An AI core drilling identification and automatic cataloging system for outdoor exploration operation robots

Abstract

The present invention discloses an AI drill core recognition and automatic cataloging system for field survey robots, which relates to the field of geological survey technology and solves the technical problems of difficulty in analyzing different feature data of drill cores and constructing fused feature values, and difficulty in constructing an AI model containing multiple feature branches to identify drill cores. The system includes the following modules: a multi-source data acquisition module for installing cameras and multiple sensors to collect corresponding data; a sensor collaborative work module for the main control system to establish a clock source, synchronize camera and sensor data acquisition, and add tags; a data processing and analysis module for analyzing the three types of data separately and constructing corresponding feature values; an AI recognition module for constructing an AI model containing texture, spectrum, and mechanics branches, and deploying it after feature fusion and cross entropy loss training to output category probability distribution; and an automatic cataloging module for automatically generating cataloging information after the AI model recognizes the drill core.

Description

Technical Field

[0001] The present invention belongs to the technical field of geological exploration, and in particular relates to an AI drill core recognition and automatic cataloging system for an outdoor exploration operation robot. Background Art

[0002] In geological field surveys, the collection and analysis of drill core samples is an important means of obtaining geological information. With the development of artificial intelligence technology, it has become possible to use AI technology to realize the automatic identification and cataloging of drill cores. An AI drill core identification and automatic cataloging system for field survey robots collects features associated with drill cores through multiple devices, analyzes and constructs corresponding feature values, inputs them into the AI recognition model for training, and deploys the trained model into the system to realize automatic identification of drill cores.

[0003] However, traditional drill core processing processes often rely on manual classification, labeling and cataloging, which is not only inefficient but also prone to errors and omissions due to human factors. Although the existing drill core processing and identification meet the drill core identification needs to a certain extent, there are still problems such as low recognition accuracy and incomplete cataloging information. It cannot meet the needs of actual survey operations. It is difficult to query the corresponding radius value in the mapping table through the comprehensive texture complexity coefficient, and use the automatically adjusted radius to extract texture features of the drill core image. It is also difficult to analyze the spectral and mechanical characteristics of the drill core and construct corresponding feature values. There is a lack of building an AI model that includes texture, spectrum and mechanical branches to automatically classify drill core types. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes an AI drill core recognition and automatic cataloging system for an outdoor survey robot, which is used to solve the following technical problems:

[0005] Traditional drill core processing processes often rely on manual classification, labeling, and cataloging, which is not only inefficient but also prone to errors and omissions due to human factors. Although the existing drill core processing and identification meet the needs of drill core identification to a certain extent, there are still problems such as low recognition accuracy and incomplete cataloging information. It cannot meet the needs of actual survey operations. It is difficult to query the corresponding radius value in the mapping table through the comprehensive texture complexity coefficient and use the automatically adjusted radius to extract texture features of the drill core image. It is also difficult to analyze the spectral and mechanical characteristics of the drill core and construct the corresponding feature values. There is a lack of building an AI model that includes texture, spectrum, and mechanical branches to automatically classify drill core types.

[0006] To solve the above problems, the present invention provides an AI drill core recognition and automatic cataloging system for field survey robots, comprising the following modules:

[0007] Multi-source data acquisition module: A high-resolution camera is installed on the field survey robot to collect images of the drill core in real time. A hyperspectral sensor and a six-dimensional force sensor are integrated into the same device to collect corresponding data.

[0008] Sensor collaboration module: The main control system establishes a clock source, synchronizes camera and sensor data collection and adds markers, starts the robot according to the motion plan, and triggers the acquisition when the core drill reaches the sensor;

[0009] Data processing and analysis module: Analyzes drill core images, determines the relationship between the comprehensive texture complexity coefficient and the LBP radius, uses the automatically adjusted radius to extract texture features from drill core images, analyzes spectral curves, combines multiple features to obtain fused spectral features, analyzes force and torque signals, and integrates time domain features and frequency domain mechanical features;

[0010] AI recognition module: This module builds a unified timestamp coordinate system for drill core samples, aligns data and performs Z-score normalization. It then constructs an AI model encompassing texture, spectrum, and mechanics branches. After training with feature fusion and cross-entropy loss, it is deployed and outputs a category probability distribution.

[0011] Automatic cataloging module: After the AI model identifies the drill core, it automatically generates cataloging information, builds a database to store relevant information, and designs a user interface to display the drill core identification information.

[0012] As a further solution of the present invention, a high-resolution camera is installed on an outdoor survey robot to collect images of the drill core in real time, and a hyperspectral sensor and a six-dimensional force sensor are integrated into the same device to collect corresponding data, including the following steps:

[0013] A high-resolution camera was fixed with an anti-vibration bracket, and the camera's power and data transmission cables were connected to the robot's main control system. The installation location of the hyperspectral sensor was determined based on the measured distance and angle between the sensor and the drill core. The power and signal output cables of the hyperspectral sensor were connected to the robot's main control system, and a corresponding analog-to-digital conversion module was added to the main control system to convert the signals into digital signals. A six-dimensional force sensor was installed at the junction of the drill bit and the robotic arm, and its signal cable was connected to the robot's main control system.

[0014] Initialize the camera, hyperspectral sensor, and six-dimensional force sensor in the main control system.

[0015] As a further solution of the present invention, the main control system establishes a clock source, synchronizes the camera and sensor data acquisition and adds markers, starts the robot according to the motion plan, and triggers the acquisition when the core is drilled to the sensor, including the following steps:

[0016] A clock source is established in the main control system. Using the clock synchronization signal, the camera and various sensors start collecting data at the same time, adding synchronization markers to the collected data. The survey robot is activated according to the preset motion plan. When the drill core reaches the sensor's measurement position, the high-resolution camera, hyperspectral sensor, and six-dimensional force sensor are triggered to begin collecting data.

[0017] The high-resolution camera continuously captures images of the drill core at a set frame rate, and the image data is transmitted to the main control system in the form of digital signals via a data transmission line. After receiving the trigger signal, the hyperspectral sensor collects spectral information on the drill core surface. The six-dimensional force sensor converts the measured force and torque signals into electrical signals.

[0018] The storage format of image data, spectral data and mechanical data is determined in the main control system, and metadata is added, including acquisition time, location and drill core number information.

[0019] As a further solution of the present invention, the analysis of the drill core image comprises the following steps:

[0020] The collected drill core image is preprocessed, including denoising, white balance adjustment, and image enhancement. The color image is converted into a grayscale image, the gray-level co-occurrence matrix of the drill core image is calculated, and texture features, including contrast, correlation, and energy, are extracted. After the extracted texture features are normalized, multiple texture feature parameters are combined using weighted summation to obtain a first texture complexity coefficient.

[0021] At the same time, the gradient vector of each pixel point in the drill core image is calculated to obtain the gradient matrix. The structure tensor is obtained by smoothing the gradient matrix. The structure tensor is decomposed into two eigenvalues. The eigenvalues of multiple pixel points in different directions are averaged to obtain the mean of the eigenvalues in the corresponding direction. The variance of the eigenvalue in each direction is calculated, the variance in each direction is sorted, and the corresponding weight coefficients are assigned in order from large to small. Weighted addition is performed to obtain the second texture complexity coefficient.

[0022] As a further solution of the present invention, determining the relationship between the comprehensive texture complexity coefficient and the LBP radius and using the automatically adjusted radius to extract texture features from the drill core image includes the following steps:

[0023] The first texture complexity coefficient and the second texture complexity coefficient are averaged to obtain the comprehensive texture complexity coefficient. Based on the experimental data, the relationship between the comprehensive texture complexity coefficient and the LBP radius is determined. A coefficient-radius mapping table is established to record the optimal LBP radius corresponding to each comprehensive texture complexity coefficient.

[0024] According to the calculated comprehensive texture complexity coefficient, the corresponding radius value is queried in the mapping table, and the automatically adjusted radius is used to extract texture features from the drill core image.

[0025] As a further solution of the present invention, the analysis of the spectral curve comprises the following steps:

[0026] Determine that the spectrum covers a range of 350-2500 nanometers, obtain a complete spectral curve for each acquisition point, record its wavelength and corresponding reflectance, and perform preprocessing operations on the spectral data, including noise removal and baseline correction;

[0027] Use Gaussian kernel filters of different sizes to perform convolution operations on the spectral curve, detect peaks at different scales, integrate the detection results of each scale, and determine the final absorption peak position by weighted averaging; after determining the absorption peak position, take the reflectivity of that position as the starting point, and obtain the absorption peak depth along the vertical distance to the baseline, and then standardize the absorption peak depth; search for points where the reflectivity reaches half-height along the wavelength axis on both sides of the absorption peak, and record the wavelength difference between these two points as the half-height full width. Use the same method to obtain one-tenth full width and one-quarter full width, and determine the weight coefficients corresponding to the three width features based on experimental experience, and give weighted addition to obtain the comprehensive width feature.

[0028] As a further solution of the present invention, the combining of multiple features to obtain a fused spectral feature comprises the following steps:

[0029] Traverse the entire spectral curve data, find all the positions identified as absorption peaks, determine whether the distribution of absorption peak positions is uniform, determine the corresponding mapping function based on whether the distribution is uniform, and convert the absorption peak positions into values between 0 and 1;

[0030] The weight coefficients corresponding to the position, depth and width features are determined, and the sum of the three weight coefficients is equal to 1. The position, depth and width features are weighted and added according to the determined weight coefficients to obtain the fused spectral features.

[0031] As a further solution of the present invention, the analysis of the force and torque signals and the fusion of the time domain characteristics and the frequency domain characteristics of the mechanical characteristics include the following steps:

[0032] After preprocessing the force and torque signals collected by the six-dimensional force sensor, the statistical features of the signals, including the mean, impact factor, and energy ratio, are extracted. The values of the three features are mapped between 0 and 1 using the Min-Max normalization method, and the corresponding weights of the three features are determined. Weighted addition is performed to obtain the fused feature values. The waveform features of the signal are extracted, including the skewness coefficient and kurtosis coefficient of the waveform. The two coefficients are divided by the corresponding standard deviation to obtain the standardized skewness coefficient and kurtosis coefficient. Weighted addition is performed to obtain the fused waveform features, which are then combined with the fused statistical features to obtain the fused time domain vector.

[0033] Perform FTT transformation on the preprocessed force and torque signals to obtain the signal's spectrum representation, extract the frequency components and amplitudes, determine the effective frequency range, and obtain the overall signal kurtosis value based on the spectrum graph. The frequency components, their amplitudes, and the overall kurtosis value are fused to construct a feature vector containing multiple features. This feature vector is normalized and each eigenvalue in the normalized feature vector is fused through weighted fusion to obtain the fused frequency domain feature.

[0034] The fused time domain features and frequency domain features are weighted and added together to obtain the fused mechanical features.

[0035] As a further solution of the present invention, the AI model comprising texture, spectrum and mechanics branches is constructed and deployed after feature fusion and cross entropy loss training to output category probability distribution, including the following steps:

[0036] A unified timestamp coordinate system was established for each drill core sample. The texture image and spectral data were aligned within a ±50ms time window based on the triggering moment of the mechanical sensor. The extracted texture, spectral, and mechanical feature dimensions were then Z-score normalized.

[0037] Constructing the AI model architecture: The input layer includes a texture branch (224*224*3 RGB image input), a spectral branch (420-dimensional spectral curve), and a mechanical branch (6-dimensional force signal * 1000 time steps). The texture network uses an improved ResNet18 architecture, with the last fully connected layer changed to a 512-dimensional output. The spectral network uses a 1D convolutional layer and adds a spectral attention module. The mechanical network uses a bidirectional LSTM layer and adds a temporal attention mechanism. The feature fusion layer uses a cross-modal attention mechanism to concatenate high-order features from each modality to form a 1024-dimensional joint feature vector.

[0038] The weighted cross-entropy loss was selected as the loss function for model training. In the first stage, the texture network was frozen and the spectral or mechanical branch was trained. In the second stage, the entire network was unfrozen. End-to-end training was performed using the Lookahead optimizer with k=5 and α=0.5. The trained AI model was deployed to the actual drill core recognition task, and the model output was the probability distribution of each category.

[0039] As a further solution of the present invention, the AI model automatically generates catalog information after identifying the drill core, builds a database to store relevant information, and designs a user interface to display the drill core identification information, including the following steps:

[0040] Automatically generate catalog information for drill cores based on the AI model's recognition results. This catalog information includes the drill core's category, source, feature description, and recognition date, and supports custom catalog fields and templates.

[0041] Design a user interface to display the drill core images, spectral data, mechanical characteristics and identification results, provide editing and viewing functions for catalog information, and allow users to modify and supplement the catalog information.

[0042] Beneficial effects of the present invention:

[0043] The present invention determines the LBP radius by integrating the texture complexity coefficient, which can make the scale of texture feature extraction more accurately adapt to the texture changes in the image. A fixed radius may not provide the optimal feature representation under different texture complexities. By dynamically adjusting the radius according to the integrated texture complexity coefficient, the features extracted by LBP can better reflect the actual texture structure of the image, thereby providing more discriminative feature vectors for subsequent classification, recognition or analysis tasks.

[0044] The present invention sets three types of feature data related to drill cores through the input layer. The feature fusion layer adopts a cross-modal attention mechanism to splice the high-order features of each modality to form a 1024-dimensional joint feature vector, and optimizes the model performance through a phased training strategy. Setting a weighted cross-entropy loss can assign different weights to different categories, so that the model pays more attention to those difficult-to-identify or minority class samples during the training process, avoiding the model's biased learning towards majority class samples due to class imbalance, thereby improving the model's generalization ability on different data sets. When the model is deployed in actual drill core recognition tasks, it can better adapt to various unseen data situations and ensure a high recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 It is a module flow chart of the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0048] See also Figure 1 As shown, the present invention is an AI drill core recognition and automatic cataloging system for field survey robots, including the following modules:

[0049] Multi-source data acquisition module: A high-resolution camera is installed on the field survey robot to collect images of the drill core in real time. A hyperspectral sensor and a six-dimensional force sensor are integrated into the same device to collect corresponding data.

[0050] Sensor collaboration module: The main control system establishes a clock source, synchronizes camera and sensor data collection and adds markers, starts the robot according to the motion plan, and triggers the acquisition when the core drill reaches the sensor;

[0051] Data processing and analysis module: Analyzes drill core images, determines the relationship between the comprehensive texture complexity coefficient and the LBP radius, uses the automatically adjusted radius to extract texture features from drill core images, analyzes spectral curves, combines multiple features to obtain fused spectral features, analyzes force and torque signals, and integrates time domain features and frequency domain mechanical features;

[0052] AI recognition module: This module builds a unified timestamp coordinate system for drill core samples, aligns data and performs Z-score normalization. It then constructs an AI model encompassing texture, spectrum, and mechanics branches. After training with feature fusion and cross-entropy loss, it is deployed and outputs a category probability distribution.

[0053] Automatic cataloging module: After the AI model identifies the drill core, it automatically generates cataloging information, builds a database to store relevant information, and designs a user interface to display the drill core identification information.

[0054] Specifically, a camera and multiple sensors are installed in the system respectively, a synchronization signal generator module is created in the main control system, and the module is configured so that it can generate a synchronization pulse signal according to the signal of the clock source, the storage format of image data, spectral data and mechanical data is determined in the main control system, and a metadata database or data structure is created in the main control system to store information such as acquisition time, location and drill core number; the image data collected by the camera is analyzed, the first and second texture complexity coefficients are calculated, and the comprehensive texture complexity coefficient is obtained by adding them, and a corresponding mapping table is constructed with the LBP radius. After determining the optimal radius, the texture characteristics of the drill core are analyzed; the spectral data collected by the spectral sensor is analyzed, and the pre-processed position, depth and width features are weighted and added according to the determined weight coefficient to obtain the fused spectral characteristics; the force and torque signals collected by the six-dimensional force sensor are analyzed, and the analyzed time domain characteristics and frequency domain characteristics are fused to obtain the final mechanical characteristics; the three characteristics are used as the input of the model to construct an AI model that can identify the drill core type; and an automatic cataloging system is designed based on the recognition structure of the AI model.

[0055] In one embodiment of the present invention, the high-resolution camera is installed on an outdoor survey robot to collect images of the drill core in real time, and a hyperspectral sensor and a six-dimensional force sensor are integrated into the same device to collect corresponding data, including the following steps:

[0056] A high-resolution camera was fixed with an anti-vibration bracket, and the camera's power and data transmission cables were connected to the robot's main control system. The installation location of the hyperspectral sensor was determined based on the measured distance and angle between the sensor and the drill core. The power and signal output cables of the hyperspectral sensor were connected to the robot's main control system, and a corresponding analog-to-digital conversion module was added to the main control system to convert the signals into digital signals. A six-dimensional force sensor was installed at the junction of the drill bit and the robotic arm, and its signal cable was connected to the robot's main control system.

[0057] Initialize the camera, hyperspectral sensor, and six-dimensional force sensor in the main control system.

[0058] Specifically, check whether the anti-seismic bracket firmly fixes the high-resolution camera to ensure that it will not be displaced due to vibration and other factors during subsequent operations, and reconfirm that the camera's power cord and data transmission line are correctly and firmly connected to the robot's main control system. After the installation is complete, according to the configuration file or setting interface provided by the driver, according to the measurement distance and measurement angle requirements between the hyperspectral sensor and the drill core, carefully confirm whether its installation position is correct, and confirm that it is successfully connected to the main control system. In the main control system, find the corresponding analog-to-digital conversion module configuration interface, set various parameters for analog-to-digital conversion, so that it can correctly convert the analog signal output by the hyperspectral sensor into a digital signal, check whether the six-dimensional force sensor is correctly installed at the connection between the drill bit and the robotic arm, and whether the installation is firm and reliable, and send instructions through the main control system to read the data of the six-dimensional force sensor.

[0059] The initialization settings for the camera, hyperspectral sensor, and six-dimensional force sensor in the main control system are shown in Table 1:

[0060] Table 1.

[0061]

[0062] In one embodiment of the present invention, the main control system establishes a clock source, synchronizes the camera and sensor data acquisition and adds markers, starts the robot according to the motion plan, and triggers the acquisition when the core is drilled to the sensor, including the following steps:

[0063] A clock source is established in the main control system. Using the clock synchronization signal, the camera and various sensors start collecting data at the same time, adding synchronization markers to the collected data. The survey robot is activated according to the preset motion plan. When the drill core reaches the sensor's measurement position, the high-resolution camera, hyperspectral sensor, and six-dimensional force sensor are triggered to begin collecting data.

[0064] The high-resolution camera continuously captures images of the drill core at a set frame rate, and the image data is transmitted to the main control system in the form of digital signals via a data transmission line. After receiving the trigger signal, the hyperspectral sensor collects spectral information on the drill core surface. The six-dimensional force sensor converts the measured force and torque signals into electrical signals.

[0065] The storage format of image data, spectral data and mechanical data is determined in the main control system, and metadata is added, including acquisition time, location and drill core number information.

[0066] Specifically, connect the clock source device to the main control system through a suitable interface, and install the corresponding driver so that it can be recognized by the main control system. Enter the operating system of the main control system, find the clock source setting option, and set the clock source parameters, including frequency and phase, according to the technical manual of the clock source. Create a synchronization signal generator module in the main control system and configure the module so that it can generate a synchronization pulse signal according to the signal of the clock source. According to the requirements of the field survey operation, use the robot programming language to write a motion planning program. Define the robot's motion trajectory, speed, acceleration and other parameters in detail in the program to ensure that it can accurately move to the core collection position according to the preset path. Integrate the written motion planning program into the robot's control system. After confirming that the motion planning program and the robot control system are working properly, control the robot through the control interface of the main control system or the remote control terminal. A start command is sent to start the movement of the outdoor survey robot. The position sensor is enabled in the main control system to monitor the position of the drill core. When the drill core approaches the measurement position of the sensor, the position sensor can promptly feedback the position information to the main control system. According to the information fed back by the position sensor, when the drill core reaches the set measurement position, the main control system sends a trigger signal to the high-resolution camera, hyperspectral sensor and six-dimensional force sensor through the control line. According to the type of data acquisition and the requirements of subsequent processing, the storage format of image data, spectral data and mechanical data is determined in the main control system. A metadata database or data structure is created in the main control system to store information such as acquisition time, location and drill core number. When a group of data is acquired, the corresponding metadata is associated with the group of data. For example, this information can be recorded in the header of the data file or through a separate metadata file.

[0067] In one embodiment of the present invention, the analyzing of the drill core image comprises the following steps:

[0068] The collected drill core image is preprocessed, including denoising, white balance adjustment, and image enhancement. The color image is converted into a grayscale image, the gray-level co-occurrence matrix of the drill core image is calculated, and texture features, including contrast, correlation, and energy, are extracted. After the extracted texture features are normalized, multiple texture feature parameters are combined using weighted summation to obtain a first texture complexity coefficient.

[0069] At the same time, the gradient vector of each pixel point in the drill core image is calculated to obtain the gradient matrix. The structure tensor is obtained by smoothing the gradient matrix. The structure tensor is decomposed into two eigenvalues. The eigenvalues of multiple pixel points in different directions are averaged to obtain the mean of the eigenvalues in the corresponding direction. The variance of the eigenvalue in each direction is calculated, the variance in each direction is sorted, and the corresponding weight coefficients are assigned in order from large to small. Weighted addition is performed to obtain the second texture complexity coefficient.

[0070] Specifically, according to the selected conversion algorithm, each pixel in the image is traversed, and its RGB value is converted into a grayscale value to obtain a grayscale image. The direction and distance for calculating the grayscale co-occurrence matrix are selected. The directions include 0°, 45°, 90° and 135°, and the distance is 1 pixel. For a given direction and distance, the number of pixel pairs that meet the corresponding position relationship in the grayscale image is counted to construct a grayscale co-occurrence matrix. The contrast, correlation and energy are calculated using the corresponding calculation formulas respectively. The extracted texture features are normalized using methods such as min-max normalization, including contrast, correlation and energy texture features. The normalized texture feature values are obtained, and the contrast weight is set to 0.5, the correlation weight is set to 0.3, and the energy weight is set to 0.2. , multiply the normalized texture feature value by the corresponding weight coefficient, and then add the results of each feature to obtain the first texture complexity coefficient. For each pixel in the grayscale image, calculate its gradient vector, combine the gradient vectors of each pixel, construct a gradient matrix, select Gaussian filtering and other methods to smooth the gradient matrix to obtain a smoothed gradient matrix. According to the smoothed gradient matrix, calculate the structure tensor, perform eigenvalue decomposition on the structure tensor, and obtain two eigenvalues λ1 and λ2. Average the eigenvalues of multiple pixels in the 0° and 90° directions respectively to obtain the mean of the eigenvalues in the corresponding directions. According to the mean of the eigenvalues in each direction, calculate the variance of the eigenvalues in each direction, and sort the variances of the eigenvalues in each direction from large to small. Then assign weight coefficients according to the sorting results, with a weight of 0.6 for the first direction and a weight of 0.4 for the second direction. Multiply the variance in each direction by the corresponding weight coefficient, and then add the results in each direction to obtain the second texture complexity coefficient.

[0071] The first texture complexity coefficient calculation formula:

[0072]

[0073] in, is the first texture complexity coefficient, 、 and are the contrast, correlation and energy after normalization, 、 and is the corresponding weight coefficient;

[0074] The second texture complexity coefficient calculation formula:

[0075]

[0076] in, is the second texture complexity coefficient, and is the variance of the eigenvalue in each direction, and is the corresponding weight coefficient.

[0077] In one embodiment of the present invention, determining the relationship between the comprehensive texture complexity coefficient and the LBP radius, and extracting texture features from the drill core image using the automatically adjusted radius, comprises the following steps:

[0078] The first texture complexity coefficient and the second texture complexity coefficient are averaged to obtain the comprehensive texture complexity coefficient. Based on the experimental data, the relationship between the comprehensive texture complexity coefficient and the LBP radius is determined. A coefficient-radius mapping table is established to record the optimal LBP radius corresponding to each comprehensive texture complexity coefficient.

[0079] According to the calculated comprehensive texture complexity coefficient, the corresponding radius value is queried in the mapping table, and the automatically adjusted radius is used to extract texture features from the drill core image.

[0080] Specifically, a series of representative drill core image samples are collected, covering different texture complexities and annotated with category labels, including low complexity, medium complexity and high complexity. For each image sample, a cross-validation SVM classifier is trained to determine the optimal LBP radius: for each candidate radius R∈ , extract the LBP histogram features of the image, use 5-fold cross validation to train the SVM classifier, calculate the classification accuracy of each fold validation set, take the average of the classification accuracy to get the average accuracy as the score of the radius, compare the average accuracy of different radii, and select the radius with the highest average accuracy as the optimal LBP radius. At the same time, calculate the comprehensive texture complexity coefficient of each sample image according to the previous steps, record (comprehensive texture complexity coefficient, optimal LBP radius), draw a scatter plot of the two, observe whether there is interval division or continuous relationship, and find that the two are obviously divided into intervals. Equal-width bins are divided according to the range of the comprehensive texture complexity coefficient, and the samples are assigned to the corresponding intervals according to the bin boundaries. The mode of the optimal LBP radius in each interval is counted, and a coefficient-radius mapping table is constructed.

[0081] The mapping coefficient-radius mapping table is shown in Table 2:

[0082] Table 2

[0083]

[0084] In one embodiment of the present invention, the analyzing of the spectral curve comprises the following steps:

[0085] Determine that the spectrum covers a range of 350-2500 nanometers, obtain a complete spectral curve for each acquisition point, record its wavelength and corresponding reflectance, and perform preprocessing operations on the spectral data, including noise removal and baseline correction;

[0086] Use Gaussian kernel filters of different sizes to perform convolution operations on the spectral curve, detect peaks at different scales, integrate the detection results of each scale, and determine the final absorption peak position by weighted averaging; after determining the absorption peak position, take the reflectivity of that position as the starting point, and obtain the absorption peak depth along the vertical distance to the baseline, and then standardize the absorption peak depth; search for points where the reflectivity reaches half-height along the wavelength axis on both sides of the absorption peak, and record the wavelength difference between these two points as the half-height full width. Use the same method to obtain one-tenth full width and one-quarter full width, and determine the weight coefficients corresponding to the three width features based on experimental experience, and give weighted addition to obtain the comprehensive width feature.

[0087] Specifically, during the acquisition process, ensure that the spectrometer can cover the spectral range of 350-2500 nanometers, and obtain a complete spectral curve for each acquisition point, perform median filtering on the collected spectral data in the wavelength direction to remove pulse noise and spike noise, and subtract the baseline value obtained by fitting from the original spectral data to obtain the corrected spectral data, select different sizes from 3×3 to 11×11, with a step size of 2, and perform convolution operation on the spectral curve with a Gaussian kernel of each size. For each data point on the spectral curve, multiply the value of the corresponding adjacent data point according to the weight coefficient of the Gaussian kernel and then sum them to obtain the convolution result. In the convolution results of different scales, detect the peaks respectively, and perform weighted averaging on the detection results of different scales according to experimental experience or pre-set rules to determine the absorption peak position. Finally, taking the reflectivity at that position as the starting point, the absorption peak depth is obtained along the vertical direction to the baseline. All detected absorption peak depths are standardized to eliminate the influence of overall reflectivity differences or other factors between different samples. The reflectivity reaches half-height positions along the wavelength axis on both sides of the absorption peak, and the wavelength difference between the two points is recorded as the half-height full width. Similarly, the points where the reflectivity reaches one-tenth and one-quarter of the total absorption peak depth are found, and the wavelength difference is calculated to obtain the one-tenth full width and one-quarter full width. The weight coefficients corresponding to the three width features are determined based on experimental experience. The half-height full width weight is 0.5, the one-tenth full width weight is 0.3, and the one-quarter full width weight is 0.2. Each width feature is multiplied by the corresponding weight coefficient and added together to obtain the comprehensive width feature.

[0088] Comprehensive width feature calculation formula:

[0089]

[0090] in, is the comprehensive width feature, when When it is 1, 2, and 3, it corresponds to the three width features of half-height full width, one-tenth of the height full width, and one-quarter of the height full width. For the Width features, For the The weight coefficient corresponding to the width feature.

[0091] In one embodiment of the present invention, the combining of multiple features to obtain a fused spectral feature comprises the following steps:

[0092] Traverse the entire spectral curve data, find all the positions identified as absorption peaks, determine whether the distribution of absorption peak positions is uniform, determine the corresponding mapping function based on whether the distribution is uniform, and convert the absorption peak positions into values between 0 and 1;

[0093] The weight coefficients corresponding to the position, depth and width features are determined, and the sum of the three weight coefficients is equal to 1. The position, depth and width features are weighted and added according to the determined weight coefficients to obtain the fused spectral features.

[0094] Specifically, along the wavelength axis direction of the spectral curve, with a certain step size set to 1 nanometer, each data point is checked in turn. If the reflectivity of the point is less than that of the surrounding neighboring points and the absorption peak reaches a certain depth threshold, it is marked as an absorption peak position, and the wavelength value of the position is recorded. The wavelength difference between adjacent absorption peak positions is calculated to obtain a wavelength difference data set. These wavelength difference data are statistically analyzed. If the coefficient of variation is small, it is considered that the absorption peak position distribution is relatively uniform; otherwise, it is uneven. If the absorption peak position distribution is uniform, a linear mapping function can be used to convert the absorption peak position to 0 to 1 interval; if the absorption peak position is unevenly distributed, a mapping based on the cumulative distribution function is adopted. First, the cumulative distribution function of the wavelengths of all absorption peak positions is calculated. Then, for the wavelength of each absorption peak position, the mapping to the interval of 0 to 1 is completed by finding its corresponding value in the CDF. According to experimental experience or theoretical analysis, the weight coefficient range of the position, depth and width features is preliminarily set. The weight coefficient corresponding to the position feature is 0.4, and the depth feature and width feature are both 0.3. The preprocessed position, depth and width features are weighted and added according to the determined weight coefficient to obtain the fused spectral features.

[0095] In one embodiment of the present invention, the analysis of the force and torque signals and the integration of the time domain characteristics and the frequency domain characteristics of the mechanical characteristics include the following steps:

[0096] After preprocessing the force and torque signals collected by the six-dimensional force sensor, the statistical features of the signals, including the mean, impact factor, and energy ratio, are extracted. The values of the three features are mapped between 0 and 1 using the Min-Max normalization method, and the corresponding weights of the three features are determined. Weighted addition is performed to obtain the fused feature values. The waveform features of the signal are extracted, including the skewness coefficient and kurtosis coefficient of the waveform. The two coefficients are divided by the corresponding standard deviation to obtain the standardized skewness coefficient and kurtosis coefficient. Weighted addition is performed to obtain the fused waveform features, which are then combined with the fused statistical features to obtain the fused time domain vector.

[0097] Perform FTT transformation on the preprocessed force and torque signals to obtain the signal's spectrum representation, extract the frequency components and amplitudes, determine the effective frequency range, and obtain the overall signal kurtosis value based on the spectrum graph. The frequency components, their amplitudes, and the overall kurtosis value are fused to construct a feature vector containing multiple features. This feature vector is normalized and each eigenvalue in the normalized feature vector is fused through weighted fusion to obtain the fused frequency domain feature.

[0098] The fused time domain features and frequency domain features are weighted and added together to obtain the fused mechanical features.

[0099] Specifically, all data points are added together and divided by the number of data points to obtain the mean vector, and the corresponding eigenvalues are calculated according to the calculation formulas of the impact factor and the energy ratio. The minimum and maximum normalization formula is used to map the values of the three features of mean, impact factor and energy ratio to between 0 and 1. According to experimental experience or relevant field knowledge, the weight coefficients corresponding to the three features of mean, impact factor and energy ratio are determined. The standardized features are added to the weights to obtain the fused statistical features. For each force and torque signal, its skewness coefficient and kurtosis coefficient are calculated and divided by the corresponding standard deviation to obtain the standardized skewness coefficient and standardized kurtosis coefficient. According to experimental experience or relevant field knowledge, the weight coefficients corresponding to the skewness coefficient and kurtosis coefficient are determined. The eigenvalues of the standardized skewness coefficient and the standardized kurtosis coefficient are multiplied by the corresponding weight coefficients, and then added to obtain the fused waveform features. The fused The statistical features after the fusion and the waveform features after fusion are combined to obtain a fused time domain feature vector; the preprocessed force and torque signals are fast Fourier transformed to obtain the signal's spectrum representation, and the main frequency components of the signal and their corresponding amplitudes are determined according to the spectrum diagram. A frequency range containing the main frequency components and a high energy ratio is selected as the effective frequency interval, and the overall kurtosis value of the signal is calculated according to the spectrum diagram. A feature vector F_frequency=[frequency component 1, amplitude 1, frequency component 2, amplitude 2, ..., overall kurtosis value] containing multiple features is constructed, and the feature vector is normalized to determine the weight coefficient corresponding to each feature in the feature vector. Then, each eigenvalue in the normalized feature vector is multiplied by the corresponding weight coefficient, and the fused frequency domain features are obtained by adding them together. The weight coefficients corresponding to the fused time domain and frequency domain features are both 0.5.

[0100] In one embodiment of the present invention, the construction of an AI model including texture, spectral, and mechanical branches, and deployment after feature fusion and cross-entropy loss training to output category probability distribution includes the following steps:

[0101] A unified timestamp coordinate system was established for each drill core sample. The texture image and spectral data were aligned within a ±50ms time window based on the triggering moment of the mechanical sensor. The extracted texture, spectral, and mechanical feature dimensions were then Z-score normalized.

[0102] Constructing the AI model architecture: The input layer includes a texture branch (224*224*3 RGB image input), a spectral branch (420-dimensional spectral curve), and a mechanical branch (6-dimensional force signal * 1000 time steps). The texture network uses an improved ResNet18 architecture, with the last fully connected layer changed to a 512-dimensional output. The spectral network uses a 1D convolutional layer and adds a spectral attention module. The mechanical network uses a bidirectional LSTM layer and adds a temporal attention mechanism. The feature fusion layer uses a cross-modal attention mechanism to concatenate high-order features from each modality to form a 1024-dimensional joint feature vector.

[0103] The weighted cross-entropy loss was selected as the loss function for model training. In the first stage, the texture network was frozen and the spectral or mechanical branch was trained. In the second stage, the entire network was unfrozen. End-to-end training was performed using the Lookahead optimizer with k=5 and α=0.5. The trained AI model was deployed to the actual drill core recognition task, and the model output was the probability distribution of each category.

[0104] Specifically, the input layer settings in the construction of the AI model include texture branch, spectral branch and mechanical branch. The texture branch sets the input size to 224×224×3 RGB image input, the spectral branch sets the input size to 420-dimensional spectral curve, and the mechanical branch sets the input size to 6-dimensional force signal × 1000 time steps; the network architecture construction includes texture network, spectral network and mechanical network. The texture network uses the improved ResNet18 architecture, and changes the last fully connected layer to 512-dimensional output. The spectral network uses a 1D convolution layer to process the spectral curve, and adds a spectral attention module to automatically assign different weights according to the importance of different wavelength points of the spectral data. The mechanical network uses a bidirectional LSTM layer to process the mechanical signal and adds a time attention mechanism; the feature fusion layer uses a cross-modal attention mechanism for feature fusion, and automatically learns the relationship and weights between different modal features. In this way, the high-order features output by the three branches are concatenated into a 1024-dimensional joint feature vector. The weighted cross entropy loss is selected as the loss function, and the Lookahead optimizer is used for model training. The trained model is then deployed to the actual drill core recognition task to obtain the probability distribution results for each category, which facilitates user decision-making and analysis.

[0105] In one embodiment of the present invention, the AI model automatically generates catalog information after identifying the drill core, builds a database to store relevant information, and designs a user interface to display the drill core identification information, including the following steps:

[0106] Automatically generate catalog information for drill cores based on the AI model's recognition results. This catalog information includes the drill core's category, source, feature description, and recognition date, and supports custom catalog fields and templates.

[0107] Design a user interface to display the drill core images, spectral data, mechanical characteristics and identification results, provide editing and viewing functions for catalog information, and allow users to modify and supplement the catalog information.

[0108] Specifically, the recognition results of the drill core are obtained from the trained AI model, and the category with the highest probability is selected as the category of the drill core. At the same time, the source information of the drill core, such as sampling location, sampling depth, etc., is extracted from the original data. Based on the previously extracted and standardized feature data, a feature description of the drill core is generated, and a cataloging information template is designed, including fixed fields such as "category", "source", "feature description", "identification date" and customizable cataloging fields. The appropriate data type and format are set for each field. According to the cataloging information template and the extracted information, the cataloging information of the drill core is automatically generated; a user interface is designed, which is divided into four main areas: image display area, spectral data display area, and force The image display area supports the loading and display of drill core images in multiple image formats, such as JPEG, PNG, etc. The spectral data display area displays the spectral data of the drill core in a graphical manner, with the horizontal axis representing the wavelength and the vertical axis representing the spectral intensity. The mechanical feature display area displays the mechanical feature data of the drill core in a graphical form. The recognition result display area displays the recognition results of the AI model in an intuitive manner, such as category name and recognition probability. In the catalog information table, click on a catalog field to pop up an edit box, allowing the user to modify the field content and set the "Save" button. After the user completes the catalog information modification, click Save to store the updated catalog information in the database.

[0109] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An AI drill core recognition and automatic cataloging system for field survey robots, characterized in that: Includes the following modules: Multi-source data acquisition module: A high-resolution camera is installed on the field survey robot to collect images of the drill core in real time. A hyperspectral sensor and a six-dimensional force sensor are integrated into the same device to collect corresponding data. Sensor collaboration module: The main control system establishes a clock source, synchronizes camera and sensor data collection and adds markers, starts the robot according to the motion plan, and triggers the acquisition when the core drill reaches the sensor; Data processing and analysis module: Analyzes drill core images, determines the relationship between the comprehensive texture complexity coefficient and the LBP radius, uses the automatically adjusted radius to extract texture features from drill core images, analyzes spectral curves, combines multiple features to obtain fused spectral features, analyzes force and torque signals, and integrates time domain features and frequency domain mechanical features; AI recognition module: This module builds a unified timestamp coordinate system for drill core samples, aligns data and performs Z-score normalization. It then constructs an AI model encompassing texture, spectrum, and mechanics branches. After training with feature fusion and cross-entropy loss, it is deployed and outputs a category probability distribution. Automatic cataloging module: After the AI model identifies the drill core, it automatically generates cataloging information, builds a database to store relevant information, and designs a user interface to display the drill core identification information; The analysis of the drill core image comprises the following steps: The collected drill core image is preprocessed, including denoising, white balance adjustment, and image enhancement. The color image is converted into a grayscale image, the gray-level co-occurrence matrix of the drill core image is calculated, and texture features, including contrast, correlation, and energy, are extracted. After the extracted texture features are normalized, multiple texture feature parameters are combined using weighted summation to obtain a first texture complexity coefficient. At the same time, the gradient vector of each pixel in the drill core image is calculated to obtain a gradient matrix. The gradient matrix is smoothed to obtain a structure tensor. The structure tensor is subjected to eigenvalue decomposition to obtain two eigenvalues. The eigenvalues of multiple pixels in different directions are averaged to obtain the mean of the eigenvalues in the corresponding direction. The variance of the eigenvalues in each direction is calculated, the variances in each direction are sorted, and the corresponding weight coefficients are assigned in descending order. The weighted addition is performed to obtain the second texture complexity coefficient. Determining the relationship between the comprehensive texture complexity coefficient and the LBP radius, and extracting texture features from the drill core image using the automatically adjusted radius, comprises the following steps: The first texture complexity coefficient and the second texture complexity coefficient are averaged to obtain a comprehensive texture complexity coefficient.

2. The AI drill core recognition and automatic cataloging system for outdoor survey robots according to claim 1 is characterized in that: The method includes installing a high-resolution camera on an outdoor survey robot to collect images of the drill core in real time, and integrating a hyperspectral sensor and a six-dimensional force sensor on the same device to collect corresponding data, including the following steps: A high-resolution camera was fixed with an anti-vibration bracket, and the camera's power and data transmission cables were connected to the robot's main control system. The installation location of the hyperspectral sensor was determined based on the measured distance and angle between the sensor and the drill core. The power and signal output cables of the hyperspectral sensor were connected to the robot's main control system, and a corresponding analog-to-digital conversion module was added to the main control system to convert the signals into digital signals. A six-dimensional force sensor was installed at the junction of the drill bit and the robotic arm, and its signal cable was connected to the robot's main control system. Initialize the camera, hyperspectral sensor, and six-dimensional force sensor in the main control system.

3. The AI drill core recognition and automatic cataloging system for outdoor survey robots according to claim 1 is characterized in that: The main control system establishes a clock source, synchronizes the camera and sensor data collection and adds markers, starts the robot according to the motion plan, and triggers the collection when the core is drilled to the sensor, including the following steps: A clock source is established in the main control system. Using the clock synchronization signal, the camera and various sensors start collecting data at the same time, adding synchronization markers to the collected data. The survey robot is activated according to the preset motion plan. When the drill core reaches the sensor's measurement position, the high-resolution camera, hyperspectral sensor, and six-dimensional force sensor are triggered to begin collecting data. The high-resolution camera continuously captures images of the drill core at a set frame rate, and the image data is transmitted to the main control system in the form of digital signals via a data transmission line. After receiving the trigger signal, the hyperspectral sensor collects spectral information on the drill core surface. The six-dimensional force sensor converts the measured force and torque signals into electrical signals. The storage format of image data, spectral data and mechanical data is determined in the main control system, and metadata is added, including acquisition time, location and drill core number information.

4. The AI drill core recognition and automatic cataloging system for outdoor survey robots according to claim 1 is characterized in that: The determining of the relationship between the comprehensive texture complexity coefficient and the LBP radius, and extracting texture features from the drill core image using the automatically adjusted radius, further includes: Based on the experimental data, the relationship between the comprehensive texture complexity coefficient and the LBP radius is determined, and a coefficient-radius mapping table is established to record the optimal LBP radius corresponding to each comprehensive texture complexity coefficient; According to the calculated comprehensive texture complexity coefficient, the corresponding radius value is queried in the mapping table, and the automatically adjusted radius is used to extract texture features from the drill core image.

5. The AI drill core recognition and automatic cataloging system for outdoor survey robots according to claim 1 is characterized in that: The analysis of the spectral curve comprises the following steps: Determine that the spectrum covers a range of 350-2500 nanometers, obtain a complete spectral curve for each acquisition point, record its wavelength and corresponding reflectance, and perform preprocessing operations on the spectral data, including noise removal and baseline correction; Use Gaussian kernel filters of different sizes to perform convolution operations on the spectral curve, detect peaks at different scales, integrate the detection results of each scale, and determine the final absorption peak position by weighted averaging; after determining the absorption peak position, take the reflectivity of that position as the starting point, and obtain the absorption peak depth along the vertical distance to the baseline, and then standardize the absorption peak depth; search for points where the reflectivity reaches half-height along the wavelength axis on both sides of the absorption peak, and record the wavelength difference between these two points as the half-height full width. Use the same method to obtain one-tenth full width and one-quarter full width, and determine the weight coefficients corresponding to the three width features based on experimental experience, and give weighted addition to obtain the comprehensive width feature.

6. The AI drill core recognition and automatic cataloging system for outdoor survey robots according to claim 1 is characterized in that: The method of combining multiple features to obtain a fused spectral feature comprises the following steps: Traverse the entire spectral curve data, find all the positions identified as absorption peaks, determine whether the distribution of absorption peak positions is uniform, determine the corresponding mapping function based on whether the distribution is uniform, and convert the absorption peak positions into values between 0 and 1; The weight coefficients corresponding to the position, depth and width features are determined, and the sum of the three weight coefficients is equal to 1. The position, depth and width features are weighted and added according to the determined weight coefficients to obtain the fused spectral features.

7. The AI drill core recognition and automatic cataloging system for outdoor survey robots according to claim 1 is characterized in that: The analysis of the force and torque signals and the integration of the time domain characteristics and the frequency domain characteristics of the mechanical characteristics include the following steps: After preprocessing the force and torque signals collected by the six-dimensional force sensor, the statistical features of the signals, including the mean, impact factor, and energy ratio, are extracted. The values of the three features are mapped between 0 and 1 using the Min-Max normalization method, and the corresponding weights of the three features are determined. Weighted addition is performed to obtain the fused feature values. The waveform features of the signal are extracted, including the skewness coefficient and kurtosis coefficient of the waveform. The two coefficients are divided by the corresponding standard deviation to obtain the standardized skewness coefficient and kurtosis coefficient. Weighted addition is performed to obtain the fused waveform features, which are then combined with the fused statistical features to obtain the fused time domain vector. Perform FTT transformation on the preprocessed force and torque signals to obtain the signal's spectrum representation, extract the frequency components and amplitudes, determine the effective frequency range, and obtain the overall signal kurtosis value based on the spectrum graph. The frequency components, their amplitudes, and the overall kurtosis value are fused to construct a feature vector containing multiple features. This feature vector is normalized and each eigenvalue in the normalized feature vector is fused through weighted fusion to obtain the fused frequency domain feature. The fused time domain features and frequency domain features are weighted and added together to obtain the fused mechanical features.

8. The AI drill core recognition and automatic cataloging system for outdoor survey robots according to claim 1 is characterized in that: The construction of an AI model that includes texture, spectrum, and mechanics branches, and deploying it after feature fusion and cross-entropy loss training to output category probability distribution, includes the following steps: A unified timestamp coordinate system was established for each drill core sample. The texture image and spectral data were aligned within a ±50ms time window based on the triggering moment of the mechanical sensor. The extracted texture, spectral, and mechanical feature dimensions were then Z-score normalized. Constructing the AI model architecture: The input layer includes a texture branch (224*224*3 RGB image input), a spectral branch (420-dimensional spectral curve), and a mechanical branch (6-dimensional force signal * 1000 time steps). The texture network uses an improved ResNet18 architecture, with the last fully connected layer changed to a 512-dimensional output. The spectral network uses a 1D convolutional layer and adds a spectral attention module. The mechanical network uses a bidirectional LSTM layer and adds a temporal attention mechanism. The feature fusion layer uses a cross-modal attention mechanism to concatenate high-order features from each modality to form a 1024-dimensional joint feature vector. The weighted cross-entropy loss was selected as the loss function for model training. In the first stage, the texture network was frozen and the spectral or mechanical branch was trained. In the second stage, the entire network was unfrozen. End-to-end training was performed using the Lookahead optimizer with k=5 and α=0.

5. The trained AI model was deployed to the actual drill core recognition task, and the model output was the probability distribution of each category.

9. The AI drill core recognition and automatic cataloging system for outdoor survey robots according to claim 1 is characterized in that: After the AI model identifies the drill core, it automatically generates catalog information, builds a database to store relevant information, and designs a user interface to display the drill core identification information, including the following steps: Automatically generate catalog information for drill cores based on the AI model's recognition results. This catalog information includes the drill core's category, source, feature description, and recognition date, and supports custom catalog fields and templates. Design a user interface to display the drill core images, spectral data, mechanical characteristics and identification results, provide editing and viewing functions for catalog information, and allow users to modify and supplement the catalog information.

Images