Exploration data anomaly detection method and device based on artificial intelligence

Abstract

The application provides an exploration data anomaly detection method and device based on artificial intelligence, relates to the technical field of data processing, and through self-adaptive frequency band division driven by information entropy, lets the frequency band range dynamically adjust with the data complexity, and completely gets rid of the restraint of the fixed frequency band on the non-steady-state logging signal. Further, the high-frequency dynamic characteristics are reserved through normalized logging values, the global information is supplemented through window statistics, the synergistic information of the associated channels is effectively utilized in combination with the cross-channel correlation, and the feature fusion is carried out by using the frequency band energy vector, so that the high-frequency abnormal characteristics still maintain the significant recognition degree after fusion and are not suppressed by the low-frequency background or statistical characteristics. In summary, the model can accurately distinguish the confusing abnormalities and improve the abnormal recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an artificial intelligence-based method and apparatus for detecting anomalies in exploration data. Background Technology

[0002] In the field of geophysical exploration, well logging data, as a crucial tool for understanding subsurface geological structures and identifying anomalous strata, is widely used in oil and gas exploration, mineral resource exploration, and geological hazard monitoring. Its analysis results have a critical impact on exploration decisions. Currently, existing technologies for processing well logging data typically analyze the frequency distribution by pre-setting fixed frequency band divisions and rely on a single energy feature as the core basis for anomaly identification. However, well logging curves generally possess complex inherent characteristics, such as significant non-steady-state features. Different geological events often correspond to different frequency distributions; for example, faults typically exhibit high-frequency abrupt signals, while lithological gradations often show low-frequency, gradual changes, causing the frequency characteristics of the data to dynamically change with geological events. Furthermore, the energy of effective signals and noise easily overlaps, increasing the difficulty of feature extraction. Moreover, the data simultaneously contains short-period high-frequency features and long-period low-frequency features, exhibiting a coexistence of high and low frequency characteristics.

[0003] Existing fixed-frequency band division methods are ill-suited to the dynamic and unsteady characteristics of well logging data as geological events change, failing to accurately capture the true frequency distribution corresponding to different geological events. Furthermore, the single-energy-feature judgment model has significant limitations, not only failing to effectively distinguish between valid signals and noise but also missing multi-dimensional key information required for anomaly identification. This results in an inability to comprehensively characterize anomalies in scenarios where high and low-frequency features coexist. These issues directly impact the accuracy of anomaly identification, making existing technologies generally insufficient to meet the practical needs of geophysical exploration for accurate anomaly identification and reliable early warning of high-risk anomalies. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide an exploration data anomaly detection method and apparatus based on artificial intelligence, which can accurately distinguish easily confused anomalies and improve the accuracy of anomaly identification.

[0005] In a first aspect, embodiments of the present invention provide an artificial intelligence-based method for detecting anomalies in exploration data. The method includes: acquiring the formation physical response of a target exploration segment, recording the depth coordinates, instrument status, and environmental parameters of the formation physical response, and generating raw exploration data for the target exploration segment; determining the adaptive frequency band energy ratio of the raw exploration data based on its information entropy; determining the frequency band energy vector contained in the raw exploration data based on the adaptive frequency band energy ratio; the frequency band energy vector includes low-frequency vectors, mid-frequency vectors, and high-frequency vectors; constructing a basic feature matrix for the raw exploration data based on normalized logging values ​​and window statistics; performing cross-channel feature fusion on the basic feature matrix using the frequency band energy vectors based on the cross-channel correlation of the basic feature matrix to generate fused features of the raw exploration data; inputting the fused features as features to be tested into a pre-constructed artificial intelligence model, and using the artificial intelligence model to identify anomalies in the features to be tested to determine the anomaly identification result corresponding to the target exploration segment; and determining the anomaly category to which the target exploration segment belongs based on the anomaly identification result.

[0006] In conjunction with the first aspect, the present invention provides a first implementation of the first aspect, wherein the step of determining the anomaly category to which the target exploration segment belongs based on the anomaly identification result includes: determining the anomaly confidence level corresponding to the anomaly identification result of the target exploration segment in each test area by means of a sliding window; and determining the anomaly category to which the target exploration segment belongs based on the cumulative result of the anomaly confidence level.

[0007] In conjunction with the first aspect, this embodiment of the invention provides a second implementation of the first aspect, wherein the step of determining the adaptive frequency band energy ratio of the original exploration data based on the information entropy of the original exploration data includes: performing wavelet packet decomposition on the original exploration data to divide the original exploration data into multiple layers of data; each layer of data contains multiple frequency band nodes; calculating the information entropy of each frequency band node; determining the optimal subtree node set from the decomposition tree to which the frequency band node belongs based on the information entropy; calculating the node energy of the current frequency band node based on the wavelet packet coefficients of each frequency band node in the optimal subtree node set; and determining the adaptive frequency band energy ratio of the original exploration data for each frequency band node based on the node energy and the information entropy.

[0008] In conjunction with the first aspect, this embodiment of the invention provides a third implementation of the first aspect, wherein the step of performing cross-channel feature fusion on the basic feature matrix based on the frequency band energy vector and the cross-channel correlation of the basic feature matrix to generate fused features of the original exploration data includes: calculating the cross-channel correlation coefficient matrix of the basic feature matrix, and constructing a diagonal matrix of the frequency band energy vector; using the diagonal matrix to perform channel weighting on the cross-channel correlation coefficient matrix to generate a cross-channel feature fusion matrix; selectively filtering the frequency domain features of the original exploration data, and superimposing the selectively filtered original exploration data with the cross-channel feature fusion matrix to generate fused features of the original exploration data, and the feature to be measured after feature fusion.

[0009] In conjunction with the first aspect, this invention provides a fourth implementation of the first aspect, wherein the step of inputting fused features as test features into a pre-constructed artificial intelligence model, and using the artificial intelligence model to perform anomaly identification on the test features and determine the anomaly identification result corresponding to the target exploration layer includes: adaptively selecting features for the test features based on a preset feature channel attention mechanism to determine the attention output feature matrix of the test features; extracting short-period features and long-period features from the attention output feature matrix respectively, and performing gated residual connection on the features extracted by the short-period and long-period features to generate a dual-path residual output feature matrix; performing frequency band modulation processing on the dual-path residual output feature matrix based on the frequency band energy vector to generate a frequency domain modulation feature matrix; performing attention-gated probability transformation on the frequency domain modulation feature matrix to determine the anomaly probability distribution corresponding to the test features; and determining the anomaly identification result corresponding to the test features based on the anomaly probability distribution.

[0010] In conjunction with the first aspect, this invention provides a fifth implementation of the first aspect, wherein the step of adaptively selecting features based on a preset feature channel attention mechanism to determine the attention output feature matrix of the feature to be tested includes: inputting the feature to be tested into a single hidden layer perceptron; performing linear activation processing on the feature to be tested based on the weight matrix of the single hidden layer perceptron with respect to the hidden layer; performing feature channel normalization on the linearly activated feature to be tested to generate an attention weight matrix corresponding to the feature to be tested; and adaptively weighting the feature to be tested using the attention weight matrix to generate an attention output feature matrix corresponding to the feature to be tested.

[0011] In conjunction with the first aspect, this invention provides a sixth implementation of the first aspect, wherein the step of extracting short-period features from the attention output feature matrix includes: applying a depthwise separable convolution operator to extract short-period features from the attention output feature matrix, and outputting a short-period feature output matrix of the attention output feature matrix; the depthwise separable convolution operator uses a small convolution kernel time size and a unit stride to extract features; the step of extracting long-period features from the attention output feature matrix includes: performing standard convolution and max pooling operations sequentially on the attention output feature matrix, and then extracting features from the attention output feature matrix to output a long-period feature output matrix of the attention output feature matrix.

[0012] In conjunction with the first aspect, this invention provides a seventh implementation of the first aspect, wherein the method further includes: acquiring a pre-constructed training sample set; using a mini-batch gradient descent strategy to input the training sample set into an artificial intelligence model for model training; calculating the training gradient through a pre-constructed joint loss function, and dynamically adjusting the learning rate of the artificial intelligence model using a preset adaptive moment estimator optimizer to update the network parameters of the artificial intelligence model.

[0013] In conjunction with the first aspect, this embodiment of the invention provides an eighth implementation of the first aspect, wherein the method further includes: constructing a depth continuity penalty term for the training sample set based on the JS divergence of the predicted probability distributions of neighboring samples in the training sample set and the absolute value of the depth difference between neighboring samples; obtaining the misjudgment cost matrix corresponding to the training sample set; calculating the cost-sensitive weights corresponding to the training sample set based on the misjudgment cost matrix; weighting the cross-entropy loss corresponding to the training sample set using the cost-sensitive weights; and constructing a joint loss function for the training sample set by combining the depth continuity penalty term.

[0014] Secondly, embodiments of the present invention also provide an exploration data anomaly detection device based on artificial intelligence. The device includes: a data acquisition module, used to acquire the formation physical response of a target exploration segment, record the depth coordinates, instrument status, and environmental parameters of the formation physical response, and generate raw exploration data for the target exploration segment; a data processing module, used to determine the adaptive frequency band energy ratio of the raw exploration data based on the information entropy of the raw exploration data; determine the frequency band energy vector contained in the raw exploration data based on the adaptive frequency band energy ratio; the frequency band energy vector includes a low-frequency vector, a mid-frequency vector, and a high-frequency vector; and feature fusion. The module is used to construct a basic feature matrix for the original exploration data based on the normalized logging values ​​and window statistics of the original exploration data; according to the cross-channel correlation of the basic feature matrix, the module uses frequency band energy vectors to perform cross-channel feature fusion on the basic feature matrix to generate fused features of the original exploration data; the execution module is used to input the fused features as the features to be tested into a pre-built artificial intelligence model, and the artificial intelligence model performs anomaly identification on the features to be tested to determine the anomaly identification result corresponding to the target exploration segment; the output module is used to determine the anomaly category to which the target exploration segment belongs based on the anomaly identification result.

[0015] This invention provides an artificial intelligence-based method and apparatus for detecting anomalies in exploration data. Through information entropy-driven adaptive frequency band partitioning, the frequency band range is dynamically adjusted according to data complexity, completely eliminating the constraints of fixed frequency bands on unsteady logging signals. Furthermore, by preserving high-frequency dynamic features through normalized logging values, supplementing global information through window statistics, effectively utilizing collaborative information from associated channels by combining cross-channel correlation, and using frequency band energy vectors for feature fusion, high-frequency anomaly features retain significant distinguishability after fusion, and are not suppressed by low-frequency background or statistical features. In summary, the model can accurately distinguish easily confused anomalies, improving anomaly identification accuracy. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating an artificial intelligence-based method for detecting anomalies in exploration data, provided as an embodiment of the present invention;

[0018] Figure 2 A flowchart of another artificial intelligence-based exploration data anomaly detection method provided in an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram illustrating the decomposition effect of a frequency band decomposition method provided in an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram comparing the anomaly detection performance of different methods according to an embodiment of the present invention;

[0021] Figure 5 This is a performance comparison diagram of different feature fusion methods provided in an embodiment of the present invention;

[0022] Figure 6 This is a schematic diagram of the frequency domain energy vector distribution of different geological events provided in an embodiment of the present invention;

[0023] Figure 7 A schematic diagram of the structure of an exploration data anomaly detection device based on artificial intelligence provided in an embodiment of the present invention;

[0024] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention provides an artificial intelligence-based method and apparatus for detecting anomalies in exploration data. It can ensure that anomaly information is not obscured and can take into account key correlations across channels, time and frequency domains, and geology and engineering, ensuring that features are not fragmented and information is not isolated, thereby achieving a high capture rate, low false judgment rate, and accurate classification rate for geological anomalies.

[0027] To facilitate understanding, the present invention will first describe an artificial intelligence-based method for detecting anomalies in exploration data, as provided in an embodiment of the invention. (Refer to...) Figure 1 The method includes the following steps:

[0028] Step S102: Obtain the formation physical response of the target exploration section, record the depth coordinates, instrument status and environmental parameters of the formation physical response, and generate the original exploration data of the target exploration section.

[0029] This invention acquires raw exploration data through a multi-channel logging instrument system, specifically including five types of logging curves: sonic transit time, gamma rays, resistivity, neutron porosity, and density. The data acquisition depth covers the target exploration layer, such as 0-1500 meters, and the formation physical response is continuously recorded at a fixed sampling interval, such as 0.1 meters / point, forming an initial logging dataset. The depth coordinates, instrument status, and environmental parameters are recorded simultaneously during the data acquisition process to ensure the spatiotemporal consistency of the data.

[0030] Step S104: Determine the adaptive frequency band energy ratio of the original exploration data based on the information entropy of the original exploration data; determine the frequency band energy vector contained in the original exploration data based on the adaptive frequency band energy ratio.

[0031] The frequency band energy vector includes low-frequency, mid-frequency, and high-frequency vectors. Conventional wavelet decomposition or Fourier methods use fixed frequency bands, which cannot adapt to non-steady-state signals, resulting in insufficient accuracy in detecting transient anomalies such as faults. This invention uses information entropy to filter sensitive frequency bands from the original exploration data and further calculates the energy ratio of each frequency band. Based on the energy ratio, the contribution of different frequency bands to the total signal energy is determined to highlight the importance of high-entropy frequency bands. This ensures that the determined frequency band energy vector not only reflects the energy distribution of each frequency band but also highlights high-uncertainty areas based on information entropy, effectively capturing transient anomalies such as faults.

[0032] Step S106: Based on the normalized logging values ​​and window statistics of the original exploration data, construct a basic feature matrix for the original exploration data; according to the cross-channel correlation of the basic feature matrix, use the frequency band energy vector to perform cross-channel feature fusion on the basic feature matrix to generate the fused features of the original exploration data.

[0033] By constructing a basic feature matrix using normalized logging values ​​and window statistics, the original feature information of the data can be preserved to a certain extent, including high-frequency transient features. Window statistics can capture the statistical characteristics of data within different time windows, while normalization processing makes different logging values ​​comparable, avoiding the situation where certain features are ignored due to differences in numerical magnitude, thereby reducing the possibility of high-frequency transient features being masked by statistical features. Using frequency band energy vectors to perform cross-channel feature fusion on the basic feature matrix can fully explore the correlation information between channels. Frequency band energy vectors reflect the energy distribution of different frequency bands in each channel. Through frequency band energy vectors, it is possible to determine which channels have strong correlations in specific frequency bands, and then fuse these related information, so that the fused features can more comprehensively reflect the true situation of the data, improving the ability to represent complex information in logging data. This embodiment of the invention not only considers the features of data within a single channel, but also comprehensively utilizes cross-channel correlation information, making the generated fused features more accurate and reliable. In practical applications, it can more accurately identify changes in geological structure, anomalies in formation parameters, and other information, providing stronger support for geological exploration and resource evaluation.

[0034] Step S108: Input the fused features as the features to be tested into the pre-built artificial intelligence model, and use the artificial intelligence model to identify anomalies in the features to be tested to determine the anomaly identification results corresponding to the target exploration layer.

[0035] Step S110: Based on the anomaly identification results, determine the anomaly category to which the target exploration section belongs.

[0036] The aforementioned data processing transforms raw exploration data into feature associations that the model can recognize, providing multi-dimensional evidence for determining lithological gradations and fluid-bearing layers. Furthermore, it avoids the model prioritizing large numerical features (such as high-resistivity layers) and ignoring small but crucial anomalous signals (such as minor drifts caused by instrument malfunctions) due to differences in feature magnitude. The fusion feature AI model provides a complete information view that includes both local transients and global statistics, and correlates single-channel features while integrating cross-channel couplings.

[0037] The final output is a probability distribution vector for the anomaly category. In one implementation, this includes four anomalies: fault, fluid-bearing layer, lithological gradation, and instrument malfunction. This invention provides clear category indications while preserving uncertainties. The collaborative decision-making process of model identification and expert verification offers flexible support, significantly improving the accuracy and reliability of anomaly interpretation for target exploration sections.

[0038] In summary, this invention, based on the information entropy of the original exploration data of the target exploration layer, reflects the data complexity, performs adaptive frequency band division, dynamically matches data fluctuation characteristics, and provides a precise frequency band range. Furthermore, it preserves high-frequency dynamic features by normalizing logging values, supplements global information through window statistics, effectively utilizes the collaborative information of associated channels by combining cross-channel correlation, and uses frequency band energy vectors for feature fusion. This ensures that high-frequency anomaly features retain significant distinguishability after fusion, are not suppressed by low-frequency background or statistical features, and enables the model to accurately distinguish easily confused anomalies, improving anomaly identification accuracy.

[0039] Based on the above embodiments, this invention also provides another method for detecting anomalies in exploration data based on artificial intelligence. Figure 2 This diagram illustrates a flowchart of another artificial intelligence-based exploration data anomaly detection method provided by an embodiment of the present invention, with reference to... Figure 2 The method includes the following steps:

[0040] Step S202: Obtain the formation physical response of the target exploration section, record the depth coordinates, instrument status and environmental parameters of the formation physical response, and generate the original exploration data of the target exploration section.

[0041] Step S204: Determine the adaptive frequency band energy ratio of the original exploration data based on the information entropy of the original exploration data.

[0042] The original frequency domain analysis method does not differentiate between specific frequency bands corresponding to different geological events. For example, faults typically correspond to high-frequency signals, while lithological gradients correspond to low-frequency signals. Conventional processing uses a fixed frequency band division strategy, which is difficult to adapt to the non-steady-state characteristics of well logging signals, leading to the omission of transient anomaly information detection. This invention dynamically divides wavelet packet subtrees based on information entropy and extracts adaptive frequency band energy ratios. In specific implementation, the embodiments of this invention use wavelet packet decomposition to divide the original exploration data into frequency bands, possessing unique characteristics such as dynamic frequency band division and integration of geological significance. This overcomes the limitations of conventional fixed frequency band analysis and provides adaptive, highly targeted frequency domain energy information. Furthermore, by using the information entropy determined based on the statistical characteristics of wavelet packet coefficients of sub-nodes, the complexity / uncertainty of the coefficient distribution of each sub-node after wavelet packet decomposition is measured, which can accurately screen geologically effective signals and suppress noise interference. After further determining the optimal subtree node set, the corresponding adaptive frequency band energy ratio is determined, which can combine the physical energy of the sub-node with geological effectiveness to determine the importance of the frequency band for anomaly detection. Specifically, the embodiments of this invention include the following steps:

[0043] 1) Wavelet packet decomposition is performed on the original exploration data to divide it into multiple layers. Each layer contains multiple frequency band nodes.

[0044] The original logging signal is decomposed into J-levels using wavelet packet decomposition algorithm to obtain the wavelet packet coefficient vectors of each sub-node. These vectors are used to store the time-frequency details of the signal in a specific frequency band. The J-level decomposition is defined as follows: In the layer decomposition The wavelet packet coefficients of the child nodes are: It contains detailed information about the signal in a specific frequency band, and at the same time, defines... It is the first Layer The coefficient vector of each child node stores the time-frequency coefficients of that frequency band. This represents the layer index of the wavelet packet decomposition. ; This indicates the maximum number of wavelet packet decomposition levels, for example, when the signal length is greater than... At that time, it is usually set Ensure that the bandwidth resolution is appropriate; This represents the child node index, generated at each level of decomposition. Number of child nodes The lowest level corresponding to time has a total of Number of child nodes .

[0045] It should be noted that a child node is a frequency band node in the wavelet packet decomposition tree, containing a set of coefficient vectors that represent detailed information of the signal in a specific frequency band, rather than a single value.

[0046] 2) Calculate the information entropy of each frequency band node, and based on the information entropy, determine the optimal subtree node set from the decomposition tree to which the frequency band node belongs.

[0047] Calculate the information entropy of each child node, measure the complexity of its wavelet packet coefficient distribution, and then select the optimal subtree node set from all possible subtrees by maximizing the entropy value. This determines the optimal subtree containing the key frequency band nodes, expressed as:

[0048]

[0049] In the formula, Indicates the first The information entropy of a child node measures the complexity or uncertainty of the coefficient distribution of that node. Indicates the first The wavelet packet coefficient histogram of the nth child node The probability of the i-th interval is obtained through histogram statistics, specifically by... The wavelet packet coefficients of the child nodes are divided into: Calculate the interval. Frequency of coefficients within each interval ,but b is the interval index of the histogram. Binning corresponding to the coefficient values; It is a logarithmic function, with the default base being the natural constant; This indicates the number of intervals in the histogram, i.e., the number of bins. The number of histogram intervals affects the accuracy of entropy calculation. When the range of coefficient values ​​is large, it is usually set to a higher value. When computing resources are limited, it is usually set to . This means finding the subtree that maximizes the objective function value among all possible subtrees of the wavelet packet decomposition tree. This represents a subtree in the wavelet packet decomposition tree, which is the... A subset of child nodes; This represents the set of nodes contained in the optimal subtree.

[0050] It should be noted that a child node is the basic unit of a decomposition tree, and a subtree is a tree structure composed of multiple child nodes, such as a parent node and its descendants.

[0051] 3) Calculate the node energy of the current frequency band node based on the wavelet packet coefficients of each frequency band node in the optimal subtree node set.

[0052] 4) Based on node energy and information entropy, determine the adaptive frequency band energy ratio of the original exploration data for each frequency band node.

[0053] For each child node in the optimal subtree, the sum of squared moduli of its wavelet packet coefficients is calculated as the energy. Then, the normalized adaptive energy ratio is calculated by combining the information entropy of that node to measure the importance of the frequency band for anomaly detection, expressed as:

[0054]

[0055] In the formula, Indicates the number of elements belonging to the optimal subtree. The energy of each child node is calculated as the sum of the squares of the moduli of all wavelet packet coefficients of that node; This represents the sequence index of the wavelet packet coefficients within the node. ; Indicates the first The length of the wavelet packet coefficients of each child node is determined by the length of the original signal. The maximum number of wavelet packet decomposition levels, J, is determined by the formula: ,like, , hour, ; Indicates the first In the layer decomposition The first child node Wavelet packet coefficient values; It represents the square of the absolute value, and in vector calculations, it represents the modulo-square operation. Indicates the first The adaptive energy ratio of each child node; the larger the value, the more important the frequency band is for anomaly detection. Indicates the number of elements belonging to the optimal subtree. The energy of each child node; The set of nodes contained in the optimal subtree Node index in .

[0056] It should be noted that geological anomalies such as faults often exhibit low entropy values ​​and high signal purity in specific frequency bands. By using the inverse logarithm of entropy as a weighted amplifier, the energy ratio in the low-entropy frequency band is nonlinearly enhanced, which can amplify the contribution of transient events such as faults, while suppressing background noise in the high-entropy noise band, such as the background noise mixed in the lithological gradient zone. By combining information entropy, the energy ratio can simultaneously carry physical strength and geological significance, and significantly improve anomaly sensitivity without increasing computational complexity.

[0057] Step S206: Based on the adaptive frequency band energy ratio, determine the frequency band energy vector contained in the original exploration data.

[0058] This can be based on the prior knowledge of the frequency domain characteristics of geological events (such as faults corresponding to high frequencies and lithological gradations corresponding to low frequencies), using the center frequency of the sub-nodes. Based on this, all child nodes in the optimal subtree M are divided into three non-overlapping frequency band sets. Furthermore, for the three sets of low / medium / high frequency, the adaptive frequency band energy ratio of all child nodes in each set is summed. The three summed results are used as the three dimensions of the vector to finally form the frequency band energy vector.

[0059] In practical implementation, Denotes the set of low-frequency child nodes whose center frequency satisfies ; Denotes the set of intermediate frequency child nodes whose center frequency satisfies ; Denotes the set of high-frequency child nodes whose center frequency satisfies . This indicates the center frequency of the frequency band represented by the child node; This represents the threshold for distinguishing between low and mid frequencies, such as... ; This represents the threshold for distinguishing between intermediate and high frequencies, such as... ; This represents a vector construction operation that combines three scalars into a three-dimensional vector.

[0060] Then, the adaptive energy ratios of each set are aggregated to form a frequency band energy vector containing three dimensions: low frequency, mid frequency, and high frequency, expressed as:

[0061]

[0062] In the formula, This represents a frequency band energy vector, containing the aggregated energy ratio across three dimensions: low frequency, mid frequency, and high frequency. It should be noted that the frequency band energy vector is generated using a dynamically selected set of nodes from the optimal subtree. During the process, based on the frequency distribution characteristics of geological events, such as faults corresponding to high frequencies and lithological gradations corresponding to low frequencies, through... Adaptive classification, while employing Instead of using raw energy as the aggregation object, the vector simultaneously integrates frequency band physical energy and geological importance weights. When a frequency band has high energy but also high entropy, such as in the case of instrument noise, its adaptive energy ratio... The value is suppressed to avoid interference vector representation; the dual filtering mechanism makes the frequency band energy vector... It can more purely characterize the frequency domain energy distribution with geological significance.

[0063] In one embodiment, the present invention also analyzes the difference in effectiveness between adaptive bandwidth decomposition and conventional fixed bandwidth decomposition when processing logging signals. Figure 3 A schematic diagram illustrating the decomposition effect of a frequency band decomposition method is shown. Figure 3 The diagram contains four sub-plots: the original logging signal shows the comprehensive signal characteristics within a depth range of 0-1000 meters, with the fault location marked at approximately 450 meters and the lithological gradient initiation point at 700 meters; the conventional fixed frequency band decomposition sub-plot shows that its low-frequency, mid-frequency, and high-frequency components have weak responses in the fault region and produce false high-frequency fluctuations in the lithological gradient zone; the decomposition sub-plot of the present invention significantly enhances the high-frequency response in the fault region while maintaining smooth low-frequency characteristics in the lithological gradient zone; the high-frequency component comparison sub-plot directly proves that the present invention's technology produces sharper and more significant high-frequency peaks at the fault location near 450 meters, while maintaining lower background noise in non-anomaly areas. Experimental results show that the dynamic frequency band division mechanism based on information entropy of the present invention can adaptively adjust the frequency band for different geological events, significantly improving the ability to detect transient anomalies.

[0064] Step S208: Based on the normalized logging values ​​and window statistics of the original exploration data, construct a basic feature matrix for the original exploration data.

[0065] By integrating normalized well logging values ​​and window statistics as basic features, a basic feature matrix is ​​constructed, which contains multiple feature dimensions. The basic feature matrix is ​​defined as follows: , dimension It includes basic information such as normalized well logging values ​​and window statistics. This represents the total number of samples, i.e., the total number of sampling points in the well logging data. Each sampling point corresponds to one sample. The window is used to indicate the depth neighborhood range centered on a single sampling point in the original exploration data, including a fixed number of adjacent sampling points. It should be noted that the basic feature matrix contains 15 feature dimensions, meaning that the total number of integrated basic features is 15, divided into two main categories: normalized logging values ​​and window statistics. The normalized logging values ​​include normalized sonic transit time, normalized gamma rays, normalized resistivity, normalized neutron porosity, and normalized density. The window statistics include window mean, window standard deviation, window skewness, window kurtosis, window minimum, window maximum, window energy, window first-order gradient mean, window second-order gradient mean, and window autocorrelation coefficient.

[0066] Step S210: Based on the cross-channel correlation of the basic feature matrix, the cross-channel feature fusion of the basic feature matrix is ​​performed using the frequency band energy vector to generate the fused features of the original exploration data.

[0067] Well logging data contains dynamic numerical values, static statistics, and cross-channel correlation features. Conventional feature stitching methods ignore the spatiotemporal coupling relationships between features, resulting in high-frequency transient features being masked by statistical features, and cross-channel correlation information not being effectively utilized. This invention uses a feature interaction matrix to achieve dynamic weighted fusion, and determines the fused features by allocating weights through frequency domain energy constraints.

[0068] In practical implementation, the fusion features are determined through the following steps:

[0069] 1) Calculate the cross-channel correlation coefficient matrix of the basic feature matrix, and construct the diagonal matrix of the frequency band energy vector.

[0070] The cross-channel correlation coefficient matrix contains the correlation coefficients calculated between different logging channels. It is obtained by calculating the Pearson correlation coefficients between different logging channels. For example, for the sonic transit time and gamma-ray channels, their correlation coefficients within the sliding window are calculated to form matrix elements. The matrix dimension is... ; The number of channel combinations, i.e., the number of different logging channel pairs involved in the correlation calculation, depends on the number of logging channels. ,like Corresponding to five channels: acoustic transit time, gamma rays, resistivity, neutron porosity, and density, the number of channel combinations is obtained by calculating all possible channel pairs. .

[0071] The diagonal weight matrix is ​​used to perform channel weighting on the cross-channel correlation coefficient matrix. Its diagonal elements are determined by the frequency domain energy, and are expressed as follows: ;in, Represents the diagonal weight matrix. It is a natural constant; It represents the frequency band energy vector, which includes the aggregated energy ratio of the three dimensions of low frequency, mid frequency, and high frequency; This represents the operation of constructing a diagonal matrix from a vector; This represents the weight adjustment factor, used to control the strength of the influence of frequency domain energy on the weights. The optimal value, selected through cross-validation, is typically [value missing]. When its value is greater than 1, the representation value is too large, which can easily lead to excessive weighting of high frequencies. When its value is less than 0.1, the representation value is too small, which can easily lead to weakening of dynamic weighting. It serves as a frequency band index, taking energy information from different frequency bands, specifically including low frequency, mid frequency, and high frequency; It represents the sum of frequency domain energy vectors, that is, the sum of frequency domain energy vectors of low frequency, mid frequency, and high frequency.

[0072] 2) Use a diagonal matrix to perform channel weighting on the cross-channel correlation coefficient matrix to generate a cross-channel feature fusion matrix.

[0073] The cross-channel correlation coefficient matrix is ​​weighted using a weight matrix determined by the frequency domain energy vector, and then concatenated with the basic feature matrix to form a cross-channel feature fusion matrix, expressed as:

[0074]

[0075] In the formula, The cross-channel feature fusion matrix represents the concatenation of the basic feature matrix and the weighted cross-channel correlation coefficient matrix. Represents the cross-channel correlation coefficient matrix; Represents the diagonal weight matrix; Indicates a splicing operation; This represents the Hadamard product, which is the element-wise multiplication of a matrix.

[0076] 3) Selectively filter the frequency domain features of the original exploration data, and superimpose the selectively filtered original exploration data with the cross-channel feature fusion matrix to generate the fused features of the original exploration data and the features to be measured after feature fusion.

[0077] In practical implementation, the gradient feature matrix is ​​transformed to the frequency domain, selective filtering is performed using a frequency band mask vector, and then inverse Fourier transform is applied back to the time domain. This result is then superimposed onto the cross-channel feature fusion matrix to form a fused feature matrix, represented as:

[0078]

[0079] In the formula, This represents the fused feature matrix, which includes basic features, cross-channel weighted features, and reconstructed gradient features, used to enhance high-frequency transient information; The gradient feature matrix contains the first-order gradient information of the original well logging data. The first-order gradient information of the original well logging data is the rate of change of the well logging value with depth, which is calculated as the difference between adjacent sampling points. The gradient feature matrix stores the gradient values ​​of all sampling points. This represents the Fourier transform operator, which transforms a time-domain signal into the frequency domain; This represents the inverse Fourier transform operator, which converts a frequency domain signal back to the time domain; This represents a frequency band mask vector, used for selective filtering of gradient features in the frequency domain, and is defined as follows: ; The weighting coefficients for low-frequency components control the contribution of low-frequency gradient information in the reconstruction. They are set based on geological priors, such as the stratigraphic trend corresponding to low frequencies. To enhance the low-frequency signal of lithological gradation; These represent the weighting coefficients of the intermediate frequency components, controlling the contribution of the intermediate frequency gradient information to the reconstruction. For example, the default weight for the intermediate frequency components can be set... ; The weighting coefficients for high-frequency components control the contribution of high-frequency gradient information in the reconstruction. For example, since high frequencies correspond to transient events, the weighting coefficients are set accordingly. To enhance the high-frequency signal of the fault.

[0080] By defining frequency band mask vectors This achieves frequency band selective enhancement by assigning weighting coefficients to high-frequency gradients that are sensitive to faults. Weighting coefficients were set for the low-frequency gradients related to lithological gradation. ,at the same time, The term represents the gradient features after the inverse Fourier transform, retaining only the geologically valid frequency band information, and then fused with the cross-channel feature matrix. Superposition can achieve a balance between "high-frequency transient enhancement" and "low-frequency trend fidelity", solving the problem of high-frequency noise drowning out real anomalies in conventional methods.

[0081] Furthermore, Figure 4This diagram illustrates a performance comparison of different anomaly detection methods, used to compare the detection accuracy of feature fusion and different data preprocessing methods. In one embodiment, the performance comparison of different anomaly detection methods can be performed by using a heatmap to quantitatively compare the detection accuracy of four processing techniques for four types of geological anomalies, and then comparing them with feature fusion methods. The horizontal axis represents the anomaly type (fault, fluid-bearing layer, lithological gradient, instrument malfunction), and the vertical axis represents the processing method (fixed-band wavelet, Fourier transform, conventional wavelet packet, and the technique of this invention). The color gradient from blue to red indicates accuracy from low to high. The most significant feature is that the row containing the technique of this invention shows a full deep red / yellow tone, which is superior to the conventional wavelet packet, which is the best performing conventional method. This indicates that the frequency domain energy constraint mechanism and feature fusion strategy of the technique of this invention effectively enhance the sensitivity to high-frequency transient anomalies while preserving the integrity of low-frequency gradient features.

[0082] Furthermore, this embodiment of the invention also provides a performance comparison diagram of different feature fusion methods, referring to... Figure 5 The impact of different feature fusion methods on model training can be evaluated by comparing and analyzing the effects of feature fusion. Experiments compared the performance changes of three methods during training: conventional feature concatenation, static weighted fusion, and the adaptive fusion method of this invention. The figure contains two sub-figures: the upper figure shows the accuracy change curve, and the lower figure shows the F1 score change curve; the horizontal axis represents the number of training epochs (unitless), and the vertical axis represents the index score (unitless). Each method is represented by a line with different markings, and a semi-transparent error band shows the training fluctuation range. As can be seen from the accuracy curve, the method of this invention rises rapidly in the early stages of training, stabilizing at a high level above 0.9 after about 30 epochs, with the narrowest error band. Static weighted fusion reaches around 0.9 after 40 epochs, and conventional feature concatenation (blue square line) is similar. The F1 score curve shows the same trend: the method of this invention eventually stabilizes above 0.9, significantly higher than other methods, proving that this invention effectively solves the problem of high-frequency transient features being masked by statistical features through dynamic weighting of the feature interaction matrix and selective filtering of frequency band masks. The narrow error band indicates that the method of this invention has more stable training and faster convergence speed.

[0083] Step S212: Based on the preset feature channel attention mechanism, adaptive feature selection is performed on the feature to be tested to determine the attention output feature matrix of the feature to be tested.

[0084] Different logging channels exhibit significantly different sensitivities to geological anomalies. Manual feature selection methods are highly subjective and cannot adapt to sample specificity, leading to the weakening or neglect of key anomaly response features. This invention constructs a lightweight attention mechanism to achieve adaptive optimization of sample-level feature channels. In its implementation, it includes the following steps:

[0085] 1) Input the features to be tested into a single hidden layer perceptron, and perform linear activation processing on the features to be tested based on the weight matrix of the single hidden layer perceptron with respect to the hidden layer.

[0086] 2) Normalize the feature channels of the features to be tested after linear activation processing to generate the attention weight matrix corresponding to the features to be tested.

[0087] In one implementation, the ReLU activation function is used for linear activation. This represents the rectified linear unit activation function. In this embodiment of the invention, the fused feature matrix (i.e., the features to be tested mentioned above) is input into a single hidden layer perceptron and processed using the rectified linear unit activation function; then, an attention weight matrix is ​​generated using the softmax function, and the feature channel weights of each sample are normalized, expressed as:

[0088]

[0089] In the formula, Represents the activation function of the rectified linear unit; This represents the weight matrix from the input layer to the hidden layer, with dimension 1. ; The number of hidden units is typically set to a value of [value to be filled in]. ; For the fusion feature matrix The number of feature channels, since the basic feature matrix contains 15 feature dimensions, then . This represents the bias vector of the hidden layer, with dimension . . This represents the weight matrix from the hidden layer to the output layer, with dimension 1. ; This represents the bias vector of the output layer, with dimension . . This represents the Softmax normalized exponential function. The item representation is normalized according to the sample dimension to ensure that the sum of all channel weights for each sample is 1. This represents the attention weight matrix, with dimensions equal to the attention input feature matrix. The same, its i-th row, the first The elements of a column represent elements. Indicates the first The first sample The importance weights of each feature channel satisfy the following conditions. ; For feature channel index, .

[0090] 3) Adaptively weight the features to be tested using the attention weight matrix to generate the attention output feature matrix corresponding to the features to be tested.

[0091] In the specific implementation, the attention weight matrix and the fused feature matrix (i.e., the features to be tested mentioned above) are multiplied element-wise to achieve adaptive weighting of feature channels, and the attention output feature matrix is ​​output as follows:

[0092]

[0093] In the formula, This represents the attention output feature matrix, with dimension 1. It should be noted that the feature selection module is a lightweight single-hidden-layer perceptron architecture, consisting of an input layer, hidden layers, and an output layer. This lightweight single-hidden-layer perceptron architecture achieves sample-level feature channel adaptation by calculating the attention weight matrix. Generate sample-specific weights, and define the first... The importance weight vector of each sample is: It dynamically allocates channel importance based on its fusion characteristics, such as automatically increasing the weight of neutron porosity channels in fluid layers, enabling the model to adjust the channel importance based on frequency band energy vectors. The implicit information, attention to the transient feature channels of high-frequency dominant samples, and the dynamic mechanism enable the model to learn the association rules of "channel-anomaly type" autonomously without explicit input of geological labels.

[0094] Furthermore, refer to Figure 6 This invention also provides a schematic diagram of the frequency domain energy vector distribution of different geological events, used to analyze the frequency domain energy vector distribution and display the distribution patterns of different geological events in the low-frequency, mid-frequency, and high-frequency energy space in three-dimensional space. The X-axis represents low-frequency energy (0-1 dimensionless ratio), the Y-axis represents mid-frequency energy, and the Z-axis represents high-frequency energy. Fault events (red dot cloud) are concentrated in areas with high-frequency energy greater than 0.5, forming obvious clusters; fluid-bearing layers (blue dot cloud) are mainly distributed in areas with higher mid-frequency energy; lithological gradations (green dot cloud) are concentrated in areas with higher low-frequency energy; while instrument noise (purple dot cloud) is randomly scattered throughout the space. Experimental results show that the frequency band energy vector constructed based on the frequency band distribution characteristics of geological events (faults correspond to high frequencies, and lithological gradations correspond to low frequencies) has significant characterization capabilities.

[0095] Step S214: Short-cycle feature extraction and long-cycle feature extraction are performed on the attention output feature matrix, and gated residual connections are made between the features extracted by the short-cycle feature extraction and the long-cycle feature extraction to generate a dual-path residual output feature matrix.

[0096] Geological anomaly patterns exist simultaneously in short-term instantaneous values ​​and long-term trends. Conventional single convolutional kernel structures struggle to extract features across different time scales, easily leading to the loss of short-term abrupt changes or long-term trend features. This invention employs parallel convolutional pathways and an adaptive residual connection mechanism to fuse short-term transient features and long-term trend features.

[0097] The steps for short-period feature extraction are as follows:

[0098] A depthwise separable convolution operator is applied to the attention output feature matrix, using a small kernel temporal size and unit stride to extract high-frequency transient features and output a short-period feature output matrix. The depthwise separable convolution operator is a lightweight convolution operation used to extract high-frequency transient features from the attention output feature matrix. It separates spatial (or temporal) convolution from channel convolution, reducing computational complexity while maintaining the accuracy of high-frequency feature extraction, thus meeting the detection requirements of "high-frequency transient anomalies such as faults" in well logging data. Specifically, it is expressed as follows:

[0099]

[0100] In the formula, This represents a depthwise separable convolution operator, where This indicates that the kernel time size is 3. This indicates a time step of 1, which enables lightweight design and is suitable for capturing local patterns. The time dimension of the convolution kernel; This represents the stride of the convolution operation; It is a short-period feature output matrix that characterizes high-frequency transient features.

[0101] Long-period feature extraction is as follows:

[0102] The attention output feature matrix is ​​first subjected to standard convolution to expand the receptive field, and then max pooling is performed to downsample and extract low-frequency trend features, outputting a long-period feature output matrix, as follows:

[0103]

[0104] In the formula, Denotes the standard convolution operator, where This indicates that the kernel time size is 7. This indicates a time step of 2, used to expand the receptive field to capture longer-period patterns; Denotes the max pooling operator, where This indicates that the pooling window time size is 5, and further sampling and extraction of significant features are performed by reducing the resolution to extract coarse-grained features. This represents the long-period feature output matrix, characterizing low-frequency trend features. Represents operator combination symbols, The term indicates that standard convolution is performed first, followed by max pooling.

[0105] Furthermore, the short-period feature output matrix and the long-period feature output matrix are added together, processed by the rectified linear unit activation function, and then added to the attention output feature matrix filtered by the gated residual function to form a dual-path residual output feature matrix, expressed as:

[0106]

[0107] In the formula, This represents the gating weight matrix, which consists of trainable parameters used to learn the gating signal. This represents the Sigmoid activation function, which compresses the input to the range (0,1). This represents element-wise multiplication; Represents the gated residual function. Item-to-attention output feature matrix Perform adaptive gating filtering to retain important information; This is the feature matrix of the dual-path residual output.

[0108] It should be noted that the structure of the dual-path residual temporal convolution module is a dual-path parallel structure. The short-period path uses depthwise separable convolution to capture transient features, while the long-period path uses standard convolution and pooling to capture trend features. Then, gated residual connections are used to fuse the dual-path outputs and retain key information.

[0109] It should also be noted that by first outputting the short-period feature matrix... and long-period feature output matrix After summing and re-activating with ReLU, the summation is then added to the attention output feature matrix filtered by the gated residual function. The resulting residual connection is the dual-path residual output feature matrix. .

[0110] It should also be noted that, The representation utilizes learnable gating to perform soft selection on the original features, retaining only those components important to the current task. Simultaneously, it combines dual-path convolutional output to achieve triple feature fusion, resulting in a short-period feature output matrix. Long-period feature output matrix and key original features after gated filtering It is suitable for sampling points with large depth intervals. When the convolution path loses details due to downsampling, the gating mechanism can supplement the high-frequency abnormal signals in the original features.

[0111] Step S216: Based on the frequency band energy vector, perform frequency band modulation processing on the dual-path residual output feature matrix to generate a frequency domain modulation feature matrix.

[0112] The probability of geological anomalies is affected by the energy distribution across frequency bands, and different channels contribute differently to the discrimination of anomaly types. Conventional fully connected layers that directly map probabilities weaken this crucial information. This invention employs a frequency domain modulation layer and an attention gating mechanism, combined with the optimal subtree node set, to generate the final predicted probability distribution vector.

[0113] In a specific implementation, the execution steps of the frequency-domain modulation fully connected layer in this embodiment of the invention are as follows:

[0114] The Hadamard product is calculated between the dual-path residual output feature matrix and the projected frequency band energy vector. This product is then input into a fully connected layer, processed using the ELU activation function, and the output frequency domain modulation feature matrix is ​​expressed as:

[0115]

[0116] In the formula, This represents the frequency domain modulation feature matrix, with dimension 1. ; ELU activation function, used to enhance nonlinearity; The frequency domain modulation weight matrix is ​​a trainable parameter with dimension . ; Let be the frequency domain modulation bias vector, which are trainable parameters with dimension . ; The feature matrix of the dual-path residual output The number of feature channels; For frequency band energy projection operators, The term will be a three-dimensional frequency band energy vector. Mapped to the dual-path residual output feature matrix For the same dimension, the mapping method is represented as : The frequency band modulation matrix has a dimension of . The row vector of the d-th dimension of the frequency band modulation matrix is Control the first The sensitivity of dimensional features to low-frequency / mid-frequency / high-frequency energy; For the first The sensitivity weights of each feature channel to low-frequency energy are trainable parameters. For the first The sensitivity weights of each feature channel to mid-frequency energy are trainable parameters; For the first The sensitivity weights of each feature channel to high-frequency energy are trainable parameters. Copy along the timeline Dimensional vector to One sample; It represents the frequency band energy vector, which includes the aggregated energy ratio of the three dimensions of low frequency, mid frequency, and high frequency; for The transpose of .

[0117] It should be noted that the frequency band energy vector Output after projection by the frequency band energy projection operator A dimensional vector, but the dual-path residual output feature matrix The dimension is ,need Performing the Hadamard product operation on a 3D matrix by copying along the time axis Dimensional vector to Generate samples A dimensional matrix ensures that each sample shares the same frequency band energy weights, guaranteeing that the frequency band energy vector can modulate all sample features and enhance the model's utilization of frequency domain priors.

[0118] It should also be noted that, through a trainable frequency band modulation matrix Learn the sensitivity of feature channels to each frequency band, and at the same time, The term enables each feature channel to be modulated by the corresponding frequency band energy. When the high-frequency energy is significant, it indicates a fault. The model automatically increases the response intensity of the high-frequency sensitive channel. The coupling mechanism of "frequency domain-feature domain" enables the subsequent fully connected layer to make more accurate use of the frequency domain prior of geological events.

[0119] Step S218: Perform attention-gated probability transformation on the frequency domain modulation feature matrix to determine the abnormal probability distribution corresponding to the feature to be tested.

[0120] Step S220: Determine the anomaly identification result corresponding to the feature to be tested based on the anomaly probability distribution.

[0121] The frequency domain modulation feature matrix is ​​weighted and summed using an extended attention weight matrix, and then the original probability distribution vector is generated using the Softmax function, as follows:

[0122]

[0123] In the formula, For the first The original probability distribution vector of each sample is a 4-dimensional predicted probability distribution vector, corresponding to 4 types of anomalies, namely faults, fluid-bearing layers, instrument malfunctions, and lithological gradations. This represents the Softmax normalized exponential function. The item representation is normalized along the category dimension. To expand the attention matrix, a pre-defined single-hidden-layer perceptron network is used, employing the ReLU activation function, with the input feature being a frequency-domain modulation feature matrix. The output features are the extended attention matrix. Expand the attention matrix in the i-th row. Column elements are , indicating the first The first sample The importance weights of each feature channel are determined by the fact that a higher weight indicates that the channel is more sensitive to anomaly detection in the current sample. Frequency domain modulation feature matrix The i-th row The column elements are the feature elements processed by the frequency domain modulation fully connected layer.

[0124] Furthermore, in this embodiment of the invention, an adaptive temperature coefficient is calculated based on the average information entropy of the optimal subtree node set, and then the original probability distribution is adjusted to generate a predicted probability distribution vector, represented as:

[0125]

[0126] In the formula, For the first The predicted probability distribution vector of each sample; For the first The original probability distribution vector of each sample; Number of categories; For category indexing, ; For the first The adaptive temperature coefficient for each sample. The calculation method is expressed as ; Indicates the first The set of nodes contained in the optimal subtree of the sample Average information entropy; For entropy-sensitive factors, such as, Controlling the average information entropy for adaptive temperature coefficient The intensity of the impact.

[0127] It should be noted that the set of nodes contained in the optimal subtree is used. The average information entropy on the adaptive temperature coefficient is dynamically adjusted. This allows high-entropy samples with complex signals to correspond to a larger adaptive temperature coefficient. By smoothing the probability distribution, low-entropy samples with pure signals reduce the adaptive temperature coefficient. Sharpening predictions can solve the problem of misjudgment in geologically ambiguous areas. For example, in lithological gradient zones with high average information entropy, the model outputs a relatively flat probability distribution, avoiding overconfident predictions of the "fault / normal" category. At clear faults with low average information entropy, it outputs a high-confidence anomalous probability.

[0128] Step S222: Based on the anomaly identification results, determine the anomaly category to which the target exploration section belongs.

[0129] In conjunction with the above steps, this embodiment of the invention also uses a sliding window method to determine the anomaly confidence level corresponding to the anomaly identification results of the target exploration segment in each area to be tested. Based on the cumulative result of the anomaly confidence level, the anomaly category to which the target exploration segment belongs is determined. Through the above steps S212-S220, the features to be tested are identified, and the probability distribution vector of multiple anomaly categories for the target exploration segment is output, such as the following four types of anomalies: fault / fluid-bearing layer / lithological gradient / instrument malfunction. These categories can be determined by data annotation on their training sample set. In one embodiment, data annotation can be achieved manually. Specifically, the location of geological anomalies can be identified based on borehole core data, seismic profiles, and formation dip logging data, and four types of labels can be labeled: fault, corresponding to high-frequency abrupt signal within the depth range; fluid-bearing layer, which is the fluid response in a specific lithological combination; lithological gradient, which is a zone of slow change in physical parameters; and instrument malfunction, which is noise interference of non-geological origin. In the specific implementation, the well logging profile is continuously processed in a sliding window manner (window length of 512 sampling points, step size of 64 points). For each depth point, the anomaly confidence score in the interval [0,1] is output. When the confidence score of a specific anomaly category continuously exceeds the depth threshold, such as a fault confidence score greater than 0.7 and a continuous depth greater than 0.5 meters, anomaly event marking is automatically triggered to determine the final anomaly category.

[0130] In summary, this invention proposes an artificial intelligence-based anomaly detection method for exploration data, which is innovative compared to existing technologies in the following aspects: 1. This invention selects the optimal subtree by maximizing entropy and adaptively determines the frequency band, effectively enhancing the distinction between high-frequency faults and low-frequency lithological gradients. 2. Existing methods only use energy magnitude as a feature, failing to distinguish meaningful high-energy signals from noise energy, resulting in false anomalies or missed detections. This invention combines energy and information entropy, nonlinearly amplifying low-entropy high-frequency signals, such as transient events like faults, while suppressing high-entropy noise, such as instrument interference, thus improving anomaly detection sensitivity. 3. Conventional methods often involve feature splicing or static weighting, ignoring cross-channel correlation and spatiotemporal coupling, making high-frequency transient features easily masked by statistical features. This invention integrates basic features, cross-channel correlation, and frequency domain enhanced gradient features to construct a dynamically weighted interaction matrix, taking into account both high-frequency transients and low-frequency trends, avoiding the masking of anomalies by a single feature. 4. Conventional convolutional neural network models often rely on a single convolutional kernel or a fixed receptive field, resulting in an imbalance between short-period and long-period feature extraction, easily leading to the loss of anomaly information. This invention employs sample-level channel attention to achieve adaptive channel weighting, dual-path convolution to simultaneously extract short-period transient features and long-period trend features, and gated residuals to preserve key original signals.

[0131] Furthermore, this invention also proposes a model training method for the aforementioned artificial intelligence model. By acquiring a pre-constructed training sample set, a mini-batch gradient descent strategy is used to input the training sample set into the artificial intelligence model for training. In specific implementation, a fixed number of samples, such as 256 depth points, are input in each batch. During the forward propagation phase, the aforementioned frequency-domain energy-constrained wavelet packet decomposition, multi-source feature adaptive fusion, channel attention feature selection, dual-path residual convolution, and joint probability estimation modules are executed sequentially to generate a predicted probability distribution. Further, during the backpropagation phase, the gradient is calculated based on the joint loss function, and the learning rate is dynamically adjusted to update the network parameters using an adaptive moment estimation optimizer. The adaptive moment estimation optimizer is an adaptive optimizer that combines two classic optimization ideas: momentum gradient descent and adaptive learning rate. During training, the validation set performance is monitored in real time. When the validation set loss does not decrease or the anomaly classification F1 score fluctuates by less than 0.5% for 10 consecutive training rounds, an early stopping mechanism is triggered to save the optimal model parameters. Based on the gradient calculated by the joint loss function, the update step size (learning rate) of each network parameter is dynamically adjusted to ensure that the model can quickly converge to the optimal parameters during training, while avoiding overfitting or under-convergence caused by a fixed learning rate. This adapts to the complex training requirements of multiple modules (channel attention, dual-path convolution, frequency domain modulation) and multiple parameters in the well logging data anomaly detection model.

[0132] Well logging anomalies exhibit depth continuity, and the cost of misjudging different anomaly types varies significantly. The standard cross-entropy loss function ignores spatial continuity constraints and cost sensitivity, easily leading to misjudgments of isolated points and missed detections of high-cost anomalies. This invention integrates depth continuity constraints and dynamic cost weights to construct a loss function, combining weighted cross-entropy loss, a depth continuity penalty term, and a regularization term to form a joint loss function. In specific implementation, the steps for constructing the joint loss function in this embodiment are as follows:

[0133] a- Construct a depth continuity penalty term for the training sample set based on the JS divergence of the predicted probability distributions of neighboring samples and the absolute value of the depth difference between neighboring samples.

[0134] Construction of the depth continuity penalty term: Calculate the JS divergence of the predicted probability distributions of adjacent samples, combine it with the absolute value of the depth difference, and apply a depth decay factor to form a depth continuity penalty term, which encourages consistency in predictions along the depth direction, expressed as:

[0135]

[0136] In the formula, This represents a depth continuity penalty term, used to penalize situations where the predicted probability distributions of adjacent depth points differ too much, encouraging the model to predict depth continuity. Indicates the number of samples in the batch training; The model represents the first The predicted probability distribution vector of each sample is a 4-dimensional predicted probability distribution vector, corresponding to 4 types of anomalies, namely faults, fluid-bearing layers, instrument malfunctions, and lithological gradations. The model represents the first The predicted probability distribution vector of each sample; This represents the JS divergence, used to measure the difference between two probability distributions; Indicates the first The depth value of each sample; Indicates the first The depth value of each sample, Characterizes the absolute value of the depth difference between adjacent samples; This represents the depth attenuation coefficient, which controls the rate at which the continuity constraint weakens as the depth difference increases. A larger value indicates a more significant impact of the depth difference. It should be noted that, through The term implements adaptive penalty strength, such as when the depth difference between adjacent samples is small. If the forced probability distributions are similar, then when the depth difference is large, such as This allows for distribution jumps, a strategy that aligns with geological stratum variation patterns, maintaining predictive consistency within homogeneous strata while tolerating abrupt changes at strata interfaces.

[0137] b - Obtain the misjudgment cost matrix corresponding to the training sample set, and calculate the cost-sensitive weights corresponding to the training sample set based on the misjudgment cost matrix.

[0138] The misclassification cost is obtained from the cost matrix based on the true and predicted labels, and the cost-sensitive weight is calculated by combining the class sample frequency, as follows:

[0139]

[0140] In the formula, Indicates the first The cost-sensitive weights of each sample are used in the weighted cross-entropy loss term; Indicates according to the first The true label of each sample and the Predicted label for each sample Extracting the corresponding misjudgment cost from the cost matrix is... The cost matrix, where row indices correspond to the true class labels and column indices correspond to the predicted class labels, is used to define the cost of misclassification between different classes. For example, the cost matrix is ​​defined as follows: , of which elements A value of 0.3 indicates the cost of mispredicting a sample whose true class is the second class as the third class. The diagonal elements indicate that the prediction is correct and there is no cost. Indicates the total number of samples; Indicates the first The true label of each sample is a scalar value, which can be 1, 2, 3 or 4; Indicates the first The predicted label for each sample is a scalar value, which can be 1, 2, 3 or 4; Indicates the first The number of samples corresponding to the true category of each sample; c is to distinguish from Category index, It should be noted that, The characterization cost matrix defines the geological requirements, while The adaptive boosting of minority class weights, and the product of the two, gives high-cost and sparse samples a significantly higher loss weight, solving the problem of fixed weights being out of sync with sample distribution in conventional cost-sensitive learning, and improving the model’s recall rate for “high-cost-low-frequency” anomalies.

[0141] c-We use cost-sensitive weights to weight the cross-entropy loss corresponding to the training sample set, and combine it with a depth continuity penalty term to construct a joint loss function for the training sample set.

[0142] The weighted cross-entropy loss, depth continuity penalty term, and L2 regularization term are combined to form a joint loss function, which serves as the model optimization objective, expressed as:

[0143]

[0144] In the formula, Represents the joint loss function; Indicates the first One-hot encoded vectors of the true labels of each sample; The model represents the first The predicted probability distribution vector of each sample; The intensity coefficient representing the depth continuity penalty term is set by prior geological knowledge and is calculated as follows: ; It is the expected average thickness of a typical geological anomaly layer, in meters, set by domain knowledge, such as the average thickness of a fluid-containing layer being 2 meters; It is the sampling interval of well logging data, which makes the penalty intensity related to the actual geological scale. The unit is meters, which comes from the process parameters of data acquisition, such as sampling once every 0.1 meters. The coefficients of the L2 regularization term are used to prevent the model from overfitting. Represents all trainable parameters of the model The squared L2 norm is used as a weight decay term; This represents all trainable parameters of the model, in set form. It should be noted that in the joint loss function... In the calculation process, the intensity coefficient of the depth continuity penalty term is... Compared with the expected average thickness of typical geological anomaly layers Sampling interval of logging data Binding allows the constraint strength to adapt to the data sampling density, when The continuity constraint is enhanced when the data is in a certain state and weakened when it is in a certain state, thereby giving the model the ability to perceive geological scales and maintain optimal performance on logging data with different sampling densities.

[0145] In summary, the embodiments of the present invention construct a joint loss function based on the adaptive temperature calibration probability distribution of entropy, combined with depth continuity penalty and cost-sensitive weights, which can effectively reduce false positives and high-cost anomaly misses.

[0146] Furthermore, based on the above embodiments, this invention also provides an artificial intelligence-based exploration data anomaly detection device. Figure 7 This diagram illustrates the structure of an artificial intelligence-based exploration data anomaly detection device according to an embodiment of the present invention. (Refer to...) Figure 7 The device includes: a data acquisition module 100, used to acquire the formation physical response of the target exploration section, record the depth coordinates, instrument status, and environmental parameters of the formation physical response, and generate raw exploration data of the target exploration section; a data processing module 200, used to determine the adaptive frequency band energy ratio of the raw exploration data based on the information entropy of the raw exploration data; and to determine the frequency band energy vector contained in the raw exploration data based on the adaptive frequency band energy ratio; the frequency band energy vector includes low-frequency vector, mid-frequency vector, and high-frequency vector; and a feature fusion module 300, used to perform feature fusion based on the raw exploration data. The method uses well logging values ​​and window statistics to construct a basic feature matrix from the original exploration data. Based on the cross-channel correlation of the basic feature matrix, it uses frequency band energy vectors to perform cross-channel feature fusion on the basic feature matrix, generating fused features from the original exploration data. Execution module 400 inputs the fused features as features to be tested into a pre-built artificial intelligence model, which then performs anomaly identification on the features to be tested to determine the anomaly identification result corresponding to the target exploration layer. Output module 500 determines the anomaly category of the target exploration layer based on the anomaly identification result. The exploration data anomaly detection device based on artificial intelligence provided in this embodiment of the invention has the same technical features as the above method embodiment, and therefore can solve the same technical problems and achieve the same technical effects.

[0147] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described... Figures 1 to 2 The steps of the method are shown. Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the above-described steps. Figures 1 to 2 The steps of the method are shown. This invention also provides a schematic diagram of the structure of an electronic device, as shown. Figure 8 The diagram shows the structure of the electronic device, which includes a processor 81 and a memory 80. The memory 80 stores computer-executable instructions that can be executed by the processor 81. The processor 81 executes the computer-executable instructions to implement the above-mentioned... Figures 1 to 2 The method shown. In Figure 8 In the illustrated embodiment, the electronic device further includes a bus 82 and a communication interface 83, wherein the processor 81, the communication interface 83, and the memory 80 are connected via the bus 82. The memory 80 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk drive. Communication between this system network element and at least one other network element is achieved through at least one communication interface 83 (which can be wired or wireless), using the Internet, wide area network, local area network, metropolitan area network, etc. Bus 82 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, or an AMBA (Advanced Microcontroller Bus Architecture) bus. AMBA defines three types of buses: APB (Advanced Peripheral Bus), AHB (Advanced High-performance Bus), and AXI (Advanced deXtensible Interface). Bus 82 can be divided into address bus, data bus, and control bus, etc. For ease of representation, Figure 8 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0148] Processor 81 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 81 or by instructions in software form. The processor 81 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in the memory. The processor 81 reads the information in the memory and, in conjunction with its hardware, completes the aforementioned task. Figures 1 to 2 Any of the methods shown.

[0149] The computer program product of the exploration data anomaly detection method and apparatus based on artificial intelligence provided in this invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. Specific implementations can be found in the method embodiments and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. Furthermore, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0150] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An artificial intelligence-based method for detecting anomalies in exploration data, characterized in that, The method includes: Acquire the formation physical response of the target exploration section, record the depth coordinates, instrument status and environmental parameters of the formation physical response, and generate the raw exploration data of the target exploration section; Based on the information entropy of the original exploration data, the adaptive frequency band energy ratio of the original exploration data is determined; based on the adaptive frequency band energy ratio, the frequency band energy vector contained in the original exploration data is determined; the frequency band energy vector includes a low-frequency vector, a mid-frequency vector, and a high-frequency vector; Based on the normalized logging values ​​and window statistics of the original exploration data, a basic feature matrix is ​​constructed for the original exploration data; according to the cross-channel correlation of the basic feature matrix, the frequency band energy vector is used to perform cross-channel feature fusion on the basic feature matrix to generate the fused features of the original exploration data. The fused features are input as features to be tested into a pre-built artificial intelligence model. The artificial intelligence model is used to identify anomalies in the features to be tested, and the anomaly identification result corresponding to the target exploration layer is determined. Based on the anomaly identification results, the anomaly category to which the target exploration section belongs is determined; The step of performing cross-channel feature fusion on the basic feature matrix based on the frequency band energy vector and the cross-channel correlation of the basic feature matrix to generate the fused features of the original exploration data includes: Calculate the cross-channel correlation coefficient matrix of the basic feature matrix, and construct the diagonal matrix of the frequency band energy vector; The cross-channel correlation coefficient matrix is ​​weighted by the diagonal matrix to generate a cross-channel feature fusion matrix; Selective filtering is applied to the frequency domain features of the original exploration data. The selectively filtered original exploration data is then superimposed with the cross-channel feature fusion matrix to generate the fused features of the original exploration data and the features to be measured after feature fusion. The step of inputting the fused features as features to be tested into a pre-built artificial intelligence model, and using the artificial intelligence model to perform anomaly identification on the features to be tested, and determining the anomaly identification result corresponding to the target exploration layer, includes: Based on a preset feature channel attention mechanism, adaptive feature selection is performed on the feature to be tested to determine the attention output feature matrix of the feature to be tested. Short-cycle feature extraction and long-cycle feature extraction are performed on the attention output feature matrix respectively, and gated residual connections are made between the features extracted by short-cycle feature extraction and long-cycle feature extraction to generate a dual-path residual output feature matrix; Based on the frequency band energy vector, the dual-path residual output feature matrix is ​​subjected to frequency band modulation processing to generate a frequency domain modulation feature matrix; The frequency domain modulation feature matrix is ​​subjected to attention-gated probability transformation to determine the anomaly probability distribution corresponding to the feature to be tested; Based on the anomaly probability distribution, determine the anomaly identification result corresponding to the feature to be tested; The step of adaptively selecting features for the features to be tested based on a preset feature channel attention mechanism and determining the attention output feature matrix of the features to be tested includes: The feature to be tested is input into a single hidden layer perceptron, and linear activation processing is performed on the feature to be tested based on the weight matrix of the single hidden layer perceptron with respect to the hidden layer. The feature channel of the linear activation processing is normalized to generate the attention weight matrix corresponding to the feature. The attention weight matrix is ​​used to adaptively weight the feature to be tested, thereby generating the attention output feature matrix corresponding to the feature to be tested.

2. The method according to claim 1, characterized in that, Based on the anomaly identification results, the step of determining the anomaly category to which the target exploration section belongs includes: The anomaly confidence level corresponding to the anomaly identification result of the target exploration segment in each area to be tested is determined by using a sliding window method. Based on the cumulative results of the anomaly confidence scores, the anomaly category to which the target exploration section belongs is determined.

3. The method according to claim 1, characterized in that, The step of determining the adaptive frequency band energy ratio of the original exploration data based on the information entropy of the original exploration data includes: The original exploration data is decomposed using wavelet packet decomposition to divide it into multiple layers; each layer contains multiple frequency band nodes. Calculate the information entropy of each frequency band node, and based on the information entropy, determine the optimal subtree node set from the decomposition tree to which the frequency band node belongs; The node energy of the current frequency band node is calculated based on the wavelet packet coefficients of each frequency band node in the optimal subtree node set. Based on the node energy and the information entropy, the adaptive frequency band energy ratio of the original exploration data for each frequency band node is determined.

4. The method according to claim 1, characterized in that, The step of extracting short-period features from the attention output feature matrix includes: A depthwise separable convolution operator is applied to extract short-period features from the attention output feature matrix, and outputs a short-period feature output matrix of the attention output feature matrix; the depthwise separable convolution operator uses a small convolution kernel time size and a unit stride for feature extraction; The step of extracting long-period features from the attention output feature matrix includes: After performing standard convolution and max pooling operations on the attention output feature matrix in sequence, feature extraction is performed on the attention output feature matrix to output the long-period feature output matrix of the attention output feature matrix.

5. The method according to claim 1, characterized in that, The method further includes: Obtain a pre-built training sample set; The training sample set is input into the artificial intelligence model using a mini-batch gradient descent strategy for model training. The training gradient is calculated using a pre-built joint loss function, and the learning rate of the artificial intelligence model is dynamically adjusted using a preset adaptive moment estimator to update the network parameters of the artificial intelligence model.

6. The method according to claim 5, characterized in that, The method further includes: Based on the JS divergence of the predicted probability distribution of neighboring samples in the training sample set and the absolute value of the depth difference between neighboring samples, a depth continuity penalty term for the training sample set is constructed. Obtain the misjudgment cost matrix corresponding to the training sample set, and calculate the cost-sensitive weights corresponding to the training sample set based on the misjudgment cost matrix; The cost-sensitive weights are used to weight the cross-entropy loss corresponding to the training sample set, and combined with the depth continuity penalty term, a joint loss function for the training sample set is constructed.

7. An exploration data anomaly detection device based on artificial intelligence, characterized in that, The device includes: The data acquisition module is used to acquire the formation physical response of the target exploration section, record the depth coordinates, instrument status and environmental parameters of the formation physical response, and generate the raw exploration data of the target exploration section. The data processing module is used to determine the adaptive frequency band energy ratio of the original exploration data based on the information entropy of the original exploration data; and to determine the frequency band energy vector contained in the original exploration data based on the adaptive frequency band energy ratio; the frequency band energy vector includes a low-frequency vector, a mid-frequency vector, and a high-frequency vector. The feature fusion module is used to construct a basic feature matrix for the original exploration data based on the normalized logging values ​​and window statistics of the original exploration data; and to perform cross-channel feature fusion on the basic feature matrix using the frequency band energy vector according to the cross-channel correlation of the basic feature matrix to generate the fused features of the original exploration data. The execution module is used to input the fused features as features to be tested into a pre-built artificial intelligence model, and to perform anomaly identification on the features to be tested through the artificial intelligence model to determine the anomaly identification result corresponding to the target exploration layer. The output module is used to determine the anomaly category to which the target exploration section belongs based on the anomaly identification results; The feature fusion module is further configured to: calculate the cross-channel correlation coefficient matrix of the basic feature matrix, and construct the diagonal matrix of the frequency band energy vector; use the diagonal matrix to perform channel weighting on the cross-channel correlation coefficient matrix to generate a cross-channel feature fusion matrix; selectively filter the frequency domain features of the original exploration data, and superimpose the selectively filtered original exploration data with the cross-channel feature fusion matrix to generate the fused features of the original exploration data, and the feature to be measured after feature fusion; The execution module is further configured to: perform adaptive feature selection on the feature to be tested based on a preset feature channel attention mechanism to determine the attention output feature matrix of the feature to be tested; perform short-period feature extraction and long-period feature extraction on the attention output feature matrix respectively, and perform gated residual connection on the features extracted by the short-period feature extraction and long-period feature extraction to generate a dual-path residual output feature matrix; perform frequency band modulation processing on the dual-path residual output feature matrix based on the frequency band energy vector to generate a frequency domain modulation feature matrix; perform attention-gated probability transformation on the frequency domain modulation feature matrix to determine the abnormal probability distribution corresponding to the feature to be tested; and determine the abnormal identification result corresponding to the feature to be tested based on the abnormal probability distribution. The execution module is further configured to: input the feature to be tested into a single hidden layer perceptron; perform linear activation processing on the feature to be tested based on the weight matrix of the single hidden layer perceptron with respect to the hidden layer; perform feature channel normalization on the linearly activated feature to be tested to generate an attention weight matrix corresponding to the feature to be tested; and use the attention weight matrix to adaptively weight the feature to be tested to generate an attention output feature matrix corresponding to the feature to be tested.

Images