Method and system for detecting performance of graphene heat-conducting thick film

Abstract

The invention discloses a method and a system for detecting the performance of a graphene heat-conducting thick film, and relates to the technical field of thick film detection. Performing a time sequence heat-conducting property test on the graphene heat-conducting thick film to obtain an atom out-of-plane displacement sequence and an atom in-plane displacement sequence corresponding to each atom in the heat-conducting thick film; carrying out area division on the graphene thick film, then carrying out fault defect probability detection on the atomic out-of-plane displacement sequence and the atomic in-plane displacement sequence in an area, and if the detected fault defect probability is greater than a preset defect threshold value, marking the area as a fault area; performing preprocessing and crystal object identification on the atomic crystal image of each fault area to obtain a crystal identification image; and determining a performance result of the heat-conducting thick film according to all crystal identification images. The time sequence heat-conducting property test is performed on the graphene heat-conducting thick film to determine the abnormal position, crystal identification is performed on the atomic crystal image of the abnormal position, the graphene heat-conducting thick film property is determined, and the graphene heat-conducting thick film property detection efficiency is improved.

Description

Technical Field

[0001] This invention belongs to the field of thick film testing technology, specifically relating to a method and system for testing the performance of graphene thermally conductive thick films. Background Technology

[0002] Graphene, a two-dimensional carbon nanomaterial with excellent thermal conductivity, far surpasses traditional metals and inorganic non-metallic materials, and has shown great application potential in fields such as heat dissipation of electronic devices, thermal management of new energy batteries, and thermal control in aerospace. With the development of industrial technology, single graphene sheets are gradually being replaced by graphene thermally conductive thick films (usually referring to graphene-based composite films with thicknesses ranging from micrometers to millimeters) due to limitations in mechanical properties and difficulties in large-scale application.

[0003] However, the performance of graphene thermally conductive thick films is significantly affected by the fabrication process, and their actual thermal conductivity often differs greatly from the theoretical value. Furthermore, localized thermal resistance can easily form within the film due to defects in the graphene sheets, leading to a decrease in overall thermal uniformity. Therefore, accurately measuring the thermal conductivity of graphene thermally conductive thick films is a core step in evaluating their quality, optimizing the fabrication process, and ensuring application reliability.

[0004] Patent CN117388312A discloses a testing device and method for graphene thermally conductive films, including an insulated chamber. A thermally conductive insulating sleeve is fixedly installed on the inner wall of the chamber, and a rotating detection mechanism is installed inside the chamber to test the thermal conductivity of the graphene thermally conductive film. A clamping mechanism is provided on the upper side of the chamber to clamp the graphene thermally conductive film. An operating window corresponding to the position of the clamping mechanism is opened on the side wall of the chamber. However, this solution is inefficient when testing the performance of large-area graphene thermally conductive thick films. Summary of the Invention

[0005] The purpose of this invention is to solve the problem of low efficiency in testing the performance of large-area graphene thermally conductive thick films, and to propose a method and system for testing the performance of graphene thermally conductive thick films.

[0006] In a first aspect of this invention, a method for detecting the thermally conductive thick film properties of graphene is first proposed, the method comprising:

[0007] A graphene thermally conductive thick film was obtained, and the time-series thermal conductivity performance of the graphene thermally conductive thick film was tested to obtain the out-of-plane displacement sequence and in-plane displacement sequence of each atom in the graphene thermally conductive thick film.

[0008] The graphene thick film is divided into sub-region sets by using preset rules;

[0009] Fault probability detection is performed based on the out-of-plane displacement sequence and in-plane displacement sequence of atoms within the target sub-region. If the detected fault probability is greater than a preset defect threshold, the target sub-region is recorded as a fault region. The target sub-region is any one of the sub-regions in the set.

[0010] Acquire atomic crystal images of all fault areas, and preprocess each atomic crystal image to obtain an initial crystal image set;

[0011] A crystal recognition image is obtained by performing crystal object recognition on the target crystal image; the target crystal image is any one of the initial crystal image sets;

[0012] The performance results of the graphene thermally conductive thick film were determined based on all crystal identification images.

[0013] Optionally, fault probability detection based on out-of-plane and in-plane displacement sequences within the target sub-region includes:

[0014] Fast Fourier transform is performed on the out-of-plane displacement sequence and the in-plane displacement sequence of atoms within the target sub-region to obtain the horizontal frequency response, vertical frequency response and vertical frequency response;

[0015] The energy scalars corresponding to the horizontal frequency response, the vertical frequency response, and the vertical angle frequency response are calculated using the energy integral formula, and the vibrational energy of each atom is obtained based on the calculated energy scalars.

[0016] The average energy of the target is obtained by calculating the vibrational energy of all atoms in the target sub-region;

[0017] The target average energy is substituted into a preset model to obtain the probability of failure.

[0018] Optionally, substituting the target average energy into a preset model to obtain the fault probability includes:

[0019] The kernel vector is obtained by calculating the kernel function value of the target average energy and all preset training samples;

[0020] The probability of a fault or defect is obtained by multiplying the kernel vector by a preset coefficient.

[0021] The acquisition of the preset coefficient specifically includes:

[0022] Obtain the failure probability corresponding to the average energy of different targets in historical data, and map the average energy of each target and the corresponding failure probability through a polynomial kernel function to obtain the target kernel function; through ridge regression regularization, minimize the target kernel function to obtain the preset coefficient.

[0023] Optionally, preprocessing the crystal image for each atom to obtain an initial crystal image set includes:

[0024] For each atomic crystal image, a guided filter with a preset-sized window is applied to the image to obtain a denoised image;

[0025] The denoised image is normalized to obtain an initial crystal image, and all initial crystal images are obtained to form an initial crystal image set.

[0026] Optionally, crystal object recognition is performed on the target crystal image to obtain a crystal recognition image, including:

[0027] The target crystal image is sequentially substituted into three residual blocks for downsampling to obtain the first residual feature, the second residual feature, and the third residual feature;

[0028] After performing a 1×1 convolution operation on the third residual feature, the first attention feature is obtained by inputting it into the position attention module;

[0029] The first attention feature is upsampled to obtain a first upsampled image, and the first upsampled image and the third residual feature are substituted into a multi-scale fusion attention block to obtain a second attention feature.

[0030] The second attention feature is upsampled to obtain a second upsampled image. The second upsampled image and the second residual feature are substituted into a multi-scale fusion attention block to obtain a third attention feature.

[0031] The third attention feature is upsampled to obtain a third upsampled image. The third upsampled image and the first residual feature are substituted into a multi-scale fusion attention block to obtain a fourth attention feature.

[0032] The fourth attention feature is deconvolved to identify all crystal objects in the target crystal image, resulting in a crystal recognition image.

[0033] In a second aspect of this invention, a system for testing the thermally conductive thick film properties of graphene is provided, comprising:

[0034] A thermal conductivity testing module is used to obtain a graphene thermally conductive thick film and to perform time-series thermal conductivity testing on the graphene thermally conductive thick film to obtain the out-of-plane displacement sequence and in-plane displacement sequence of each atom in the graphene thermally conductive thick film.

[0035] The region division module is used to divide the graphene thick film into sub-region sets according to preset rules;

[0036] The fault probability determination module is used to detect fault probability based on the out-of-plane displacement sequence and in-plane displacement sequence of atoms in the target sub-region. If the detected fault probability is greater than a preset fault threshold, the target sub-region is recorded as a fault region. The target sub-region is any one of the sub-regions in the set.

[0037] The preprocessing module is used to acquire atomic crystal images of all fault areas and preprocess each atomic crystal image to obtain an initial crystal image set.

[0038] A crystal recognition module is used to perform crystal object recognition on a target crystal image to obtain a crystal recognition image; the target crystal image is any one of the initial crystal images in the set.

[0039] The performance result determination module is used to determine the performance result of the graphene thermally conductive thick film based on all crystal recognition images.

[0040] Optionally, the fault defect probability determination module includes:

[0041] The frequency response determination module is used to perform fast Fourier transform on the out-of-plane displacement sequence and the in-plane displacement sequence of atoms in the target sub-region to obtain the horizontal frequency response, vertical frequency response and vertical frequency response.

[0042] The thermal conductivity testing module is used to calculate the energy scalars corresponding to the horizontal frequency response, the vertical frequency response, and the vertical angle frequency response using the energy integration formula, and to obtain the vibrational energy of each atom based on the calculated energy scalars.

[0043] The target average energy determination module is used to calculate the vibrational energy of all atoms in the target sub-region to obtain the target average energy;

[0044] The fault probability calculation module is used to substitute the target average energy into a preset model to obtain the fault probability.

[0045] Optionally, the fault probability calculation module includes:

[0046] The kernel vector determination module is used to calculate the kernel vector by comparing the target average energy with the kernel function values ​​of all preset training samples.

[0047] The fault / defect probability generation module is used to obtain the fault / defect probability by multiplying the kernel vector by a preset coefficient.

[0048] The acquisition of the preset coefficient specifically includes:

[0049] Obtain the failure probability corresponding to the average energy of different targets in historical data, and map the average energy of each target and the corresponding failure probability through a polynomial kernel function to obtain the target kernel function; through ridge regression regularization, minimize the target kernel function to obtain the preset coefficient.

[0050] Optionally, the preprocessing module includes:

[0051] The denoising module is used to apply guided filtering with a preset-size window to each atomic crystal image to obtain a denoised image.

[0052] The normalization processing module is used to normalize the denoised image to obtain an initial crystal image, and to obtain an initial crystal image set by acquiring all the initial crystal images.

[0053] Optionally, the crystal recognition module includes:

[0054] The residual feature extraction module is used to sequentially substitute the target crystal image into three residual blocks for downsampling to obtain the first residual feature, the second residual feature, and the third residual feature;

[0055] The first attention feature determination module is used to perform a 1×1 convolution operation on the third residual feature and then input it into the position attention module to obtain the first attention feature;

[0056] The second attention feature determination module is used to upsample the first attention feature to obtain a first upsampled image, and substitute the first upsampled image and the third residual feature into the multi-scale fusion attention module to obtain the second attention feature.

[0057] The third attention feature determination module is used to upsample the second attention feature to obtain a second upsampled image, and substitute the second upsampled image and the second residual feature into the multi-scale fusion attention module to obtain the third attention feature.

[0058] The fourth attention feature determination module is used to upsample the third attention feature to obtain a third upsampled image, and substitute the third upsampled image and the first residual feature into the multi-scale fusion attention module to obtain the fourth attention feature;

[0059] The deconvolution module is used to perform a deconvolution operation on the fourth attention feature to identify all crystal objects in the target crystal image and obtain a crystal recognition image.

[0060] The beneficial effects of this invention are:

[0061] This invention proposes a method for performance testing of graphene thermally conductive thick films. The method involves acquiring a graphene thermally conductive thick film, performing time-series thermal conductivity tests to obtain the out-of-plane and in-plane displacement sequences of each atom in the film, and dividing the graphene thick film into sub-region sets according to preset rules. Fault probability detection is performed based on the out-of-plane and in-plane displacement sequences within the target sub-regions. If the detected fault probability is greater than a preset defect threshold, the target sub-region is designated as a fault region. The target sub-region can be any one of the sub-region sets. Atomic crystal images of all fault regions are acquired, and each atomic crystal image is preprocessed to obtain an initial crystal image set. Crystal object recognition is performed on the target crystal images to obtain crystal recognition images. The target crystal image can be any one of the initial crystal image sets. The performance results of the graphene thermally conductive thick film are determined based on all crystal recognition images. By conducting time-series thermal conductivity tests on graphene thermally conductive thick films to identify abnormal locations, and then performing crystal identification on atomic crystal images of the abnormal locations, a comprehensive judgment is made on the entire graphene thermally conductive thick film to determine its performance. This approach avoids large-area performance testing while ensuring the accuracy of the test and improving the efficiency of graphene thermally conductive thick film performance testing. Attached Figure Description

[0062] The invention will now be further described with reference to the accompanying drawings.

[0063] Figure 1 A flowchart of a method for testing the thermally conductive thick film performance of graphene provided in an embodiment of the present invention;

[0064] Figure 2 This is a framework diagram of a system for testing the performance of thermally conductive thick films of graphene, provided as an embodiment of the present invention. Detailed Implementation

[0065] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0066] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] This invention provides a method for testing the thermally conductive thick film properties of graphene. See also... Figure 1 , Figure 1 This is a flowchart illustrating a method for testing the thermally conductive thick film performance of graphene, provided as an embodiment of the present invention. The method includes the following steps:

[0068] S101, Obtain a graphene thermally conductive thick film, and perform time-series thermal conductivity performance testing on the graphene thermally conductive thick film to obtain the out-of-plane displacement sequence and in-plane displacement sequence of each atom in the graphene thermally conductive thick film.

[0069] S102, the graphene thick film is divided into regions according to preset rules to obtain a set of sub-regions;

[0070] S103, perform fault probability detection based on the out-of-plane displacement sequence and in-plane displacement sequence of the atomic sub-region. If the detected fault probability is greater than the preset fault threshold, the target sub-region is recorded as a fault region.

[0071] S104: Collect atomic crystal images of all fault areas, and preprocess each atomic crystal image to obtain an initial crystal image set;

[0072] S105, Perform crystal object recognition on the target crystal image to obtain a crystal recognition image;

[0073] S106, determine the performance results of the graphene thermally conductive thick film based on all crystal recognition images;

[0074] The target sub-region is any one of the sub-regions in the set; the target crystal image is any one of the initial crystal images in the set.

[0075] The method for testing the performance of graphene thermally conductive thick films provided by this invention involves determining abnormal locations by performing time-series thermal conductivity tests on the graphene thermally conductive thick film, then identifying crystals in the atomic crystal images of the abnormal locations, and finally making a comprehensive judgment on the entire graphene thermally conductive thick film to determine its performance. This method avoids large-area performance testing while ensuring the accuracy of the test and improving the efficiency of graphene thermally conductive thick film performance testing.

[0076] In one implementation, time-series thermal conductivity tests are performed on graphene thermally conductive thick films to obtain the out-of-plane and in-plane displacement sequences of each atom, enabling the capture of microscopic dynamic changes in the material at the atomic level. This microscale monitoring allows even minute atomic displacement anomalies to be detected, providing extremely detailed basic data for subsequent defect detection.

[0077] In one implementation, after dividing the film into sub-regions according to preset rules, a fault probability detection is performed on each sub-region to accurately locate problematic areas. When the fault probability exceeds a preset threshold, the sub-region is marked as a faulty area. This avoids making a general judgment on the entire thick film and clarifies the specific location and extent of the defect. The performance of the graphene thermally conductive thick film can be determined by analyzing the crystal state at the location of the defect.

[0078] In one implementation, if the detected fault probability is less than or equal to a preset defect threshold, it indicates that the area is normal and no operation is performed.

[0079] In one implementation, the preset defect threshold is determined by technicians and is usually 50%. The smaller the threshold, the more accurate the detection, but the more areas need to be detected. The preset rule is to divide the graphene thick film into regions by a preset square size, which is determined by technicians.

[0080] In one implementation, atomic crystal images of all fault regions are acquired and preprocessed to obtain an initial crystal image set, which can remove irrelevant information such as noise and interference from the images, ensuring the reliability of subsequent analysis data.

[0081] In one implementation, crystal object recognition is performed on the target crystal image to obtain a crystal recognition image. The performance result of the graphene thermally conductive thick film is then determined by combining all the crystal recognition images. This is a comprehensive judgment based on a large amount of specific and accurate image information, rather than relying on a single or small amount of data, which greatly improves the accuracy of performance evaluation.

[0082] In one implementation, after identifying defective crystals in all crystal recognition images, the proportion of defective crystals in the current sub-region and the proportion of defective crystals in the entire graphene thermally conductive thick film can be obtained. The proportions in the sub-regions are set with corresponding thresholds A1 and A2 (A1 is less than A2, and the values ​​are both between 0 and 1, with specific values ​​determined by technical personnel). If the proportion in a sub-region is less than or equal to A1, the current sub-region is recorded as a good sub-region (can be used normally, for scenarios with high requirements for thermal conductivity). If the proportion in a sub-region is greater than A1 but less than A2, the current sub-region is recorded as a general sub-region (quality is lower than that of the good sub-region, used for scenarios with less stringent requirements for thermal conductivity). If the proportion in a sub-region is greater than A2, the current sub-region is recorded as a poor-quality sub-region (cannot be used normally).

[0083] In one embodiment, fault probability detection based on out-of-plane displacement sequences and in-plane displacement sequences within the target sub-region includes:

[0084] Fast Fourier transform is performed on the out-of-plane and in-plane displacement sequences of atoms within the target sub-region to obtain the horizontal, vertical, and longitudinal frequency responses.

[0085] The energy scalars corresponding to the horizontal, vertical, and triangular frequency responses are calculated using the energy integral formula, and the vibrational energy of each atom is obtained based on the calculated energy scalars.

[0086] The average energy of the target is obtained by calculating the vibrational energy of all atoms in the target sub-region;

[0087] Substitute the target average energy into the preset model to obtain the probability of failure.

[0088] In one implementation, the out-of-plane and in-plane displacement sequences of atoms reflect the dynamic vibration state of atoms during heat conduction. By using FFT to convert the displacement sequences into frequency responses in the horizontal, vertical, and perpendicular directions, the complex vibrations in the time domain can be decomposed into the superposition of different frequency components, clearly presenting the energy distribution characteristics of atomic vibrations. The in-plane displacement of atoms is the sliding of points on the thin film left and right or back and forth within the plane of the thin film (parallel to the film surface); the out-of-plane displacement of atoms is the upward bulging or downward depression of points on the thin film (perpendicular to the film surface).

[0089] In one implementation, the energy scalar corresponding to the frequency response in each direction is calculated by the energy integral formula, and then the vibrational energy of a single atom is obtained by summing them up. This realizes the quantification of the intensity of atomic vibration. The energy analysis in the horizontal, vertical and vertical directions can comprehensively cover the vibrational characteristics within and between graphene layers, avoiding the one-sidedness of single-dimensional analysis.

[0090] In one implementation, the thermal conductivity of graphene primarily depends on the transmission of phonons (energy quanta of lattice vibrations), and the atomic vibration energy directly reflects the activity and transmission efficiency of phonons. Defects scatter phonons, leading to abnormal local vibrational energy distribution. By setting a defect threshold to determine the fault area, the stringency of the threshold can be adjusted according to the actual application scenario, enabling customization of the detection standard and improving the applicability of the process.

[0091] In one implementation, defects can cause local atomic vibration modes to deviate from the normal state, but they may appear as weak fluctuations in the time-domain displacement sequence, which are difficult to identify through direct analysis. FFT can transform these tiny abnormal vibrations into energy anomalies of specific frequencies through frequency domain transformation, significantly improving the identifiability of defect signals.

[0092] In one implementation, the energy integral formula is used. The energy scalar of the vertical frequency response is obtained; z(f) is the vertical frequency response; the energy integral formula is used to obtain the energy scalar of the vertical frequency response. The energy scalar of the horizontal frequency response is obtained; x(f) is the horizontal frequency response; the energy integral formula is used to obtain the energy scalar of the horizontal frequency response. The energy scalar of the vertical frequency response is obtained; y(f) is the vertical frequency response; the out-of-plane and in-plane energy features are fused into a multidimensional feature vector [S]. x (f),S y (f),S z (f)].

[0093] In one implementation, the average vibrational energy (target average energy) of all atoms in the target sub-region is calculated and substituted into a preset model to obtain the fault probability. The defect risk is judged by statistically analyzing the degree to which the overall vibrational energy of the region deviates from the normal range. Based on the statistical analysis of the region average, the interference of random vibration of individual atoms can be reduced, misjudgment can be reduced (such as misjudging normal thermal vibration as a defect), and the stability of detection can be improved.

[0094] In one embodiment, substituting the target average energy into a preset model to obtain the fault probability includes:

[0095] The kernel vector is obtained by calculating the target average energy and the kernel function values ​​of all preset training samples;

[0096] The probability of a fault or defect is obtained by multiplying the kernel vector by a preset coefficient.

[0097] The acquisition of preset coefficients specifically includes:

[0098] Obtain the failure and defect probabilities corresponding to the average energy of different targets in historical data, and map the average energy of each target and the corresponding failure and defect probabilities using a polynomial kernel function to obtain the target kernel function; through ridge regression regularization, minimize the target kernel function to obtain the preset coefficients.

[0099] In one implementation, the relationship between the target average energy (input feature) and the fault probability (output label) is often nonlinear. The polynomial kernel function can map the original linear feature space to a higher-dimensional feature space, transforming the originally nonlinear relationship into a linearly separable relationship in the high-dimensional space, thereby capturing the complex relationship between the two more accurately.

[0100] In one implementation, when solving for the preset coefficients using historical data, ridge regression effectively limits the absolute value of the coefficients by adding an L2 regularization term to the loss function, thus avoiding overfitting the model to noise in the historical data. The kernel vector is obtained by calculating the target average energy and the kernel function value of the preset training samples, and then multiplied by the preset coefficients to obtain the fault probability. The kernel trick avoids complex calculations in high-dimensional space (only the kernel function needs to be calculated in the original space), which greatly reduces the computational complexity while ensuring the nonlinear expressive power of the model, thus meeting the high efficiency requirements of batch detection of sub-regions.

[0101] In one implementation, the failure / defect probabilities corresponding to different target average energies in historical data are obtained, and a target kernel function is obtained by mapping each target average energy and its corresponding failure / defect probability using a polynomial kernel function. The target kernel function is then minimized through ridge regression regularization to obtain preset coefficients. Specifically, the target average energy corresponding to sub-regions under different states in historical data is denoted as x = [x1, x2, ..., xn]T (where n is the number of samples and xi is the average energy of the i-th sample); the failure / defect probability of each target average energy is denoted as y = [y1, y2, ..., yn]. T (yi∈[0,1], true value known); for any two samples xi (target average energy in historical data) and xj (target average energy in historical data), the kernel function is K(xi,xj)=(xi·xj+c). p Where C and P are constants (the specific values ​​are determined by technical personnel), the kernel matrix K∈R is obtained based on all kernel functions. n×n Then, the polynomial kernel function maps the average energy of each target to the corresponding fault probability. Where ||y-Kα|| 2 The prediction error (in kernel space, the squared difference between the model's predicted value Kα and the true label y) is... For ridge regression, α is the regularization term (λ≥0 is the regularization parameter, controlling the "size" of the coefficient α; the larger λ is, the smoother the coefficient, avoiding overfitting). T This is the transpose of the control coefficient α, where α is a preset coefficient to be solved.

[0102] In one embodiment, preprocessing each atomic crystal image to obtain an initial crystal image set includes:

[0103] For each atomic crystal image, a guided filter with a preset-sized window is applied to the image to obtain a denoised image;

[0104] The denoised image is normalized to obtain the initial crystal image, and all the initial crystal images are obtained to form the initial crystal image set.

[0105] In one implementation, guided filtering is an edge-preserving denoising algorithm. Its core is to use the structural information of the image itself as the filtering basis. While removing noise, it can better preserve key structures such as edges and textures in the image. Guided filtering with a preset window size can customize window parameters according to the microscopic scale of the crystal image. It can effectively smooth out high-frequency noise without blurring the boundaries and regular structures of atomic arrangements, providing a clear and reliable image foundation for subsequent crystal object recognition. Compared with traditional methods such as mean filtering and Gaussian filtering, which may cause edge blurring, guided filtering can achieve a better balance between denoising and edge preservation, ensuring that defects are not misjudged due to loss of details in the subsequent recognition process. Normalization processing (mapping gray values ​​to the [0,1] interval).

[0106] In one embodiment, obtaining a crystal recognition image by performing crystal object recognition on a target crystal image includes:

[0107] The target crystal image is sequentially substituted into three residual blocks for downsampling to obtain the first residual feature, the second residual feature, and the third residual feature;

[0108] The first attention feature is obtained by performing a 1×1 convolution operation on the third residual feature and then inputting it into the position attention module.

[0109] The first attention feature is upsampled to obtain the first upsampled image. The first upsampled image and the third residual feature are substituted into the multi-scale fusion attention block to obtain the second attention feature.

[0110] The second attention feature is upsampled to obtain the second upsampled image. The second upsampled image and the second residual feature are then substituted into the multi-scale fused attention block to obtain the third attention feature.

[0111] The third attention feature is upsampled to obtain the third upsampled image. The third upsampled image and the first residual feature are substituted into the multi-scale fusion attention block to obtain the fourth attention feature.

[0112] By performing a deconvolution operation on the fourth attention feature, all crystal objects in the target crystal image are identified, resulting in a crystal recognition image.

[0113] In one implementation, the three residual blocks are connected by skip connections to effectively alleviate the gradient vanishing and exploding problems in deep network training. This enables the stable extraction of multi-dimensional features from crystal images, from shallow atomic edges and local brightness changes to deep lattice periodic structures and abnormal patterns in defect regions. The progressive feature extraction adapts to the hierarchical information of crystal images from microscopic atoms (small scale) to regional structures (large scale).

[0114] In one implementation, the image size is reduced in each downsampling operation, so that the feature map contains a wider range of spatial context information. After the third residual feature is reduced in dimensionality by 1×1 convolution, the spatial dependency between features is calculated by the position attention module, which can automatically focus on the region that is more critical to defect identification, suppress interference from irrelevant background or normal lattice, and improve the sensitivity to small defects.

[0115] In one implementation, during the upsampling process, high-resolution features (first and second residual features, containing fine details of atomic positions) are combined with upsampled low-resolution features (first and second upsampled images, containing global structure) through a fusion attention block. This approach can utilize both the semantic information of high-level features and the spatial details of low-level features. When identifying misalignments, it is necessary to know both the macroscopic direction of the dislocation line (high-level features) and the precise offset position of individual atoms (low-level features). After fusion, the positioning accuracy can be significantly improved.

[0116] In one implementation, the feature map size is gradually restored to near the original image by multiple upsampling steps from the third residual feature to the third upsampled image. This ensures that when fusing multi-scale features, high-level semantic information can be mapped back to the spatial coordinates of the original image, avoiding feature misalignment caused by scale differences and marking defects at incorrect atomic positions.

[0117] In one implementation, a crystal recognition image of the same size as the original image is generated through deconvolution, which can accurately mark the position and boundary of each crystal object (normal atom, missing atom, misaligned atom).

[0118] Based on the same inventive concept, this invention also provides a system for testing the performance of thermally conductive thick films of graphene. See also Figure 2 , Figure 2 A framework diagram of a system for testing the thermally conductive thick film performance of graphene, provided in an embodiment of the present invention, includes:

[0119] The thermal conductivity testing module is used to obtain a graphene thermally conductive thick film and to perform time-series thermal conductivity testing on the graphene thermally conductive thick film to obtain the out-of-plane displacement sequence and in-plane displacement sequence of each atom in the graphene thermally conductive thick film.

[0120] The region division module is used to divide the graphene thick film into sub-region sets according to preset rules;

[0121] The fault probability determination module is used to detect fault probability based on the out-of-plane displacement sequence and in-plane displacement sequence of atoms in the target sub-region. If the detected fault probability is greater than the preset fault threshold, the target sub-region is recorded as a fault region. The target sub-region is any one of the sub-regions in the set.

[0122] The preprocessing module is used to acquire atomic crystal images of all fault areas and preprocess each atomic crystal image to obtain an initial crystal image set.

[0123] The crystal recognition module is used to perform crystal object recognition on the target crystal image to obtain a crystal recognition image; the target crystal image is any one of the initial crystal images in the set.

[0124] The performance result determination module is used to determine the performance results of the graphene thermally conductive thick film based on all crystal recognition images.

[0125] The system for testing the performance of graphene thermally conductive thick films provided by this invention determines abnormal locations by performing time-series thermal conductivity tests on the graphene thermally conductive thick film, then identifies crystals in the atomic crystal images of the abnormal locations, and finally makes a comprehensive judgment on the entire graphene thermally conductive thick film to determine its performance. This avoids large-area performance testing while ensuring the accuracy of the test and improving the efficiency of graphene thermally conductive thick film performance testing.

[0126] In one embodiment, the fault probability determination module includes:

[0127] The frequency response determination module is used to perform fast Fourier transform on the out-of-plane displacement sequence and in-plane displacement sequence of atoms in the target sub-region to obtain the horizontal frequency response, vertical frequency response and vertical frequency response.

[0128] The thermal conductivity testing module is used to calculate the energy scalars corresponding to the horizontal, vertical, and triangular frequency responses using the energy integration formula, and to obtain the vibrational energy of each atom based on the calculated energy scalars.

[0129] The target average energy determination module is used to calculate the vibrational energy of all atoms in the target sub-region to obtain the target average energy;

[0130] The fault probability calculation module is used to substitute the target average energy into a preset model to obtain the fault probability.

[0131] In one embodiment, the fault probability calculation module includes:

[0132] The kernel vector determination module is used to calculate the kernel vector by combining the target average energy with the kernel function value of all preset training samples.

[0133] The fault / defect probability generation module is used to obtain the fault / defect probability by multiplying the kernel vector by a preset coefficient.

[0134] The acquisition of preset coefficients specifically includes:

[0135] Obtain the failure and defect probabilities corresponding to the average energy of different targets in historical data, and map the average energy of each target and the corresponding failure and defect probabilities using a polynomial kernel function to obtain the target kernel function; through ridge regression regularization, minimize the target kernel function to obtain the preset coefficients.

[0136] In one embodiment, the preprocessing module includes:

[0137] The denoising module is used to apply guided filtering with a preset-size window to each atomic crystal image to obtain a denoised image.

[0138] The normalization processing module is used to normalize the denoised image to obtain the initial crystal image, and to obtain the initial crystal image set by acquiring all the initial crystal images.

[0139] In one embodiment, the crystal recognition module includes:

[0140] The residual feature extraction module is used to sequentially substitute the target crystal image into three residual blocks for downsampling to obtain the first residual feature, the second residual feature, and the third residual feature;

[0141] The first attention feature determination module is used to perform a 1×1 convolution operation on the third residual feature and then input it into the position attention module to obtain the first attention feature;

[0142] The second attention feature determination module is used to upsample the first attention feature to obtain the first upsampled image, and substitute the first upsampled image and the third residual feature into the multi-scale fusion attention module to obtain the second attention feature.

[0143] The third attention feature determination module is used to upsample the second attention feature to obtain the second upsampled image, and then substitute the second upsampled image and the second residual feature into the multi-scale fusion attention module to obtain the third attention feature.

[0144] The fourth attention feature determination module is used to upsample the third attention feature to obtain the third upsampled image, and substitute the third upsampled image and the first residual feature into the multi-scale fusion attention module to obtain the fourth attention feature;

[0145] The deconvolution module is used to perform deconvolution operations on the fourth attention features to identify all crystal objects in the target crystal image and obtain a crystal recognition image.

[0146] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for testing the thermally conductive thick film properties of graphene, characterized in that, The method includes: A graphene thermally conductive thick film was obtained, and the time-series thermal conductivity performance of the graphene thermally conductive thick film was tested to obtain the out-of-plane displacement sequence and in-plane displacement sequence of each atom in the graphene thermally conductive thick film. The graphene thick film is divided into sub-region sets by using preset rules; Fault probability detection is performed based on the out-of-plane displacement sequence and in-plane displacement sequence of atoms within the target sub-region. If the detected fault probability is greater than a preset defect threshold, the target sub-region is recorded as a fault region. The target sub-region is any one of the sub-regions in the set. Acquire atomic crystal images of all fault areas, and preprocess each atomic crystal image to obtain an initial crystal image set; A crystal recognition image is obtained by performing crystal object recognition on the target crystal image; the target crystal image is any one of the initial crystal image sets; The performance results of the graphene thermally conductive thick film were determined based on all crystal identification images.

2. The method for testing the thermal conductivity of graphene thick films according to claim 1, characterized in that, Fault and defect probability detection based on out-of-plane and in-plane displacement sequences within the target sub-region includes: Fast Fourier transform is performed on the out-of-plane displacement sequence and the in-plane displacement sequence of atoms within the target sub-region to obtain the horizontal frequency response, vertical frequency response and vertical frequency response; The energy scalars corresponding to the horizontal frequency response, the vertical frequency response, and the vertical angle frequency response are calculated using the energy integral formula, and the vibrational energy of each atom is obtained based on the calculated energy scalars. The average energy of the target is obtained by calculating the vibrational energy of all atoms in the target sub-region; The target average energy is substituted into a preset model to obtain the probability of failure.

3. The method for testing the thermal conductivity of graphene thick films according to claim 2, characterized in that, Substituting the target average energy into a preset model yields the fault probability, which includes: The kernel vector is obtained by calculating the kernel function value of the target average energy and all preset training samples; The probability of a fault or defect is obtained by multiplying the kernel vector by a preset coefficient. The acquisition of the preset coefficient specifically includes: Obtain the failure probability corresponding to the average energy of different targets in historical data, and map the average energy of each target and the corresponding failure probability through a polynomial kernel function to obtain the target kernel function; through ridge regression regularization, minimize the target kernel function to obtain the preset coefficient.

4. The method for testing the thermal conductivity of graphene thick films according to claim 1, characterized in that, Preprocessing the crystal image of each atom yields an initial crystal image set including: For each atomic crystal image, a guided filter with a preset-sized window is applied to the image to obtain a denoised image; The denoised image is normalized to obtain an initial crystal image, and all initial crystal images are obtained to form an initial crystal image set.

5. The method for testing the thermal conductivity of graphene thick films according to claim 1, characterized in that, Crystal object recognition obtained by performing crystal object recognition on the target crystal image includes: The target crystal image is sequentially substituted into three residual blocks for downsampling to obtain the first residual feature, the second residual feature, and the third residual feature; After performing a 1×1 convolution operation on the third residual feature, the first attention feature is obtained by inputting it into the position attention module; The first attention feature is upsampled to obtain a first upsampled image, and the first upsampled image and the third residual feature are substituted into a multi-scale fusion attention block to obtain a second attention feature. The second attention feature is upsampled to obtain a second upsampled image. The second upsampled image and the second residual feature are substituted into a multi-scale fusion attention block to obtain a third attention feature. The third attention feature is upsampled to obtain a third upsampled image. The third upsampled image and the first residual feature are substituted into a multi-scale fusion attention block to obtain a fourth attention feature. The fourth attention feature is deconvolved to identify all crystal objects in the target crystal image, resulting in a crystal recognition image.

6. A system for testing the performance of thermally conductive thick films of graphene, characterized in that, The system includes: A thermal conductivity testing module is used to obtain a graphene thermally conductive thick film and to perform time-series thermal conductivity testing on the graphene thermally conductive thick film to obtain the out-of-plane displacement sequence and in-plane displacement sequence of each atom in the graphene thermally conductive thick film. The region division module is used to divide the graphene thick film into sub-region sets according to preset rules; The fault probability determination module is used to detect fault probability based on the out-of-plane displacement sequence and in-plane displacement sequence of atoms in the target sub-region. If the detected fault probability is greater than a preset fault threshold, the target sub-region is recorded as a fault region. The target sub-region is any one of the sub-regions in the set. The preprocessing module is used to acquire atomic crystal images of all fault areas and preprocess each atomic crystal image to obtain an initial crystal image set. A crystal recognition module is used to perform crystal object recognition on a target crystal image to obtain a crystal recognition image; the target crystal image is any one of the initial crystal images in the set. The performance result determination module is used to determine the performance result of the graphene thermally conductive thick film based on all crystal recognition images.

7. The system for testing the thermal conductivity of graphene thick films according to claim 6, characterized in that, The fault / defect probability determination module includes: The frequency response determination module is used to perform fast Fourier transform on the out-of-plane displacement sequence and the in-plane displacement sequence of atoms in the target sub-region to obtain the horizontal frequency response, vertical frequency response and vertical frequency response. The thermal conductivity testing module is used to calculate the energy scalars corresponding to the horizontal frequency response, the vertical frequency response, and the vertical angle frequency response using the energy integration formula, and to obtain the vibrational energy of each atom based on the calculated energy scalars. The target average energy determination module is used to calculate the vibrational energy of all atoms in the target sub-region to obtain the target average energy; The fault probability calculation module is used to substitute the target average energy into a preset model to obtain the fault probability.

8. The system for testing the thermally conductive thick film properties of graphene according to claim 7, characterized in that, The fault / defect probability calculation module includes: The kernel vector determination module is used to calculate the kernel vector by comparing the target average energy with the kernel function values ​​of all preset training samples. The fault / defect probability generation module is used to obtain the fault / defect probability by multiplying the kernel vector by a preset coefficient. The acquisition of the preset coefficient specifically includes: Obtain the failure probability corresponding to the average energy of different targets in historical data, and map the average energy of each target and the corresponding failure probability through a polynomial kernel function to obtain the target kernel function; through ridge regression regularization, minimize the target kernel function to obtain the preset coefficient.

9. A system for testing the thermal conductivity of graphene thick films according to claim 6, characterized in that, The preprocessing module includes: The denoising module is used to apply guided filtering with a preset-size window to each atomic crystal image to obtain a denoised image. The normalization processing module is used to normalize the denoised image to obtain an initial crystal image, and to obtain an initial crystal image set by acquiring all the initial crystal images.

10. A system for testing the thermal conductivity of graphene thick films according to claim 6, characterized in that, The crystal recognition module includes: The residual feature extraction module is used to sequentially substitute the target crystal image into three residual blocks for downsampling to obtain the first residual feature, the second residual feature, and the third residual feature; The first attention feature determination module is used to perform a 1×1 convolution operation on the third residual feature and then input it into the position attention module to obtain the first attention feature; The second attention feature determination module is used to upsample the first attention feature to obtain a first upsampled image, and substitute the first upsampled image and the third residual feature into the multi-scale fusion attention module to obtain the second attention feature. The third attention feature determination module is used to upsample the second attention feature to obtain a second upsampled image, and substitute the second upsampled image and the second residual feature into the multi-scale fusion attention module to obtain the third attention feature. The fourth attention feature determination module is used to upsample the third attention feature to obtain a third upsampled image, and substitute the third upsampled image and the first residual feature into the multi-scale fusion attention module to obtain the fourth attention feature; The deconvolution module is used to perform a deconvolution operation on the fourth attention feature to identify all crystal objects in the target crystal image and obtain a crystal recognition image.

Images