Anomaly analysis method for mineral images within boreholes

Abstract

The application discloses an abnormality analysis method for mineral images in drill holes, comprising the following steps: firstly, collecting spectral images of drill holes in multiple mining areas, and constructing a training set through pre-processing and data augmentation guided by physical knowledge; secondly, constructing an interpretable network model containing a physical constraint attention and a feature decoupling double branch, and introducing an adversarial domain invariant learning and physical consistency verification module; training the model by jointly optimizing the classification, domain adversarial, physical consistency and feature decoupling loss; finally, applying the trained model to new mining area data, realizing mineral identification by fusing data-driven prediction and physical matching degree, and synchronously outputting explanatory evidence such as key band attention map and matching standard spectrum, and automatically identifying abnormal samples needing review based on confidence difference and domain feature strength; the application effectively improves the cross-mining area generalization ability of mineral identification, the interpretability of the decision-making process and the automation level of abnormal perception.

Description

Technical Field

[0001] This invention relates to the field of digital rock engineering, and more particularly to an anomaly analysis method for mineral images inside boreholes. Background Technology

[0002] In deep mineral resource exploration and engineering geological surveys, borehole wall images obtained through borehole imaging technology are a crucial data source for revealing the mineral composition and structure of underground rock masses. To obtain clear original rock mass information, borehole washing is necessary in actual operations to remove drilling mud covering the borehole wall. Therefore, the borehole wall images processed in this invention specifically refer to hyperspectral or visible-near-infrared spectral images obtained after borehole washing. However, even after washing, borehole wall images from different mining areas and different boreholes still face significant challenges. First, due to differences in mineralization geological background, wall rock alteration type, post-modification intensity, and the degree of microfracture development, the spectral characteristics of similar minerals in different exploration areas may exhibit significant but not essential shifts, i.e., cross-mineralization area shifts. Second, on-site imaging conditions and the extremely thin residual material adhering to the borehole wall, which is difficult to completely avoid, further distort the spectral curves and introduce noise. Existing technologies typically design feature extraction and classification algorithms for single mining areas or limited data, or employ simple image enhancement and threshold segmentation. These methods are insufficient to adaptively overcome the complex and variable cross-mining area interference, resulting in weak generalization ability of mineral identification models and a sharp decline in accuracy when applied to new mining areas.

[0003] While existing deep learning-based mineral identification methods can automatically extract features, they are mostly end-to-end black-box models. Their internal decision-making processes are opaque, failing to provide physical and spectral evidence for classifying a region as a specific mineral. This makes it difficult for geologists to verify and trust automated results, hindering their widespread application in practical exploration decisions. Furthermore, existing methods typically output only a mineral category label and a single confidence level, lacking inherent measures of the reliability of the identification results and anomaly warning mechanisms. When faced with mineral types exhibiting blurred spectral features, complex mineral coexistence, or completely unknown mineral types in post-well cleaning images, the model's incorrect classifications often carry high confidence levels, easily leading to misinterpretations. Summary of the Invention

[0004] This application provides an anomaly analysis method for mineral images inside boreholes, which is adaptable to changes across mining areas, allows for interpretable decision-making, and can automatically detect and alert to geological anomalies.

[0005] This application provides a method for anomaly analysis of borehole mineral images, including: S1, collects borehole spectral image data from multiple mining areas, performs preprocessing and physical knowledge augmentation, and constructs a multi-mining area spectral training dataset; S2 constructs an interpretable network model that includes a physical constraint attention mechanism and a feature decoupling dual branch, and introduces adversarial domain invariant learning and physical consistency verification. S3, the network model is trained by jointly optimizing the classification loss, domain adversarial loss, physical consistency loss and feature decoupling loss; S4 applies the trained model to borehole data from new mining areas to perform mineral identification and anomaly detection, and outputs mineral categories, fusion confidence scores, and interpretive evidence.

[0006] S5 automatically identifies anomalous samples that require review based on confidence level differences and domain feature strength.

[0007] Preferably, in step S1, the step of collecting borehole spectral image data from multiple mining areas and performing physical knowledge augmentation further includes: S11, based on the mineral profiles of identified boreholes in the region, construct a structured regional geological knowledge map, and summarize standard vertical mineral sequence templates through sequence alignment and clustering; S12, Sequence encoding is performed on the mineral identification profile of a single hole to extract a depth feature vector containing contextual information; S13, the extracted sequence depth features are matched and compared with the standard template library to identify abnormal patterns, and the geological rules in the knowledge graph are called to perform reasoning to generate potential geological genesis hypotheses; S14: Extract the original spectral data corresponding to the sequence anomaly segment, label it with anomaly type and geological hypothesis, and generate special augmented samples to be fed back into the training set.

[0008] Preferably, in step S2, the construction of an interpretable network model comprising a physical constraint attention mechanism and a feature decoupling dual branch includes: constructing a shared encoder to extract spectral features; constructing a mineral essential feature branch, which includes a spectral attention module with physical knowledge constraints to weight spectral channels; constructing a mining area feature branch to capture spectral shift features related to a specific mining area; designing a physical consistency verification module to calculate the similarity between the essential features and a standard spectral library; introducing a domain discriminator network to extract mining area information from the essential features through adversarial training; and concatenating the essential features with the domain features, inputting them into a classifier, and outputting a mineral category prediction.

[0009] Preferably, in step S3, the joint optimization of classification loss, domain adversarial loss, physical consistency loss, and feature decoupling loss includes: calculating the classification loss using the cross-entropy loss function; setting up an adversarial training mechanism, where the domain discriminator attempts to distinguish the source mining area of ​​the feature, and the mineral essential feature branch maximizes the error rate of the discriminator; calculating the physical consistency loss based on the output of the physical consistency verification module, encouraging the essential features to be similar to the standard spectral features; calculating the mutual information estimate between the mineral essential features and the mining area features, or minimizing their correlation through the discriminator, as the feature decoupling loss; and weighting and merging the above loss functions into a total loss, and optimizing the model parameters through backpropagation.

[0010] Preferably, in step S4, applying the trained model to borehole data in the new mining area for mineral identification and anomaly detection includes: inputting the preprocessed data from the new mining area into the model to obtain the initial predicted probability of the mineral category, the physical matching score, the attention weight, and the domain feature vector; weighting and fusing the initial predicted probability and the physical matching score to obtain the final fused confidence probability, and selecting the category with the highest confidence as the identification result; outputting the key band attention map and the best matching standard spectrum as explanatory evidence; and automatically identifying abnormal samples that need to be reviewed based on the final fused confidence, the difference between the initial predicted probability and the physical matching score, and the magnitude of the domain feature vector.

[0011] Preferably, in S11, the construction of a structured regional geological knowledge map specifically includes: defining mineral types, lithological types, alteration types, mineralization stages, and geological tectonic events as entities; and establishing symbiotic relationships, hierarchical relationships, genetic relationships, and spatiotemporal associations among entities based on regional geological data and expert knowledge.

[0012] Preferably, in step S12, the step of sequential encoding the mineral identification profile of a single hole specifically includes: mapping the mineral category label into a continuous feature vector through an embedding layer to form a two-dimensional feature matrix; inputting the feature matrix into a sequence encoding model to extract depth context features; and aggregating the output of the sequence encoding model to obtain a fixed-length depth feature vector.

[0013] Preferably, in step S13, identifying abnormal patterns includes: calculating the similarity between sequence depth features and a standard template, marking segments with similarity below a threshold as candidate abnormal sequences; identifying inverted sequences, missing bands, or abnormal symbiotic patterns in the candidate abnormal sequences using a sequence alignment algorithm; and invoking predefined geological rules in the knowledge graph to reason about the identified abnormal patterns, generate one or more potential geological genesis hypotheses, and calculate their confidence levels.

[0014] Preferably, in step S14, generating special augmented samples and feeding them back to the training set specifically includes: extracting corresponding spectral data blocks from the original spectral dataset based on the depth coordinates of the anomalous sections; generating triplet labels for each data block containing mineral category, sequence anomalous type, and geological hypothesis to form augmented samples; merging the labeled augmented samples into the multi-mining area spectral training dataset; highlighting the sequence anomalous sections on the mineral histogram and outputting an explanation card containing the anomalous type and geological hypothesis.

[0015] One or more technical solutions provided in this application have at least the following technical effects or advantages: By employing a data augmentation and feature decoupling network guided by physical knowledge, this method can automatically and accurately identify mineral composition from borehole wall images exhibiting significant cross-mining-region spectral shifts. Its core advantages lie in the fact that the model not only possesses strong cross-mining-region generalization ability and robustness, effectively overcoming interference from different mining environments and instruments, but also, by integrating physical spectral matching and an interpretable attention mechanism, makes the automated decision-making process transparent and reliable. Furthermore, it can simultaneously output key evidence and automatically identify low-confidence anomalous samples, greatly enhancing the method's practicality and reliability. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an anomaly analysis method for borehole mineral images according to an embodiment of the present invention. Detailed Implementation

[0017] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of the present invention may be more thorough and complete.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terminology used herein includes any and all combinations of one or more of the associated listed items.

[0019] Example 1: Figure 1 This is a flowchart illustrating an anomaly analysis method for borehole mineral images according to an embodiment of the present invention.

[0020] like Figure 1 As shown, an anomaly analysis method for borehole mineral images includes the following steps: S1 aggregates borehole spectral image data from multiple mining areas, performs relative reflectance conversion and band alignment, and augments the processed data with physical knowledge.

[0021] The physical knowledge augmentation involves performing the following augmentation operations on each training sample (i.e., the spectral curve of a specific mineral) to simulate cross-mineral region migration and enhance essential features: First, domain migration simulation applies multiplicative noise of random intensity to the spectral curve to simulate random fluctuations in reflectance intensity caused by different lighting conditions and surface roughness. Slight, nonlinear band shifts (micro-shifts along the wavelength axis) and stretching / compressions are applied to the entire spectral curve to simulate minor calibration errors or wavelength positioning deviations between different instruments. A small number of background spectra from other mining areas (such as typical spectra of surrounding rocks and alteration products) are randomly selected and linearly mixed with the current sample spectrum with low weights to simulate environmental material adhesion or mixed pixel effects. Second, essential feature enhancement determines the range of key diagnostic absorption or reflectance characteristic bands (e.g., the location of absorption valleys for specific ions) of the mineral category to which the current sample belongs, based on known mineral standard spectral databases. Constraints are imposed on the range of the located diagnostic bands during all the above domain migration simulation operations. To ensure that the added noise intensity is lower, the translation and stretching amplitude is smaller, and the background spectral weighting is weaker within this specific band range, the spectral morphology characteristics (such as absorption depth, width, and symmetry) of these key regions are not destructively distorted, so that they can still maintain a stable physical identification function after augmentation.

[0022] Specifically, raw visible-near-infrared spectral image data from boreholes in multiple target mining areas were collected. For the data from each mining area, relative reflectance conversion (using standard whiteboard reference data collected at the same time, the original DN values ​​(digital quantization values) were converted into relative reflectance data to eliminate the influence of the instrument's own dark current and the absolute difference in light intensity) and instrument response correction and band alignment (based on the spectral response function of each spectrometer, the spectral range of all data was resampled and unified to the same band sequence and spectral resolution to ensure that data from different sources can be directly compared in the spectral dimension).

[0023] For each preprocessed mining area data, based on geological logging or core analysis results, each spectral pixel or spectral unit (such as the average spectrum of a homogeneous mineral region) is labeled with its corresponding mineral category tag. Simultaneously, the mining area tag from which each sample originates is recorded. This forms a structured training sample set. For each training sample in the set (i.e., a labeled spectral curve), domain shift simulation and essential feature enhancement are performed sequentially.

[0024] After completing the above augmentation operations on all the original training samples, the newly generated large number of augmented samples are merged with the original samples to form the final multi-mining area spectral training dataset used for model training, which has been standardized and enhanced with physical knowledge guidance.

[0025] The data, which compiles borehole spectral image data from multiple mining areas, also includes: S11, based on the mineral profiles of identified boreholes in the region, constructs a structured regional geological knowledge map, and summarizes standard vertical mineral sequence templates through sequence alignment and clustering.

[0026] Specifically, it collects data from multiple borehole profiles within the target area that have been identified as minerals. Each profile contains information on mineral categories distributed along the borehole depth and their corresponding depth coordinates.

[0027] Define the entity types of the knowledge graph, including mineral types, lithological types, alteration types, mineralization stages, and geological tectonic events. Based on regional geological data, ore deposit models, and expert knowledge, establish relationships between entities, such as symbiotic relationships, stratigraphic relationships, genetic correlations, and spatiotemporal associations, forming a machine-readable structured knowledge network.

[0028] From the input borehole profiles, sections representing typical mineralization models or normal alteration zoning are selected based on geological understanding. The mineral-depth sequences of these sections are extracted and standardized, such as by normalization based on depth or by zoning according to geological significance.

[0029] Sequence alignment algorithms (such as dynamic time warping or its variants) are applied to compare and analyze multiple extracted mineral sequences, identifying common and stable mineral occurrence orders, transition relationships, and relative thickness patterns among different sequences. Highly similar sequences are grouped into the same pattern class using clustering methods (such as hierarchical clustering). For each pattern class, a standard vertical mineral sequence template is generated, explicitly recording the mineral type, occurrence order, possible thickness range, and tolerance for variation. Finally, these templates and their feature representations (such as embedding vectors and state transition probability matrices) are associated with corresponding nodes in the knowledge graph and stored in a standard sequence template library.

[0030] S12, the mineral identification profile of a single hole is sequentially encoded, converted into sequential data, and input into the sequence model to extract a deep feature vector containing contextual information.

[0031] Specifically, the mineral identification results profile output from the target borehole is read. This profile consists of a series of data points along the borehole depth direction, with each data point containing depth coordinates and its corresponding mineral category label.

[0032] Discrete mineral category labels are mapped to continuous, fixed-dimensional feature vectors through an embedding layer, thereby transforming the entire borehole profile into a two-dimensional feature matrix, where rows correspond to depth order and columns correspond to the dimensions of the embedding vectors.

[0033] The aforementioned feature matrix is ​​input into a sequence encoding model, such as a multi-layer Transformer encoder or a bidirectional gated recurrent unit network. This model performs computation along the depth sequence dimension, capturing the dependencies and contextual information between mineral features at each depth point and their adjacent depth point features through self-attention mechanisms or recurrent connections.

[0034] Typically, the output of the last layer of the sequence coding model is extracted at a specific position (such as the position corresponding to a special classification token added at the beginning of the sequence), or the output of all depth positions of the last layer is globally pooled (such as average pooling), and aggregated into a fixed-length, high-dimensional depth feature vector. This vector integrates the mineral composition and sequence structure information of the entire borehole profile.

[0035] S13 involves matching and comparing the extracted sequence depth features with a standard template library to identify abnormal patterns, and then using geological rules from the knowledge graph to perform reasoning and generate potential geological origin hypotheses.

[0036] Specifically, the borehole sequence depth feature vector extracted in step S12 is compared with the feature representations of all templates in the standard vertical sequence template library established in step S11 to calculate similarity. A similarity threshold is set, and borehole segments with similarity below the threshold are initially marked as candidate abnormal sequences.

[0037] For candidate anomalous sequences, a refined analysis is performed using sequence alignment algorithms to specifically identify the types of anomalous patterns that violate the standard template. The main patterns include: reverse sequence detection, determining whether the relative order of mineral occurrence is opposite to the expected stratigraphic sequence of the standard template; missing band detection, determining whether a specific mineral band or combination expected to be present in the standard template is completely missing or has a significantly inconsistent thickness in the sequence; and anomalous symbiosis detection, identifying mineral pairs or mineral combinations appearing in the sequence that are not defined as common symbiotic combinations in the knowledge graph.

[0038] Based on the specific anomaly pattern type identified, the regional geological knowledge map constructed in step S11 is invoked. Logical reasoning is performed on the anomalies according to the predefined geological rules in the map (for example, if a high-temperature mineral A overlies a low-temperature mineral B, it may indicate a hydrothermal superposition event; if an alteration zone C is missing, it may indicate tectonic truncation).

[0039] The reasoning results are transformed into one or more structured potential geological genetic hypotheses. Each hypothesis includes a description of the cause of the anomaly (e.g., possible reversal of the alteration sequence due to subsequent hydrothermal activity) and a quantitative confidence assessment. The confidence level is calculated based on the degree of matching between the anomaly and the rule conditions, the strength of the association in the knowledge graph, and the degree of deviation of the sequence from the template. Finally, the generated hypotheses are sorted and output according to their confidence levels.

[0040] S14: Extract the original spectral data corresponding to the sequence anomaly segments, label them with anomaly type and geological hypothesis, generate special augmented samples and return them to the training set, visualize and highlight the identified sequence anomaly segments on the mineral column chart, and output an interpretation card containing anomaly type and geological hypothesis.

[0041] Specifically, based on the sequence anomaly detection results output in step S13, the borehole number, starting depth, and ending depth information corresponding to each anomaly segment are obtained. Based on this depth coordinate information, the original spectral image data blocks corresponding to the depth intervals are located and extracted from the original multi-mining area spectral image dataset collected in step S1 of Example 1.

[0042] A set of composite labels is generated for each extracted raw spectral data block. The labels include: a) the main mineral category label identified in step S4 of Example 1; b) the sequence anomaly type label determined in step S13 (such as inverted sequence); c) one or more potential geological genetic hypothesis labels generated in step S13 and their confidence levels. This triplet label, together with the raw spectral data, constitutes a special augmented sample.

[0043] All the generated labeled special augmented samples are treated as a separate batch or dataset and merged or appended to the multi-mining region spectral training dataset used to train the model in Example 1, in the same format as the other training samples in step S1 of Example 1. This ensures that this data containing anomalous patterns can be used for network parameter updates in subsequent model training cycles.

[0044] Based on the output standard mineral histogram, according to the depth range of the anomalous section, use a distinct color or fill pattern that is different from the background to highlight or draw a bounding box in the corresponding depth range, and explain in the legend or side note that the highlighted area represents the identified sequence anomalous.

[0045] For each highlighted anomalous section, a structured interpretation card is automatically generated. The card content is recorded in text or structured data format and includes at least: the borehole ID and depth range of the anomalous section, the specific sequence anomaly type, the main geological hypothesis derived from the inference, and the confidence score of the hypothesis. This card is associated with the highlighted area in the bar chart through depth coordinates or a unique ID.

[0046] S2 constructs an interpretable network model that includes a physical constraint attention mechanism and a feature decoupling dual branch, and introduces adversarial domain invariant learning and physical consistency verification.

[0047] Specifically, a convolutional neural network (CNN) is constructed as a shared encoder. The input is a normalized spectral image (or a flattened spectral vector) processed in step S1. This encoder automatically extracts local and global features of the spectrum to generate a shared primary feature map.

[0048] Above the shared encoder, two structurally symmetric but functionally decoupled branch networks are constructed: a mineral essence feature branch, which receives shared features and whose core is a physically knowledge-constrained spectral attention module. Internally, this module reweights different spectral channels (bands) of the input features using a learnable attention weight vector. The learning process of this weight vector is subject to soft constraints from a standard mineral spectral library (for example, through an auxiliary loss function that encourages assigning higher attention weights to channels corresponding to known diagnostic absorption bands), thus achieving adaptive attention guided by physical knowledge. The mineral region feature branch runs parallel to the essence branch but has a different training objective. It aims to specifically capture and characterize non-universal features in the data related to specific mineral regions, such as spectral shifts and background interference; its output is domain features.

[0049] Design a differentiable physical consistency verification module. This module receives the feature vector output from the mineral intrinsic feature branch and internally maintains a queryable standard spectral feature library for minerals. The module calculates the similarity (e.g., cosine similarity) between the intrinsic feature vector and all spectral templates in the standard library and outputs the index of the best-matching template and a similarity score. This similarity score serves as a direct measure of whether the learned features conform to known physical laws and participates in guiding network training.

[0050] A lightweight domain discriminator network is introduced. During model training, the discriminator aims to distinguish which mining area the feature vectors output by the mineral essential feature branch originate from (i.e., perform mining area classification). Simultaneously, an adversarial training objective for the mineral essential feature branch is to maximize the confusion its output features cause with the domain discriminator, preventing the discriminator from accurately determining the feature source. This adversarial training forces the essential feature branch to remove mining area information from the learned features as much as possible, thereby enhancing its cross-mining area invariance.

[0051] The feature vectors output from the essential mineral features branch and the feature vectors output from the mining area features branch are concatenated to form a fused feature. This fused feature is then input into a fully connected classifier, which outputs an initial predicted probability distribution for the mineral category. Finally, the network model simultaneously outputs the mineral category prediction (classification results based on the fused feature), a physical matching reference (the best matching standard spectral index and matching degree output by the physical consistency verification module), and interpretability weights (channel weights of the spectral attention module in the essential features branch, which can be used to visualize key bands of interest).

[0052] S3 drives the model to simultaneously perform high-precision mineral classification, cross-mining area generalization, feature physical interpretability, and feature decoupling by jointly optimizing classification loss, domain adversarial loss, physical consistency loss, and feature decoupling loss.

[0053] Specifically, the cross-entropy loss function is used to calculate the difference between the mineral category probability distribution predicted by the model and the actual mineral labels of the samples. This loss directly drives the model to learn the correct mineral classification.

[0054] An adversarial training mechanism is set up. First, the domain discriminator attempts to correctly distinguish the source mining area labels corresponding to the features extracted by the mineral essential feature branch, and its loss function is the classification loss of the discriminator. At the same time, the training objective of the mineral essential feature branch includes an adversarial loss, that is, to maximize the error rate of the domain discriminator, thereby encouraging it to generate features that are not affected by differences in mining areas.

[0055] The calculation is based on the output of the physical consistency verification module. This loss function encourages feature vectors extracted from the mineral intrinsic feature branch to have higher similarity to standard spectral feature vectors of the corresponding category in the mineral standard spectral feature library (e.g., calculating and maximizing cosine similarity). This ensures that the learned features are consistent with known physical spectral knowledge.

[0056] The loss function aims to minimize the correlation between two vectors: the intrinsic mineral characteristics and the regional mineral characteristics. This is achieved by calculating the mutual information between them, or by designing a discriminator to distinguish them. The goal is to ensure that the feature representations extracted by the two branches are as independent as possible, each carrying different and unrelated information.

[0057] The four loss functions are linearly weighted according to preset weight coefficients and combined into a single total loss function. Using gradient descent and its variants (such as the Adam optimizer), the parameters of the main network (shared encoder, dual-branch, classification head), the physical consistency verification module, and the domain discriminator are simultaneously optimized via backpropagation to minimize this total loss. The entire training process drives the model to synchronously approach multiple objectives: high-precision classification, cross-mining area feature invariance, feature physical rationality, and decoupling between features.

[0058] S4, when the model is applied to borehole data from a brand new mining area, the standard preprocessed spectral image of the new mining area is input into the trained model.

[0059] Specifically, the raw borehole spectral image data of the target new mining area is processed using the same preprocessing procedure as in step S1, including relative reflectance conversion, band alignment, and resampling, to ensure that the input data format is consistent with that of the training phase. The standardized new data is then input into the pre-trained complete model network. The model will sequentially perform calculations such as shared encoding, bi-branch feature extraction, and physical consistency verification, and simultaneously output the following set of results: the initial predicted probability distribution of mineral categories, the essential feature vectors of minerals and the feature vectors of the mining area, the best matching standard spectral index and matching degree score calculated by the physical consistency verification module, and the channel attention weight vector generated by the spectral attention module in the essential feature branch.

[0060] The initial predicted probabilities and physical matching scores output by the model are weighted and fused according to the optimal weight ratio determined during the training phase to generate the final fused confidence probability for each mineral category. The category with the highest fused confidence probability is selected as the model's final mineral identification result.

[0061] Based on the intermediate data generated by the model, the system automatically outputs key band attention maps (generating visual charts based on the spectral attention weight vectors to identify the spectral band positions that the model focuses on in its decisions) and best-match physical references (outputting the mineral names and matching degrees of the best-match standard spectra given by the physical consistency verification module).

[0062] The system employs automated logical rules to validate the identification results. If any of the following conditions occur, the sample is automatically marked as an anomalous sample requiring further review, and the final fused confidence probability falls below a preset reliability threshold: A significant difference exists between the model's initial predicted probability distribution and the physical matching score (e.g., the difference exceeds a set threshold). The magnitude or energy of the mining area feature vector exceeds a specific threshold, indicating a significant difference between the current sample and the training data distribution.

[0063] S5, the model outputs mineral categories and fusion confidence scores by driving the prediction probability and physical matching degree through fused data; it simultaneously generates key band attention maps and interpretive evidence for matching standard spectra; and automatically identifies abnormal samples that need to be reviewed based on confidence score differences and domain feature strength.

[0064] Specifically, the system receives all intermediate results generated from the model forward propagation in step S4, including the initial predicted probability distribution of mineral categories, the best-matching standard spectral index and its matching score calculated by the physical consistency verification module, the spectral attention weight vector of the essential feature branch, and the feature vector of the mining area. Next, decision fusion is performed, where the initial predicted probabilities and physical matching scores are weighted and summed according to preset fusion weights to calculate the fusion confidence score for each mineral category. The category with the highest fusion confidence score is selected as the final mineral type. Simultaneously, a keyband attention map is generated based on the spectral attention weight vector to visually display the spectral channels upon which the network decision relies. At the same time, according to the best-matching standard spectral index, the corresponding standard spectral curve is retrieved from the built-in standard library as a physical reference and output. Finally, automated anomaly detection is performed, and judgment is made according to preset logical rules: if the final fusion confidence is lower than the set threshold, or the difference between the initial prediction probability and the physical matching degree exceeds the allowable range, or the norm of the calculated mineral area feature vector exceeds a specific threshold, then the sample is automatically marked as an anomaly sample that needs expert review, and together with the above mineral category, fusion confidence and interpretive evidence, it constitutes the final output result.

[0065] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By employing a data augmentation and feature decoupling network guided by physical knowledge, this method can automatically and accurately identify mineral composition from borehole wall images exhibiting significant cross-mining-region spectral shifts. Its core advantages lie in the fact that the model not only possesses strong cross-mining-region generalization ability and robustness, effectively overcoming interference from different mining environments and instruments, but also, by integrating physical spectral matching and an interpretable attention mechanism, makes the automated decision-making process transparent and reliable. Furthermore, it can simultaneously output key evidence and automatically identify low-confidence anomalous samples, greatly enhancing the method's practicality and reliability.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for anomaly analysis of mineral images inside boreholes, characterized in that, include: S1, collects borehole spectral image data from multiple mining areas, performs preprocessing and physical knowledge augmentation, and constructs a multi-mining area spectral training dataset; S2 constructs an interpretable network model that includes a physical constraint attention mechanism and a feature decoupling dual branch, and introduces adversarial domain invariant learning and physical consistency verification. S3, the network model is trained by jointly optimizing the classification loss, domain adversarial loss, physical consistency loss and feature decoupling loss; S4 applies the trained model to borehole data from new mining areas to perform mineral identification and anomaly detection, and outputs mineral categories, fusion confidence scores, and interpretive evidence. S5 automatically identifies anomalous samples that require review based on confidence level differences and domain feature strength.

2. The anomaly analysis method for borehole mineral images according to claim 1, characterized in that, In step S1, the process of collecting borehole spectral image data from multiple mining areas and performing physical knowledge augmentation also includes: S11, based on the mineral profiles of identified boreholes in the region, construct a structured regional geological knowledge map, and summarize standard vertical mineral sequence templates through sequence alignment and clustering; S12, Sequence encoding is performed on the mineral identification profile of a single hole to extract a depth feature vector containing contextual information; S13, the extracted sequence depth features are matched and compared with the standard template library to identify abnormal patterns, and the geological rules in the knowledge graph are called to perform reasoning to generate potential geological genesis hypotheses; S14: Extract the original spectral data corresponding to the sequence anomaly segment, label it with anomaly type and geological hypothesis, and generate special augmented samples to be fed back into the training set.

3. The anomaly analysis method for borehole mineral images according to claim 1, characterized in that, In step S2, the construction of an interpretable network model comprising a physical constraint attention mechanism and a feature decoupling dual branch includes: constructing a shared encoder to extract spectral features; constructing a mineral essential feature branch, which includes a spectral attention module with physical knowledge constraints to weight spectral channels; constructing a mining area feature branch to capture spectral shift features related to a specific mining area; designing a physical consistency verification module to calculate the similarity between essential features and a standard spectral library; introducing a domain discriminator network to extract mining area information from essential features through adversarial training; and concatenating essential features with domain features, inputting the concatenation into a classifier, and outputting a mineral category prediction.

4. The anomaly analysis method for borehole mineral images according to claim 1, characterized in that, In step S3, the joint optimization of classification loss, domain adversarial loss, physical consistency loss, and feature decoupling loss includes: calculating the classification loss using the cross-entropy loss function; setting up an adversarial training mechanism, where the domain discriminator attempts to distinguish the source mining area of ​​the features, and the mineral essential feature branch maximizes the error rate of the discriminator; calculating the physical consistency loss based on the output of the physical consistency verification module to encourage the similarity between essential features and standard spectral features; calculating the mutual information estimate between the mineral essential features and the mining area features or minimizing their correlation through the discriminator as the feature decoupling loss; and weighting and merging the above loss functions into a total loss, and optimizing the model parameters through backpropagation.

5. The anomaly analysis method for borehole mineral images according to claim 1, characterized in that, In step S4, applying the trained model to borehole data in the new mining area for mineral identification and anomaly detection includes: inputting the preprocessed data from the new mining area into the model to obtain the initial predicted probability of the mineral category, the physical matching score, the attention weight, and the domain feature vector; weighting and fusing the initial predicted probability and the physical matching score to obtain the final fused confidence probability, and selecting the category with the highest confidence as the identification result; outputting the key band attention map and the best matching standard spectrum as explanatory evidence; and automatically identifying abnormal samples that need to be reviewed based on the final fused confidence, the difference between the initial predicted probability and the physical matching score, and the magnitude of the domain feature vector.

6. The anomaly analysis method for borehole mineral images according to claim 2, characterized in that, In S11, the construction of a structured regional geological knowledge map specifically includes: defining mineral types, lithological types, alteration types, mineralization stages, and geological tectonic events as entities; and establishing symbiotic relationships, hierarchical relationships, genetic relationships, and spatiotemporal associations among entities based on regional geological data and expert knowledge.

7. The anomaly analysis method for borehole mineral images according to claim 2, characterized in that, In step S12, the sequential encoding of the mineral identification profile of a single hole specifically includes: mapping the mineral category label into a continuous feature vector through an embedding layer to form a two-dimensional feature matrix; inputting the feature matrix into a sequential encoding model to extract depth context features; and aggregating the output of the sequential encoding model to obtain a fixed-length depth feature vector.

8. The anomaly analysis method for borehole mineral images according to claim 2, characterized in that, In step S13, identifying abnormal patterns includes: calculating the similarity between sequence depth features and standard templates, marking segments with similarity below a threshold as candidate abnormal sequences; identifying inverted sequences, missing bands, or abnormal co-occurrence patterns in the candidate abnormal sequences using a sequence alignment algorithm; and invoking predefined geological rules in the knowledge graph to reason about the identified abnormal patterns, generate one or more potential geological genesis hypotheses, and calculate their confidence levels.

9. The anomaly analysis method for borehole mineral images according to claim 2, characterized in that, In step S14, generating special augmented samples and feeding them back to the training set specifically includes: extracting corresponding spectral data blocks from the original spectral dataset based on the depth coordinates of the anomalous sections; generating triplet labels for each data block containing mineral category, sequence anomalous type, and geological hypothesis to form augmented samples; merging the labeled augmented samples into the multi-mining area spectral training dataset; highlighting the sequence anomalous sections on the mineral histogram and outputting an explanation card containing the anomalous type and geological hypothesis.

Images