A deep learning-based drilling fracture type prediction method, system and medium

Abstract

The present application relates to the technical field of drilling data processing, and particularly relates to a drilling fracture type prediction method and system based on deep learning and a medium. The method comprises the following steps: obtaining a fracture scene original data set; training a generative adversarial network model based on the fracture scene original data set, generating a fracture enhanced sample set, and obtaining an expanded fracture image data set; performing multi-modal information fusion modeling on the expanded fracture image data set to obtain multi-modal fracture fusion feature data; constructing a fracture evolution time series model based on the multi-modal fracture fusion feature data to obtain time series evolution feature embedding data; and performing deep classification reasoning on the time series evolution feature embedding data to obtain fracture type prediction result data. Through image enhancement, multi-modal information fusion, time series evolution modeling and visual path rendering, the present application effectively improves the accuracy, robustness and dynamic adaptability of fracture type prediction, and enhances the interpretability of the results.

Description

Technical Field

[0001] This invention relates to the field of drilling data processing technology, and in particular to a method, system and medium for predicting drilling fracture types based on deep learning. Background Technology

[0002] Early methods for predicting fracture types relied heavily on physical detection techniques such as sonic logging and resistivity logging, combined with expert interpretation. While these methods achieved fracture identification to some extent, they suffered from problems such as limited information dimensions, strong subjectivity, and low processing efficiency.

[0003] Current research has attempted to use deep classification networks for pattern recognition of fracture images. However, most methods rely solely on single-modal image input, making it difficult to integrate logging, image, and structural information. Furthermore, they lack modeling of fracture evolution over time, resulting in limited accuracy in predicting fracture types in complex drilling environments. In addition, existing technologies suffer from data scarcity, especially when samples of different fracture types are unevenly distributed, leading to overfitting during model training and affecting generalization ability. Moreover, the prediction results are difficult to present intuitively, hindering understanding and decision support for engineers. Summary of the Invention

[0004] Therefore, the present invention needs to provide a method, system and medium for predicting drilling fracture types based on deep learning, in order to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, a deep learning-based method for predicting drilling fracture types includes the following steps:

[0006] Step S1: Obtain the original dataset of the crack scene; train a generative adversarial network model based on the original dataset of the crack scene to generate a crack enhancement sample set, and obtain an extended crack image dataset;

[0007] Step S2: Perform multimodal information fusion modeling on the extended crack image dataset to obtain multimodal crack fusion feature data;

[0008] Step S3: Construct a crack evolution time series model based on multimodal crack fusion feature data to obtain time series evolution feature embedding data;

[0009] Step S4: Perform deep classification inference on the temporal evolution feature embedding data to obtain crack type prediction results;

[0010] Step S5: Construct a visualization path based on the temporal evolution feature embedding data and crack type prediction results to obtain the crack type discrimination path visualization results.

[0011] This invention, by introducing a joint processing mechanism of image enhancement, information fusion, temporal modeling, and visualization, demonstrates several beneficial effects in crack type prediction tasks: First, by expanding the number and type distribution of samples, it significantly alleviates the model overfitting problem caused by sample scarcity and class imbalance, enhancing the model's ability to learn multiple crack morphologies; second, by fusing multi-dimensional feature information, including texture features, morphological boundaries, and spatial structure representation, it elevates crack discrimination from a single image visual approach to a multi-modal collaborative analysis, effectively improving robustness in complex backgrounds; furthermore, by introducing crack... Time-series modeling of fracture evolution enables the model to continuously recognize fracture development trends and state changes, enhancing its adaptability to fracture type changes in dynamic drilling scenarios. Furthermore, through dimensional regularization of deep features and construction of inference networks, effective mapping of high-dimensional information to a low-dimensional decision space is achieved, improving the stability and accuracy of classification. Finally, by rendering the prediction results as a path graph, the model gains interpretability and intuitiveness, facilitating engineers to quickly understand fracture type distribution and evolution trends during actual drilling operations, thereby improving on-site decision-making efficiency and safety assurance levels.

[0012] Preferably, the present invention also provides a deep learning-based drilling fracture type prediction system for performing the above-described deep learning-based drilling fracture type prediction method, wherein the deep learning-based drilling fracture type prediction system includes:

[0013] A generative enhancement module is used to obtain the original dataset of the crack scene; a generative adversarial network model is trained based on the original dataset of the crack scene to generate a crack enhancement sample set, resulting in an extended crack image dataset.

[0014] The feature fusion module is used to perform multimodal information fusion modeling on the extended crack image dataset to obtain multimodal crack fusion feature data;

[0015] The time-series modeling module is used to construct a time-series model of crack evolution based on multimodal crack fusion feature data, and obtain time-series evolution feature embedding data;

[0016] The classification and reasoning module is used to perform deep classification and reasoning on the embedded data of temporal evolution features to obtain crack type prediction results.

[0017] The path visualization module is used to construct a visual path based on the embedded data of temporal evolution features and the crack type prediction results, and obtain the crack type discrimination path visualization results.

[0018] This invention effectively expands the diversity and quantity of crack data through an enhancement module, alleviating the data scarcity problem and improving the basic quality of model training. Multimodal information fusion significantly enhances the comprehensiveness and richness of feature representation, improving the ability to capture complex crack features. The temporal modeling module deeply mines the temporal dynamic features of crack evolution, enhancing the understanding of crack development patterns and the accuracy of prediction. The classification reasoning module achieves efficient and accurate crack type identification through deep learning methods, improving the reliability of automated classification. The path visualization module intuitively presents complex temporal and classification information, enhancing the interpretability of results and user interaction experience, significantly improving the intelligence level and application value of crack monitoring.

[0019] Preferably, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the deep learning-based drilling fracture type prediction method. Attached Figure Description

[0020] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0021] Figure 1 This is a schematic diagram of the steps of the deep learning-based drilling fracture type prediction method of the present invention;

[0022] Figure 2 for Figure 1 A detailed flowchart of step S1;

[0023] Figure 3 This is a time-series evolution diagram of the crack type discrimination path in an embodiment of the present invention. Detailed Implementation

[0024] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0025] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0026] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0027] To achieve the above objectives, please refer to Figures 1 to 3 This invention provides a deep learning-based method for predicting drilling fracture types, the method comprising the following steps:

[0028] Step S1: Obtain the original dataset of the crack scene; train a generative adversarial network model based on the original dataset of the crack scene to generate a crack enhancement sample set, and obtain an extended crack image dataset;

[0029] Step S2: Perform multimodal information fusion modeling on the extended crack image dataset to obtain multimodal crack fusion feature data;

[0030] Step S3: Construct a crack evolution time series model based on multimodal crack fusion feature data to obtain time series evolution feature embedding data;

[0031] Step S4: Perform deep classification inference on the temporal evolution feature embedding data to obtain crack type prediction results;

[0032] Step S5: Construct a visualization path based on the temporal evolution feature embedding data and crack type prediction results to obtain the crack type discrimination path visualization results.

[0033] Preferably, step S1 includes the following steps:

[0034] Step S11: Obtain the raw data of the fracture scene in the drilling area and preprocess it to obtain a structured fracture image dataset;

[0035] Step S12: Construct a generative adversarial network model based on the structured crack image dataset. Design a generator network to generate crack image samples, design a discriminator network to determine the authenticity of the images, and optimize the network parameters through adversarial training to obtain the generative adversarial network model.

[0036] Step S13: Input the structured crack image dataset into the generative adversarial network model to generate crack enhancement sample sets under various crack morphologies, textures, and background interference conditions, and obtain crack enhancement image data.

[0037] Step S14: Based on the structured crack image dataset and the crack enhancement image data, perform sample fusion, label unification and format normalization to obtain the extended crack image dataset.

[0038] In this embodiment of the invention, high-resolution images of areas such as the casing wall and wellbore rock mass are acquired during drilling operations using industrial endoscopic imaging equipment. The acquired image resolution is no less than 1920×1080 pixels, and the sampling frame rate is no less than 10 frames per second to ensure that the acquired fracture images have discernible details. After acquisition, edge sharpening enhancement methods are used to improve the clarity of the fracture outline, and histogram equalization is used to adjust the brightness and contrast to eliminate image noise interference. Then, image smoothing processing based on median filtering is combined to remove light spot artifacts caused by the downhole environment, completing the preprocessing of the fracture images and obtaining a structured fracture image dataset containing multiple fracture structures in the drilling area. Based on this structured image data, a dual-channel neural network construction method is used to build two independent but interactive network units for image generation and image discrimination. The image generation unit has a fixed input size of 128×128 pixel image tensor, and the output size remains consistent. The core structure consists of three layers of convolutional processing modules and two layers of deconvolutional reconstruction modules, with the convolution kernel size set to 5×5 and the stride to 1. The discrimination unit adopts a four-layer cascaded structure, which realizes image authenticity recognition through layer-by-layer convolution processing. Its output is a single-channel binary image judgment result. The discrimination result is used to correct the output quality of the generation unit through an error backpropagation mechanism. The structured image data is input into the above network structure at a ratio of 70% as input image and 30% as discrimination reference image. The error feedback update process is executed cyclically for no less than 3000 iterations to generate enhanced image data under complex interference conditions such as crack direction distortion, texture discontinuity, and background reflection disturbance, ensuring that the enhanced samples cover the typical crack morphology distribution characteristics. The original structured crack image data and the enhanced image data are matched and fused according to image size, texture feature labels, and morphological boundary categories. The label mapping matrix is ​​used to uniformly label samples from different sources into a unified classification standard system. The fused data is uniformly adjusted to a 256×256 pixel format and re-encoded and grouped according to a unified naming rule. Finally, an extended crack image dataset containing multiple crack morphologies, multiple interference background conditions, and a unified labeling system is constructed.

[0039] This invention improves the data quality and consistency of original crack images through structured preprocessing, laying a unified data foundation for subsequent feature extraction and classification. It utilizes an adversarial generation mechanism to synthesize diverse crack images, effectively expanding the coverage of training samples in dimensions such as crack direction, texture coarseness, and background illumination, enhancing the model's adaptability to complex environments. By introducing multiple types of interference factors during crack image synthesis, the model can actively learn anti-interference features during training, improving the robustness and generalization of crack recognition. Through label mapping and format standardization of the original and enhanced images, a high-quality extended image dataset with unified structure and semantic consistency is constructed, providing rich and reliable data support for subsequent multimodal feature fusion and temporal modeling, fundamentally improving the accuracy and stability of the entire prediction process.

[0040] Preferably, step S2 includes the following steps:

[0041] Step S21: Extract texture features, morphological features and edge features from each sample in the extended crack image dataset to obtain a crack feature map set;

[0042] Step S22: Construct a three-branch multimodal feature fusion network architecture based on the crack feature map set, and modularly encode the texture, edge and shape information of the image to obtain the branch encoded feature vector group;

[0043] Step S23: Input the branch-encoded feature vector group into the preset attention-guided feature interaction component, and perform inter-channel weight allocation and spatial correlation aggregation to obtain the fused enhanced feature tensor;

[0044] Step S24: Integrate and reduce the dimensionality of the fusion enhancement feature tensor and perform label mapping to generate multimodal crack fusion feature data.

[0045] In this embodiment of the invention, based on an extended crack image dataset, texture feature extraction, morphological feature extraction, and edge feature extraction are sequentially performed on each image. Texture feature extraction uses a gray-level co-occurrence matrix to calculate the image's energy, contrast, entropy, and correlation indices. The image block size is set to 32×32 pixels, and the gray-level levels are limited to 64. Morphological feature extraction uses morphological closing operations combined with contour fitting to extract the crack's principal axis length, number of branches, and boundary envelope area. A 5×5 elliptical structuring element is used for image dilation and erosion. Edge feature extraction uses the Sobel edge operator to extract edge gradient images in both the horizontal and vertical directions. Noise pixels are filtered out using a double threshold. The edge amplitude map is normalized and stored in a unified data format, ultimately forming a crack feature map set composed of texture, morphological, and edge feature maps. This feature map set is input into three feature encoding paths. Each path contains two local feature extraction layers and one global feature compression layer. The local extraction layers use a 3×3 convolution kernel with a stride of 1, and the number of output channels is set to 32 and 32 respectively. 64. The global compression layer uses a global averaging method to compress the two-dimensional features into fixed-length vectors. The three paths correspond to the three dimensions of texture, edge, and morphology, respectively, and finally output three independent sets of branch-encoded feature vectors. The three sets of branch-encoded feature vectors are concatenated into a three-dimensional array of 3×64 along the channel direction and input into the attention-guided feature interaction component. This component includes a channel response calculation module and a spatial aggregation module. The channel response calculation module performs mean normalization on each channel and calculates its correlation coefficient with all channels, outputting a response coefficient matrix between channels. The spatial aggregation module performs weighted superposition based on the response coefficients and introduces a two-dimensional position information matrix to encode weights for each position in the feature map, finally forming a fusion enhancement feature tensor of size 64×64. Two-dimensional principal component analysis is performed on the fusion enhancement feature tensor to compress the original 64-dimensional channels to 32-dimensional while retaining more than 95% of the information. The compressed tensor is categorized through a preset label mapping table, with each tensor corresponding one-to-one with the type label of the corresponding crack image, generating multimodal crack fusion feature data.

[0046] This invention decomposes and extracts the texture, morphology, and edge information of crack images in a fine-grained manner, effectively compensating for the limitations of single feature dimensions in complex image structures and enhancing the ability to capture details of different crack structures. Modular encoding is achieved through a three-branch structure, allowing each type of feature to fully learn its spatial pattern in independent channels, maintaining structural integrity and semantic independence between features, and laying a clear semantic foundation for subsequent interactive fusion. The introduction of channel attention and spatial weighting mechanisms eliminates the reliance on static superposition in the feature fusion process, instead dynamically adjusting channel responses and spatial positional associations, effectively improving the model's perception accuracy and information aggregation capabilities for key crack regions. Through feature dimensionality reduction and label mapping operations, high-dimensional features are mapped to low-dimensional representations while maintaining semantic integrity, providing structurally concise, semantically complete, and information-rich multimodal crack fusion feature data for subsequent temporal modeling and type classification.

[0047] Preferably, step S23 includes the following steps:

[0048] Step S231: Perform channel dimension expansion processing on the branch coding feature vector group to obtain the channel expanded feature vector group;

[0049] Step S232: Perform channel weight response analysis based on the channel expanded feature vector group to obtain the channel attention weight map;

[0050] Step S233: Perform weighted fusion on the expanded feature vector groups of the channels to obtain the channel attention-enhanced feature groups;

[0051] Step S234: Perform location association encoding based on channel attention-enhanced feature groups to obtain a spatial location information encoding matrix;

[0052] Step S235: Perform spatial weight allocation on the channel attention enhancement feature group to obtain the spatial attention enhancement feature group;

[0053] Step S236: Perform joint aggregation of channel and spatial dimensions on the spatial attention-enhanced feature group to obtain the fused enhanced feature tensor.

[0054] In this embodiment of the invention, the three sets of branch-encoded feature vectors are expanded according to the channel dimension, with each set of feature vectors having a dimension of 64. After expansion, the three sets of feature vectors are concatenated in the column direction to form a channel expanded feature vector group of size 3×64. Maximum value normalization is performed on each channel vector to ensure the numerical stability of subsequent response calculations. Using the channel expanded feature vector group as input, the cosine similarity between each channel vector and all channel vectors is calculated sequentially to obtain a symmetric channel similarity matrix. This matrix is ​​then normalized after row summation and used as the channel response value. All channel response values ​​are sorted and mapped to the [0,1] interval in normalized order to construct a channel attention weight map of size 1×64. This channel attention weight map is multiplied element-wise with the channel expanded feature vector group to form a weighted channel attention enhancement feature group, maintaining a size of 3×64. The enhancement process improves channel representation ability by strengthening high-weight channels and suppressing low-response channels. The feature groups are sequentially mapped to a two-dimensional spatial location matrix. Based on the activation center position of each channel feature, a spatial mapping coordinate system is constructed. A Gaussian distribution function with a radius of 5 is generated with the centroid of the activation region of each channel as the spatial correlation kernel function. The spatial response results of all channels are superimposed to form a spatial location information encoding matrix of size 64×64. The above spatial location information encoding matrix is ​​used as a spatial weight factor and applied to the two-dimensional feature map corresponding to each channel in the channel attention enhancement feature group (which is restored to a two-dimensional structure through reconstruction). After pixel-by-pixel weighting, the output is a spatial attention enhancement feature group of size 64×64. This group of data reflects the response intensity and distribution differences of the spatial region to the crack feature. The spatial attention enhancement feature group is jointly aggregated according to the spatial location dimension and the channel dimension. First, a weighted average is performed along the channel dimension to obtain a spatial aggregated feature map. Then, it is compressed into a two-dimensional tensor of uniform size through spatial pooling operation. Finally, a 64-dimensional fusion enhancement feature tensor is constructed by combining the channel weighting results.

[0055] This invention expands response signals in different feature subspaces independently through channel-dimensional expansion, enhancing the model's discriminability of various semantic features and providing a data foundation for accurate attention weight allocation. The channel weight response analysis process strengthens the identification of key channels, effectively highlighting representative textures, boundaries, or morphological expressions and suppressing the interference of redundant information channels on the final judgment result. The weighted fusion process dynamically enhances high-response channel information, improving the overall expression strength of salient crack features. At the spatial construction level, the introduction of a positional encoding mechanism guides features to cluster towards densely distributed or complex crack structures, giving the model significant spatial discrimination capabilities. The spatial weight allocation process strengthens the perception priority of crack structure regions, effectively extracting key spatial distribution information such as crack direction, extension, and inflection points. Through joint channel and spatial aggregation operations, high-value features are globally integrated in a two-dimensional dimension, significantly improving the stability of subsequent feature inference processes and the completeness of the prediction foundation.

[0056] Preferably, step S3 includes the following steps:

[0057] Step S31: Perform time label alignment processing on the multimodal crack fusion feature data to obtain structured time-series crack feature data;

[0058] Step S32: Construct a crack state transition map model based on structured time-series crack feature data to obtain a crack state transition map;

[0059] Step S33: Initialize the time series modeling of the crack state transition map to obtain the initial parameter set of crack evolution;

[0060] Step S34: Train a preset long short-term memory network model based on the initial parameter set of crack evolution to obtain a crack evolution time series prediction model;

[0061] Step S35: Input the multimodal crack fusion feature data into the crack evolution time series prediction model and perform feature inference to obtain the time series evolution feature embedding data.

[0062] In this embodiment of the invention, multimodal fracture fusion feature data is grouped and organized according to drilling depth and corresponding acquisition time. Each group of data must include three pieces of information: acquisition timestamp, well depth location identifier, and fracture feature code. A time axis template with a minimum time interval of 5 seconds is used to uniformly interpolate and align all acquired data. Missing data is filled in using linear interpolation, and data from repeated time nodes are merged using an average fusion method. Finally, structured time-series fracture feature data with continuous time dimension is formed, with the dimension set to N×D, where N is the number of time steps and D is the fracture feature dimension of each step. Based on this structured data, the feature distance difference between any two consecutive time points is calculated using Euclidean distance calculation, with a distance threshold set to 0.25. If the difference between consecutive time points is greater than this threshold, it is marked as a state transition event, and directed edges are established in the transition graph. All time points are used as graph nodes to establish connections, forming a fracture state transition graph, where each edge is accompanied by a transition weight and transition direction. The graph is encoded and stored in an adjacency matrix manner. The transition graph is segmented into time windows, with each segment having a set length. For each of the five time steps, the frequency of state transitions, the average amplitude of transitions, and the duration of transitions are statistically analyzed within each window. The maximum feature change, the mean change trend, and the standard deviation within the corresponding window are extracted to form an initial parameter set for crack evolution containing six numerical indicators, with a quantity of N / 5 groups. Based on the above parameter set, it is sequentially input into a three-layer temporal memory structure. The structure parameters are set as follows: input dimension 6, hidden unit dimension 128, and output dimension 64. Each layer of the structure contains a temporal memory unit, a gating structure, and a state update unit. The training cycle is set to 200 rounds, and the batch size is 32. The difference between the output sequence and the actual evolution trend is evaluated by the mean square error index, and the training process is terminated when the error is less than 0.01, finally obtaining the crack evolution temporal prediction model. The multimodal crack fusion feature data is input into the trained prediction structure in chronological order, pushed in one by one according to the time step. Each step outputs a feature vector with a dimension of 64 after joint judgment of the current state and the previous state. Feature generation is completed at all time steps, finally forming a temporal evolution feature embedding data with a dimension of N×64.

[0063] This invention assigns a unified temporal index to multimodal fusion features through time-label alignment, constructing a feature sequence structure that conforms to the real evolutionary laws of the drilling process, providing basic semantic continuity for subsequent dynamic behavior modeling. By constructing a crack state transition map, it realizes a graph structure representation of crack features changing over time, accurately capturing the evolutionary trends of crack appearance, expansion, and convergence, enhancing the model's ability to characterize complex temporal behaviors. Through the temporal modeling initialization process, it extracts state transition frequency, directionality, and amplitude features, quantifying the structural features of the evolutionary pattern and providing a stable parameter starting point for the prediction mechanism. Combined with the training process of the evolutionary prediction mechanism, it establishes a prediction structure with strong temporal correlation and sufficient memory capacity, effectively improving the learning ability of historical feature change patterns and the accuracy of capturing future crack state trends. By generating embedded evolutionary representations through temporal reasoning operations on input features, it constructs a continuous mapping relationship between high-dimensional temporal semantics and crack morphological changes, providing a dynamic and structured temporal feature expression foundation for subsequent crack type reasoning and visualization path construction.

[0064] Preferably, step S4 includes the following steps:

[0065] Step S41: Perform high-dimensional feature normalization on the temporal evolution feature embedding data to obtain feature data with unified embedding dimensions;

[0066] Step S42: Construct a deep classification inference network structure based on unified feature data of embedding dimension to obtain a crack type classification network model;

[0067] Step S43: Initialize and optimize the parameters of the crack type classification network model to obtain the parameter set of the deployable inference model;

[0068] Step S44: Load the crack type classification network model based on the deployable inference model parameter set to obtain the loaded inference model;

[0069] Step S45: Input the unified feature data of the embedded dimension into the loaded inference model and perform crack type inference to obtain crack type prediction result data.

[0070] In this embodiment of the invention, a dimension normalization operation is performed on the embedded data of time-series evolution features. The original embedded data is an N×64 time-series feature matrix. First, a sliding window method is applied to each time series to extract local feature statistics. The window length is set to 5 time steps with a step size of 1. The extracted indicators include maximum value, minimum value, mean, variance, and skewness. Each window outputs a 5-dimensional feature vector. After flattening and concatenation operations, a high-dimensional sequence feature set of size (N-4)×320 is formed. Subsequently, Z-score normalization is performed on this feature set along the feature dimension, with the mean set to 0 and the standard deviation set to 1. Finally, unified feature data with embedded dimensions is formed. Using this unified feature data as input, a crack type classification network based on a multi-level feature mapping structure is constructed. The network structure consists of two convolutional processing layers, two normalization and activation layers, and one fully connected classification layer. The first layer has 64 convolutional kernels with a kernel size of 1×3 and a stride of 1. The second layer has 128 convolutional kernels with a kernel size of 1×3. The normalization layer uses batch normalization to distribute and normalize the data for each channel dimension. The activation layer uses ReLU to enhance non-linear expression. The output vector dimension of the fully connected classification layer is the number of crack type categories C, where C is limited to 5. Classes were defined, corresponding to linear cracks, curved cracks, composite cracks, network cracks, and closed cracks, respectively. All learnable parameters in the above classification structure were initialized. Convolutional kernel weights were assigned values ​​using a truncated normal distribution with a mean of 0 and a standard deviation of 0.02. Bias values ​​were uniformly initialized to a constant of 0. During the optimization phase, cross-entropy was used as the loss evaluation criterion, the learning rate was set to 0.001, the momentum parameter was set to 0.9, the training period was 100 rounds, and the batch size was 64. After completion, all weight coefficients, biases, and normalized parameters were extracted and organized to form a deployable inference model parameter set. The standard loading interface was then called. The parameter set of the inference model is bound to the crack type classification network structure layer by layer. The parameter mapping, structure locking and input / output channel declaration are completed strictly according to the hierarchical order. Finally, it is loaded into an inference network with a fixed structure and preset parameters, which is called the loaded inference model. The unified feature data of the embedding dimension is input into the loaded inference model one by one. The feature mapping, nonlinear transformation and multi-level transformation are completed through the forward propagation mechanism. The crack type classification prediction result with dimension 1×C is output, where each dimension value corresponds to the confidence probability of a crack type. Finally, the category with the highest probability is used as the crack type prediction result data of the feature sequence.

[0071] This invention unifies the feature dimensions by performing high-dimensional regularization on temporal evolution feature data, providing a solid data foundation for subsequent model training and inference. The constructed and optimized deep classification inference network structure effectively captures complex patterns of crack features, improving classification accuracy and robustness. Through parameter initialization and optimization, the model ensures good generalization ability and inference efficiency, making it suitable for practical deployment. After loading the optimized model, crack type inference can be completed quickly and efficiently, improving the real-time performance and reliability of predictions. This invention achieves an effective transformation from multi-dimensional temporal data to accurate crack type prediction, significantly enhancing the automation level and practical value of crack monitoring and identification.

[0072] Preferably, step S42 includes the following steps:

[0073] Step S421: Perform temporal expansion on the unified feature data of the embedding dimension to obtain the time series feature mapping matrix;

[0074] Step S422: Construct a cross-time step attention mechanism based on the time series feature mapping matrix to obtain multi-time step response weight data;

[0075] Step S423: Weighted fusion of the unified feature data of the embedded dimension and the multi-time response weight data is performed to obtain the temporal augmentation feature tensor;

[0076] Step S424: Construct a multi-layer convolution-normalization-activation structure based on the temporal enhancement feature tensor, and extract deep feature representations to obtain temporal embedding classification feature groups;

[0077] Step S425: Perform a fully connected mapping on the temporal embedded classification feature group to obtain the initial classification output vector of crack type;

[0078] Step S426: Encapsulate the network structure and set the input / output interface for the initial crack type classification output vector to obtain the crack type classification network model.

[0079] In this embodiment of the invention, the embedded dimension unified feature data are arranged in chronological order, with each time step having a data dimension of 1×320. These are sequentially expanded to form a time series feature mapping matrix. The total length of the time series is set to N, resulting in a two-dimensional feature matrix of size N×320. Based on this time series feature mapping matrix, a global dependency response structure is constructed for each time step. By calculating the weighted inner product similarity between feature vectors of any two time steps, an N×N time-relatedness matrix is ​​formed. Subsequently, softmax normalization is performed on each row to obtain the relationship between each time step and the current time step. The response weights at each time step generate multi-time step response weight data of size N×N. Using the unified feature data with embedded dimensions as the main input, the data is weighted and fused with the corresponding multi-time step response weights according to the time step. That is, the final feature at each time step is the result of a linear combination of its original features and features from all historical time steps according to the response weights. The matrix composed of all fused time step data is the temporal augmentation feature tensor, maintaining a size of N×320. This temporal augmentation feature tensor is input into a processing structure consisting of three sets of convolution-normalization-activation combinations, each containing a one-dimensional convolutional layer. The system consists of a batch normalization layer and a ReLU activation layer. The one-dimensional convolutional kernel sizes are set to 3, 5, and 3 respectively, and the number of kernels is 128, 128, and 64 respectively. The stride is uniformly 1. The batch normalization layer standardizes the distribution of each feature channel using channel normalization. The ReLU activation operation suppresses negative features to improve non-linear expressive power. The output three-dimensional tensor after this processing is the temporal embedding classification feature set, with a size of N×64×1. This feature set is then flattened along the time dimension, compressing the N×64 two-dimensional tensor into a vector of length N×64. The input is then fed into a fully connected mapping structure. The fully connected layer outputs a dimension of 5, corresponding to five types of cracks: linear cracks, curved cracks, composite cracks, network cracks, and closed cracks. The output is an initial classification vector for crack types, containing the corresponding confidence scores for each type. The above output vectors are then encapsulated and integrated with a convolutional normalization structure. The structure name, input interface parameter size (N×320), and output interface size (1×5) are declared. The inference execution order, data channel name, and external interface standard are defined. Finally, a crack type classification network model with a structural description document and complete input and output declarations is generated.

[0080] This invention effectively captures the dynamic correlation between different time points through temporal unrolling and cross-time step attention mechanisms, enhancing the model's sensitivity and expressive ability to temporal changes. Weighted fusion of multi-time step response weights strengthens the processing of temporal features, improving the richness and discriminative power of feature representation. The design of a multi-layer convolution-normalization-activation structure helps extract deep-level, multi-scale temporal features, enhancing the model's feature extraction and non-linear expressive capabilities. Fully connected mapping and network structure encapsulation ensure that the output classification results have good interpretability and adaptability, facilitating subsequent model deployment and interface calls. The overall technical solution improves the accuracy and stability of crack type identification, while enhancing the model's temporal processing capabilities and application flexibility, making it suitable for crack classification tasks in complex dynamic environments.

[0081] Preferably, step S5 includes the following steps:

[0082] Step S51: Perform time-leveling processing on the temporal evolution feature embedding data to obtain hierarchical evolution feature sequence data;

[0083] Step S52: Based on the crack type prediction results, perform label mapping on the hierarchical evolution feature sequence data to obtain crack type label sequence data;

[0084] Step S53: Reconstruct the feature channel dimension of the crack type label sequence data to obtain visualized path mapping vector data;

[0085] Step S54: Construct a crack type discrimination path based on the visualized path mapping vector data to obtain a crack type discrimination path map;

[0086] Step S55: Visualize and render the crack type discrimination path map to obtain the crack type discrimination path visualization result.

[0087] In this embodiment of the invention, the temporal evolution feature embedding data is equidistantly layered according to the time dimension, with each layer having a fixed span of 10 time steps. For each layer, five types of indicators are statistically analyzed: maximum eigenvalue, minimum eigenvalue, mean, variance, and slope. The start and end time labels of each layer are retained, resulting in a layered evolution feature sequence data containing time indexes and five-dimensional statistical features. The number of sequences is the total number of time steps divided by 10. The crack type prediction results are mapped one-to-one with the layered sequences according to the time steps, and the prediction labels are mapped to each layer's feature structure. If a layer corresponds to multiple labels, the label with the highest proportion in that layer is selected as the primary label. The output is a crack type label sequence data consisting of one crack type label for each layer, with the sequence length consistent with the number of layers. Based on this label sequence data, a channel dimension reconstruction mapping structure is constructed, encoding each layer label into a five-dimensional one-hot vector. Linear cracks, curved cracks, composite cracks, network cracks, and closed cracks correspond to positions 1 to 5, forming a unique "crack type" in the five-dimensional vector. The data is set to "1" for all positions and "0" for the rest. This data is then concatenated with the corresponding hierarchical evolution feature statistics in the column direction to obtain a visual path mapping vector data with a size of 10 times the number of hierarchies. This vector data is then input into a two-dimensional coordinate mapping construction module. The hierarchical index is used as the horizontal axis, and the unique hot vector position corresponding to each type of crack label is used as the vertical axis. A separate trajectory line is created for each category. Each trajectory connects the crack labels of the same type in adjacent hierarchies in the form of line segments. If a category is missing in an adjacent layer, the trajectory is broken. All trajectories together constitute a crack type discrimination path map. The width of each path line in the map is adjusted according to the confidence level of the label in the corresponding hierarchical layer, and the color is set to a fixed RGB value according to the crack type code. A two-dimensional image rendering tool is used to visualize the map. The coordinate axis range is fixed so that the horizontal axis equals the number of hierarchies and the vertical axis equals the total number of crack types. The background is set to white, and the path line colors are red, blue, green, orange, and purple, respectively. The transparency of each trajectory in the map is controlled to enhance the contrast. The final output image size is 1920×1080 pixels, and the image format is PNG.

[0088] This invention systematically reveals the hierarchical structure and dynamic changes in crack evolution through time-layer processing, enhancing the understanding and analysis of crack development stages. By combining crack type prediction results with label mapping, it achieves the organic integration of features and classification information, improving data expression and semantic relevance. The reconstruction of feature channel dimensions provides an efficient and structured data foundation for the subsequent generation of visualization paths, facilitating the intuitive display of crack evolution trajectories. The construction and visualization rendering of crack type discrimination path maps make the crack identification process more transparent and easier to interpret, helping users intuitively grasp the crack classification decision-making process and evolution trends. It significantly improves the interpretability and visualization effect of crack type discrimination, providing strong auxiliary support for crack monitoring and maintenance.

[0089] Preferably, the present invention also provides a deep learning-based drilling fracture type prediction system for performing the above-described deep learning-based drilling fracture type prediction method, wherein the deep learning-based drilling fracture type prediction system includes:

[0090] A generative enhancement module is used to obtain the original dataset of the crack scene; a generative adversarial network model is trained based on the original dataset of the crack scene to generate a crack enhancement sample set, resulting in an extended crack image dataset.

[0091] The feature fusion module is used to perform multimodal information fusion modeling on the extended crack image dataset to obtain multimodal crack fusion feature data;

[0092] The time-series modeling module is used to construct a time-series model of crack evolution based on multimodal crack fusion feature data, and obtain time-series evolution feature embedding data;

[0093] The classification and reasoning module is used to perform deep classification and reasoning on the embedded data of temporal evolution features to obtain crack type prediction results.

[0094] The path visualization module is used to construct a visual path based on the embedded data of temporal evolution features and the crack type prediction results, and obtain the crack type discrimination path visualization results.

[0095] Preferably, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the deep learning-based drilling fracture type prediction method.

[0096] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is not limited by the foregoing description. Thus, all changes falling within the meaning and scope of the equivalents of the application are intended to be included within the scope of the invention.

[0097] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for predicting drilling fracture types based on deep learning, characterized in that, Includes the following steps: Step S1: Obtain the original dataset of the crack scene; A generative adversarial network model is trained based on the original dataset of crack scenes to generate a crack enhancement sample set, resulting in an extended crack image dataset. Step S2: Perform multimodal information fusion modeling on the extended crack image dataset to obtain multimodal crack fusion feature data; Step S2 includes the following steps: Step S21: Extract texture features, morphological features and edge features from each sample in the extended crack image dataset to obtain a crack feature map set; Step S22: Construct a three-branch multimodal feature fusion network architecture based on the crack feature map set, and modularly encode the texture, edge and shape information of the image to obtain the branch encoded feature vector group; Step S23: Input the branch-encoded feature vector group into the preset attention-guided feature interaction component, and perform inter-channel weight allocation and spatial correlation aggregation to obtain the fused enhanced feature tensor; Step S23 includes the following steps: Step S231: Perform channel dimension expansion processing on the branch coding feature vector group to obtain the channel expanded feature vector group; Step S232: Perform channel weight response analysis based on the channel expanded feature vector group to obtain the channel attention weight map; Step S233: Perform weighted fusion on the expanded feature vector groups of the channels to obtain the channel attention-enhanced feature groups; Step S234: Perform location association encoding based on channel attention-enhanced feature groups to obtain a spatial location information encoding matrix; Step S235: Perform spatial weight allocation on the channel attention enhancement feature group to obtain the spatial attention enhancement feature group; Step S236: Perform joint aggregation of channel and spatial dimensions on the spatial attention-enhanced feature group to obtain the fused enhanced feature tensor; Step S24: Integrate and reduce the dimensionality of the fusion enhancement feature tensor and perform label mapping to generate multimodal crack fusion feature data; Step S3: Construct a crack evolution time series model based on multimodal crack fusion feature data to obtain time series evolution feature embedding data; Step S4: Perform deep classification inference on the temporal evolution feature embedding data to obtain crack type prediction results; Step S5: Construct a visualization path based on the temporal evolution feature embedding data and crack type prediction results to obtain the crack type discrimination path visualization results.

2. The drilling fracture type prediction method based on deep learning according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Obtain the raw data of the fracture scene in the drilling area and preprocess it to obtain a structured fracture image dataset; Step S12: Construct a generative adversarial network model based on the structured crack image dataset. Design a generator network to generate crack image samples, design a discriminator network to determine the authenticity of the images, and optimize the network parameters through adversarial training to obtain the generative adversarial network model. Step S13: Input the structured crack image dataset into the generative adversarial network model to generate crack enhancement sample sets under various crack morphologies, textures, and background interference conditions, and obtain crack enhancement image data. Step S14: Based on the structured crack image dataset and the crack enhancement image data, perform sample fusion, label unification and format normalization to obtain the extended crack image dataset.

3. The drilling fracture type prediction method based on deep learning according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Perform time label alignment processing on the multimodal crack fusion feature data to obtain structured time-series crack feature data; Step S32: Construct a crack state transition map model based on structured time-series crack feature data to obtain a crack state transition map; Step S33: Initialize the time series modeling of the crack state transition map to obtain the initial parameter set of crack evolution; Step S34: Train a preset long short-term memory network model based on the initial parameter set of crack evolution to obtain a crack evolution time series prediction model; Step S35: Input the multimodal crack fusion feature data into the crack evolution time series prediction model and perform feature inference to obtain the time series evolution feature embedding data.

4. The drilling fracture type prediction method based on deep learning according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Perform high-dimensional feature normalization on the temporal evolution feature embedding data to obtain feature data with unified embedding dimensions; Step S42: Construct a deep classification inference network structure based on unified feature data of embedding dimension to obtain a crack type classification network model; Step S43: Initialize and optimize the parameters of the crack type classification network model to obtain the parameter set of the deployable inference model; Step S44: Load the crack type classification network model based on the deployable inference model parameter set to obtain the loaded inference model; Step S45: Input the unified feature data of the embedded dimension into the loaded inference model and perform crack type inference to obtain crack type prediction result data.

5. The drilling fracture type prediction method based on deep learning according to claim 4, characterized in that, Step S42 includes the following steps: Step S421: Perform temporal expansion on the unified feature data of the embedding dimension to obtain the time series feature mapping matrix; Step S422: Construct a cross-time step attention mechanism based on the time series feature mapping matrix to obtain multi-time step response weight data; Step S423: Weighted fusion of the unified feature data of the embedded dimension and the multi-time response weight data is performed to obtain the temporal augmentation feature tensor; Step S424: Construct a multi-layer convolution-normalization-activation structure based on the temporal enhancement feature tensor, and extract deep feature representations to obtain temporal embedding classification feature groups; Step S425: Perform a fully connected mapping on the temporal embedded classification feature group to obtain the initial classification output vector of crack type; Step S426: Encapsulate the network structure and set the input / output interface for the initial crack type classification output vector to obtain the crack type classification network model.

6. The drilling fracture type prediction method based on deep learning according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Perform time-leveling processing on the temporal evolution feature embedding data to obtain hierarchical evolution feature sequence data; Step S52: Based on the crack type prediction results, perform label mapping on the hierarchical evolution feature sequence data to obtain crack type label sequence data; Step S53: Reconstruct the feature channel dimension of the crack type label sequence data to obtain visualized path mapping vector data; Step S54: Construct a crack type discrimination path based on the visualized path mapping vector data to obtain a crack type discrimination path map; Step S55: Visualize and render the crack type discrimination path map to obtain the crack type discrimination path visualization result.

7. A drilling fracture type prediction system based on deep learning, characterized in that, For executing the deep learning-based drilling fracture type prediction method as described in claim 1, the deep learning-based drilling fracture type prediction system comprises: A generative enhancement module is used to obtain the original dataset of the crack scene; a generative adversarial network model is trained based on the original dataset of the crack scene to generate a crack enhancement sample set, resulting in an extended crack image dataset. The feature fusion module is used to perform multimodal information fusion modeling on the extended crack image dataset to obtain multimodal crack fusion feature data; The time-series modeling module is used to construct a time-series model of crack evolution based on multimodal crack fusion feature data, and obtain time-series evolution feature embedding data; The classification and reasoning module is used to perform deep classification and reasoning on the embedded data of temporal evolution features to obtain crack type prediction results. The path visualization module is used to construct a visual path based on the embedded data of temporal evolution features and the crack type prediction results, and obtain the crack type discrimination path visualization results.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed, implements the deep learning-based drilling fracture type prediction method as described in any one of claims 1-6.

Images