Method for predicting pore size distribution of composite fiber membrane based on deep learning

Abstract

The invention provides a composite fiber membrane aperture distribution prediction method based on deep learning, and relates to the technical field of distribution prediction.The method comprises the steps that multiple kinds of image data are collected and coded respectively, and a composite coding feature vector is generated through an attention mechanism; based on the composite coding feature vector, performing region segmentation on a fiber cross-linked region and a non-cross-linked region, and extracting and quantifying pore size distribution dominant features; performing associated coding by combining the pore size distribution dominant characteristics and spinning process parameters; and based on the associated codes, optimizing the depth prediction network, and adjusting the parameter weights of the associated codes by adopting an error correction mechanism. The pore size distribution of the composite fiber membrane can be accurately predicted, and powerful support is provided for material design optimization and production process control.

Description

Technical Field

[0001] The present invention relates to the field of distribution prediction technology, and more specifically, to a method for predicting the pore size distribution of a composite fiber membrane based on deep learning. Background Art

[0002] In the field of textile materials and functional membrane materials, composite fiber membranes have been widely used in water treatment, medical filtration, gas separation and other fields due to their excellent mechanical properties, filtration characteristics and biocompatibility. The pore size distribution of fiber membranes is a key parameter that determines their filtration efficiency, air permeability and structural stability. Therefore, accurate prediction of pore size distribution is of great significance for optimizing spinning processes and improving product quality. Traditional pore size distribution measurement methods mainly rely on experimental detection, such as bubble method, mercury intrusion method and scanning electron microscope (SEM) imaging analysis. Although these methods have high accuracy, the testing process is complex, time-consuming and costly, and it is difficult to meet the needs of industrial production for real-time monitoring and dynamic optimization. In recent years, machine learning and computer vision technologies have developed rapidly in the field of material characterization, especially the application of deep learning methods, which provides an efficient and intelligent solution for the prediction of fiber membrane pore size distribution.

[0003] At present, the pore size distribution prediction based on deep learning mainly adopts the method of combining image feature extraction and regression modeling, but it has the following limitations. For example, the traditional image feature extraction method is difficult to effectively capture the complexity of the fiber membrane microstructure, especially in terms of the pore size gradient difference between the cross-linked area and the non-cross-linked area, which easily leads to a decrease in prediction accuracy; most of the existing studies rely on static image features, and ignore the dynamic influence of spinning process parameters on the pore size distribution, resulting in insufficient model generalization ability and difficulty in adapting to pore size changes under different process conditions; in addition, the deep prediction model often lacks an effective correction mechanism during the training process, and fails to make full use of actual measurement data to optimize the model parameters, which may lead to a large deviation between the prediction results and the true value.

[0004] Therefore, how to improve the accuracy and adaptability of pore size distribution prediction is an urgent problem to be solved. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method for predicting the pore size distribution of composite fiber membranes based on deep learning.

[0006] According to one aspect of the present invention, a method for predicting the pore size distribution of composite fiber membranes based on deep learning is provided, which includes: collecting a variety of image data and encoding them separately, and generating a composite coding feature vector through an attention mechanism; based on the composite coding feature vector, performing regional segmentation on the cross-linked area and the non-cross-linked area of ​​the fiber, extracting and quantifying the dominant characteristics of the pore size distribution; combining the dominant characteristics of the pore size distribution and the spinning process parameters to perform associative coding; based on the associative coding, optimizing the deep prediction network, and using an error correction mechanism to adjust the parameter weights of the associative coding.

[0007] Furthermore, the multiple image data include scanning electron microscope images and atomic force microscope morphology data; the separate encoding includes: for scanning electron microscope images, hierarchical encoding according to grayscale pixel distribution; for atomic force microscope morphology data, gradient encoding is constructed according to the height gradient change in the Z-axis direction.

[0008] Furthermore, the generation of the composite coding feature vector includes: preliminary quantization of grayscale distribution and Z-axis gradient at the pixel level to form a basic feature unit; aggregating primary coding nodes, calculating the local area feature mean, gradient direction and change rate, extracting fiber structure features, and forming a low-level coding layer; in the high-level coding layer, introducing an attention mechanism, dynamically adjusting the weights of different areas, highlighting key areas, and forming a composite coding feature vector.

[0009] Furthermore, the regional segmentation of the fiber cross-linked area and the non-cross-linked area includes: calculating the local gradient change rate to identify the cross-linked area, and using an adaptive threshold to divide the candidate area; analyzing the fiber topological connectivity, constructing a network topology map, and preliminarily completing the regional segmentation; using hierarchical morphological operations to optimize the segmentation results and improve the boundary continuity.

[0010] Furthermore, the extraction of the dominant feature of the pore size distribution includes: performing normalization processing on the pore size morphology in different regions, and calculating the pore size gradient information of each region.

[0011] Furthermore, the quantification of the dominant characteristics of the pore size distribution includes: constructing an inter-regional feature comparison matrix based on the pore size gradient information, and using a dynamic weight allocation mechanism to quantify the pore size distribution difference between the cross-linked area and the non-cross-linked area.

[0012] Furthermore, the association coding includes: extracting the pore size gradient characteristics of the cross-linked area and the non-cross-linked area based on the dominant characteristics of the pore size distribution; obtaining the spinning process parameters, including the spinning solution concentration, the spinning speed and the coagulation bath conditions, and standardizing the spinning process parameters; based on the standardized spinning process parameters, a mapping strategy is adopted to nonlinearly associate the spinning process parameters with the pore size gradient characteristics.

[0013] Furthermore, the optimized deep prediction network includes: adopting a contrastive learning strategy, constructing positive and negative sample pairs, extracting the similarity of aperture characteristics under the same process parameters, calculating the cross-scale distribution differences under different process parameters, and generating a contrastive loss function; based on the contrastive loss function, optimizing the feature embedding space of the deep prediction network so that the aperture distribution under different process conditions remains relatively stable in the high-dimensional feature space.

[0014] Furthermore, the error correction mechanism includes: calculating the deviation between the output of the depth prediction network and the actual measurement data, and using a dynamic error decomposition strategy to divide the deviation into structural error and random error.

[0015] Furthermore, the error correction mechanism also includes: for the structural error, retrospectively analyzing the impact of different process parameters on the prediction deviation, and generating a correction factor matrix to make preliminary adjustments to the association coding parameter weights; for the random error, using the Bayesian optimization method to dynamically update the parameter distribution of the association coding, enhance the stability of parameter adjustment, and avoid local convergence problems.

[0016] Compared with the prior art, the composite fiber membrane pore size distribution prediction method based on deep learning provided by the present invention realizes the accurate encoding of the microstructure information of the fiber membrane and improves the extraction accuracy of the pore size distribution characteristics. Through the regional segmentation strategy, the cross-linked area and the non-cross-linked area can be accurately distinguished to ensure that the quantitative analysis of the pore size characteristics is more targeted, thereby improving the resolution ability of the prediction model. The association coding matrix is ​​established in combination with the spinning process parameters, so that the pore size distribution under different process conditions remains consistent, and the applicability of the prediction method is enhanced. The deep prediction network is optimized by contrastive learning, so that the pore size characteristics remain stable in the high-dimensional feature space, and the generalization ability of the model is improved. In addition, the error correction mechanism effectively reduces the prediction deviation and improves the overall prediction accuracy by dynamically adjusting the structural error and random error. Finally, this method can accurately predict the pore size distribution of the composite fiber membrane, provide strong support for material design optimization and production process control, and improve the quality stability and production efficiency of the fiber membrane. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings: Figure 1 This is a flow chart of a method for predicting the pore size distribution of a composite fiber membrane based on deep learning according to an embodiment of the present invention.

[0018] Figure 2 It is a schematic diagram of the process of segmenting the fiber cross-linked area and the non-cross-linked area in the composite fiber membrane pore size distribution prediction method based on deep learning according to an embodiment of the present invention.

[0019] Figure 3 It is a schematic diagram of the process of association coding in the composite fiber membrane pore size distribution prediction method based on deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] Below, the exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention, and it should be understood that the present invention is not limited to the exemplary embodiments described here.

[0021] As mentioned in the background technology above, at present, the pore size distribution prediction based on deep learning mainly adopts a method that combines image feature extraction and regression modeling, but it has the following limitations. For example, traditional image feature extraction methods are difficult to effectively capture the complexity of the fiber membrane microstructure, especially in terms of the pore size gradient difference between the cross-linked area and the non-cross-linked area, which easily leads to a decrease in prediction accuracy; most existing studies rely on static image features, but ignore the dynamic influence of spinning process parameters on pore size distribution, resulting in insufficient model generalization ability and difficulty in adapting to pore size changes under different process conditions; in addition, deep prediction models often lack an effective correction mechanism during the training process, and fail to make full use of actual measurement data to optimize model parameters, which may lead to a large deviation between the prediction results and the true value.

[0022] Therefore, how to improve the accuracy and adaptability of pore size distribution prediction is an urgent problem to be solved.

[0023] Figure 1 Flow chart of a method for predicting pore size distribution of composite fiber membrane based on deep learning according to an embodiment of the present invention. Figure 1 As shown, in the composite fiber membrane pore size distribution prediction method based on deep learning, it includes: S1: collecting multiple image data and encoding them separately, and generating a composite coding feature vector through an attention mechanism; S2: based on the composite coding feature vector, the fiber cross-linked area and the non-cross-linked area are segmented, and the dominant characteristics of the pore size distribution are extracted and quantified; S3: combining the dominant characteristics of the pore size distribution and the spinning process parameters, and performing associative coding; S4: based on the associative coding, optimizing the deep prediction network, and using the error correction mechanism to adjust the parameter weights of the associative coding.

[0024] Scanning electron microscopy (SEM) and atomic force microscopy (AFM) are the two main imaging methods for characterizing the pore size distribution of composite fiber membranes. Scanning electron microscopy forms a high-resolution two-dimensional grayscale image by scanning the sample surface with an electron beam, which can clearly show the contour and distribution of the pore size. In contrast, atomic force microscopy is based on scanning the sample surface with a probe, which can provide three-dimensional morphological information at the nanometer level, and is particularly suitable for measuring the microscopic undulations of the fiber surface and the depth information of the pore size. In conventional research, these two methods are usually used independently, resulting in insufficient data complementarity and difficulty in comprehensive utilization. However, in the present invention, we adopted a multimodal fusion strategy to jointly encode the two-dimensional grayscale information of SEM and the three-dimensional morphological data of AFM, thereby constructing a more complete pore size feature representation.

[0025] In an embodiment of the present invention, step S1 specifically includes: collecting scanning electron microscope images and atomic force microscope morphology data, performing grayscale pixel encoding and Z-axis gradient encoding respectively, and dynamically allocating weights through an attention mechanism to generate a composite encoding feature vector.

[0026] It should be noted that the imaging of the scanning electron microscope depends on the interaction between the electron beam and the sample surface, and the sample usually needs to be pre-treated, such as gold plating or platinum plating, to enhance the recovery rate of the electron signal. In the present invention, a high-resolution (at least 10 nm level) scanning electron microscope is used to image the composite fiber membrane at different magnifications (e.g., 500×, 1000×, 5000×) to ensure that both the overall structure and the local details of the aperture can be captured. In addition, we use backscattered electron imaging mode (BSE) to enhance the contrast of different density areas inside the material, so as to more clearly define the aperture edge.

[0027] Atomic force microscopy measures height information by changing the force between the probe and the sample surface. In order to obtain high-precision pore depth and fiber surface morphology, a scanning method combining contact mode and tapping mode is adopted. Among them, the tapping mode can effectively reduce the damage to the probe and the sample, ensuring that the pore structure will not be destroyed during the scanning process. During the data acquisition process, the scanning area size is set to 5μm×5μm to ensure sufficient resolution to characterize the surface topological features of the fiber. In addition, a moderate loading force is applied to the probe to balance the imaging accuracy and probe life, and a denoising algorithm is used to reduce the noise interference that may occur during the scanning process.

[0028] The collected scanning electron microscope images and atomic force microscope morphology data were subjected to feature coding processing. For the scanning electron microscope images, hierarchical coding was performed according to the grayscale pixel distribution to highlight the aperture morphology and contrast information. For the atomic force microscope morphology data, gradient coding was constructed according to the height gradient change in the Z-axis direction to enhance the expression ability of the fiber surface undulation characteristics.

[0029] In the embodiment of the present invention, the main feature of the scanning electron microscope image is reflected in the grayscale distribution, that is, the brightness contrast of different aperture areas. Traditional image analysis methods usually use global threshold segmentation, but this method may not accurately reflect the morphological changes of the aperture. Therefore, the present invention uses a hierarchical coding method to quantize the grayscale pixel distribution at multiple scales to capture the aperture morphological characteristics more finely.

[0030] Specifically, first, the grayscale value (range 0-255) of the SEM image is mapped to a multi-level coding space. Specifically, the local contrast of each pixel can be calculated based on the local histogram equalization method, and encoded according to different grayscale ranges (for example, 0-50, 51-100, 101-150, 151-200, 201-255). The encoding of each pixel not only includes its own grayscale value, but also includes the mean and variance of its adjacent areas to enhance the local contrast characteristics. In traditional methods, the boundary detection of the aperture mainly relies on operators such as Sobel and Canny for edge extraction, but these methods are prone to false detection for complex backgrounds. To this end, the present invention introduces a gradient enhancement convolution kernel to give higher weights to the edge area to highlight the morphological boundary of the aperture.

[0031] Specifically, the core idea of ​​the gradient enhanced convolution kernel is to introduce learnable weight parameters based on the traditional edge detection operator so that the convolution kernel can dynamically adjust the response intensity of different regions. Specifically, the present invention multiplies the traditional Sobel operator with a trainable weight matrix. In the present invention, the trainable weight matrix adopts a Gaussian distribution form and is adaptively adjusted according to local gradient changes. In this way, in the edge area, the response of the convolution kernel will be amplified, thereby improving the recognition ability of the aperture boundary.

[0032] For example, compared with the two-dimensional grayscale information of the scanning electron microscope, the atomic force microscope provides three-dimensional morphological data, and its core feature is the height change in the Z-axis direction. Therefore, it is necessary to perform gradient encoding on the Z-axis information to enhance the expression ability of the aperture morphological features.

[0033] First, the AFM scanning data is reconstructed in three dimensions, and the Z-axis gradient of each pixel is calculated. By calculating the Z-axis gradient, the undulating characteristics of the aperture edge can be obtained. Especially in the fiber cross-linking area, the depth change of the aperture is more significant, so local weighted processing is required to ensure that the information of the cross-linking area is not ignored. After the Z-axis gradient is calculated, a hierarchical quantization method is used to map different gradient values ​​to a unified encoding space. For example, the gradient range is divided into three levels: low (0-10 nm), medium (10-50 nm), and high (>50 nm). The encoding of each level not only considers the gradient value of the current pixel, but also the mean and change trend of the surrounding pixels to construct a more stable representation of the morphological features.

[0034] Exemplarily, calculating the Z-axis gradient of each pixel point includes: the morphological data obtained by the atomic force microscope is usually stored in a matrix form, where each pixel point corresponds to a height value. The traditional method only uses the Z-axis data for direct analysis, and the present invention introduces a three-dimensional reconstruction method based on neighborhood interpolation to improve spatial consistency. The specific steps are as follows: collect AFM data to obtain an initial height matrix; use bicubic interpolation to spatially interpolate the data to fill possible scanning gaps; calculate the mean and variance of the local area to measure the surface roughness; and construct a three-dimensional point cloud model: Specifically, during the AFM scanning process, the device will scan the sample surface point by point according to the set step size. Each scanning point will record its horizontal and vertical coordinates in the horizontal plane (i.e., the X-axis and Y-axis positions), and obtain the height information corresponding to the point (i.e., Z-axis data). All data points in the entire scanning area can be arranged in the three directions of X, Y, and Z to form a spatially distributed data set, i.e., a three-dimensional point cloud model. When establishing the point cloud model, the position coordinates of the X-axis and Y-axis must be clarified first. The values ​​of these two coordinates are usually calculated based on the scanning step size. For example, if the scanning device collects data at intervals of 1 micron, the X-axis or Y-axis coordinates between adjacent pixels will differ by 1 micron. Therefore, the horizontal and vertical coordinate information of each pixel can be directly determined by the set parameters of the scanning device, while the height data of the Z-axis is directly measured by the AFM device, which reflects the actual height of the sample surface at that point.

[0035] Furthermore, the Z-axis gradient is a key parameter used to measure the changes in the height direction of the sample surface. It indicates the rate of change of the height value between different pixel points, which can reflect the degree of fluctuation of the surface morphology. For composite fiber membranes, the Z-axis gradient can help identify the concave-convex structure of the fiber surface and the depth change of the pore size, which is an important basis for morphological analysis.

[0036] In actual operation, in order to improve the stability of calculation, it is usually not just relying on the height change of a single pixel, but integrating the information of multiple pixels around it. For example, a 3×3 or 5×5 local area can be selected, and the average height change rate can be calculated in this area to obtain a smoother and more accurate Z-axis gradient value. This method can effectively reduce the calculation deviation caused by the error of a single pixel point, making the overall gradient information more reliable. When calculating the Z-axis gradient, it is also necessary to perform a certain smoothing process on the data, such as using the mean filter or median filter method to remove outliers, so that the Z-axis gradient data is more consistent with the actual surface morphology. Finally, the calculated Z-axis gradient information will be combined with the grayscale encoding information to form a composite encoding feature vector.

[0037] Based on the above-mentioned grayscale pixel encoding and Z-axis gradient encoding, it is necessary to construct a composite coding feature vector to uniformly characterize the structural information of the aperture, including the following steps: preliminarily quantify the grayscale distribution and Z-axis gradient at the pixel level to form basic feature units, each feature unit corresponds to the grayscale value or gradient value of a local pixel area; perform regional aggregation on the primary coding nodes, calculate the feature mean, gradient direction and change rate of the local area based on the spatial neighborhood information, extract the local fiber structure features, and form a low-level coding layer; at a higher level, introduce the attention mechanism to dynamically adjust the weights of different regions to highlight the areas that contribute more to the aperture distribution characteristics, and form the final composite coding feature vector.

[0038] Exemplarily, preliminary quantification of grayscale distribution and Z-axis gradient includes: using K-means clustering method to divide grayscale values ​​into k categories, each category corresponds to a coding label; standardizing gradient values ​​and performing multi-level encoding according to set thresholds; each feature unit is formed by a combination of corresponding grayscale coding values ​​and Z-axis gradient coding values.

[0039] In the embodiment of the present invention, step S2 specifically includes: It should be noted that when analyzing the pore size distribution of the composite fiber membrane, it is first necessary to distinguish between the fiber cross-linked area and the non-cross-linked area. This division is a key step, because the pore size distribution of the cross-linked area is often greatly affected by the fiber interlacing, while the non-cross-linked area shows a relatively regular pore size morphology. Therefore, the present invention designs an efficient and accurate regional segmentation method to ensure that subsequent feature extraction and quantification can be performed based on accurate regional division.

[0040] First, if Figure 2As shown, based on the composite coding feature vector generated in the previous step S1, the local gradient change rate is calculated. The local gradient change rate reflects the degree of change in pixel intensity or height in a certain area. Generally, a large gradient change means that there is a structural mutation on the surface of the material, such as the boundary area of ​​fiber cross-linking. Therefore, it is necessary to find gradient mutation points in the composite coding feature vector and use these mutation points as preliminary candidate points for the cross-linking area.

[0041] In traditional methods, commonly used Sobel, Laplacian or Prewitt operators can be used to detect local gradient changes, but these methods are easily affected by noise when processing complex fiber structures, resulting in misjudgment. Therefore, in the present invention, an adaptive gradient calculation method is introduced to adjust the size of the calculation window according to the distribution of local pixel gradient changes. For example, a larger calculation window is used in a flat area to reduce the impact of local fluctuations on the results, and a smaller window is used in a complex cross-linked area to improve the accuracy of boundary recognition, so that the boundaries of the cross-linked area can be more accurately identified, while reducing the misidentification of non-cross-linked areas. Since the cross-linked area is usually manifested as a prominent gradient change, the recognition accuracy of the cross-linked area can be further optimized according to these gradient mutation points, thereby providing a more reliable initial area for subsequent segmentation.

[0042] After identifying the possible cross-linked areas, the regional division needs to be further refined to ensure that the boundaries between the cross-linked areas and the non-cross-linked areas are clear. The present invention adopts an adaptive threshold method to analyze the changing trends of grayscale pixel coding and Z-axis gradient coding, thereby preliminarily dividing the candidate areas of the cross-linked areas and the non-cross-linked areas.

[0043] It should be noted that the traditional fixed threshold method has certain limitations, because the surface morphology of the composite fiber membrane may have a large range of variation due to different manufacturing processes. If a fixed threshold is used, some cross-linked areas may be missed, or some non-cross-linked areas may be misjudged. Therefore, the present invention uses an adaptive threshold method to dynamically adjust the segmentation threshold according to the grayscale distribution of the entire image and the Z-axis gradient distribution characteristics.

[0044] For example, the grayscale histogram distribution of the entire image can be calculated to find the local maximum and minimum values, and the Otsu method or K-means clustering algorithm can be used to determine the optimal threshold to achieve more accurate region segmentation. The advantage of this adaptive threshold method is that it can automatically adjust the segmentation parameters according to different samples, avoiding the instability caused by artificially setting the threshold.

[0045] After completing the preliminary region segmentation, it is necessary to further analyze the topological structure of the fiber to determine the true boundary of the cross-linked region. In the present invention, by calculating the topological connectivity of the fiber structure, a fiber network topology map is constructed to analyze the connection relationship between fiber nodes.

[0046] Among them, the core idea of ​​topological connectivity analysis is to regard the fiber structure as an undirected graph, in which each fiber intersection is used as a node of the graph, and the adjacent fiber segments are used as the edges of the graph. By calculating the degree and connectivity components of the nodes, it is possible to determine which areas belong to the cross-linked area and which belong to the non-cross-linked area. For example, if a node has a high degree (that is, it is connected to multiple other fiber nodes), the area can be judged as a cross-linked area; if a node has a low degree (only connected to one or two nodes), it is more likely to belong to a non-cross-linked area. The advantage of this method is that it can combine the geometric characteristics of the fiber arrangement to improve the accuracy of cross-linked area identification. At the same time, this method can also be used to evaluate the complexity of the cross-linked area. For example, by calculating the clustering coefficient of the network, the compactness of the fiber can be judged, providing valuable information for the subsequent pore size distribution analysis.

[0047] Optionally, after the segmentation of the cross-linked area and the non-cross-linked area is initially completed, further optimization is required to remove isolated noise points and improve the continuity of the boundary. In the present invention, hierarchical morphological operations such as expansion and corrosion are used to refine the segmentation results. The role of the expansion operation is to fill small holes to make the boundary of the cross-linked area more complete, while the corrosion operation is used to remove small isolated noise points to reduce misidentification. By alternating between expansion and corrosion, the segmentation results can be optimized, the boundaries of the cross-linked area and the non-cross-linked area can be made clearer, and the accuracy of subsequent feature extraction can be improved.

[0048] In an embodiment of the present invention, the extraction and quantification of the dominant characteristics of the pore size distribution include: adopting a topological feature matching strategy, normalizing the pore morphology in different regions, and calculating the pore gradient information of each region to obtain the dominant characteristics of the pore size distribution; based on the pore gradient information, constructing an inter-regional feature comparison matrix, and adopting a dynamic weight allocation mechanism to quantify the pore size distribution difference between the cross-linked area and the non-cross-linked area.

[0049] In one embodiment of the present invention, calculating the aperture gradient information of each region includes the following steps: extracting the aperture boundary based on the aperture distribution image, and calculating the spatial gradient change rate of the boundary to measure the morphological change of the aperture edge. In order to improve the accuracy of the calculation, the height change of the aperture region is analyzed in combination with the Z-axis gradient encoding data, and the aperture gradient calculation is optimized. For example, in the area where the boundary mutation is more obvious, a higher weight is given to make it better reflect the actual aperture distribution characteristics.

[0050] In one embodiment of the present invention, each element of the inter-region feature comparison matrix represents the pore size feature difference between two regions, and the rows and columns of the matrix correspond to different regions. By calculating the overall distribution characteristics of the matrix, the pore size difference between the cross-linked region and the non-cross-linked region can be quantified, thereby providing more detailed pore size distribution information.

[0051] Exemplarily, the weights in the feature contrast matrix are adjusted according to the contribution of different regions to the overall pore size distribution. For example, the proportion of the number of pores in each region can be calculated, that is, the ratio of the number of pores in a certain region to the global number of pores. The weight ratio is set so that the contribution of the pore-dense region in the feature contrast matrix is ​​greater, which can ensure that when calculating regional differences, high-density regions will not be diluted by low-density regions, thereby better reflecting the overall distribution characteristics.

[0052] Through the composite coding feature vector, different regions of the fiber structure can be distinguished more accurately, the errors of simple morphological methods can be avoided, and the reliability of regional segmentation can be improved.

[0053] In an embodiment of the present invention, the purpose of step S3 is to associate the dominant characteristics of the pore size distribution with the spinning process parameters to establish a mathematical model between the process parameters and the final fiber structure. The pore size gradient characteristics are used to describe the rate and direction of pore size change in the cross-linked area and the non-cross-linked area. The cross-linked area often has a higher pore size change rate, while the pore size distribution in the non-cross-linked area is more uniform. Therefore, by analyzing the gradient information, the pore size distribution differences in different regions can be quantified, and basic data can be provided for subsequent association coding. Commonly used pore size analysis methods generally include directly calculating pore size distribution, statistical pore size density, and extracting pore size morphological characteristics using Fourier transform or wavelet transform. However, these methods often cannot accurately characterize the gradient changes in the cross-linked area. Therefore, the present invention adopts a spatial distribution matrix to more systematically describe the pore size characteristics.

[0054] like Figure 3 As shown, the association coding includes: based on the dominant characteristics of the pore size distribution, extracting the pore size gradient characteristics of the cross-linked area and the non-cross-linked area; obtaining the spinning process parameters, including spinning solution concentration, spinning speed, coagulation bath conditions, etc., and standardizing the spinning process parameters to make them comparable with the pore size gradient characteristics; based on the standardized spinning process parameters, a mapping strategy is adopted to nonlinearly associate the spinning process parameters with the pore size gradient characteristics.

[0055] Among them, the pore gradient characteristics of the cross-linked area and the non-cross-linked area are extracted by normalization processing, and the gradient data of different areas are converted into dimensionless values ​​for subsequent feature matching and calculation. The dimensionality reduction method is used to extract the most representative pore gradient characteristics and reduce the calculation complexity.

[0056] In the embodiment of the present invention, the spinning process parameters are obtained as follows: the spinning process parameters directly affect the microstructure of the fiber, for example, the spinning solution concentration affects the size and uniformity of the pore size during the fiber formation process; the spinning speed affects the stretching ratio of the fiber, thereby changing the pore size distribution; the coagulation bath conditions affect the phase separation process, thereby changing the pore size structure. Since the physical units of these process parameters are different and the numerical ranges are also quite different, they need to be standardized so that they can be directly compared with the pore size gradient characteristics.

[0057] Furthermore, the relationship between the spinning process parameters and the pore size distribution is not a simple linear relationship. For example, the spinning speed may have a linear effect on the pore size within a certain range, but after exceeding a certain range, it may lead to enhanced phase separation and form a nonlinear effect. Therefore, it is necessary to use a nonlinear mapping method to accurately describe the effect of the spinning process on the pore size distribution. Conventional association modeling methods include linear regression, neural networks, support vector machines (SVM), etc. However, due to the complexity of the spinning process, the present invention is more inclined to use a multivariate influencing factor matrix and combine it with a nonlinear mapping strategy. Specifically, the standardized spinning process parameters and pore size gradient characteristics are selected as input variables, and the process parameters with greater influence are screened to reduce redundant information to form an N×M dimensional matrix (where N is the number of process parameters and M is the number of pore size gradient characteristics). Preferably, Gaussian process regression (GPR) can be used to perform nonlinear mapping between spinning process parameters and pore size gradient characteristics, and the contribution weight of each spinning process parameter to the pore size gradient characteristics can be obtained, that is, the mapping relationship between spinning parameters and pore size characteristics.

[0058] It can be seen that through associative coding, the process parameters (such as spinning solution concentration, spinning speed, coagulation bath conditions, etc.) are mapped to the pore size characteristics, revealing the nonlinear influence between the two and improving the model's sensitivity to process changes. By standardizing the process parameters and performing nonlinear association, the data deviation under different experimental conditions can be effectively reduced, making the prediction results more stable.

[0059] In the embodiment of the present invention, step S4 specifically includes: The core of optimizing the deep prediction network lies in how to effectively learn the impact of different process parameters on pore size distribution and construct a stable feature expression in a high-dimensional feature space. To this end, a contrastive learning strategy is adopted to enhance the model's adaptability to pore size changes under different process conditions by constructing positive and negative sample pairs. Specifically: For the same spinning process parameters (such as spinning solution concentration, spinning speed, coagulation bath conditions, etc.), the corresponding pore size distribution characteristics are extracted to form positive sample pairs to ensure that the model can learn the stable feature expression under the same process conditions; for different process parameter conditions, the corresponding cross-scale differences in pore size distribution are calculated to form negative sample pairs to enhance the model's ability to recognize pore size changes caused by process parameter changes. In conventional methods, contrastive learning is mainly used in the field of image classification or natural language processing, but in the present invention, its core role is to establish a deep association between process parameters and pore size characteristics, so that the model can learn that the pore size distribution under the same process conditions is similar, while the pore size distribution under different process conditions has significant differences, thereby improving the prediction accuracy.

[0060] Preferably, the contrast loss function is mainly used to learn the pore size distribution characteristics under different spinning process parameters, so that the features extracted under similar process conditions have a higher similarity, and have a greater degree of distinction under different process conditions. In the present invention, the contrast loss is constructed by combining the pore size distribution characteristics + spinning process parameters, mainly including: selecting the pore size characteristics under the same process parameters as the positive sample pair, and the pore size characteristics under different process parameters as the negative sample pair; using Euclidean distance or cosine similarity to measure the proximity of the features; and optimizing based on contrast loss, such as Hinge Loss.

[0061] In addition, during the training process, the feature embedding space of the deep prediction network is optimized based on the distribution of process parameters, so that it remains relatively stable under different process conditions and avoids large fluctuations in prediction results due to parameter changes.

[0062] Furthermore, the deep prediction network is often affected by errors in practical applications, and the sources of errors include but are not limited to data noise, process parameter deviations, and the learning ability limitations of the model itself. In order to improve the accuracy of the prediction results, the present invention adopts an error correction mechanism to divide the prediction error into structural errors and random errors, and adopts different processing methods for optimization, wherein the error correction mechanism includes: calculating the deviation between the output of the deep prediction network and the actual measurement data, and using a dynamic error decomposition strategy to divide the deviation into structural errors and random errors; for the structural error, retrospectively analyzing the impact of different process parameters on the prediction deviation, and generating a correction factor matrix, the correction factor matrix makes a preliminary adjustment to the weights of the associated coding parameters; and for the random error, using the Bayesian optimization method to dynamically update the parameter distribution of the associated coding, enhance the stability of the parameter adjustment, and avoid local convergence problems.

[0063] In one embodiment of the present invention, after the prediction model training is completed, the predicted pore size distribution result is compared with the actual measurement data to calculate the error size. Since the prediction error is not uniformly distributed, but is affected by different process parameters, it is necessary to further analyze the source of the error and make targeted corrections.

[0064] Among them, the errors can be divided into the following two categories: structural errors: mainly caused by the influence of process parameters on pore size characteristics, with a certain regularity. For example, a high concentration of spinning solution may lead to an overall decrease in pore size distribution, and this effect can be quantified through statistical analysis; and random errors: mainly caused by random factors such as experimental measurement errors and equipment noise, which cannot be directly modeled by simple mathematical formulas. Different correction methods are adopted for these two types of errors.

[0065] For structural errors, the retrospective analysis method is used to establish the mapping relationship between errors and process parameters, and generate a correction factor matrix. Specifically: the distribution of prediction errors under different process parameters is statistically analyzed, and the trend of errors changing with parameters is analyzed; based on regression analysis or other methods, a quantitative relationship between process parameters and errors can be established to form a correction factor matrix; the correction factor matrix is ​​used to make preliminary adjustments to the weights of associated coding parameters to reduce systematic deviations caused by structural errors. The correction factor matrix can effectively reduce the impact of process parameters on the prediction model and make the prediction results closer to the actual measured data.

[0066] In one embodiment of the present invention, the correction factor matrix may be used for the preliminary adjustment of the associated coding parameter weights through simple weighted adjustment or proportional adjustment.

[0067] Random errors are often difficult to correct through fixed rules, so the Bayesian optimization method is used to dynamically adjust the parameter distribution of the association coding to enhance the stability of parameter adjustment and avoid local convergence problems. The basic steps are as follows: establish a probability distribution model of the association coding parameters, that is, assume that the optimal value of the parameter obeys a certain probability distribution and continuously update it; use the Bayesian optimization method to dynamically adjust the parameter distribution of the association coding to make it converge to the optimal solution; dynamically adjust the parameter weights to ensure the generalization ability of the model on different data sets.

[0068] In one embodiment of the present invention, if the experiment finds that the adjustment of the correction factor matrix cannot effectively reduce the error, or the error is unstable as the process parameters change, a simple Bayesian optimization method can be used to dynamically adjust the parameter distribution of the association code, such as selecting a parameter combination with a smaller error, and increasing its sampling probability, etc., gradually narrowing the parameter search range, so that it converges to the optimal solution. In summary, the composite fiber membrane pore size distribution prediction method based on deep learning according to the embodiment of the present invention is explained, and the composite fiber membrane pore size distribution prediction method based on deep learning provided by the present invention realizes the accurate encoding of the microstructure information of the fiber membrane and improves the extraction accuracy of the pore size distribution characteristics. Through the regional segmentation strategy, the cross-linked area and the non-cross-linked area can be accurately distinguished, ensuring that the quantitative analysis of the pore size characteristics is more targeted, thereby improving the resolution ability of the prediction model. The association coding matrix is ​​established in combination with the spinning process parameters, so that the pore size distribution under different process conditions remains consistent, and the applicability of the prediction method is enhanced. The contrast learning is used to optimize the deep prediction network, so that the pore size characteristics remain stable in the high-dimensional feature space, and the generalization ability of the model is improved. In addition, the error correction mechanism effectively reduces the prediction deviation and improves the overall prediction accuracy by dynamically adjusting the structural error and random error. Finally, this method can accurately predict the pore size distribution of the composite fiber membrane, provide strong support for material design optimization and production process control, and improve the quality stability and production efficiency of the fiber membrane.

Claims

1. A method for predicting pore size distribution of composite fiber membranes based on deep learning, characterized in that: include: Collect multiple image data and encode them separately, and generate composite encoding feature vectors through attention mechanism; Based on the composite coding feature vector, the fiber cross-linked area and the non-cross-linked area are segmented, and the dominant features of the pore size distribution are extracted and quantified; Combining the dominant characteristics of the pore size distribution and the spinning process parameters, performing correlation coding; Based on the association coding, the deep prediction network is optimized, and the parameter weights of the association coding are adjusted by using an error correction mechanism.

2. The method for predicting pore size distribution of composite fiber membranes based on deep learning according to claim 1, characterized in that: The multiple image data include scanning electron microscope images and atomic force microscope morphology data; the separate encoding includes: for scanning electron microscope images, hierarchical encoding according to grayscale pixel distribution; for atomic force microscope morphology data, gradient encoding is constructed according to the height gradient change in the Z-axis direction.

3. The method for predicting pore size distribution of composite fiber membranes based on deep learning according to claim 2, characterized in that: The generation of the composite coding feature vector includes: preliminarily quantizing the grayscale distribution and Z-axis gradient at the pixel level to form a basic feature unit; aggregating primary coding nodes, calculating the local area feature mean, gradient direction and change rate, extracting fiber structure features, and forming a low-level coding layer; in the high-level coding layer, introducing an attention mechanism, dynamically adjusting the weights of different areas, highlighting key areas, and forming a composite coding feature vector.

4. The method for predicting pore size distribution of composite fiber membranes based on deep learning according to claim 1, characterized in that: The segmentation of the fiber cross-linked area and the non-cross-linked area comprises: The local gradient change rate is calculated to identify the cross-linked area, and the candidate area is divided using an adaptive threshold. The fiber topological connectivity is analyzed, a network topology map is constructed, and the region segmentation is preliminarily completed. Hierarchical morphological operations are used to optimize the segmentation results and improve boundary continuity.

5. The method for predicting pore size distribution of composite fiber membranes based on deep learning according to claim 1, characterized in that: The extraction of the dominant feature of the pore size distribution includes: performing normalization processing on the pore size morphology in different regions, and calculating the pore size gradient information of each region.

6. The method for predicting pore size distribution of composite fiber membranes based on deep learning according to claim 5, characterized in that: The quantification of the dominant feature of the pore size distribution includes: constructing an inter-regional feature comparison matrix based on the pore size gradient information, and using a dynamic weight allocation mechanism to quantify the pore size distribution difference between the cross-linked area and the non-cross-linked area.

7. The method for predicting pore size distribution of composite fiber membranes based on deep learning according to claim 1, characterized in that: The association coding includes: extracting the pore size gradient characteristics of the cross-linked area and the non-cross-linked area based on the dominant characteristics of the pore size distribution; obtaining the spinning process parameters, including the spinning solution concentration, the spinning speed and the coagulation bath conditions, and standardizing the spinning process parameters; based on the standardized spinning process parameters, adopting a mapping strategy, nonlinearly associating the spinning process parameters with the pore size gradient characteristics.

8. The method for predicting pore size distribution of composite fiber membranes based on deep learning according to claim 1, characterized in that: The optimized deep prediction network includes: adopting a contrastive learning strategy, constructing positive and negative sample pairs, extracting the similarity of aperture characteristics under the same process parameters, calculating the cross-scale distribution difference under different process parameters, and generating a contrastive loss function; based on the contrastive loss function, optimizing the feature embedding space of the deep prediction network, so that the aperture distribution under different process conditions remains relatively stable in the high-dimensional feature space.

9. The method for predicting pore size distribution of composite fiber membranes based on deep learning according to claim 1, characterized in that: The error correction mechanism includes: calculating the deviation between the output of the depth prediction network and the actual measurement data, and using a dynamic error decomposition strategy to divide the deviation into structural error and random error.

10. The method for predicting pore size distribution of composite fiber membranes based on deep learning according to claim 9, characterized in that: The error correction mechanism also includes: for the structural error, retrospectively analyzing the impact of different process parameters on the prediction deviation, and generating a correction factor matrix to make preliminary adjustments to the association coding parameter weights; for the random error, using the Bayesian optimization method to dynamically update the parameter distribution of the association coding, enhance the stability of parameter adjustment, and avoid local convergence problems.

Images