the three-dimensional local binary pattern, the neighborhood of each voxel is divided into a plurality of first-class sub-regions according to a preset neighborhood radius and a preset number of neighborhood points. For each first-class sub-region, the gray value difference between the central voxel and the neighborhood voxels in the first-class sub-region is calculated to obtain a local difference vector in the first-class sub-region. The local difference vector is compared with a preset threshold to obtain a binary coding sequence in the first-class sub-region. The binary coding sequences of all the first-class sub-regions are spliced ​​to obtain a three-dimensional local binary pattern feature descriptor of the voxel, and a local texture feature is determined.

[0041] Based on the preset three-dimensional shape context radius and angle interval, the spherical area centered on the current voxel is divided into multiple second-class sub-areas. For each second-class sub-area, the radial distance and angle of all neighboring voxels in the second-class sub-area relative to the current voxel are calculated, and the radial distance and angle are quantified to obtain the shape context distribution in the second-class sub-area. The shape context distributions of all the second-class sub-areas are spliced ​​to obtain the shape context descriptor of the voxel.

[0042] Perform motion estimation on voxel grids of adjacent time steps to obtain a motion vector field of the voxel grid, calculate the gradient of the motion vector field, obtain a motion boundary image, divide the motion boundary image into multiple three-category sub-regions, calculate the gradient direction histogram for each of the three-category sub-regions, concatenate the gradient direction histograms of all the three-category sub-regions, obtain a three-dimensional motion boundary histogram of the voxel grid, and determine the temporal dynamic characteristics of the voxel grid;

[0043] Through a multi-scale feature fusion strategy, local texture features, shape context descriptors and temporal dynamic features at different scales are fused to obtain multi-scale dynamic features within an adaptive voxel grid.

[0044] In an optional embodiment,

[0045] The multi-scale dynamic features are input into the pre-trained spatiotemporal convolutional neural network, and the spatiotemporal coupling features of the surface morphology are extracted through temporal causal convolution and spatial graph convolution. The spatiotemporal coupling features are dynamically mapped into defect probability sequences corresponding to the adaptive voxel grid through the time recursive fully connected layer, including:

[0046] Inputting the multi-scale dynamic features into the spatiotemporal convolutional neural network, the spatiotemporal convolutional neural network includes a temporal causal convolution layer and a spatial graph convolution layer, the temporal causal convolution layer constructs a causal convolution kernel in the temporal dimension, performs a temporal causal convolution operation on the multi-scale dynamic features, and extracts the causal dependency in the temporal dimension; the spatial graph convolution layer constructs a graph convolution kernel in the spatial dimension, performs a spatial graph convolution operation on the multi-scale dynamic features, and extracts the local structural information in the spatial dimension;

[0047] An activation function and a pooling layer are constructed between the temporal causal convolution layer and the spatial graph convolution layer, and a spatiotemporal coupling result is obtained through nonlinear transformation and pooling operation;

[0048] By alternately stacking multiple temporal causal convolutional layers and multiple spatial graph convolutional layers, the spatiotemporal coupling results are extracted layer by layer to determine the spatiotemporal coupling characteristics;

[0049] Based on the pre-trained long short-term memory network, the dynamic patterns and dependencies of the spatiotemporal coupling features changing over time are captured, and the spatiotemporal coupling features are mapped into defect probability values ​​of each voxel in combination with the weight matrix and bias to generate a defect probability sequence.

[0050] In an optional embodiment,

[0051] Based on the multi-physics field coupling data in the progressive synchronous stretching process, dynamic numerical simulation is performed through the adaptive finite element method to obtain the multi-physics field instability indicators, and the multi-physics field instability indicators are dynamically associated with the defect probability sequence across scales to build a defect evolution prediction model based on multivariate time series. The defect evolution prediction model is trained and updated in real time through the adaptive dynamic Bayesian network inference algorithm to generate photovoltaic cell film quality warning signals including:

[0052] The mechanical sensors, thermal sensors and electrical sensors installed on the progressive synchronous stretching equipment are used to collect multi-physical field parameters including mechanical field, thermal field and electrical field, and the multi-physical field parameters are synchronously sampled, quantified and preprocessed to obtain a synchronous multi-physical field coupling data sequence;

[0053] Based on the multi-physics coupling data sequence, a multi-physics coupling finite element model including mechanical field, thermal field and electrical field is established. The multi-physics coupling finite element model is adaptively optimized and dynamically numerically solved using adaptive meshing technology and dynamic time step control technology to obtain the spatiotemporal evolution results of the multi-physics field during the progressive synchronous stretching process.

[0054] Analyze the spatiotemporal evolution results of multi-physical fields, extract mechanical instability indicators through the nonlinear change characteristics of stress-strain curves, extract thermal instability indicators through the local mutation characteristics of temperature fields, and extract electrical instability indicators through the abnormal fluctuation characteristics of electrical parameters, thus forming multi-physical field instability indicators;

[0055] The time-space scale transformation technology is used to map the instability indicators of multi-physical fields and the defect probability sequence to the same time-space scale, and the dynamic mapping relationship between the instability indicators of multi-physical fields and the defect probability sequence is established through the dynamic data fusion technology;

[0056] Based on the dynamic mapping relationship, a multivariate time series data set including a multi-physical field instability indication time series and a defect probability time series is determined, a defect evolution prediction model is constructed, and the defect evolution prediction model is trained by a pre-constructed long short-term memory neural network algorithm;

[0057] The input and output of the defect evolution prediction model are mapped to the nodes and edges of the adaptive dynamic Bayesian network. According to the newly collected multi-physics field coupling data and defect detection data, the adaptive dynamic Bayesian network is adjusted and updated in real time through the dynamic structure learning algorithm and parameter estimation algorithm to determine the optimal defect evolution prediction model.

[0058] The multi-physical field coupling data collected in real time is input into the optimal defect evolution prediction model to predict the future defect evolution trend, calculate the defect density and defect size distribution, determine the risk index, and judge the abnormal state and potential risks of the photovoltaic cell film quality based on the preset quality threshold and risk assessment rules, and generate a quality warning signal including warning level, warning time and warning location.

[0059] According to a second aspect of the embodiments of the present invention,

[0060] A photovoltaic cell film quality detection system based on synchronous stretching is provided, comprising:

[0061] The first unit is used to obtain a surface morphology image sequence of a photovoltaic cell film based on progressive synchronous stretching, perform multi-scale analysis on the surface morphology image sequence based on an adaptive fractal dimension analysis algorithm, extract surface features, map the surface morphology features to a low-dimensional manifold space through a feature mapping method based on manifold learning, construct a surface morphology dynamic evolution manifold model, reconstruct the corresponding surface morphology based on the surface morphology dynamic evolution manifold model through a deep learning three-dimensional space reconstruction algorithm, generate time-varying surface morphology point cloud data, and divide the time-varying surface morphology point cloud data into an adaptive voxel grid through an adaptive voxelization strategy;

[0062] The second unit is used to extract multi-scale dynamic features in each adaptive voxel grid, input the multi-scale dynamic features into a pre-trained spatiotemporal convolutional neural network, extract the spatiotemporal coupling features of the surface morphology through temporal causal convolution and spatial graph convolution, and dynamically map the spatiotemporal coupling features into a defect probability sequence corresponding to the adaptive voxel grid through a time-recursive fully connected layer;

[0063] The third unit is used to perform dynamic numerical simulation based on the multi-physical field coupling data in the progressive synchronous stretching process through the adaptive finite element method, obtain the multi-physical field instability indicators, dynamically associate the multi-physical field instability indicators with the defect probability sequence across scales, and build a defect evolution prediction model based on multivariate time series. The defect evolution prediction model is trained and updated in real time through the adaptive dynamic Bayesian network inference algorithm to generate a warning signal for the quality of photovoltaic cell thin films.

[0064] According to a third aspect of the embodiments of the present invention,

[0065] An electronic device is provided, comprising:

[0066] processor;

[0067] a memory for storing processor-executable instructions;

[0068] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0069] In an embodiment of the present invention, based on an adaptive fractal dimension analysis algorithm, a multi-scale analysis is performed on the surface morphology of a photovoltaic cell film, which effectively captures the structural information of the surface at different scales and accurately identifies surface roughness, defects and other features; through a manifold learning method, complex high-dimensional surface morphology features are mapped to a low-dimensional space, which simplifies the data structure and reduces the computational complexity, while retaining key features, reducing storage and computing overhead, and helping to better understand the changing pattern of the surface morphology; a surface morphology dynamic evolution manifold model is constructed to capture and visualize the changes in the surface morphology of photovoltaic films over time, which is helpful for analyzing the dynamic evolution characteristics of the film such as the aging process and stress effects; adaptive voxel division is performed on the point cloud data of the time-varying surface morphology to ensure a balance between the accuracy and efficiency of the voxel division, maintain a high resolution in dense areas, and reduce unnecessary computing resource consumption in sparse areas; through multi-scale dynamic feature extraction and spatiotemporal convolutional neural networks, it is possible to effectively capture subtle changes in the surface morphology in the time and space dimensions, ensuring the accuracy of the surface morphology. Comprehensive analysis of features; using dynamic mapping of spatiotemporal coupling features and recursive fully connected layers to associate complex changes in surface morphology with defects, generate defect probability sequences, and improve the detection accuracy of potential defects in adaptive voxel grids; through temporal causal convolution and spatial graph convolution, the spatiotemporal correlation of features is enhanced, and dynamic tracking and analysis of the evolution of complex morphology is achieved, which helps to identify potential defect trends; through the adaptive finite element method combined with multi-physics field coupling data, the multi-physics field instability of photovoltaic cell films during progressive synchronous stretching can be dynamically simulated and analyzed, so as to accurately predict the evolution of defects; multi-physics field instability indicators are dynamically associated with defect probability sequences across scales, realizing a comprehensive analysis from microscopic defects to macroscopic instability indicators, and improving the comprehensiveness and accuracy of defect prediction; using an adaptive dynamic Bayesian network inference algorithm, the defect evolution prediction model is trained and updated in real time, so that the model can continuously adapt to new data and changes, improving the timeliness and reliability of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic diagram of the flow of a photovoltaic cell film quality detection method based on synchronous stretching according to an embodiment of the present invention;

[0071] Figure 2 It is a schematic diagram of the structure of a photovoltaic cell film quality detection system based on synchronous stretching according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0074] Figure 1 FIG. 1 is a flow chart of a photovoltaic cell film quality detection method based on synchronous stretching according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0075] S101. Based on progressive synchronous stretching, a surface morphology image sequence of a photovoltaic cell film is obtained; based on an adaptive fractal dimension analysis algorithm, a multi-scale analysis is performed on the surface morphology image sequence to extract surface features; the surface morphology features are mapped to a low-dimensional manifold space through a feature mapping method based on manifold learning; a surface morphology dynamic evolution manifold model is constructed; based on the surface morphology dynamic evolution manifold model, a deep learning three-dimensional space reconstruction algorithm is used to reconstruct the corresponding surface morphology to generate time-varying surface morphology point cloud data; and the time-varying surface morphology point cloud data is divided into an adaptive voxel grid through an adaptive voxelization strategy;

[0076] In this embodiment, based on an adaptive fractal dimension analysis algorithm, a multi-scale analysis is performed on the surface morphology of the photovoltaic cell film, which effectively captures the structural information of the surface at different scales and accurately identifies surface roughness, defects and other features; through a manifold learning method, complex high-dimensional surface morphology features are mapped to a low-dimensional space, which simplifies the data structure and reduces the computational complexity, while retaining key features, reducing storage and computing overhead, and helping to better understand the changing pattern of the surface morphology; a surface morphology dynamic evolution manifold model is constructed to capture and visualize the changes in the surface morphology of the photovoltaic film over time, which is helpful for analyzing the dynamic evolution characteristics of the film such as the aging process and stress effect; based on the manifold model, the surface morphology is reconstructed through a deep learning three-dimensional reconstruction algorithm to generate high-precision three-dimensional point cloud data of the time-varying surface morphology, thereby improving the visualization and evaluation capabilities of the film surface morphology; the point cloud data of the time-varying surface morphology is adaptively voxel-divided to ensure a balance between the accuracy and efficiency of the voxel division, maintain a high resolution in dense areas, and reduce unnecessary computing resource consumption in sparse areas.

[0077] In an optional embodiment, based on progressive synchronous stretching, a surface morphology image sequence of a photovoltaic cell film is obtained, and based on an adaptive fractal dimension analysis algorithm, a multi-scale analysis is performed on the surface morphology image sequence to extract surface features, and the surface morphology features are mapped to a low-dimensional manifold space through a feature mapping method based on manifold learning, and a surface morphology dynamic evolution manifold model is constructed, including:

[0078] Based on progressive synchronous stretching, the high-speed camera device and the stretching device are controlled to synchronously collect data through a synchronous trigger control circuit to obtain continuous image frames. Combined with the optical flow motion compensation algorithm, the continuous image frames are aligned with motion vectors to obtain the surface morphology image sequence of the photovoltaic cell film.

[0079] Through non-subsampled contourlet transform, each image in the surface morphology image sequence is decomposed in a multi-scale and multi-directional manner to obtain transformation subbands, and the transformation coefficients corresponding to each of the transformation subbands are determined. The transformation coefficient energy spectrum is determined by calculating the energy of the transformation coefficient, and the fractal dimension of each transformation subband is calculated by using a box counting method to construct a multi-scale fractal dimension spectrum. Through multi-scale gradient analysis of the transformation coefficient energy spectrum, the optimal scale range is adaptively determined, and the statistical features of the multi-scale fractal dimension spectrum are extracted to generate surface morphology features.

[0080] A high-dimensional data dimensionality reduction method is adopted to maintain local and global structures by adjusting the KL divergence, realize nonlinear mapping of surface morphology features to low-dimensional manifolds, and generate a low-dimensional manifold space. In the low-dimensional manifold space, the similarity between different surface morphologies is measured by geodesic distance, and a time series transition probability matrix is ​​generated. A manifold model of dynamic evolution of surface morphology is constructed to determine the dynamic evolution law of surface morphology.

[0081] Acquire a sequence of surface morphology images based on progressive synchronous stretching; build a synchronous trigger control circuit, select a high-precision programmable controller, and realize synchronous trigger control of the high-speed camera device and the progressive stretching device through programming; select a high-resolution high-speed camera as the high-speed camera device, determine the appropriate lens magnification and depth of field according to the surface characteristic scale of the photovoltaic cell film, and adjust the light source arrangement to obtain a clear and stable surface image; fix the photovoltaic cell film on the fixture of the progressive stretching device, and realize the progressive uniaxial stretching of the photovoltaic cell film by controlling the driving motor of the stretching device; set a reasonable stretching rate and strain range, through a synchronous trigger control circuit, control the high-speed camera device to continuously collect the surface image of the photovoltaic cell film at a fixed time interval, preferably 0.1s, until the stretching process within the preset strain range is completed; for the obtained continuous image frame sequence, use the optical flow method, preferably the Horn-Schunck algorithm, to calculate the motion vector field between adjacent image frames; according to the motion vector field, perform motion compensation alignment on the continuous image frames to correct the image position offset caused by the displacement of the photovoltaic cell film during the stretching process; sort the image frames after motion compensation alignment according to the size of the stretching strain, and obtain an image sequence of the surface morphology of the photovoltaic cell film at different stretching stages.

[0082] Surface morphology features are extracted based on non-subsampled contourlet transform; for each image in the surface morphology image sequence, multi-scale and multi-directional decomposition is performed using non-subsampled contourlet transform; in the process of non-subsampled contourlet transform decomposition, sparse representation of images at different scales and directions is achieved by designing appropriate non-subsampled pyramid filters and non-subsampled directional filters; for the non-subsampled contourlet transform decomposition results of each image, a low-frequency subband and a series of high-frequency subbands are obtained, and the non-subsampled contourlet transform coefficients corresponding to each subband are extracted; the energy of the non-subsampled contourlet transform coefficients in each subband is calculated. The energy spectrum of the transformation coefficients is obtained to reflect the significance of the image detail features at different scales and directions; the fractal dimension of the non-subsampled contourlet transform coefficients in each sub-band is estimated by using the differential box counting method, and the multi-scale fractal dimension spectrum of the surface morphology of the photovoltaic cell film is constructed through the collection of multi-scale fractal dimensions; the multi-scale fractal dimension spectrum is analyzed, and the optimal scale range that can best reflect the surface morphology characteristics is adaptively determined by calculating the fractal dimension gradient between different scales; within the optimal scale range, the statistical characteristics of the multi-scale fractal dimension spectrum, such as mean, variance, entropy, etc., are extracted and combined into a surface morphology feature vector.

[0083] Based on high-dimensional data dimensionality reduction, a manifold model of the dynamic evolution of surface morphology is constructed; the obtained high-dimensional surface morphology feature vector is subjected to dimensionality reduction analysis using the manifold learning method; in the dimensionality reduction process, the KL divergence between the low-dimensional manifold after dimensionality reduction and the original high-dimensional manifold is minimized to balance the local structure and the global structure, and realize the nonlinear mapping of surface morphology features to the low-dimensional manifold space; the hyperparameters of manifold learning, such as the manifold dimension, are optimized by cross-validation and other methods to obtain the optimal low-dimensional manifold embedding; in the generated low-dimensional manifold space, the geodesic distance is introduced to measure the similarity between the surface morphologies of photovoltaic cell films at different stretching stages, and the geodesic distance matrix of the surface morphology is constructed; based on the geodesic distance matrix The transfer similarity between surface morphologies at different stretching stages is calculated through dynamic time warping, Gaussian kernel function and other methods, and the time series transfer probability matrix of the dynamic evolution of surface morphology is generated; based on the time series transfer probability matrix, a Markov chain model of the dynamic evolution of surface morphology is constructed, and the dynamic evolution behavior of the surface morphology of photovoltaic cell thin film is predicted and analyzed by solving the stationary distribution and absorption state of the Markov chain; combining the material mechanics theory and failure mechanism analysis, the Markov chain model is associated with the macroscopic mechanical behavior, and a manifold mapping model of multi-scale surface morphology dynamic evolution and mechanical failure is established, revealing the evolution law of the microscopic morphology of photovoltaic cell thin films and its intrinsic connection with the macroscopic mechanical properties.

[0084] Exemplarily, a flexible perovskite photovoltaic cell is placed on a stretching device, and through synchronous trigger control, during the progressive stretching process with a strain of 10% per step, a high-speed camera continuously collects surface images at a speed of 1000 frames per second, and finally obtains 100 frames of continuous images; for the obtained continuous image frame sequence, the optical flow method is used to estimate and align the motion vector, and a surface morphology image sequence in 10 stages is obtained; a non-subsampled contourlet transform is used to perform 5-scale and 8-directional decomposition on each surface morphology image, and the transformation coefficients of each transformation subband are extracted, the transformation coefficient energy spectrum is calculated, and the box counting method is used to calculate the fractal dimension of each subband, and a multi-scale fractal dimension spectrum is constructed; through the 3-scale gradient analysis of the transformation coefficient energy spectrum, the scale 3-5 is determined to be the optimal scale, and the scale is extracted. The 40-dimensional surface morphology feature vector is obtained by using the statistical characteristics of the mean, variance and other statistical characteristics of the multi-scale fractal dimension spectrum within the scale range; the 40-dimensional surface morphology features are reduced to the 3-dimensional manifold space by using the local linear embedding method to obtain the low-dimensional manifold embedding of the surface morphology; in the low-dimensional manifold space, the similarity between the surface morphologies at different stretching stages is calculated by the geodesic distance to generate a 10×10 timing transition probability matrix; based on the timing transition probability matrix, a manifold model of the dynamic evolution of the surface morphology is constructed, and it is found through analysis that: with the increase of tensile strain, microcracks gradually appear and expand on the surface of the flexible perovskite photovoltaic cell. When the strain exceeds 60%, the density and size of the microcracks increase rapidly, and the surface morphology changes significantly, until macro cracks appear at 80% strain, leading to device failure.

[0085] In this embodiment, the surface morphology image is decomposed and analyzed in a multi-scale and multi-directional manner based on the non-subsampled contourlet transform, and combined with the fractal dimension extraction method, the complex features of the film surface at different scales can be efficiently extracted, and the sensitivity to detail changes is enhanced; through the manifold learning method, a manifold model of the dynamic evolution of the surface morphology is constructed to achieve dimensionality reduction of high-dimensional surface features, effectively reduce the complexity of the data dimension, and retain the key information of the morphological features, so as to facilitate the analysis of the law of film morphology evolution with stretching; based on the geodesic distance matrix and the Markov chain model, the transfer similarity of the surface morphology of the photovoltaic cell film between different stretching stages can be accurately calculated, providing an effective method to predict the dynamic evolution process of the film; through the multi-scale manifold mapping model, the dynamic evolution of the microscopic morphology is associated with the macroscopic mechanical properties, revealing the intrinsic connection between the morphological changes of the film and its mechanical properties and failure behavior, which is helpful to optimize the material design and service life prediction of photovoltaic films.

[0086] In an optional embodiment, the high-dimensional data dimensionality reduction method includes:

[0087] Input the surface morphology features into the high-dimensional feature space, determine the high-dimensional sample points, calculate the Euclidean distance between each high-dimensional sample point, and construct the Euclidean distance matrix;

[0088] Based on the low-dimensional manifold space, randomly select position coordinates, determine low-dimensional sample points, and set the t-distribution degree of freedom parameters to determine the initial low-dimensional manifold space;

[0089] In the high-dimensional feature space, based on the Euclidean distance matrix, the conditional probability between the high-dimensional sample points is calculated, the corresponding distance values ​​are converted into similarity metrics using a Gaussian kernel function, and the probability distribution of the high-dimensional sample points is constructed;

[0090] In the initial low-dimensional manifold space, according to the position coordinates of the current low-dimensional sample points, the conditional probabilities between the low-dimensional sample points are calculated, and the probability distribution of the low-dimensional sample points is constructed using t-distribution as a similarity measure;

[0091] Calculate the KL divergence between the probability distribution of high-dimensional sample points and the probability distribution of low-dimensional sample points, construct the optimization objective function, iteratively adjust the coordinates of the low-dimensional sample points through the gradient descent method, combine with the dynamic adjustment of the t distribution degree of freedom parameter, control the balance between the local structure and the global structure, minimize the optimization objective function, obtain the probability distribution of the low-dimensional sample points that best conforms to the probability distribution of the high-dimensional sample points, and determine the optimal low-dimensional manifold space.

[0092] Construct a Euclidean distance matrix of high-dimensional sample points; input the multi-scale fractal dimension spectrum feature vectors of the photovoltaic cell film surface morphology into the high-dimensional feature space, each feature vector corresponds to a sample point in the high-dimensional space; calculate the Euclidean distance between any two sample points in the high-dimensional space, that is, the square root of the sum of the squares of the differences of the components of the corresponding feature vectors; construct a Euclidean distance matrix between the high-dimensional sample points based on the calculated Euclidean distance, the number of rows and columns of the matrix is ​​equal to the number of sample points, and the matrix element (i, j) represents the Euclidean distance from the i-th sample point to the j-th sample point.

[0093] Initialize the low-dimensional manifold space and determine the dimension of the target low-dimensional manifold space, which is generally selected based on the inherent dimension of the high-dimensional space and the information retention rate after dimensionality reduction; in the low-dimensional space of the target dimension, randomly select position coordinates equal to the number of high-dimensional sample points, and use these coordinate points as the initial sample points in the low-dimensional space; set the degree of freedom parameter of the t distribution to control the similarity measure between sample points in the low-dimensional space. The smaller the degree of freedom, the more concentrated the similarity is in the local neighborhood, and the larger the degree of freedom, the more evenly distributed the similarity is.

[0094] Construct the probability distribution of high-dimensional sample points; calculate the conditional probability between any two high-dimensional sample points based on the Euclidean distance matrix in high-dimensional space. The conditional probability reflects the preference of one sample point for choosing another sample point as its neighborhood; use the Gaussian kernel function to convert the Euclidean distance into a similarity measure. The smaller the distance between sample point pairs, the higher the similarity, and the larger the distance between sample point pairs, the lower the similarity; obtain the probability distribution of high-dimensional sample points by normalizing the conditional probabilities between all sample point pairs. The probability distribution reflects the neighborhood structure relationship of sample points in high-dimensional space.

[0095] Construct the probability distribution of low-dimensional sample points; in the low-dimensional manifold space, calculate the conditional probability between any two low-dimensional sample points according to the current sample point coordinates; use t-distribution as the similarity measure between sample points in the low-dimensional space. Compared with the Gaussian kernel function, t-distribution has a higher probability density when the distance is large and is more suitable for capturing the global structure of the low-dimensional manifold; obtain the probability distribution of low-dimensional sample points by normalizing the conditional probabilities between all pairs of low-dimensional sample points. The probability distribution reflects the neighborhood relationship of sample points in the low-dimensional manifold space.

[0096] Optimize the low-dimensional manifold space; calculate the KL divergence between the probability distribution of high-dimensional sample points and the probability distribution of low-dimensional sample points, and the KL divergence measures the difference between the two probability distributions; take minimizing the KL divergence as the optimization goal, construct the objective function, and the independent variable of the objective function is the coordinates of the low-dimensional sample points; use the gradient descent method to optimize and solve the objective function, and gradually reduce the KL divergence by iteratively updating the coordinates of the low-dimensional sample points; in the optimization process, dynamically adjust the degree of freedom parameter of the t distribution, use a smaller degree of freedom in the early stage to maintain the local structure, and gradually increase the degree of freedom in the later stage to take into account the global structure; when the KL divergence is reduced to the set threshold or reaches the maximum number of iterations, stop the optimization process and obtain the optimal low-dimensional manifold space, at which time the probability distribution of the low-dimensional sample points is closest to the probability distribution of the high-dimensional sample points, and the low-dimensional manifold maintains the inherent structure of the high-dimensional space to the greatest extent.

[0097] Exemplarily, 10 perovskite photovoltaic cell films with different formulations were taken as the research objects, 5 samples were prepared for each formulation, and a total of 50 film samples were obtained; multi-scale fractal analysis was performed on the surface morphology image of each film sample to obtain 50 40-dimensional feature vectors to constitute high-dimensional sample points; the Euclidean distances between the 50 high-dimensional sample points were calculated to construct a 50*50 distance matrix; the target dimension was selected as 3 dimensions, and the coordinates of 50 sample points were randomly initialized in the low-dimensional space, and the t-distribution degree of freedom parameter was set to 1; the Gaussian kernel conditional probability of the Euclidean distance in the high-dimensional space was calculated, and normalized to obtain the probability distribution of the high-dimensional sample points; the conditional probability of the t-distribution in the low-dimensional space was calculated, and normalized to obtain the probability distribution of the low-dimensional sample points; the KL divergence of the high-dimensional probability distribution and the low-dimensional probability distribution was calculated, the low-dimensional coordinates were optimized based on the gradient descent method, and the t-distribution degree of freedom was dynamically increased to 100; after 500 iterative optimizations, the KL divergence dropped from the initial 5.12 to 0.35, and the optimal 3D manifold space was obtained. In the visualized 3D space, the film samples of 10 formulations showed a clear clustering structure, indicating that the low-dimensional manifold well maintained the inherent pattern of high-dimensional surface morphology features, laying the foundation for further analysis of the differences between different formulations.

[0098] In an optional embodiment, it also includes:

[0099] The probability distribution of low-dimensional sample points is as follows:

[0100]

[0101] Among them, q ij Represents the similarity probability between low-dimensional sample points, y i Indicates the coordinate position of the i-th sample point, y j represents the coordinate position of the jth sample point, ν represents the degree of freedom parameter of the t distribution, ∑k≠l represents the sum of the relationships between any different sample points among all sample points, yk Indicates the coordinate position of the kth sample point, y l Represents the coordinate position of the lth sample point;

[0102] Dynamically adjust the t distribution degree of freedom parameter, the formula is as follows:

[0103]

[0104] Where ν(c) represents the degree of freedom parameter of the t distribution at the cth iteration, c represents the cth iteration, and v min Indicates the preset minimum value of the degree of freedom parameter, v max Represents the maximum value of the preset degree of freedom parameter, C represents the total number of iterations, α represents the steepness of the change, and β represents the starting point of dynamic adjustment.

[0105] The similarity probability formula of low-dimensional sample points is used to calculate the similarity probability between low-dimensional sample points. The numerator in the formula represents the similarity measure of the distance between two sample points after t-distribution processing. The larger the distance, the smaller the similarity. The denominator is the sum of the similarities between all pairs of sample points, which is used to normalize the similarity so that it meets the requirements of probability distribution.

[0106] The formula for dynamically adjusting the degrees of freedom parameter of the t-distribution describes how to dynamically adjust the degrees of freedom parameter of the t-distribution during the iteration process. The formula gradually adjusts the degrees of freedom within a certain range, that is, between a minimum and a maximum value, through an S-shaped function based on the number of iterations. As the number of iterations increases, the degrees of freedom parameter gradually transitions from a smaller value to a larger value, and the speed and starting point of the transition are controlled by two parameters. This process ensures that the degrees of freedom of the t-distribution are small in the early stages of the iteration, making the model more focused on local structure; while the degrees of freedom gradually increase in the later stages of the iteration, increasing the focus on global structure.

[0107] In this embodiment, by dynamically adjusting the degree of freedom parameter of the t distribution, the similarity probability distribution of sample points in the low-dimensional space can be adaptively adjusted according to the iterative process, so as to more accurately reflect the similarity of sample points in the high-dimensional space; the dynamic adjustment formula of the degree of freedom parameter ensures that the t distribution can be gradually optimized as the number of iterations changes, so that at different iteration stages, it can not only capture global structural features but also optimize local details, thereby improving the stability and flexibility of the dimensionality reduction model; through the gradual dynamic adjustment of the degree of freedom parameter, the importance of global structure and local structure in the low-dimensional manifold space can be balanced, so that the dimensionality reduction results are more consistent with the characteristic distribution of the high-dimensional space at different scales; using the sigmoid function to control the change of the degree of freedom parameter, the adjustment process of the t distribution can be more smoothly controlled, avoiding excessively drastic parameter fluctuations, and ensuring that the dimensionality reduction process is more stable and efficient.

[0108] In an optional embodiment, based on the surface morphology dynamic evolution manifold model, the corresponding surface morphology is reconstructed by a deep learning three-dimensional space reconstruction algorithm to generate time-varying surface morphology point cloud data, and the time-varying surface morphology point cloud data is divided into an adaptive voxel grid by an adaptive voxelization strategy, including:

[0109] A 3D reconstruction model is constructed based on a conditional generative adversarial network, wherein the conditional generative adversarial network includes a generator and a discriminator. The generator takes low-dimensional manifold coordinates and timestamps as conditional inputs, and generates 3D surface morphology samples through multi-layer deconvolution and jump connections. The discriminator takes the 3D surface morphology samples and the real surface morphology as inputs, and determines the authenticity measure of the 3D surface morphology samples through multi-layer convolution and downsampling. The final 3D surface morphology is generated through adversarial loss function combined with temporal consistency loss function and adversarial iteration of the generator and the discriminator. Based on the final 3D surface morphology corresponding to each timestamp, the 3D coordinate data is obtained according to the predetermined surface morphology reference point, and the time-varying surface morphology point cloud data is generated.

[0110] Based on the time-varying surface morphology point cloud data, the global three-dimensional surface morphology corresponding to the time-varying surface morphology point cloud data is recursively divided into eight sub-regions through octree space partitioning. If the granularity of the sub-region is greater than a preset coarse-grained threshold, the recursive partitioning is continued until all sub-regions reach the preset coarse-grained threshold, thereby generating coarse-grained voxels.

[0111] Performing KD tree partitioning on the local three-dimensional surface morphology corresponding to the coarse-grained voxels, selecting the coordinate axis with the maximum variance as the segmentation axis, and recursively dividing the local three-dimensional surface morphology into two sub-areas until a preset fine-grained threshold is met, thereby generating an adaptive voxelization result;

[0112] For each voxel in the adaptive voxelization result, the internal point cloud density and curvature change are calculated, the voxel size is adaptively adjusted, the voxel merging criterion based on entropy is combined, redundant voxels are merged, a time-varying adaptive voxel grid index is generated, and the spatial position, size and timestamp of each voxel are determined.

[0113] Generate three-dimensional surface morphology based on conditional generative adversarial network; construct the generator and discriminator of conditional generative adversarial network, the generator takes low-dimensional manifold coordinates and timestamp as conditional input, and the discriminator takes the generated three-dimensional surface morphology sample and the real surface morphology as input; in the generator, the conditional input is mapped to the high-dimensional feature space through a fully connected layer, and then through multi-layer deconvolution operations, the feature map size is gradually enlarged, and at the same time, the shallow features are fused with the deep features through jump connections, and finally a three-dimensional sample with the same size as the real surface morphology is generated; in the discriminator, the generated three-dimensional sample and the real surface morphology are respectively input into the convolutional neural network, and multi-scale features are extracted through multi-layer convolution and downsampling operations, and finally through the fully connected The connection layer is mapped to the binary classification probability, which indicates the probability that the input sample is the real surface morphology. An adversarial loss function is designed to make the three-dimensional samples generated by the generator as close to the real surface morphology as possible. At the same time, a temporal consistency loss function is introduced to ensure the continuity of the generated surface morphology in the time dimension. Through alternating training of the generator and the discriminator, the loss function of the generator is minimized and the classification accuracy of the discriminator is maximized. After multiple rounds of iterative optimization, a generative model that can generate high-quality and temporally coherent three-dimensional surface morphology is obtained. Using the trained generative model, the corresponding three-dimensional surface morphology is generated for each timestamp, and the three-dimensional coordinate data is extracted according to the pre-set surface morphology reference points to construct a time-varying surface morphology point cloud dataset.

[0114] Coarse-grained voxel partitioning: for each time stamp corresponding to the surface morphology point cloud data, determine its three-dimensional space boundary and construct a bounding box; adopt the octree space partitioning method, with the bounding box as the root node, recursively divide the three-dimensional space into eight sub-regions, each sub-region corresponds to a voxel; calculate the number and density of point clouds in each sub-region, if the density is greater than the preset coarse-grained threshold, then continue to recursively divide the sub-region; repeat the execution until the point cloud density of all sub-regions is less than or equal to the coarse-grained threshold, generate a coarse-grained voxel set, each voxel represents the local three-dimensional surface morphology.

[0115] Adaptive voxelization; for each voxel in the coarse-grained voxel set, extract the local point cloud data inside it; perform principal component analysis on the local point cloud data, calculate the variance of the point cloud on the three coordinate axes, and select the coordinate axis with the largest variance as the segmentation axis; recursively divide the local point cloud data into two sub-regions with the segmentation axis as the boundary, and generate a new voxel in each sub-region; calculate the point cloud density and curvature change in each new voxel, and if the density is greater than the preset fine-grained threshold and the curvature change is greater than the preset threshold, continue the recursive division; repeat until all voxels meet the fine-grained and curvature change requirements, and generate an adaptive voxelization result.

[0116] Time-varying adaptive voxel grid index; for each voxel in the adaptive voxelization result, calculate its internal point cloud density and curvature change; according to the point cloud density and curvature change, adaptively adjust the size of each voxel, for areas with high density and large curvature change, use smaller voxels, for areas with low density and small curvature change, use larger voxels; based on the entropy-based voxel merging criterion, calculate the information entropy of adjacent voxels, if the reduction in information entropy after merging is less than the preset threshold, merge the adjacent voxels into a larger voxel; repeat until the voxels can no longer be merged, and obtain the final time-varying adaptive voxel grid model; for each voxel, record its spatial position, size and timestamp information, and generate a time-varying adaptive voxel grid index to facilitate subsequent three-dimensional reconstruction and visualization analysis.

[0117] In this embodiment, a conditional generative adversarial network is used to gradually generate high-resolution, time-coherent three-dimensional surface morphology in combination with low-dimensional manifold coordinates and timestamps as input conditions, ensuring that the generated results can accurately reflect the surface features and maintain consistency in the time dimension; through coarse-grained voxel division and adaptive voxelization methods, the voxel sizes of different regions are dynamically adjusted, and voxels are adaptively refined or merged according to changes in point cloud density and curvature, thereby improving the accuracy and efficiency of three-dimensional reconstruction and ensuring accurate capture of local details; using a time-varying adaptive voxel grid indexing method, the generated three-dimensional surface morphology not only considers the spatial dimension, but also performs adaptive adjustments based on the time dimension information, thereby achieving more accurate dynamic change tracking and facilitating subsequent three-dimensional reconstruction and visualization analysis; through the entropy voxel merging criterion and the calculation of density, curvature and other information, the generation of redundant voxels is avoided, the complexity of data processing and storage is reduced, and the overall calculation efficiency is improved.

[0118] S102. For each adaptive voxel grid, extract multi-scale dynamic features within the adaptive voxel grid, input the multi-scale dynamic features into a pre-trained spatiotemporal convolutional neural network, extract the spatiotemporal coupling features of the surface morphology through temporal causal convolution and spatial graph convolution, and dynamically map the spatiotemporal coupling features into a defect probability sequence corresponding to the adaptive voxel grid through a time-recursive fully connected layer;

[0119] In this embodiment, through multi-scale dynamic feature extraction and spatiotemporal convolutional neural network, subtle changes in surface morphology in time and space dimensions can be effectively captured, ensuring a comprehensive analysis of surface features; by utilizing spatiotemporal coupling features and dynamic mapping of recursive fully connected layers, complex changes in surface morphology are associated with defects, and defect probability sequences are generated, thereby improving the detection accuracy of potential defects within the adaptive voxel grid; through temporal causal convolution and spatial graph convolution, the spatiotemporal correlation of features is enhanced, and dynamic tracking and analysis of the evolution process of complex morphology is achieved, which helps to identify potential defect trends; the extraction of multi-scale dynamic features enables the method to simultaneously analyze large-scale and local details, which helps to better understand the evolution of surface morphology and the formation mechanism of defects.

[0120] In an optional embodiment, for each adaptive voxel grid, multi-scale dynamic features in the adaptive voxel grid are extracted, the multi-scale dynamic features are input into a pre-trained spatiotemporal convolutional neural network, and the spatiotemporal coupling features of the surface morphology are extracted through temporal causal convolution and spatial graph convolution. The spatiotemporal coupling features are dynamically mapped into a defect probability sequence corresponding to the adaptive voxel grid through a time-recursive fully connected layer, including:

[0121] Performing multi-scale decomposition on the adaptive voxel grid by three-dimensional wavelet transform to obtain voxel grid representations at different scales;

[0122] Based on the three-dimensional local binary pattern, the neighborhood of each voxel is divided into a plurality of first-class sub-regions according to a preset neighborhood radius and a preset number of neighborhood points. For each first-class sub-region, the gray value difference between the central voxel and the neighborhood voxels in the first-class sub-region is calculated to obtain a local difference vector in the first-class sub-region. The local difference vector is compared with a preset threshold to obtain a binary coding sequence in the first-class sub-region. The binary coding sequences of all the first-class sub-regions are spliced ​​to obtain a three-dimensional local binary pattern feature descriptor of the voxel, and a local texture feature is determined.

[0123] Based on the preset three-dimensional shape context radius and angle interval, the spherical area centered on the current voxel is divided into multiple second-class sub-areas. For each second-class sub-area, the radial distance and angle of all neighboring voxels in the second-class sub-area relative to the current voxel are calculated, and the radial distance and angle are quantified to obtain the shape context distribution in the second-class sub-area. The shape context distributions of all the second-class sub-areas are spliced ​​to obtain the shape context descriptor of the voxel.

[0124] Perform motion estimation on voxel grids of adjacent time steps to obtain a motion vector field of the voxel grid, calculate the gradient of the motion vector field, obtain a motion boundary image, divide the motion boundary image into multiple three-category sub-regions, calculate the gradient direction histogram for each of the three-category sub-regions, concatenate the gradient direction histograms of all the three-category sub-regions, obtain a three-dimensional motion boundary histogram of the voxel grid, and determine the temporal dynamic characteristics of the voxel grid;

[0125] Through a multi-scale feature fusion strategy, local texture features, shape context descriptors and temporal dynamic features at different scales are fused to obtain multi-scale dynamic features within an adaptive voxel grid.

[0126] Select appropriate three-dimensional wavelet basis functions, such as Haar wavelet and Daubechies wavelet, and determine the scale layer number of wavelet transform according to the spatial size and resolution requirements of the voxel grid; fill the adaptive voxel grid data so that its size in the three coordinate axis directions is an integer power of 2, which is convenient for wavelet transform; perform three-dimensional wavelet forward transform on the filled voxel grid data to obtain wavelet coefficients at different scales, and the wavelet coefficients include low-frequency approximation coefficients and high-frequency detail coefficients; filter the wavelet coefficients according to a preset threshold to remove noise and redundant information to obtain denoised wavelet coefficients; perform three-dimensional wavelet inverse transform on the denoised wavelet coefficients to reconstruct voxel grid representations at different scales, and each scale corresponds to a spatial resolution and detail level.

[0127] For each voxel grid at each scale, a voxel is selected as the central voxel, and its neighborhood range is determined according to the preset neighborhood radius; within the neighborhood range, uniform sampling is performed according to the preset number of neighborhood points to obtain multiple discrete neighborhood voxels; the neighborhood voxels are divided into multiple first-class sub-regions with the central voxel as the boundary, and each first-class sub-region corresponds to a direction; for each first-class sub-region, the grayscale value difference between all neighboring voxels and the central voxel in the sub-region is calculated to obtain a local difference vector in the first-class sub-region; the local difference vector is compared with a preset threshold, and the difference greater than the threshold is assigned a value of 1, and the difference less than or equal to the threshold is assigned a value of 0, to obtain a binary coding sequence in the first-class sub-region; the binary coding sequences of all first-class sub-regions are spliced ​​in a clockwise or counterclockwise direction to obtain a three-dimensional local binary pattern feature descriptor of the central voxel, which reflects the local texture characteristics in the voxel neighborhood.

[0128] For each voxel grid at each scale, a voxel is selected as the central voxel, and the spherical neighborhood range is determined according to the preset shape context radius with it as the center of the sphere; the spherical neighborhood is divided into multiple Class II sub-regions along the radial and angular directions according to the preset angle interval, and each Class II sub-region corresponds to a radial distance and angle range; for each Class II sub-region, the radial distance and angle of all neighboring voxels in the sub-region relative to the central voxel are calculated; the radial distance is quantified and divided into multiple discrete distance intervals, and the angle is quantified and divided into multiple discrete angle intervals; the number of neighboring voxels in each quantized distance interval and angle interval is counted to obtain the shape context distribution in the Class II sub-region, which reflects the distribution characteristics of the neighboring voxels in the radial and angular directions; the shape context distributions of all Class II sub-regions are spliced ​​in a clockwise or counterclockwise direction to obtain the three-dimensional shape context descriptor of the central voxel, which reflects the geometric shape characteristics in the voxel neighborhood.

[0129] For the voxel grids of two adjacent time steps, the motion vector of the voxels is estimated by the optical flow method or the block matching method to obtain the three-dimensional motion vector field of the voxel grid; the gradient of the motion vector field is calculated to obtain a gradient image reflecting the motion change, the gradient direction represents the motion direction, and the gradient amplitude represents the motion speed; the gradient image is threshold segmented to extract significant motion boundaries to obtain a motion boundary image; the motion boundary image is divided into multiple three-category sub-regions, each of which corresponds to a spatial position range; for each of the three-category sub-regions, the gradient direction of all pixels in the sub-region is calculated, the gradient direction is quantized into multiple discrete direction intervals, the number of pixels in each direction interval is counted, and the gradient direction histogram of the three-category sub-region is obtained, which reflects the directional distribution characteristics of the local motion boundary; the gradient direction histograms of all the three-category sub-regions are spliced ​​to obtain the three-dimensional motion boundary histogram of the voxel grid, which reflects the dynamic change characteristics of the voxel grid in the time dimension.

[0130] For each voxel grid at each scale, its local texture features, shape context descriptors and temporal dynamic features are extracted to obtain multiple feature vectors; the feature vectors at different scales are normalized to make their scales and ranges consistent; the feature vectors at different scales are fused by weighted averaging or cascading to obtain a fused multi-scale feature vector; the fused multi-scale feature vector is subjected to dimensionality reduction processing, such as principal component analysis, independent component analysis, etc., to remove redundant information and extract a compact feature representation; the multi-scale features after dimensionality reduction are used as dynamic feature descriptors of the adaptive voxel grid for subsequent 3D reconstruction, recognition and analysis tasks.

[0131] In this embodiment, through three-dimensional wavelet transform, local binary pattern, shape context descriptor and time dynamic feature extraction, the local texture, geometric shape and dynamic changes in the time dimension of the surface morphology are comprehensively captured, which helps to generate more refined and diverse feature expressions; through three-dimensional wavelet transform and threshold filtering, noise and redundant information can be removed to ensure that the voxel grid representation at different scales is clearer, which is helpful for subsequent analysis and recognition tasks; spatial texture features, geometric shape features and time dynamic features are fused to form a powerful spatiotemporal feature descriptor, which enhances the tracking and analysis capabilities of the dynamic changes of the three-dimensional voxel grid; through the fusion and dimensionality reduction of multi-scale features, redundant information is removed, the feature expression is simplified, and the subsequent three-dimensional reconstruction, recognition and analysis are more efficient and accurate.

[0132] In an optional embodiment, the multi-scale dynamic features are input into a pre-trained spatiotemporal convolutional neural network, and the spatiotemporal coupling features of the surface morphology are extracted through temporal causal convolution and spatial graph convolution. The spatiotemporal coupling features are dynamically mapped into a defect probability sequence corresponding to an adaptive voxel grid through a time-recursive fully connected layer, including:

[0133] Inputting the multi-scale dynamic features into the spatiotemporal convolutional neural network, the spatiotemporal convolutional neural network includes a temporal causal convolution layer and a spatial graph convolution layer, the temporal causal convolution layer constructs a causal convolution kernel in the temporal dimension, performs a temporal causal convolution operation on the multi-scale dynamic features, and extracts the causal dependency in the temporal dimension; the spatial graph convolution layer constructs a graph convolution kernel in the spatial dimension, performs a spatial graph convolution operation on the multi-scale dynamic features, and extracts the local structural information in the spatial dimension;

[0134] An activation function and a pooling layer are constructed between the temporal causal convolution layer and the spatial graph convolution layer, and a spatiotemporal coupling result is obtained through nonlinear transformation and pooling operation;

[0135] By alternately stacking multiple temporal causal convolutional layers and multiple spatial graph convolutional layers, the spatiotemporal coupling results are extracted layer by layer to determine the spatiotemporal coupling characteristics;

[0136] Based on the pre-trained long short-term memory network, the dynamic patterns and dependencies of the spatiotemporal coupling features changing over time are captured, and the spatiotemporal coupling features are mapped into defect probability values ​​of each voxel in combination with the weight matrix and bias to generate a defect probability sequence.

[0137] According to the time dimension length of the input multi-scale dynamic features, the number of layers of the temporal causal convolution layer and the number of causal convolution kernels in each layer are determined; the causal convolution kernel slides on the time dimension, only considering the time step before the current moment, avoiding the use of information at future moments, and reflecting the temporal causality; the size of the causal convolution kernel is designed, usually an odd number, such as 3×3×3, to ensure that the center of the causal convolution kernel is located at the current time step; through the causal convolution operation, the causal dependency of the multi-scale dynamic features in the time dimension is extracted to obtain the feature representation in the time dimension;

[0138] According to the spatial dimension of the input multi-scale dynamic features, the number of spatial graph convolution layers and the number of graph convolution kernels in each layer are determined; the graph convolution kernel slides in the spatial dimension, considering the connection relationship between neighboring voxels to extract local structural information; the size of the graph convolution kernel is designed, usually an odd number, such as 5×5×5, to ensure that the center of the graph convolution kernel is located at the current voxel position; through the graph convolution operation, the local structural information of the multi-scale dynamic features in the spatial dimension is extracted to obtain the feature representation in the spatial dimension;

[0139] Insert activation functions and pooling layers between the temporal causal convolution layer and the spatial graph convolution layer to introduce nonlinear transformation and feature aggregation operations; the activation function can be ReLU, Leaky ReLU, ELU, etc., to improve the nonlinear expression ability of the network and enhance the discriminability of features; the pooling layer can select maximum pooling, average pooling, etc. to aggregate adjacent features, reduce the size of the feature map, and improve the robustness of the features; through the combination of activation functions and pooling layers, the spatiotemporal coupling results are obtained, reflecting the interactive influence of the time dimension and the spatial dimension;

[0140] Multiple temporal causal convolutional layers and spatial graph convolutional layers are stacked alternately to build a deep spatiotemporal convolutional neural network. The spatiotemporal convolutions of the first few layers mainly extract local spatiotemporal features, while the spatiotemporal convolutions of the latter layers mainly extract global spatiotemporal features. The final spatiotemporal coupling features are obtained by extracting and fusing spatiotemporal features layer by layer, fully mining the potential patterns and dependencies in spatiotemporal data.

[0141] The pre-trained LSTM network is used as the top-level structure of the spatiotemporal convolutional neural network to capture temporal dynamic patterns. The spatiotemporal convolutional features are used as the input of the LSTM network, and the gating mechanism of the LSTM network is used to adaptively selectively forget and update temporal information. The hidden state of the LSTM network contains the temporal dynamic patterns of long-term and short-term dependencies, reflecting the influence of temporal context. The hidden states of different time steps are connected through the loop structure of the LSTM network to form a temporal dynamic feature sequence.

[0142] The output layer of the LSTM network is connected to a fully connected layer to map the hidden state of the LSTM network to the defect probability value of each voxel. The weight matrix and bias term of the fully connected layer are used as learnable parameters and are optimized and adjusted through training data. The Sigmoid function is used as the activation function of the fully connected layer to map the defect probability value to the interval [0, 1], indicating the confidence that the voxel is a defect. Through threshold judgment, voxels above the threshold are marked as defects, and voxels below the threshold are marked as normal, and a defect probability sequence is obtained. The defect probability sequence corresponds one-to-one with the original voxel grid, which intuitively reflects whether each voxel has a defect and the possibility of a defect occurring.

[0143] In this embodiment, by alternately stacking temporal causal convolutional layers and spatial graph convolutional layers, the dependencies of surface morphology in the temporal and spatial dimensions can be captured simultaneously, deep fusion of spatiotemporal features can be achieved, and richer spatiotemporal coupling features can be extracted; the use of long short-term memory networks enhances the ability to capture temporal dynamic patterns, can flexibly handle long-term dependencies and short-term changes, and effectively model complex dependencies in temporal information; through the combination of convolutional neural networks and long short-term memory networks, a defect probability sequence for each voxel is generated, and defects are intuitively marked through threshold judgment, thereby achieving efficient prediction and identification of surface defects; through the introduction of activation functions and pooling layers, the nonlinear expression ability of the network and the robustness of the features are improved, and potential defects in spatiotemporal data can be detected more accurately; the generation of defect probability sequences provides an intuitive way to show whether each voxel has defects, and quantifies the possibility of defects, which is convenient for further analysis and processing.

[0144] S103. Based on the multi-physical field coupling data in the progressive synchronous stretching process, dynamic numerical simulation is performed through the adaptive finite element method to obtain the multi-physical field instability indicators, and the multi-physical field instability indicators are dynamically correlated with the defect probability sequence across scales to construct a defect evolution prediction model based on multivariate time series. The defect evolution prediction model is trained and updated in real time through the adaptive dynamic Bayesian network inference algorithm to generate a warning signal for the quality of photovoltaic cell thin films.

[0145] In this embodiment, the multi-physical field instability of the photovoltaic cell film during the progressive synchronous stretching process can be dynamically simulated and analyzed by combining the adaptive finite element method with the multi-physical field coupling data, so as to accurately predict the evolution process of the defect; the multi-physical field instability indicators are dynamically correlated with the defect probability sequence across scales, realizing a comprehensive analysis from microscopic defects to macroscopic instability indicators, thereby improving the comprehensiveness and accuracy of defect prediction; the defect evolution prediction model is trained and updated in real time using the adaptive dynamic Bayesian network inference algorithm, so that the model can continuously adapt to new data and changes, thereby improving the timeliness and reliability of the prediction; through the generated quality warning signals, potential quality problems of the photovoltaic cell film can be discovered and responded to in a timely manner, providing strong support for quality control and optimization in the production process.

[0146] In an optional embodiment, based on the multi-physical field coupling data in the progressive synchronous stretching process, dynamic numerical simulation is performed by an adaptive finite element method to obtain multi-physical field instability indicators, and the multi-physical field instability indicators are dynamically associated with the defect probability sequence across scales to construct a defect evolution prediction model based on a multivariate time series. The defect evolution prediction model is trained and updated in real time by an adaptive dynamic Bayesian network inference algorithm to generate a photovoltaic cell film quality warning signal including:

[0147] The mechanical sensors, thermal sensors and electrical sensors installed on the progressive synchronous stretching equipment are used to collect multi-physical field parameters including mechanical field, thermal field and electrical field, and the multi-physical field parameters are synchronously sampled, quantified and preprocessed to obtain a synchronous multi-physical field coupling data sequence;

[0148] Based on the multi-physics field coupling data sequence, a multi-physics field coupling finite element model including mechanical field, thermal field and electrical field is established. The multi-physics field coupling finite element model is adaptively optimized and dynamically numerically solved using adaptive meshing technology and dynamic time step control technology to obtain the spatiotemporal evolution results of the multi-physics field during the progressive synchronous stretching process.

[0149] Analyze the spatiotemporal evolution results of multi-physical fields, extract mechanical instability indicators through the nonlinear change characteristics of stress-strain curves, extract thermal instability indicators through the local mutation characteristics of temperature fields, and extract electrical instability indicators through the abnormal fluctuation characteristics of electrical parameters, thus forming multi-physical field instability indicators;

[0150] The time-space scale transformation technology is used to map the instability indicators of multi-physical fields and the defect probability sequence to the same time-space scale, and the dynamic mapping relationship between the instability indicators of multi-physical fields and the defect probability sequence is established through the dynamic data fusion technology;

[0151] Based on the dynamic mapping relationship, a multivariate time series data set including a multi-physical field instability indication time series and a defect probability time series is determined, a defect evolution prediction model is constructed, and the defect evolution prediction model is trained by a pre-constructed long short-term memory neural network algorithm;

[0152] The input and output of the defect evolution prediction model are mapped to the nodes and edges of the adaptive dynamic Bayesian network. According to the newly collected multi-physics field coupling data and defect detection data, the adaptive dynamic Bayesian network is adjusted and updated in real time through the dynamic structure learning algorithm and parameter estimation algorithm to determine the optimal defect evolution prediction model.

[0153] The multi-physical field coupling data collected in real time is input into the optimal defect evolution prediction model to predict the future defect evolution trend, calculate the defect density and defect size distribution, determine the risk index, and judge the abnormal state and potential risks of the photovoltaic cell film quality based on the preset quality threshold and risk assessment rules, and generate a quality warning signal including warning level, warning time and warning location.

[0154] Mechanical sensors, thermal sensors and electrical sensors are installed on the progressive synchronous stretching equipment to collect multi-physical field parameters of mechanical field, thermal field and electrical field respectively; mechanical sensors collect mechanical parameters such as stress and strain, thermal sensors collect thermal parameters such as temperature and heat flow, and electrical sensors collect electrical parameters such as voltage and current; multi-physical field parameters are sampled synchronously to ensure that data from different physical fields are aligned in the time dimension to ensure data consistency and integrity; the collected multi-physical field parameters are quantified, analog signals are converted into digital signals, and data formats and data units are unified; the quantized multi-physical field parameters are preprocessed, including data cleaning, denoising, normalization and other operations to improve data quality and reliability; through data synchronization, quantization and preprocessing, a synchronized multi-physical field coupling data sequence is obtained to provide a data basis for subsequent analysis and modeling.

[0155] Based on the synchronized multi-physics coupling data sequence, a multi-physics coupling finite element model including mechanical field, thermal field and electrical field is established; the mechanical, thermal and electrical boundary conditions and initial conditions, such as displacement constraints, temperature boundaries, voltage boundaries, etc., are defined in the geometric model; the adaptive meshing technology is used to adaptively adjust the mesh size and shape according to the coupling characteristics and complexity of the multi-physics fields; the dynamic time step control technology is used to adaptively adjust the time step and the number of iterations according to the time-varying characteristics and convergence of the multi-physics fields; the multi-physics coupling finite element model is adaptively optimized through adaptive meshing and dynamic time step control to improve the computational efficiency and accuracy; the dynamic numerical solution of the multi-physics coupling finite element model is performed to obtain the spatiotemporal evolution results of the multi-physics fields during the progressive synchronous stretching process.

[0156] The spatiotemporal evolution results of multi-physical fields are analyzed to extract mechanical instability indicators, thermal instability indicators and electrical instability indicators. Among them, the mechanical instability indicators judge the mechanical instability state of the material through the nonlinear change characteristics of the stress-strain curve, such as yield point, fracture point, etc.; the thermal instability indicators judge the thermal instability state of the material through the local mutation characteristics of the temperature field, such as hot spots, thermal enthalpy changes, etc.; the electrical instability indicators judge the electrical instability state of the material through the abnormal fluctuation characteristics of electrical parameters, such as resistance mutation, capacitance drift, etc.; the multi-physical field instability indicators are constructed by comprehensively considering the mechanical, thermal and electrical instability indicators, which reflect the instability characteristics of the material under the coupling of multiple fields.

[0157] Spatiotemporal scale transformation techniques, such as wavelet transform and Fourier transform, are used to map the instability indicators of multiple physical fields and defect probability sequences to the same spatiotemporal scale. Dynamic data fusion techniques, such as Kalman filtering and Bayesian fusion, are used to establish a dynamic mapping relationship between instability indicators of multiple physical fields and defect probability sequences. The dynamic mapping relationship reflects the influence mechanism of instability indicators of multiple physical fields on defect evolution, and quantifies the correlation and causality between instability indicators and defect probabilities. Spatiotemporal scale mapping and dynamic data fusion are used to achieve effective integration of multi-source heterogeneous data, providing data support for defect evolution prediction.

[0158] Based on the dynamic mapping relationship, a multivariate time series data set including multi-physical field instability indicator time series and defect probability time series is determined; a defect evolution prediction model is constructed, the model input is the multi-physical field instability indicator time series, and the model output is the defect probability time series; the defect evolution prediction model is trained and optimized using a pre-built long short-term memory neural network algorithm; the long short-term memory neural network can effectively process time series data and capture the long-term and short-term dependencies between multi-physical field instability indicators and defect evolution; through the gating mechanism and memory units of the long short-term memory neural network, the timing features are adaptively selected and updated to improve the dynamic adaptability of defect evolution prediction; based on the training data, the weights and biases of the long short-term memory neural network are optimized to minimize the prediction error and obtain the optimized defect evolution prediction model.

[0159] The input and output of the defect evolution prediction model are mapped to the nodes and edges of the adaptive dynamic Bayesian network to construct a causal relationship diagram between multi-physical field instability indicators and defect probability; the topological structure of the Bayesian network is adaptively adjusted through the dynamic structure learning algorithm according to the newly collected multi-physical field coupling data and defect detection data; the conditional probability parameters of the Bayesian network are adaptively updated through the parameter estimation algorithm to quantify the causal strength and uncertainty between nodes; the Bayesian inference algorithm is used to calculate the optimal defect evolution prediction model, comprehensively considering the influence of multi-physical field instability indicators and the probability distribution of defect evolution; the adaptive dynamic Bayesian network can dynamically adjust the prediction model according to real-time data to improve the robustness and generalization performance of the prediction.

[0160] Input the multi-physical field coupling data collected in real time into the optimal defect evolution prediction model to predict the future defect evolution trend; calculate the quality indicators such as defect density and defect size distribution within the prediction time period through statistical analysis and numerical calculation; determine the risk level of photovoltaic cell film based on the predicted quality indicators, such as high risk, medium risk, low risk, etc.; judge the abnormal state and potential risks of photovoltaic cell film quality based on the preset quality threshold and risk assessment rules; generate quality warning signals, including warning level, warning time and warning location, etc., to provide decision support for quality control and maintenance; timely discover and solve quality problems through real-time monitoring and dynamic warning, and improve the reliability and service life of photovoltaic cell film.

[0161] In this embodiment, by installing mechanical, thermal and electrical sensors, and performing synchronous sampling and data preprocessing, the consistency and integrity of multi-physical field data are ensured, providing a high-quality data basis for subsequent analysis; a multi-physical field coupling finite element model is established, and through adaptive meshing and dynamic time step control, the spatiotemporal evolution of multi-physical fields in the progressive synchronous stretching process can be efficiently and accurately simulated, thereby improving the efficiency and accuracy of the calculation; by extracting mechanical, thermal and electrical instability indicators and conducting comprehensive analysis, the instability characteristics of the material under the action of multi-field coupling can be fully understood, thereby improving the ability to predict instability phenomena; by using spatiotemporal scale transformation and dynamic data fusion technology, a dynamic mapping relationship between multi-physical field instability indicators and defect probabilities is established, providing a deep understanding and prediction of defect evolution; a defect evolution prediction model based on multivariate time series is constructed, and combined with an adaptive dynamic Bayesian network, the prediction model can be updated and optimized in real time, and quality warning signals of photovoltaic cell thin films can be accurately generated, providing decision support for quality control and maintenance; through real-time monitoring and dynamic warning, quality problems can be discovered and solved in a timely manner, thereby improving the reliability and service life of photovoltaic cell thin films and reducing potential risks.

[0162] Figure 2FIG. 1 is a schematic diagram of the structure of a photovoltaic cell film quality detection system based on synchronous stretching according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0163] The first unit is used to obtain a surface morphology image sequence of a photovoltaic cell film based on progressive synchronous stretching, perform multi-scale analysis on the surface morphology image sequence based on an adaptive fractal dimension analysis algorithm, extract surface features, map the surface morphology features to a low-dimensional manifold space through a feature mapping method based on manifold learning, construct a surface morphology dynamic evolution manifold model, reconstruct the corresponding surface morphology based on the surface morphology dynamic evolution manifold model through a deep learning three-dimensional space reconstruction algorithm, generate time-varying surface morphology point cloud data, and divide the time-varying surface morphology point cloud data into an adaptive voxel grid through an adaptive voxelization strategy;

[0164] The second unit is used to extract multi-scale dynamic features in each adaptive voxel grid, input the multi-scale dynamic features into a pre-trained spatiotemporal convolutional neural network, extract the spatiotemporal coupling features of the surface morphology through temporal causal convolution and spatial graph convolution, and dynamically map the spatiotemporal coupling features into a defect probability sequence corresponding to the adaptive voxel grid through a time-recursive fully connected layer;

[0165] The third unit is used to perform dynamic numerical simulation based on the multi-physical field coupling data in the progressive synchronous stretching process through the adaptive finite element method, obtain the multi-physical field instability indicators, dynamically associate the multi-physical field instability indicators with the defect probability sequence across scales, and build a defect evolution prediction model based on multivariate time series. The defect evolution prediction model is trained and updated in real time through the adaptive dynamic Bayesian network inference algorithm to generate a warning signal for the quality of photovoltaic cell thin films.

[0166] According to a third aspect of the embodiments of the present invention,

[0167] An electronic device is provided, comprising:

[0168] processor;

[0169] a memory for storing processor-executable instructions;

[0170] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0171] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A photovoltaic cell film quality detection method based on synchronous stretching, characterized in that: include: Based on progressive synchronous stretching, a surface morphology image sequence of a photovoltaic cell film is obtained; based on an adaptive fractal dimension analysis algorithm, a multi-scale analysis is performed on the surface morphology image sequence to extract surface features; the surface morphology features are mapped to a low-dimensional manifold space through a feature mapping method based on manifold learning, and a surface morphology dynamic evolution manifold model is constructed; based on the surface morphology dynamic evolution manifold model, the corresponding surface morphology is reconstructed through a deep learning three-dimensional space reconstruction algorithm to generate time-varying surface morphology point cloud data; and the time-varying surface morphology point cloud data is divided into an adaptive voxel grid through an adaptive voxelization strategy; For each adaptive voxel grid, extract the multi-scale dynamic features in the adaptive voxel grid, input the multi-scale dynamic features into the pre-trained spatiotemporal convolutional neural network, extract the spatiotemporal coupling features of the surface morphology through temporal causal convolution and spatial graph convolution, and dynamically map the spatiotemporal coupling features into the defect probability sequence corresponding to the adaptive voxel grid through the time recursive fully connected layer; Based on the multi-physical field coupling data in the progressive synchronous stretching process, dynamic numerical simulation is carried out through the adaptive finite element method to obtain the multi-physical field instability indicators, which are dynamically correlated with the defect probability sequence across scales, and a defect evolution prediction model based on multivariate time series is constructed. The defect evolution prediction model is trained and updated in real time through the adaptive dynamic Bayesian network inference algorithm to generate early warning signals for photovoltaic cell film quality.

2. The method according to claim 1, characterized in that Based on progressive synchronous stretching, the surface morphology image sequence of photovoltaic cell film is obtained. Based on the adaptive fractal dimension analysis algorithm, the surface morphology image sequence is analyzed at multiple scales to extract surface features. The surface morphology features are mapped to low-dimensional manifold space through the feature mapping method based on manifold learning. The surface morphology dynamic evolution manifold model is constructed, including: Based on progressive synchronous stretching, the high-speed camera device and the stretching device are controlled to synchronously collect data through a synchronous trigger control circuit to obtain continuous image frames. Combined with the optical flow motion compensation algorithm, the continuous image frames are aligned with motion vectors to obtain the surface morphology image sequence of the photovoltaic cell film. Through non-subsampled contourlet transform, each image in the surface morphology image sequence is decomposed in a multi-scale and multi-directional manner to obtain transformation subbands, and the transformation coefficients corresponding to each of the transformation subbands are determined. The transformation coefficient energy spectrum is determined by calculating the energy of the transformation coefficient, and the fractal dimension of each transformation subband is calculated by using a box counting method to construct a multi-scale fractal dimension spectrum. Through multi-scale gradient analysis of the transformation coefficient energy spectrum, the optimal scale range is adaptively determined, and the statistical features of the multi-scale fractal dimension spectrum are extracted to generate surface morphology features. A high-dimensional data dimensionality reduction method is adopted to maintain local and global structures by adjusting the KL divergence, realize nonlinear mapping of surface morphology features to low-dimensional manifolds, and generate a low-dimensional manifold space. In the low-dimensional manifold space, the similarity between different surface morphologies is measured by geodesic distance, and a time series transition probability matrix is ​​generated. A manifold model of dynamic evolution of surface morphology is constructed to determine the dynamic evolution law of surface morphology.

3. The method according to claim 2, characterized in that The high-dimensional data dimensionality reduction method comprises: Input the surface morphology features into the high-dimensional feature space, determine the high-dimensional sample points, calculate the Euclidean distance between each high-dimensional sample point, and construct the Euclidean distance matrix; Based on the low-dimensional manifold space, randomly select position coordinates, determine low-dimensional sample points, and set the t-distribution degree of freedom parameters to determine the initial low-dimensional manifold space; In the high-dimensional feature space, based on the Euclidean distance matrix, the conditional probability between the high-dimensional sample points is calculated, the corresponding distance values ​​are converted into similarity metrics using a Gaussian kernel function, and the probability distribution of the high-dimensional sample points is constructed; In the initial low-dimensional manifold space, according to the position coordinates of the current low-dimensional sample points, the conditional probabilities between the low-dimensional sample points are calculated, and the probability distribution of the low-dimensional sample points is constructed using t-distribution as a similarity measure; Calculate the KL divergence between the probability distribution of high-dimensional sample points and the probability distribution of low-dimensional sample points, construct the optimization objective function, iteratively adjust the coordinates of the low-dimensional sample points through the gradient descent method, combine with the dynamic adjustment of the t distribution degree of freedom parameter, control the balance between the local structure and the global structure, minimize the optimization objective function, obtain the probability distribution of the low-dimensional sample points that best conforms to the probability distribution of the high-dimensional sample points, and determine the optimal low-dimensional manifold space.

4. The method according to claim 3, characterized in that Also includes: The probability distribution of low-dimensional sample points is as follows: Among them, q ij Represents the similarity probability between low-dimensional sample points, y i Indicates the coordinate position of the i-th sample point, y j represents the coordinate position of the jth sample point, ν represents the degree of freedom parameter of the t distribution, ∑ k≠l Represents the sum of the relationships between any different sample points among all sample points, y k Indicates the coordinate position of the kth sample point, y l Represents the coordinate position of the lth sample point; Dynamically adjust the t distribution degree of freedom parameter, the formula is as follows: Where ν(c) represents the degree of freedom parameter of the t distribution at the cth iteration, c represents the cth iteration, and v min Indicates the preset minimum value of the degree of freedom parameter, v max Represents the maximum value of the preset degree of freedom parameter, C represents the total number of iterations, α represents the steepness of the change, and β represents the starting point of dynamic adjustment.

5. The method according to claim 2, characterized in that: Based on the surface topography dynamic evolution manifold model, the corresponding surface topography is reconstructed through a deep learning three-dimensional space reconstruction algorithm to generate time-varying surface topography point cloud data, and the time-varying surface topography point cloud data is divided into adaptive voxel grids through an adaptive voxelization strategy, including: A 3D reconstruction model is constructed based on a conditional generative adversarial network, wherein the conditional generative adversarial network includes a generator and a discriminator. The generator takes low-dimensional manifold coordinates and timestamps as conditional inputs, and generates 3D surface morphology samples through multi-layer deconvolution and jump connections. The discriminator takes the 3D surface morphology samples and the real surface morphology as inputs, and determines the authenticity measure of the 3D surface morphology samples through multi-layer convolution and downsampling. The final 3D surface morphology is generated through adversarial loss function combined with temporal consistency loss function and adversarial iteration of the generator and the discriminator. Based on the final 3D surface morphology corresponding to each timestamp, the 3D coordinate data is obtained according to the predetermined surface morphology reference point, and the time-varying surface morphology point cloud data is generated. Based on the time-varying surface morphology point cloud data, the global three-dimensional surface morphology corresponding to the time-varying surface morphology point cloud data is recursively divided into eight sub-regions through octree space partitioning. If the granularity of the sub-region is greater than a preset coarse-grained threshold, the recursive partitioning is continued until all sub-regions reach the preset coarse-grained threshold, thereby generating coarse-grained voxels. Performing KD tree partitioning on the local three-dimensional surface morphology corresponding to the coarse-grained voxels, selecting the coordinate axis with the maximum variance as the segmentation axis, and recursively dividing the local three-dimensional surface morphology into two sub-areas until a preset fine-grained threshold is met, thereby generating an adaptive voxelization result; For each voxel in the adaptive voxelization result, the internal point cloud density and curvature change are calculated, the voxel size is adaptively adjusted, the voxel merging criterion based on entropy is combined, redundant voxels are merged, a time-varying adaptive voxel grid index is generated, and the spatial position, size and timestamp of each voxel are determined.

6. The method according to claim 1, characterized in that For each adaptive voxel grid, the multi-scale dynamic features in the adaptive voxel grid are extracted, and the multi-scale dynamic features are input into the pre-trained spatiotemporal convolutional neural network. The spatiotemporal coupling features of the surface morphology are extracted through temporal causal convolution and spatial graph convolution. The spatiotemporal coupling features are dynamically mapped to the defect probability sequence corresponding to the adaptive voxel grid through the time recursive fully connected layer, including: Performing multi-scale decomposition on the adaptive voxel grid by three-dimensional wavelet transform to obtain voxel grid representations at different scales; Based on the three-dimensional local binary pattern, the neighborhood of each voxel is divided into a plurality of first-class sub-regions according to a preset neighborhood radius and a preset number of neighborhood points. For each first-class sub-region, the gray value difference between the central voxel and the neighborhood voxels in the first-class sub-region is calculated to obtain a local difference vector in the first-class sub-region. The local difference vector is compared with a preset threshold to obtain a binary coding sequence in the first-class sub-region. The binary coding sequences of all the first-class sub-regions are spliced ​​to obtain a three-dimensional local binary pattern feature descriptor of the voxel, and a local texture feature is determined. Based on the preset three-dimensional shape context radius and angle interval, the spherical area centered on the current voxel is divided into multiple second-class sub-areas. For each second-class sub-area, the radial distance and angle of all neighboring voxels in the second-class sub-area relative to the current voxel are calculated, and the radial distance and angle are quantified to obtain the shape context distribution in the second-class sub-area. The shape context distributions of all the second-class sub-areas are spliced ​​to obtain the shape context descriptor of the voxel. Perform motion estimation on voxel grids of adjacent time steps to obtain a motion vector field of the voxel grid, calculate the gradient of the motion vector field, obtain a motion boundary image, divide the motion boundary image into multiple three-category sub-regions, calculate the gradient direction histogram for each of the three-category sub-regions, concatenate the gradient direction histograms of all the three-category sub-regions, obtain a three-dimensional motion boundary histogram of the voxel grid, and determine the temporal dynamic characteristics of the voxel grid; Through a multi-scale feature fusion strategy, local texture features, shape context descriptors and temporal dynamic features at different scales are fused to obtain multi-scale dynamic features within an adaptive voxel grid.

7. The method according to claim 6, characterized in that The multi-scale dynamic features are input into the pre-trained spatiotemporal convolutional neural network, and the spatiotemporal coupling features of the surface morphology are extracted through temporal causal convolution and spatial graph convolution. The spatiotemporal coupling features are dynamically mapped into defect probability sequences corresponding to the adaptive voxel grid through the time recursive fully connected layer, including: Inputting the multi-scale dynamic features into the spatiotemporal convolutional neural network, the spatiotemporal convolutional neural network includes a temporal causal convolution layer and a spatial graph convolution layer, the temporal causal convolution layer constructs a causal convolution kernel in the temporal dimension, performs a temporal causal convolution operation on the multi-scale dynamic features, and extracts the causal dependency in the temporal dimension; the spatial graph convolution layer constructs a graph convolution kernel in the spatial dimension, performs a spatial graph convolution operation on the multi-scale dynamic features, and extracts the local structural information in the spatial dimension; An activation function and a pooling layer are constructed between the temporal causal convolution layer and the spatial graph convolution layer, and a spatiotemporal coupling result is obtained through nonlinear transformation and pooling operation; By alternately stacking multiple temporal causal convolutional layers and multiple spatial graph convolutional layers, the spatiotemporal coupling results are extracted layer by layer to determine the spatiotemporal coupling characteristics; Based on the pre-trained long short-term memory network, the dynamic patterns and dependencies of the spatiotemporal coupling features changing over time are captured, and the spatiotemporal coupling features are mapped into defect probability values ​​of each voxel in combination with the weight matrix and bias to generate a defect probability sequence.

8. The method according to claim 1, characterized in that Based on the multi-physics field coupling data in the progressive synchronous stretching process, dynamic numerical simulation is performed through the adaptive finite element method to obtain the multi-physics field instability indicators, and the multi-physics field instability indicators are dynamically associated with the defect probability sequence across scales to build a defect evolution prediction model based on multivariate time series. The defect evolution prediction model is trained and updated in real time through the adaptive dynamic Bayesian network inference algorithm to generate photovoltaic cell film quality warning signals including: The mechanical sensors, thermal sensors and electrical sensors installed on the progressive synchronous stretching equipment are used to collect multi-physical field parameters including mechanical field, thermal field and electrical field, and the multi-physical field parameters are synchronously sampled, quantified and preprocessed to obtain a synchronous multi-physical field coupling data sequence; Based on the multi-physics field coupling data sequence, a multi-physics field coupling finite element model including mechanical field, thermal field and electrical field is established. The multi-physics field coupling finite element model is adaptively optimized and dynamically numerically solved using adaptive meshing technology and dynamic time step control technology to obtain the spatiotemporal evolution results of the multi-physics field during the progressive synchronous stretching process. Analyze the spatiotemporal evolution results of multi-physical fields, extract mechanical instability indicators through the nonlinear change characteristics of stress-strain curves, extract thermal instability indicators through the local mutation characteristics of temperature fields, and extract electrical instability indicators through the abnormal fluctuation characteristics of electrical parameters, thus forming multi-physical field instability indicators; The time-space scale transformation technology is used to map the instability indicators of multi-physical fields and the defect probability sequence to the same time-space scale, and the dynamic mapping relationship between the instability indicators of multi-physical fields and the defect probability sequence is established through the dynamic data fusion technology; Based on the dynamic mapping relationship, a multivariate time series data set including a multi-physical field instability indication time series and a defect probability time series is determined, a defect evolution prediction model is constructed, and the defect evolution prediction model is trained by a pre-constructed long short-term memory neural network algorithm; The input and output of the defect evolution prediction model are mapped to the nodes and edges of the adaptive dynamic Bayesian network. According to the newly collected multi-physics field coupling data and defect detection data, the adaptive dynamic Bayesian network is adjusted and updated in real time through the dynamic structure learning algorithm and parameter estimation algorithm to determine the optimal defect evolution prediction model. The multi-physical field coupling data collected in real time is input into the optimal defect evolution prediction model to predict the future defect evolution trend, calculate the defect density and defect size distribution, determine the risk index, and judge the abnormal state and potential risks of the photovoltaic cell film quality based on the preset quality threshold and risk assessment rules, and generate a quality warning signal including warning level, warning time and warning location.

9. A photovoltaic cell film quality detection system based on synchronous stretching, used to implement the method described in any one of claims 1 to 8, characterized in that: include: The first unit is used to obtain a surface morphology image sequence of a photovoltaic cell film based on progressive synchronous stretching, perform multi-scale analysis on the surface morphology image sequence based on an adaptive fractal dimension analysis algorithm, extract surface features, map the surface morphology features to a low-dimensional manifold space through a feature mapping method based on manifold learning, construct a surface morphology dynamic evolution manifold model, reconstruct the corresponding surface morphology based on the surface morphology dynamic evolution manifold model through a deep learning three-dimensional space reconstruction algorithm, generate time-varying surface morphology point cloud data, and divide the time-varying surface morphology point cloud data into an adaptive voxel grid through an adaptive voxelization strategy; The second unit is used to extract multi-scale dynamic features in each adaptive voxel grid, input the multi-scale dynamic features into a pre-trained spatiotemporal convolutional neural network, extract the spatiotemporal coupling features of the surface morphology through temporal causal convolution and spatial graph convolution, and dynamically map the spatiotemporal coupling features into a defect probability sequence corresponding to the adaptive voxel grid through a time-recursive fully connected layer; The third unit is used to perform dynamic numerical simulation based on the multi-physical field coupling data in the progressive synchronous stretching process through the adaptive finite element method, obtain the multi-physical field instability indicators, dynamically associate the multi-physical field instability indicators with the defect probability sequence across scales, and build a defect evolution prediction model based on multivariate time series. The defect evolution prediction model is trained and updated in real time through the adaptive dynamic Bayesian network inference algorithm to generate a warning signal for the quality of photovoltaic cell thin films.

10. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 8.

Images