Fractured man-made rock printing method based on optimization algorithm

Abstract

The invention discloses a crack-containing man-made rock printing method based on an optimization algorithm, and belongs to the technical field of building construction, and the crack-containing man-made rock printing method comprises the following steps: S1, collecting a reservoir crack logging data set, and generating a cavity topological structure by using a generator module and a discriminator module; s2, the printing path of the cavity topological structure is adjusted, the concrete proportion of the cavity topological structure is optimized, and the optimal concrete proportion is obtained; and S3, according to the optimal concrete proportion of the cavity topological structure and the adjusted printing path, man-made rock printing is completed, and joint repairing is conducted on the crack position. The concrete proportion is dynamically regulated and controlled by combining the crack fluid resistivity contrast model, so that the inner wall and the outer wall of the cavity are continuously formed, and the problem of leakage or cracking caused by uneven material accumulation is avoided.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of building construction, and particularly relates to a printing method for artificial rock containing cracks based on an optimization algorithm. BACKGROUND

[0002] The structure of artificial rock containing cracks is widely used in the fields of geological engineering and underground engineering, but its traditional preparation method has long faced technical bottlenecks. The existing process relies on mold pouring or mechanical cutting, and there are three major problems: first, the crack morphology replication precision is low, and the irregular topological structure of natural cracks is difficult to accurately reproduce through a mold; second, the material performance is difficult to control, and the mechanical properties (such as compressive strength and elastic modulus) of the rock matrix and the crack filling material are difficult to match, and stress concentration easily occurs at the interface, which reduces the overall bearing capacity of the structure; third, the contradiction between permeability and stability is prominent, and the porous structure designed to simulate the permeability of natural cracks often leads to weakening of the rock matrix deformation resistance, which easily causes sudden damage under hydraulic fracturing and other working conditions.

[0003] The development of 3D printing technology provides new possibilities for the preparation of complex structures, but there are still significant limitations in the field of artificial rock containing cracks. The path planning of the existing printing method does not consider the influence of crack distribution on material accumulation, and the printing layer is prone to slip or void; the printing material ratio uses a fixed formula, which cannot be dynamically adjusted according to the crack position (such as higher toughness material at the crack tip), resulting in secondary cracks in the printed part during the curing stage; quality detection relies on post-test physical tests (such as permeability test and mechanical test), and lacks real-time monitoring of the crack-matrix interface state, which cannot correct defects in the printing process.

[0004] Deep learning algorithms have achieved breakthroughs in reservoir crack prediction, and through the use of generative adversarial networks, crack topological structures that conform to mechanical laws can be generated. However, the printing of artificial rock containing cracks still relies on traditional CAD modeling, and neural networks are not used to intelligently match crack morphology and printing parameters, resulting in a design cycle of several weeks and difficulty in dealing with complex crack combination scenarios. Therefore, an integrated solution that combines optimization algorithms and 3D printing is urgently needed to solve the core problems of artificial rock containing cracks in terms of morphology replication, performance control, and quality assurance. SUMMARY

[0005] The application is proposed to solve the above problems, and a printing method for artificial rock containing cracks based on an optimization algorithm is provided.

[0006] The technical solution of the application is as follows: a printing method for artificial rock containing cracks based on an optimization algorithm includes the following steps:

[0007] S1, collecting a data set of reservoir crack logging, and generating a cavity topological structure using a generator module and a discriminator module;

[0008] S2, adjust the printing path of the cavity topology and optimize the concrete proportion of the cavity topology to obtain an optimal concrete proportion;

[0009] S3, complete the printing of the artificial rock according to the optimal concrete proportion of the cavity topology and the adjusted printing path, and repair the node position of the crack.

[0010] Further, S1 includes the following sub-steps:

[0011] S11, collect the crack width, resistivity response and mechanical property data of the reservoir crack logging as a data set;

[0012] S12, construct a crack spacing constraint condition for the data set, and encode the crack spacing constraint condition into a condition vector;

[0013] S13, input the data set and the condition vector into the generator module to generate a topological image;

[0014] S14, input the topological image into the discriminator module to generate a mechanical score and a permeability score;

[0015] S15, according to the mechanical score and the permeability score, the generator module and the discriminator module are optimized by using the adversarial training optimization module to generate a cavity topology.

[0016] Further, in S12, the crack spacing constraint condition is encoded into a condition vector, specifically: when the crack spacing is less than or equal to a specific threshold of 5mm, the specific threshold is encoded as 0, otherwise the specific threshold is encoded as 1.

[0017] Further, in S13, the generator module includes a first fully connected layer, a second fully connected layer, a third fully connected layer and a plurality of deconvolution layers connected in sequence;

[0018] The first fully connected layer, the second fully connected layer and the third fully connected layer are used to map the data set and the condition vector to a high-dimensional latent space;

[0019] The deconvolution layer is used to upsample the high-dimensional latent space to generate a topological image;

[0020] The expression of the deconvolution layer is:

[0021] ;

[0022] In the formula, represents the output data of the layer deconvolution layer, represents the input data of the layer deconvolution layer, represents the condition vector, denotes a concatenation function.

[0023] Further, in S14, the discriminator module comprises a feature extraction layer, a correlation layer and a global pooling layer connected in sequence;

[0024] The feature extraction layer is configured to extract multi-scale structural features of the topological image to generate a first feature map and a second feature map;

[0025] The correlation layer is configured to generate local features according to the first feature map and the second feature map;

[0026] The global pooling layer is configured to obtain a mechanical score and a permeability score according to the local features;

[0027] The expression of the correlation layer is:

[0028] ;

[0029] In the formula, denotes the output of the correlation layer, denotes the first feature map, denotes the second feature map, denotes the first attention weight matrix, denotes the second attention weight matrix, denotes the third attention weight matrix, denotes the dimension, denotes an activation function, denotes a transposition operation, denotes a conditional vector.

[0030] Further, in S15, the maximum deformation is calculated according to the mechanical score, the Darcy permeability is calculated according to the permeability score, and the generator module and the discriminator module are optimized by using the maximum deformation and the Darcy permeability, and the cavity topological structure is generated after optimization.

[0031] Further, S2 comprises the following sub-steps:

[0032] S21, optimizing the printing path by using a graph convolution network; wherein the graph convolution network comprises a first graph convolution layer, a second graph convolution layer and a third graph convolution layer connected in sequence;

[0033] S22, adjusting the concrete proportion of the cavity topological structure by using a Transformer model based on the optimized printing path to obtain the best concrete proportion.

[0034] Further, in S22, the Transformer model comprises an encoder and a decoder;

[0035] The encoder comprises a first input layer, a first adder, an environmental factor embedding layer, a spatial-frequency mixed convolution layer, a first multi-head attention layer, an instance normalization layer, a first feedforward network layer, and a layer normalization unit;

[0036] An input end of the first input layer is connected with an output end of the encoder, and an input end of the first adder is connected with an output end of the first input layer; a first output end of the first adder, an input end of the environmental factor embedding layer, and an input end of the spatial-frequency mixed convolution layer are sequentially connected; a first output end, a second output end, and a third output end of the spatial-frequency mixed convolution layer are connected with an input end of the first multi-head attention layer; an output end of the first multi-head attention layer is connected with a first input end of the instance normalization layer; a second output end of the first adder is connected with a second input end of the instance normalization layer; a first output end of the instance normalization layer, the first feedforward network layer, and a first input end of the layer normalization unit are connected; a second output end of the instance normalization layer is connected with a second input end of the layer normalization unit; and an output end of the layer normalization unit is connected with an output end of the encoder.

[0037] An expression of the spatial-frequency mixed convolution layer is as follows:

[0038] ;

[0039] In the formula, x represents an input of the spatial-frequency mixed convolution layer, y represents an output of the spatial-frequency mixed convolution layer, x represents an output of the environmental factor embedding layer, represents a spatial convolution operation, represents a first convolution kernel of the spatial-frequency mixed convolution layer, represents a number of convolution kernels of the spatial-frequency mixed convolution layer, represents a bias term, represents a spatial convolution operation, represents a number of convolution kernels of the spatial-frequency mixed convolution layer.

[0040] Further, the decoder comprises a second input layer, a second adder, a second multi-head attention layer, a first accelerated normalization layer, a third multi-head attention layer, a second accelerated normalization layer, a second feedforward network layer, and an adaptive normalization layer.

[0041] The input end of the second input layer is an input end of the decoder, the output end thereof is connected with the input end of the second adder; the first output end, the second output end and the third output end of the second adder are all connected with the input end of the second multi-head attention layer; the output end of the second multi-head attention layer is connected with the first input end of the first accelerated normalization layer; the fourth output end of the second adder is connected with the second input end of the first accelerated normalization layer; the first output end of the first accelerated normalization layer is connected with the first input end of the third multi-head attention layer; the output end of the third multi-head attention layer is connected with the first input end of the second accelerated normalization layer; the second output end of the first accelerated normalization layer is connected with the second input end of the second accelerated normalization layer; the third output end of the first accelerated normalization layer is connected with the third input end of the second accelerated normalization layer; the output end of the encoder is connected with the second input end of the third multi-head attention layer; the first output end of the second accelerated normalization layer, the second feedforward network layer and the first input end of the adaptive normalization layer are sequentially connected; the second output end of the second accelerated normalization layer is connected with the second input end of the adaptive normalization layer; the output end of the adaptive normalization layer is an output end of the decoder;

[0042] Loss function of the decoder The expression is:

[0043] ;

[0044] In the formula, represents the actual monitoring concrete density measurement value, represents the total number of samples in the training batch, represents the model predicted concrete density value, represents the density change reference threshold defined by the construction specification, represents the first dynamic weight coefficient, represents the second dynamic weight coefficient, represents a small constant to prevent division by zero error, represents an exponential function.

[0045] Further, S3 includes the following sub-steps:

[0046] S31, based on the adjusted printing path and the optimal concrete proportion, collecting temperature data, strain data and conductivity data of the cavity topological structure by using a sensor;

[0047] S32, pre-processing the temperature data, the strain data and the conductivity data, specifically: eliminating the noise of the temperature data by using Kalman filtering method, extracting the main frequency component of the strain data by using fast Fourier transform, and calibrating the conductivity data according to the crack fluid resistivity contrast;

[0048] S33, convert the pre-processed temperature data, strain data and conductivity data into a state vector;

[0049] S34, generate an action vector according to the state vector, print the cavity topology, the action vector including a printing speed adjustment amount, a layer thickness adjustment amount and a carbon fiber slurry injection ratio;

[0050] S35, after printing, use an ultrasonic flaw detector to receive an original signal, and perform denoising processing on the original signal;

[0051] S36, based on the denoised original signal, generate a crack position using a 3D convolutional neural network, and repair the crack position;

[0052] In S33, the calculation formula for denoising the original signal is:

[0053] ;

[0054] In the formula, denotes the denoised original signal, denotes the original signal, denotes the setting time, denotes the reference setting period, denotes the real-time frequency component of the signal, denotes the structure reference frequency specified by the design.

[0055] The beneficial effects of the present application are:

[0056] (1) The present application realizes a systematic breakthrough in cavity design, construction and quality control by fusing reservoir fracture logging numerical simulation technology and deep learning algorithm; based on the nonlinear law of fracture width and spacing in reservoir fracture logging on resistivity response, a conditional generative adversarial network is used to generate a biomimetic cavity topology, the cavity rate is dynamically adjustable, and at the same time, a graph convolution network is used to optimize the printing path, combined with a fracture fluid resistivity contrast model to dynamically control the concrete ratio, so that the inner and outer walls of the cavity are continuously formed, avoiding leakage or cracking problems caused by uneven material accumulation;

[0057] (2) The present application uses a 3D printed interlayer embedded scale test well design fiber sensor array to monitor the conductivity and temperature distribution in real time, combined with wavelet transform-3D convolutional neural network and resistivity imaging technology, the defect recognition accuracy is 97%, and the positioning error is less than 2mm;

[0058] (3) When detecting defects, the present application uses gradient material printing technology for local reinforcement, so that the cavity node compressive strength is greatly improved; through an optimization algorithm, a cavity shape with high degree of freedom and reasonable mechanics is generated, ensuring the unity of shape and function. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A flow chart of a method for printing a fractured artificial rock based on an optimization algorithm;

[0060] Figure 2 A structural diagram of a Transformer model;

[0061] Figure 3 A structural diagram of a multi-head self-attention mechanism;

[0062] Figure 4 A structural diagram of a DRL adaptive printing path planning system architecture. DETAILED DESCRIPTION

[0063] The embodiments of the present application will be further described below with reference to the accompanying drawings.

[0064] As shown in Figure 1 , the present application provides a method for printing a fractured artificial rock based on an optimization algorithm, comprising the following steps:

[0065] S1, collecting a data set of reservoir fracture logging, and generating a cavity topology structure using a generator module and a discriminator module;

[0066] S2, adjusting the printing path of the cavity topology structure, and optimizing the concrete proportion of the cavity topology structure to obtain the best concrete proportion;

[0067] S3, according to the best concrete proportion of the cavity topology structure and the adjusted printing path, completing the printing of the artificial rock, and repairing the node position of the fracture.

[0068] In the embodiments of the present application, S1 comprises the following sub-steps:

[0069] S11, collecting the fracture width, resistivity response and mechanical property data of reservoir fracture logging as a data set;

[0070] S12, constructing a fracture spacing constraint condition for the data set, and encoding the fracture spacing constraint condition into a condition vector;

[0071] S13, inputting the data set and the condition vector into the generator module to generate a topology image;

[0072] S14, inputting the topology image into the discriminator module to generate a mechanical score and a permeability score;

[0073] S15, according to the mechanical score and the permeability score, using an adversarial training optimization module to optimize the generator module and the discriminator module to generate a cavity topology structure.

[0074] First, a "fracture parameter-mechanical property" dataset was constructed based on numerical simulations of reservoir fracture logging. This dataset encompasses 100,000 data points related to fracture width, resistivity response, and their corresponding mechanical properties (stiffness, permeability). Statistical analysis of the fracture parameters revealed that when the fracture spacing is less than a specific threshold (5 mm), the electrical imaging signal exhibits a single response characteristic due to current density merging. This threshold was converted into a binary condition vector, encoded as 0 when the input spacing d ≤ 5 mm and 1 otherwise, thus constraining the generator to avoid signal merging regions in the topology. After normalization to eliminate dimensional differences, the fracture spacing constraint was encoded into a condition vector, which, along with the design parameters, was input into the generator module.

[0075] In this embodiment of the invention, in S12, encoding the crack spacing constraint condition into a condition vector specifically means: when the crack spacing is less than or equal to a specific threshold of 5mm, the specific threshold is encoded as 0; otherwise, the specific threshold is encoded as 1.

[0076] In this embodiment of the invention, in S13, the generator module includes a first fully connected layer, a second fully connected layer, a third fully connected layer and several deconvolutional layers connected in sequence.

[0077] The first fully connected layer, the second fully connected layer, and the third fully connected layer are used to map the dataset and conditional vectors to a high-dimensional latent space;

[0078] Deconvolutional layers are used to upsample the high-dimensional latent space to generate a topological image. In the convolutional layers of this invention, the convolutional kernel is 3×3 with a stride of 2. This setting progressively upsamples the data output from the fully connected layers to restore spatial resolution. Furthermore, conditional vectors are injected into each deconvolutional layer through a concatenation operation.

[0079] The expression for a deconvolutional layer is:

[0080] ;

[0081] In the formula, Indicates the first The output data of the deconvolution layer Indicates the first Input data for each deconvolution layer Represents a condition vector. This represents the concatenation function.

[0082] The generator module employs a hybrid architecture of fully connected neural networks and deconvolutional layers. In the fully connected layer section, the input vector is... Its dimension is d in The weight matrix of the l-th fully connected layer is The bias vector is Then the first Output of the layer fully connected layer The calculation formula is:

[0083] ;

[0084] In the formula, represents an activation function, and in the present application, a ReLU function can be selected, that is, In the generator of the present application, three fully connected layers are included, and the number of layer neurons is 512→256→128, which means that from the input layer to the first fully connected layer, The dimension of is 512xd in , The dimension of is 512; from the first layer to the second layer, The dimension of is 256x512, The dimension of is 256; from the second layer to the third layer, The dimension of is 128x256, The dimension of is 128. After being mapped to a high-dimensional latent space through the fully connected layer, the data enters the deconvolution layer. The role of the deconvolution layer is to gradually up-sample to generate an initial topological structure.

[0085] Residual connection is introduced to effectively fuse different scale features:

[0086] ;

[0087] In the formula, represents an output result, represents context encoding, represents latent encoding, represents context up-sampling, represents detail enhancement.

[0088] In this way, the breakage or discontinuity of the generated structure is avoided, and the generator finally outputs a topological model in the form of a gray-scale image, wherein the white area represents solid concrete, and the black area represents a cavity.

[0089] In the embodiment of the present application, in S14, the discriminator module includes a feature extraction layer, a correlation layer and a global pooling layer connected in sequence;

[0090] The feature extraction layer is used to extract multi-scale structural features of the topological image to generate a first feature map and a second feature map;

[0091] The correlation layer is used to generate local features according to the first feature map and the second feature map;

[0092] The global pooling layer is used to obtain a mechanical score and a permeation score according to the local features; the mechanical score is based on a stiffness formula to predict whether the deformation amount meets the standard, and the permeation score is evaluated by a resistivity response model to evaluate the uniformity of the permeation path. If the generated topology deviates from the design target in terms of mechanical or permeation performance, the discriminator will give a low score feedback.

[0093] The expression of the association layer is:

[0094] ;

[0095] In the formula, represents the output of the association layer, represents the first feature map, represents the second feature map, represents the first attention weight matrix, represents the second attention weight matrix, represents the third attention weight matrix, represents the dimension, represents the activation function, represents the transposition operation, represents the conditional vector.

[0096] In the embodiment of the application, in S15, the maximum deformation amount is calculated according to the mechanical score, the Darcy permeability is calculated according to the permeation score, and the generator module and the discriminator module are optimized by using the maximum deformation amount and the Darcy permeability, and the cavity topology structure is generated after optimization.

[0097] The adversarial training optimization module adopts a Wasserstein GAN (WGAN) framework, and the model parameters are optimized through a dynamic game between the generator and the discriminator. The loss function of the generator is designed to maximize the score of the output structure of the discriminator, while the discriminator needs to distinguish between real data and generated data, and a gradient penalty term is introduced to improve the stability of training. The optimizer selects Adam (learning rate 0.0002, momentum parameter β1=0.5), and the parameters of the generator and the discriminator are updated alternately to gradually approach the Nash equilibrium. In this process, the generator constantly explores the high-dimensional design space, and the discriminator eliminates unreasonable schemes through adversarial feedback.

[0098] The feasibility of the generated topology is verified by finite element analysis (FEA). In the stiffness verification stage, a vertical load of 500kN / m² is applied to calculate the maximum deformation amount; in the permeation verification stage, the flow path of the fluid in the cavity is simulated to calculate the Darcy permeability. The structure that does not meet the standard will be fed back to the generator for re-optimization to form a closed-loop process of “generation-evaluation-iteration”. The final output cavity topology model has a cavity rate of 38%.

[0099] From data input to final output, through the adversarial game of generator and discriminator, combined with physical constraints and measured data, the mechanical and permeability performance of the cavity topology is gradually optimized. The generator is responsible for exploring the high-dimensional design space, and the discriminator eliminates unreasonable schemes, and finally outputs a topology model with a cavity ratio of 38% and double-layer reinforcement at key nodes.

[0100] In the embodiment of the present application, S2 comprises the following sub-steps:

[0101] S21, optimizing the printing path by using a graph convolution network; wherein the graph convolution network comprises a first graph convolution layer, a second graph convolution layer and a third graph convolution layer connected in turn;

[0102] S22, based on the optimized printing path, adjusting the concrete mix proportion of the cavity topology structure by using a Transformer model to obtain the best concrete mix proportion.

[0103] In the embodiment of the present application, in S22, the Transformer model comprises an encoder and a decoder;

[0104] Modeling as a graph data containing 120 nodes and 300 edges, the node weight is related to the influence parameter of the crack spacing on the well logging signal combination. Through 3 layers of graph convolution layer to extract spatial features, combined with attention mechanism to dynamically allocate printing path priority, the total length of the path is shortened by 28%. Based on the crack fluid resistivity contrast, the Transformer model as shown in Figure 2 and the multi-head self-attention mechanism as shown in Figure 3 are used to predict the best material proportion under the environment humidity of 70%. The input is the crack fluid resistivity contrast, the environment humidity and the historical proportion data. The spatial-frequency hybrid convolution module is introduced in the encoder part, which can effectively extract the local texture, shape and frequency domain convolution of the crack, and can obtain the crack features from multiple dimensions, complement the multi-head attention, and improve the feature extraction capability.

[0105] The encoder comprises a first input layer, a first adder, an environmental factor embedding layer, a spatial-frequency hybrid convolution layer, a first multi-head attention layer, an instance normalization layer, a first feedforward network layer and a layer normalization unit;

[0106] The input of the first input layer serves as the input of the encoder, and its output is connected to the input of the first adder. The first output of the first adder, the input of the environmental factor embedding layer, and the input of the spatial-frequency hybrid convolutional layer are connected sequentially. The first, second, and third outputs of the spatial-frequency hybrid convolutional layer are all connected to the input of the first multi-head attention layer. The output of the first multi-head attention layer is connected to the first input of the instance normalization layer. The second output of the first adder is connected to the second input of the instance normalization layer. The first output of the instance normalization layer, the first feedforward network layer, and the first input of the layer normalization unit are connected. The second output of the instance normalization layer is connected to the second input of the layer normalization unit. The output of the layer normalization unit serves as the output of the encoder.

[0107] The expression for a spatial-frequency hybrid convolutional layer is:

[0108] ;

[0109] In the formula, This represents the output of the spatial-frequency hybrid convolutional layer. This represents the output of the environmental factors embedding layer. This represents the spatial convolution operation. The first spatial-frequency hybrid convolutional layer represents the... One convolutional kernel, Indicates the bias term. This represents the spatial convolution operation. This indicates the number of convolution kernels in a spatial-frequency hybrid convolutional layer.

[0110] In this embodiment of the invention, the decoder includes a second input layer, a second adder, a second multi-head attention layer, a first acceleration normalization layer, a third multi-head attention layer, a second acceleration normalization layer, a second feedforward network layer, and an adaptive normalization layer.

[0111] The input terminal of the second input layer serves as the input terminal of the decoder, and its output terminal is connected to the input terminal of the second adder. The first, second, and third output terminals of the second adder are all connected to the input terminal of the second multi-head attention layer. The output terminal of the second multi-head attention layer is connected to the first input terminal of the first acceleration normalization layer. The fourth output terminal of the second adder is connected to the second input terminal of the first acceleration normalization layer. The first output terminal of the first acceleration normalization layer is connected to the first input terminal of the third multi-head attention layer. The output terminal of the third multi-head attention layer is connected to the first input terminal of the second acceleration normalization layer. The second output terminal of the first acceleration normalization layer is connected to the second input terminal of the second acceleration normalization layer. The third output terminal of the first acceleration normalization layer is connected to the third input terminal of the second acceleration normalization layer. The output terminal of the encoder is connected to the second input terminal of the third multi-head attention layer. The first output terminal of the second acceleration normalization layer, the first input terminal of the second feedforward network layer, and the first input terminal of the adaptive normalization layer are connected sequentially. The second output terminal of the second acceleration normalization layer is connected to the second input terminal of the adaptive normalization layer. The output terminal of the adaptive normalization layer serves as the output terminal of the decoder.

[0112] The input sequence (length 24, feature number 5) is then injected with temporal information through positional encoding, and features are extracted through a 3-layer self-attention module. Each layer's self-attention... The calculation formula is:

[0113] ;

[0114] In the formula, Represents the query matrix. Represents the key matrix. Represents a value matrix, This indicates a location encoding representation related to layer thickness. Indicates the location related to layer thickness. Indicates dimension.

[0115] The decoder receives the encoder output and historical pairing sequences, and generates a predicted sequence through masked self-attention and cross-attention. The decoder's loss function... The expression is:

[0116] ;

[0117] In the formula, This represents the actual measured value of concrete density. This represents the total number of samples in the training batch. This represents the concrete density value predicted by the model. This represents the reference threshold for density variation as defined in the construction specifications. This represents the first dynamic weighting coefficient. This represents the second dynamic weighting coefficient. a small constant representing the prevention of zero errors, representing an exponential function.

[0118] In the embodiment of the present application, S3 comprises the following sub-steps:

[0119] S31, based on the adjusted printing path and the optimal concrete proportion, collecting temperature data, strain data and conductivity data of the cavity topology by using a sensor;

[0120] S32, pre-processing the temperature data, strain data and conductivity data, specifically: using Kalman filtering method to eliminate the noise of the temperature data, using fast Fourier transform to extract the main frequency component of the strain data, and calibrating the conductivity data according to the crack fluid resistivity contrast;

[0121] S33, converting the pre-processed temperature data, strain data and conductivity data into a state vector;

[0122] S34, generating an action vector according to the state vector to print the cavity topology, the action vector including printing speed adjustment amount, layer thickness adjustment amount and carbon fiber slurry injection ratio;

[0123] S35, after printing, receiving the original signal by using an ultrasonic flaw detector, and carrying out denoising processing on the original signal;

[0124] S36, based on the original signal after denoising processing, generating a crack position by using a 3D convolutional neural network, and repairing the crack position;

[0125] In S33, the calculation formula for denoising processing of the original signal is:

[0126] ;

[0127] In the formula, denotes the original signal after denoising processing, denotes the original signal, denotes the setting time, denotes the reference setting period, denotes the real-time frequency component of the signal, denotes the structure reference frequency specified by design.

[0128] In the embodiment of the present application, during the printing process, 24 groups of distributed optical fiber sensors embedded along the inner and outer walls of the cavity collect temperature, strain and conductivity data in real time. When the conductivity of a certain area is abnormal or the temperature gradient exceeds the safe range, the DRL system triggers dynamic regulation and control.

[0129] As Figure 4As shown, the DRL system consists of a perception module, a decision module, and an execution module. The perception module receives real-time data through a fiber sensor network and performs preprocessing: temperature data is filtered by Kalman filter to eliminate noise, strain data is extracted by Fast Fourier Transform (FFT) to extract the main frequency component, and conductivity data is dynamically calibrated according to the contrast of crack fluid resistivity. The preprocessed data is integrated into a state vector, including conductivity deviation, temperature gradient, and interlayer strain difference, etc. 12-dimensional features.

[0130] The decision module uses the Proximal Policy Optimization (PPO) algorithm based on the Actor-Critic architecture. The Actor network inputs the state vector t , and outputs the action vector, including the print speed adjustment, the layer thickness adjustment, and the carbon fiber slurry injection ratio. The Critic network (value network) evaluates the value of the current state (V(st)), guiding the Actor to optimize the strategy. The reward function is designed in a composite form, specifically:

[0131] ;

[0132] wherein, represents the first weight coefficient, represents the second weight coefficient, represents the third weight coefficient, represents the crack width prediction value, represents the real-time resistivity of the printing material, represents the design reference resistivity, represents the time difference.

[0133] The execution module converts the decision into physical control instructions: the print head speed is reduced from 50 mm / s to 42 mm / s, and precise speed regulation is achieved through servo motor closed-loop control; the layer thickness is adjusted from 10 mm to 8 mm, and the nozzle opening is adjusted by a piezoelectric ceramic driver; the carbon fiber reinforced slurry is injected into the abnormal area through a high-pressure injection pump, and the fiber content is increased to 2.5%. At the same time, the state data after adjustment is recorded and fed back to the DRL system, forming a "perception-decision-execution-feedback" closed loop, continuously optimizing the adjustment strategy.

[0134] After the cavity structure is printed, joint detection is performed by an ultrasonic flaw detector and a resistivity imager. The ultrasonic flaw detector emits a pulse wave with a frequency of 5 MHz, and after receiving the reflected signal, the original signal is first denoised by wavelet transform based on the Daubechies wavelet basis, and a t cureThe adaptive filter function is used to denoise the signal. The denoised signal is input into a 3D convolutional neural network (3D-CNN), and the network structure includes four 3D convolutional layers (convolution kernel size 3x3x3, step size 1x1x1), two maximum pooling layers (pooling size 2x2x2), and one fully connected layer. The convolutional layer extracts the three-dimensional crack features inside the cavity through the ReLU activation function, and finally outputs the crack position and geometric parameters. The resistivity data is analyzed by the graph attention network to locate the abnormal area of conductivity. After cross verification with the ultrasonic wave results, the defect positioning error is ≤1.5mm. For the defect area, the gradient printing scheme generated by GAN is used: in the crack extension direction, the fiber direction density is adjusted according to the 0-45-90 gradient, and nano-silicon dioxide is locally added for node repair.

[0135] Those skilled in the art will appreciate that the embodiments described herein are presented for the purpose of aiding the reader in understanding the principles of the present application, and should be understood as not limiting the scope of protection of the present application to such specific recitations and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A method for printing a fractured artificial rock based on an optimization algorithm, characterized by, The method comprises the following steps: S1, collecting a data set of reservoir fracture logging, and generating a cavity topology structure by using a generator module and a discriminator module; S2, adjusting a printing path of the cavity topology structure, and optimizing a concrete proportion of the cavity topology structure to obtain an optimal concrete proportion; S3, completing artificial rock printing according to the optimal concrete proportion of the cavity topology structure and the adjusted printing path, and repairing a node at a fracture position.

2. The method of claim 1, wherein, The S1 comprises the following sub-steps: S11, collecting fracture width, resistivity response and mechanical property data of reservoir fracture logging as a data set; S12, constructing a fracture spacing constraint condition for the data set, and encoding the fracture spacing constraint condition into a condition vector; S13, inputting the data set and the condition vector into the generator module to generate a topology image; S14, inputting the topology image into the discriminator module to generate a mechanical score and a permeability score; S15, optimizing the generator module and the discriminator module by using an adversarial training optimization module according to the mechanical score and the permeability score to generate the cavity topology structure.

3. The method of claim 2, wherein the optimization algorithm-based printing of a fractured artificial rock is characterized by, In the S12, the fracture spacing constraint condition is encoded into the condition vector, specifically: when the fracture spacing is less than or equal to a specific threshold of 5 mm, the specific threshold is encoded as 0, otherwise the specific threshold is encoded as 1.

4. The method of claim 2, wherein the optimization algorithm-based fractured artificial rock printing method is characterized by, In the S13, the generator module comprises a first full connection layer, a second full connection layer, a third full connection layer and a plurality of layers of deconvolution layers connected in sequence; The first full connection layer, the second full connection layer and the third full connection layer are used to map the data set and the condition vector to a high-dimensional latent space; The deconvolution layer is used to up-sample the high-dimensional latent space to generate the topology image; The expression of the deconvolution layer is: ； In the formula, denotes the output data of the layer deconvolution layer, denotes the input data of the layer deconvolution layer, denotes the conditional vector, denotes the concatenation function.

5. The method of claim 2, wherein the optimization algorithm-based printing of a fractured artificial rock is characterized by, In the S14, the discriminator module comprises a feature extraction layer, a correlation layer and a global pooling layer connected in sequence; The feature extraction layer is used to extract multi-scale structural features of the topology image to generate a first feature map and a second feature map; The correlation layer is used to generate a local feature according to the first feature map and the second feature map; The global pooling layer is used to obtain the mechanical score and the permeability score according to the local feature; The expression of the correlation layer is: ； wherein denotes the output of the association layer, denotes the first feature map, denotes the second feature map, denotes the first attention weight matrix, denotes the second attention weight matrix, denotes the third attention weight matrix, denotes the dimension, denotes the activation function, denotes the transpose operation, denotes the condition vector.

6. The method of printing an optimized algorithm based fractured artificial rock according to claim 2, wherein, In the S15, the maximum deformation is calculated according to the mechanical score, the Darcy permeability is calculated according to the permeability score, and the generator module and the discriminator module are optimized by using the maximum deformation and the Darcy permeability to generate the cavity topology structure after optimization.

7. The method of claim 1, wherein the optimization algorithm-based fractured artificial rock printing method is characterized by, The S2 comprises the following sub-steps: S21, optimizing the printing path by using a graph convolution network; wherein the graph convolution network comprises a first graph convolution layer, a second graph convolution layer and a third graph convolution layer connected in sequence; S22, adjusting the concrete proportion of the cavity topology structure based on the optimized printing path by using a Transformer model to obtain the optimal concrete proportion.

8. The method of printing an optimized algorithm-based fractured artificial rock according to claim 7, wherein, In the S22, the Transformer model comprises an encoder and a decoder; The encoder comprises a first input layer, a first adder, an environmental factor embedding layer, a space-frequency hybrid convolution layer, a first multi-head attention layer, an instance normalization layer, a first feedforward network layer and a layer normalization unit; The input end of the first input layer is connected with the input end of the encoder, and the output end is connected with the input end of the first adder; the first output end of the first adder, the input end of the environmental factor embedding layer and the input end of the spatial-frequency mixed convolution layer are sequentially connected; the first output end, the second output end and the third output end of the spatial-frequency mixed convolution layer are all connected with the input end of the first multi-head attention layer; the output end of the first multi-head attention layer is connected with the first input end of the instance normalization layer; The second output end of the first adder is connected with the second input end of the instance normalization layer; The first output end of the instance normalization layer is connected with the first input end of the first feedforward network layer and the layer normalization unit; The second output end of the instance normalization layer is connected with the second input end of the layer normalization unit; and the output end of the layer normalization unit is the output end of the encoder; The expression of the spatial-frequency mixed convolution layer is: ； wherein denotes the output of the spatial-frequency hybrid convolutional layer, denotes the output of the environmental factor embedding layer, denotes a spatial convolution operation, denotes the number of convolution kernels of the spatial-frequency hybrid convolutional layer, denotes the number of convolution kernels of the spatial-frequency hybrid convolutional layer, denotes a bias term, denotes a spatial convolution operation, denotes the number of convolution kernels of the spatial-frequency hybrid convolutional layer.

9. The method of printing an optimized algorithm-based fractured artificial rock according to claim 8, wherein, The decoder comprises a second input layer, a second adder, a second multi-head attention layer, a first accelerated normalization layer, a third multi-head attention layer, a second accelerated normalization layer, a second feedforward network layer and an adaptive normalization layer; The input end of the second input layer is the input end of the decoder, and the output end is connected with the input end of the second adder; The first output end, the second output end and the third output end of the second adder are all connected with the input end of the second multi-head attention layer; The output end of the second multi-head attention layer is connected with the first input end of the first accelerated normalization layer; The fourth output end of the second adder is connected with the second input end of the first accelerated normalization layer; the first output end of the first accelerated normalization layer is connected with the first input end of the third multi-head attention layer; The output end of the third multi-head attention layer is connected with the first input end of the second accelerated normalization layer; the second output end of the first accelerated normalization layer is connected with the second input end of the second accelerated normalization layer; the third output end of the first accelerated normalization layer is connected with the third input end of the second accelerated normalization layer; the output end of the encoder is connected with the second input end of the third multi-head attention layer; the first output end of the second accelerated normalization layer is sequentially connected with the second feedforward network layer and the first input end of the adaptive normalization layer; the second output end of the second accelerated normalization layer is connected with the second input end of the adaptive normalization layer; The output end of the adaptive normalization layer is the output end of the decoder; Loss function of the decoder The expression is: ； wherein, denotes the actual monitored concrete density measurement value, denotes the total number of samples in the training batch, denotes the model predicted concrete density value, denotes the density variation reference threshold defined by the construction specification, denotes the first dynamic weight coefficient, denotes the second dynamic weight coefficient, denotes a small constant to prevent division by zero errors, denotes the exponential function.

10. The method of claim 1, wherein, The S3 comprises the following sub-steps: S31, based on the adjusted printing path and the optimal concrete proportion, collecting temperature data, strain data and conductivity data of the cavity topological structure by using a sensor; S32, preprocessing the temperature data, the strain data and the conductivity data, specifically: eliminating the noise of the temperature data by using a Kalman filtering method, extracting the main frequency component of the strain data by using a fast Fourier transform, and calibrating the conductivity data according to the crack fluid resistivity contrast; S33, converting the preprocessed temperature data, strain data and conductivity data into a state vector; S34, generating an action vector according to the state vector, the action vector comprising a printing speed adjustment amount, a layer thickness adjustment amount and a carbon fiber slurry injection ratio, and printing the cavity topological structure; S35, after printing, receiving an original signal by using an ultrasonic flaw detector, and performing denoising processing on the original signal; S36, based on the original signal after denoising processing, generating a crack position by using a 3D convolutional neural network, and performing node repair on the crack position; In the S33, the calculation formula for denoising processing of the original signal is: ； In the formula, denotes the original signal after the de-noising process, denotes the original signal, denotes the coagulation time, denotes the reference coagulation period, denotes the real-time frequency component of the signal, denotes the structure reference frequency prescribed by design.

Images