A specific geological discrimination and discrimination system and method based on a multi-modal sensor

Abstract

The application provides a specific geological discrimination and discrimination system and method based on a multi-modal sensor, the system comprising a cloud computing node, a data persistent storage unit and a user interaction terminal, the data persistent storage unit being used for storing historical exploration data, geological parameters and system configurations, the user interaction terminal being used for graphical user interface display, abnormal working condition triggering, data reporting and early warning receiving, the cloud computing node being used for extracting historical geological exploration parameters and in-situ sampling parameters for topology analysis, generating a three-dimensional stratigraphic reference profile and issuing it to the user interaction terminal for graphical rendering display. Through multi-modal feature fusion technology, the application realizes accurate capture of key physical anomalies such as "drill drop" and "liquid leakage" without the need for high-cost Internet of Things hardware upgrade and reconstruction, significantly reducing the technical implementation threshold and deployment cost.

Description

Technical Field

[0001] This invention relates to the field of digital engineering exploration and data processing technology, specifically a special geological identification and discrimination system and method based on a multimodal sensor. Background Technology

[0002] In the current engineering geological exploration workflow, borehole stratigraphic logging mainly relies on manual operations by field surveyors. Technicians typically retrieve some historical borehole data from the surrounding area as a reference, combine it with soil and rock samples collected on-site, and manually delineate and enter the boundaries between soil and rock strata based on their personal professional experience.

[0003] Existing stratigraphic prediction technologies mainly employ the following schemes:

[0004] (1) Spatial interpolation methods: such as ordinary kriging interpolation, inverse distance weighted interpolation, etc., based on the stratigraphic elevation data of known borehole locations, predict the stratigraphic interface location of unknown areas.

[0005] (2) Machine learning methods: Using algorithms such as neural networks and support vector machines, based on historical exploration data, the formation distribution of new boreholes is predicted.

[0006] (3) Manual verification process: The survey personnel manually review and correct the prediction results on site.

[0007] The existing technology has the following problems:

[0008] (1) Multidimensional spatial data fitting efficiency is low and easily affected by noise: The massive historical borehole data often contains early subjective recording errors (such as coordinate offset, depth reading errors, etc.). Traditional processing methods lack an automatic verification mechanism for abnormal data, which easily introduces erroneous references into interpolation calculations. Existing smoothing and noise reduction algorithms (such as mean filtering and Gaussian smoothing) use a globally unified threshold, which cannot distinguish between "discrete human error" and "real geological mutations". It is easy to forcibly erase real special geological bodies (such as lenses, underground rivers, and open streams) as noise, causing serious geological accidents.

[0009] (2) The lack of IoT sensing in existing old equipment leads to the loss of physical anomalies: Old drilling rigs (such as XY-1A, XY-2, etc.) widely used in the field lack digital sensor interfaces. When encountering extreme physical anomalies such as drill falling into underground rivers or slurry leakage in open streams, the system cannot automatically capture key working condition characteristics such as the sudden drop rate of drill bit and the amount of flushing fluid consumed, which severely limits the input dimension of geological judgment and cannot provide a complete feature space for the formation prediction model.

[0010] (3) The algorithm is weakly adaptive to complex sites and easily loses gradual features: Conventional static interpolation algorithms use fixed weight parameters and cannot dynamically adjust the interpolation sensitivity according to the geological complexity of the site (such as the variance of stratum thickness and the frequency of lithological changes). At the same time, the lithological descriptions of strata (such as “strongly weathered granite” and “hard plastic clay”) are discrete text labels. Conventional one-hot encoding easily leads to a highly sparse feature matrix, which causes gradient vanishing and “neuron death” problems during neural network training, greatly affecting the accuracy and consistency of strata division. Summary of the Invention

[0011] To address the shortcomings of the existing technologies, this invention provides a system and method for identifying and judging special geological features based on a multimodal sensor. The aim is to achieve accurate identification of special geological features and intelligent prediction of stratigraphic boundaries by constructing a dual verification link of "spatial topology deduction + terminal human-computer interaction" without modifying the hardware of old drilling rigs.

[0012] The technical solution provided by this invention: a method for identifying and discriminating specific geological features based on a multimodal sensor, comprising the following steps:

[0013] (1) The cloud computing node obtains the spatial positioning coordinates and elevation data of the target exploration site and historical geological exploration parameters, uses the preset spatial interpolation algorithm to calculate the predicted value of the burial depth of each historical point, and then calculates the interpolation residual and extracts the candidate anomaly point set.

[0014] (2) A density-based spatial clustering algorithm is used to perform topological correlation analysis on the candidate anomaly set, calculate the number of neighboring anomalies in the ε neighborhood of each candidate point, use interpolated predicted values ​​for smooth replacement, generate corrected parameters and adjacent stratigraphic reference profile data, and confirm whether there are special topological candidate clusters.

[0015] (3) Receive structured abnormal working condition tags triggered by operators through the graphical user interface of the user interaction terminal and convert them into a working condition abnormal Boolean vector;

[0016] (4) Obtain the in-situ sampling parameters of the current borehole. The discrete text features in the in-situ sampling parameters are mapped into dense feature vectors through the entity embedding layer. Each dense feature vector is concatenated with the working condition abnormal Boolean vector to construct a complete multimodal fusion feature vector.

[0017] (5) If there is a unique topological candidate cluster and the multimodal fusion feature vector contains an activated working condition abnormal Boolean vector, the dual verification mechanism is triggered to confirm the corresponding area as a spatially discontinuous geological body, and a three-dimensional closed envelope surface is independently generated for the geological body, triggering the early warning signal to be sent to the user interaction terminal.

[0018] (6) If the dual verification mechanism is not triggered, the multimodal fusion feature vector and the reference profile data of the adjacent strata are input into the multilayer perceptron neural network model. The environmental complexity index is calculated based on the stratum thickness variance. The feature weight coefficients of the hidden layer neurons of the multilayer perceptron neural network model are dynamically adjusted through the environmental complexity. The predicted result of the stratum interface position elevation is output. The graded early warning is triggered according to the prediction result and the corresponding level signal is sent to the user interaction terminal.

[0019] Furthermore, in step (1), the parameters for the burial depth of the bearing layer are... The predicted values ​​for each point are calculated using ordinary kriging interpolation. ,

[0020] The residual verification formula is set as follows:

[0021] ,

[0022] Calculate the standard deviation of the residuals at all points. ,like If so, that point is extracted as a candidate outlier and added to the set. .

[0023] Furthermore, in step (2), the spatial clustering algorithm sets the neighborhood radius parameter ε and the minimum number of contained points MinPts, introducing a parameter adaptive adjustment mechanism:

[0024] Adaptive calculation of neighborhood radius ε:

[0025] ,

[0026] in, The distance from each candidate anomaly to its k-th nearest neighbor is the mean distance, with the coefficient k dynamically adjusted according to the site type.

[0027] Minimum number of points (MinPts) adaptive calculation:

[0028] ,

[0029] Where N is the total number of candidate outliers. If the density of outliers in the local space is less than MinPts, it is determined to be discrete distorted data. If the density in the local space is greater than or equal to MinPts, it is retained and marked as a candidate cluster of special topology.

[0030] Furthermore, in step (3), the system predefines a fixed list of working conditions, each working condition corresponds to a bit of a Boolean vector, and a Boolean vector vbool is generated according to the working condition. If the operator triggers at least one structured abnormal working condition label through the user interaction terminal, that is, if at least one dimension of the Boolean vector has a value of 1, then there is an activated abnormal working condition Boolean vector.

[0031] Furthermore, in step (4), the discrete text features in the in-situ sampling parameters are mapped into low-dimensional dense vectors by the system using an entity embedding layer. The configuration of the embedding dimension follows the following rules: ,

[0032] in, This represents the total number of categories for this discrete feature. This is the preset maximum embedding dimension.

[0033] Furthermore, the network structure configuration of the multilayer perceptron neural network model is as follows: Input layer: dimension equal to the length of the multimodal fusion feature vector; First hidden layer: 64 neurons, ReLU activation; Second hidden layer: 128 neurons, PReLU activation; Third hidden layer: 64 neurons, PReLU activation; Output layer: 1 neuron, linear activation.

[0034] Furthermore, the formula for calculating the environmental complexity index in step (6) is as follows:

[0035] ,

[0036] in This refers to the reference stratum number within the current borehole neighborhood. For the first Variance of stratum thickness For the first Mean thickness of the strata.

[0037] The dynamic weight adjustment mechanism is as follows:

[0038] in, For the first The original weight vectors of the hidden layer neurons. For the Sigmoid function, For learnable scaling parameters, This is element-wise multiplication.

[0039] Another technical solution provided by this invention: a special geological identification and discrimination system based on a multimodal sensor, which implements the above method, includes:

[0040] The persistent data storage unit is used to store historical exploration data, geological parameters, and system configuration.

[0041] User interaction terminal, used for graphical user interface display, abnormal operating condition triggering, data reporting and early warning reception;

[0042] The cloud computing node includes an anomaly topology verification module, a multimodal feature fusion module, an anomalous geological body identification module, an adaptive neural network module, and an early warning grading response module. The anomaly topology verification module calculates the spatial interpolation residuals of historical geological exploration parameters, extracts candidate anomalies, and performs topology analysis based on a density clustering algorithm. The multimodal feature fusion module maps discrete text features into dense vectors and concatenates them with anomaly Boolean vectors to construct a complete multimodal fusion feature vector. The anomalous geological body identification module performs dual verification logic using the multimodal fusion feature vector, the residuals obtained from the anomaly topology verification module, and the clustering results to confirm spatially discontinuous geological bodies and generate a three-dimensional envelope surface. The adaptive neural network module dynamically adjusts weights based on environmental complexity using the multimodal fusion feature vector and adjacent stratigraphic reference profile data to output stratigraphic interface predictions. The early warning grading response module generates graded early warning signals based on the degree of geological anomalies and the level of operational hazard.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] (1) Overcoming the hardware constraints of old drilling rigs lacking sensors: This invention transforms the engineering experience of the recording personnel into structured human-computer interaction control signals. Through multimodal feature fusion technology, it achieves accurate capture of key physical anomalies such as "drill drop and fluid leakage" without the need for expensive IoT hardware upgrades, significantly reducing the technical implementation threshold and deployment cost.

[0045] (2) Effectively prevents the misfiltering of special geological disaster characteristics: The innovative introduction of a "dual verification mechanism" combining spatial clustering topology analysis and terminal manual calibration breaks the technical bias of traditional smoothing algorithms that tend to remove abrupt geological bodies such as lenses and underground rivers as noise, and realizes the accurate identification of artificially distorted data and real special geological bodies, effectively avoiding the risk of geological exploration accidents.

[0046] (3) Significantly enhances the robustness of the prediction model to complex geological environments: The weights of the hidden layers of the neural network are dynamically adjusted with the stratigraphic spatial variation coefficient as a constraint, and the entity embedding layer is combined to solve the sparsity problem of discrete text features. The parameterized nonlinear activation function (PReLU) effectively alleviates the gradient vanishing phenomenon, deeply explores the gradual lithological laws, and objectively and accurately predicts the nodes of soil and rock transformation.

[0047] (4) Adaptive configuration of DBSCAN clustering parameters is realized: the neighborhood radius and minimum number of points threshold are dynamically adjusted according to the site type (plain, mountain, sea) and the size of candidate anomalies, avoiding subjective bias of manually setting parameters and improving the accuracy and generalization ability of anomaly identification. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall system architecture provided in the embodiments of the present invention, showing the connection relationship between cloud computing nodes, persistent data storage units and user interaction terminals, as well as the deployment location of each functional module.

[0049] Figure 2 This is a flowchart of the special geological identification and stratigraphic identification method provided in the embodiments of the present invention, which shows the complete execution path from data acquisition to the final output prediction result.

[0050] Figure 3 This is a multi-terminal interaction and closed-loop data flow timing diagram provided in the embodiments of the present invention, which shows the data interaction timing between user interaction terminals, cloud computing nodes and persistent data storage units.

[0051] Figure 4 This is a flowchart of the graded early warning response mechanism provided in the embodiments of the present invention, showing the triggering conditions and response measures for four levels of early warning. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] like Figure 1 The illustrated system for identifying and discriminating specific geological features based on a multimodal sensor mainly comprises the following core modules:

[0054] Anomaly topology verification module Cloud computing nodes Calculate the spatial interpolation residuals of historical geological exploration parameters, extract candidate anomalies, and perform topology analysis based on density clustering algorithm. Multimodal feature fusion module Cloud computing nodes Discrete text features are mapped to dense vectors and then concatenated with boolean vectors representing abnormal operating conditions in a multidimensional manner. Module for Identifying Special Geological Bodies Cloud computing nodes Perform dual verification logic to confirm spatially discontinuous geological bodies and generate a three-dimensional envelope surface. Adaptive Neural Network Module Cloud computing nodes Multilayer perceptron networks dynamically adjust weights based on environmental complexity to output stratigraphic interface predictions. Early warning graded response module Cloud computing nodes Based on the degree of geological anomalies and the level of operational hazard, graded early warning signals are generated. persistent data storage unit cloud server Stores historical exploration data, geological parameters, and system configuration. User Interaction Terminal Field operation site Graphical user interface display, abnormal condition triggering, data reporting and early warning reception

[0055] The system of this invention acquires location coordinates and retrieves historical geological parameters. It then uses interpolation residuals and DBSCAN density clustering to extract candidate anomaly clusters. DBSCAN parameters are adaptively adjusted based on site type and anomaly scale. The user terminal acquires structured anomaly condition tags triggered on-site, generating Boolean vectors. Entity embedding technology is used to convert lithological text into dense vectors, which are then concatenated into multimodal fusion features. If the anomaly topology density meets the standard and the Boolean vector is activated, it is confirmed as a unique geological body and triggers a red-level Level 1 warning; otherwise, an MLP neural network prediction is executed, dynamically adjusting the weights of hidden layers based on the stratigraphic spatial variation coefficient. The system establishes a four-level warning response mechanism, achieving a data closed loop from on-site physical anomaly recording to terminal disaster warning.

[0056] like Figure 2 The method for identifying and discriminating specific geological features based on a multimodal perceptron, as shown, includes the following steps:

[0057] Step 1: Historical Data Acquisition and Residual Extraction

[0058] The cloud computing nodes acquire the spatial coordinates and elevation data of the target exploration site, and extract historical geological exploration parameters within a preset spatial coverage area (such as a circular area with a radius of 500 meters) from the persistent data storage unit. Using a preset spatial interpolation algorithm (such as ordinary kriging interpolation), the predicted stratigraphic depth of each historical point is calculated, and then the interpolation residuals are calculated. Points with an absolute value of the interpolation residual greater than a preset standard deviation multiple (such as 3σ) are extracted as a set of candidate anomalies.

[0059] Step 2: Density clustering and topology analysis (including adaptive parameter adjustment mechanism)

[0060] A density-based spatial clustering algorithm (DBSCAN) is used to perform topological association analysis on the candidate outlier set. The DBSCAN algorithm requires setting the neighborhood radius parameter ε and the minimum number of contained points MinPts. This invention introduces a parameter adaptive adjustment mechanism:

[0061] Adaptive calculation of neighborhood radius ε:

[0062] in The distance from all candidate anomalies to their k-th nearest neighbor is the average distance (k=4 in this embodiment). The coefficient k is dynamically adjusted according to the site type: - Plains: k=1.0 (relatively uniform borehole distribution) - Mountainous and hilly areas: k=1.5 (undulating terrain leads to uneven borehole distribution) - Marine exploration: k=2.0 (sparse points, requiring a wider neighborhood range)

[0063] Minimum number of points (MinPts) adaptive calculation:

[0064] Where N is the total number of candidate outliers, ensuring that MinPts is at least 3 and grows moderately with the data size.

[0065] This adaptive mechanism is based on the engineering experience of exploration technicians: in densely drilled urban plains, a smaller neighborhood radius can effectively identify local anomalies; in sparsely drilled mountainous or marine areas, the neighborhood radius needs to be appropriately widened to avoid missed detections.

[0066] The algorithm calculates the number of neighboring anomalies in the ε-neighborhood for each candidate point. If the anomaly density in the local space is less than MinPts, it is determined to be discrete distorted data (human recording error), and the interpolated predicted value is used for smooth replacement to generate corrected parameters and reference profile data of nearby strata; if the density is greater than or equal to MinPts, it is retained and marked as a special topological candidate cluster (which may be a real geological anomaly).

[0067] Step 3: On-site multimodal data acquisition and feature stitching

[0068] Obtain the in-situ sampling parameters of the current borehole (such as lithological description, SPT blow count, groundwater level, etc.), and then map the in-situ sampling parameters (lithological description, SPT blow count, groundwater level, etc.) into dense vectors (i.e., vrock, vstate, etc. in the text) through the entity embedding layer.

[0069] At the same time, through the graphical user interface of the user interaction terminal, the system receives structured abnormal working condition labels triggered by the operators (such as "sudden drop in drill string", "fluid leakage", "stuck drill string", etc.) and converts them into Boolean vectors of abnormal working conditions.

[0070] Conversion method: The system predefines a fixed list of working conditions (e.g., [drill string descent, flushing fluid loss, stuck drill]), with each condition corresponding to a bit in a Boolean vector. When the operator clicks the corresponding touch button on the user interface terminal, the system sets the corresponding bit to 1, and the untriggered bit to 0. For example, if only a drill string descent occurs, vbool=[1,0,0] is generated.

[0071] For discrete text features (such as lithological descriptions and state descriptions) in in-situ sampling parameters, the multimodal feature fusion module uses an entity embedding layer to map them into low-dimensional dense vectors.

[0072] The mapping process can be summarized into three steps:

[0073] Step 1: Determine the dimensions:

[0074] Based on the total number of categories of this discrete feature Automatically calculate the embedding dimension, with an upper limit of 1. .

[0075] Step 2: Embedding layer pre-training

[0076] The weight matrix of the embedding layer is not randomly initialized, but is obtained by pre-training the Skip-gram model using historical exploration data. The Skip-gram model can learn the semantic relationships between lithological tags, so that tags with similar geological meanings (such as "strongly weathered granite" and "moderately weathered granite") are closer in the Euclidean distance in the embedding space, while tags with large differences in meaning (such as "clay" and "granite") are farther apart.

[0077] Step 3: Forward Mapping

[0078] In actual reasoning, for a discrete text feature (such as the lithological description "clay"), the system inputs its corresponding category index into the pre-trained embedding layer weight matrix, and extracts the vector of the corresponding row through the embedding lookup operation, thus obtaining the low-dimensional dense vector representation of the label.

[0079] The configuration of the embedding dimension follows these rules:

[0080] in This represents the total number of categories for this discrete feature. The maximum embedding dimension is preset (32 dimensions in this embodiment). This rule is based on the balance between feature expressiveness and computational efficiency: a lower dimension is used when the number of categories is small to avoid overfitting, and a higher dimension is used when the number of categories is large to ensure expressiveness.

[0081] Lithological description (e.g., clay, sand, granite, etc.) Approximately 200 categories 16-dimensional State description (e.g., rigid plastic, malleable plastic, soft plastic, etc.) Approximately 10 categories 8D Degree of weathering (e.g., complete weathering, strong weathering, moderate weathering, etc.) Approximately 6 categories 8D

[0082] The computing nodes concatenate each dense feature vector with the abnormal Boolean vector of the operating condition to construct a complete multimodal fusion feature vector.

[0083] Step 4: Specific Geological Dual Verification Mechanism

[0084] The system performs a dual-verification logic judgment: if the density of spatial topological anomalies meets the standard (i.e., there are special topological candidate clusters), and the multimodal fusion feature vector contains activated working condition anomaly Boolean vectors, then the dual-verification mechanism is triggered, the corresponding area is identified as a spatial discontinuous geological body (such as underground rivers, lenses, and open streams), and a three-dimensional closed envelope surface is independently generated for the geological body, triggering an early warning signal to be sent to the user interaction terminal.

[0085] "Activated" means that at least one dimension of the abnormal working condition Boolean vector has a value of 1, that is, the operator has triggered at least one structured abnormal working condition label through the user interaction terminal.

[0086] Specific triggering scenarios include:

[0087] (1) The operator perceives physical anomalies during the drilling process (such as the drill bit suddenly falling, the flushing fluid leaking, abnormal vibration, etc.).

[0088] (2) The operator actively clicks the corresponding touch button preset on the terminal's graphical user interface.

[0089] (3) The terminal encapsulates the triggering event into a Boolean vector and reports it to the cloud.

[0090] Step 5: Adaptive MLP Neural Network Prediction

[0091] If the dual verification mechanism is not triggered, the conventional gradual stratigraphic prediction process will proceed. The multimodal fusion feature vector and the reference profile data of the adjacent stratigraphy will be input into the multilayer perceptron (MLP) neural network model.

[0092] This invention selects a multilayer perceptron (MLP) as the core network architecture for geological formation prediction, mainly based on the following considerations:

[0093] (1) Input feature matching: The input of the formation prediction task is a structured multi-source feature vector (lithological embedding vector, working condition Boolean vector, spatial coordinates, etc.), which is tabular data. MLP has a natural advantage in processing this type of data, without the need to introduce convolutional layers to process spatial correlation or cyclic layers to process temporal dependence.

[0094] (2) Computational resource efficiency: Field survey scenarios have high real-time requirements. MLP network parameters are relatively small and forward propagation latency is low (the time for a single prediction is less than 50ms under typical configuration), making it suitable for deployment on cloud servers for high-concurrency response.

[0095] (3) Interpretability advantage: The neurons in each hidden layer of the MLP can correspond to the abstract level of geological features (such as the first layer corresponding to lithological combination features and the second layer corresponding to spatial distribution patterns), which facilitates geological experts to optimize the model and analyze the results.

[0096] The network structure is configured as follows: - Input layer: dimension equal to the length of the multimodal fusion feature vector (64 dimensions in this embodiment) - First hidden layer: 64 neurons, ReLU activation - Second hidden layer: 128 neurons, PReLU activation (adaptive negative interval slope) - Third hidden layer: 64 neurons, PReLU activation - Output layer: 1 neuron (stratum boundary elevation), linear activation

[0097] During forward propagation, the environmental complexity index is calculated based on the formation thickness variance, and the feature weight coefficients of the hidden layer neurons are dynamically adjusted accordingly, ultimately outputting the predicted elevation of the formation interface position.

[0098] "Feature weight coefficients" refer to the weighting coefficients of each dimension of the multimodal fusion feature vector by the hidden layer neurons of the MLP, i.e., the weighting coefficients in the hidden layer. weight matrix Specifically, it refers to the weighting coefficients of each neuron in the first, second, and third hidden layers for its input features.

[0099] These "features" include all dimensions of the concatenated multimodal fusion feature vector:

[0100] - Lithology embedding vector (e.g., 32-dimensional)

[0101] - State embedding vector (e.g., 32-dimensional)

[0102] - Boolean vector of abnormal operating conditions (e.g., 3D)

[0103] - and adjacent stratigraphic reference profile data

[0104] Calculation of environmental complexity index:

[0105] in This refers to the reference stratum number within the current borehole neighborhood. For the first Variance of stratum thickness For the first Mean thickness of the strata. The larger the value, the stronger the spatial variability of the geological strata at that site.

[0106] Dynamic weight adjustment mechanism:

[0107] in, For the first The original weight vectors of the hidden layer neurons. For the Sigmoid function, For learnable scaling parameters, This is element-wise multiplication. The physical meaning of this mechanism is: when the site's strata undergo drastic changes ( (Larger), with enhanced gating signals, the model redistributes its attention to features across different dimensions, improving its sensitivity to local mutation features.

[0108] Dynamic weight adjustment:

[0109] in For the original weights, The preset adjustment coefficient (e.g.) ), The adjusted weights.

[0110] This invention achieves a closed-loop collaboration between the cloud and multiple terminals. User interaction terminals report exploration coordinates and in-situ sampling data to cloud computing nodes in real time via wireless communication networks (such as 4G / 5G mobile networks or satellite communication links). The cloud computing nodes retrieve historical exploration data from the persistent data storage unit, perform topological analysis, generate a three-dimensional stratigraphic reference profile, and send it to the terminal for graphic rendering and display.

[0111] This invention establishes a hardware-software integrated dimensionality reduction feedback mechanism. To compensate for the hardware deficiency of older drilling rigs lacking digital sensors, the system encapsulates physical anomalies into structured trigger commands for the terminal. Specifically, the graphical user interface of the user terminal pre-sets touch buttons for "sudden drill string drop," "flushing fluid leakage," and "abnormal vibration." After the operator perceives a physical anomaly, they click the corresponding button. The terminal encapsulates the trigger event as a Boolean vector (e.g., [1,0,0] indicates only drill string drop occurred) and transmits it to the cloud computing node via a wireless communication link. After feature fusion is completed in the cloud, if a specific geological hazard rule is matched, the cloud immediately sends three-dimensional closed envelope bounding box data and graded early warning signals to the on-site terminal, achieving a complete data loop from "on-site physical anomaly recording → cloud topology verification → terminal hazard early warning."

[0112] This invention establishes a tiered early warning system. Based on the degree of geological anomalies and the level of operational hazard, the system establishes a four-level early warning response mechanism.

[0113] Red (Level 1) The spatial topological anomaly density meets the standard, and the Boolean vector for the abnormal operating condition is activated (double verification passed). Sound and light blocking alarm + 3D envelope surface highlighting + terminal vibration Immediately stop drilling, evacuate personnel, and activate the emergency response plan. Orange (Level 2) Spatial topological anomaly density meets the standard or operating condition anomaly Boolean vector activation (single condition satisfied). Voice prompts + yellow warning markers Slow down the drilling speed and increase the monitoring frequency Yellow (Level 3) Prediction confidence level below the threshold (e.g., <0.7) or abnormal formation thickness variation coefficient. Text prompt + question mark Manual review and supplementation of exploration points are recommended. Blue (Level 4) Normal prediction process completed Green confirmation mark Automatically fill in the cataloging form and continue the work.

[0114] The warning signal is issued in a structured JSON data packet format, which includes: warning level code, geological anomaly type description, three-dimensional envelope vertex coordinates (if present), suggested response measures text, timestamp, and other fields.

[0115] Abnormal state handling mechanism:

[0116] Network outage scenario: The user terminal has local caching capabilities and automatically retransmits any unreported data packets after the network is restored. Note: Currently, the system does not support a fully offline working mode and requires a network connection to access cloud computing services. Future versions may consider introducing edge computing nodes to enable offline degradation capabilities.

[0117] Data verification failed: The cloud performs integrity verification on the reported data. If a missing field is detected, a supplementary entry instruction is issued.

[0118] Insufficient prediction confidence: When the confidence of the neural network output is lower than the preset threshold, the system marks it as "awaiting manual review" and triggers a yellow warning.

[0119] Example 1: Complete Execution Process in a Smart Survey Scenario

[0120] This embodiment details the complete execution logic and mathematical derivation process of a multimodal sensor-based method for identifying and judging specific geological features in a certain urban rail transit survey project.

[0121] Step 1: Extraction of interpolation residuals

[0122] After acquiring the target borehole coordinates (e.g., longitude 116.405°, latitude 39.905°), the system retrieves historical borehole data from 85 locations within a 500-meter radius from the cloud computing node. This data is then used to determine the bearing stratum depth parameters. The predicted values ​​for each point are calculated using ordinary kriging interpolation. .

[0123] The residual verification formula is set as follows:

[0124] Calculate the standard deviation of the residuals at all points. ,like If so, that point is extracted as a candidate outlier and added to the set. In this embodiment, a total of 12 candidate anomalies were extracted.

[0125] Step 2: Adaptive calculation of DBSCAN parameters and topology clustering

[0126] First, the system identifies the current survey site type as "plain area" (determined by an elevation difference of less than 50 meters), and the coefficient k is set to 1.0.

[0127] Calculate the distance from each candidate outlier to its fourth nearest neighbor, and obtain... Then the neighborhood radius is:

[0128] Calculate the minimum number of contained points:

[0129] DBSCAN clustering was performed using the above adaptive parameters to calculate the number of neighboring outliers within a 12.5-meter neighborhood for each candidate point:

[0130] Point A (coordinates [116.403, 39.904]): There are 4 outliers in its neighborhood. Marked as core point

[0131] Point B (coordinates [116.408, 39.907]): Only one outlier in its neighborhood. It was determined to be discrete distorted data.

[0132] Ultimately, one unique topological candidate cluster (containing 5 cluster points) was identified, which was determined to be an area where an underground river may exist; at the same time, 7 discrete distortion points were identified, and interpolated predicted values ​​were used for smooth replacement.

[0133] Step 3: Multimodal feature fusion

[0134] On-site operators noticed that the drill string suddenly dropped 0.8 meters during drilling. They clicked the "Drill String Drop" button on the user interface terminal, which generated an abnormal Boolean vector for the working condition. .

[0135] Meanwhile, the lithological description obtained from the current borehole sampling is "clay, stiff plastic, brownish-yellow". The system extracts discrete feature categories: - Lithology: "clay", with 200 categories and embedding dimensions... However, limited by the maximum dimension of 32, 32 dimensions are actually used - State: "Hard Plastic", number of categories: 10, embedding dimension Similarly, take 32 dimensions.

[0136] The "clay" is mapped to a 32-dimensional dense vector through a pre-trained entity embedding layer. Mapping "hard plastic" to a 32-dimensional vector .

[0137] Concatenate to generate a multimodal fusion feature vector:

[0138] Step 4: Dual Verification Mechanism for Specific Geological Bodies

[0139] The cloud computing node logically correlates the spatial topology results from step 2 with the fusion vector from step 3: - Spatial topology analysis results: There is one anomalous topology candidate cluster (5 clustering anomalies), satisfying the density condition - Boolean vector of operating conditions: The first digit being 1 indicates that "drill string drop" is activated.

[0140] With both conditions met simultaneously, the system confirmed with high confidence that the area had encountered a discontinuous geological hazard body in the underground river space. The system forcibly skipped the conventional prediction branch and used the Alpha Shapes algorithm to calculate the three-dimensional morphological boundary of the anomalous cluster, generating an independent three-dimensional closed envelope surface (approximately 280 cubic meters in volume).

[0141] Step 5: Early Warning and Tiered Response

[0142] Based on the dual verification results, the early warning grading response module determines that a red level one early warning has been triggered: - Issues an audio file for sound and light blocking alarm - Issues a terminal vibration command (lasting 500ms) - Issues a three-dimensional closed envelope surface vertex coordinate array - Issues an emergency warning pop-up window with a red background: "An underground river has been detected. Please stop drilling immediately and evacuate personnel!"

[0143] After receiving the warning signal, the user terminal automatically pauses the form filling process and switches to the emergency response interface, displaying a three-dimensional geological anomaly model and evacuation route suggestions.

[0144] Example 2: Parameter Adaptation Application in Mountainous and Hilly Areas

[0145] This embodiment illustrates the application of the present invention in a mountainous highway survey project, focusing on demonstrating the actual effect of the DBSCAN parameter adaptive adjustment mechanism.

[0146] Site characteristics: The survey area is located in a mountainous area of ​​the Yunnan-Guizhou Plateau, with an elevation difference of 280 meters. The distribution of borehole points is uneven (sparse points on the ridge and dense points in the valley).

[0147] Adaptive parameter calculation:

[0148] The system identifies the site type as "mountainous and hilly" (judged by an elevation difference greater than 50 meters), and the coefficient k is set to 1.5.

[0149] Calculate the mean distance of the fourth nearest neighbor of candidate outliers: Due to the generally large spacing between boreholes in mountainous areas, .

[0150] Adaptive neighborhood radius:

[0151] Minimum number of points (18 candidate outliers in total):

[0152] Comparative analysis:

[0153] Fixed parameters (default values ​​for plains) 12.5m 1 67% (2 landslides were missed) Adaptive parameters (mountain mode) 42.5m 3 0% (All landslides successfully identified)

[0154] The comparison results show that the adaptive parameter mechanism effectively improves the accuracy of geological anomaly identification under complex mountainous terrain conditions.

[0155] Example 3: Yellow Alert Scenario with Insufficient Prediction Confidence

[0156] This embodiment demonstrates the process of triggering a yellow level 3 warning when the prediction confidence is insufficient.

[0157] Scenario description: There is limited historical borehole data for a certain site (only 5 reference points in the surrounding area), and the formation thickness varies greatly, resulting in low confidence of the neural network prediction results.

[0158] Execution process:

[0159] After the multimodal fusion feature vector is input into the MLP network, the output stratigraphic interface elevation is 15.2 meters, and the output prediction confidence is 0.58.

[0160] The warning grading response module determines that if the confidence level is 0.58 < the threshold of 0.7, a yellow level 3 warning is triggered.

[0161] The system issues an early warning signal:

[0162] The text message reads: "The prediction confidence level is low (58%), manual review is recommended."

[0163] Question mark markers superimposed on the predicted stratigraphic boundary.

[0164] Recommendation: "The number of reference boreholes in the surrounding area is insufficient. It is recommended to supplement the exploration sites or consult geological experts."

[0165] After receiving the warning, the corresponding field in the record form is marked as "pending review" and can only be submitted after manual confirmation.

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for identifying and discriminating specific geological features based on a multimodal perceptron, characterized in that, Includes the following steps: (1) The cloud computing node obtains the spatial positioning coordinates and elevation data of the target exploration site and historical geological exploration parameters, uses the preset spatial interpolation algorithm to calculate the predicted value of the burial depth of each historical point, and then calculates the interpolation residual and extracts the candidate anomaly point set. (2) A density-based spatial clustering algorithm is used to perform topological correlation analysis on the candidate anomaly set, calculate the number of neighboring anomalies in the ε neighborhood of each candidate point, use interpolated predicted values ​​for smooth replacement, generate corrected parameters and adjacent stratigraphic reference profile data, and confirm whether there are special topological candidate clusters. (3) Receive structured abnormal working condition tags triggered by operators through the graphical user interface of the user interaction terminal and convert them into a working condition abnormal Boolean vector; (4) Obtain the in-situ sampling parameters of the current borehole. The discrete text features in the in-situ sampling parameters are mapped into dense feature vectors through the entity embedding layer. Each dense feature vector is concatenated with the working condition abnormal Boolean vector to construct a complete multimodal fusion feature vector. (5) If there is a unique topological candidate cluster and the multimodal fusion feature vector contains an activated working condition abnormal Boolean vector, the dual verification mechanism is triggered to confirm the corresponding area as a spatially discontinuous geological body, and a three-dimensional closed envelope surface is independently generated for the geological body, triggering the early warning signal to be sent to the user interaction terminal. (6) If the dual verification mechanism is not triggered, the multimodal fusion feature vector and the reference profile data of the adjacent strata are input into the multilayer perceptron neural network model. The environmental complexity index is calculated based on the stratum thickness variance. The feature weight coefficients of the hidden layer neurons of the multilayer perceptron neural network model are dynamically adjusted through the environmental complexity. The predicted result of the stratum interface position elevation is output. The graded early warning is triggered according to the prediction result and the corresponding level signal is sent to the user interaction terminal.

2. The method for identifying and discriminating specific geological features based on a multimodal sensor according to claim 1, characterized in that, In step (1), the parameters for the burial depth of the bearing layer are... The predicted values ​​for each point are calculated using ordinary kriging interpolation. , The residual verification formula is set as follows: ， Calculate the standard deviation of the residuals at all points. ,like If so, that point is extracted as a candidate outlier and added to the set. .

3. The method for identifying and discriminating specific geological features based on a multimodal perceptron according to claim 1, characterized in that, In step (2), the spatial clustering algorithm sets the neighborhood radius parameter ε and the minimum number of contained points MinPts, and introduces a parameter adaptive adjustment mechanism: Adaptive calculation of neighborhood radius ε: ， in, The distance from each candidate anomaly to its k-th nearest neighbor is the mean distance, with the coefficient k dynamically adjusted according to the site type. Minimum number of points (MinPts) adaptive calculation: ， Where N is the total number of candidate outliers. If the density of outliers in the local space is less than MinPts, it is determined to be discrete distorted data. If the density in the local space is greater than or equal to MinPts, it is retained and marked as a candidate cluster of special topology.

4. The method for identifying and discriminating specific geological features based on a multimodal perceptron according to claim 1, characterized in that, In step (3), the system predefines a fixed list of working conditions. Each working condition corresponds to a bit of a Boolean vector. A Boolean vector vbool is generated based on the working condition. If the operator triggers at least one structured abnormal working condition label through the user interaction terminal, that is, if at least one dimension of the Boolean vector has a value of 1, then there is an activated abnormal working condition Boolean vector.

5. The method for identifying and discriminating specific geological features based on a multimodal sensor according to claim 1, characterized in that, In step (4), the discrete text features in the in-situ sampling parameters are mapped into low-dimensional dense vectors by the multimodal feature fusion module using an entity embedding layer. The configuration of the embedding dimension follows the following rules: ， in, This represents the total number of categories for this discrete feature. This is the preset maximum embedding dimension.

6. The method for identifying and discriminating specific geological features based on a multimodal perceptron according to claim 1, characterized in that, The network structure configuration of the multilayer perceptron neural network model is as follows: Input layer: dimension equal to the length of the multimodal fusion feature vector; First hidden layer: 64 neurons, ReLU activation; Second hidden layer: 128 neurons, PReLU activation; Third hidden layer: 64 neurons, PReLU activation; Output layer: 1 neuron, linear activation.

7. The method for identifying and discriminating specific geological features based on a multimodal perceptron according to claim 1, characterized in that, The formula for calculating the environmental complexity index in step (6) is as follows: ， in This refers to the reference stratum number within the current borehole neighborhood. For the first Variance of stratum thickness For the first Mean thickness of the strata. The dynamic weight adjustment mechanism is as follows: ， in, For the first The original weight vectors of the hidden layer neurons. For the Sigmoid function, For learnable scaling parameters, This is element-wise multiplication.

8. A special geological identification and discrimination system based on a multimodal sensor, implementing the method described in any one of claims 1-7, characterized in that, include: The persistent data storage unit is used to store historical exploration data, geological parameters, and system configuration. User interaction terminal, used for graphical user interface display, abnormal operating condition triggering, data reporting and early warning reception; The cloud computing node includes an anomaly topology verification module, a multimodal feature fusion module, an anomalous geological body identification module, an adaptive neural network module, and an early warning grading response module. The anomaly topology verification module calculates the spatial interpolation residuals of historical geological exploration parameters, extracts candidate anomalies, and performs topology analysis based on a density clustering algorithm. The multimodal feature fusion module maps discrete text features into dense vectors and concatenates them with anomaly Boolean vectors to construct a complete multimodal fusion feature vector. The anomalous geological body identification module performs dual verification logic using the multimodal fusion feature vector, the residuals obtained from the anomaly topology verification module, and the clustering results to confirm spatially discontinuous geological bodies and generate a three-dimensional envelope surface. The adaptive neural network module dynamically adjusts weights based on environmental complexity using the multimodal fusion feature vector and adjacent stratigraphic reference profile data to output stratigraphic interface predictions. The early warning grading response module generates graded early warning signals based on the degree of geological anomalies and the level of operational hazard.

Images