A Prediction Method for Pore Size Distribution of Composite Fiber Membrane Based on Deep Learning

Abstract

The present invention provides a method for predicting the pore size distribution of a composite fiber membrane based on deep learning, which relates to the technical field of distribution prediction. It encodes various image data respectively and generates a composite encoded feature vector through an attention mechanism; based on the composite encoded feature vector, it performs regional segmentation on the fiber crosslinking region and the non-crosslinking region, extracts and quantifies the dominant features of the pore size distribution; combines the dominant features of the pore size distribution and the spinning process parameters for associated encoding; based on the associated encoding, optimizes the deep prediction network, and adjusts the parameter weights of the associated encoding by using an error correction mechanism. It can accurately predict the pore size distribution of the composite fiber membrane, providing strong support for material design optimization and production process control.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution prediction, 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 fields 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 the fiber membrane is a key parameter that determines its filtration efficiency, air permeability and structural stability. Therefore, accurately predicting the pore size distribution is of great significance for optimizing the spinning process and improving product quality. Traditional methods for measuring pore size distribution mainly rely on experimental detection, such as the bubble method, mercury intrusion method and scanning electron microscope (SEM) imaging analysis. Although these methods have high precision, the testing process is complex, time-consuming and costly, and it is difficult to meet the requirements 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 provides an efficient and intelligent solution for predicting the pore size distribution of fiber membranes.

[0003] Currently, the prediction of pore size distribution based on deep learning mainly adopts the method of combining image feature extraction and regression modeling, but there are the following limitations. For example, traditional image feature extraction methods are difficult to effectively capture the complexity of the microstructure of the fiber membrane, 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 and ignore the dynamic influence of spinning process parameters on the pore size distribution, resulting in insufficient generalization ability of the model 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 fully utilize actual measurement data to optimize model parameters, which may lead to a large deviation between the prediction result 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 at present. 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 a composite fiber membrane based on deep learning.

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

[0007] Further, the variety of image data includes scanning electron microscope images and atomic force microscope topography data; the separate encoding includes: for the scanning electron microscope images, performing hierarchical encoding according to the gray-level pixel distribution; for the atomic force microscope topography data, constructing a gradient encoding based on the height gradient change in the Z-axis direction.

[0008] Further, the generation of the composite encoded feature vector includes: preliminarily quantifying the gray-level distribution and the Z-axis gradient at the pixel level to form basic feature units; aggregating primary encoding nodes, calculating the local region feature mean, gradient direction and change rate, extracting fiber structure features, and forming a low-level encoding layer; at the high-level encoding layer, introducing an attention mechanism to dynamically adjust the weights of different regions, highlighting key regions, and forming a composite encoded feature vector.

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

[0010] Further, the extraction of the dominant features of the pore size distribution includes: normalizing the pore size morphology in different regions and calculating the pore size gradient information of each region.

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

[0012] Further, the associated encoding includes: based on the dominant features of the pore size distribution, extracting the pore size gradient features of the cross-linked region and the non-cross-linked region; obtaining the spinning process parameters, including the spinning solution concentration, spinning speed and coagulation bath conditions, and performing standardization processing on the spinning process parameters; based on the standardized spinning process parameters, using a mapping strategy to non-linearly associate the spinning process parameters with the pore size gradient features.

[0013] Further, the optimized depth prediction network includes: adopting a contrast learning strategy to construct positive and negative sample pairs, extracting the similarity of pore size features under the same process parameters, calculating the cross-scale distribution difference under different process parameters, and generating a contrast loss function; based on the contrast loss function, optimizing the feature embedding space of the depth prediction network to keep the pore size distributions under different process conditions relatively stable in the high-dimensional feature space.

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

[0015] Further, the error correction mechanism also includes: for the structural error, backtracking and analyzing the influence of different process parameters on the prediction deviation, and generating a correction factor matrix to preliminarily adjust the weights of the associated coding parameters; for the random error, adopting a Bayesian optimization method to dynamically update the parameter distribution of the associated coding, enhance the stability of parameter adjustment, and avoid the problem of local convergence.

[0016] Compared with the prior art, the method for predicting the pore size distribution of a composite fiber membrane based on deep learning provided by the present invention realizes the accurate coding of the microscopic structure information of the fiber membrane and improves the extraction accuracy of the pore size distribution characteristics. Through the region 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 enhancing the resolution ability of the prediction model. An associated coding matrix is established in combination with the spinning process parameters, so that the pore size distributions under different process conditions are kept consistent, enhancing the applicability of the prediction method. The depth prediction network is optimized by contrast learning, so that the pore size features are kept stable in the high-dimensional feature space, improving the generalization ability of the model. In addition, the error correction mechanism effectively reduces the prediction deviation and improves the overall prediction accuracy through the dynamic adjustment of the structural error and the 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 technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to these drawings without creative efforts. In the drawings:

[0018] Figure 1 It is a flowchart 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.

[0019] Figure 2 It is a schematic flowchart for region segmentation of the fiber crosslinked region and non-crosslinked region in the pore size distribution prediction method of the composite fiber membrane based on deep learning according to an embodiment of the present invention.

[0020] Figure 3 It is a schematic flowchart for associated encoding in the pore size distribution prediction method of the composite fiber membrane based on deep learning according to an embodiment of the present invention. Detailed implementation manners

[0021] Next, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments of the present invention. It should be understood that the present invention is not limited by the exemplary embodiments described herein.

[0022] As mentioned in the above background art, currently, the pore size distribution prediction based on deep learning mainly adopts the method of combining image feature extraction and regression modeling, but there are the following limitations. For example, traditional image feature extraction methods are difficult to effectively capture the complexity of the microstructure of the fiber membrane, especially in terms of the pore size gradient difference between the crosslinked region and the non-crosslinked region, which easily leads to a decrease in prediction accuracy; most existing studies rely on static image features and ignore the dynamic influence of spinning process parameters on the pore size distribution, resulting in insufficient generalization ability of the model 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 fully utilize the actual measurement data to optimize the model parameters, which may lead to a large deviation between the prediction result and the true value.

[0023] Therefore, how to improve the accuracy and adaptability of pore size distribution prediction is an urgent problem to be solved currently.

[0024] Figure 1 It is a flowchart of the pore size distribution prediction method of the composite fiber membrane based on deep learning according to an embodiment of the present invention. As Figure 1 shown, in the pore size distribution prediction method of the composite fiber membrane based on deep learning, it includes: S1: Collect a variety of image data and perform encoding respectively, and generate a composite encoding feature vector through an attention mechanism; S2: Based on the composite encoding feature vector, perform region segmentation on the fiber crosslinked region and non-crosslinked region, and extract and quantify the dominant features of the pore size distribution; S3: Combine the dominant features of the pore size distribution and the spinning process parameters to perform associated encoding; S4: Based on the associated encoding, optimize the deep prediction network, and adopt an error correction mechanism to adjust the parameter weights of the associated encoding.

[0025] Scanning electron microscopy (SEM) and atomic force microscopy (AFM) are two main imaging techniques for characterizing the pore size distribution of composite fiber membranes. SEM scans the surface of the sample with an electron beam to form a high-resolution two-dimensional grayscale image, which can clearly show the contour and distribution pattern of the pore size. In contrast, AFM is based on a probe scanning the sample surface and can provide three-dimensional topography information at the nanoscale, especially suitable for measuring the microscopic undulations on 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 adopt a multimodal fusion strategy to jointly encode the two-dimensional grayscale information of SEM and the three-dimensional topography data of AFM, thereby constructing a more complete pore size feature representation.

[0026] In an embodiment of the present invention, step S1 specifically includes: collecting SEM images and AFM topography data, respectively performing grayscale pixel encoding and Z-axis gradient encoding, and dynamically allocating weights through an attention mechanism to generate a composite encoded feature vector.

[0027] It should be noted that the imaging of SEM depends on the interaction between the electron beam and the sample surface. Usually, the sample 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) SEM is used to image the composite fiber membrane at different magnifications (such as 500×, 1000×, 5000×) to ensure that both the overall structure and the local details of the pore size can be obtained. In addition, we adopt the backscattered electron imaging mode (BSE) to enhance the contrast of different density regions inside the material, so as to more clearly define the pore edge.

[0028] AFM measures the height information by the change in the force between the probe and the sample surface. In order to obtain high-precision pore depth and fiber surface topography, a scanning method combining the contact mode and the 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 damaged during the scanning process. During the data acquisition process, the scanning area size is set to 5μm×5 μm to ensure sufficient resolution to depict the surface topological features of the fiber. In addition, an appropriate loading force is applied to the probe to balance the imaging accuracy and the probe life, and a denoising algorithm is used to reduce the possible noise interference during the scanning process.

[0029] Perform feature encoding processing on the collected SEM images and AFM topography data. Among them, for the SEM images, hierarchical encoding is performed according to the grayscale pixel distribution to highlight the pore morphology and contrast information; for the AFM topography data, gradient encoding is constructed based on the height gradient change in the Z-axis direction to enhance the expression ability of the fiber surface undulation features.

[0030] In the embodiments of the present invention, the main features of the scanning electron microscope images are reflected in the gray-scale distribution, that is, the brightness contrast in different pore-size regions. Traditional image analysis methods usually adopt global threshold segmentation, but this method may not accurately reflect the morphological changes of the pore size. Therefore, the present invention adopts a hierarchical coding method to perform multi-scale quantization on the gray-level pixel distribution to more precisely capture the morphological features of the pore size.

[0031] Specifically, first, the gray-scale values (range 0 - 255) of the SEM images are mapped to a multi-level coding space. Specifically, based on the local histogram equalization method, the local contrast of each pixel point can be calculated and encoded according to different gray-scale ranges (such as 0 - 50, 51 - 100, 101 - 150, 151 - 200, 201 - 255). The encoding of each pixel point not only includes its own gray-scale value but also the mean and variance of its adjacent regions to enhance the local contrast features. In traditional methods, the boundary detection of the pore size mainly relies on operators such as Sobel and Canny for edge extraction, but these methods are prone to false detection for complex backgrounds. For this reason, the present invention introduces a gradient-enhanced convolution kernel, which assigns higher weights to the edge regions to highlight the morphological boundaries of the pore size.

[0032] Specifically, the core idea of the gradient-enhanced convolution kernel is to introduce learnable weight parameters on the basis of traditional edge detection operators, enabling the convolution kernel to dynamically adjust the response intensity of different regions. Specifically, the present invention multiplies the traditional Sobel operator by 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 regions, the response of the convolution kernel will be amplified, thereby improving the ability to identify the pore-size boundaries.

[0033] Exemplarily, compared with the two-dimensional gray-scale information of the scanning electron microscope, the atomic force microscope provides three-dimensional topography data, and its core feature lies in the height change in the Z-axis direction. Therefore, gradient coding needs to be performed on the Z-axis information to enhance the expression ability of the pore-size topography features.

[0034] First, perform three-dimensional reconstruction on the AFM scanning data, calculate the Z-axis gradient of each pixel point. By calculating the Z-axis gradient, the undulation characteristics of the pore edge can be obtained. Especially in the fiber cross-linking region, the depth change of the pore is more significant. Therefore, local weighted processing is required to ensure that the information in the cross-linking region is not ignored. After calculating the Z-axis gradient, a hierarchical quantization method is adopted to map different gradient values to a unified coding space. For example, the gradient range is divided into three levels: low (0 - 10 nm), medium (10 - 50 nm), and high (>50 nm). The coding of each level not only considers the gradient value of the current pixel point but also the mean value and change trend of its surrounding pixel points to construct a more stable representation of the topography features.

[0035] Exemplarily, calculating the Z-axis gradient of each pixel point includes: The topography data obtained by the atomic force microscope is usually stored in matrix form, where each pixel point corresponds to a height value. Traditional methods only use the Z-axis data for direct analysis, while 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 the initial height matrix; perform spatial interpolation on the data using bicubic interpolation to fill possible scanning gaps; calculate the mean value and variance of the local area to measure the surface roughness; construct a three-dimensional point cloud model:

[0036] Specifically, during the AFM scanning process, the device scans the sample surface point by point according to the set step size. Each scanning point records its abscissa and ordinate in the horizontal plane (i.e., the X-axis and Y-axis positions), and at the same time obtains the corresponding height information of this point (i.e., the Z-axis data). All data points in the entire scanning area can be arranged in the X, Y, and Z directions to form a data set with spatial distribution, that is, a three-dimensional point cloud model. When establishing this point cloud model, it is first necessary to clarify the position coordinates of the X-axis and Y-axis. The values of these two coordinates are usually calculated according to the scanning step size. For example, if the scanning device collects data at an interval of 1 micron, then the X-axis or Y-axis coordinates between adjacent pixel points will differ by 1 micron. Therefore, the horizontal and vertical coordinate information of each pixel point can be directly determined by the set parameters of the scanning device, while the Z-axis height data is directly measured by the AFM device, which reflects the true height of the sample surface at this point.

[0037] Furthermore, the Z-axis gradient is a key parameter used to measure the change of the sample surface in the height direction. It represents the change rate of the height value between different pixel points, thus being able to reflect the undulation degree of the surface topography. For the composite fiber membrane, the Z-axis gradient can help identify the concave and convex structures on the fiber surface and the depth change of the pore diameter, which is an important basis for morphological analysis.

[0038] In actual operation, to improve the stability of calculations, one usually does not rely solely on the height change of a single pixel point but rather synthesizes the information of multiple surrounding pixel points. For example, a local area of 3×3 or 5×5 can be selected, and the average height change rate is calculated within 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, the data also needs to be smoothed, such as using mean filtering or median filtering methods to remove outliers, making the Z-axis gradient data more conform to the true surface morphology. Finally, the calculated Z-axis gradient information will be combined with the gray-level coding information to form a composite coding feature vector.

[0039] Based on the above gray-level pixel coding and Z-axis gradient coding, it is necessary to construct a composite coding feature vector to uniformly represent the structural information of the pore diameter, including the following operation steps: initially quantify the gray-level distribution and Z-axis gradient at the pixel level to form basic feature units, where each feature unit corresponds to the gray 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 spatial neighborhood information, extract the local fiber structure features, and form a low-level coding layer; at a higher level, introduce an attention mechanism to dynamically adjust the weights of different regions to highlight the regions that contribute more to the pore diameter distribution features, forming the final composite coding feature vector.

[0040] Exemplarily, the initial quantification of the gray-level distribution and Z-axis gradient includes: using the K-means clustering method to divide the gray values into k categories, with each category corresponding to a coding label; standardize the gradient values and perform multi-level coding according to a set threshold; each feature unit is formed by combining the corresponding gray coding value and Z-axis gradient coding value.

[0041] In the embodiment of the present invention, step S2 specifically includes:

[0042] It should be noted that when analyzing the pore diameter distribution of the composite fiber membrane, it is first necessary to distinguish between the fiber crosslinking area and the non-crosslinking area. This division is a key step because the pore diameter distribution in the crosslinking area is often greatly affected by fiber interlacing, while the non-crosslinking area shows a relatively regular pore morphology. Therefore, the present invention designs an efficient and accurate region segmentation method to ensure that subsequent feature extraction and quantification can be carried out based on an accurate region division.

[0043] First, as Figure 2As shown, based on the composite coding feature vector generated in the previous step S1, calculate the local gradient change rate. The local gradient change rate reflects the degree of change in pixel intensity or height within a certain area. Usually, a larger gradient change means there is a structural mutation on the material surface, such as the boundary area of fiber crosslinking. 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 crosslinking regions.

[0044] In traditional methods, commonly used Sobel, Laplacian, or Prewitt operators can be used to detect local gradient changes. However, when dealing with complex fiber structures, these methods are easily affected by noise, resulting in misjudgments. Therefore, in the present invention, an adaptive gradient calculation method is introduced. According to the distribution of local pixel gradient changes, the size of the calculation window is adjusted. For example, a larger calculation window is used in flat areas to reduce the impact of local fluctuations on the results, while a smaller window is used in complex crosslinking regions to improve the accuracy of boundary recognition. It can more accurately identify the boundaries of crosslinking regions and reduce misidentifications of non-crosslinking regions. Since crosslinking regions usually exhibit prominent gradient changes, the recognition accuracy of crosslinking regions can be further optimized based on these gradient mutation points, thereby providing a more reliable initial region for subsequent segmentation.

[0045] After identifying the possible crosslinking regions, it is necessary to further refine the region division to ensure clear boundaries between crosslinking regions and non-crosslinking regions. The present invention adopts an adaptive threshold method to analyze the change trends of gray-level pixel coding and Z-axis gradient coding, thereby preliminarily dividing the candidate regions of crosslinking regions and non-crosslinking regions.

[0046] 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 variation range due to different manufacturing processes. If a fixed threshold is used, some crosslinking regions may be missed, or some non-crosslinking regions may be misjudged. Therefore, the present invention uses an adaptive threshold method to dynamically adjust the segmentation threshold according to the overall gray distribution and Z-axis gradient distribution characteristics of the image.

[0047] Exemplarily, the gray histogram distribution of the entire image can be calculated, the local maximum and minimum values can be found, and the Otsu method or K-means clustering algorithm can be used to determine the optimal threshold to achieve more accurate region division. 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 thresholds.

[0048] After completing the preliminary region segmentation, it is also necessary to further analyze the topological structure of the fibers to determine the true boundaries of the crosslinking regions. In the present invention, by calculating the topological connectivity of the fiber structure, a fiber network topology graph is constructed to analyze the connection relationship between fiber nodes.

[0049] Among them, the core idea of topological connectivity analysis is to regard the fiber structure as an undirected graph, where each fiber intersection is used as a node of the graph, and adjacent fiber segments are used as the edges of the graph. By calculating the degrees of the nodes and the connected components, it is possible to determine which regions belong to the crosslinked region and which belong to the non-crosslinked region. For example, if a node has a high degree (i.e., is connected to multiple other fiber nodes), it can be determined that this region is a crosslinked region; while if a node has a low degree (only connected to one or two nodes), it is more likely to belong to the non-crosslinked region. The advantage of this method is that it can combine the geometric characteristics of the fiber arrangement to improve the accuracy of crosslinked region identification. At the same time, this method can also be used to evaluate the complexity of the crosslinked region. For example, by calculating the clustering coefficient of the network, the tightness of the fibers can be judged, providing valuable information for subsequent pore size distribution analysis.

[0050] Optionally, after initially completing the segmentation of the crosslinked region and the non-crosslinked region, 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 dilation and erosion, are used to refine the segmentation result. The role of the dilation operation is to fill small holes and make the boundary of the crosslinked region more complete, while the erosion operation is used to remove small isolated noise points to reduce misidentification. By alternately using dilation and erosion, the segmentation result can be optimized to make the boundary between the crosslinked region and the non-crosslinked region clearer, improving the accuracy of subsequent feature extraction.

[0051] In an embodiment of the present invention, the extraction and quantification of the dominant features of the pore size distribution include: adopting a topological feature matching strategy, normalizing the pore size morphology in different regions, and calculating the pore size gradient information of each region to obtain the dominant features of the pore size distribution; based on the pore size gradient information, constructing an inter-region feature comparison matrix, and adopting a dynamic weight allocation mechanism to quantify the pore size distribution difference between the crosslinked region and the non-crosslinked region.

[0052] In an embodiment of the present invention, calculating the pore size gradient information of each region includes the following steps: extracting the pore size boundary based on the pore size distribution image, and calculating the spatial gradient change rate of the boundary to measure the morphological change of the pore size edge. To improve the calculation accuracy, combining the Z-axis gradient encoded data, analyzing the height change of the pore size region, and optimizing the pore size gradient calculation. For example, in regions where the boundary mutation is more obvious, a higher weight is given to make it better reflect the actual pore size distribution characteristics.

[0053] In an 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 respectively. By calculating the overall distribution characteristics of the matrix, the pore size difference between the crosslinked region and the non-crosslinked region can be quantified, thereby providing more detailed pore size distribution information.

[0054] Exemplarily, according to the contribution degree of different regions to the overall pore size distribution, the weights in the feature comparison matrix are adjusted. For example, the proportion of the pore size number in each region can be calculated, that is, the ratio of the pore size number in a certain region to the global pore size number; a weight ratio is set so that the regions with dense pore sizes have a greater contribution degree in the feature comparison matrix, which can ensure that when calculating the regional differences, the high-density regions will not be diluted by the low-density regions, thus better reflecting the overall distribution characteristics.

[0055] By compound-coding the feature vectors, different regions of the fiber structure can be more accurately distinguished, avoiding the errors of simple morphological methods and improving the reliability of regional segmentation.

[0056] In the embodiment of the present invention, the purpose of step S3 is to associate the dominant features 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 feature is used to describe the rate and direction of the pore size change in the cross-linked region and the non-cross-linked region. The cross-linked region often has a higher pore size change rate, while the pore size distribution in the non-cross-linked region is relatively uniform. Therefore, by analyzing the gradient information, the pore size distribution differences in different regions can be quantified and basic data for subsequent associated coding can be provided. Commonly used pore size analysis methods usually include directly calculating the pore size distribution, statistically analyzing the pore size density, and extracting the pore size morphological features using Fourier transform or wavelet transform. However, these methods often cannot accurately characterize the gradient changes in the cross-linked region. Therefore, the present invention adopts the method of spatial distribution matrix to more systematically describe the pore size characteristics.

[0057] As Figure 3 shown, the associated coding includes: based on the dominant features of the pore size distribution, extracting the pore size gradient features of the cross-linked region and the non-cross-linked region; obtaining the spinning process parameters, including the spinning solution concentration, the spinning speed, the coagulation bath conditions, etc., and performing standardization processing on the spinning process parameters to make them comparable with the pore size gradient features; based on the standardized spinning process parameters, adopting a mapping strategy to non-linearly associate the spinning process parameters with the pore size gradient features.

[0058] Among them, extracting the pore size gradient features of the cross-linked region and the non-cross-linked region is to perform normalization processing to convert the gradient data in different regions into dimensionless values for subsequent feature matching and calculation, and adopting a dimensionality reduction method to extract the most representative pore size gradient features and reduce the calculation complexity.

[0059] In the embodiments of the present invention, the obtaining of the spinning process parameters includes: The spinning process parameters directly affect the microstructure of the fiber. For example, the concentration of the spinning solution affects the size and uniformity of the pore diameter during the fiber formation process; the spinning speed affects the draw ratio of the fiber, thereby changing the pore diameter distribution; the coagulation bath conditions affect the phase separation process, thereby changing the pore structure. Since the physical units of these process parameters are different and the numerical ranges also vary greatly, standardization processing is required so that they can be directly compared with the pore diameter gradient characteristics.

[0060] Furthermore, the relationship between the spinning process parameters and the pore diameter distribution is not a simple linear relationship. For example, the spinning speed may have a linear effect on the pore diameter size within a certain range, but beyond a certain range, it may lead to enhanced phase separation, forming a non-linear effect. Therefore, a non-linear mapping method needs to be adopted to accurately describe the influence of the spinning process on the pore diameter distribution. Conventional correlation modeling methods include linear regression, neural networks, support vector machines (SVM), etc. However, due to the complexity of the spinning process, the present invention prefers to use a multi-variable influence factor matrix and combine it with a non-linear mapping strategy. Specifically, the standardized spinning process parameters and the pore diameter gradient characteristics are selected as input variables, and the process parameters with greater influence are screened to reduce redundant information, forming an N×M-dimensional matrix (where N is the number of process parameters and M is the number of pore diameter gradient characteristics). Preferably, Gaussian process regression (GPR) can be used for the non-linear mapping between the spinning process parameters and the pore diameter gradient characteristics, and the contribution weight of each spinning process parameter to the pore diameter gradient characteristics, that is, the mapping relationship between the spinning parameters and the pore diameter characteristics, can be obtained.

[0061] It can be seen that through correlation coding, the process parameters (such as the concentration of the spinning solution, the spinning speed, the coagulation bath conditions, etc.) are mapped to the pore diameter characteristics, revealing the non-linear influence between the two, improving the sensitivity of the model to process changes, and through standardizing the process parameters and performing non-linear correlation, it is possible to effectively reduce the data deviation under different experimental conditions, making the prediction results more stable.

[0062] In the embodiments of the present invention, step S4 specifically includes:

[0063] The core of optimizing the depth prediction network lies in how to effectively learn the influence of different process parameters on the pore diameter distribution and construct a stable feature representation in the high-dimensional feature space. For this purpose, a contrastive learning strategy is adopted. By constructing positive and negative sample pairs, the adaptability of the model to the pore diameter changes under different process conditions is enhanced. Specifically:

[0064] For the same spinning process parameters (such as spinning solution concentration, spinning speed, coagulation bath conditions, etc.), extract the corresponding pore size distribution characteristics to form positive sample pairs, ensuring that the model can learn stable feature expressions under the same process conditions; for different process parameter conditions, calculate the cross-scale differences in their corresponding pore size distributions to form negative sample pairs, so as to enhance the model's recognition ability for pore size changes caused by process parameter variations. In conventional methods, contrastive learning is mainly used in the fields 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, enabling the model to learn that the pore size distributions are similar under the same process conditions, while there are significant differences in the pore size distributions under different process conditions, thereby improving the prediction accuracy.

[0065] Preferably, the contrastive 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 high degree of similarity, while having a large degree of distinguishability under different process conditions. In the present invention, the contrastive 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 positive sample pairs, and those under different process parameters as negative sample pairs; using Euclidean distance or cosine similarity to measure the closeness of features; and optimizing based on the contrastive loss, such as in the form of Hinge Loss.

[0066] In addition, during the training process, in combination with the distribution of process parameters, optimize the feature embedding space of the deep prediction network to keep it relatively stable under different process conditions, and avoid large fluctuations in the prediction results caused by parameter changes.

[0067] Furthermore, in practical applications, the deep prediction network is often affected by errors, and the sources of errors include but are not limited to data noise, process parameter deviations, and limitations in the model's own learning ability. To improve the accuracy of the prediction results, the present invention adopts an error correction mechanism, divides the prediction error into structural error and random error, and respectively adopts different processing methods for optimization. Among them, 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 error and random error; for the structural error, retrospectively analyze the influence of different process parameters on the prediction deviation, and generate a correction factor matrix, which preliminarily adjusts the weights of the associated coding parameters; and for the random error, adopt the Bayesian optimization method to dynamically update the parameter distribution of the associated coding, enhance the stability of parameter adjustment, and avoid the problem of local convergence.

[0068] In one embodiment of the present invention, after the prediction model is trained, the predicted pore size distribution result is compared with the actual measurement data to calculate the error magnitude. 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.

[0069] Among them, the errors can be divided into the following two categories: Structural errors: mainly caused by the influence of process parameters on pore characteristics, with a certain regularity. For example, a higher spinning solution concentration may lead to an overall reduction in the pore size distribution, and this influence can be quantified through statistical analysis; and Random errors: mainly caused by random factors such as experimental measurement errors and equipment noise, and cannot be directly modeled through simple mathematical formulas. Different correction methods are adopted for these two types of errors.

[0070] For structural errors, a backtracking analysis method is adopted to establish the mapping relationship between the error and process parameters and generate a correction factor matrix. Specifically: statistically analyze the prediction error distribution under different process parameters and analyze the trend of the error changing with the parameters; a quantitative relationship between the process parameters and the error can be established based on regression analysis or other methods to form a correction factor matrix; the correction factor matrix is applied to the preliminary adjustment of the weights of the associated coding parameters to reduce the systematic deviation caused by structural errors. The correction factor matrix can effectively reduce the influence of process parameters on the prediction model and make the prediction result closer to the actual measurement data.

[0071] In one embodiment of the present invention, the correction factor matrix can be applied to the preliminary adjustment of the weights of the associated coding parameters through simple weighted adjustment or proportional adjustment.

[0072] Random errors are often difficult to correct through fixed rules. Therefore, the Bayesian optimization method is used to dynamically adjust the parameter distribution of the associated coding to enhance the stability of parameter adjustment and avoid local convergence problems. The basic steps are as follows: establish a probability distribution model for the parameters of the associated coding, that is, assume that the optimal value of the parameter follows a certain probability distribution and continuously update it; use the Bayesian optimization method to dynamically adjust the parameter distribution of the associated 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.

[0073] In one embodiment of the present invention, if experiments find that the adjustment of the correction factor matrix cannot effectively reduce the error, or the error is unstable with the change of process parameters, a simple Bayesian optimization method can be used to dynamically adjust the parameter distribution of the associated coding. For example, select the parameter combination with smaller error and increase its sampling probability, etc., gradually narrow the parameter search range and make it converge to the optimal solution. In summary, the prediction method of the pore size distribution of the composite fiber membrane based on deep learning according to the embodiments of the present invention is elucidated. The prediction method of the pore size distribution of the composite fiber membrane based on deep learning provided by the present invention realizes the accurate coding of the microscopic structure information of the fiber membrane and improves the extraction accuracy of the pore size distribution characteristics. Through the region 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 of the prediction model. By establishing an associated coding matrix in combination with the spinning process parameters, the pore size distribution under different process conditions is kept consistent, enhancing the applicability of the prediction method. Using contrastive learning to optimize the deep prediction network makes the pore size characteristics stable in the high-dimensional feature space and improves the generalization ability of the model. In addition, the error correction mechanism effectively reduces the prediction deviation and improves the overall prediction accuracy through the dynamic adjustment of the structural error and the 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 the pore size distribution of a composite fiber membrane based on deep learning, characterized in that, Including: Collecting multiple types of image data, encoding them separately, and generating a composite encoded feature vector through an attention mechanism; Based on the composite encoded feature vector, performing regional segmentation on the fiber crosslinked region and the non-crosslinked region, and extracting and quantifying the dominant features of the pore size distribution; Combining the dominant features of the pore size distribution and the spinning process parameters for associated encoding; The associated encoding includes: based on the dominant features of the pore size distribution, extracting the pore size gradient features of the crosslinked region and the non-crosslinked region; obtaining the spinning process parameters, including the spinning solution concentration, the spinning speed, and the coagulation bath conditions, and performing standardization processing on the spinning process parameters; based on the standardized spinning process parameters, adopting a mapping strategy to nonlinearly associate the standardized spinning process parameters with the pore size gradient features; Based on the associated encoding, optimizing the depth prediction network and using an error correction mechanism to adjust the parameter weights of the associated encoding.

2. The method for predicting the pore size distribution of the composite fiber membrane based on deep learning according to claim 1, wherein The multiple types of image data include scanning electron microscope images and atomic force microscope topography data; the separate encoding includes: for the scanning electron microscope images, performing hierarchical encoding according to the gray-level pixel distribution; for the atomic force microscope topography data, constructing gradient encoding based on the height gradient change in the Z-axis direction.

3. The method for predicting the pore size distribution of the composite fiber membrane based on deep learning according to claim 2, wherein The generation of the composite encoded feature vector includes: preliminarily quantifying the gray-level distribution and the Z-axis gradient at the pixel level to form basic feature units; aggregating primary encoding nodes, calculating the local region feature mean, the gradient direction, and the change rate, and extracting the fiber structure features to form a low-level encoding layer; at the high-level encoding layer, introducing an attention mechanism to dynamically adjust the weights of different regions, highlighting the key regions, and forming a composite encoded feature vector.

4. The method for predicting the pore size distribution of a composite fiber membrane based on deep learning according to claim 1, wherein The regional segmentation of the fiber crosslinked region and the non-crosslinked region includes: Calculating the local gradient change rate to identify the crosslinked region, and using an adaptive threshold to divide the candidate regions; analyzing the fiber topological connectivity, constructing a network topology graph, and initially completing the regional segmentation; adopting hierarchical morphological operations to optimize the segmentation result and improve the boundary continuity.

5. The method for predicting the pore size distribution of the composite fiber membrane based on deep learning according to claim 1, wherein The extraction of the dominant features 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 the pore size distribution of the composite fiber membrane based on deep learning according to claim 5, wherein The quantification of the dominant features of the pore size distribution includes: based on the pore size gradient information, constructing an inter-region feature comparison matrix and using a dynamic weight allocation mechanism to quantify the pore size distribution difference between the crosslinked region and the non-crosslinked region.

7. The method for predicting the pore size distribution of the composite fiber membrane based on deep learning according to claim 1, wherein The optimization of the depth prediction network includes: adopting a contrast learning strategy, constructing positive and negative sample pairs, extracting the similarity of the pore size features under the same process parameters, calculating the cross-scale distribution difference under different process parameters, and generating a contrast loss function; based on the contrast loss function, optimizing the feature embedding space of the depth prediction network to keep the pore size distribution under different process conditions relatively stable in the high-dimensional feature space.

8. The method for predicting the pore size distribution of the composite fiber membrane based on deep learning according to claim 1, wherein, 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.

9. The method for predicting the pore size distribution of a composite fiber membrane based on deep learning according to claim 8, wherein The error correction mechanism further includes: for the structural error, retrospectively analyzing the influence of different process parameters on the prediction deviation, and generating a correction factor matrix to preliminarily adjust the weights of the associated coding parameters; for the random error, adopting a Bayesian optimization method to dynamically update the parameter distribution of the associated coding, enhance the stability of parameter adjustment, and avoid the problem of local convergence.

Images