A cement and a method and system for non-direct measurement of elemental content of a cementitious material

Abstract

The application discloses a cement and a non-direct measurement method and system for element content of a cement-based material element; wherein the method comprises the following steps: collecting microstructure images and element surface scanning images of a detection object; the detection object is cement or a cement-based material; using the collected data to make a data set; constructing a neural network and training the network with the data set; the neural network is a network model for predicting the element content of pixel points or voxel points of cement; collecting microstructure data of a to-be-detected object, inputting the microstructure data of the to-be-detected object into the trained neural network, and obtaining a preliminary prediction result; sampling the preliminary prediction result to obtain the prediction result of the element surface scanning or element three-dimensional body corresponding to the to-be-detected object.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing for the purpose of prediction, and in particular, to a method and system for non-direct measurement of cement and its base material element content. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Cement, an important material indispensable to the modern construction industry, is self-evident in importance. With the development of science and technology and the demand for special cement in special environments in production, cement has not been limited to the role of bonding material. More special cements have been developed, such as Portland clinker cement, aluminate clinker cement, sulphoaluminate clinker cement and fluoroaluminate clinker cement; sulphate-resistant cement for sulphate-resistant, freeze-thaw and dry-wet port water conservancy engineering, expansion cement with expansion properties, oil well cement specially used for oil and gas well cementing, refractory cement, radiation-resistant cement and the like. Like other materials, the properties of cement are basically determined by the compounds and constituent elements inside. At present, the research on high-performance cement and special cement in the international community mainly starts from two aspects: one is the evolution mechanism of cement hydration, and the other is the optimization of cement formula. No matter which aspect of research, the microcosmic world of cement needs to be analyzed, because the microcosm determines the macrocosm. In the process of studying the microcosm, researchers can only obtain the internal microstructure characterization of the material through large instruments such as scanning electron microscope or CT tomography. In order to obtain the internal element situation, EDS energy spectrum analyzer can only be used. However, many scientific research institutions do not have EDS energy spectrum analyzer, and cannot obtain the internal element situation of cement. In addition, the time cost and money cost in the process of collection are also very high. Therefore, some academic works in the international community are trying to predict the content of elements in different materials. However, their methods have obvious shortcomings and defects: first, the existing methods are coarse-grained content prediction, that is, starting from the whole of the material to be predicted, the element content prediction is regarded as a general regression problem, and the overall content of the element in the current sample is predicted. The prediction results of this series of methods are coarse-grained and lack of precision; second, the existing methods are deterministic prediction, that is, starting from the whole of the material to be predicted, a single value prediction result is obtained, while the content of elements in the material is often complex. The lack of information and the uncertainty introduced by the error in the prediction process greatly reduce the effectiveness of this kind of prediction method; third, the existing methods are from the macroscopic point of view to predict the content of elements, while researchers almost always use cement microstructure images to analyze, so the existing methods cannot well assist researchers in studying the microcosmic situation such as cement hydration mechanism; fourth, the EDS energy spectrum analyzer can only obtain the distribution of elements within the two-dimensional image range, which is a great limitation to the research on high-performance cement. SUMMARY

[0004] In order to solve the problems in the prior art, the application provides a cement and a non-direct measurement method and system for element content of cement base material; the method is directed to a cement microstructure image or a three-dimensional cement microstructure body to predict element content; the method is pixel point level, can predict element content of each pixel of the microstructure image or each voxel of the three-dimensional microstructure body, and has very fine granularity; the method obtains all possibilities and probabilities of element content instead of a single element content value; the method can not only predict a two-dimensional cement microstructure image, but also can be extended to predict a three-dimensional cement microstructure body, and solves the problem that an EDS energy spectrum analyzer can only obtain a two-dimensional image.

[0005] In a first aspect, the application provides a non-direct measurement method for element content of cement and cement base material;

[0006] The non-direct measurement method for element content of cement and cement base material comprises the following steps.

[0007] S1: collecting a microstructure image and an element area scanning image of a detection object; the detection object is cement or cement base material; using the collected data, a data set is prepared;

[0008] S2: constructing a neural network and training the network with the data set; the neural network is a network model for predicting element content at a pixel point or voxel point level of cement;

[0009] S3: collecting microstructure data of a to-be-detected object, inputting the microstructure data of the to-be-detected object into the trained neural network, and obtaining a preliminary prediction result;

[0010] S4: sampling the preliminary prediction result to obtain a prediction result of element area scanning or volume scanning corresponding to the to-be-detected object.

[0011] Further, S1: collecting a microstructure image and an element area scanning image of a detection object comprises the following steps.

[0012] A cement microstructure image is obtained, and a cement element area scanning image at the same position is obtained by using a scanning electron microscope combined with an EDS energy spectrum analyzer;

[0013] In the step of obtaining the cement microstructure image, a scanning electron microscope, a tomographic CT or any instrument capable of obtaining a cement microstructure image is used.

[0014] Further, when the cement microstructure image is obtained, the gray value is not set to have an upper and lower limit; when the cement element area scanning image is obtained, the number of sampling frames is set to a specified number of frames; the higher the specified number of frames is, the better.

[0015] Further, the collected data is used to make a data set, including:

[0016] S11: According to the characteristics of Markov random field and isotropy of cement microstructure, a sliding window of a specified size is designed, and the first point in the upper left corner of the collected microstructure data is taken as the initial center point of the sliding window. The sliding window is sequentially slid from left to right and from top to bottom, and the two-dimensional image or three-dimensional body in the sliding window is intercepted as the input data corresponding to the first point. The label corresponding to the first point is the gray value of the position corresponding to the first point in the element surface scanning data corresponding to the microstructure data.

[0017] S12: Then, according to the method of S11, all points of the cement microstructure data are traversed to obtain the input data and labels of all points, and a data set is obtained.

[0018] Among them, the collected data used to make a data set has different meanings for different problem scenarios. For two-dimensional element content prediction, the collected data refers to the collected original microstructure image and original element surface scanning image. For three-dimensional element content prediction, the collected data refers to a three-dimensional microstructure body and a three-dimensional element scanning body.

[0019] Further, the three-dimensional microstructure body refers to a three-dimensional microstructure body containing the internal microstructure of cement, which is obtained by sequentially stacking the collected two-dimensional cement microstructure images in sequence and on the channel. According to the same stacking sequence of the two-dimensional cement microstructure image, the corresponding element surface scanning image is stacked to obtain a three-dimensional scanning body of cement elements.

[0020] Further, S2: constructing a neural network and training the network with the data set, including:

[0021] S21: constructing a neural network, wherein the neural network includes a feature extraction module and a probability modeling module connected in sequence;

[0022] S22: After the construction of the neural network, the data set is used to select a log-likelihood loss function, and the neural network is trained by using back propagation.

[0023] Further, the feature extraction module refers to a neural network for feature extraction of input cement data. The feature extraction module is a fully connected network, a two-dimensional convolutional neural network or a three-dimensional convolutional neural network, and any one of them. For a two-dimensional convolutional neural network or a three-dimensional convolutional neural network, the obtained feature map is finally unfolded into a one-dimensional vector.

[0024] Furthermore, the probability modeling module refers to an output layer containing multiple parallel fully connected branches, with each fully connected branch having only one fully connected layer.

[0025] Furthermore, the specific construction steps for the output layer of the multiple parallel fully connected branches include:

[0026] S211: Select a kernel function; the kernel function includes, but is not limited to, Gaussian function, Laplace function, etc., different kernel functions include different numbers of parameters, the number of parameters of each kernel function is denoted as N, and the number of fully connected branches is set as N+1; the selection principle of the kernel function is based on the form of the probability distribution followed by the target problem in the problem to be solved;

[0027] S212: Set the number of kernel functions to M, and the number of neurons in each fully connected branch to M. The output of the first fully connected branch is the weights (α1, α2, ..., α...) of all M kernel functions. M The output of the second fully connected branch is the parameters of all M kernel functions. 11 parameter 21 , ..., parameter M1 ), and so on, the output of the (N+1)th fully connected branch is the parameters of all M kernel functions. 1N parameter 2N , ..., parameter MN ).

[0028] Further, S3: Collect microstructure data of the object under test, input the microstructure data of the object under test into the trained neural network, and obtain preliminary prediction results, including:

[0029] S31: Collect the microstructure data of the object to be tested, design a sliding window of a specified size, take the first point in the upper left corner of the cement microstructure data as the initial center point of the sliding window, slide it sequentially from left to right and from top to bottom, and capture the two-dimensional image or three-dimensional volume within the sliding window as the input data corresponding to the first point. Similarly, obtain the input data corresponding to each point of the microstructure data of the object to be tested.

[0030] S32: Input data corresponding to each point of the microstructure data of the object under test is fed into the trained neural network to obtain the prediction result; the prediction result is the total parameters of all kernel functions at each point.

[0031] S33: Perform a weighted summation of all M kernel functions at each point:

[0032]

[0033] wherein the i-th kernel function F(i) is a kernel function composed of corresponding parameters parameter i1 , parameter i2 , …, parameter iN ; parameter i1 represents the 1st parameter of the i-th kernel function; parameter i2 represents the 2nd parameter of the i-th kernel function; parameter iN represents the Nth parameter of the i-th kernel function; and i is the weight of the i-th kernel function obtained by the first full connection branch.

[0034] After the weighted summation, a mixing function F at each point of the input data is obtained, and the mixing function F represents a probability density function of the element content at the current point.

[0035] Further, S4: sampling the preliminary prediction result to obtain a prediction result of an element surface scan or a volume scan of the object to be measured, comprising:

[0036] S41: sampling the probability density function at each point to obtain a maximum probability value of the element content at the current point;

[0037] S42: splicing the maximum probability value of the element content at each point according to the corresponding position to obtain a matrix, and obtaining a prediction result of an element surface scan or a three-dimensional scanning volume of the object to be measured.

[0038] In a second aspect, the present application provides a cement and its base material element content non-direct measurement system.

[0039] A cement and its base material element content non-direct measurement system, comprising:

[0040] A data set making module configured to: collect microstructure images and element surface scan images of a detection object; the detection object is cement or cement base material; and use the collected data to make a data set;

[0041] A neural network construction module configured to: construct a neural network and train the network with the data set; and the neural network is a network model for predicting the element content at the pixel or voxel level of the cement;

[0042] A prediction module configured to: collect microstructure data of an object to be measured, input the microstructure data of the object to be measured into the trained neural network, and obtain a preliminary prediction result;

[0043] The sampling module is configured to sample the preliminary prediction result to obtain a prediction result of an element surface scan or a volume scan corresponding to the to-be-tested object.

[0044] In a third aspect, the present application provides a method for non-direct measurement of element content of cement and base material thereof,

[0045] A method for non-direct measurement of element content of cement and base material thereof comprises:

[0046] Collecting a microstructure image and an element surface scan image of a detection object; the detection object is cement or cement base material; using the collected data, a data set is made;

[0047] A microstructure three-dimensional voxel element content prediction neural network is constructed, and the neural network is used to predict the probability density function of the element content of each voxel of the microstructure three-dimensional body; the data set is used to train the network;

[0048] The microstructure three-dimensional body of the to-be-tested cement or cement base material is pretreated, the pretreated microstructure three-dimensional body is input into the trained neural network, and the probability density function of the chemical element content at each voxel is output.

[0049] Compared with the prior art, the present application has the following beneficial effects:

[0050] The use object of the present application is not only a two-dimensional microstructure image of cement, but also a three-dimensional microstructure body of cement, which breaks the limitation and barrier that the EDS energy spectrum analyzer can only obtain a two-dimensional element surface scan image, greatly helps researchers to study high-performance cement, and breaks the monopoly position of the EDS energy spectrum analyzer to obtain an element surface scan image, so that researchers can obtain the element pixel level content and distribution in the cement without using the EDS energy spectrum analyzer.

[0051] The present application introduces probability theory into element prediction for the first time from the perspective of statistics, breaks the method that all existing methods can only predict the overall content of elements in the current sample, and for the first time, the probability density function of the element content at each pixel of the cement microstructure image or at each voxel of the cement microstructure three-dimensional body is predicted.

[0052] The method in the present application adopts a neural network to construct an element prediction model, the neural network is specially designed and has strong speciality, fully adapts to the texture characteristics and isotropic characteristics of the cement microstructure image and the cement microstructure three-dimensional body, fully utilizes the powerful feature extraction capability of the neural network, and effectively improves the prediction accuracy. The method in the present application is a real-time prediction method, which can quickly return a prediction result for a cement microstructure image provided by a user, and has good real-time performance. BRIEF DESCRIPTION OF DRAWINGS

[0053] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification. The embodiments of the application, together with its

[0054] Figure 1 is a flow chart of the whole method of the embodiment one of the present application;

[0055] Figure 2 is a structural schematic diagram of the neural network of the embodiment one of the present application;

[0056] Figure 3 is a structural schematic diagram of the neural network of the embodiment three of the present application. DETAILED DESCRIPTION

[0057] It should be noted that the following detailed description is merely exemplary and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0058] It is to be understood that the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. It should be noted that the terms "comprises", "comprising", "includes", "including", "contains", "containing" or variations thereof do not specify an exhaustive inclusion or list, and that the terms are not intended to exclude other possible elements or methods. It should be noted that the use of the term "or" includes a variety of implementations. For example, the term "X employs A or B" means that X employs A; X employs B; or X employs both A and B. It should be noted that the use of the term "at least" followed by a set of one or more items suggests that the set of items include at least one of the listed items. For example, "at least one of A or B" means A; B; or A and B. It should be noted that the use of the term "about" means that a value includes the standard deviation of the mean. For example, "about 5" means 5 ± 1. It should be noted that the use of the term "set" includes a variety of implementations. For example, the term "at least one of a set of items" means at least one of one or more of the items; a set of one or more of the items; or the items themselves. It should be noted that the use of the term "set" includes a variety of implementations. For example, the term "at least one of a set of items" means at least one of one or more of the items; a set of one or more of the items; or the items themselves.

[0059] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0060] All data of the present embodiment are acquired on the basis of compliance with laws and regulations and user consent, and legal application of the data.

[0061] In the field of material science, element content and spatial distribution are a crucial step in basic research, which plays a very important auxiliary role in structure inference and optimization of simulation design, and is also the case in cement concrete design and optimization. No matter in the aspects of cement type, hydration speed, strength, life span, etc., the element content and spatial distribution play a decisive role. The present application discloses a cement element content non-direct measurement method and system, aiming to optimize the problem that element content analysis is extremely dependent on equipment, high investment and low efficiency, the problem that the existing method is not fine enough in granularity, the problem that the existing method is inaccurate in predicting single value of element content, and the problem that the existing method can only make two-dimensional prediction. The method introduces probability theory into the prediction of element content, effectively realizes the function of predicting the probability density function of the element content of each pixel point according to the microstructure image, controls the error within an acceptable range, and saves a lot of time and financial cost.

[0062] Embodiment one

[0063] The present embodiment provides a cement and its base material element content non-direct measurement method;

[0064] As shown in Figure 1 , a cement and its base material element content non-direct measurement method comprises:

[0065] S101: Collecting microstructure images and element surface scanning images of a detection object; the detection object is cement or cement base material; using the collected data, a data set is made;

[0066] S102: Constructing a neural network and training the network with the data set; the neural network is a network model for predicting the element content of pixel points of cement;

[0067] S103: Collecting microstructure images of a to-be-detected object, inputting the microstructure images of the to-be-detected object into the trained neural network, and obtaining a preliminary prediction result;

[0068] S104: Sampling the preliminary prediction result to obtain a prediction result of element surface scanning corresponding to the to-be-detected object.

[0069] Further, the detection object refers to cement samples or cement base material samples with different proportions; wherein, the preparation process of the samples comprises:

[0070] Making cement samples or cement base material samples with different proportions, so that the hydration time of the samples is different;

[0071] According to different proportions and different hydration times, the samples are divided into different groups.

[0072] Exemplarily, in order to ensure the accuracy and robustness of the model in the later stage, the cement data set is required to be comprehensive and balanced according to the theoretical requirements. For example, cement and cement-based materials, as many types of cement material data sets as possible are made. At least cover the commonly used cement types today, such as Portland cement, ordinary Portland cement, slag Portland cement, pozzolanic Portland cement, coal powder Portland cement, composite Portland cement and part of aluminate cement, etc. Therefore, different formula cement samples are made, and multiple samples are made for each formula. The samples in the same formula have different hydration times.

[0073] After the completion of the above-mentioned various control experiment groups of cement and its base materials, in order to ensure that the subsequent image sampling process is not disturbed by external factors, the same standing, curing, packaging operation is carried out, and once it is completed, it is not unpacked until the instrument obtains the image data set.

[0074] Further, S101: collecting microstructure images and element surface scanning images of the detection object, specifically comprising:

[0075] Using a scanning electron microscope to collect microstructure images of the sample;

[0076] Using a scanning electron microscope and an energy spectrum analyzer to analyze the microstructure images to obtain various chemical element surface scanning maps.

[0077] Exemplarily, the various chemical element surface scanning maps include but are not limited to: calcium element surface scanning map, silicon element surface scanning map, magnesium element surface scanning map, aluminum element surface scanning map, iron element surface scanning map, sulfur element surface scanning map, etc., and each surface scanning map only displays one chemical element.

[0078] Exemplarily, the microstructure images include but are not limited to backscattering images, CT images and all images that can display the microstructure characteristics of cement.

[0079] Exemplarily, after the cement material is made, a scanning electron microscope scanning process is uniformly carried out, and this process takes a model JSM-7610F electron microscope as an example. JSM-7610F, a high-resolution Schottky field emission scanning electron microscope, adopts a high-power optical illumination system, can perform high-temperature and high-precision element analysis, is also suitable for high spatial resolution element analysis, and observes the sample through a few hundred electron volts of extremely low acceleration voltage in a gentle beam mode. As an important tool for electron microscope micro-area composition analysis, the energy spectrum analyzer has always been the most commonly used accessory of scanning and transmission electron microscopes. Combined with the high speed and high sensitivity advantages of large area energy spectrum, and integrating the high-performance and easy-to-use Aztec EDS analysis platform, it is easy to analyze and detect particulate objects. The cement sample is polished and polished in turn, placed in the electron microscope carrier, focused and calibrated, backscattering electron images are collected, and element surface scanning images are obtained by combining EDS energy spectrum analyzer.

[0080] Further, S101: using the collected data, making a data set; specifically including:

[0081] Taking the pixel point i of the microstructure image as the center point, a sliding window with a side length of 2N+1 is designed, which is sequentially slid from left to right and from top to bottom, the two-dimensional image in the sliding window is intercepted and expanded into a vector, which is taken as the input vector corresponding to the current pixel point i, and the corresponding label is the gray value of the pixel point i in the element surface scanning image, the gray value represents the content of the element;

[0082] Similarly, the input vector and label corresponding to all pixel points of the microstructure image are obtained; the input vector corresponding to all pixel points of the microstructure image is taken as the data of the data set; i is a positive integer, and N is a positive integer;

[0083] According to the data of the data set and the label of the data set, the data set is obtained;

[0084] The data set is divided into a training set, a validation set and a test set according to a set proportion.

[0085] The principle of the above data set making method is based on the fact that cement has Markov random field and isotropy, therefore, the focus of the neural network is the local information of the cement microstructure image, and the global information is not concerned, therefore, in order to make the network focus more on the local information, the original data is processed as described above.

[0086] Exemplarily, a point a is randomly selected from the backscattering image, and a square with a side length of (2N+1) is intercepted with the point a as the center point, here N=12, and a total of 625 pixels in the intercepted area are covered. The pixel values from the top left to the bottom right of the intercepted area are recorded in sequence, and the value of the pixel point corresponding to the position a in the element surface scanning image is appended at the end of the data as a data (each data has (2N+1)*(2N+1)+1 values, here N=12, a total of 626).

[0087] According to the proportion of 0.7:0.05:0.25, the training set, the validation set and the test set are made, and the data set is collected for a specified number of times according to the need (this time a total of one million times, the training set, the test set and the validation set in the data set are divided according to the proportion of 0.7:0.05:0.25, and the number is 700000, 50000 and 250000 respectively), which are used for training, verification and final test respectively.

[0088] Further, as shown in Figure 2 S102: constructing a neural network and training the network with the data set; the neural network is a network model for predicting the element content of the pixel point level of cement, wherein the network structure of the neural network specifically includes:

[0089] the feature extraction module and the probability modeling module are connected in sequence;

[0090] the feature extraction module comprises first, second, third, fourth, fifth, sixth and seventh fully connected layers connected in sequence;

[0091] the probability modeling module comprises three fully connected branches in parallel, which are a weight output branch, a mean output branch and a variance output branch; each branch has only one fully connected layer and comprises M neurons; wherein M is a positive integer, and a Gaussian function is selected as a kernel function.

[0092] the weight output branch is the first fully connected layer in the three fully connected branches in parallel and is used for outputting the weights of all Gaussian functions; the mean output branch is the second fully connected layer in the three fully connected branches in parallel and is used for outputting the mean of each Gaussian function; and the variance output branch is the third fully connected layer in the three fully connected branches in parallel and is used for outputting the variance of each Gaussian function.

[0093] Exemplarily, the element content prediction neural network of the embodiment of the application is different from a common neural network for predicting purposes. First, the output of the element content prediction neural network is not a single category or a regression task value, but parameters of a probability density function of a to-be-predicted variable (here, sub-distribution mean, sub-distribution standard deviation and sub-distribution weight value) under the condition of given input features. The probability density function of the element content can be obtained by combining the parameters. The uncertainty and regular constraint of the cement element content prediction make the model have good adaptability. Second, the input of the network is not global image information but local image information after interception. Therefore, the output of the network is not traditional global prediction information but local prediction information. Further, the local image information is used to output the probability density function of the element content corresponding to the center pixel point of the local image.

[0094] The number of network input layer neurons is (2N+1)*(2N+1), and the first (2N+1)*(2N+1) items of each piece of data, i.e., the gray values of all pixel points in the backscattering image acquisition region, are read in. The label is the last item (the (2N+1)*(2N+1)+1 item) in each piece of data.

[0095] The feature extraction module comprises seven fully connected layers, and each fully connected layer comprises [1024, 2048, 2048, 4096, 2048, 1024, 512] neurons.

[0096] The probability modeling module includes three parallel full-connection branches, which are used to generate the weights, means and variances of all sub-distributions respectively.

[0097] The number of neurons in the three output branches is M, i.e. M distributions are superimposed, and each distribution has three parameters: weight, mean and standard deviation.

[0098] Here, M is set to 30, i.e. the probability density function of the element content at each pixel point is superimposed by 30 sub-distributions according to different weights. The first branch is used to generate the weights of the 30 distributions, the second branch is used to generate the means of the 30 distributions, and the third branch is used to generate the variances of the 30 distributions.

[0099] The activation function uses the Relu function, the distribution type used in the mixed distribution is selected as the Gaussian distribution, no cluster features are added in the data set, the network optimization algorithm is Adam, the learning rate is fixed at 0.001, and in order to avoid overfitting, the EarlyStopping strategy is set, and EarlyStopping is tentatively set to 250.

[0100] The reason for designing the probability modeling module according to the above structure is that the features of the cement microstructure image and the corresponding element surface scanning image are visualized, and it is found that it is not a traditional regression problem, because the independent variable features and dependent variable features of the regression problem will present a certain regular distribution, and it is easy to fit a linear function or a multiple function. However, the features of the cement microstructure image and the element surface scanning image do not conform to the above feature rules, but present a kind of local cluster state. When the cement microstructure image feature takes a certain value, the element content will be distributed in a certain interval range, and will be sporadically distributed outside the interval range. Therefore, the element content prediction is not a single value problem, and the traditional regression problem cannot solve the element content prediction problem. The above characteristics of the cement element content make the prediction interval a better choice, so the output layer is designed in the network structure, which first converts the traditional single value prediction of the cement element content into the prediction of the probability density function, i.e. the aggregation interval and the maximum possible value of the element content at each pixel point can be clearly seen. This approach is more in line with the current situation of cement element prediction and better meets the needs of the industry and academia. Especially for the industry, whether it is element content prediction or cement strength prediction, etc., single value prediction lacks a suitable confidence level, and the judgment mode of black and white or right and wrong is more arbitrary and prone to errors. However, by giving the probability density function of the element content, the probability of the element content in a certain interval range can be obtained, which can better guide workers to judge whether the element content meets the standard and has a better fault tolerance.

[0101] Further, S2: construct a neural network, and train the network with a data set, the training process comprising:

[0102] The input vector corresponding to each pixel point of the microstructure image is taken as the input value of the element content prediction neural network; the pixel value corresponding to the center pixel point of the microstructure image (i.e. the intercepted 2N+1*2N+1 square) in the chemical element area scan image is taken as the label of the neural network; the output value of the weight output branch of the element content prediction neural network is the weight of the M sub-probability density functions; the output value of the mean output branch of the element content prediction neural network is the mean of the M sub-probability density functions; the output value of the variance output branch of the element content prediction neural network is the variance of the M sub-probability density functions;

[0103] The jth mean and the jth variance are combined to generate the jth Gaussian function; similarly, M Gaussian functions are obtained;

[0104] The M Gaussian functions are weighted and summed according to the M weights to obtain the probability density function of the element content of each pixel point;

[0105] The element content prediction neural network is trained several times using the training set, and the accuracy of the model is continuously verified through the validation set during the training process; after the network model is optimized, the final element content prediction neural network is tested using the test set, and the tested neural network model is saved.

[0106] Further, the S103: collecting the microstructure image of the to-be-measured object, inputting the microstructure image of the to-be-measured object into the trained neural network to obtain a preliminary prediction result, specifically comprising:

[0107] The microstructure image of the to-be-measured cement or cement-based material is preprocessed to obtain an input vector corresponding to each pixel point of the microstructure image;

[0108] The input vector corresponding to the pixel point i of the preprocessed microstructure image is input into the trained element prediction neural network;

[0109] The trained element prediction neural network outputs the weights (α1, α2, …, αM) of the M probability density functions; M );

[0110] The trained element prediction neural network outputs the means (b1, b2, …, bM) of the M probability density functions; M );

[0111] The trained element prediction neural network outputs the variances (c1, c2, …, cM) of the M probability density functions; M ​

[0112] M probability density functions based on the mean (b1, b2, ···, b M ) and the variance (c1, c2, ···, c M ) of the M probability density functions; the probability density functions are Gaussian functions;

[0113] Then, the M probability density functions are weighted and summed based on the weight (a1, a2, ···, a M ) of the M probability density functions to obtain a weighted Gaussian probability density function of the chemical element content of the pixel point i:

[0114]

[0115] Wherein, G(i) represents the i-th Gaussian function; F(i) represents the probability density function of the chemical element content of the pixel point i, and M represents the number of sub-Gaussian distributions constituting the mixed distribution F(i).

[0116] Further, the probability density function also allows the probability density function of the Laplace distribution, the probability density function of the F distribution, the probability density function of the T distribution, the probability density function of the Log-Normal distribution, or the probability density function of the Chi-Square distribution, etc. Wherein, i takes all values in [1, w*h], w is the width of the to-be-detected picture, and h is the height of the to-be-detected picture.

[0117] Further, the S104: sampling the preliminary prediction result to obtain the prediction result of the element area scan corresponding to the to-be-detected object, further comprises:

[0118] The probability density function of the chemical element content of the pixel point i is sampled to obtain the chemical element content of the pixel point i, and further to obtain the chemical element content values of all pixel points. The chemical element content values of all pixel points are grouped into an image according to the positions of the pixel points to obtain the predicted element area scan image.

[0119] Further, the sampling is performed in the roulette mode or in the random sampling mode.

[0120] In summary, the cement backscatter image without the corresponding element area scan image is input into the model, the backscatter image is sequentially intercepted from the top left to the bottom right, and is input into the model for prediction to sequentially obtain the weight, mean and variance of 30 sub-distributions at each pixel point from the top left to the bottom right. The 30 sub-distributions at each pixel point are fused to obtain a mixed distribution. Roulette sampling is performed on each mixed distribution to obtain the element content value at each pixel point. The element content values at all pixel points form an image, which is the prediction result of the element area scan image corresponding to the input backscatter image.

[0121] Embodiment Two

[0122] The embodiment provides a cement and base material element content non-direct measurement system;

[0123] A cement and base material element content non-direct measurement system comprises the following steps:

[0124] A data set making module is configured to: collect microstructure images and element area scanning images of a detection object; the detection object is cement or cement base material; and use the collected data to make a data set;

[0125] A neural network construction module is configured to: construct a neural network, and train the network with the data set; the neural network is a network model for predicting element content at a pixel point or voxel point level of cement;

[0126] A prediction module is configured to: collect microstructure data of a to-be-detected object, input the microstructure data of the to-be-detected object into the trained neural network, and obtain a preliminary prediction result;

[0127] A sampling module is configured to: sample the preliminary prediction result, and obtain a prediction result of element area scanning or volume scanning corresponding to the to-be-detected object.

[0128] It should be noted that the data set making module, the neural network construction module, the prediction module and the sampling module correspond to steps S101 to S104 in Embodiment One, and the above modules and the corresponding steps have the same examples and application scenarios, but are not limited to the content disclosed in Embodiment One.

[0129] Embodiment Three

[0130] The embodiment provides a cement and base material element content non-direct measurement method;

[0131] As shown in Figure 3 A cement and base material element content non-direct measurement method comprises the following steps:

[0132] S301: Collect microstructure images and element area scanning images of a detection object; the detection object is cement or cement base material; and use the collected data to make a data set;

[0133] S302: Construct a microstructure three-dimensional voxel element content prediction neural network; the neural network is used to predict a probability density function of element content of each voxel of the microstructure three-dimensional body; and train the network with the data set;

[0134] S303: Pretreat the microstructure three-dimensional body of the cement or cement-based material to be tested, input the pretreated microstructure three-dimensional body into the trained neural network, and output the probability density function of the chemical element content at each voxel point.

[0135] Further, the method further comprises:

[0136] S304: Sampling the probability density function of the element content at each voxel point obtained to obtain the element content with the maximum probability at the current voxel point.

[0137] Further, the sample refers to cement samples or cement-based material samples with different proportions; wherein the preparation process of the sample comprises:

[0138] Different proportions of cement samples or cement-based material samples are made, and the hydration time of the samples is different;

[0139] According to different proportions and different hydration times, the samples are divided into different groups.

[0140] Illustratively, in order to ensure the accuracy and robustness of the model in the later stage, according to the theoretical requirements, the cement data set should be comprehensive and balanced. For example, cement and cement-based materials, as many types of cement material data sets as possible are made. At least cover the commonly used cement types today, such as Portland cement, ordinary Portland cement, slag Portland cement, pozzolanic Portland cement, coal powder Portland cement, composite Portland cement and part of aluminate cement, etc. Therefore, different formula cement samples are made, and multiple samples are made for each formula. The samples in the same formula have different hydration times.

[0141] After completing the preparation of the various control experimental group cements and their base materials, in order to ensure that the subsequent image sampling process is not disturbed by external factors, the same standing, curing, and packaging operations are performed. Once completed, the samples are not unpacked until the instrument obtains the image data set.

[0142] Further, the S301: collecting the microstructure image and element surface scanning image of the detection object, specifically comprising:

[0143] The microstructure image of the sample is collected using a scanning electron microscope;

[0144] The scanning electron microscope and the energy spectrum analyzer are used to analyze the microstructure image to obtain various chemical element surface scanning images.

[0145] Illustratively, the various chemical element surface scanning images include but are not limited to: calcium element surface scanning image, silicon element surface scanning image, magnesium element surface scanning image, aluminum element surface scanning image, iron element surface scanning image, sulfur element surface scanning image, etc. One surface scanning image only shows one chemical element.

[0146] Exemplarily, the microstructure image includes but is not limited to a backscattering image, a CT image and the like.

[0147] Exemplarily, after the cement material is made, a scanning electron microscope scanning process is uniformly performed. This process takes a JSM-7610F electron microscope as an example. The JSM-7610F is a high-resolution Schottky field emission scanning electron microscope. A high-power optical illumination system can be used for high-temperature and high-precision element analysis, and is also suitable for high spatial resolution element analysis. The sample is observed through a few hundred electron volts of extremely low acceleration voltage in a gentle beam mode. The energy spectrometer is an important tool for micro-area component analysis of the electron microscope. It is the most commonly used accessory of the scanning and transmission electron microscope. Combined with the high speed and high sensitivity advantages of large area energy spectrum, and the integration of the high-performance and easy-to-use Aztec EDS analysis platform, it is easy to analyze and detect particulate objects. The cement sample is sequentially polished and polished, placed in the electron microscope carrier, focused and calibrated, and the backscattering electron image is collected. At the same time, the corresponding element surface scanning image is obtained by combining the EDS energy spectrum analyzer.

[0148] Further, S301: using the collected data, making a data set, specifically including:

[0149] The microstructure images are stacked in sequence, that is, a plurality of two-dimensional images are stacked into a microstructure three-dimensional body. Similarly, the corresponding element surface scanning images are stacked in the same sequence, that is, an element surface scanning three-dimensional body is formed;

[0150] Taking a voxel point i of the microstructure three-dimensional body as a center point, a cube with a side length of 2N+1 is intercepted, and the cube is taken as the input of the model. The corresponding label is the gray value of the voxel point i in the element surface scanning three-dimensional body. The gray value represents the content of the element;

[0151] Similarly, the three-dimensional body corresponding to each voxel point and the corresponding label are obtained to form a data set.

[0152] The data set is divided into a training set, a validation set and a test set according to a set proportion.

[0153] The principle of the above data set making method is based on the fact that the cement has a Markov random field and isotropy. Therefore, the focus of the neural network is the local information of the cement microstructure three-dimensional body, and the global information is not concerned. Therefore, in order to make the network focus more on the local information, the data set is processed as described above.

[0154] Exemplarily, a point a is randomly selected from the microstructure three-dimensional body, and a cube with a side length of (2N+1) is intercepted with the point a as the center point. Here, N=12 is taken as a training data.

[0155] The training set, the verification set and the test set are made in a ratio of 0.7:0.05:0.25, and the data set is collected for a specified number of times as required (in this case, one million times, and the training set, the test set and the verification set are divided in a ratio of 0.7:0.05:0.25, and the number of each is 700000, 50000 and 250000 respectively), and are used for training, verification and final testing respectively.

[0156] Further, as shown in S302, a microstructure three-dimensional body voxel element content prediction neural network is constructed, and the neural network is used to predict the probability density function of the element content of each voxel of the microstructure three-dimensional body; wherein the network structure of the neural network specifically comprises: Figure 3 a feature extraction module and a probability modeling module connected in sequence;

[0157] The feature extraction module comprises a first three-dimensional convolutional layer, a second three-dimensional convolutional layer, a third three-dimensional convolutional layer, a fourth three-dimensional convolutional layer and a first full connection layer connected in sequence.

[0158] The probability modeling module comprises three full connection branches in parallel, which are a weight output branch, a mean output branch and a variance output branch respectively; each branch has only one full connection layer, including M neurons; wherein M is a positive integer, and a Gaussian function is selected as the kernel function here.

[0159] The weight output branch is the first full connection layer in the parallel full connection layers, which is used to output the weights of all Gaussian functions, the mean output branch is the second full connection layer in the parallel full connection layers, which is used to output the mean of each Gaussian function, and the variance output branch is the third full connection layer in the parallel full connection layers, which is used to output the variance of each Gaussian function.

[0160] Exemplarily, the element content prediction neural network of the embodiment of the present application is different from the ordinary neural network for prediction purposes. Firstly, the output is not a single category or a regression task value, but the parameters of the probability density function of the prediction result (here, the sub-distribution mean, the sub-distribution standard deviation and the sub-distribution weight value) under the condition of given some input features, and the probability density function of the element content can be obtained by combining the parameters, and the uncertainty and regular constraint of the cement element content prediction make the model have good adaptability; secondly, the input of the network is not the global information of the three-dimensional body, but the local three-dimensional body information after interception, so the output of the network is not the traditional global prediction information, but the local prediction information, and further, the probability density function of the element content corresponding to the center pixel point of the local three-dimensional body is output according to the local three-dimensional body information.

[0161]

[0162] ​Each convolutional layer is composed of a three-dimensional convolution kernel, each convolution kernel has a size of 3*3*3, and the convolution step is 2; the number of neurons in the first fully connected layer is 64.

[0163] The feature extraction module includes three parallel fully connected branches, and the three branches are used to generate the weights, mean and variance of all sub-distributions.

[0164] The number of neurons in the three parallel fully connected branches is M, that is, M distributions are superimposed, and each distribution has three parameters: weight, mean and standard deviation.

[0165] Here, M is set to 30, that is, the probability density function of the element content of each voxel point is superimposed by 30 sub-distributions according to different weights.

[0166] The first branch is used to generate the weights of the 30 distributions, the second branch is used to generate the mean of the 30 distributions, and the third branch is used to generate the variance of the 30 distributions.

[0167] The activation function uses the Relu function, the distribution type used in the mixed distribution selects the Gaussian distribution, no cluster features are added in the data set, the network optimization algorithm is Adam, the learning rate is fixed at 0.001, and in order to avoid overfitting, the EarlyStopping strategy is set, and EarlyStopping is temporarily set to 250.

[0168] The reason for using the above network structure is the same as in Example 1.

[0169] Further, S302: training the network with the data set, the training process includes:

[0170] Each microstructure three-dimensional body intercepted is taken as the input value of the element content prediction neural network; the gray value of the element surface scanning three-dimensional body corresponding to the center voxel is taken as the label of the neural network; the output value of the weight output branch of the element content prediction neural network is the weight of the M sub-probability density functions; the output value of the mean output branch of the element content prediction neural network is the mean of the M sub-probability density functions; the output value of the variance output branch of the element content prediction neural network is the variance of the M sub-probability density functions;

[0171] The element content prediction neural network is trained several times using the training set, and the accuracy of the model is continuously verified through the validation set during the training process; after the network model optimization is completed, the final element content prediction neural network is tested using the test set, and the neural network model that passes the test is saved.

[0172] Further, step S303 involves: preprocessing the three-dimensional microstructure of the cement or cement-based material to be tested, inputting the preprocessed three-dimensional microstructure into a trained neural network, and outputting the probability density function of the chemical element content at each voxel point; specifically including:

[0173] The microstructure of the cement or cement-based material to be tested is preprocessed to obtain the local three-dimensional volume corresponding to each voxel point of the microstructure.

[0174] The three-dimensional microstructure corresponding to voxel point i of the preprocessed microstructure is input into the trained element prediction neural network.

[0175] The trained element prediction neural network outputs the weights (α1, α2, ..., α) of M probability density functions. M );

[0176] The trained element prediction neural network outputs the mean (b1, b2, ..., b...) of M probability density functions. M );

[0177] The variances (c1, c2, ..., c) of the M probability density functions output by the trained element prediction neural network. M );

[0178] Based on the mean (b1, b2, ..., b) of M probability density functions M ) and the variances (c1, c2, ..., c) of the M probability density functions M M probability density functions are obtained; the probability density functions adopt the probability density function of Gaussian distribution;

[0179] Then, based on the weights (α1, α2, ..., α) of the M probability density functions. M By taking a weighted sum of the M probability density functions, we obtain the weighted Gaussian probability density function of the chemical element content at voxel point i:

[0180]

[0181] Where G(i) represents the i-th Gaussian function, F(i) is the probability density function of the element content at the i-th voxel, and M represents the number of sub-distributions that make up the mixed distribution.

[0182] Furthermore, the probability density function can also be selected from the probability density functions of the Laplace distribution, the F distribution, the T distribution, the Log-Normal distribution, or the Chi-Square distribution, etc.

[0183] Further, the S304: sampling the probability density function of the obtained element content at each voxel point to obtain the element content with the maximum probability at the current voxel point, comprising:

[0184] Sampling the probability density function of the chemical element content of the voxel point i to obtain the chemical element content of the voxel point i; and further obtaining the chemical element content values of all voxel points, and grouping the chemical element content values of all voxel points according to the positions of the voxel points to obtain a predicted element spatial distribution map.

[0185] Further, the sampling is performed in the roulette mode or the random sampling mode.

[0186] In summary, the cement microstructure three-dimensional body without the corresponding element surface scanning three-dimensional body is input into the model, the three-dimensional body is sequentially intercepted and input into the model for prediction, and the weight, mean value and variance of 30 sub-distributions at each voxel point are sequentially obtained. The 30 sub-distributions at each voxel point are fused to obtain a mixed distribution. Roulette sampling is performed on each mixed distribution to obtain the element content value at each voxel point. The element content values at all voxel points form a three-dimensional body, which is the prediction result of the input cement microstructure three-dimensional body corresponding to the element surface scanning three-dimensional body.

[0187] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for indirect measurement of elemental content in cement and its base materials, characterized in that, include: Acquire microstructure images and elemental surface scan images of the object being tested; The object being tested is cement or cement-based materials; Use the collected data to create a dataset; A neural network is constructed and trained using a dataset; the neural network is a network model that predicts the elemental content of cement at the pixel or voxel level. Collect microstructure data of the object under test, input the microstructure data of the object under test into the trained neural network, and obtain preliminary prediction results; The preliminary prediction results are sampled to obtain the prediction results of the element surface scan or volume scan corresponding to the object to be measured; Constructing a neural network and training it with a dataset includes: Construct a neural network, the neural network comprising a feature extraction module and a probability modeling module connected in sequence; After the neural network is built, a dataset is used, the log-likelihood loss function is selected, and backpropagation is used to train the neural network. The feature extraction module refers to a neural network that extracts features from the input cement data. The feature extraction module is a fully connected network, which can be either a two-dimensional convolutional neural network or a three-dimensional convolutional neural network. For the two-dimensional convolutional neural network or the three-dimensional convolutional neural network, the feature map obtained is finally unfolded into a one-dimensional vector. The probabilistic modeling module refers to an output layer containing multiple parallel fully connected branches, with each fully connected branch having only one fully connected layer. The specific construction steps for the output layer of the multiple parallel fully connected branches include: Select a kernel function; the kernel function includes, but is not limited to, Gaussian function, Laplace function, etc., different kernel functions include different numbers of parameters, the number of parameters of each kernel function is denoted as N, and the number of fully connected branches is set as N+1; the selection principle of the kernel function is based on the form of the probability distribution followed by the target problem in the problem to be solved; The number of kernel functions is set to M, and the number of neurons in each fully connected branch is also M; the output of the first fully connected branch is the weights of the M kernel functions ( , , ···, The output of the second fully connected branch is the parameters of all M kernel functions. , , ···, ), and so on, the output of the (N+1)th fully connected branch is the parameters of all M kernel functions ( , , ···, ).

2. The method for non-direct measurement of elemental content in cement and its base materials as described in claim 1, characterized in that, Acquire microscopic structural images and elemental surface scan images of the object being tested, including: To obtain images of the cement microstructure, a scanning electron microscope combined with an energy dispersive spectroscopy (EDS) instrument was used to obtain surface scan images of cement elements at the same locations. Among these methods, images of the cement microstructure are obtained using scanning electron microscopes, computed tomography (CT) scans, or any instrument capable of acquiring images of the cement microstructure.

3. The method for non-direct measurement of elemental content in cement and its base materials as described in claim 1, characterized in that, The process of creating a dataset using the collected data includes: Based on the principles of Markov random fields and the isotropic properties of cement microstructure, a sliding window of a specified size is designed. The first point in the upper left corner of the cement microstructure data is taken as the initial center point of the sliding window. The window slides sequentially from left to right and from top to bottom, capturing two-dimensional images or three-dimensional volumes within the sliding window as the input data corresponding to the first point. The label corresponding to the first point is the gray value of the position corresponding to the first pixel in the element surface scan image or three-dimensional volume. Following the same method, traverse all points of the cement microstructure data to obtain the input data and labels for all points, thus obtaining the dataset.

4. The method for non-direct measurement of elemental content in cement and its base materials as described in claim 1, characterized in that, The data collected is used to create a dataset. For different problem scenarios, the collected data has different meanings. For two-dimensional element content prediction, the collected data refers to the original microstructure image and the original element surface scan image. For three-dimensional element content prediction, the collected data refers to the three-dimensional microstructure and the three-dimensional scan. The three-dimensional microstructure refers to the process of sequentially superimposing acquired two-dimensional cement microstructure images on a channel according to their layer order to obtain a three-dimensional microstructure containing the internal microstructure of cement; and superimposing corresponding element surface scan images according to the same superposition layer order of the two-dimensional cement microstructure images to obtain a three-dimensional cement scan volume.

5. The method for non-direct measurement of elemental content in cement and its base materials as described in claim 1, characterized in that, Microstructural data of the object under test is collected and input into a trained neural network to obtain preliminary prediction results, including: Collect microstructure data of the object under test, design a sliding window of a specified size, take the first point in the upper left corner of the cement microstructure data as the initial center point of the sliding window, slide it sequentially from left to right and from top to bottom, and capture the two-dimensional image or three-dimensional volume within the sliding window as the input data corresponding to the first point. Similarly, obtain the input data corresponding to each point of the microstructure image of the object under test. The input data corresponding to each point of the microstructure data of the object under test is fed into the trained neural network to obtain the prediction result; the prediction result is the total parameters of all kernel functions at each point. Take a weighted sum of all M kernel functions at each point: Among them, the Kernel functions It is determined by the corresponding parameters , , ···, The kernel function is composed of; Indicates the first The first parameter of the kernel function; Indicates the first The second parameter of the kernel function; Indicates the first The first kernel function N One parameter; The first fully connected branch obtained by finding the first fully connected branch The weights of each kernel function; After weighted summation, the mixture function F at each point of the input data is obtained. The mixture function F represents the probability density function of the element content at the current point.

6. The method for indirect measurement of elemental content in cement and its base materials as described in claim 1, characterized in that, The preliminary prediction results are sampled to obtain the prediction results of the elemental surface scan or volume scan corresponding to the object to be measured, including: By sampling the probability density function at each point, the maximum probability value of the element content at the current point is obtained; The maximum probability values ​​of the element content at each point are concatenated into a matrix according to their corresponding positions to obtain the prediction results of the element surface scan or 3D scan volume corresponding to the object under test.

7. A system for indirect measurement of elemental content in cement and its base materials, characterized in that, include: The dataset creation module is configured to: acquire microscopic structural images and elemental surface scan images of the detected object; The object being tested is cement or cement-based materials; the collected data is used to create a dataset. The neural network building module is configured to: build a neural network and train the network with a dataset; the neural network is a network model for predicting the elemental content at the pixel or voxel level of cement. The prediction module is configured to: collect microstructure data of the object under test, input the microstructure data of the object under test into the trained neural network, and obtain preliminary prediction results; The sampling module is configured to sample the preliminary prediction results to obtain the prediction results of the element surface scan or volume scan corresponding to the object to be measured. Constructing a neural network and training it with a dataset includes: Construct a neural network, the neural network comprising a feature extraction module and a probability modeling module connected in sequence; After the neural network is built, a dataset is used, the log-likelihood loss function is selected, and backpropagation is used to train the neural network. The feature extraction module refers to a neural network that extracts features from the input cement data. The feature extraction module is a fully connected network, which can be either a two-dimensional convolutional neural network or a three-dimensional convolutional neural network. For the two-dimensional convolutional neural network or the three-dimensional convolutional neural network, the feature map obtained is finally unfolded into a one-dimensional vector. The probabilistic modeling module refers to an output layer containing multiple parallel fully connected branches, with each fully connected branch having only one fully connected layer. The specific construction steps for the output layer of the multiple parallel fully connected branches include: Select a kernel function; the kernel function includes, but is not limited to, Gaussian function, Laplace function, etc., different kernel functions include different numbers of parameters, the number of parameters of each kernel function is denoted as N, and the number of fully connected branches is set as N+1; the selection principle of the kernel function is based on the form of the probability distribution followed by the target problem in the problem to be solved; The number of kernel functions is set to M, and the number of neurons in each fully connected branch is also M; the output of the first fully connected branch is the weights of the M kernel functions ( , , ···, The output of the second fully connected branch is the parameters of all M kernel functions. , , ···, ), and so on, the output of the (N+1)th fully connected branch is the parameters of all M kernel functions ( , , ···, ).

8. A method for indirect measurement of elemental content in cement and its base materials, characterized in that, include: Acquire microstructure images and elemental surface scan images of the object being tested; The object being tested is cement or cement-based materials; Use the collected data to create a dataset; A neural network for predicting the element content of voxels in a three-dimensional volumetric microstructure is constructed. The neural network is used to predict the probability density function of the element content of each voxel in the three-dimensional volumetric microstructure. The network is trained using a dataset. The three-dimensional microstructure of the cement or cement-based material to be tested is preprocessed, and the preprocessed three-dimensional microstructure is input into the trained neural network, which outputs the probability density function of the chemical element content at each voxel point. Constructing a neural network and training it with a dataset includes: Construct a neural network, the neural network comprising a feature extraction module and a probability modeling module connected in sequence; After the neural network is built, a dataset is used, the log-likelihood loss function is selected, and backpropagation is used to train the neural network. The feature extraction module refers to a neural network that extracts features from the input cement data. The feature extraction module is a fully connected network, which can be either a two-dimensional convolutional neural network or a three-dimensional convolutional neural network. For the two-dimensional convolutional neural network or the three-dimensional convolutional neural network, the feature map obtained is finally unfolded into a one-dimensional vector. The probabilistic modeling module refers to an output layer containing multiple parallel fully connected branches, with each fully connected branch having only one fully connected layer. The specific construction steps for the output layer of the multiple parallel fully connected branches include: Select a kernel function; the kernel function includes, but is not limited to, Gaussian function, Laplace function, etc., different kernel functions include different numbers of parameters, the number of parameters of each kernel function is denoted as N, and the number of fully connected branches is set as N+1; the selection principle of the kernel function is based on the form of the probability distribution followed by the target problem in the problem to be solved; The number of kernel functions is set to M, and the number of neurons in each fully connected branch is also M; the output of the first fully connected branch is the weights of the M kernel functions ( , , ···, The output of the second fully connected branch is the parameters of all M kernel functions. , , ···, ), and so on, the output of the (N+1)th fully connected branch is the parameters of all M kernel functions ( , , ···, ).

Images